all: release v0.17.0

transform: show better error message in coroutines lowering
A common error is when someone tries to export a blocking function. This is not possible with the coroutines scheduler. Previously, it resulted in an error like this: panic: trying to make exported function async: messageHandler With this change, the error is much better and shows where it comes from exactly: /home/ayke/tmp/export-async.go:8: blocking operation in exported function: messageHandler traceback: messageHandler /home/ayke/tmp/export-async.go:9:5 main.foo /home/ayke/tmp/export-async.go:15:2 runtime.chanSend /home/ayke/src/github.com/tinygo-org/tinygo/src/runtime/chan.go:494:12 This should make it easier to identify and fix the problem. And it avoids a compiler panic, which is a really bad way of showing diagnostics.
2026-08-01 17:47:46 +00:00 · 2021-03-05 17:41:06 +01:00 · 2021-03-05 14:42:43 +01:00 · 2021-03-05 13:02:01 +01:00 · 2021-03-05 11:16:42 +01:00 · 2021-03-04 17:25:22 +01:00
626 changed files with 42258 additions and 11094 deletions
@@ -14,20 +14,21 @@ commands:
      - run:
          name: "Install apt dependencies"
          command: |
-            echo 'deb http://apt.llvm.org/stretch/ llvm-toolchain-stretch<<parameters.llvm>> main' | sudo tee /etc/apt/sources.list.d/llvm.list
+            echo 'deb https://apt.llvm.org/buster/ llvm-toolchain-buster-<<parameters.llvm>> main' | sudo tee /etc/apt/sources.list.d/llvm.list
            wget -O - https://apt.llvm.org/llvm-snapshot.gpg.key|sudo apt-key add -
            sudo apt-get update
            sudo apt-get install \
-                llvm<<parameters.llvm>>-dev \
-                clang<<parameters.llvm>> \
-                libclang<<parameters.llvm>>-dev \
-                lld<<parameters.llvm>> \
+                llvm-<<parameters.llvm>>-dev \
+                clang-<<parameters.llvm>> \
+                libclang-<<parameters.llvm>>-dev \
+                lld-<<parameters.llvm>> \
                gcc-arm-linux-gnueabihf \
                gcc-aarch64-linux-gnu \
                qemu-system-arm \
                qemu-user \
                gcc-avr \
                avr-libc
+            sudo apt-get install --no-install-recommends libc6-dev-i386 lib32gcc-8-dev
  install-node:
    steps:
      - run:
@@ -37,59 +38,91 @@ commands:
            sudo tar -C /usr/local -xf node-v10.15.1-linux-x64.tar.xz
            sudo ln -s /usr/local/node-v10.15.1-linux-x64/bin/node /usr/bin/node
            rm node-v10.15.1-linux-x64.tar.xz
+  install-chrome:
+    steps:
+      - run:
+          name: "Install Chrome"
+          command: |
+            wget https://dl.google.com/linux/direct/google-chrome-stable_current_amd64.deb
+            sudo apt install ./google-chrome-stable_current_amd64.deb
+  install-wasmtime:
+    steps:
+      - run:
+          name: "Install wasmtime"
+          command: |
+            curl https://wasmtime.dev/install.sh -sSf | bash
+            sudo ln -s ~/.wasmtime/bin/wasmtime /usr/local/bin/wasmtime
+  install-xtensa-toolchain:
+    parameters:
+      variant:
+        type: string
+    steps:
+      - run:
+          name: "Install Xtensa toolchain"
+          command: |
+            curl -L https://github.com/espressif/crosstool-NG/releases/download/esp-2020r2/xtensa-esp32-elf-gcc8_2_0-esp-2020r2-<<parameters.variant>>.tar.gz -o xtensa-esp32-elf-gcc8_2_0-esp-2020r2-<<parameters.variant>>.tar.gz
+            sudo tar -C /usr/local -xf xtensa-esp32-elf-gcc8_2_0-esp-2020r2-<<parameters.variant>>.tar.gz
+            sudo ln -s /usr/local/xtensa-esp32-elf/bin/xtensa-esp32-elf-ld /usr/local/bin/xtensa-esp32-elf-ld
+            rm xtensa-esp32-elf-gcc8_2_0-esp-2020r2-<<parameters.variant>>.tar.gz
  llvm-source-linux:
    steps:
      - restore_cache:
          keys:
-            - llvm-source-9-v0
+            - llvm-source-11-v1
      - run:
          name: "Fetch LLVM source"
          command: make llvm-source
      - save_cache:
-          key: llvm-source-9-v0
+          key: llvm-source-11-v1
          paths:
            - llvm-project
  build-wasi-libc:
    steps:
      - restore_cache:
          keys:
-            - wasi-libc-sysroot-v0
+            - wasi-libc-sysroot-v4
      - run:
          name: "Build wasi-libc"
-          command: make wasi-libc CLANG=$PWD/llvm-build/bin/clang LLVM_AR=$PWD/llvm-build/bin/llvm-ar LLVM_NM=$PWD/llvm-build/bin/llvm-nm
+          command: make wasi-libc
      - save_cache:
-          key: wasi-libc-sysroot-v0
+          key: wasi-libc-sysroot-v4
          paths:
            - lib/wasi-libc/sysroot
  test-linux:
+    parameters:
+      llvm:
+        type: string
    steps:
      - checkout
      - submodules
      - apt-dependencies:
-          llvm: "-9"
+          llvm: "<<parameters.llvm>>"
      - install-node
+      - install-chrome
+      - install-wasmtime
      - restore_cache:
          keys:
            - go-cache-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_PREVIOUS_BUILD_NUM }}
            - go-cache-v2-{{ checksum "go.mod" }}
      - llvm-source-linux
-      - run: go install .
+      - run: go install -tags=llvm<<parameters.llvm>> .
      - restore_cache:
          keys:
-            - wasi-libc-sysroot-systemclang-v0
+            - wasi-libc-sysroot-systemclang-v3
      - run: make wasi-libc
      - save_cache:
-          key: wasi-libc-sysroot-systemclang-v0
+          key: wasi-libc-sysroot-systemclang-v3
          paths:
            - lib/wasi-libc/sysroot
-      - run: go test -v ./cgo ./compileopts ./interp ./transform .
+      - run: go test -v -tags=llvm<<parameters.llvm>> ./cgo ./compileopts ./compiler ./interp ./transform .
      - run: make gen-device -j4
-      - run: make smoketest
+      - run: make smoketest XTENSA=0
+      - run: make tinygo-test
+      - run: make wasmtest
      - save_cache:
          key: go-cache-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_BUILD_NUM }}
          paths:
            - ~/.cache/go-build
-            - ~/.cache/tinygo
            - /go/pkg/mod
      - run: make fmt-check
  assert-test-linux:
@@ -99,9 +132,9 @@ commands:
      - run:
          name: "Install apt dependencies"
          command: |
+            sudo apt-get update
            sudo apt-get install \
                gcc-arm-linux-gnueabihf \
-                binutils-arm-none-eabi \
                libc6-dev-armel-cross \
                gcc-aarch64-linux-gnu \
                libc6-dev-arm64-cross \
@@ -109,7 +142,11 @@ commands:
                qemu-user \
                gcc-avr \
                avr-libc
+            sudo apt-get install --no-install-recommends libc6-dev-i386 lib32gcc-6-dev
      - install-node
+      - install-wasmtime
+      - install-xtensa-toolchain:
+          variant: "linux-amd64"
      - restore_cache:
          keys:
            - go-cache-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_PREVIOUS_BUILD_NUM }}
@@ -117,17 +154,14 @@ commands:
      - llvm-source-linux
      - restore_cache:
          keys:
-            - llvm-build-9-linux-v0-assert
+            - llvm-build-11-linux-v2-assert
      - run:
          name: "Build LLVM"
          command: |
            if [ ! -f llvm-build/lib/liblldELF.a ]
            then
              # install dependencies
-              sudo apt-get install cmake clang ninja-build
-              # make build faster
-              export CC=clang
-              export CXX=clang++
+              sudo apt-get install cmake ninja-build
              # hack ninja to use less jobs
              echo -e '#!/bin/sh\n/usr/bin/ninja -j3 "$@"' > /go/bin/ninja
              chmod +x /go/bin/ninja
@@ -135,7 +169,7 @@ commands:
              make ASSERT=1 llvm-build
            fi
      - save_cache:
-          key: llvm-build-9-linux-v0-assert
+          key: llvm-build-11-linux-v2-assert
          paths:
            llvm-build
      - run: make ASSERT=1
@@ -147,7 +181,6 @@ commands:
          key: go-cache-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_BUILD_NUM }}
          paths:
            - ~/.cache/go-build
-            - ~/.cache/tinygo
            - /go/pkg/mod
      - run: make gen-device -j4
      - run: make smoketest TINYGO=build/tinygo
@@ -158,9 +191,9 @@ commands:
      - run:
          name: "Install apt dependencies"
          command: |
+            sudo apt-get update
            sudo apt-get install \
                gcc-arm-linux-gnueabihf \
-                binutils-arm-none-eabi \
                libc6-dev-armel-cross \
                gcc-aarch64-linux-gnu \
                libc6-dev-arm64-cross \
@@ -168,7 +201,11 @@ commands:
                qemu-user \
                gcc-avr \
                avr-libc
+            sudo apt-get install --no-install-recommends libc6-dev-i386 lib32gcc-6-dev
      - install-node
+      - install-wasmtime
+      - install-xtensa-toolchain:
+          variant: "linux-amd64"
      - restore_cache:
          keys:
            - go-cache-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_PREVIOUS_BUILD_NUM }}
@@ -176,17 +213,14 @@ commands:
      - llvm-source-linux
      - restore_cache:
          keys:
-            - llvm-build-9-linux-v0
+            - llvm-build-11-linux-v2-noassert
      - run:
          name: "Build LLVM"
          command: |
            if [ ! -f llvm-build/lib/liblldELF.a ]
            then
              # install dependencies
-              sudo apt-get install cmake clang ninja-build
-              # make build faster
-              export CC=clang
-              export CXX=clang++
+              sudo apt-get install cmake ninja-build
              # hack ninja to use less jobs
              echo -e '#!/bin/sh\n/usr/bin/ninja -j3 "$@"' > /go/bin/ninja
              chmod +x /go/bin/ninja
@@ -194,25 +228,32 @@ commands:
              make llvm-build
            fi
      - save_cache:
-          key: llvm-build-9-linux-v0
+          key: llvm-build-11-linux-v2-noassert
          paths:
            llvm-build
      - build-wasi-libc
      - run:
          name: "Test TinyGo"
          command: make test
+      - run:
+          name: "Install fpm"
+          command: |
+            sudo apt-get install ruby ruby-dev
+            sudo gem install --no-document fpm
      - run:
          name: "Build TinyGo release"
          command: |
-            make release -j3
+            make release deb -j3
            cp -p build/release.tar.gz /tmp/tinygo.linux-amd64.tar.gz
+            cp -p build/release.deb    /tmp/tinygo_amd64.deb
      - store_artifacts:
          path: /tmp/tinygo.linux-amd64.tar.gz
+      - store_artifacts:
+          path: /tmp/tinygo_amd64.deb
      - save_cache:
          key: go-cache-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_BUILD_NUM }}
          paths:
            - ~/.cache/go-build
-            - ~/.cache/tinygo
            - /go/pkg/mod
      - run:
          name: "Extract release tarball"
@@ -229,27 +270,29 @@ commands:
      - run:
          name: "Install dependencies"
          command: |
-            curl https://dl.google.com/go/go1.13.darwin-amd64.tar.gz -o go1.13.darwin-amd64.tar.gz
-            sudo tar -C /usr/local -xzf go1.13.darwin-amd64.tar.gz
+            curl https://dl.google.com/go/go1.16.darwin-amd64.tar.gz -o go1.16.darwin-amd64.tar.gz
+            sudo tar -C /usr/local -xzf go1.16.darwin-amd64.tar.gz
            ln -s /usr/local/go/bin/go /usr/local/bin/go
            HOMEBREW_NO_AUTO_UPDATE=1 brew install qemu
+      - install-xtensa-toolchain:
+          variant: "macos"
      - restore_cache:
          keys:
            - go-cache-macos-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_PREVIOUS_BUILD_NUM }}
            - go-cache-macos-v2-{{ checksum "go.mod" }}
      - restore_cache:
          keys:
-            - llvm-source-9-macos-v0
+            - llvm-source-11-macos-v1
      - run:
          name: "Fetch LLVM source"
          command: make llvm-source
      - save_cache:
-          key: llvm-source-9-macos-v0
+          key: llvm-source-11-macos-v1
          paths:
            - llvm-project
      - restore_cache:
          keys:
-            - llvm-build-9-macos-v0
+            - llvm-build-11-macos-v2
      - run:
          name: "Build LLVM"
          command: |
@@ -261,17 +304,17 @@ commands:
              make llvm-build
            fi
      - save_cache:
-          key: llvm-build-9-macos-v0
+          key: llvm-build-11-macos-v2
          paths:
            llvm-build
      - restore_cache:
          keys:
-            - wasi-libc-sysroot-macos-v0
+            - wasi-libc-sysroot-macos-v3
      - run:
          name: "Build wasi-libc"
-          command: make wasi-libc CLANG=$PWD/llvm-build/bin/clang LLVM_AR=$PWD/llvm-build/bin/llvm-ar LLVM_NM=$PWD/llvm-build/bin/llvm-nm
+          command: make wasi-libc
      - save_cache:
-          key: wasi-libc-sysroot-macos-v0
+          key: wasi-libc-sysroot-macos-v3
          paths:
            - lib/wasi-libc/sysroot
      - run:
@@ -291,44 +334,58 @@ commands:
            tar -C /usr/local/opt -xf /tmp/tinygo.darwin-amd64.tar.gz
            ln -s /usr/local/opt/tinygo/bin/tinygo /usr/local/bin/tinygo
            tinygo version
-      - run:
-          name: "Download SiFive GNU toolchain"
-          command: |
-            curl -O https://static.dev.sifive.com/dev-tools/riscv64-unknown-elf-gcc-8.2.0-2019.05.3-x86_64-apple-darwin.tar.gz
-            sudo tar -C /usr/local --strip-components=1 -xf riscv64-unknown-elf-gcc-8.2.0-2019.05.3-x86_64-apple-darwin.tar.gz
      - run: make smoketest AVR=0
      - save_cache:
          key: go-cache-macos-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_BUILD_NUM }}
          paths:
            - ~/.cache/go-build
-            - ~/.cache/tinygo
            - /go/pkg/mod

-
 jobs:
-  test-llvm9-go111:
+  test-llvm10-go111:
    docker:
-      - image: circleci/golang:1.11-stretch
+      - image: circleci/golang:1.11-buster
    steps:
-      - test-linux
-  test-llvm9-go112:
+      - test-linux:
+          llvm: "10"
+  test-llvm10-go112:
    docker:
-      - image: circleci/golang:1.12-stretch
+      - image: circleci/golang:1.12-buster
    steps:
-      - test-linux
-  test-llvm9-go113:
+      - test-linux:
+          llvm: "10"
+  test-llvm10-go113:
    docker:
-      - image: circleci/golang:1.13-stretch
+      - image: circleci/golang:1.13-buster
    steps:
-      - test-linux
+      - test-linux:
+          llvm: "10"
+  test-llvm10-go114:
+    docker:
+      - image: circleci/golang:1.14-buster
+    steps:
+      - test-linux:
+          llvm: "10"
+  test-llvm11-go115:
+    docker:
+      - image: circleci/golang:1.15-buster
+    steps:
+      - test-linux:
+          llvm: "11"
+  test-llvm11-go116:
+    docker:
+      - image: circleci/golang:1.16-buster
+    steps:
+      - test-linux:
+          llvm: "11"
  assert-test-linux:
    docker:
-      - image: circleci/golang:1.13-stretch
+      - image: circleci/golang:1.14-stretch
    steps:
      - assert-test-linux
  build-linux:
    docker:
-      - image: circleci/golang:1.13-stretch
+      - image: circleci/golang:1.14-stretch
    steps:
      - build-linux
  build-macos:
@@ -339,13 +396,15 @@ jobs:



-
 workflows:
  test-all:
    jobs:
-      - test-llvm9-go111
-      - test-llvm9-go112
-      - test-llvm9-go113
+      - test-llvm10-go111
+      - test-llvm10-go112
+      - test-llvm10-go113
+      - test-llvm10-go114
+      - test-llvm11-go115
+      - test-llvm11-go116
      - build-linux
      - build-macos
      - assert-test-linux
@@ -3,14 +3,19 @@ docs/_build
 src/device/avr/*.go
 src/device/avr/*.ld
 src/device/avr/*.s
+src/device/esp/*.go
 src/device/nrf/*.go
 src/device/nrf/*.s
+src/device/nxp/*.go
+src/device/nxp/*.s
 src/device/sam/*.go
 src/device/sam/*.s
 src/device/sifive/*.go
 src/device/sifive/*.s
 src/device/stm32/*.go
 src/device/stm32/*.s
+src/device/kendryte/*.go
+src/device/kendryte/*.s
 vendor
 llvm-build
 llvm-project
@@ -17,3 +17,9 @@
 [submodule "lib/wasi-libc"]
 	path = lib/wasi-libc
 	url = https://github.com/CraneStation/wasi-libc
+[submodule "lib/picolibc"]
+	path = lib/picolibc
+	url = https://github.com/keith-packard/picolibc.git
+[submodule "lib/stm32-svd"]
+	path = lib/stm32-svd
+	url = https://github.com/tinygo-org/stm32-svd
@@ -10,6 +10,14 @@ This guide describes how to statically link TinyGo against LLVM, libclang and
 lld so that the binary can be easily moved between systems. It also shows how to
 build a release tarball that includes this binary and all necessary extra files.

+**Note**: this documentation describes how to build a statically linked release
+tarball. If you want to develop TinyGo, you will probably want to follow a
+different guide:
+
+  * [Linux](https://tinygo.org/getting-started/linux/#source-install)
+  * [macOS](https://tinygo.org/getting-started/macos/#source-install)
+  * [Windows](https://tinygo.org/getting-started/windows/#source-install)
+
 ## Dependencies

 LLVM, Clang and LLD are quite light on dependencies, requiring only standard
@@ -84,6 +92,10 @@ Now that we have a working static build, it's time to make a release tarball:

    make release

+If you did not clone the repository with the `--recursive` option, you will get errors until you initialize the project submodules:
+
+    git submodule update --init
+
 The release tarball is stored in build/release.tar.gz, and can be extracted with
 the following command (for example in ~/lib):

@@ -1,3 +1,410 @@
+0.17.0
+
+---
+* **command line**
+  - switch to LLVM 11 for static builds
+  - support gdb debugging with AVR
+  - add support for additional openocd commands
+  - add `-x` flag to print commands
+  - use LLVM 11 by default when linking LLVM dynamically
+  - update go-llvm to use LLVM 11 on macOS
+  - bump go.bug.st/serial to version 1.1.2
+  - do not build LLVM with libxml to work around a bugo on macOS
+  - add support for Go 1.16
+  - support gdb daemonization on Windows
+  - remove support for LLVM 9, to fix CI
+  - kill OpenOCD if it does not exit with a regular quit signal
+  - support `-ocd-output` on Windows
+* **compiler**
+  - `builder`: parallelize most of the build
+  - `builder`: remove unused cacheKey parameter
+  - `builder`: add -mcpu flag while building libraries
+  - `builder`: wait for running jobs to finish
+  - `cgo`: add support for variadic functions
+  - `compiler`: fix undefined behavior in wordpack
+  - `compiler`: fix incorrect "exported function" panic
+  - `compiler`: fix non-int integer constants (fixing a crash)
+  - `compiler`: refactor and add tests
+  - `compiler`: emit a nil check when slicing an array pointer
+  - `compiler`: saturate float-to-int conversions
+  - `compiler`: test float to int conversions and fix upper-bound calculation
+  - `compiler`: support all kinds of deferred builtins
+  - `compiler`: remove ir package
+  - `compiler`: remove unnecessary main.main call workaround
+  - `compiler`: move the setting of attributes to getFunction
+  - `compiler`: create runtime types lazily when needed
+  - `compiler`: move settings to a separate Config struct
+  - `compiler`: work around an ARM backend bug in LLVM
+  - `interp`: rewrite entire package
+  - `interp`: fix alignment of untyped globals
+  - `loader`: use name "main" for the main package
+  - `loader`: support imports from vendor directories
+  - `stacksize`: add support for DW_CFA_offset_extended
+  - `transform`: show better error message in coroutines lowering
+* **standard library**
+  - `machine`: accept configuration struct for ADC parameters
+  - `machine`: make I2C.Configure signature consistent
+  - `reflect`: implement PtrTo
+  - `runtime`: refactor to simplify stack switching
+  - `runtime`: put metadata at the top end of the heap
+* **targets**
+  - `atsam`: add a length check to findPinPadMapping
+  - `atsam`: improve USBCDC
+  - `atsam`: avoid infinite loop when USBCDC is disconnected
+  - `avr`: add SPI support for Atmega based chips
+  - `avr`: use Clang for compiling C and assembly files
+  - `esp32`: implement task based scheduler
+  - `esp32`: enable the FPU
+  - `esp8266`: implement task based scheduler
+  - `esp`: add compiler-rt library
+  - `esp`: add picolibc
+  - `nrf`: refactor code a bit to reduce duplication
+  - `nrf`: use SPIM peripheral instead of the legacy SPI peripheral
+  - `nrf`: update nrfx submodule to latest commit
+  - `nrf52840`: ensure that USB CDC interface is only initialized once
+  - `nrf52840`: improve USBCDC
+  - `stm32`: use stm32-rs SVDs which are of much higher quality
+  - `stm32`: harmonization of UART logic
+  - `stm32`: replace I2C addressable interface with simpler type
+  - `stm32`: fix i2c and add stm32f407 i2c
+  - `stm32`: revert change that adds support for channels in interrupts
+  - `wasm`: implement a growable heap
+  - `wasm`: fix typo in wasm_exec.js, syscall/js.valueLoadString()
+  - `wasm`: Namespaced Wasm Imports so they don't conflict across modules, or reserved LLVM IR
+  - `wasi`: support env variables based on libc
+  - `wasi`: specify wasi-libc in a different way, to improve error message
+* **boards**
+  - `matrixportal-m4`: add support for board Adafruit Matrix Portal M4
+  - `mkr1000`: add this board
+  - `nucleo-f722ze`: add this board
+  - `clue`: correct volume name and add alias for release version of Adafruit Clue board
+  - `p1am-100`: add support for the P1AM-100 (similar to Arduino MKR)
+  - `microbit-v2`: add initial support based on work done by @alankrantas thank you!
+  - `lgt92`: support for STM32L0 MCUs and Dragino LGT92 device
+  - `nicenano`: nice!nano board support
+  - `circuitplay-bluefruit`: correct internal I2C pin mapping
+  - `clue`: correct for lack of low frequency crystal
+  - `digispark`: split off attiny85 target
+  - `nucleo-l552ze`: implementation with CLOCK, LED, and UART
+  - `nrf52840-mdk-usb-dongle`: add this board
+
+0.16.0
+---
+
+* **command-line**
+  - add initial support for LLVM 11
+  - make lib64 clang include path check more robust
+  - `build`: improve support for GOARCH=386 and add tests
+  - `gdb`: add support for qemu-user targets
+  - `test`: support non-host tests
+  - `test`: add support for -c and -o flags
+  - `test`: implement some benchmark stubs
+* **compiler**
+  - `builder`: improve detection of clang on Fedora
+  - `compiler`: fix floating point comparison bugs
+  - `compiler`: implement negate for complex numbers
+  - `loader`: fix linkname in test binaries
+  - `transform`: add missing return pointer restore for regular coroutine tail
+    calls
+* **standard library**
+  - `machine`: switch default frequency to 4MHz
+  - `machine`: clarify caller's responsibility in `SetInterrupt`
+  - `os`: add `LookupEnv()` stub
+  - `reflect`: implement `Swapper`
+  - `runtime`: fix UTF-8 decoding
+  - `runtime`: gc: use raw stack access whenever possible
+  - `runtime`: use dedicated printfloat32
+  - `runtime`: allow ranging over a nil map
+  - `runtime`: avoid device/nxp dependency in HardFault handler
+  - `testing`: implement dummy Helper method
+  - `testing`: add Run method
+* **targets**
+  - `arm64`: add support for SVCall intrinsic
+  - `atsamd51`: avoid panic when configuring SPI with SDI=NoPin
+  - `avr`: properly support the `.rodata` section
+  - `esp8266`: implement `Pin.Get` function
+  - `nintendoswitch`: fix crash when printing long lines (> 120)
+  - `nintendoswitch`: add env parser and removed unused stuff
+  - `nrf`: add I2C error checking
+  - `nrf`: give more flexibility in picking SPI speeds
+  - `nrf`: fix nrf52832 flash size
+  - `stm32f103`: support wakeups from interrupts
+  - `stm32f405`: add SPI support
+  - `stm32f405`: add I2C support
+  - `wasi`: add support for this target
+  - `wasi`: use 'generic' ABI by default
+  - `wasi`: remove --no-threads flag from wasm-ld
+  - `wasm`: add instanceof support for WebAssembly
+  - `wasm`: use fixed length buffer for putchar
+* **boards**
+  - `d1mini`: add this ESP8266 based board
+  - `esp32`: use board definitions instead of chip names
+  - `qtpy`: add board definition for Adafruit QTPy
+  - `teensy40`: add this board
+
+0.15.0
+---
+
+* **command-line**
+  - add cached GOROOT to info subcommand
+  - embed git-hash in tinygo-dev executable
+  - implement tinygo targets to list usable targets
+  - use simpler file copy instead of file renaming to avoid issues on nrf52840 UF2 bootloaders
+  - use ToSlash() to specify program path
+  - support flashing esp32/esp8266 directly from tinygo
+  - when flashing call PortReset only on other than openocd
+* **compiler**
+  - `compileopts`: add support for custom binary formats
+  - `compiler`: improve display of goroutine wrappers
+  - `interp`: don't panic in the Store method
+  - `interp`: replace some panics with error messages
+  - `interp`: show error line in first line of the traceback
+  - `loader`: be more robust when creating the cached GOROOT
+  - `loader`: rewrite/refactor much of the code to use go list directly
+  - `loader`: use ioutil.TempDir to create a temporary directory
+  - `stacksize`: deal with DW_CFA_advance_loc1
+* **standard library**
+  - `runtime`: use waitForEvents when appropriate
+* **wasm**
+  - `wasm`: Remove --no-threads from wasm-ld calls.
+  - `wasm`: update wasi-libc dependency
+* **targets**
+  - `arduino-mega2560`: fix flashing on Windows
+  - `arm`: automatically determine stack sizes
+  - `arm64`: make dynamic loader structs and constants private
+  - `avr`: configure emulator in board files
+  - `cortexm`: fix stack size calculation with interrupts
+  - `flash`: add openocd settings to atsamd21 / atsamd51
+  - `flash`: add openocd settings to nrf5
+  - `microbit`: reelboard: flash using OpenOCD when needed
+  - `nintendoswitch`: Add dynamic loader for runtime loading PIE sections
+  - `nintendoswitch`: fix import cycle on dynamic_arm64.go
+  - `nintendoswitch`: Fix invalid memory read / write in print calls
+  - `nintendoswitch`: simplified assembly code
+  - `nintendoswitch`: support outputting .nro files directly
+* **boards**
+  - `arduino-zero`: Adding support for the Arduino Zero (#1365)
+  - `atsamd2x`: fix BAUD value
+  - `atsamd5x`: fix BAUD value
+  - `bluepill`: Enable stm32's USART2 for the board and map it to UART1 tinygo's device
+  - `device/atsamd51x`: add all remaining bitfield values for PCHCTRLm Mapping
+  - `esp32`: add libgcc ROM functions to linker script
+  - `esp32`: add SPI support
+  - `esp32`: add support for basic GPIO
+  - `esp32`: add support for the Espressif ESP32 chip
+  - `esp32`: configure the I/O matrix for GPIO pins
+  - `esp32`: export machine.PortMask* for bitbanging implementations
+  - `esp8266`: add support for this chip
+  - `machine/atsamd51x,runtime/atsamd51x`: fixes needed for full support for all PWM pins. Also adds some useful constants to clarify peripheral clock usage
+  - `machine/itsybitsy-nrf52840`: add support for Adafruit Itsybitsy nrf52840 (#1243)
+  - `machine/stm32f4`: refactor common code and add new build tag stm32f4 (#1332)
+  - `nrf`: add SoftDevice support for the Circuit Playground Bluefruit
+  - `nrf`: call sd_app_evt_wait when the SoftDevice is enabled
+  - `nrf52840`: add build tags for SoftDevice support
+  - `nrf52840`: use higher priority for USB-CDC code
+  - `runtime/atsamd51x`: use PCHCTRL_GCLK_SERCOMX_SLOW for setting clocks on all SERCOM ports
+  - `stm32f405`: add basic UART handler
+  - `stm32f405`: add STM32F405 machine/runtime, and new board/target feather-stm32f405
+* **build**
+  - `all`: run test binaries in the correct directory
+  - `build`: Fix arch release job
+  - `ci`: run `tinygo test` for known-working packages
+  - `ci`: set git-fetch-depth to 1
+  - `docker`: fix the problem with the wasm build (#1357)
+  - `Makefile`: check whether submodules have been downloaded in some common cases
+* **docs**
+  - add ESP32, ESP8266, and Adafruit Feather STM32F405 to list of supported boards
+
+0.14.1
+---
+* **command-line**
+  - support for Go 1.15
+* **compiler**
+  - loader:  work around Windows symlink limitation
+
+0.14.0
+---
+* **command-line**
+  - fix `getDefaultPort()` on non-English Windows locales
+  - compileopts: improve error reporting of unsupported flags
+  - fix test subcommand
+  - use auto-retry to locate MSD for UF2 and HEX flashing
+  - fix touchSerialPortAt1200bps on Windows
+  - support package names with backslashes on Windows
+* **compiler**
+  - fix a few crashes due to named types
+  - add support for atomic operations
+  - move the channel blocked list onto the stack
+  - fix -gc=none
+  - fix named string to `[]byte` slice conversion
+  - implement func value and builtin defers
+  - add proper parameter names to runtime.initAll, to fix a panic
+  - builder: fix picolibc include path
+  - builder: use newer version of gohex
+  - builder: try to determine stack size information at compile time
+  - builder: remove -opt=0
+  - interp: fix sync/atomic.Value load/store methods
+  - loader: add Go module support
+  - transform: fix debug information in func lowering pass
+  - transform: do not special-case zero or one implementations of a method call
+  - transform: introduce check for method calls on nil interfaces
+  - transform: gc: track 0-index GEPs to fix miscompilation
+* **cgo**
+  - Add LDFlags support
+* **standard library**
+  - extend stdlib to allow import of more packages
+  - replace master/slave terminology with appropriate alternatives (MOSI->SDO
+    etc)
+  - `internal/bytealg`: reimplement bytealg in pure Go
+  - `internal/task`: fix nil panic in (*internal/task.Stack).Pop
+  - `os`: add Args and stub it with mock data
+  - `os`: implement virtual filesystem support
+  - `reflect`: add Cap and Len support for map and chan
+  - `runtime`: fix return address in scheduler on RISC-V
+  - `runtime`: avoid recursion in printuint64 function
+  - `runtime`: replace ReadRegister with AsmFull inline assembly
+  - `runtime`: fix compilation errors when using gc.extalloc
+  - `runtime`: add cap and len support for chans
+  - `runtime`: refactor time handling (improving accuracy)
+  - `runtime`: make channels work in interrupts
+  - `runtime/interrupt`: add cross-chip disable/restore interrupt support
+  - `sync`: implement `sync.Cond`
+  - `sync`: add WaitGroup
+* **targets**
+  - `arm`: allow nesting in DisableInterrupts and EnableInterrupts
+  - `arm`: make FPU configuraton consistent
+  - `arm`: do not mask fault handlers in critical sections
+  - `atmega2560`: fix pin mapping for pins D2, D5 and the L port
+  - `atsamd`: return an error when an incorrect PWM pin is used
+  - `atsamd`: add support for pin change interrupts
+  - `atsamd`: add DAC support
+  - `atsamd21`: add more ADC pins
+  - `atsamd51`: fix ROM / RAM size on atsamd51j20
+  - `atsamd51`: add more pins
+  - `atsamd51`: add more ADC pins
+  - `atsamd51`: add pin change interrupt settings
+  - `atsamd51`: extend pinPadMapping
+  - `arduino-nano33`: use (U)SB flag to ensure that device can be found when
+     not on default port
+  - `arduino-nano33`: remove (d)ebug flag to reduce console noise when flashing
+  - `avr`: use standard pin numbering
+  - `avr`: unify GPIO pin/port code
+  - `avr`: add support for PinInputPullup
+  - `avr`: work around codegen bug in LLVM 10
+  - `avr`: fix target triple
+  - `fe310`: remove extra println left in by mistake
+  - `feather-nrf52840`: add support for the Feather nRF52840
+  - `maixbit`: add board definition and dummy runtime
+  - `nintendoswitch`: Add experimental Nintendo Switch support without CRT
+  - `nrf`: expose the RAM base address
+  - `nrf`: add support for pin change interrupts
+  - `nrf`: add microbit-s110v8 target
+  - `nrf`: fix bug in SPI.Tx
+  - `nrf`: support debugging the PCA10056
+  - `pygamer`: add Adafruit PyGamer suport
+  - `riscv`: fix interrupt configuration bug
+  - `riscv`: disable linker relaxations during gp init
+  - `stm32f4disco`: add new target with ST-Link v2.1 debugger
+  - `teensy36`: add Teensy 3.6 support
+  - `wasm`: fix event handling
+  - `wasm`: add --no-demangle linker option
+  - `wioterminal`: add support for the Seeed Wio Terminal
+  - `xiao`: add support for the Seeed XIAO
+
+0.13.1
+---
+* **standard library**
+  - `runtime`: do not put scheduler and GC code in the same section
+  - `runtime`: copy stack scan assembly for GBA
+* **boards**
+  - `gameboy-advance`: always use ARM mode instead of Thumb mode
+
+
+0.13.0
+---
+* **command line**
+  - use `gdb-multiarch` for debugging Cortex-M chips
+  - support `tinygo run` with simavr
+  - support LLVM 10
+  - support Go 1.14
+  - retry 3 times when attempting to do a 1200-baud reset
+* **compiler**
+  - mark the `abort` function as noreturn
+  - fix deferred calls to exported functions
+  - add debug info for local variables
+  - check for channel size limit
+  - refactor coroutine lowering
+  - add `dereferenceable_or_null` attribute to pointer parameters
+  - do not perform nil checking when indexing slices and on `unsafe.Pointer`
+  - remove `runtime.isnil` hack
+  - use LLVM builtins for runtime `memcpy`/`memmove`/`memzero` functions
+  - implement spec-compliant shifts on negative/overflow
+  - support anonymous type asserts
+  - track pointer result of string concatenation for GC
+  - track PHI nodes for GC
+  - add debug info to goroutine start wrappers
+  - optimize comparing interface values against nil
+  - fix miscompilation when deferring an interface call
+  - builder: include picolibc for most baremetal targets
+  - builder: run tools (clang, lld) as separate processes
+  - builder: use `-fshort-enums` consistently
+  - interp: add support for constant type asserts
+  - interp: better support for interface operations
+  - interp: include backtrace with error
+  - transform: do not track const globals for GC
+  - transform: replace panics with source locations
+  - transform: fix error in interface lowering pass
+  - transform: make coroutine lowering deterministic
+  - transform: fix miscompilation in func lowering
+* **cgo**
+  - make `-I` and `-L` paths absolute
+* **standard library**
+  - `machine`: set the USB VID and PID to the manufacturer values
+  - `machine`: correct USB CDC composite descriptors
+  - `machine`: move `errors.New` calls to globals
+  - `runtime`: support operations on nil maps
+  - `runtime`: fix copy builtin return value on AVR
+  - `runtime`: refactor goroutines
+  - `runtime`: support `-scheduler=none` on most platforms
+  - `runtime`: run package initialization in the main goroutine
+  - `runtime`: export `malloc` / `free` for use from C
+  - `runtime`: add garbage collector that uses an external allocator
+  - `runtime`: scan callee-saved registers while marking the stack
+  - `runtime`: remove recursion from conservative GC
+  - `runtime`: fix blocking select on nil channel
+  - `runtime/volatile`: include `ReplaceBits` method
+  - `sync`: implement trivial `sync.Map`
+* **targets**
+  - `arm`: use `-fomit-frame-pointer`
+  - `atmega1284`: support this chip for testing purposes
+  - `atsamd51`: make QSPI available on all boards
+  - `atsamd51`: add support for ADC1
+  - `atsamd51`: use new interrupt registration in UART code
+  - `attiny`: clean up pin definitions
+  - `avr`: use the correct RAM start address
+  - `avr`: pass the correct `-mmcu` flag to the linker
+  - `avr`: add support for tasks scheduler (disabled by default)
+  - `avr`: fix linker problem with overlapping program/data areas
+  - `nrf`: fix typo in pin configuration options
+  - `nrf`: add lib/nrfx/mdk to include dirs
+  - `nrf52840`: implement USB-CDC
+  - `riscv`: implement VirtIO target and add RISC-V integration test
+  - `riscv`: add I2C support for the HiFive1 rev B board
+  - `stm32`: refactor GPIO pin handling
+  - `stm32`: refactor UART code
+  - `stm32f4`: add SPI
+  - `wasm`: support Go 1.14 (breaking previous versions)
+  - `wasm`: support `syscall/js.CopyBytesToJS`
+  - `wasm`: sync polyfills from Go 1.14.
+* **boards**
+  - `arduino-mega2560`: add the Arduino Mega 2560
+  - `clue-alpha`: add the Adafruit CLUE Alpha
+  - `gameboy-advance`: enable debugging with GDB
+  - `particle-argon`: add the Particle Argon board
+  - `particle-boron`: add the Particle Boron board
+  - `particle-xenon`: add the Particle Xenon board
+  - `reelboard`: add `reelboard-s140v7` SoftDevice target
+
 0.12.0
 ---
 * **command line**
@@ -32,7 +32,7 @@ Microcontrollers have lots of peripherals (I2C, SPI, ADC, etc.) and many don't h

 ## How to use our Github repository

-The `master` branch of this repo will always have the latest released version of TinyGo. All of the active development work for the next release will take place in the `dev` branch. TinyGo will use semantic versioning and will create a tag/release for each release.
+The `release` branch of this repo will always have the latest released version of TinyGo. All of the active development work for the next release will take place in the `dev` branch. TinyGo will use semantic versioning and will create a tag/release for each release.

 Here is how to contribute back some code or documentation:

@@ -15,4 +15,4 @@ Ayke van Laethem <aykevanlaethem@gmail.com>
 Daniel Esteban <conejo@conejo.me>
 Loon, LLC.
 Ron Evans <ron@hybridgroup.com>
-Jaden Weiss <jaden@jadendw.dev>
+Nia Weiss <niaow1234@gmail.com>
@@ -1,10 +1,10 @@
-# TinyGo base stage installs Go 1.13, LLVM 9 and the TinyGo compiler itself.
-FROM golang:1.13 AS tinygo-base
+# TinyGo base stage installs the most recent Go 1.15.x, LLVM 11 and the TinyGo compiler itself.
+FROM golang:1.15 AS tinygo-base

 RUN wget -O- https://apt.llvm.org/llvm-snapshot.gpg.key| apt-key add - && \
-    echo "deb http://apt.llvm.org/buster/ llvm-toolchain-buster-9 main" >> /etc/apt/sources.list && \
+    echo "deb http://apt.llvm.org/buster/ llvm-toolchain-buster-11 main" >> /etc/apt/sources.list && \
    apt-get update && \
-    apt-get install -y llvm-9-dev libclang-9-dev git
+    apt-get install -y llvm-11-dev libclang-11-dev lld-11 git

 COPY . /tinygo

@@ -12,8 +12,11 @@ COPY . /tinygo
 # after copying the tinygo directory in the previous step.
 RUN cd /tinygo/ && \
    rm -rf ./lib/* && \
+    git submodule sync && \
    git submodule update --init --recursive --force

+COPY ./lib/picolibc-include/* /tinygo/lib/picolibc-include/
+
 RUN cd /tinygo/ && \
    go install /tinygo/

@@ -24,10 +27,10 @@ COPY --from=tinygo-base /go/bin/tinygo /go/bin/tinygo
 COPY --from=tinygo-base /tinygo/src /tinygo/src
 COPY --from=tinygo-base /tinygo/targets /tinygo/targets

-RUN wget -O- https://apt.llvm.org/llvm-snapshot.gpg.key| apt-key add - && \
-    echo "deb http://apt.llvm.org/buster/ llvm-toolchain-buster-9 main" >> /etc/apt/sources.list && \
+RUN cd /tinygo/ && \
    apt-get update && \
-    apt-get install -y libllvm9 lld-9
+    apt-get install -y make clang-11 libllvm11 lld-11 && \
+    make wasi-libc

 # tinygo-avr stage installs the needed dependencies to compile TinyGo programs for AVR microcontrollers.
 FROM tinygo-base AS tinygo-avr
@@ -58,7 +61,7 @@ COPY --from=tinygo-base /tinygo/lib /tinygo/lib

 RUN cd /tinygo/ && \
    apt-get update && \
-    apt-get install -y apt-utils make clang-9 && \
+    apt-get install -y apt-utils make clang-11 && \
    make gen-device-nrf && make gen-device-stm32

 # tinygo-all stage installs the needed dependencies to compile TinyGo programs for all platforms.
@@ -70,7 +73,7 @@ COPY --from=tinygo-base /tinygo/lib /tinygo/lib

 RUN cd /tinygo/ && \
    apt-get update && \
-    apt-get install -y apt-utils make clang-9 binutils-avr gcc-avr avr-libc && \
+    apt-get install -y apt-utils make clang-11 binutils-avr gcc-avr avr-libc && \
    make gen-device

 CMD ["tinygo"]
@@ -1,7 +1,7 @@
-Copyright (c) 2018-2019 TinyGo Authors. All rights reserved.
+Copyright (c) 2018-2021 TinyGo Authors. All rights reserved.

 TinyGo includes portions of the Go standard library.
-Copyright (c) 2009-2019 The Go Authors. All rights reserved.
+Copyright (c) 2009-2021 The Go Authors. All rights reserved.

 TinyGo includes portions of LLVM, which is under the Apache License v2.0 with
 LLVM Exceptions. See https://llvm.org/LICENSE.txt for license information.
@@ -4,13 +4,15 @@ all: tinygo

 # Default build and source directories, as created by `make llvm-build`.
 LLVM_BUILDDIR ?= llvm-build
-CLANG_SRC ?= llvm-project/clang
-LLD_SRC ?= llvm-project/lld
+LLVM_PROJECTDIR ?= llvm-project
+CLANG_SRC ?= $(LLVM_PROJECTDIR)/clang
+LLD_SRC ?= $(LLVM_PROJECTDIR)/lld

-# Default tool selection.
-CLANG ?= clang-9
-LLVM_AR ?= llvm-ar-9
-LLVM_NM ?= llvm-nm-9
+# Try to autodetect LLVM build tools.
+detect = $(shell command -v $(1) 2> /dev/null && echo $(1))
+CLANG ?= $(word 1,$(abspath $(call detect,llvm-build/bin/clang))$(call detect,clang-11)$(call detect,clang-10)$(call detect,clang))
+LLVM_AR ?= $(word 1,$(abspath $(call detect,llvm-build/bin/llvm-ar))$(call detect,llvm-ar-11)$(call detect,llvm-ar-10)$(call detect,llvm-ar))
+LLVM_NM ?= $(word 1,$(abspath $(call detect,llvm-build/bin/llvm-nm))$(call detect,llvm-nm-11)$(call detect,llvm-nm-10)$(call detect,llvm-nm))

 # Go binary and GOROOT to select
 GO ?= go
@@ -20,10 +22,10 @@ export GOROOT = $(shell $(GO) env GOROOT)
 MD5SUM = md5sum

 # tinygo binary for tests
-TINYGO ?= tinygo
+TINYGO ?= $(word 1,$(call detect,tinygo)$(call detect,build/tinygo))

 # Use CCACHE for LLVM if possible
-ifneq (, $(shell which ccache))
+ifneq (, $(shell command -v ccache 2> /dev/null))
    LLVM_OPTION += '-DLLVM_CCACHE_BUILD=ON'
 endif

@@ -34,9 +36,9 @@ else
    LLVM_OPTION += '-DLLVM_ENABLE_ASSERTIONS=OFF'
 endif

-.PHONY: all tinygo test $(LLVM_BUILDDIR) llvm-source clean fmt gen-device gen-device-nrf gen-device-avr
+.PHONY: all tinygo test $(LLVM_BUILDDIR) llvm-source clean fmt gen-device gen-device-nrf gen-device-nxp gen-device-avr

-LLVM_COMPONENTS = all-targets analysis asmparser asmprinter bitreader bitwriter codegen core coroutines coverage debuginfodwarf executionengine instrumentation interpreter ipo irreader linker lto mc mcjit objcarcopts option profiledata scalaropts support target
+LLVM_COMPONENTS = all-targets analysis asmparser asmprinter bitreader bitwriter codegen core coroutines coverage debuginfodwarf executionengine frontendopenmp instrumentation interpreter ipo irreader linker lto mc mcjit objcarcopts option profiledata scalaropts support target

 ifeq ($(OS),Windows_NT)
    EXE = .exe
@@ -46,64 +48,69 @@ ifeq ($(OS),Windows_NT)
    # LLVM compiled using MinGW on Windows appears to have problems with threads.
    # Without this flag, linking results in errors like these:
    #     libLLVMSupport.a(Threading.cpp.obj):Threading.cpp:(.text+0x55): undefined reference to `std::thread::hardware_concurrency()'
-    LLVM_OPTION += -DLLVM_ENABLE_THREADS=OFF
+    LLVM_OPTION += -DLLVM_ENABLE_THREADS=OFF -DLLVM_ENABLE_PIC=OFF

+    CGO_CPPFLAGS += -DCINDEX_NO_EXPORTS
    CGO_LDFLAGS += -static -static-libgcc -static-libstdc++
    CGO_LDFLAGS_EXTRA += -lversion

-    # Build libclang manually because the CMake-based build system on Windows
-    # doesn't allow building libclang as a static library.
-    LIBCLANG_PATH = $(abspath build/libclang-custom.a)
-    LIBCLANG_FILES = $(abspath $(wildcard $(LLVM_BUILDDIR)/tools/clang/tools/libclang/CMakeFiles/libclang.dir/*.cpp.obj))
-
-    # Add the libclang dependency to the tinygo binary target.
-tinygo: $(LIBCLANG_PATH)
-test: $(LIBCLANG_PATH)
-    # Build libclang.
-$(LIBCLANG_PATH): $(LIBCLANG_FILES)
-	@mkdir -p build
-	ar rcs $(LIBCLANG_PATH) $^
+    LIBCLANG_NAME = libclang

 else ifeq ($(shell uname -s),Darwin)
    MD5SUM = md5
-    LIBCLANG_PATH = $(abspath $(LLVM_BUILDDIR))/lib/libclang.a
+    LIBCLANG_NAME = clang
 else ifeq ($(shell uname -s),FreeBSD)
    MD5SUM = md5
-    LIBCLANG_PATH = $(abspath $(LLVM_BUILDDIR))/lib/libclang.a
+    LIBCLANG_NAME = clang
    START_GROUP = -Wl,--start-group
    END_GROUP = -Wl,--end-group
 else
-    LIBCLANG_PATH = $(abspath $(LLVM_BUILDDIR))/lib/libclang.a
+    LIBCLANG_NAME = clang
    START_GROUP = -Wl,--start-group
    END_GROUP = -Wl,--end-group
 endif

-CLANG_LIBS = $(START_GROUP) -lclangAnalysis -lclangARCMigrate -lclangAST -lclangASTMatchers -lclangBasic -lclangCodeGen -lclangCrossTU -lclangDriver -lclangDynamicASTMatchers -lclangEdit -lclangFormat -lclangFrontend -lclangFrontendTool -lclangHandleCXX -lclangHandleLLVM -lclangIndex -lclangLex -lclangParse -lclangRewrite -lclangRewriteFrontend -lclangSema -lclangSerialization -lclangStaticAnalyzerCheckers -lclangStaticAnalyzerCore -lclangStaticAnalyzerFrontend -lclangTooling -lclangToolingASTDiff -lclangToolingCore -lclangToolingInclusions $(END_GROUP) -lstdc++
+# Libraries that should be linked in for the statically linked Clang.
+CLANG_LIB_NAMES = clangAnalysis clangARCMigrate clangAST clangASTMatchers clangBasic clangCodeGen clangCrossTU clangDriver clangDynamicASTMatchers clangEdit clangFormat clangFrontend clangFrontendTool clangHandleCXX clangHandleLLVM clangIndex clangLex clangParse clangRewrite clangRewriteFrontend clangSema clangSerialization clangStaticAnalyzerCheckers clangStaticAnalyzerCore clangStaticAnalyzerFrontend clangTooling clangToolingASTDiff clangToolingCore clangToolingInclusions
+CLANG_LIBS = $(START_GROUP) $(addprefix -l,$(CLANG_LIB_NAMES)) $(END_GROUP) -lstdc++

-LLD_LIBS = $(START_GROUP) -llldCOFF -llldCommon -llldCore -llldDriver -llldELF -llldMachO -llldMinGW -llldReaderWriter -llldWasm -llldYAML $(END_GROUP)
+# Libraries that should be linked in for the statically linked LLD.
+LLD_LIB_NAMES = lldCOFF lldCommon lldCore lldDriver lldELF lldMachO lldMinGW lldReaderWriter lldWasm lldYAML
+LLD_LIBS = $(START_GROUP) $(addprefix -l,$(LLD_LIB_NAMES)) $(END_GROUP)

+# Other libraries that are needed to link TinyGo.
+EXTRA_LIB_NAMES = LLVMInterpreter
+
+# These build targets appear to be the only ones necessary to build all TinyGo
+# dependencies. Only building a subset significantly speeds up rebuilding LLVM.
+# The Makefile rules convert a name like lldELF to lib/liblldELF.a to match the
+# library path (for ninja).
+# This list also includes a few tools that are necessary as part of the full
+# TinyGo build.
+NINJA_BUILD_TARGETS = clang llvm-config llvm-ar llvm-nm $(addprefix lib/lib,$(addsuffix .a,$(LIBCLANG_NAME) $(CLANG_LIB_NAMES) $(LLD_LIB_NAMES) $(EXTRA_LIB_NAMES)))

 # For static linking.
 ifneq ("$(wildcard $(LLVM_BUILDDIR)/bin/llvm-config*)","")
-    CGO_CPPFLAGS=$(shell $(LLVM_BUILDDIR)/bin/llvm-config --cppflags) -I$(abspath $(LLVM_BUILDDIR))/tools/clang/include -I$(abspath $(CLANG_SRC))/include -I$(abspath $(LLD_SRC))/include
-    CGO_CXXFLAGS=-std=c++11
-    CGO_LDFLAGS+=$(LIBCLANG_PATH) -std=c++11 -L$(abspath $(LLVM_BUILDDIR)/lib) $(CLANG_LIBS) $(LLD_LIBS) $(shell $(LLVM_BUILDDIR)/bin/llvm-config --ldflags --libs --system-libs $(LLVM_COMPONENTS)) -lstdc++ $(CGO_LDFLAGS_EXTRA)
+    CGO_CPPFLAGS+=$(shell $(LLVM_BUILDDIR)/bin/llvm-config --cppflags) -I$(abspath $(LLVM_BUILDDIR))/tools/clang/include -I$(abspath $(CLANG_SRC))/include -I$(abspath $(LLD_SRC))/include
+    CGO_CXXFLAGS=-std=c++14
+    CGO_LDFLAGS+=$(abspath $(LLVM_BUILDDIR))/lib/lib$(LIBCLANG_NAME).a -L$(abspath $(LLVM_BUILDDIR)/lib) $(CLANG_LIBS) $(LLD_LIBS) $(shell $(LLVM_BUILDDIR)/bin/llvm-config --ldflags --libs --system-libs $(LLVM_COMPONENTS)) -lstdc++ $(CGO_LDFLAGS_EXTRA)
 endif


 clean:
 	@rm -rf build

-FMT_PATHS = ./*.go builder cgo compiler interp ir loader src/device/arm src/examples src/machine src/os src/reflect src/runtime src/sync src/syscall src/internal/reflectlite transform
+FMT_PATHS = ./*.go builder cgo compiler interp loader src/device/arm src/examples src/machine src/os src/reflect src/runtime src/sync src/syscall src/internal/reflectlite transform
 fmt:
 	@gofmt -l -w $(FMT_PATHS)
 fmt-check:
 	@unformatted=$$(gofmt -l $(FMT_PATHS)); [ -z "$$unformatted" ] && exit 0; echo "Unformatted:"; for fn in $$unformatted; do echo "  $$fn"; done; exit 1


-gen-device: gen-device-avr gen-device-nrf gen-device-sam gen-device-sifive gen-device-stm32
+gen-device: gen-device-avr gen-device-esp gen-device-nrf gen-device-sam gen-device-sifive gen-device-stm32 gen-device-kendryte gen-device-nxp

 gen-device-avr:
+	@if [ ! -e lib/avr/README.md ]; then echo "Submodules have not been downloaded. Please download them using:\n  git submodule update --init"; exit 1; fi
 	$(GO) build -o ./build/gen-device-avr ./tools/gen-device-avr/
 	./build/gen-device-avr lib/avr/packs/atmega src/device/avr/
 	./build/gen-device-avr lib/avr/packs/tiny src/device/avr/
@@ -112,10 +119,18 @@ gen-device-avr:
 build/gen-device-svd: ./tools/gen-device-svd/*.go
 	$(GO) build -o $@ ./tools/gen-device-svd/

+gen-device-esp: build/gen-device-svd
+	./build/gen-device-svd -source=https://github.com/posborne/cmsis-svd/tree/master/data/Espressif-Community -interrupts=software lib/cmsis-svd/data/Espressif-Community/ src/device/esp/
+	GO111MODULE=off $(GO) fmt ./src/device/esp
+
 gen-device-nrf: build/gen-device-svd
 	./build/gen-device-svd -source=https://github.com/NordicSemiconductor/nrfx/tree/master/mdk lib/nrfx/mdk/ src/device/nrf/
 	GO111MODULE=off $(GO) fmt ./src/device/nrf

+gen-device-nxp: build/gen-device-svd
+	./build/gen-device-svd -source=https://github.com/posborne/cmsis-svd/tree/master/data/NXP lib/cmsis-svd/data/NXP/ src/device/nxp/
+	GO111MODULE=off $(GO) fmt ./src/device/nxp
+
 gen-device-sam: build/gen-device-svd
 	./build/gen-device-svd -source=https://github.com/posborne/cmsis-svd/tree/master/data/Atmel lib/cmsis-svd/data/Atmel/ src/device/sam/
 	GO111MODULE=off $(GO) fmt ./src/device/sam
@@ -124,48 +139,70 @@ gen-device-sifive: build/gen-device-svd
 	./build/gen-device-svd -source=https://github.com/posborne/cmsis-svd/tree/master/data/SiFive-Community -interrupts=software lib/cmsis-svd/data/SiFive-Community/ src/device/sifive/
 	GO111MODULE=off $(GO) fmt ./src/device/sifive

+gen-device-kendryte: build/gen-device-svd
+	./build/gen-device-svd -source=https://github.com/posborne/cmsis-svd/tree/master/data/Kendryte-Community -interrupts=software lib/cmsis-svd/data/Kendryte-Community/ src/device/kendryte/
+	GO111MODULE=off $(GO) fmt ./src/device/kendryte
+
 gen-device-stm32: build/gen-device-svd
-	./build/gen-device-svd -source=https://github.com/posborne/cmsis-svd/tree/master/data/STMicro lib/cmsis-svd/data/STMicro/ src/device/stm32/
+	./build/gen-device-svd -source=https://github.com/tinygo-org/stm32-svd lib/stm32-svd/svd src/device/stm32/
 	GO111MODULE=off $(GO) fmt ./src/device/stm32


 # Get LLVM sources.
-llvm-project/README.md:
-	git clone -b release/9.x https://github.com/llvm/llvm-project
-llvm-source: llvm-project/README.md
+$(LLVM_PROJECTDIR)/README.md:
+	git clone -b xtensa_release_11.0.0 --depth=1 https://github.com/tinygo-org/llvm-project $(LLVM_PROJECTDIR)
+llvm-source: $(LLVM_PROJECTDIR)/README.md

 # Configure LLVM.
 TINYGO_SOURCE_DIR=$(shell pwd)
 $(LLVM_BUILDDIR)/build.ninja: llvm-source
-	mkdir -p $(LLVM_BUILDDIR); cd $(LLVM_BUILDDIR); cmake -G Ninja $(TINYGO_SOURCE_DIR)/llvm-project/llvm "-DLLVM_TARGETS_TO_BUILD=X86;ARM;AArch64;RISCV;WebAssembly" "-DLLVM_EXPERIMENTAL_TARGETS_TO_BUILD=AVR" -DCMAKE_BUILD_TYPE=Release -DLIBCLANG_BUILD_STATIC=ON -DLLVM_ENABLE_TERMINFO=OFF -DLLVM_ENABLE_ZLIB=OFF -DLLVM_ENABLE_PROJECTS="clang;lld" -DLLVM_TOOL_CLANG_TOOLS_EXTRA_BUILD=OFF $(LLVM_OPTION)
+	mkdir -p $(LLVM_BUILDDIR); cd $(LLVM_BUILDDIR); cmake -G Ninja $(TINYGO_SOURCE_DIR)/$(LLVM_PROJECTDIR)/llvm "-DLLVM_TARGETS_TO_BUILD=X86;ARM;AArch64;RISCV;WebAssembly" "-DLLVM_EXPERIMENTAL_TARGETS_TO_BUILD=AVR;Xtensa" -DCMAKE_BUILD_TYPE=Release -DLIBCLANG_BUILD_STATIC=ON -DLLVM_ENABLE_TERMINFO=OFF -DLLVM_ENABLE_ZLIB=OFF -DLLVM_ENABLE_LIBEDIT=OFF -DLLVM_ENABLE_Z3_SOLVER=OFF -DLLVM_ENABLE_OCAMLDOC=OFF -DLLVM_ENABLE_LIBXML2=OFF -DLLVM_ENABLE_PROJECTS="clang;lld" -DLLVM_TOOL_CLANG_TOOLS_EXTRA_BUILD=OFF $(LLVM_OPTION)

 # Build LLVM.
 $(LLVM_BUILDDIR): $(LLVM_BUILDDIR)/build.ninja
-	cd $(LLVM_BUILDDIR); ninja
+	cd $(LLVM_BUILDDIR); ninja $(NINJA_BUILD_TARGETS)


 # Build wasi-libc sysroot
 .PHONY: wasi-libc
 wasi-libc: lib/wasi-libc/sysroot/lib/wasm32-wasi/libc.a
 lib/wasi-libc/sysroot/lib/wasm32-wasi/libc.a:
+	@if [ ! -e lib/wasi-libc/Makefile ]; then echo "Submodules have not been downloaded. Please download them using:\n  git submodule update --init"; exit 1; fi
 	cd lib/wasi-libc && make -j4 WASM_CC=$(CLANG) WASM_AR=$(LLVM_AR) WASM_NM=$(LLVM_NM)


 # Build the Go compiler.
 tinygo:
 	@if [ ! -f "$(LLVM_BUILDDIR)/bin/llvm-config" ]; then echo "Fetch and build LLVM first by running:"; echo "  make llvm-source"; echo "  make $(LLVM_BUILDDIR)"; exit 1; fi
-	CGO_CPPFLAGS="$(CGO_CPPFLAGS)" CGO_CXXFLAGS="$(CGO_CXXFLAGS)" CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GO) build -o build/tinygo$(EXE) -tags byollvm .
+	CGO_CPPFLAGS="$(CGO_CPPFLAGS)" CGO_CXXFLAGS="$(CGO_CXXFLAGS)" CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GO) build -buildmode exe -o build/tinygo$(EXE) -tags byollvm -ldflags="-X main.gitSha1=`git rev-parse --short HEAD`" .

 test: wasi-libc
-	CGO_CPPFLAGS="$(CGO_CPPFLAGS)" CGO_CXXFLAGS="$(CGO_CXXFLAGS)" CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GO) test -v -tags byollvm ./cgo ./compileopts ./interp ./transform .
+	CGO_CPPFLAGS="$(CGO_CPPFLAGS)" CGO_CXXFLAGS="$(CGO_CXXFLAGS)" CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GO) test -v -buildmode exe -tags byollvm ./cgo ./compileopts ./compiler ./interp ./transform .

+# Test known-working standard library packages.
+# TODO: do this in one command, parallelize, and only show failing tests (no
+# implied -v flag).
+.PHONY: tinygo-test
 tinygo-test:
-	cd tests/tinygotest && tinygo test
+	$(TINYGO) test container/heap
+	$(TINYGO) test container/list
+	$(TINYGO) test container/ring
+	$(TINYGO) test crypto/des
+	$(TINYGO) test encoding/ascii85
+	$(TINYGO) test encoding/base32
+	$(TINYGO) test encoding/hex
+	$(TINYGO) test hash/adler32
+	$(TINYGO) test hash/fnv
+	$(TINYGO) test hash/crc64
+	$(TINYGO) test math
+	$(TINYGO) test math/cmplx
+	$(TINYGO) test text/scanner
+	$(TINYGO) test unicode/utf8

 .PHONY: smoketest
 smoketest:
 	$(TINYGO) version
-	# test all examples
+	# test all examples (except pwm)
 	$(TINYGO) build -size short -o test.hex -target=pca10040            examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=pca10040            examples/adc
@@ -186,7 +223,9 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=microbit            examples/microbit-blink
 	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=pca10040            examples/pwm
+	$(TINYGO) build -size short -o test.hex -target=microbit-v2         examples/microbit-blink
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=pca10040            examples/pininterrupt
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=pca10040            examples/serial
 	@$(MD5SUM) test.hex
@@ -212,6 +251,8 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=microbit            examples/echo
 	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=microbit-s110v8     examples/echo
+	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=nrf52840-mdk        examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=pca10031            examples/blinky1
@@ -238,10 +279,16 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=stm32f4disco        examples/blinky2
 	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=stm32f4disco-1      examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=feather-stm32f405   examples/blinky1
+	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=circuitplay-bluefruit examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=circuitplay-express examples/i2s
 	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=clue-alpha          examples/blinky1
+	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.gba -target=gameboy-advance     examples/gba-display
 	@$(MD5SUM) test.gba
 	$(TINYGO) build -size short -o test.hex -target=itsybitsy-m4        examples/blinky1
@@ -254,35 +301,107 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=pyportal            examples/blinky1
 	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=particle-argon      examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=particle-boron      examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=particle-xenon      examples/blinky1
+	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=nucleo-f103rb       examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=pinetime-devkit0    examples/blinky1
 	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=lgt92    examples/blinky1
+	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=x9pro               examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=pca10056-s140v7     examples/blinky1
 	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=reelboard-s140v7    examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=wioterminal         examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=pygamer             examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=xiao                examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=circuitplay-express examples/dac
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=pyportal            examples/dac
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=feather-nrf52840  	examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=itsybitsy-nrf52840  examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=qtpy  				examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=teensy40            examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=teensy36            examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=nucleo-f722ze       examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=nucleo-l552ze       examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=p1am-100            examples/blinky1
+	@$(MD5SUM) test.hex
+	# test pwm
+	$(TINYGO) build -size short -o test.hex -target=itsybitsy-m0        examples/pwm
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=itsybitsy-m4        examples/pwm
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=feather-m4          examples/pwm
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=pyportal            examples/pwm
+	@$(MD5SUM) test.hex
 ifneq ($(AVR), 0)
+	$(TINYGO) build -size short -o test.hex -target=atmega1284p         examples/serial
+	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=arduino             examples/blinky1
 	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=arduino             examples/pwm
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=arduino -scheduler=tasks  examples/blinky1
+	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=arduino-nano        examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=digispark           examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=digispark -gc=leaking examples/blinky1
 	@$(MD5SUM) test.hex
+endif
+ifneq ($(XTENSA), 0)
+	$(TINYGO) build -size short -o test.bin -target=esp32-mini32      	examples/blinky1
+	@$(MD5SUM) test.bin
+	$(TINYGO) build -size short -o test.bin -target=nodemcu             examples/blinky1
+	@$(MD5SUM) test.bin
 endif
 	$(TINYGO) build -size short -o test.hex -target=hifive1b            examples/blinky1
 	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=hifive1-qemu        examples/serial
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=maixbit             examples/blinky1
+	@$(MD5SUM) test.hex
 	$(TINYGO) build             -o wasm.wasm -target=wasm               examples/wasm/export
 	$(TINYGO) build             -o wasm.wasm -target=wasm               examples/wasm/main
+	# test various compiler flags
+	$(TINYGO) build -size short -o test.hex -target=pca10040 -gc=none -scheduler=none examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=pca10040 -opt=1     examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build             -o test.nro -target=nintendoswitch      examples/serial
+	@$(MD5SUM) test.nro

-release: tinygo gen-device wasi-libc
+wasmtest:
+	$(GO) test ./tests/wasm
+
+build/release: tinygo gen-device wasi-libc
 	@mkdir -p build/release/tinygo/bin
 	@mkdir -p build/release/tinygo/lib/clang/include
 	@mkdir -p build/release/tinygo/lib/CMSIS/CMSIS
 	@mkdir -p build/release/tinygo/lib/compiler-rt/lib
 	@mkdir -p build/release/tinygo/lib/nrfx
+	@mkdir -p build/release/tinygo/lib/picolibc/newlib/libc
 	@mkdir -p build/release/tinygo/lib/wasi-libc
 	@mkdir -p build/release/tinygo/pkg/armv6m-none-eabi
 	@mkdir -p build/release/tinygo/pkg/armv7m-none-eabi
@@ -296,10 +415,28 @@ release: tinygo gen-device wasi-libc
 	@cp -rp lib/compiler-rt/LICENSE.TXT  build/release/tinygo/lib/compiler-rt
 	@cp -rp lib/compiler-rt/README.txt   build/release/tinygo/lib/compiler-rt
 	@cp -rp lib/nrfx/*                   build/release/tinygo/lib/nrfx
+	@cp -rp lib/picolibc/newlib/libc/ctype       build/release/tinygo/lib/picolibc/newlib/libc
+	@cp -rp lib/picolibc/newlib/libc/include     build/release/tinygo/lib/picolibc/newlib/libc
+	@cp -rp lib/picolibc/newlib/libc/locale      build/release/tinygo/lib/picolibc/newlib/libc
+	@cp -rp lib/picolibc/newlib/libc/string      build/release/tinygo/lib/picolibc/newlib/libc
+	@cp -rp lib/picolibc/newlib/libc/tinystdio   build/release/tinygo/lib/picolibc/newlib/libc
+	@cp -rp lib/picolibc-include         build/release/tinygo/lib
 	@cp -rp lib/wasi-libc/sysroot        build/release/tinygo/lib/wasi-libc/sysroot
 	@cp -rp src                          build/release/tinygo/src
 	@cp -rp targets                      build/release/tinygo/targets
-	./build/tinygo build-builtins -target=armv6m-none-eabi  -o build/release/tinygo/pkg/armv6m-none-eabi/compiler-rt.a
-	./build/tinygo build-builtins -target=armv7m-none-eabi  -o build/release/tinygo/pkg/armv7m-none-eabi/compiler-rt.a
-	./build/tinygo build-builtins -target=armv7em-none-eabi -o build/release/tinygo/pkg/armv7em-none-eabi/compiler-rt.a
+	./build/tinygo build-library -target=armv6m-none-eabi  -o build/release/tinygo/pkg/armv6m-none-eabi/compiler-rt.a compiler-rt
+	./build/tinygo build-library -target=armv7m-none-eabi  -o build/release/tinygo/pkg/armv7m-none-eabi/compiler-rt.a compiler-rt
+	./build/tinygo build-library -target=armv7em-none-eabi -o build/release/tinygo/pkg/armv7em-none-eabi/compiler-rt.a compiler-rt
+	./build/tinygo build-library -target=armv6m-none-eabi  -o build/release/tinygo/pkg/armv6m-none-eabi/picolibc.a picolibc
+	./build/tinygo build-library -target=armv7m-none-eabi  -o build/release/tinygo/pkg/armv7m-none-eabi/picolibc.a picolibc
+	./build/tinygo build-library -target=armv7em-none-eabi -o build/release/tinygo/pkg/armv7em-none-eabi/picolibc.a picolibc
+
+release: build/release
 	tar -czf build/release.tar.gz -C build/release tinygo
+
+deb: build/release
+	@mkdir -p build/release-deb/usr/local/bin
+	@mkdir -p build/release-deb/usr/local/lib
+	cp -ar build/release/tinygo build/release-deb/usr/local/lib/tinygo
+	ln -sf ../lib/tinygo/bin/tinygo build/release-deb/usr/local/bin/tinygo
+	fpm -f -s dir -t deb -n tinygo -v $(shell grep "const Version = " goenv/version.go | awk '{print $$NF}') -m '@tinygo-org' --description='TinyGo is a Go compiler for small places.' --license='BSD 3-Clause' --url=https://tinygo.org/ --deb-changelog CHANGELOG.md -p build/release.deb -C ./build/release-deb
@@ -43,33 +43,59 @@ See the [getting started instructions](https://tinygo.org/getting-started/) for

 You can compile TinyGo programs for microcontrollers, WebAssembly and Linux.

-The following 27 microcontroller boards are currently supported:
+The following 53 microcontroller boards are currently supported:

 * [Adafruit Circuit Playground Bluefruit](https://www.adafruit.com/product/4333)
 * [Adafruit Circuit Playground Express](https://www.adafruit.com/product/3333)
+* [Adafruit CLUE](https://www.adafruit.com/product/4500)
 * [Adafruit Feather M0](https://www.adafruit.com/product/2772)
 * [Adafruit Feather M4](https://www.adafruit.com/product/3857)
+* [Adafruit Feather nRF52840 Express](https://www.adafruit.com/product/4062)
+* [Adafruit Feather STM32F405 Express](https://www.adafruit.com/product/4382)
 * [Adafruit ItsyBitsy M0](https://www.adafruit.com/product/3727)
 * [Adafruit ItsyBitsy M4](https://www.adafruit.com/product/3800)
+* [Adafruit ItsyBitsy nRF52840](https://www.adafruit.com/product/4481)
+* [Adafruit Matrix Portal M4](https://www.adafruit.com/product/4745)
 * [Adafruit Metro M4 Express Airlift](https://www.adafruit.com/product/4000)
 * [Adafruit PyBadge](https://www.adafruit.com/product/4200)
+* [Adafruit PyGamer](https://www.adafruit.com/product/4242)
 * [Adafruit PyPortal](https://www.adafruit.com/product/4116)
 * [Adafruit Trinket M0](https://www.adafruit.com/product/3500)
+* [Arduino Mega 2560](https://store.arduino.cc/arduino-mega-2560-rev3)
+* [Arduino MKR1000](https://store.arduino.cc/arduino-mkr1000-wifi)
 * [Arduino Nano](https://store.arduino.cc/arduino-nano)
 * [Arduino Nano33 IoT](https://store.arduino.cc/nano-33-iot)
 * [Arduino Uno](https://store.arduino.cc/arduino-uno-rev3)
+* [Arduino Zero](https://store.arduino.cc/usa/arduino-zero)
 * [BBC micro:bit](https://microbit.org/)
+* [BBC micro:bit v2](https://microbit.org/new-microbit/)
 * [Digispark](http://digistump.com/products/1)
+* [Dragino LoRaWAN GPS Tracker LGT-92](http://www.dragino.com/products/lora-lorawan-end-node/item/142-lgt-92.html)
+* [ESP32](https://www.espressif.com/en/products/socs/esp32)
+* [ESP8266](https://www.espressif.com/en/products/socs/esp8266)
 * [Game Boy Advance](https://en.wikipedia.org/wiki/Game_Boy_Advance)
 * [Makerdiary nRF52840-MDK](https://wiki.makerdiary.com/nrf52840-mdk/)
+* [Makerdiary nRF52840-MDK USB Dongle](https://wiki.makerdiary.com/nrf52840-mdk-usb-dongle/)
+* [nice!nano](https://docs.nicekeyboards.com/#/nice!nano/)
+* [Nintendo Switch](https://www.nintendo.com/switch/)
 * [Nordic Semiconductor PCA10031](https://www.nordicsemi.com/eng/Products/nRF51-Dongle)
 * [Nordic Semiconductor PCA10040](https://www.nordicsemi.com/eng/Products/Bluetooth-low-energy/nRF52-DK)
 * [Nordic Semiconductor PCA10056](https://www.nordicsemi.com/Software-and-Tools/Development-Kits/nRF52840-DK)
+* [Particle Argon](https://docs.particle.io/datasheets/wi-fi/argon-datasheet/)
+* [Particle Boron](https://docs.particle.io/datasheets/cellular/boron-datasheet/)
+* [Particle Xenon](https://docs.particle.io/datasheets/discontinued/xenon-datasheet/)
 * [Phytec reel board](https://www.phytec.eu/product-eu/internet-of-things/reelboard/)
 * [PineTime DevKit](https://www.pine64.org/pinetime/)
+* [PJRC Teensy 3.6](https://www.pjrc.com/store/teensy36.html)
+* [ProductivityOpen P1AM-100](https://facts-engineering.github.io/modules/P1AM-100/P1AM-100.html)
+* [Seeed Wio Terminal](https://www.seeedstudio.com/Wio-Terminal-p-4509.html)
+* [Seeed Seeeduino XIAO](https://www.seeedstudio.com/Seeeduino-XIAO-Arduino-Microcontroller-SAMD21-Cortex-M0+-p-4426.html)
+* [Seeed Sipeed MAix BiT](https://www.seeedstudio.com/Sipeed-MAix-BiT-for-RISC-V-AI-IoT-p-2872.html)
 * [SiFIve HiFive1](https://www.sifive.com/boards/hifive1)
-* [ST Micro "Nucleo F103RB"](https://www.st.com/en/evaluation-tools/nucleo-f103rb.html)
-* [ST Micro STM32F103XX "Bluepill"](http://wiki.stm32duino.com/index.php?title=Blue_Pill)
+* [ST Micro "Nucleo" F103RB](https://www.st.com/en/evaluation-tools/nucleo-f103rb.html)
+* [ST Micro "Nucleo" F722ZE](https://www.st.com/en/evaluation-tools/nucleo-f722ze.html)
+* [ST Micro "Nucleo" L552ZE](https://www.st.com/en/evaluation-tools/nucleo-l552ze-q.html)
+* [ST Micro STM32F103XX "Bluepill"](https://stm32-base.org/boards/STM32F103C8T6-Blue-Pill)
 * [ST Micro STM32F407 "Discovery"](https://www.st.com/en/evaluation-tools/stm32f4discovery.html)
 * [X9 Pro smartwatch](https://github.com/curtpw/nRF5x-device-reverse-engineering/tree/master/X9-nrf52832-activity-tracker/)

@@ -131,4 +157,6 @@ The original reasoning was: if [Python](https://micropython.org/) can run on mic

 This project is licensed under the BSD 3-clause license, just like the [Go project](https://golang.org/LICENSE) itself.

-Some code has been copied from the LLVM project and is therefore licensed under [a variant of the Apache 2.0 license](http://releases.llvm.org/9.0.0/LICENSE.TXT). This has been clearly indicated in the header of these files.
+Some code has been copied from the LLVM project and is therefore licensed under [a variant of the Apache 2.0 license](http://releases.llvm.org/11.0.0/LICENSE.TXT). This has been clearly indicated in the header of these files.
+
+Some code has been copied and/or ported from Paul Stoffregen's Teensy libraries and is therefore licensed under PJRC's license. This has been clearly indicated in the header of these files.
@@ -1,7 +1,7 @@
 # Avoid lengthy LLVM rebuilds on each newly pushed branch. Pull requests will
 # be built anyway.
 trigger:
- master
+- release
 - dev

 jobs:
@@ -10,11 +10,15 @@ jobs:
  pool:
    vmImage: 'VS2017-Win2016'
  steps:
+    - task: GoTool@0
+      inputs:
+        version: '1.16'
    - checkout: self
-    - task: CacheBeta@0
+      fetchDepth: 1
+    - task: Cache@2
      displayName: Cache LLVM source
      inputs:
-        key: llvm-source-9-windows-v0
+        key: llvm-source-11-windows-v1
        path: llvm-project
    - task: Bash@3
      displayName: Download LLVM source
@@ -24,7 +28,7 @@ jobs:
    - task: CacheBeta@0
      displayName: Cache LLVM build
      inputs:
-        key: llvm-build-9-windows-v0
+        key: llvm-build-11-windows-v4
        path: llvm-build
    - task: Bash@3
      displayName: Build LLVM
@@ -33,34 +37,34 @@ jobs:
        script: |
          if [ ! -f llvm-build/lib/liblldELF.a ]
          then
+            # install dependencies
            choco install ninja
-            # LLVM 9 cannot be built with MinGW 8.
-            # For details: https://reviews.llvm.org/D70266
-            choco uninstall mingw
-            choco install mingw --version=7.3.0
+            # hack ninja to use fewer jobs
+            echo -e 'C:\\ProgramData\\Chocolatey\\bin\\ninja -j4 %*' > /usr/bin/ninja.bat
+            # build!
            make llvm-build
          fi
    - task: Bash@3
      displayName: Install QEMU
      inputs:
        targetType: inline
-        script: choco install qemu
+        script: choco install qemu --version=2020.06.12
    - task: CacheBeta@0
      displayName: Cache wasi-libc sysroot
      inputs:
-        key: wasi-libc-sysroot-v0
+        key: wasi-libc-sysroot-v4
        path: lib/wasi-libc/sysroot
    - task: Bash@3
      displayName: Build wasi-libc
      inputs:
        targetType: inline
-        script: PATH=/usr/bin:$PATH make wasi-libc CLANG=$PWD/llvm-build/bin/clang LLVM_AR=$PWD/llvm-build/bin/llvm-ar LLVM_NM=$PWD/llvm-build/bin/llvm-nm
+        script: PATH=/usr/bin:$PATH make wasi-libc
    - task: Bash@3
      displayName: Test TinyGo
      inputs:
        targetType: inline
        script: |
-          export PATH="/c/Go1.13/bin:$PATH:./llvm-build/bin:/c/Program Files/qemu"
+          export PATH="$PATH:./llvm-build/bin:/c/Program Files/qemu"
          unset GOROOT
          make test
    - task: Bash@3
@@ -68,9 +72,9 @@ jobs:
      inputs:
        targetType: inline
        script: |
-          export PATH="/c/Go1.13/bin:$PATH:./llvm-build/bin:/c/Program Files/qemu"
+          export PATH="$PATH:./llvm-build/bin:/c/Program Files/qemu"
          unset GOROOT
-          make release -j4
+          make build/release -j4
    - publish: $(System.DefaultWorkingDirectory)/build/release/tinygo
      displayName: Publish zip as artifact
      artifact: tinygo
@@ -79,6 +83,6 @@ jobs:
      inputs:
        targetType: inline
        script: |
-          export PATH="/c/Go1.13/bin:$PATH:./llvm-build/bin:/c/Program Files/qemu"
+          export PATH="$PATH:./llvm-build/bin:/c/Program Files/qemu"
          unset GOROOT
-          make smoketest TINYGO=build/tinygo AVR=0
+          make smoketest TINYGO=build/tinygo AVR=0 XTENSA=0
@@ -4,11 +4,15 @@
 package builder

 import (
+	"debug/elf"
+	"encoding/binary"
 	"errors"
 	"fmt"
+	"go/types"
 	"io/ioutil"
 	"os"
 	"path/filepath"
+	"sort"
 	"strconv"
 	"strings"

@@ -16,44 +20,379 @@ import (
 	"github.com/tinygo-org/tinygo/compiler"
 	"github.com/tinygo-org/tinygo/goenv"
 	"github.com/tinygo-org/tinygo/interp"
+	"github.com/tinygo-org/tinygo/loader"
+	"github.com/tinygo-org/tinygo/stacksize"
 	"github.com/tinygo-org/tinygo/transform"
+	"tinygo.org/x/go-llvm"
 )

+// BuildResult is the output of a build. This includes the binary itself and
+// some other metadata that is obtained while building the binary.
+type BuildResult struct {
+	// A path to the output binary. It will be removed after Build returns, so
+	// if it should be kept it must be copied or moved away.
+	Binary string
+
+	// The directory of the main package. This is useful for testing as the test
+	// binary must be run in the directory of the tested package.
+	MainDir string
+}
+
 // Build performs a single package to executable Go build. It takes in a package
 // name, an output path, and set of compile options and from that it manages the
 // whole compilation process.
 //
 // The error value may be of type *MultiError. Callers will likely want to check
 // for this case and print such errors individually.
-func Build(pkgName, outpath string, config *compileopts.Config, action func(string) error) error {
-	c, err := compiler.NewCompiler(pkgName, config)
+func Build(pkgName, outpath string, config *compileopts.Config, action func(BuildResult) error) error {
+	compilerConfig := &compiler.Config{
+		Triple:          config.Triple(),
+		CPU:             config.CPU(),
+		Features:        config.Features(),
+		GOOS:            config.GOOS(),
+		GOARCH:          config.GOARCH(),
+		CodeModel:       config.CodeModel(),
+		RelocationModel: config.RelocationModel(),
+
+		Scheduler:          config.Scheduler(),
+		FuncImplementation: config.FuncImplementation(),
+		AutomaticStackSize: config.AutomaticStackSize(),
+		DefaultStackSize:   config.Target.DefaultStackSize,
+		NeedsStackObjects:  config.NeedsStackObjects(),
+		Debug:              config.Debug(),
+	}
+
+	// Load the target machine, which is the LLVM object that contains all
+	// details of a target (alignment restrictions, pointer size, default
+	// address spaces, etc).
+	machine, err := compiler.NewTargetMachine(compilerConfig)
 	if err != nil {
 		return err
 	}

-	// Compile Go code to IR.
-	errs := c.Compile(pkgName)
-	if len(errs) != 0 {
-		return newMultiError(errs)
+	// Load entire program AST into memory.
+	lprogram, err := loader.Load(config, []string{pkgName}, config.ClangHeaders, types.Config{
+		Sizes: compiler.Sizes(machine),
+	})
+	if err != nil {
+		return err
 	}
+	err = lprogram.Parse()
+	if err != nil {
+		return err
+	}
+
+	// The slice of jobs that orchestrates most of the build.
+	// This is somewhat like an in-memory Makefile with each job being a
+	// Makefile target.
+	var jobs []*compileJob
+
+	// Add job to compile and optimize all Go files at once.
+	// TODO: parallelize this.
+	var mod llvm.Module
+	var stackSizeLoads []string
+	programJob := &compileJob{
+		description: "compile Go files",
+		run: func() (err error) {
+			mod, err = compileWholeProgram(pkgName, config, compilerConfig, lprogram, machine)
+			if err != nil {
+				return
+			}
+			// Make sure stack sizes are loaded from a separate section so they can be
+			// modified after linking.
+			if config.AutomaticStackSize() {
+				stackSizeLoads = transform.CreateStackSizeLoads(mod, config)
+			}
+			return
+		},
+	}
+	jobs = append(jobs, programJob)
+
+	// Check whether we only need to create an object file.
+	// If so, we don't need to link anything and will be finished quickly.
+	outext := filepath.Ext(outpath)
+	if outext == ".o" || outext == ".bc" || outext == ".ll" {
+		// Run jobs to produce the LLVM module.
+		err := runJobs(jobs)
+		if err != nil {
+			return err
+		}
+		// Generate output.
+		switch outext {
+		case ".o":
+			llvmBuf, err := machine.EmitToMemoryBuffer(mod, llvm.ObjectFile)
+			if err != nil {
+				return err
+			}
+			return ioutil.WriteFile(outpath, llvmBuf.Bytes(), 0666)
+		case ".bc":
+			data := llvm.WriteBitcodeToMemoryBuffer(mod).Bytes()
+			return ioutil.WriteFile(outpath, data, 0666)
+		case ".ll":
+			data := []byte(mod.String())
+			return ioutil.WriteFile(outpath, data, 0666)
+		default:
+			panic("unreachable")
+		}
+	}
+
+	// Act as a compiler driver, as we need to produce a complete executable.
+	// First add all jobs necessary to build this object file, then afterwards
+	// run all jobs in parallel as far as possible.
+
+	// Create a temporary directory for intermediary files.
+	dir, err := ioutil.TempDir("", "tinygo")
+	if err != nil {
+		return err
+	}
+	defer os.RemoveAll(dir)
+
+	// Add job to write the output object file.
+	objfile := filepath.Join(dir, "main.o")
+	outputObjectFileJob := &compileJob{
+		description:  "generate output file",
+		dependencies: []*compileJob{programJob},
+		run: func() error {
+			llvmBuf, err := machine.EmitToMemoryBuffer(mod, llvm.ObjectFile)
+			if err != nil {
+				return err
+			}
+			return ioutil.WriteFile(objfile, llvmBuf.Bytes(), 0666)
+		},
+	}
+	jobs = append(jobs, outputObjectFileJob)
+
+	// Prepare link command.
+	linkerDependencies := []*compileJob{outputObjectFileJob}
+	executable := filepath.Join(dir, "main")
+	tmppath := executable // final file
+	ldflags := append(config.LDFlags(), "-o", executable, objfile)
+
+	// Add compiler-rt dependency if needed. Usually this is a simple load from
+	// a cache.
+	if config.Target.RTLib == "compiler-rt" {
+		path, job, err := CompilerRT.load(config.Triple(), config.CPU(), dir)
+		if err != nil {
+			return err
+		}
+		if job != nil {
+			// The library was not loaded from cache so needs to be compiled
+			// (and then stored in the cache).
+			jobs = append(jobs, job.dependencies...)
+			jobs = append(jobs, job)
+			linkerDependencies = append(linkerDependencies, job)
+		}
+		ldflags = append(ldflags, path)
+	}
+
+	// Add libc dependency if needed.
+	root := goenv.Get("TINYGOROOT")
+	switch config.Target.Libc {
+	case "picolibc":
+		path, job, err := Picolibc.load(config.Triple(), config.CPU(), dir)
+		if err != nil {
+			return err
+		}
+		if job != nil {
+			// The library needs to be compiled (cache miss).
+			jobs = append(jobs, job.dependencies...)
+			jobs = append(jobs, job)
+			linkerDependencies = append(linkerDependencies, job)
+		}
+		ldflags = append(ldflags, path)
+	case "wasi-libc":
+		path := filepath.Join(root, "lib/wasi-libc/sysroot/lib/wasm32-wasi/libc.a")
+		if _, err := os.Stat(path); os.IsNotExist(err) {
+			return errors.New("could not find wasi-libc, perhaps you need to run `make wasi-libc`?")
+		}
+		ldflags = append(ldflags, path)
+	case "":
+		// no library specified, so nothing to do
+	default:
+		return fmt.Errorf("unknown libc: %s", config.Target.Libc)
+	}
+
+	// Add jobs to compile extra files. These files are in C or assembly and
+	// contain things like the interrupt vector table and low level operations
+	// such as stack switching.
+	for i, path := range config.ExtraFiles() {
+		abspath := filepath.Join(root, path)
+		outpath := filepath.Join(dir, "extra-"+strconv.Itoa(i)+"-"+filepath.Base(path)+".o")
+		job := &compileJob{
+			description: "compile extra file " + path,
+			run: func() error {
+				err := runCCompiler(config.Target.Compiler, append(config.CFlags(), "-c", "-o", outpath, abspath)...)
+				if err != nil {
+					return &commandError{"failed to build", path, err}
+				}
+				return nil
+			},
+		}
+		jobs = append(jobs, job)
+		linkerDependencies = append(linkerDependencies, job)
+		ldflags = append(ldflags, outpath)
+	}
+
+	// Add jobs to compile C files in all packages. This is part of CGo.
+	// TODO: do this as part of building the package to be able to link the
+	// bitcode files together.
+	for i, pkg := range lprogram.Sorted() {
+		for j, filename := range pkg.CFiles {
+			file := filepath.Join(pkg.Dir, filename)
+			outpath := filepath.Join(dir, "pkg"+strconv.Itoa(i)+"."+strconv.Itoa(j)+"-"+filepath.Base(file)+".o")
+			job := &compileJob{
+				description: "compile CGo file " + file,
+				run: func() error {
+					err := runCCompiler(config.Target.Compiler, append(config.CFlags(), "-c", "-o", outpath, file)...)
+					if err != nil {
+						return &commandError{"failed to build", file, err}
+					}
+					return nil
+				},
+			}
+			jobs = append(jobs, job)
+			linkerDependencies = append(linkerDependencies, job)
+			ldflags = append(ldflags, outpath)
+		}
+	}
+
+	// Linker flags from CGo lines:
+	//     #cgo LDFLAGS: foo
+	if len(lprogram.LDFlags) > 0 {
+		ldflags = append(ldflags, lprogram.LDFlags...)
+	}
+
+	// Create a linker job, which links all object files together and does some
+	// extra stuff that can only be done after linking.
+	jobs = append(jobs, &compileJob{
+		description:  "link",
+		dependencies: linkerDependencies,
+		run: func() error {
+			err = link(config.Target.Linker, ldflags...)
+			if err != nil {
+				return &commandError{"failed to link", executable, err}
+			}
+
+			var calculatedStacks []string
+			var stackSizes map[string]functionStackSize
+			if config.Options.PrintStacks || config.AutomaticStackSize() {
+				// Try to determine stack sizes at compile time.
+				// Don't do this by default as it usually doesn't work on
+				// unsupported architectures.
+				calculatedStacks, stackSizes, err = determineStackSizes(mod, executable)
+				if err != nil {
+					return err
+				}
+			}
+			if config.AutomaticStackSize() {
+				// Modify the .tinygo_stacksizes section that contains a stack size
+				// for each goroutine.
+				err = modifyStackSizes(executable, stackSizeLoads, stackSizes)
+				if err != nil {
+					return fmt.Errorf("could not modify stack sizes: %w", err)
+				}
+			}
+
+			if config.Options.PrintSizes == "short" || config.Options.PrintSizes == "full" {
+				sizes, err := loadProgramSize(executable)
+				if err != nil {
+					return err
+				}
+				if config.Options.PrintSizes == "short" {
+					fmt.Printf("   code    data     bss |   flash     ram\n")
+					fmt.Printf("%7d %7d %7d | %7d %7d\n", sizes.Code, sizes.Data, sizes.BSS, sizes.Code+sizes.Data, sizes.Data+sizes.BSS)
+				} else {
+					fmt.Printf("   code  rodata    data     bss |   flash     ram | package\n")
+					for _, name := range sizes.sortedPackageNames() {
+						pkgSize := sizes.Packages[name]
+						fmt.Printf("%7d %7d %7d %7d | %7d %7d | %s\n", pkgSize.Code, pkgSize.ROData, pkgSize.Data, pkgSize.BSS, pkgSize.Flash(), pkgSize.RAM(), name)
+					}
+					fmt.Printf("%7d %7d %7d %7d | %7d %7d | (sum)\n", sizes.Sum.Code, sizes.Sum.ROData, sizes.Sum.Data, sizes.Sum.BSS, sizes.Sum.Flash(), sizes.Sum.RAM())
+					fmt.Printf("%7d       - %7d %7d | %7d %7d | (all)\n", sizes.Code, sizes.Data, sizes.BSS, sizes.Code+sizes.Data, sizes.Data+sizes.BSS)
+				}
+			}
+
+			// Print goroutine stack sizes, as far as possible.
+			if config.Options.PrintStacks {
+				printStacks(calculatedStacks, stackSizes)
+			}
+
+			return nil
+		},
+	})
+
+	// Run all jobs to compile and link the program.
+	// Do this now (instead of after elf-to-hex and similar conversions) as it
+	// is simpler and cannot be parallelized.
+	err = runJobs(jobs)
+	if err != nil {
+		return err
+	}
+
+	// Get an Intel .hex file or .bin file from the .elf file.
+	outputBinaryFormat := config.BinaryFormat(outext)
+	switch outputBinaryFormat {
+	case "elf":
+		// do nothing, file is already in ELF format
+	case "hex", "bin":
+		// Extract raw binary, either encoding it as a hex file or as a raw
+		// firmware file.
+		tmppath = filepath.Join(dir, "main"+outext)
+		err := objcopy(executable, tmppath, outputBinaryFormat)
+		if err != nil {
+			return err
+		}
+	case "uf2":
+		// Get UF2 from the .elf file.
+		tmppath = filepath.Join(dir, "main"+outext)
+		err := convertELFFileToUF2File(executable, tmppath, config.Target.UF2FamilyID)
+		if err != nil {
+			return err
+		}
+	case "esp32", "esp8266":
+		// Special format for the ESP family of chips (parsed by the ROM
+		// bootloader).
+		tmppath = filepath.Join(dir, "main"+outext)
+		err := makeESPFirmareImage(executable, tmppath, outputBinaryFormat)
+		if err != nil {
+			return err
+		}
+	default:
+		return fmt.Errorf("unknown output binary format: %s", outputBinaryFormat)
+	}
+	return action(BuildResult{
+		Binary:  tmppath,
+		MainDir: lprogram.MainPkg().Dir,
+	})
+}
+
+// compileWholeProgram compiles the entire *loader.Program to a LLVM module and
+// applies most necessary optimizations and transformations.
+func compileWholeProgram(pkgName string, config *compileopts.Config, compilerConfig *compiler.Config, lprogram *loader.Program, machine llvm.TargetMachine) (llvm.Module, error) {
+	// Compile AST to IR.
+	mod, errs := compiler.CompileProgram(lprogram, machine, compilerConfig, config.DumpSSA())
+	if errs != nil {
+		return mod, newMultiError(errs)
+	}
+
 	if config.Options.PrintIR {
 		fmt.Println("; Generated LLVM IR:")
-		fmt.Println(c.IR())
+		fmt.Println(mod.String())
 	}
-	if err := c.Verify(); err != nil {
-		return errors.New("verification error after IR construction")
+	if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
+		return mod, errors.New("verification error after IR construction")
 	}

-	err = interp.Run(c.Module(), config.DumpSSA())
+	err := interp.Run(mod, config.DumpSSA())
 	if err != nil {
-		return err
+		return mod, err
 	}
-	if err := c.Verify(); err != nil {
-		return errors.New("verification error after interpreting runtime.initAll")
+	if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
+		return mod, errors.New("verification error after interpreting runtime.initAll")
 	}

 	if config.GOOS() != "darwin" {
-		c.ApplyFunctionSections() // -ffunction-sections
+		transform.ApplyFunctionSections(mod) // -ffunction-sections
 	}

 	// Browsers cannot handle external functions that have type i64 because it
@@ -61,10 +400,10 @@ func Build(pkgName, outpath string, config *compileopts.Config, action func(stri
 	// keep functions interoperable, pass int64 types as pointers to
 	// stack-allocated values.
 	// Use -wasm-abi=generic to disable this behaviour.
-	if config.Options.WasmAbi == "js" && strings.HasPrefix(config.Triple(), "wasm") {
-		err := transform.ExternalInt64AsPtr(c.Module())
+	if config.WasmAbi() == "js" {
+		err := transform.ExternalInt64AsPtr(mod)
 		if err != nil {
-			return err
+			return mod, err
 		}
 	}

@@ -72,147 +411,276 @@ func Build(pkgName, outpath string, config *compileopts.Config, action func(stri
 	// exactly.
 	errs = nil
 	switch config.Options.Opt {
-	case "none:", "0":
-		errs = c.Optimize(0, 0, 0) // -O0
+	/*
+		Currently, turning optimizations off causes compile failures.
+		We rely on the optimizer removing some dead symbols.
+		Avoid providing an option that does not work right now.
+		In the future once everything has been fixed we can re-enable this.
+
+		case "none", "0":
+			errs = transform.Optimize(mod, config, 0, 0, 0) // -O0
+	*/
 	case "1":
-		errs = c.Optimize(1, 0, 0) // -O1
+		errs = transform.Optimize(mod, config, 1, 0, 0) // -O1
 	case "2":
-		errs = c.Optimize(2, 0, 225) // -O2
+		errs = transform.Optimize(mod, config, 2, 0, 225) // -O2
 	case "s":
-		errs = c.Optimize(2, 1, 225) // -Os
+		errs = transform.Optimize(mod, config, 2, 1, 225) // -Os
 	case "z":
-		errs = c.Optimize(2, 2, 5) // -Oz, default
+		errs = transform.Optimize(mod, config, 2, 2, 5) // -Oz, default
 	default:
 		errs = []error{errors.New("unknown optimization level: -opt=" + config.Options.Opt)}
 	}
 	if len(errs) > 0 {
-		return newMultiError(errs)
+		return mod, newMultiError(errs)
 	}
-	if err := c.Verify(); err != nil {
-		return errors.New("verification failure after LLVM optimization passes")
+	if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
+		return mod, errors.New("verification failure after LLVM optimization passes")
 	}

-	// On the AVR, pointers can point either to flash or to RAM, but we don't
-	// know. As a temporary fix, load all global variables in RAM.
-	// In the future, there should be a compiler pass that determines which
-	// pointers are flash and which are in RAM so that pointers can have a
-	// correct address space parameter (address space 1 is for flash).
-	if strings.HasPrefix(config.Triple(), "avr") {
-		c.NonConstGlobals()
-		if err := c.Verify(); err != nil {
-			return errors.New("verification error after making all globals non-constant on AVR")
-		}
+	// LLVM 11 by default tries to emit tail calls (even with the target feature
+	// disabled) unless it is explicitly disabled with a function attribute.
+	// This is a problem, as it tries to emit them and prints an error when it
+	// can't with this feature disabled.
+	// Because as of september 2020 tail calls are not yet widely supported,
+	// they need to be disabled until they are widely supported (at which point
+	// the +tail-call target feautre can be set).
+	if strings.HasPrefix(config.Triple(), "wasm") {
+		transform.DisableTailCalls(mod)
 	}

-	// Generate output.
-	outext := filepath.Ext(outpath)
-	switch outext {
-	case ".o":
-		return c.EmitObject(outpath)
-	case ".bc":
-		return c.EmitBitcode(outpath)
-	case ".ll":
-		return c.EmitText(outpath)
-	default:
-		// Act as a compiler driver.
+	return mod, nil
+}

-		// Create a temporary directory for intermediary files.
-		dir, err := ioutil.TempDir("", "tinygo")
-		if err != nil {
-			return err
-		}
-		defer os.RemoveAll(dir)
+// functionStackSizes keeps stack size information about a single function
+// (usually a goroutine).
+type functionStackSize struct {
+	humanName        string
+	stackSize        uint64
+	stackSizeType    stacksize.SizeType
+	missingStackSize *stacksize.CallNode
+}

-		// Write the object file.
-		objfile := filepath.Join(dir, "main.o")
-		err = c.EmitObject(objfile)
-		if err != nil {
-			return err
-		}
-
-		// Load builtins library from the cache, possibly compiling it on the
-		// fly.
-		var librt string
-		if config.Target.RTLib == "compiler-rt" {
-			librt, err = loadBuiltins(config.Triple())
-			if err != nil {
-				return err
-			}
-		}
-
-		// Prepare link command.
-		executable := filepath.Join(dir, "main")
-		tmppath := executable // final file
-		ldflags := append(config.LDFlags(), "-o", executable, objfile)
-		if config.Target.RTLib == "compiler-rt" {
-			ldflags = append(ldflags, librt)
-		}
-
-		// Compile extra files.
-		root := goenv.Get("TINYGOROOT")
-		for i, path := range config.ExtraFiles() {
-			abspath := filepath.Join(root, path)
-			outpath := filepath.Join(dir, "extra-"+strconv.Itoa(i)+"-"+filepath.Base(path)+".o")
-			err := runCCompiler(config.Target.Compiler, append(config.CFlags(), "-c", "-o", outpath, abspath)...)
-			if err != nil {
-				return &commandError{"failed to build", path, err}
-			}
-			ldflags = append(ldflags, outpath)
-		}
-
-		// Compile C files in packages.
-		for i, pkg := range c.Packages() {
-			for _, file := range pkg.CFiles {
-				path := filepath.Join(pkg.Package.Dir, file)
-				outpath := filepath.Join(dir, "pkg"+strconv.Itoa(i)+"-"+file+".o")
-				err := runCCompiler(config.Target.Compiler, append(config.CFlags(), "-c", "-o", outpath, path)...)
-				if err != nil {
-					return &commandError{"failed to build", path, err}
+// determineStackSizes tries to determine the stack sizes of all started
+// goroutines and of the reset vector. The LLVM module is necessary to find
+// functions that call a function pointer.
+func determineStackSizes(mod llvm.Module, executable string) ([]string, map[string]functionStackSize, error) {
+	var callsIndirectFunction []string
+	gowrappers := []string{}
+	gowrapperNames := make(map[string]string)
+	for fn := mod.FirstFunction(); !fn.IsNil(); fn = llvm.NextFunction(fn) {
+		// Determine which functions call a function pointer.
+		for bb := fn.FirstBasicBlock(); !bb.IsNil(); bb = llvm.NextBasicBlock(bb) {
+			for inst := bb.FirstInstruction(); !inst.IsNil(); inst = llvm.NextInstruction(inst) {
+				if inst.IsACallInst().IsNil() {
+					continue
 				}
-				ldflags = append(ldflags, outpath)
-			}
-		}
-
-		// Link the object files together.
-		err = link(config.Target.Linker, ldflags...)
-		if err != nil {
-			return &commandError{"failed to link", executable, err}
-		}
-
-		if config.Options.PrintSizes == "short" || config.Options.PrintSizes == "full" {
-			sizes, err := loadProgramSize(executable)
-			if err != nil {
-				return err
-			}
-			if config.Options.PrintSizes == "short" {
-				fmt.Printf("   code    data     bss |   flash     ram\n")
-				fmt.Printf("%7d %7d %7d | %7d %7d\n", sizes.Code, sizes.Data, sizes.BSS, sizes.Code+sizes.Data, sizes.Data+sizes.BSS)
-			} else {
-				fmt.Printf("   code  rodata    data     bss |   flash     ram | package\n")
-				for _, name := range sizes.sortedPackageNames() {
-					pkgSize := sizes.Packages[name]
-					fmt.Printf("%7d %7d %7d %7d | %7d %7d | %s\n", pkgSize.Code, pkgSize.ROData, pkgSize.Data, pkgSize.BSS, pkgSize.Flash(), pkgSize.RAM(), name)
+				if callee := inst.CalledValue(); callee.IsAFunction().IsNil() && callee.IsAInlineAsm().IsNil() {
+					callsIndirectFunction = append(callsIndirectFunction, fn.Name())
 				}
-				fmt.Printf("%7d %7d %7d %7d | %7d %7d | (sum)\n", sizes.Sum.Code, sizes.Sum.ROData, sizes.Sum.Data, sizes.Sum.BSS, sizes.Sum.Flash(), sizes.Sum.RAM())
-				fmt.Printf("%7d       - %7d %7d | %7d %7d | (all)\n", sizes.Code, sizes.Data, sizes.BSS, sizes.Code+sizes.Data, sizes.Data+sizes.BSS)
 			}
 		}

-		// Get an Intel .hex file or .bin file from the .elf file.
-		if outext == ".hex" || outext == ".bin" || outext == ".gba" {
-			tmppath = filepath.Join(dir, "main"+outext)
-			err := objcopy(executable, tmppath)
-			if err != nil {
-				return err
-			}
-		} else if outext == ".uf2" {
-			// Get UF2 from the .elf file.
-			tmppath = filepath.Join(dir, "main"+outext)
-			err := convertELFFileToUF2File(executable, tmppath, config.Target.UF2FamilyID)
-			if err != nil {
-				return err
-			}
+		// Get a list of "go wrappers", small wrapper functions that decode
+		// parameters when starting a new goroutine.
+		attr := fn.GetStringAttributeAtIndex(-1, "tinygo-gowrapper")
+		if !attr.IsNil() {
+			gowrappers = append(gowrappers, fn.Name())
+			gowrapperNames[fn.Name()] = attr.GetStringValue()
+		}
+	}
+	sort.Strings(gowrappers)
+
+	// Load the ELF binary.
+	f, err := elf.Open(executable)
+	if err != nil {
+		return nil, nil, fmt.Errorf("could not load executable for stack size analysis: %w", err)
+	}
+	defer f.Close()
+
+	// Determine the frame size of each function (if available) and the callgraph.
+	functions, err := stacksize.CallGraph(f, callsIndirectFunction)
+	if err != nil {
+		return nil, nil, fmt.Errorf("could not parse executable for stack size analysis: %w", err)
+	}
+
+	// Goroutines need to be started and finished and take up some stack space
+	// that way. This can be measured by measuing the stack size of
+	// tinygo_startTask.
+	if numFuncs := len(functions["tinygo_startTask"]); numFuncs != 1 {
+		return nil, nil, fmt.Errorf("expected exactly one definition of tinygo_startTask, got %d", numFuncs)
+	}
+	baseStackSize, baseStackSizeType, baseStackSizeFailedAt := functions["tinygo_startTask"][0].StackSize()
+
+	sizes := make(map[string]functionStackSize)
+
+	// Add the reset handler function, for convenience. The reset handler runs
+	// startup code and the scheduler. The listed stack size is not the full
+	// stack size: interrupts are not counted.
+	var resetFunction string
+	switch f.Machine {
+	case elf.EM_ARM:
+		// Note: all interrupts happen on this stack so the real size is bigger.
+		resetFunction = "Reset_Handler"
+	}
+	if resetFunction != "" {
+		funcs := functions[resetFunction]
+		if len(funcs) != 1 {
+			return nil, nil, fmt.Errorf("expected exactly one definition of %s in the callgraph, found %d", resetFunction, len(funcs))
+		}
+		stackSize, stackSizeType, missingStackSize := funcs[0].StackSize()
+		sizes[resetFunction] = functionStackSize{
+			stackSize:        stackSize,
+			stackSizeType:    stackSizeType,
+			missingStackSize: missingStackSize,
+			humanName:        resetFunction,
+		}
+	}
+
+	// Add all goroutine wrapper functions.
+	for _, name := range gowrappers {
+		funcs := functions[name]
+		if len(funcs) != 1 {
+			return nil, nil, fmt.Errorf("expected exactly one definition of %s in the callgraph, found %d", name, len(funcs))
+		}
+		humanName := gowrapperNames[name]
+		if humanName == "" {
+			humanName = name // fallback
+		}
+		stackSize, stackSizeType, missingStackSize := funcs[0].StackSize()
+		if baseStackSizeType != stacksize.Bounded {
+			// It was not possible to determine the stack size at compile time
+			// because tinygo_startTask does not have a fixed stack size. This
+			// can happen when using -opt=1.
+			stackSizeType = baseStackSizeType
+			missingStackSize = baseStackSizeFailedAt
+		} else if stackSize < baseStackSize {
+			// This goroutine has a very small stack, but still needs to fit all
+			// registers to start and suspend the goroutine. Otherwise a stack
+			// overflow will occur even before the goroutine is started.
+			stackSize = baseStackSize
+		}
+		sizes[name] = functionStackSize{
+			stackSize:        stackSize,
+			stackSizeType:    stackSizeType,
+			missingStackSize: missingStackSize,
+			humanName:        humanName,
+		}
+	}
+
+	if resetFunction != "" {
+		return append([]string{resetFunction}, gowrappers...), sizes, nil
+	}
+	return gowrappers, sizes, nil
+}
+
+// modifyStackSizes modifies the .tinygo_stacksizes section with the updated
+// stack size information. Before this modification, all stack sizes in the
+// section assume the default stack size (which is relatively big).
+func modifyStackSizes(executable string, stackSizeLoads []string, stackSizes map[string]functionStackSize) error {
+	fp, err := os.OpenFile(executable, os.O_RDWR, 0)
+	if err != nil {
+		return err
+	}
+	defer fp.Close()
+
+	elfFile, err := elf.NewFile(fp)
+	if err != nil {
+		return err
+	}
+
+	section := elfFile.Section(".tinygo_stacksizes")
+	if section == nil {
+		return errors.New("could not find .tinygo_stacksizes section")
+	}
+
+	if section.Size != section.FileSize {
+		// Sanity check.
+		return fmt.Errorf("expected .tinygo_stacksizes to have identical size and file size, got %d and %d", section.Size, section.FileSize)
+	}
+
+	// Read all goroutine stack sizes.
+	data := make([]byte, section.Size)
+	_, err = fp.ReadAt(data, int64(section.Offset))
+	if err != nil {
+		return err
+	}
+
+	if len(stackSizeLoads)*4 != len(data) {
+		// Note: while AVR should use 2 byte stack sizes, even 64-bit platforms
+		// should probably stick to 4 byte stack sizes as a larger than 4GB
+		// stack doesn't make much sense.
+		return errors.New("expected 4 byte stack sizes")
+	}
+
+	// Modify goroutine stack sizes with a compile-time known worst case stack
+	// size.
+	for i, name := range stackSizeLoads {
+		fn, ok := stackSizes[name]
+		if !ok {
+			return fmt.Errorf("could not find symbol %s in ELF file", name)
+		}
+		if fn.stackSizeType == stacksize.Bounded {
+			stackSize := uint32(fn.stackSize)
+
+			// Adding 4 for the stack canary. Even though the size may be
+			// automatically determined, stack overflow checking is still
+			// important as the stack size cannot be determined for all
+			// goroutines.
+			stackSize += 4
+
+			// Add stack size used by interrupts.
+			switch elfFile.Machine {
+			case elf.EM_ARM:
+				// On Cortex-M (assumed here), this stack size is 8 words or 32
+				// bytes. This is only to store the registers that the interrupt
+				// may modify, the interrupt will switch to the interrupt stack
+				// (MSP).
+				// Some background:
+				// https://interrupt.memfault.com/blog/cortex-m-rtos-context-switching
+				stackSize += 32
+			}
+
+			// Finally write the stack size to the binary.
+			binary.LittleEndian.PutUint32(data[i*4:], stackSize)
+		}
+	}
+
+	// Write back the modified stack sizes.
+	_, err = fp.WriteAt(data, int64(section.Offset))
+	if err != nil {
+		return err
+	}
+
+	return nil
+}
+
+// printStacks prints the maximum stack depth for functions that are started as
+// goroutines. Stack sizes cannot always be determined statically, in particular
+// recursive functions and functions that call interface methods or function
+// pointers may have an unknown stack depth (depending on what the optimizer
+// manages to optimize away).
+//
+// It might print something like the following:
+//
+//     function                         stack usage (in bytes)
+//     Reset_Handler                    316
+//     examples/blinky2.led1            92
+//     runtime.run$1                    300
+func printStacks(calculatedStacks []string, stackSizes map[string]functionStackSize) {
+	// Print the sizes of all stacks.
+	fmt.Printf("%-32s %s\n", "function", "stack usage (in bytes)")
+	for _, name := range calculatedStacks {
+		fn := stackSizes[name]
+		switch fn.stackSizeType {
+		case stacksize.Bounded:
+			fmt.Printf("%-32s %d\n", fn.humanName, fn.stackSize)
+		case stacksize.Unknown:
+			fmt.Printf("%-32s unknown, %s does not have stack frame information\n", fn.humanName, fn.missingStackSize)
+		case stacksize.Recursive:
+			fmt.Printf("%-32s recursive, %s may call itself\n", fn.humanName, fn.missingStackSize)
+		case stacksize.IndirectCall:
+			fmt.Printf("%-32s unknown, %s calls a function pointer\n", fn.humanName, fn.missingStackSize)
 		}
-		return action(tmppath)
 	}
 }
@@ -30,10 +30,7 @@ func cacheTimestamp(paths []string) (time.Time, error) {
 // Try to load a given file from the cache. Return "", nil if no cached file can
 // be found (or the file is stale), return the absolute path if there is a cache
 // and return an error on I/O errors.
-//
-// TODO: the configKey is currently ignored. It is supposed to be used as extra
-// data for the cache key, like the compiler version and arguments.
-func cacheLoad(name, configKey string, sourceFiles []string) (string, error) {
+func cacheLoad(name string, sourceFiles []string) (string, error) {
 	cachepath := filepath.Join(goenv.Get("GOCACHE"), name)
 	cacheStat, err := os.Stat(cachepath)
 	if os.IsNotExist(err) {
@@ -58,9 +55,7 @@ func cacheLoad(name, configKey string, sourceFiles []string) (string, error) {

 // Store the file located at tmppath in the cache with the given name. The
 // tmppath may or may not be gone afterwards.
-//
-// Note: the configKey is ignored, see cacheLoad.
-func cacheStore(tmppath, name, configKey string, sourceFiles []string) (string, error) {
+func cacheStore(tmppath, name string, sourceFiles []string) (string, error) {
 	// get the last modified time
 	if len(sourceFiles) == 0 {
 		panic("cache: no source files")
@@ -74,24 +69,16 @@ func cacheStore(tmppath, name, configKey string, sourceFiles []string) (string,
 		return "", err
 	}
 	cachepath := filepath.Join(dir, name)
-	err = moveFile(tmppath, cachepath)
+	err = copyFile(tmppath, cachepath)
 	if err != nil {
 		return "", err
 	}
 	return cachepath, nil
 }

-// moveFile renames the file from src to dst. If renaming doesn't work (for
-// example, the rename crosses a filesystem boundary), the file is copied and
-// the old file is removed.
-func moveFile(src, dst string) error {
-	err := os.Rename(src, dst)
-	if err == nil {
-		// Success!
-		return nil
-	}
-	// Failed to move, probably a different filesystem.
-	// Do a copy + remove.
+// copyFile copies the given file from src to dst. It can copy over
+// a possibly already existing file at the destination.
+func copyFile(src, dst string) error {
 	inf, err := os.Open(src)
 	if err != nil {
 		return err
@@ -1,12 +1,7 @@
 package builder

 import (
-	"io/ioutil"
-	"os"
-	"path/filepath"
 	"strings"
-
-	"github.com/tinygo-org/tinygo/goenv"
 )

 // These are the GENERIC_SOURCES according to CMakeList.txt.
@@ -156,105 +151,20 @@ var aeabiBuiltins = []string{
 	"arm/aeabi_uldivmod.S",
 }

-func builtinFiles(target string) []string {
-	builtins := append([]string{}, genericBuiltins...) // copy genericBuiltins
-	if strings.HasPrefix(target, "arm") || strings.HasPrefix(target, "thumb") {
-		builtins = append(builtins, aeabiBuiltins...)
-	}
-	return builtins
-}
-
-// builtinsDir returns the directory where the sources for compiler-rt are kept.
-func builtinsDir() string {
-	return filepath.Join(goenv.Get("TINYGOROOT"), "lib", "compiler-rt", "lib", "builtins")
-}
-
-// Get the builtins archive, possibly generating it as needed.
-func loadBuiltins(target string) (path string, err error) {
-	// Try to load a precompiled compiler-rt library.
-	precompiledPath := filepath.Join(goenv.Get("TINYGOROOT"), "pkg", target, "compiler-rt.a")
-	if _, err := os.Stat(precompiledPath); err == nil {
-		// Found a precompiled compiler-rt for this OS/architecture. Return the
-		// path directly.
-		return precompiledPath, nil
-	}
-
-	outfile := "librt-" + target + ".a"
-	builtinsDir := builtinsDir()
-
-	builtins := builtinFiles(target)
-	srcs := make([]string, len(builtins))
-	for i, name := range builtins {
-		srcs[i] = filepath.Join(builtinsDir, name)
-	}
-
-	if path, err := cacheLoad(outfile, commands["clang"][0], srcs); path != "" || err != nil {
-		return path, err
-	}
-
-	var cachepath string
-	err = CompileBuiltins(target, func(path string) error {
-		path, err := cacheStore(path, outfile, commands["clang"][0], srcs)
-		cachepath = path
-		return err
-	})
-	return cachepath, err
-}
-
-// CompileBuiltins compiles builtins from compiler-rt into a static library.
-// When it succeeds, it will call the callback with the resulting path. The path
-// will be removed after callback returns. If callback returns an error, this is
-// passed through to the return value of this function.
-func CompileBuiltins(target string, callback func(path string) error) error {
-	builtinsDir := builtinsDir()
-
-	builtins := builtinFiles(target)
-	srcs := make([]string, len(builtins))
-	for i, name := range builtins {
-		srcs[i] = filepath.Join(builtinsDir, name)
-	}
-
-	dirPrefix := "tinygo-builtins"
-	remapDir := filepath.Join(os.TempDir(), dirPrefix)
-	dir, err := ioutil.TempDir(os.TempDir(), dirPrefix)
-	if err != nil {
-		return err
-	}
-	defer os.RemoveAll(dir)
-
-	// Compile all builtins.
-	// TODO: use builtins optimized for a given target if available.
-	objs := make([]string, 0, len(builtins))
-	for _, name := range builtins {
-		objname := name
-		if strings.LastIndexByte(objname, '/') >= 0 {
-			objname = objname[strings.LastIndexByte(objname, '/'):]
+// CompilerRT is a library with symbols required by programs compiled with LLVM.
+// These symbols are for operations that cannot be emitted with a single
+// instruction or a short sequence of instructions for that target.
+//
+// For more information, see: https://compiler-rt.llvm.org/
+var CompilerRT = Library{
+	name:      "compiler-rt",
+	cflags:    func() []string { return []string{"-Werror", "-Wall", "-std=c11", "-nostdlibinc"} },
+	sourceDir: "lib/compiler-rt/lib/builtins",
+	sources: func(target string) []string {
+		builtins := append([]string{}, genericBuiltins...) // copy genericBuiltins
+		if strings.HasPrefix(target, "arm") || strings.HasPrefix(target, "thumb") {
+			builtins = append(builtins, aeabiBuiltins...)
 		}
-		objpath := filepath.Join(dir, objname+".o")
-		objs = append(objs, objpath)
-		srcpath := filepath.Join(builtinsDir, name)
-		// Note: -fdebug-prefix-map is necessary to make the output archive
-		// reproducible. Otherwise the temporary directory is stored in the
-		// archive itself, which varies each run.
-		args := []string{"-c", "-Oz", "-g", "-Werror", "-Wall", "-std=c11", "-fshort-enums", "-nostdlibinc", "-ffunction-sections", "-fdata-sections", "-Wno-macro-redefined", "--target=" + target, "-fdebug-prefix-map=" + dir + "=" + remapDir}
-		if strings.HasPrefix(target, "riscv32-") {
-			args = append(args, "-march=rv32imac", "-mabi=ilp32", "-fforce-enable-int128")
-		}
-		err := runCCompiler("clang", append(args, "-o", objpath, srcpath)...)
-		if err != nil {
-			return &commandError{"failed to build", srcpath, err}
-		}
-	}
-
-	// Put all the object files in a single archive. This archive file will be
-	// used to statically link compiler-rt.
-	arpath := filepath.Join(dir, "librt.a")
-	err = makeArchive(arpath, objs)
-	if err != nil {
-		return err
-	}
-
-	// Give the caller the resulting file. The callback must copy the file,
-	// because after it returns the temporary directory will be removed.
-	return callback(arpath)
+		return builtins
+	},
 }
@@ -48,6 +48,7 @@
 #include "llvm/Support/Host.h"
 #include "llvm/Support/MemoryBuffer.h"
 #include "llvm/Support/Path.h"
+#include "llvm/Support/Process.h"
 #include "llvm/Support/Signals.h"
 #include "llvm/Support/SourceMgr.h"
 #include "llvm/Support/TargetRegistry.h"
@@ -70,12 +71,12 @@ bool AssemblerInvocation::CreateFromArgs(AssemblerInvocation &Opts,
  bool Success = true;

  // Parse the arguments.
-  std::unique_ptr<OptTable> OptTbl(createDriverOptTable());
+  const OptTable &OptTbl = getDriverOptTable();

  const unsigned IncludedFlagsBitmask = options::CC1AsOption;
  unsigned MissingArgIndex, MissingArgCount;
-  InputArgList Args = OptTbl->ParseArgs(Argv, MissingArgIndex, MissingArgCount,
-                                        IncludedFlagsBitmask);
+  InputArgList Args = OptTbl.ParseArgs(Argv, MissingArgIndex, MissingArgCount,
+                                       IncludedFlagsBitmask);

  // Check for missing argument error.
  if (MissingArgCount) {
@@ -88,7 +89,7 @@ bool AssemblerInvocation::CreateFromArgs(AssemblerInvocation &Opts,
  for (const Arg *A : Args.filtered(OPT_UNKNOWN)) {
    auto ArgString = A->getAsString(Args);
    std::string Nearest;
-    if (OptTbl->findNearest(ArgString, Nearest, IncludedFlagsBitmask) > 1)
+    if (OptTbl.findNearest(ArgString, Nearest, IncludedFlagsBitmask) > 1)
      Diags.Report(diag::err_drv_unknown_argument) << ArgString;
    else
      Diags.Report(diag::err_drv_unknown_argument_with_suggestion)
@@ -100,7 +101,7 @@ bool AssemblerInvocation::CreateFromArgs(AssemblerInvocation &Opts,

  // Target Options
  Opts.Triple = llvm::Triple::normalize(Args.getLastArgValue(OPT_triple));
-  Opts.CPU = Args.getLastArgValue(OPT_target_cpu);
+  Opts.CPU = std::string(Args.getLastArgValue(OPT_target_cpu));
  Opts.Features = Args.getAllArgValues(OPT_target_feature);

  // Use the default target triple if unspecified.
@@ -131,13 +132,19 @@ bool AssemblerInvocation::CreateFromArgs(AssemblerInvocation &Opts,

  Opts.RelaxELFRelocations = Args.hasArg(OPT_mrelax_relocations);
  Opts.DwarfVersion = getLastArgIntValue(Args, OPT_dwarf_version_EQ, 2, Diags);
-  Opts.DwarfDebugFlags = Args.getLastArgValue(OPT_dwarf_debug_flags);
-  Opts.DwarfDebugProducer = Args.getLastArgValue(OPT_dwarf_debug_producer);
-  Opts.DebugCompilationDir = Args.getLastArgValue(OPT_fdebug_compilation_dir);
-  Opts.MainFileName = Args.getLastArgValue(OPT_main_file_name);
+  Opts.DwarfDebugFlags =
+      std::string(Args.getLastArgValue(OPT_dwarf_debug_flags));
+  Opts.DwarfDebugProducer =
+      std::string(Args.getLastArgValue(OPT_dwarf_debug_producer));
+  Opts.DebugCompilationDir =
+      std::string(Args.getLastArgValue(OPT_fdebug_compilation_dir));
+  Opts.MainFileName = std::string(Args.getLastArgValue(OPT_main_file_name));

-  for (const auto &Arg : Args.getAllArgValues(OPT_fdebug_prefix_map_EQ))
-    Opts.DebugPrefixMap.insert(StringRef(Arg).split('='));
+  for (const auto &Arg : Args.getAllArgValues(OPT_fdebug_prefix_map_EQ)) {
+    auto Split = StringRef(Arg).split('=');
+    Opts.DebugPrefixMap.insert(
+        {std::string(Split.first), std::string(Split.second)});
+  }

  // Frontend Options
  if (Args.hasArg(OPT_INPUT)) {
@@ -153,8 +160,9 @@ bool AssemblerInvocation::CreateFromArgs(AssemblerInvocation &Opts,
    }
  }
  Opts.LLVMArgs = Args.getAllArgValues(OPT_mllvm);
-  Opts.OutputPath = Args.getLastArgValue(OPT_o);
-  Opts.SplitDwarfOutput = Args.getLastArgValue(OPT_split_dwarf_output);
+  Opts.OutputPath = std::string(Args.getLastArgValue(OPT_o));
+  Opts.SplitDwarfOutput =
+      std::string(Args.getLastArgValue(OPT_split_dwarf_output));
  if (Arg *A = Args.getLastArg(OPT_filetype)) {
    StringRef Name = A->getValue();
    unsigned OutputType = StringSwitch<unsigned>(Name)
@@ -181,8 +189,10 @@ bool AssemblerInvocation::CreateFromArgs(AssemblerInvocation &Opts,
  Opts.RelaxAll = Args.hasArg(OPT_mrelax_all);
  Opts.NoExecStack = Args.hasArg(OPT_mno_exec_stack);
  Opts.FatalWarnings = Args.hasArg(OPT_massembler_fatal_warnings);
-  Opts.RelocationModel = Args.getLastArgValue(OPT_mrelocation_model, "pic");
-  Opts.TargetABI = Args.getLastArgValue(OPT_target_abi);
+  Opts.NoWarn = Args.hasArg(OPT_massembler_no_warn);
+  Opts.RelocationModel =
+      std::string(Args.getLastArgValue(OPT_mrelocation_model, "pic"));
+  Opts.TargetABI = std::string(Args.getLastArgValue(OPT_target_abi));
  Opts.IncrementalLinkerCompatible =
      Args.hasArg(OPT_mincremental_linker_compatible);
  Opts.SymbolDefs = Args.getAllArgValues(OPT_defsym);
@@ -208,8 +218,8 @@ getOutputStream(StringRef Path, DiagnosticsEngine &Diags, bool Binary) {
    sys::RemoveFileOnSignal(Path);

  std::error_code EC;
-  auto Out = llvm::make_unique<raw_fd_ostream>(
-      Path, EC, (Binary ? sys::fs::F_None : sys::fs::F_Text));
+  auto Out = std::make_unique<raw_fd_ostream>(
+      Path, EC, (Binary ? sys::fs::OF_None : sys::fs::OF_Text));
  if (EC) {
    Diags.Report(diag::err_fe_unable_to_open_output) << Path << EC.message();
    return nullptr;
@@ -245,7 +255,9 @@ bool ExecuteAssembler(AssemblerInvocation &Opts, DiagnosticsEngine &Diags) {
  std::unique_ptr<MCRegisterInfo> MRI(TheTarget->createMCRegInfo(Opts.Triple));
  assert(MRI && "Unable to create target register info!");

-  std::unique_ptr<MCAsmInfo> MAI(TheTarget->createMCAsmInfo(*MRI, Opts.Triple));
+  MCTargetOptions MCOptions;
+  std::unique_ptr<MCAsmInfo> MAI(
+      TheTarget->createMCAsmInfo(*MRI, Opts.Triple, MCOptions));
  assert(MAI && "Unable to create target asm info!");

  // Ensure MCAsmInfo initialization occurs before any use, otherwise sections
@@ -269,7 +281,7 @@ bool ExecuteAssembler(AssemblerInvocation &Opts, DiagnosticsEngine &Diags) {
  // MCObjectFileInfo needs a MCContext reference in order to initialize itself.
  std::unique_ptr<MCObjectFileInfo> MOFI(new MCObjectFileInfo());

-  MCContext Ctx(MAI.get(), MRI.get(), MOFI.get(), &SrcMgr);
+  MCContext Ctx(MAI.get(), MRI.get(), MOFI.get(), &SrcMgr, &MCOptions);

  bool PIC = false;
  if (Opts.RelocationModel == "static") {
@@ -310,12 +322,7 @@ bool ExecuteAssembler(AssemblerInvocation &Opts, DiagnosticsEngine &Diags) {
                            SrcMgr.getMemoryBuffer(BufferIndex)->getBuffer());

  // Build up the feature string from the target feature list.
-  std::string FS;
-  if (!Opts.Features.empty()) {
-    FS = Opts.Features[0];
-    for (unsigned i = 1, e = Opts.Features.size(); i != e; ++i)
-      FS += "," + Opts.Features[i];
-  }
+  std::string FS = llvm::join(Opts.Features, ",");

  std::unique_ptr<MCStreamer> Str;

@@ -326,7 +333,8 @@ bool ExecuteAssembler(AssemblerInvocation &Opts, DiagnosticsEngine &Diags) {
  raw_pwrite_stream *Out = FDOS.get();
  std::unique_ptr<buffer_ostream> BOS;

-  MCTargetOptions MCOptions;
+  MCOptions.MCNoWarn = Opts.NoWarn;
+  MCOptions.MCFatalWarnings = Opts.FatalWarnings;
  MCOptions.ABIName = Opts.TargetABI;

  // FIXME: There is a bit of code duplication with addPassesToEmitFile.
@@ -340,7 +348,7 @@ bool ExecuteAssembler(AssemblerInvocation &Opts, DiagnosticsEngine &Diags) {
    std::unique_ptr<MCAsmBackend> MAB(
        TheTarget->createMCAsmBackend(*STI, *MRI, MCOptions));

-    auto FOut = llvm::make_unique<formatted_raw_ostream>(*Out);
+    auto FOut = std::make_unique<formatted_raw_ostream>(*Out);
    Str.reset(TheTarget->createAsmStreamer(
        Ctx, std::move(FOut), /*asmverbose*/ true,
        /*useDwarfDirectory*/ true, IP, std::move(CE), std::move(MAB),
@@ -351,7 +359,7 @@ bool ExecuteAssembler(AssemblerInvocation &Opts, DiagnosticsEngine &Diags) {
    assert(Opts.OutputType == AssemblerInvocation::FT_Obj &&
           "Invalid file type!");
    if (!FDOS->supportsSeeking()) {
-      BOS = make_unique<buffer_ostream>(*FDOS);
+      BOS = std::make_unique<buffer_ostream>(*FDOS);
      Out = BOS.get();
    }

@@ -378,7 +386,7 @@ bool ExecuteAssembler(AssemblerInvocation &Opts, DiagnosticsEngine &Diags) {
    MCSection *AsmLabel = Ctx.getMachOSection(
        "__LLVM", "__asm", MachO::S_REGULAR, 4, SectionKind::getReadOnly());
    Str.get()->SwitchSection(AsmLabel);
-    Str.get()->EmitZeros(1);
+    Str.get()->emitZeros(1);
  }

  // Assembly to object compilation should leverage assembly info.
@@ -436,7 +444,7 @@ static void LLVMErrorHandler(void *UserData, const std::string &Message,
  Diags.Report(diag::err_fe_error_backend) << Message;

  // We cannot recover from llvm errors.
-  exit(1);
+  sys::Process::Exit(1);
 }

 int cc1as_main(ArrayRef<const char *> Argv, const char *Argv0, void *MainAddr) {
@@ -464,11 +472,11 @@ int cc1as_main(ArrayRef<const char *> Argv, const char *Argv0, void *MainAddr) {
    return 1;

  if (Asm.ShowHelp) {
-    std::unique_ptr<OptTable> Opts(driver::createDriverOptTable());
-    Opts->PrintHelp(llvm::outs(), "clang -cc1as [options] file...",
-                    "Clang Integrated Assembler",
-                    /*Include=*/driver::options::CC1AsOption, /*Exclude=*/0,
-                    /*ShowAllAliases=*/false);
+    getDriverOptTable().PrintHelp(
+        llvm::outs(), "clang -cc1as [options] file...",
+        "Clang Integrated Assembler",
+        /*Include=*/driver::options::CC1AsOption, /*Exclude=*/0,
+        /*ShowAllAliases=*/false);
    return 0;
  }

@@ -485,7 +493,7 @@ int cc1as_main(ArrayRef<const char *> Argv, const char *Argv0, void *MainAddr) {
  // FIXME: Remove this, one day.
  if (!Asm.LLVMArgs.empty()) {
    unsigned NumArgs = Asm.LLVMArgs.size();
-    auto Args = llvm::make_unique<const char*[]>(NumArgs + 2);
+    auto Args = std::make_unique<const char*[]>(NumArgs + 2);
    Args[0] = "clang (LLVM option parsing)";
    for (unsigned i = 0; i != NumArgs; ++i)
      Args[i + 1] = Asm.LLVMArgs[i].c_str();
@@ -499,6 +507,7 @@ int cc1as_main(ArrayRef<const char *> Argv, const char *Argv0, void *MainAddr) {
  // If any timers were active but haven't been destroyed yet, print their
  // results now.
  TimerGroup::printAll(errs());
+  TimerGroup::clearAll();

  return !!Failed;
 }
@@ -11,8 +11,6 @@
 //
 //===----------------------------------------------------------------------===//

-#include "llvm/ADT/ArrayRef.h"
-
 /// Helper class for representing a single invocation of the assembler.
 struct AssemblerInvocation {
  /// @name Target Options
@@ -82,6 +80,7 @@ struct AssemblerInvocation {
  unsigned RelaxAll : 1;
  unsigned NoExecStack : 1;
  unsigned FatalWarnings : 1;
+  unsigned NoWarn : 1;
  unsigned IncrementalLinkerCompatible : 1;
  unsigned EmbedBitcode : 1;

@@ -107,6 +106,7 @@ public:
    RelaxAll = 0;
    NoExecStack = 0;
    FatalWarnings = 0;
+    NoWarn = 0;
    IncrementalLinkerCompatible = 0;
    DwarfVersion = 0;
    EmbedBitcode = 0;
@@ -11,6 +11,7 @@
 #include <clang/FrontendTool/Utils.h>
 #include <llvm/ADT/IntrusiveRefCntPtr.h>
 #include <llvm/Option/Option.h>
+#include <llvm/Support/Host.h>

 using namespace llvm;
 using namespace clang;
@@ -52,8 +53,7 @@ bool tinygo_clang_driver(int argc, char **argv) {
 			std::unique_ptr<clang::CompilerInstance> Clang(new clang::CompilerInstance());
 			bool success = clang::CompilerInvocation::CreateFromArgs(
 					Clang->getInvocation(),
-					const_cast<const char **>(CCArgs.data()),
-					const_cast<const char **>(CCArgs.data()) + CCArgs.size(),
+					CCArgs,
 					Diags);
 			if (!success) {
 				return false;
@@ -74,7 +74,8 @@ bool tinygo_clang_driver(int argc, char **argv) {
 		} else if (strcmp(*CCArgs.data(), "-cc1as") == 0) {
 			// This is the assembler frontend. Parse the arguments.
 			AssemblerInvocation Asm;
-			if (!AssemblerInvocation::CreateFromArgs(Asm, llvm::ArrayRef<const char*>(CCArgs).slice(1), Diags))
+			ArrayRef<const char *> Argv = llvm::ArrayRef<const char*>(CCArgs);
+			if (!AssemblerInvocation::CreateFromArgs(Asm, Argv.slice(1), Diags))
 				return false;

 			// Execute the invocation, unless there were parsing errors.
@@ -6,25 +6,28 @@ import (
 	"os/exec"
 	"runtime"
 	"strings"
+
+	"tinygo.org/x/go-llvm"
 )

 // Commands lists command alternatives for various operating systems. These
 // commands may have a slightly different name across operating systems and
 // distributions or may not even exist in $PATH, in which case absolute paths
 // may be used.
-var commands = map[string][]string{
-	"clang":   {"clang-9"},
-	"ld.lld":  {"ld.lld-9", "ld.lld"},
-	"wasm-ld": {"wasm-ld-9", "wasm-ld"},
-}
+var commands = map[string][]string{}

 func init() {
-	// Add the path to a Homebrew-installed LLVM 9 for ease of use (no need to
+	llvmMajor := strings.Split(llvm.Version, ".")[0]
+	commands["clang"] = []string{"clang-" + llvmMajor}
+	commands["ld.lld"] = []string{"ld.lld-" + llvmMajor, "ld.lld"}
+	commands["wasm-ld"] = []string{"wasm-ld-" + llvmMajor, "wasm-ld"}
+	// Add the path to a Homebrew-installed LLVM for ease of use (no need to
 	// manually set $PATH).
 	if runtime.GOOS == "darwin" {
-		commands["clang"] = append(commands["clang"], "/usr/local/opt/llvm@9/bin/clang-9")
-		commands["ld.lld"] = append(commands["ld.lld"], "/usr/local/opt/llvm@9/bin/ld.lld")
-		commands["wasm-ld"] = append(commands["wasm-ld"], "/usr/local/opt/llvm@9/bin/wasm-ld")
+		prefix := "/usr/local/opt/llvm@" + llvmMajor + "/bin/"
+		commands["clang"] = append(commands["clang"], prefix+"clang-"+llvmMajor)
+		commands["ld.lld"] = append(commands["ld.lld"], prefix+"ld.lld")
+		commands["wasm-ld"] = append(commands["wasm-ld"], prefix+"wasm-ld")
 	}
 	// Add the path for when LLVM was installed with the installer from
 	// llvm.org, which by default doesn't add LLVM to the $PATH environment
@@ -34,11 +37,12 @@ func init() {
 		commands["ld.lld"] = append(commands["ld.lld"], "lld", "C:\\Program Files\\LLVM\\bin\\lld.exe")
 		commands["wasm-ld"] = append(commands["wasm-ld"], "C:\\Program Files\\LLVM\\bin\\wasm-ld.exe")
 	}
-	// Add the path to the llvm90 installed from ports
+	// Add the path to LLVM installed from ports.
 	if runtime.GOOS == "freebsd" {
-		commands["clang"] = append(commands["clang"], "/usr/local/llvm90/bin/clang-9")
-		commands["ld.lld"] = append(commands["ld.lld"], "/usr/local/llvm90/bin/ld.lld")
-		commands["wasm-ld"] = append(commands["wasm-ld"], "/usr/local/llvm90/bin/wasm-ld")
+		prefix := "/usr/local/llvm" + llvmMajor + "/bin/"
+		commands["clang"] = append(commands["clang"], prefix+"clang-"+llvmMajor)
+		commands["ld.lld"] = append(commands["ld.lld"], prefix+"ld.lld")
+		commands["wasm-ld"] = append(commands["wasm-ld"], prefix+"wasm-ld")
 	}
 }

@@ -1,60 +0,0 @@
-// +build byollvm
-
-package builder
-
-import (
-	"errors"
-	"os"
-	"os/exec"
-	"unsafe"
-
-	"github.com/tinygo-org/tinygo/goenv"
-)
-
-/*
-#cgo CXXFLAGS: -fno-rtti
-#include <stdbool.h>
-#include <stdlib.h>
-bool tinygo_clang_driver(int argc, char **argv);
-*/
-import "C"
-
-// runCCompiler invokes a C compiler with the given arguments.
-//
-// This version invokes the built-in Clang when trying to run the Clang compiler.
-func runCCompiler(command string, flags ...string) error {
-	switch command {
-	case "clang":
-		// Compile this with the internal Clang compiler.
-		headerPath := getClangHeaderPath(goenv.Get("TINYGOROOT"))
-		if headerPath == "" {
-			return errors.New("could not locate Clang headers")
-		}
-		flags = append(flags, "-I"+headerPath)
-		flags = append([]string{"tinygo:" + command}, flags...)
-		var cflag *C.char
-		buf := C.calloc(C.size_t(len(flags)), C.size_t(unsafe.Sizeof(cflag)))
-		cflags := (*[1 << 10]*C.char)(unsafe.Pointer(buf))[:len(flags):len(flags)]
-		for i, flag := range flags {
-			cflag := C.CString(flag)
-			cflags[i] = cflag
-			defer C.free(unsafe.Pointer(cflag))
-		}
-		ok := C.tinygo_clang_driver(C.int(len(flags)), (**C.char)(buf))
-		if !ok {
-			return errors.New("failed to compile using built-in clang")
-		}
-		return nil
-	default:
-		// Running some other compiler. Maybe it has been defined in the
-		// commands map (unlikely).
-		if cmdNames, ok := commands[command]; ok {
-			return execCommand(cmdNames, flags...)
-		}
-		// Alternatively, run the compiler directly.
-		cmd := exec.Command(command, flags...)
-		cmd.Stdout = os.Stdout
-		cmd.Stderr = os.Stderr
-		return cmd.Run()
-	}
-}
@@ -1,24 +0,0 @@
-// +build !byollvm
-
-package builder
-
-// This file provides a way to a C compiler as an external command. See also:
-// clang-external.go
-
-import (
-	"os"
-	"os/exec"
-)
-
-// runCCompiler invokes a C compiler with the given arguments.
-//
-// This version always runs the compiler as an external command.
-func runCCompiler(command string, flags ...string) error {
-	if cmdNames, ok := commands[command]; ok {
-		return execCommand(cmdNames, flags...)
-	}
-	cmd := exec.Command(command, flags...)
-	cmd.Stdout = os.Stdout
-	cmd.Stderr = os.Stderr
-	return cmd.Run()
-}
@@ -21,12 +21,12 @@ func NewConfig(options *compileopts.Options) (*compileopts.Config, error) {
 	if goroot == "" {
 		return nil, errors.New("cannot locate $GOROOT, please set it manually")
 	}
-	major, minor, err := getGorootVersion(goroot)
+	major, minor, err := goenv.GetGorootVersion(goroot)
 	if err != nil {
 		return nil, fmt.Errorf("could not read version from GOROOT (%v): %v", goroot, err)
 	}
-	if major != 1 || (minor != 11 && minor != 12 && minor != 13) {
-		return nil, fmt.Errorf("requires go version 1.11, 1.12, or 1.13, got go%d.%d", major, minor)
+	if major != 1 || minor < 11 || minor > 16 {
+		return nil, fmt.Errorf("requires go version 1.11 through 1.16, got go%d.%d", major, minor)
 	}
 	clangHeaderPath := getClangHeaderPath(goenv.Get("TINYGOROOT"))
 	return &compileopts.Config{
@@ -1,71 +1,16 @@
 package builder

 import (
-	"errors"
-	"fmt"
-	"io"
 	"io/ioutil"
 	"os"
 	"os/exec"
 	"path/filepath"
-	"regexp"
 	"sort"
 	"strings"
+
+	"tinygo.org/x/go-llvm"
 )

-// getGorootVersion returns the major and minor version for a given GOROOT path.
-// If the goroot cannot be determined, (0, 0) is returned.
-func getGorootVersion(goroot string) (major, minor int, err error) {
-	s, err := GorootVersionString(goroot)
-	if err != nil {
-		return 0, 0, err
-	}
-
-	if s == "" || s[:2] != "go" {
-		return 0, 0, errors.New("could not parse Go version: version does not start with 'go' prefix")
-	}
-
-	parts := strings.Split(s[2:], ".")
-	if len(parts) < 2 {
-		return 0, 0, errors.New("could not parse Go version: version has less than two parts")
-	}
-
-	// Ignore the errors, we don't really handle errors here anyway.
-	var trailing string
-	n, err := fmt.Sscanf(s, "go%d.%d%s", &major, &minor, &trailing)
-	if n == 2 && err == io.EOF {
-		// Means there were no trailing characters (i.e., not an alpha/beta)
-		err = nil
-	}
-	if err != nil {
-		return 0, 0, fmt.Errorf("failed to parse version: %s", err)
-	}
-	return
-}
-
-// GorootVersionString returns the version string as reported by the Go
-// toolchain for the given GOROOT path. It is usually of the form `go1.x.y` but
-// can have some variations (for beta releases, for example).
-func GorootVersionString(goroot string) (string, error) {
-	if data, err := ioutil.ReadFile(filepath.Join(
-		goroot, "src", "runtime", "internal", "sys", "zversion.go")); err == nil {
-
-		r := regexp.MustCompile("const TheVersion = `(.*)`")
-		matches := r.FindSubmatch(data)
-		if len(matches) != 2 {
-			return "", errors.New("Invalid go version output:\n" + string(data))
-		}
-
-		return string(matches[1]), nil
-
-	} else if data, err := ioutil.ReadFile(filepath.Join(goroot, "VERSION")); err == nil {
-		return string(data), nil
-
-	} else {
-		return "", err
-	}
-}
-
 // getClangHeaderPath returns the path to the built-in Clang headers. It tries
 // multiple locations, which should make it find the directory when installed in
 // various ways.
@@ -84,6 +29,7 @@ func getClangHeaderPath(TINYGOROOT string) string {

 	// It looks like we are built with a system-installed LLVM. Do a last
 	// attempt: try to use Clang headers relative to the clang binary.
+	llvmMajor := strings.Split(llvm.Version, ".")[0]
 	for _, cmdName := range commands["clang"] {
 		binpath, err := exec.LookPath(cmdName)
 		if err == nil {
@@ -98,22 +44,38 @@ func getClangHeaderPath(TINYGOROOT string) string {
 			// Example executable:
 			//     /usr/lib/llvm-9/bin/clang
 			// Example include path:
-			//     /usr/lib/llvm-9/lib/clang/9.0.1/include/
+			//     /usr/lib/llvm-9/lib64/clang/9.0.1/include/
 			llvmRoot := filepath.Dir(filepath.Dir(binpath))
-			clangVersionRoot := filepath.Join(llvmRoot, "lib", "clang")
-			dirs, err := ioutil.ReadDir(clangVersionRoot)
-			if err != nil {
+			clangVersionRoot := filepath.Join(llvmRoot, "lib64", "clang")
+			dirs64, err64 := ioutil.ReadDir(clangVersionRoot)
+			// Example include path:
+			//     /usr/lib/llvm-9/lib/clang/9.0.1/include/
+			clangVersionRoot = filepath.Join(llvmRoot, "lib", "clang")
+			dirs32, err32 := ioutil.ReadDir(clangVersionRoot)
+			if err64 != nil && err32 != nil {
 				// Unexpected.
 				continue
 			}
-			dirnames := make([]string, len(dirs))
-			for i, d := range dirs {
-				dirnames[i] = d.Name()
+			dirnames := make([]string, len(dirs64)+len(dirs32))
+			dirCount := 0
+			for _, d := range dirs32 {
+				name := d.Name()
+				if name == llvmMajor || strings.HasPrefix(name, llvmMajor+".") {
+					dirnames[dirCount] = filepath.Join(llvmRoot, "lib", "clang", name)
+					dirCount++
+				}
+			}
+			for _, d := range dirs64 {
+				name := d.Name()
+				if name == llvmMajor || strings.HasPrefix(name, llvmMajor+".") {
+					dirnames[dirCount] = filepath.Join(llvmRoot, "lib64", "clang", name)
+					dirCount++
+				}
 			}
 			sort.Strings(dirnames)
 			// Check for the highest version first.
-			for i := len(dirnames) - 1; i >= 0; i-- {
-				path := filepath.Join(clangVersionRoot, dirnames[i], "include")
+			for i := dirCount - 1; i >= 0; i-- {
+				path := filepath.Join(dirnames[i], "include")
 				_, err := os.Stat(filepath.Join(path, "stdint.h"))
 				if err == nil {
 					return path
@@ -0,0 +1,153 @@
+package builder
+
+// This file implements support for writing ESP image files. These image files
+// are read by the ROM bootloader so have to be in a particular format.
+//
+// In the future, it may be necessary to implement support for other image
+// formats, such as the ESP8266 image formats (again, used by the ROM bootloader
+// to load the firmware).
+
+import (
+	"bytes"
+	"crypto/sha256"
+	"debug/elf"
+	"encoding/binary"
+	"fmt"
+	"io/ioutil"
+	"sort"
+)
+
+type espImageSegment struct {
+	addr uint32
+	data []byte
+}
+
+// makeESPFirmare converts an input ELF file to an image file for an ESP32 or
+// ESP8266 chip. This is a special purpose image format just for the ESP chip
+// family, and is parsed by the on-chip mask ROM bootloader.
+//
+// The following documentation has been used:
+// https://github.com/espressif/esptool/wiki/Firmware-Image-Format
+// https://github.com/espressif/esp-idf/blob/8fbb63c2a701c22ccf4ce249f43aded73e134a34/components/bootloader_support/include/esp_image_format.h#L58
+// https://github.com/espressif/esptool/blob/master/esptool.py
+func makeESPFirmareImage(infile, outfile, format string) error {
+	inf, err := elf.Open(infile)
+	if err != nil {
+		return err
+	}
+	defer inf.Close()
+
+	// Load all segments to be written to the image. These are actually ELF
+	// sections, not true ELF segments (similar to how esptool does it).
+	var segments []*espImageSegment
+	for _, section := range inf.Sections {
+		if section.Type != elf.SHT_PROGBITS || section.Size == 0 || section.Flags&elf.SHF_ALLOC == 0 {
+			continue
+		}
+		data, err := section.Data()
+		if err != nil {
+			return fmt.Errorf("failed to read section data: %w", err)
+		}
+		for len(data)%4 != 0 {
+			// Align segment to 4 bytes.
+			data = append(data, 0)
+		}
+		if uint64(uint32(section.Addr)) != section.Addr {
+			return fmt.Errorf("section address too big: 0x%x", section.Addr)
+		}
+		segments = append(segments, &espImageSegment{
+			addr: uint32(section.Addr),
+			data: data,
+		})
+	}
+
+	// Sort the segments by address. This is what esptool does too.
+	sort.SliceStable(segments, func(i, j int) bool { return segments[i].addr < segments[j].addr })
+
+	// Calculate checksum over the segment data. This is used in the image
+	// footer.
+	checksum := uint8(0xef)
+	for _, segment := range segments {
+		for _, b := range segment.data {
+			checksum ^= b
+		}
+	}
+
+	// Write first to an in-memory buffer, primarily so that we can easily
+	// calculate a hash over the entire image.
+	// An added benefit is that we don't need to check for errors all the time.
+	outf := &bytes.Buffer{}
+
+	// Image header.
+	switch format {
+	case "esp32":
+		// Header format:
+		// https://github.com/espressif/esp-idf/blob/8fbb63c2/components/bootloader_support/include/esp_image_format.h#L58
+		binary.Write(outf, binary.LittleEndian, struct {
+			magic          uint8
+			segment_count  uint8
+			spi_mode       uint8
+			spi_speed_size uint8
+			entry_addr     uint32
+			wp_pin         uint8
+			spi_pin_drv    [3]uint8
+			reserved       [11]uint8
+			hash_appended  bool
+		}{
+			magic:          0xE9,
+			segment_count:  byte(len(segments)),
+			spi_mode:       0, // irrelevant, replaced by esptool when flashing
+			spi_speed_size: 0, // spi_speed, spi_size: replaced by esptool when flashing
+			entry_addr:     uint32(inf.Entry),
+			wp_pin:         0xEE, // disable WP pin
+			hash_appended:  true, // add a SHA256 hash
+		})
+	case "esp8266":
+		// Header format:
+		// https://github.com/espressif/esptool/wiki/Firmware-Image-Format
+		// Basically a truncated version of the ESP32 header.
+		binary.Write(outf, binary.LittleEndian, struct {
+			magic          uint8
+			segment_count  uint8
+			spi_mode       uint8
+			spi_speed_size uint8
+			entry_addr     uint32
+		}{
+			magic:          0xE9,
+			segment_count:  byte(len(segments)),
+			spi_mode:       0,    // irrelevant, replaced by esptool when flashing
+			spi_speed_size: 0x20, // spi_speed, spi_size: replaced by esptool when flashing
+			entry_addr:     uint32(inf.Entry),
+		})
+	default:
+		return fmt.Errorf("builder: unknown binary format %#v, expected esp32 or esp8266", format)
+	}
+
+	// Write all segments to the image.
+	// https://github.com/espressif/esptool/wiki/Firmware-Image-Format#segment
+	for _, segment := range segments {
+		binary.Write(outf, binary.LittleEndian, struct {
+			addr   uint32
+			length uint32
+		}{
+			addr:   segment.addr,
+			length: uint32(len(segment.data)),
+		})
+		outf.Write(segment.data)
+	}
+
+	// Footer, including checksum.
+	// The entire image size must be a multiple of 16, so pad the image to one
+	// byte less than that before writing the checksum.
+	outf.Write(make([]byte, 15-outf.Len()%16))
+	outf.WriteByte(checksum)
+
+	if format == "esp32" {
+		// SHA256 hash (to protect against image corruption, not for security).
+		hash := sha256.Sum256(outf.Bytes())
+		outf.Write(hash[:])
+	}
+
+	// Write the image to the output file.
+	return ioutil.WriteFile(outfile, outf.Bytes(), 0666)
+}
@@ -0,0 +1,160 @@
+package builder
+
+// This file implements a job runner for the compiler, which runs jobs in
+// parallel while taking care of dependencies.
+
+import (
+	"fmt"
+	"runtime"
+	"time"
+)
+
+// Set to true to enable logging in the job runner. This may help to debug
+// concurrency or performance issues.
+const jobRunnerDebug = false
+
+type jobState uint8
+
+const (
+	jobStateQueued   jobState = iota // not yet running
+	jobStateRunning                  // running
+	jobStateFinished                 // finished running
+)
+
+// compileJob is a single compiler job, comparable to a single Makefile target.
+// It is used to orchestrate various compiler tasks that can be run in parallel
+// but that have dependencies and thus have limitations in how they can be run.
+type compileJob struct {
+	description  string // description, only used for logging
+	dependencies []*compileJob
+	run          func() error
+	state        jobState
+	err          error         // error if finished
+	duration     time.Duration // how long it took to run this job (only set after finishing)
+}
+
+// readyToRun returns whether this job is ready to run: it is itself not yet
+// started and all dependencies are finished.
+func (job *compileJob) readyToRun() bool {
+	if job.state != jobStateQueued {
+		// Already running or finished, so shouldn't be run again.
+		return false
+	}
+
+	// Check dependencies.
+	for _, dep := range job.dependencies {
+		if dep.state != jobStateFinished {
+			// A dependency is not finished, so this job has to wait until it
+			// is.
+			return false
+		}
+	}
+
+	// All conditions are satisfied.
+	return true
+}
+
+// runJobs runs all the jobs indicated in the jobs slice and returns the error
+// of the first job that fails to run.
+// It runs all jobs in the order of the slice, as long as all dependencies have
+// already run. Therefore, if some jobs are preferred to run before others, they
+// should be ordered as such in this slice.
+func runJobs(jobs []*compileJob) error {
+	// Create channels to communicate with the workers.
+	doneChan := make(chan *compileJob)
+	workerChan := make(chan *compileJob)
+	defer close(workerChan)
+
+	// Start a number of workers.
+	for i := 0; i < runtime.NumCPU(); i++ {
+		if jobRunnerDebug {
+			fmt.Println("## starting worker", i)
+		}
+		go jobWorker(workerChan, doneChan)
+	}
+
+	// Send each job in the jobs slice to a worker, taking care of job
+	// dependencies.
+	numRunningJobs := 0
+	var totalTime time.Duration
+	start := time.Now()
+	for {
+		// If there are free workers, try starting a new job (if one is
+		// available). If it succeeds, try again to fill the entire worker pool.
+		if numRunningJobs < runtime.NumCPU() {
+			jobToRun := nextJob(jobs)
+			if jobToRun != nil {
+				// Start job.
+				if jobRunnerDebug {
+					fmt.Println("## start:   ", jobToRun.description)
+				}
+				jobToRun.state = jobStateRunning
+				workerChan <- jobToRun
+				numRunningJobs++
+				continue
+			}
+		}
+
+		// When there are no jobs running, all jobs in the jobs slice must have
+		// been finished. Therefore, the work is done.
+		if numRunningJobs == 0 {
+			break
+		}
+
+		// Wait until a job is finished.
+		job := <-doneChan
+		job.state = jobStateFinished
+		numRunningJobs--
+		totalTime += job.duration
+		if jobRunnerDebug {
+			fmt.Println("## finished:", job.description, "(time "+job.duration.String()+")")
+		}
+		if job.err != nil {
+			// Wait for running jobs to finish.
+			for numRunningJobs != 0 {
+				<-doneChan
+				numRunningJobs--
+			}
+			// Return error of first failing job.
+			return job.err
+		}
+	}
+
+	// Some statistics, if debugging.
+	if jobRunnerDebug {
+		// Total duration of running all jobs.
+		duration := time.Since(start)
+		fmt.Println("## total:   ", duration)
+
+		// The individual time of each job combined. On a multicore system, this
+		// should be lower than the total above.
+		fmt.Println("## job sum: ", totalTime)
+	}
+
+	return nil
+}
+
+// nextJob returns the first ready-to-run job.
+// This is an implementation detail of runJobs.
+func nextJob(jobs []*compileJob) *compileJob {
+	for _, job := range jobs {
+		if job.readyToRun() {
+			return job
+		}
+	}
+	return nil
+}
+
+// jobWorker is the goroutine that runs received jobs.
+// This is an implementation detail of runJobs.
+func jobWorker(workerChan, doneChan chan *compileJob) {
+	for job := range workerChan {
+		start := time.Now()
+		err := job.run()
+		if err != nil {
+			job.err = err
+		}
+		job.duration = time.Since(start)
+		doneChan <- job
+	}
+}
@@ -0,0 +1,150 @@
+package builder
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+
+	"github.com/tinygo-org/tinygo/goenv"
+)
+
+// Library is a container for information about a single C library, such as a
+// compiler runtime or libc.
+type Library struct {
+	// The library name, such as compiler-rt or picolibc.
+	name string
+
+	cflags func() []string
+
+	// The source directory, relative to TINYGOROOT.
+	sourceDir string
+
+	// The source files, relative to sourceDir.
+	sources func(target string) []string
+}
+
+// fullPath returns the full path to the source directory.
+func (l *Library) fullPath() string {
+	return filepath.Join(goenv.Get("TINYGOROOT"), l.sourceDir)
+}
+
+// sourcePaths returns a slice with the full paths to the source files.
+func (l *Library) sourcePaths(target string) []string {
+	sources := l.sources(target)
+	paths := make([]string, len(sources))
+	for i, name := range sources {
+		paths[i] = filepath.Join(l.fullPath(), name)
+	}
+	return paths
+}
+
+// Load the library archive, possibly generating and caching it if needed.
+// The resulting file is stored in the provided tmpdir, which is expected to be
+// removed after the Load call.
+func (l *Library) Load(target, tmpdir string) (path string, err error) {
+	path, job, err := l.load(target, "", tmpdir)
+	if err != nil {
+		return "", err
+	}
+	if job != nil {
+		jobs := append([]*compileJob{job}, job.dependencies...)
+		err = runJobs(jobs)
+	}
+	return path, err
+}
+
+// load returns a path to the library file for the given target, loading it from
+// cache if possible. It will return a non-zero compiler job if the library
+// wasn't cached, this job (and its dependencies) must be run before the library
+// path is valid.
+// The provided tmpdir will be used to store intermediary files and possibly the
+// output archive file, it is expected to be removed after use.
+func (l *Library) load(target, cpu, tmpdir string) (path string, job *compileJob, err error) {
+	// Try to load a precompiled library.
+	precompiledPath := filepath.Join(goenv.Get("TINYGOROOT"), "pkg", target, l.name+".a")
+	if _, err := os.Stat(precompiledPath); err == nil {
+		// Found a precompiled library for this OS/architecture. Return the path
+		// directly.
+		return precompiledPath, nil, nil
+	}
+
+	var outfile string
+	if cpu != "" {
+		outfile = l.name + "-" + target + "-" + cpu + ".a"
+	} else {
+		outfile = l.name + "-" + target + ".a"
+	}
+
+	// Try to fetch this library from the cache.
+	if path, err := cacheLoad(outfile, l.sourcePaths(target)); path != "" || err != nil {
+		// Cache hit.
+		return path, nil, err
+	}
+	// Cache miss, build it now.
+
+	remapDir := filepath.Join(os.TempDir(), "tinygo-"+l.name)
+	dir := filepath.Join(tmpdir, "build-lib-"+l.name)
+	err = os.Mkdir(dir, 0777)
+	if err != nil {
+		return "", nil, err
+	}
+
+	// Precalculate the flags to the compiler invocation.
+	// Note: -fdebug-prefix-map is necessary to make the output archive
+	// reproducible. Otherwise the temporary directory is stored in the archive
+	// itself, which varies each run.
+	args := append(l.cflags(), "-c", "-Oz", "-g", "-ffunction-sections", "-fdata-sections", "-Wno-macro-redefined", "--target="+target, "-fdebug-prefix-map="+dir+"="+remapDir)
+	if cpu != "" {
+		args = append(args, "-mcpu="+cpu)
+	}
+	if strings.HasPrefix(target, "arm") || strings.HasPrefix(target, "thumb") {
+		args = append(args, "-fshort-enums", "-fomit-frame-pointer", "-mfloat-abi=soft")
+	}
+	if strings.HasPrefix(target, "riscv32-") {
+		args = append(args, "-march=rv32imac", "-mabi=ilp32", "-fforce-enable-int128")
+	}
+	if strings.HasPrefix(target, "riscv64-") {
+		args = append(args, "-march=rv64gc", "-mabi=lp64")
+	}
+
+	// Create job to put all the object files in a single archive. This archive
+	// file is the (static) library file.
+	var objs []string
+	arpath := filepath.Join(dir, l.name+".a")
+	job = &compileJob{
+		description: "ar " + l.name + ".a",
+		run: func() error {
+			// Create an archive of all object files.
+			err := makeArchive(arpath, objs)
+			if err != nil {
+				return err
+			}
+			// Store this archive in the cache.
+			_, err = cacheStore(arpath, outfile, l.sourcePaths(target))
+			return err
+		},
+	}
+
+	// Create jobs to compile all sources. These jobs are depended upon by the
+	// archive job above, so must be run first.
+	for _, srcpath := range l.sourcePaths(target) {
+		srcpath := srcpath // avoid concurrency issues by redefining inside the loop
+		objpath := filepath.Join(dir, filepath.Base(srcpath)+".o")
+		objs = append(objs, objpath)
+		job.dependencies = append(job.dependencies, &compileJob{
+			description: "compile " + srcpath,
+			run: func() error {
+				var compileArgs []string
+				compileArgs = append(compileArgs, args...)
+				compileArgs = append(compileArgs, "-o", objpath, srcpath)
+				err := runCCompiler("clang", compileArgs...)
+				if err != nil {
+					return &commandError{"failed to build", srcpath, err}
+				}
+				return nil
+			},
+		})
+	}
+
+	return arpath, job, nil
+}
@@ -1,71 +0,0 @@
-// +build byollvm
-
-package builder
-
-// This file provides a Link() function that uses the bundled lld if possible.
-
-import (
-	"errors"
-	"os"
-	"os/exec"
-	"unsafe"
-
-	"github.com/tinygo-org/tinygo/goenv"
-)
-
-/*
-#include <stdbool.h>
-#include <stdlib.h>
-bool tinygo_link_elf(int argc, char **argv);
-bool tinygo_link_wasm(int argc, char **argv);
-*/
-import "C"
-
-// link invokes a linker with the given name and flags.
-//
-// This version uses the built-in linker when trying to use lld.
-func link(linker string, flags ...string) error {
-	switch linker {
-	case "ld.lld":
-		flags = append([]string{"tinygo:" + linker}, flags...)
-		var cflag *C.char
-		buf := C.calloc(C.size_t(len(flags)), C.size_t(unsafe.Sizeof(cflag)))
-		cflags := (*[1 << 10]*C.char)(unsafe.Pointer(buf))[:len(flags):len(flags)]
-		for i, flag := range flags {
-			cflag := C.CString(flag)
-			cflags[i] = cflag
-			defer C.free(unsafe.Pointer(cflag))
-		}
-		ok := C.tinygo_link_elf(C.int(len(flags)), (**C.char)(buf))
-		if !ok {
-			return errors.New("failed to link using built-in ld.lld")
-		}
-		return nil
-	case "wasm-ld":
-		flags = append([]string{"tinygo:" + linker}, flags...)
-		var cflag *C.char
-		buf := C.calloc(C.size_t(len(flags)), C.size_t(unsafe.Sizeof(cflag)))
-		defer C.free(buf)
-		cflags := (*[1 << 10]*C.char)(unsafe.Pointer(buf))[:len(flags):len(flags)]
-		for i, flag := range flags {
-			cflag := C.CString(flag)
-			cflags[i] = cflag
-			defer C.free(unsafe.Pointer(cflag))
-		}
-		ok := C.tinygo_link_wasm(C.int(len(flags)), (**C.char)(buf))
-		if !ok {
-			return errors.New("failed to link using built-in wasm-ld")
-		}
-		return nil
-	default:
-		// Fall back to external command.
-		if cmdNames, ok := commands[linker]; ok {
-			return execCommand(cmdNames, flags...)
-		}
-		cmd := exec.Command(linker, flags...)
-		cmd.Stdout = os.Stdout
-		cmd.Stderr = os.Stderr
-		cmd.Dir = goenv.Get("TINYGOROOT")
-		return cmd.Run()
-	}
-}
@@ -1,27 +0,0 @@
-// +build !byollvm
-
-package builder
-
-// This file provides a Link() function that always runs an external command. It
-// is provided for when tinygo is built without linking to liblld.
-
-import (
-	"os"
-	"os/exec"
-
-	"github.com/tinygo-org/tinygo/goenv"
-)
-
-// link invokes a linker with the given name and arguments.
-//
-// This version always runs the linker as an external command.
-func link(linker string, flags ...string) error {
-	if cmdNames, ok := commands[linker]; ok {
-		return execCommand(cmdNames, flags...)
-	}
-	cmd := exec.Command(linker, flags...)
-	cmd.Stdout = os.Stdout
-	cmd.Stderr = os.Stderr
-	cmd.Dir = goenv.Get("TINYGOROOT")
-	return cmd.Run()
-}
@@ -8,12 +8,12 @@ extern "C" {

 bool tinygo_link_elf(int argc, char **argv) {
 	std::vector<const char*> args(argv, argv + argc);
-	return lld::elf::link(args, false);
+	return lld::elf::link(args, false, llvm::outs(), llvm::errs());
 }

 bool tinygo_link_wasm(int argc, char **argv) {
 	std::vector<const char*> args(argv, argv + argc);
-	return lld::wasm::link(args, false);
+	return lld::wasm::link(args, false, llvm::outs(), llvm::errs());
 }

 } // external "C"
@@ -4,12 +4,15 @@ import (
 	"debug/elf"
 	"io/ioutil"
 	"os"
-	"path/filepath"
 	"sort"

 	"github.com/marcinbor85/gohex"
 )

+// maxPadBytes is the maximum allowed bytes to be padded in a rom extraction
+// this value is currently defined by Nintendo Switch Page Alignment (4096 bytes)
+const maxPadBytes = 4095
+
 // objcopyError is an error returned by functions that act like objcopy.
 type objcopyError struct {
 	Op  string
@@ -58,7 +61,7 @@ func extractROM(path string) (uint64, []byte, error) {

 	progs := make(progSlice, 0, 2)
 	for _, prog := range f.Progs {
-		if prog.Type != elf.PT_LOAD || prog.Filesz == 0 {
+		if prog.Type != elf.PT_LOAD || prog.Filesz == 0 || prog.Off == 0 {
 			continue
 		}
 		progs = append(progs, prog)
@@ -70,8 +73,19 @@ func extractROM(path string) (uint64, []byte, error) {

 	var rom []byte
 	for _, prog := range progs {
+		romEnd := progs[0].Paddr + uint64(len(rom))
+		if prog.Paddr > romEnd && prog.Paddr < romEnd+16 {
+			// Sometimes, the linker seems to insert a bit of padding between
+			// segments. Simply zero-fill these parts.
+			rom = append(rom, make([]byte, prog.Paddr-romEnd)...)
+		}
 		if prog.Paddr != progs[0].Paddr+uint64(len(rom)) {
-			return 0, nil, objcopyError{"ROM segments are non-contiguous: " + path, nil}
+			diff := prog.Paddr - (progs[0].Paddr + uint64(len(rom)))
+			if diff > maxPadBytes {
+				return 0, nil, objcopyError{"ROM segments are non-contiguous: " + path, nil}
+			}
+			// Pad the difference
+			rom = append(rom, make([]byte, diff)...)
 		}
 		data, err := ioutil.ReadAll(prog.Open())
 		if err != nil {
@@ -93,7 +107,7 @@ func extractROM(path string) (uint64, []byte, error) {

 // objcopy converts an ELF file to a different (simpler) output file format:
 // .bin or .hex. It extracts only the .text section.
-func objcopy(infile, outfile string) error {
+func objcopy(infile, outfile, binaryFormat string) error {
 	f, err := os.OpenFile(outfile, os.O_RDWR|os.O_CREATE|os.O_TRUNC, 0666)
 	if err != nil {
 		return err
@@ -107,24 +121,20 @@ func objcopy(infile, outfile string) error {
 	}

 	// Write to the file, in the correct format.
-	switch filepath.Ext(outfile) {
-	case ".gba":
-		// The address is not stored in a .gba file.
-		_, err := f.Write(data)
-		return err
-	case ".bin":
-		// The address is not stored in a .bin file (therefore you
-		// should use .hex files in most cases).
-		_, err := f.Write(data)
-		return err
-	case ".hex":
+	switch binaryFormat {
+	case "hex":
+		// Intel hex file, includes the firmware start address.
 		mem := gohex.NewMemory()
 		err := mem.AddBinary(uint32(addr), data)
 		if err != nil {
 			return objcopyError{"failed to create .hex file", err}
 		}
-		mem.DumpIntelHex(f, 16) // TODO: handle error
-		return nil
+		return mem.DumpIntelHex(f, 16)
+	case "bin":
+		// The start address is not stored in raw firmware files (therefore you
+		// should use .hex files in most cases).
+		_, err := f.Write(data)
+		return err
 	default:
 		panic("unreachable")
 	}
@@ -0,0 +1,127 @@
+package builder
+
+import (
+	"path/filepath"
+
+	"github.com/tinygo-org/tinygo/goenv"
+)
+
+// Picolibc is a C library for bare metal embedded devices. It was originally
+// based on newlib.
+var Picolibc = Library{
+	name: "picolibc",
+	cflags: func() []string {
+		picolibcDir := filepath.Join(goenv.Get("TINYGOROOT"), "lib/picolibc/newlib/libc")
+		return []string{"-Werror", "-Wall", "-std=gnu11", "-D_COMPILING_NEWLIB", "-nostdlibinc", "-Xclang", "-internal-isystem", "-Xclang", picolibcDir + "/include", "-I" + picolibcDir + "/tinystdio", "-I" + goenv.Get("TINYGOROOT") + "/lib/picolibc-include"}
+	},
+	sourceDir: "lib/picolibc/newlib/libc",
+	sources: func(target string) []string {
+		return picolibcSources
+	},
+}
+
+var picolibcSources = []string{
+	"string/bcmp.c",
+	"string/bcopy.c",
+	"string/bzero.c",
+	"string/explicit_bzero.c",
+	"string/ffsl.c",
+	"string/ffsll.c",
+	"string/fls.c",
+	"string/flsl.c",
+	"string/flsll.c",
+	"string/gnu_basename.c",
+	"string/index.c",
+	"string/memccpy.c",
+	"string/memchr.c",
+	"string/memcmp.c",
+	"string/memcpy.c",
+	"string/memmem.c",
+	"string/memmove.c",
+	"string/mempcpy.c",
+	"string/memrchr.c",
+	"string/memset.c",
+	"string/rawmemchr.c",
+	"string/rindex.c",
+	"string/stpcpy.c",
+	"string/stpncpy.c",
+	"string/strcasecmp.c",
+	"string/strcasecmp_l.c",
+	"string/strcasestr.c",
+	"string/strcat.c",
+	"string/strchr.c",
+	"string/strchrnul.c",
+	"string/strcmp.c",
+	"string/strcoll.c",
+	"string/strcoll_l.c",
+	"string/strcpy.c",
+	"string/strcspn.c",
+	"string/strdup.c",
+	"string/strerror.c",
+	"string/strerror_r.c",
+	"string/strlcat.c",
+	"string/strlcpy.c",
+	"string/strlen.c",
+	"string/strlwr.c",
+	"string/strncasecmp.c",
+	"string/strncasecmp_l.c",
+	"string/strncat.c",
+	"string/strncmp.c",
+	"string/strncpy.c",
+	"string/strndup.c",
+	"string/strnlen.c",
+	"string/strnstr.c",
+	"string/strpbrk.c",
+	"string/strrchr.c",
+	"string/strsep.c",
+	"string/strsignal.c",
+	"string/strspn.c",
+	"string/strstr.c",
+	"string/strtok.c",
+	"string/strtok_r.c",
+	"string/strupr.c",
+	"string/strverscmp.c",
+	"string/strxfrm.c",
+	"string/strxfrm_l.c",
+	"string/swab.c",
+	"string/timingsafe_bcmp.c",
+	"string/timingsafe_memcmp.c",
+	"string/u_strerr.c",
+	"string/wcpcpy.c",
+	"string/wcpncpy.c",
+	"string/wcscasecmp.c",
+	"string/wcscasecmp_l.c",
+	"string/wcscat.c",
+	"string/wcschr.c",
+	"string/wcscmp.c",
+	"string/wcscoll.c",
+	"string/wcscoll_l.c",
+	"string/wcscpy.c",
+	"string/wcscspn.c",
+	"string/wcsdup.c",
+	"string/wcslcat.c",
+	"string/wcslcpy.c",
+	"string/wcslen.c",
+	"string/wcsncasecmp.c",
+	"string/wcsncasecmp_l.c",
+	"string/wcsncat.c",
+	"string/wcsncmp.c",
+	"string/wcsncpy.c",
+	"string/wcsnlen.c",
+	"string/wcspbrk.c",
+	"string/wcsrchr.c",
+	"string/wcsspn.c",
+	"string/wcsstr.c",
+	"string/wcstok.c",
+	"string/wcswidth.c",
+	"string/wcsxfrm.c",
+	"string/wcsxfrm_l.c",
+	"string/wcwidth.c",
+	"string/wmemchr.c",
+	"string/wmemcmp.c",
+	"string/wmemcpy.c",
+	"string/wmemmove.c",
+	"string/wmempcpy.c",
+	"string/wmemset.c",
+	"string/xpg_strerror_r.c",
+}
@@ -84,7 +84,18 @@ func loadProgramSize(path string) (*programSize, error) {
 		if section.Type != elf.SHT_PROGBITS && section.Type != elf.SHT_NOBITS {
 			continue
 		}
-		if section.Type == elf.SHT_NOBITS {
+		if section.Name == ".stack" {
+			// HACK: this works around a bug in ld.lld from LLVM 10. The linker
+			// marks sections with no input symbols (such as is the case for the
+			// .stack section) as SHT_PROGBITS instead of SHT_NOBITS. While it
+			// doesn't affect the generated binaries (.hex and .bin), it does
+			// affect the reported size.
+			// https://bugs.llvm.org/show_bug.cgi?id=45336
+			// https://reviews.llvm.org/D76981
+			// It has been merged in master, but it has not (yet) been
+			// backported to the LLVM 10 release branch.
+			sumBSS += section.Size
+		} else if section.Type == elf.SHT_NOBITS {
 			sumBSS += section.Size
 		} else if section.Flags&elf.SHF_EXECINSTR != 0 {
 			sumCode += section.Size
@@ -0,0 +1,54 @@
+// +build byollvm
+
+package builder
+
+import (
+	"errors"
+	"unsafe"
+)
+
+/*
+#cgo CXXFLAGS: -fno-rtti
+#include <stdbool.h>
+#include <stdlib.h>
+bool tinygo_clang_driver(int argc, char **argv);
+bool tinygo_link_elf(int argc, char **argv);
+bool tinygo_link_wasm(int argc, char **argv);
+*/
+import "C"
+
+const hasBuiltinTools = true
+
+// RunTool runs the given tool (such as clang).
+//
+// This version actually runs the tools because TinyGo was compiled while
+// linking statically with LLVM (with the byollvm build tag).
+func RunTool(tool string, args ...string) error {
+	args = append([]string{"tinygo:" + tool}, args...)
+
+	var cflag *C.char
+	buf := C.calloc(C.size_t(len(args)), C.size_t(unsafe.Sizeof(cflag)))
+	defer C.free(buf)
+	cflags := (*[1 << 10]*C.char)(unsafe.Pointer(buf))[:len(args):len(args)]
+	for i, flag := range args {
+		cflag := C.CString(flag)
+		cflags[i] = cflag
+		defer C.free(unsafe.Pointer(cflag))
+	}
+
+	var ok C.bool
+	switch tool {
+	case "clang":
+		ok = C.tinygo_clang_driver(C.int(len(args)), (**C.char)(buf))
+	case "ld.lld":
+		ok = C.tinygo_link_elf(C.int(len(args)), (**C.char)(buf))
+	case "wasm-ld":
+		ok = C.tinygo_link_wasm(C.int(len(args)), (**C.char)(buf))
+	default:
+		return errors.New("unknown tool: " + tool)
+	}
+	if !ok {
+		return errors.New("failed to run tool: " + tool)
+	}
+	return nil
+}
@@ -0,0 +1,15 @@
+// +build !byollvm
+
+package builder
+
+import "errors"
+
+const hasBuiltinTools = false
+
+// RunTool runs the given tool (such as clang).
+//
+// This version doesn't actually run the tool: TinyGo has not been compiled by
+// statically linking to LLVM.
+func RunTool(tool string, args ...string) error {
+	return errors.New("cannot run tool: " + tool)
+}
@@ -0,0 +1,59 @@
+package builder
+
+import (
+	"errors"
+	"os"
+	"os/exec"
+
+	"github.com/tinygo-org/tinygo/goenv"
+)
+
+// runCCompiler invokes a C compiler with the given arguments.
+func runCCompiler(command string, flags ...string) error {
+	if hasBuiltinTools && command == "clang" {
+		// Compile this with the internal Clang compiler.
+		headerPath := getClangHeaderPath(goenv.Get("TINYGOROOT"))
+		if headerPath == "" {
+			return errors.New("could not locate Clang headers")
+		}
+		flags = append(flags, "-I"+headerPath)
+		cmd := exec.Command(os.Args[0], append([]string{"clang"}, flags...)...)
+		cmd.Stdout = os.Stdout
+		cmd.Stderr = os.Stderr
+		return cmd.Run()
+	}
+
+	// Running some other compiler. Maybe it has been defined in the
+	// commands map (unlikely).
+	if cmdNames, ok := commands[command]; ok {
+		return execCommand(cmdNames, flags...)
+	}
+
+	// Alternatively, run the compiler directly.
+	cmd := exec.Command(command, flags...)
+	cmd.Stdout = os.Stdout
+	cmd.Stderr = os.Stderr
+	return cmd.Run()
+}
+
+// link invokes a linker with the given name and flags.
+func link(linker string, flags ...string) error {
+	if hasBuiltinTools && (linker == "ld.lld" || linker == "wasm-ld") {
+		// Run command with internal linker.
+		cmd := exec.Command(os.Args[0], append([]string{linker}, flags...)...)
+		cmd.Stdout = os.Stdout
+		cmd.Stderr = os.Stderr
+		return cmd.Run()
+	}
+
+	// Fall back to external command.
+	if cmdNames, ok := commands[linker]; ok {
+		return execCommand(cmdNames, flags...)
+	}
+
+	cmd := exec.Command(linker, flags...)
+	cmd.Stdout = os.Stdout
+	cmd.Stderr = os.Stderr
+	cmd.Dir = goenv.Get("TINYGOROOT")
+	return cmd.Run()
+}
@@ -14,7 +14,9 @@ package cgo
 import (
 	"fmt"
 	"go/ast"
+	"go/scanner"
 	"go/token"
+	"path/filepath"
 	"sort"
 	"strconv"
 	"strings"
@@ -39,6 +41,7 @@ type cgoPackage struct {
 	elaboratedTypes map[string]*elaboratedTypeInfo
 	enums           map[string]enumInfo
 	anonStructNum   int
+	ldflags         []string
 }

 // constantInfo stores some information about a CGo constant found by libclang
@@ -51,9 +54,10 @@ type constantInfo struct {
 // functionInfo stores some information about a CGo function found by libclang
 // and declared in the AST.
 type functionInfo struct {
-	args    []paramInfo
-	results *ast.FieldList
-	pos     token.Pos
+	args     []paramInfo
+	results  *ast.FieldList
+	pos      token.Pos
+	variadic bool
 }

 // paramInfo is a parameter of a CGo function (see functionInfo).
@@ -154,7 +158,7 @@ typedef unsigned long long  _Cgo_ulonglong;
 // newly created *ast.File that should be added to the list of to-be-parsed
 // files. If there is one or more error, it returns these in the []error slice
 // but still modifies the AST.
-func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string) (*ast.File, []error) {
+func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string) (*ast.File, []string, []error) {
 	p := &cgoPackage{
 		dir:             dir,
 		fset:            fset,
@@ -168,11 +172,28 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 		enums:           map[string]enumInfo{},
 	}

+	// Disable _FORTIFY_SOURCE as it causes problems on macOS.
+	// Note that it is only disabled for memcpy (etc) calls made from Go, which
+	// have better alternatives anyway.
+	cflags = append(cflags, "-D_FORTIFY_SOURCE=0")
+
 	// Add a new location for the following file.
 	generatedTokenPos := p.fset.AddFile(dir+"/!cgo.go", -1, 0)
 	generatedTokenPos.SetLines([]int{0})
 	p.generatedPos = generatedTokenPos.Pos(0)

+	// Find the absolute path for this package.
+	packagePath, err := filepath.Abs(fset.File(files[0].Pos()).Name())
+	if err != nil {
+		return nil, nil, []error{
+			scanner.Error{
+				Pos: fset.Position(files[0].Pos()),
+				Msg: "cgo: cannot find absolute path: " + err.Error(), // TODO: wrap this error
+			},
+		}
+	}
+	packagePath = filepath.Dir(packagePath)
+
 	// Construct a new in-memory AST for CGo declarations of this package.
 	unsafeImport := &ast.ImportSpec{
 		Path: &ast.BasicLit{
@@ -338,7 +359,21 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 						p.addErrorAfter(comment.Slash, comment.Text[:lineStart+colon+1], err.Error())
 						continue
 					}
+					makePathsAbsolute(flags, packagePath)
 					cflags = append(cflags, flags...)
+				case "LDFLAGS":
+					flags, err := shlex.Split(value)
+					if err != nil {
+						// TODO: find the exact location where the error happened.
+						p.addErrorAfter(comment.Slash, comment.Text[:lineStart+colon+1], "failed to parse flags in #cgo line: "+err.Error())
+						continue
+					}
+					if err := checkLinkerFlags(name, flags); err != nil {
+						p.addErrorAfter(comment.Slash, comment.Text[:lineStart+colon+1], err.Error())
+						continue
+					}
+					makePathsAbsolute(flags, packagePath)
+					p.ldflags = append(p.ldflags, flags...)
 				default:
 					startPos := strings.LastIndex(line[4:colon], name) + 4
 					p.addErrorAfter(comment.Slash, comment.Text[:lineStart+startPos], "invalid #cgo line: "+name)
@@ -392,7 +427,31 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 	// Print the newly generated in-memory AST, for debugging.
 	//ast.Print(fset, p.generated)

-	return p.generated, p.errors
+	return p.generated, p.ldflags, p.errors
+}
+
+// makePathsAbsolute converts some common path compiler flags (-I, -L) from
+// relative flags into absolute flags, if they are relative. This is necessary
+// because the C compiler is usually not invoked from the package path.
+func makePathsAbsolute(args []string, packagePath string) {
+	nextIsPath := false
+	for i, arg := range args {
+		if nextIsPath {
+			if !filepath.IsAbs(arg) {
+				args[i] = filepath.Join(packagePath, arg)
+			}
+		}
+		if arg == "-I" || arg == "-L" {
+			nextIsPath = true
+			continue
+		}
+		if strings.HasPrefix(arg, "-I") || strings.HasPrefix(arg, "-L") {
+			path := arg[2:]
+			if !filepath.IsAbs(path) {
+				args[i] = arg[:2] + filepath.Join(packagePath, path)
+			}
+		}
+	}
 }

 // addFuncDecls adds the C function declarations found by libclang in the
@@ -426,6 +485,16 @@ func (p *cgoPackage) addFuncDecls() {
 				Results: fn.results,
 			},
 		}
+		if fn.variadic {
+			decl.Doc = &ast.CommentGroup{
+				List: []*ast.Comment{
+					&ast.Comment{
+						Slash: fn.pos,
+						Text:  "//go:variadic",
+					},
+				},
+			}
+		}
 		obj.Decl = decl
 		for i, arg := range fn.args {
 			args[i] = &ast.Field{
@@ -5,12 +5,14 @@ import (
 	"flag"
 	"fmt"
 	"go/ast"
+	"go/build"
 	"go/format"
 	"go/parser"
 	"go/token"
 	"go/types"
 	"io/ioutil"
 	"path/filepath"
+	"regexp"
 	"runtime"
 	"strings"
 	"testing"
@@ -19,12 +21,41 @@ import (
 // Pass -update to go test to update the output of the test files.
 var flagUpdate = flag.Bool("update", false, "Update images based on test output.")

+// normalizeResult normalizes Go source code that comes out of tests across
+// platforms and Go versions.
+func normalizeResult(result string) string {
+	actual := strings.Replace(result, "\r\n", "\n", -1)
+
+	// Make sure all functions are wrapped, even those that would otherwise be
+	// single-line functions. This is necessary because Go 1.14 changed the way
+	// such functions are wrapped and it's important to have consistent test
+	// results.
+	re := regexp.MustCompile(`func \((.+)\)( .*?) +{ (.+) }`)
+	actual = re.ReplaceAllString(actual, "func ($1)$2 {\n\t$3\n}")
+
+	return actual
+}
+
 func TestCGo(t *testing.T) {
 	var cflags = []string{"--target=armv6m-none-eabi"}

 	for _, name := range []string{"basic", "errors", "types", "flags", "const"} {
 		name := name // avoid a race condition
 		t.Run(name, func(t *testing.T) {
+			// Skip tests that require specific Go version.
+			if name == "errors" {
+				ok := false
+				for _, version := range build.Default.ReleaseTags {
+					if version == "go1.16" {
+						ok = true
+						break
+					}
+				}
+				if !ok {
+					t.Skip("Results for errors test are only valid for Go 1.16+")
+				}
+			}
+
 			// Read the AST in memory.
 			path := filepath.Join("testdata", name+".go")
 			fset := token.NewFileSet()
@@ -34,7 +65,7 @@ func TestCGo(t *testing.T) {
 			}

 			// Process the AST with CGo.
-			cgoAST, cgoErrors := Process([]*ast.File{f}, "testdata", fset, cflags)
+			cgoAST, _, cgoErrors := Process([]*ast.File{f}, "testdata", fset, cflags)

 			// Check the AST for type errors.
 			var typecheckErrors []error
@@ -74,7 +105,7 @@ func TestCGo(t *testing.T) {
 			if err != nil {
 				t.Errorf("could not write out CGo AST: %v", err)
 			}
-			actual := strings.Replace(string(buf.Bytes()), "\r\n", "\n", -1)
+			actual := normalizeResult(string(buf.Bytes()))

 			// Read the file with the expected output, to compare against.
 			outfile := filepath.Join("testdata", name+".out.go")
@@ -15,7 +15,7 @@ import (
 )

 /*
-#include <clang-c/Index.h> // if this fails, install libclang-9-dev
+#include <clang-c/Index.h> // if this fails, install libclang-10-dev
 #include <stdlib.h>
 #include <stdint.h>

@@ -152,12 +152,10 @@ func tinygo_clang_globals_visitor(c, parent C.GoCXCursor, client_data C.CXClient
 			return C.CXChildVisit_Continue
 		}
 		cursorType := C.tinygo_clang_getCursorType(c)
-		if C.clang_isFunctionTypeVariadic(cursorType) != 0 {
-			return C.CXChildVisit_Continue // not supported
-		}
 		numArgs := int(C.tinygo_clang_Cursor_getNumArguments(c))
 		fn := &functionInfo{
-			pos: pos,
+			pos:      pos,
+			variadic: C.clang_isFunctionTypeVariadic(cursorType) != 0,
 		}
 		p.functions[name] = fn
 		for i := 0; i < numArgs; i++ {
@@ -246,9 +244,9 @@ func tinygo_clang_globals_visitor(c, parent C.GoCXCursor, client_data C.CXClient
 		}
 		value := source[len(name):]
 		// Try to convert this #define into a Go constant expression.
-		expr, err := parseConst(pos+token.Pos(len(name)), p.fset, value)
-		if err != nil {
-			p.errors = append(p.errors, err)
+		expr, scannerError := parseConst(pos+token.Pos(len(name)), p.fset, value)
+		if scannerError != nil {
+			p.errors = append(p.errors, *scannerError)
 		}
 		if expr != nil {
 			// Parsing was successful.
@@ -1,13 +1,14 @@
 // +build !byollvm
+// +build !llvm10

 package cgo

 /*
-#cgo linux  CFLAGS: -I/usr/lib/llvm-9/include
-#cgo darwin CFLAGS: -I/usr/local/opt/llvm@9/include
-#cgo freebsd CFLAGS: -I/usr/local/llvm90/include
-#cgo linux  LDFLAGS: -L/usr/lib/llvm-9/lib -lclang
-#cgo darwin LDFLAGS: -L/usr/local/opt/llvm@9/lib -lclang -lffi
-#cgo freebsd LDFLAGS: -L/usr/local/llvm90/lib -lclang
+#cgo linux   CFLAGS:  -I/usr/lib/llvm-11/include
+#cgo darwin  CFLAGS:  -I/usr/local/opt/llvm@11/include
+#cgo freebsd CFLAGS:  -I/usr/local/llvm11/include
+#cgo linux   LDFLAGS: -L/usr/lib/llvm-11/lib -lclang
+#cgo darwin  LDFLAGS: -L/usr/local/opt/llvm@11/lib -lclang -lffi
+#cgo freebsd LDFLAGS: -L/usr/local/llvm11/lib -lclang
 */
 import "C"
@@ -0,0 +1,14 @@
+// +build !byollvm
+// +build llvm10
+
+package cgo
+
+/*
+#cgo linux  CFLAGS: -I/usr/lib/llvm-10/include
+#cgo darwin CFLAGS: -I/usr/local/opt/llvm@10/include
+#cgo freebsd CFLAGS: -I/usr/local/llvm10/include
+#cgo linux  LDFLAGS: -L/usr/lib/llvm-10/lib -lclang
+#cgo darwin LDFLAGS: -L/usr/local/opt/llvm@10/lib -lclang -lffi
+#cgo freebsd LDFLAGS: -L/usr/local/llvm10/lib -lclang
+*/
+import "C"
@@ -3,7 +3,7 @@
 // are slightly different from the ones defined in libclang.go, but they
 // should be ABI compatible.

-#include <clang-c/Index.h> // if this fails, install libclang-9-dev
+#include <clang-c/Index.h> // if this fails, install libclang-10-dev

 CXCursor tinygo_clang_getTranslationUnitCursor(CXTranslationUnit tu) {
 	return clang_getTranslationUnitCursor(tu);
@@ -4,10 +4,10 @@
 //     testdata/errors.go:13:23: unexpected token )

 // Type checking errors after CGo processing:
-//     testdata/errors.go:102: 2 << 10 (untyped int constant 2048) overflows uint8
+//     testdata/errors.go:102: cannot use 2 << 10 (untyped int constant 2048) as uint8 value in variable declaration (overflows)
 //     testdata/errors.go:105: unknown field z in struct literal
 //     testdata/errors.go:108: undeclared name: C.SOME_CONST_1
-//     testdata/errors.go:110: C.SOME_CONST_3 (untyped int constant 1234) overflows byte
+//     testdata/errors.go:110: cannot use C.SOME_CONST_3 (untyped int constant 1234) as byte value in variable declaration (overflows)

 package main

@@ -9,6 +9,9 @@ package main

 #cgo CFLAGS: -DFOO

+#cgo CFLAGS: -Iinclude
+#include "foo.h"
+
 #if defined(FOO)
 #define BAR 3
 #else
@@ -18,9 +21,17 @@ package main
 #if defined(NOTDEFINED)
 #warning flag must not be defined
 #endif
+
+// Check Compiler flags
+#cgo LDFLAGS: -lc
+
+// This flag is not valid ldflags
+#cgo LDFLAGS: -does-not-exists
+
 */
 import "C"

 var (
 	_ = C.BAR
+	_ = C.FOO_H
 )
@@ -1,6 +1,7 @@
 // CGo errors:
 //     testdata/flags.go:5:7: invalid #cgo line: NOFLAGS
 //     testdata/flags.go:8:13: invalid flag: -fdoes-not-exist
+//     testdata/flags.go:29:14: invalid flag: -does-not-exists

 package main

@@ -9,6 +10,7 @@ import "unsafe"
 var _ unsafe.Pointer

 const C.BAR = 3
+const C.FOO_H = 1

 type C.int16_t = int16
 type C.int32_t = int32
@@ -0,0 +1 @@
+#define FOO_H 1
@@ -104,6 +104,10 @@ typedef struct {
 	unsigned char e : 3;
 	// Note that C++ allows bitfields bigger than the underlying type.
 } bitfield_t;
+
+// Function signatures.
+void variadic0();
+void variadic2(int x, int y, ...);
 */
 import "C"

@@ -163,3 +167,9 @@ func accessUnion() {
 	var _ *C.int = union2d.unionfield_i()
 	var _ *[2]float64 = union2d.unionfield_d()
 }
+
+// Test function signatures.
+func accessFunctions() {
+	C.variadic0()
+	C.variadic2(3, 5)
+}
@@ -4,6 +4,11 @@ import "unsafe"

 var _ unsafe.Pointer

+func C.variadic0() //go:variadic
+func C.variadic2(x C.int, y C.int) //go:variadic
+var C.variadic0$funcaddr unsafe.Pointer
+var C.variadic2$funcaddr unsafe.Pointer
+
 const C.option2A = 20
 const C.optionA = 0
 const C.optionB = 1
@@ -62,12 +67,18 @@ type C.union3_t = C.union_1
 type C.union_nested_t = C.union_3
 type C.unionarray_t = struct{ arr [10]C.uchar }

-func (s *C.struct_4) bitfield_a() C.uchar          { return s.__bitfield_1 & 0x1f }
-func (s *C.struct_4) set_bitfield_a(value C.uchar) { s.__bitfield_1 = s.__bitfield_1&^0x1f | value&0x1f<<0 }
+func (s *C.struct_4) bitfield_a() C.uchar {
+	return s.__bitfield_1 & 0x1f
+}
+func (s *C.struct_4) set_bitfield_a(value C.uchar) {
+	s.__bitfield_1 = s.__bitfield_1&^0x1f | value&0x1f<<0
+}
 func (s *C.struct_4) bitfield_b() C.uchar {
 	return s.__bitfield_1 >> 5 & 0x1
 }
-func (s *C.struct_4) set_bitfield_b(value C.uchar) { s.__bitfield_1 = s.__bitfield_1&^0x20 | value&0x1<<5 }
+func (s *C.struct_4) set_bitfield_b(value C.uchar) {
+	s.__bitfield_1 = s.__bitfield_1&^0x20 | value&0x1<<5
+}
 func (s *C.struct_4) bitfield_c() C.uchar {
 	return s.__bitfield_1 >> 6
 }
@@ -94,25 +105,45 @@ type C.struct_type1 struct {
 }
 type C.struct_type2 struct{ _type C.int }

-func (union *C.union_1) unionfield_i() *C.int   { return (*C.int)(unsafe.Pointer(&union.$union)) }
-func (union *C.union_1) unionfield_d() *float64 { return (*float64)(unsafe.Pointer(&union.$union)) }
-func (union *C.union_1) unionfield_s() *C.short { return (*C.short)(unsafe.Pointer(&union.$union)) }
+func (union *C.union_1) unionfield_i() *C.int {
+	return (*C.int)(unsafe.Pointer(&union.$union))
+}
+func (union *C.union_1) unionfield_d() *float64 {
+	return (*float64)(unsafe.Pointer(&union.$union))
+}
+func (union *C.union_1) unionfield_s() *C.short {
+	return (*C.short)(unsafe.Pointer(&union.$union))
+}

 type C.union_1 struct{ $union uint64 }

-func (union *C.union_2) unionfield_area() *C.point2d_t  { return (*C.point2d_t)(unsafe.Pointer(&union.$union)) }
-func (union *C.union_2) unionfield_solid() *C.point3d_t { return (*C.point3d_t)(unsafe.Pointer(&union.$union)) }
+func (union *C.union_2) unionfield_area() *C.point2d_t {
+	return (*C.point2d_t)(unsafe.Pointer(&union.$union))
+}
+func (union *C.union_2) unionfield_solid() *C.point3d_t {
+	return (*C.point3d_t)(unsafe.Pointer(&union.$union))
+}

 type C.union_2 struct{ $union [3]uint32 }

-func (union *C.union_3) unionfield_point() *C.point3d_t    { return (*C.point3d_t)(unsafe.Pointer(&union.$union)) }
-func (union *C.union_3) unionfield_array() *C.unionarray_t { return (*C.unionarray_t)(unsafe.Pointer(&union.$union)) }
-func (union *C.union_3) unionfield_thing() *C.union3_t     { return (*C.union3_t)(unsafe.Pointer(&union.$union)) }
+func (union *C.union_3) unionfield_point() *C.point3d_t {
+	return (*C.point3d_t)(unsafe.Pointer(&union.$union))
+}
+func (union *C.union_3) unionfield_array() *C.unionarray_t {
+	return (*C.unionarray_t)(unsafe.Pointer(&union.$union))
+}
+func (union *C.union_3) unionfield_thing() *C.union3_t {
+	return (*C.union3_t)(unsafe.Pointer(&union.$union))
+}

 type C.union_3 struct{ $union [2]uint64 }

-func (union *C.union_union2d) unionfield_i() *C.int      { return (*C.int)(unsafe.Pointer(&union.$union)) }
-func (union *C.union_union2d) unionfield_d() *[2]float64 { return (*[2]float64)(unsafe.Pointer(&union.$union)) }
+func (union *C.union_union2d) unionfield_i() *C.int {
+	return (*C.int)(unsafe.Pointer(&union.$union))
+}
+func (union *C.union_union2d) unionfield_d() *[2]float64 {
+	return (*[2]float64)(unsafe.Pointer(&union.$union))
+}

 type C.union_union2d struct{ $union [2]uint64 }
 type C.enum_option C.int
@@ -5,8 +5,8 @@ package compileopts
 import (
 	"errors"
 	"fmt"
+	"path/filepath"
 	"regexp"
-	"strconv"
 	"strings"

 	"github.com/tinygo-org/tinygo/goenv"
@@ -72,7 +72,7 @@ func (c *Config) CgoEnabled() bool {
 }

 // GC returns the garbage collection strategy in use on this platform. Valid
-// values are "none", "leaking", and "conservative".
+// values are "none", "leaking", "extalloc", and "conservative".
 func (c *Config) GC() string {
 	if c.Options.GC != "" {
 		return c.Options.GC
@@ -80,26 +80,33 @@ func (c *Config) GC() string {
 	if c.Target.GC != "" {
 		return c.Target.GC
 	}
-	return "conservative"
+	for _, tag := range c.Target.BuildTags {
+		if tag == "baremetal" || tag == "wasm" {
+			return "conservative"
+		}
+	}
+	return "extalloc"
 }

 // NeedsStackObjects returns true if the compiler should insert stack objects
 // that can be traced by the garbage collector.
 func (c *Config) NeedsStackObjects() bool {
-	if c.GC() != "conservative" {
+	switch c.GC() {
+	case "conservative", "extalloc":
+		for _, tag := range c.BuildTags() {
+			if tag == "wasm" {
+				return true
+			}
+		}
+
+		return false
+	default:
 		return false
 	}
-	for _, tag := range c.BuildTags() {
-		if tag == "baremetal" {
-			return false
-		}
-	}
-
-	return true
 }

-// Scheduler returns the scheduler implementation. Valid values are "coroutines"
-// and "tasks".
+// Scheduler returns the scheduler implementation. Valid values are "none",
+//"coroutines" and "tasks".
 func (c *Config) Scheduler() string {
 	if c.Options.Scheduler != "" {
 		return c.Options.Scheduler
@@ -111,6 +118,31 @@ func (c *Config) Scheduler() string {
 	return "coroutines"
 }

+// FuncImplementation picks an appropriate func value implementation for the
+// target.
+func (c *Config) FuncImplementation() string {
+	switch c.Scheduler() {
+	case "tasks":
+		// A func value is implemented as a pair of pointers:
+		//     {context, function pointer}
+		// where the context may be a pointer to a heap-allocated struct
+		// containing the free variables, or it may be undef if the function
+		// being pointed to doesn't need a context. The function pointer is a
+		// regular function pointer.
+		return "doubleword"
+	case "none", "coroutines":
+		// As "doubleword", but with the function pointer replaced by a unique
+		// ID per function signature. Function values are called by using a
+		// switch statement and choosing which function to call.
+		// Pick the switch implementation with the coroutines scheduler, as it
+		// allows the use of blocking inside a function that is used as a func
+		// value.
+		return "switch"
+	default:
+		panic("unknown scheduler type")
+	}
+}
+
 // PanicStrategy returns the panic strategy selected for this target. Valid
 // values are "print" (print the panic value, then exit) or "trap" (issue a trap
 // instruction).
@@ -118,6 +150,16 @@ func (c *Config) PanicStrategy() string {
 	return c.Options.PanicStrategy
 }

+// AutomaticStackSize returns whether goroutine stack sizes should be determined
+// automatically at compile time, if possible. If it is false, no attempt is
+// made.
+func (c *Config) AutomaticStackSize() bool {
+	if c.Target.AutoStackSize != nil && c.Scheduler() == "tasks" {
+		return *c.Target.AutoStackSize
+	}
+	return false
+}
+
 // CFlags returns the flags to pass to the C compiler. This is necessary for CGo
 // preprocessing.
 func (c *Config) CFlags() []string {
@@ -125,6 +167,14 @@ func (c *Config) CFlags() []string {
 	for _, flag := range c.Target.CFlags {
 		cflags = append(cflags, strings.Replace(flag, "{root}", goenv.Get("TINYGOROOT"), -1))
 	}
+	if c.Target.Libc == "picolibc" {
+		root := goenv.Get("TINYGOROOT")
+		cflags = append(cflags, "-nostdlibinc", "-Xclang", "-internal-isystem", "-Xclang", filepath.Join(root, "lib", "picolibc", "newlib", "libc", "include"))
+		cflags = append(cflags, "-I"+filepath.Join(root, "lib/picolibc-include"))
+	}
+	if c.Debug() {
+		cflags = append(cflags, "-g")
+	}
 	return cflags
 }

@@ -139,12 +189,6 @@ func (c *Config) LDFlags() []string {
 		ldflags = append(ldflags, strings.Replace(flag, "{root}", root, -1))
 	}
 	ldflags = append(ldflags, "-L", root)
-	if c.Target.GOARCH == "wasm" {
-		// Round heap size to next multiple of 65536 (the WebAssembly page
-		// size).
-		heapSize := (c.Options.HeapSize + (65536 - 1)) &^ (65536 - 1)
-		ldflags = append(ldflags, "--initial-memory="+strconv.FormatInt(heapSize, 10))
-	}
 	if c.Target.LinkerScript != "" {
 		ldflags = append(ldflags, "-T", c.Target.LinkerScript)
 	}
@@ -175,6 +219,31 @@ func (c *Config) Debug() bool {
 	return c.Options.Debug
 }

+// BinaryFormat returns an appropriate binary format, based on the file
+// extension and the configured binary format in the target JSON file.
+func (c *Config) BinaryFormat(ext string) string {
+	switch ext {
+	case ".bin", ".gba", ".nro":
+		// The simplest format possible: dump everything in a raw binary file.
+		if c.Target.BinaryFormat != "" {
+			return c.Target.BinaryFormat
+		}
+		return "bin"
+	case ".hex":
+		// Similar to bin, but includes the start address and is thus usually a
+		// better format.
+		return "hex"
+	case ".uf2":
+		// Special purpose firmware format, mainly used on Adafruit boards.
+		// More information:
+		// https://github.com/Microsoft/uf2
+		return "uf2"
+	default:
+		// Use the ELF format for unrecognized file formats.
+		return "elf"
+	}
+}
+
 // Programmer returns the flash method and OpenOCD interface name given a
 // particular configuration. It may either be all configured in the target JSON
 // file or be modified using the -programmmer command-line option.
@@ -214,6 +283,9 @@ func (c *Config) OpenOCDConfiguration() (args []string, err error) {
 		return nil, fmt.Errorf("unknown OpenOCD transport: %#v", c.Target.OpenOCDTransport)
 	}
 	args = []string{"-f", "interface/" + openocdInterface + ".cfg"}
+	for _, cmd := range c.Target.OpenOCDCommands {
+		args = append(args, "-c", cmd)
+	}
 	if c.Target.OpenOCDTransport != "" {
 		args = append(args, "-c", "transport select "+c.Target.OpenOCDTransport)
 	}
@@ -221,6 +293,34 @@ func (c *Config) OpenOCDConfiguration() (args []string, err error) {
 	return args, nil
 }

+// CodeModel returns the code model used on this platform.
+func (c *Config) CodeModel() string {
+	if c.Target.CodeModel != "" {
+		return c.Target.CodeModel
+	}
+
+	return "default"
+}
+
+// RelocationModel returns the relocation model in use on this platform. Valid
+// values are "static", "pic", "dynamicnopic".
+func (c *Config) RelocationModel() string {
+	if c.Target.RelocationModel != "" {
+		return c.Target.RelocationModel
+	}
+
+	return "static"
+}
+
+// WasmAbi returns the WASM ABI which is specified in the target JSON file, and
+// the value is overridden by `-wasm-abi` flag if it is provided
+func (c *Config) WasmAbi() string {
+	if c.Options.WasmAbi != "" {
+		return c.Options.WasmAbi
+	}
+	return c.Target.WasmAbi
+}
+
 type TestConfig struct {
 	CompileTestBinary bool
 	// TODO: Filter the test functions to run, include verbose flag, etc
@@ -1,5 +1,17 @@
 package compileopts

+import (
+	"fmt"
+	"strings"
+)
+
+var (
+	validGCOptions            = []string{"none", "leaking", "extalloc", "conservative"}
+	validSchedulerOptions     = []string{"none", "tasks", "coroutines"}
+	validPrintSizeOptions     = []string{"none", "short", "full"}
+	validPanicStrategyOptions = []string{"print", "trap"}
+)
+
 // Options contains extra options to give to the compiler. These options are
 // usually passed from the command line.
 type Options struct {
@@ -11,13 +23,64 @@ type Options struct {
 	PrintIR       bool
 	DumpSSA       bool
 	VerifyIR      bool
+	PrintCommands bool
 	Debug         bool
 	PrintSizes    string
+	PrintStacks   bool
 	CFlags        []string
 	LDFlags       []string
 	Tags          string
 	WasmAbi       string
-	HeapSize      int64
 	TestConfig    TestConfig
 	Programmer    string
 }
+
+// Verify performs a validation on the given options, raising an error if options are not valid.
+func (o *Options) Verify() error {
+	if o.GC != "" {
+		valid := isInArray(validGCOptions, o.GC)
+		if !valid {
+			return fmt.Errorf(`invalid gc option '%s': valid values are %s`,
+				o.GC,
+				strings.Join(validGCOptions, ", "))
+		}
+	}
+
+	if o.Scheduler != "" {
+		valid := isInArray(validSchedulerOptions, o.Scheduler)
+		if !valid {
+			return fmt.Errorf(`invalid scheduler option '%s': valid values are %s`,
+				o.Scheduler,
+				strings.Join(validSchedulerOptions, ", "))
+		}
+	}
+
+	if o.PrintSizes != "" {
+		valid := isInArray(validPrintSizeOptions, o.PrintSizes)
+		if !valid {
+			return fmt.Errorf(`invalid size option '%s': valid values are %s`,
+				o.PrintSizes,
+				strings.Join(validPrintSizeOptions, ", "))
+		}
+	}
+
+	if o.PanicStrategy != "" {
+		valid := isInArray(validPanicStrategyOptions, o.PanicStrategy)
+		if !valid {
+			return fmt.Errorf(`invalid panic option '%s': valid values are %s`,
+				o.PanicStrategy,
+				strings.Join(validPanicStrategyOptions, ", "))
+		}
+	}
+
+	return nil
+}
+
+func isInArray(arr []string, item string) bool {
+	for _, i := range arr {
+		if i == item {
+			return true
+		}
+	}
+	return false
+}
@@ -0,0 +1,138 @@
+package compileopts_test
+
+import (
+	"errors"
+	"testing"
+
+	"github.com/tinygo-org/tinygo/compileopts"
+)
+
+func TestVerifyOptions(t *testing.T) {
+
+	expectedGCError := errors.New(`invalid gc option 'incorrect': valid values are none, leaking, extalloc, conservative`)
+	expectedSchedulerError := errors.New(`invalid scheduler option 'incorrect': valid values are none, tasks, coroutines`)
+	expectedPrintSizeError := errors.New(`invalid size option 'incorrect': valid values are none, short, full`)
+	expectedPanicStrategyError := errors.New(`invalid panic option 'incorrect': valid values are print, trap`)
+
+	testCases := []struct {
+		name          string
+		opts          compileopts.Options
+		expectedError error
+	}{
+		{
+			name: "OptionsEmpty",
+			opts: compileopts.Options{},
+		},
+		{
+			name: "InvalidGCOption",
+			opts: compileopts.Options{
+				GC: "incorrect",
+			},
+			expectedError: expectedGCError,
+		},
+		{
+			name: "GCOptionNone",
+			opts: compileopts.Options{
+				GC: "none",
+			},
+		},
+		{
+			name: "GCOptionLeaking",
+			opts: compileopts.Options{
+				GC: "leaking",
+			},
+		},
+		{
+			name: "GCOptionExtalloc",
+			opts: compileopts.Options{
+				GC: "extalloc",
+			},
+		},
+		{
+			name: "GCOptionConservative",
+			opts: compileopts.Options{
+				GC: "conservative",
+			},
+		},
+		{
+			name: "InvalidSchedulerOption",
+			opts: compileopts.Options{
+				Scheduler: "incorrect",
+			},
+			expectedError: expectedSchedulerError,
+		},
+		{
+			name: "SchedulerOptionNone",
+			opts: compileopts.Options{
+				Scheduler: "none",
+			},
+		},
+		{
+			name: "SchedulerOptionTasks",
+			opts: compileopts.Options{
+				Scheduler: "tasks",
+			},
+		},
+		{
+			name: "SchedulerOptionCoroutines",
+			opts: compileopts.Options{
+				Scheduler: "coroutines",
+			},
+		},
+		{
+			name: "InvalidPrintSizeOption",
+			opts: compileopts.Options{
+				PrintSizes: "incorrect",
+			},
+			expectedError: expectedPrintSizeError,
+		},
+		{
+			name: "PrintSizeOptionNone",
+			opts: compileopts.Options{
+				PrintSizes: "none",
+			},
+		},
+		{
+			name: "PrintSizeOptionShort",
+			opts: compileopts.Options{
+				PrintSizes: "short",
+			},
+		},
+		{
+			name: "PrintSizeOptionFull",
+			opts: compileopts.Options{
+				PrintSizes: "full",
+			},
+		},
+		{
+			name: "InvalidPanicOption",
+			opts: compileopts.Options{
+				PanicStrategy: "incorrect",
+			},
+			expectedError: expectedPanicStrategyError,
+		},
+		{
+			name: "PanicOptionPrint",
+			opts: compileopts.Options{
+				PanicStrategy: "print",
+			},
+		},
+		{
+			name: "PanicOptionTrap",
+			opts: compileopts.Options{
+				PanicStrategy: "trap",
+			},
+		},
+	}
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			err := tc.opts.Verify()
+			if tc.expectedError != err {
+				if tc.expectedError.Error() != err.Error() {
+					t.Errorf("expected %v, got %v", tc.expectedError, err)
+				}
+			}
+		})
+	}
+}
@@ -8,6 +8,7 @@ import (
 	"io"
 	"os"
 	"path/filepath"
+	"reflect"
 	"runtime"
 	"strings"

@@ -32,11 +33,14 @@ type TargetSpec struct {
 	Compiler         string   `json:"compiler"`
 	Linker           string   `json:"linker"`
 	RTLib            string   `json:"rtlib"` // compiler runtime library (libgcc, compiler-rt)
+	Libc             string   `json:"libc"`
+	AutoStackSize    *bool    `json:"automatic-stack-size"` // Determine stack size automatically at compile time.
+	DefaultStackSize uint64   `json:"default-stack-size"`   // Default stack size if the size couldn't be determined at compile time.
 	CFlags           []string `json:"cflags"`
 	LDFlags          []string `json:"ldflags"`
 	LinkerScript     string   `json:"linkerscript"`
 	ExtraFiles       []string `json:"extra-files"`
-	Emulator         []string `json:"emulator"`
+	Emulator         []string `json:"emulator" override:"copy"` // inherited Emulator must not be append
 	FlashCommand     string   `json:"flash-command"`
 	GDB              string   `json:"gdb"`
 	PortReset        string   `json:"flash-1200-bps-reset"`
@@ -44,87 +48,57 @@ type TargetSpec struct {
 	FlashVolume      string   `json:"msd-volume-name"`
 	FlashFilename    string   `json:"msd-firmware-name"`
 	UF2FamilyID      string   `json:"uf2-family-id"`
+	BinaryFormat     string   `json:"binary-format"`
 	OpenOCDInterface string   `json:"openocd-interface"`
 	OpenOCDTarget    string   `json:"openocd-target"`
 	OpenOCDTransport string   `json:"openocd-transport"`
+	OpenOCDCommands  []string `json:"openocd-commands"`
 	JLinkDevice      string   `json:"jlink-device"`
+	CodeModel        string   `json:"code-model"`
+	RelocationModel  string   `json:"relocation-model"`
+	WasmAbi          string   `json:"wasm-abi"`
 }

-// copyProperties copies all properties that are set in spec2 into itself.
-func (spec *TargetSpec) copyProperties(spec2 *TargetSpec) {
-	// TODO: simplify this using reflection? Inherits and BuildTags are special
-	// cases, but the rest can simply be copied if set.
-	spec.Inherits = append(spec.Inherits, spec2.Inherits...)
-	if spec2.Triple != "" {
-		spec.Triple = spec2.Triple
-	}
-	if spec2.CPU != "" {
-		spec.CPU = spec2.CPU
-	}
-	spec.Features = append(spec.Features, spec2.Features...)
-	if spec2.GOOS != "" {
-		spec.GOOS = spec2.GOOS
-	}
-	if spec2.GOARCH != "" {
-		spec.GOARCH = spec2.GOARCH
-	}
-	spec.BuildTags = append(spec.BuildTags, spec2.BuildTags...)
-	if spec2.GC != "" {
-		spec.GC = spec2.GC
-	}
-	if spec2.Scheduler != "" {
-		spec.Scheduler = spec2.Scheduler
-	}
-	if spec2.Compiler != "" {
-		spec.Compiler = spec2.Compiler
-	}
-	if spec2.Linker != "" {
-		spec.Linker = spec2.Linker
-	}
-	if spec2.RTLib != "" {
-		spec.RTLib = spec2.RTLib
-	}
-	spec.CFlags = append(spec.CFlags, spec2.CFlags...)
-	spec.LDFlags = append(spec.LDFlags, spec2.LDFlags...)
-	if spec2.LinkerScript != "" {
-		spec.LinkerScript = spec2.LinkerScript
-	}
-	spec.ExtraFiles = append(spec.ExtraFiles, spec2.ExtraFiles...)
-	if len(spec2.Emulator) != 0 {
-		spec.Emulator = spec2.Emulator
-	}
-	if spec2.FlashCommand != "" {
-		spec.FlashCommand = spec2.FlashCommand
-	}
-	if spec2.GDB != "" {
-		spec.GDB = spec2.GDB
-	}
-	if spec2.PortReset != "" {
-		spec.PortReset = spec2.PortReset
-	}
-	if spec2.FlashMethod != "" {
-		spec.FlashMethod = spec2.FlashMethod
-	}
-	if spec2.FlashVolume != "" {
-		spec.FlashVolume = spec2.FlashVolume
-	}
-	if spec2.FlashFilename != "" {
-		spec.FlashFilename = spec2.FlashFilename
-	}
-	if spec2.UF2FamilyID != "" {
-		spec.UF2FamilyID = spec2.UF2FamilyID
-	}
-	if spec2.OpenOCDInterface != "" {
-		spec.OpenOCDInterface = spec2.OpenOCDInterface
-	}
-	if spec2.OpenOCDTarget != "" {
-		spec.OpenOCDTarget = spec2.OpenOCDTarget
-	}
-	if spec2.OpenOCDTransport != "" {
-		spec.OpenOCDTransport = spec2.OpenOCDTransport
-	}
-	if spec2.JLinkDevice != "" {
-		spec.JLinkDevice = spec2.JLinkDevice
+// overrideProperties overrides all properties that are set in child into itself using reflection.
+func (spec *TargetSpec) overrideProperties(child *TargetSpec) {
+	specType := reflect.TypeOf(spec).Elem()
+	specValue := reflect.ValueOf(spec).Elem()
+	childValue := reflect.ValueOf(child).Elem()
+
+	for i := 0; i < specType.NumField(); i++ {
+		field := specType.Field(i)
+		src := childValue.Field(i)
+		dst := specValue.Field(i)
+
+		switch kind := field.Type.Kind(); kind {
+		case reflect.String: // for strings, just copy the field of child to spec if not empty
+			if src.Len() > 0 {
+				dst.Set(src)
+			}
+		case reflect.Uint, reflect.Uint32, reflect.Uint64: // for Uint, copy if not zero
+			if src.Uint() != 0 {
+				dst.Set(src)
+			}
+		case reflect.Ptr: // for pointers, copy if not nil
+			if !src.IsNil() {
+				dst.Set(src)
+			}
+		case reflect.Slice: // for slices...
+			if src.Len() > 0 { // ... if not empty ...
+				switch tag := field.Tag.Get("override"); tag {
+				case "copy":
+					// copy the field of child to spec
+					dst.Set(src)
+				case "append", "":
+					// or append the field of child to spec
+					dst.Set(reflect.AppendSlice(src, dst))
+				default:
+					panic("override mode must be 'copy' or 'append' (default). I don't know how to '" + tag + "'.")
+				}
+			}
+		default:
+			panic("unknown field type : " + kind.String())
+		}
 	}
 }

@@ -173,11 +147,11 @@ func (spec *TargetSpec) resolveInherits() error {
 		if err != nil {
 			return err
 		}
-		newSpec.copyProperties(subtarget)
+		newSpec.overrideProperties(subtarget)
 	}

 	// When all properties are loaded, make sure they are properly inherited.
-	newSpec.copyProperties(spec)
+	newSpec.overrideProperties(spec)
 	*spec = *newSpec

 	return nil
@@ -243,6 +217,7 @@ func LoadTarget(target string) (*TargetSpec, error) {
 		}
 		goarch := map[string]string{ // map from LLVM arch to Go arch
 			"i386":    "386",
+			"i686":    "386",
 			"x86_64":  "amd64",
 			"aarch64": "arm64",
 			"armv7":   "arm",
@@ -258,39 +233,40 @@ func defaultTarget(goos, goarch, triple string) (*TargetSpec, error) {
 	// No target spec available. Use the default one, useful on most systems
 	// with a regular OS.
 	spec := TargetSpec{
-		Triple:      triple,
-		GOOS:        goos,
-		GOARCH:      goarch,
-		BuildTags:   []string{goos, goarch},
-		Compiler:    "clang",
-		Linker:      "cc",
-		CFlags:      []string{"--target=" + triple},
-		GDB:         "gdb",
-		PortReset:   "false",
-		FlashMethod: "native",
+		Triple:    triple,
+		GOOS:      goos,
+		GOARCH:    goarch,
+		BuildTags: []string{goos, goarch},
+		Compiler:  "clang",
+		Linker:    "cc",
+		CFlags:    []string{"--target=" + triple},
+		GDB:       "gdb",
+		PortReset: "false",
 	}
 	if goos == "darwin" {
 		spec.LDFlags = append(spec.LDFlags, "-Wl,-dead_strip")
 	} else {
 		spec.LDFlags = append(spec.LDFlags, "-no-pie", "-Wl,--gc-sections") // WARNING: clang < 5.0 requires -nopie
 	}
+	if goarch != "wasm" {
+		spec.ExtraFiles = append(spec.ExtraFiles, "src/runtime/gc_"+goarch+".S")
+	}
 	if goarch != runtime.GOARCH {
 		// Some educated guesses as to how to invoke helper programs.
+		spec.GDB = "gdb-multiarch"
 		if goarch == "arm" && goos == "linux" {
 			spec.CFlags = append(spec.CFlags, "--sysroot=/usr/arm-linux-gnueabihf")
 			spec.Linker = "arm-linux-gnueabihf-gcc"
-			spec.GDB = "arm-linux-gnueabihf-gdb"
 			spec.Emulator = []string{"qemu-arm", "-L", "/usr/arm-linux-gnueabihf"}
 		}
 		if goarch == "arm64" && goos == "linux" {
 			spec.CFlags = append(spec.CFlags, "--sysroot=/usr/aarch64-linux-gnu")
 			spec.Linker = "aarch64-linux-gnu-gcc"
-			spec.GDB = "aarch64-linux-gnu-gdb"
 			spec.Emulator = []string{"qemu-aarch64", "-L", "/usr/aarch64-linux-gnu"}
 		}
-		if goarch == "386" {
-			spec.CFlags = []string{"-m32"}
-			spec.LDFlags = []string{"-m32"}
+		if goarch == "386" && runtime.GOARCH == "amd64" {
+			spec.CFlags = append(spec.CFlags, "-m32")
+			spec.LDFlags = append(spec.LDFlags, "-m32")
 		}
 	}
 	return &spec, nil
@@ -1,6 +1,9 @@
 package compileopts

-import "testing"
+import (
+	"reflect"
+	"testing"
+)

 func TestLoadTarget(t *testing.T) {
 	_, err := LoadTarget("arduino")
@@ -17,3 +20,67 @@ func TestLoadTarget(t *testing.T) {
 		t.Error("LoadTarget failed for wrong reason:", err)
 	}
 }
+
+func TestOverrideProperties(t *testing.T) {
+	baseAutoStackSize := true
+	base := &TargetSpec{
+		GOOS:             "baseGoos",
+		CPU:              "baseCpu",
+		Features:         []string{"bf1", "bf2"},
+		BuildTags:        []string{"bt1", "bt2"},
+		Emulator:         []string{"be1", "be2"},
+		DefaultStackSize: 42,
+		AutoStackSize:    &baseAutoStackSize,
+	}
+	childAutoStackSize := false
+	child := &TargetSpec{
+		GOOS:             "",
+		CPU:              "chlidCpu",
+		Features:         []string{"cf1", "cf2"},
+		Emulator:         []string{"ce1", "ce2"},
+		AutoStackSize:    &childAutoStackSize,
+		DefaultStackSize: 64,
+	}
+
+	base.overrideProperties(child)
+
+	if base.GOOS != "baseGoos" {
+		t.Errorf("Overriding failed : got %v", base.GOOS)
+	}
+	if base.CPU != "chlidCpu" {
+		t.Errorf("Overriding failed : got %v", base.CPU)
+	}
+	if !reflect.DeepEqual(base.Features, []string{"cf1", "cf2", "bf1", "bf2"}) {
+		t.Errorf("Overriding failed : got %v", base.Features)
+	}
+	if !reflect.DeepEqual(base.BuildTags, []string{"bt1", "bt2"}) {
+		t.Errorf("Overriding failed : got %v", base.BuildTags)
+	}
+	if !reflect.DeepEqual(base.Emulator, []string{"ce1", "ce2"}) {
+		t.Errorf("Overriding failed : got %v", base.Emulator)
+	}
+	if *base.AutoStackSize != false {
+		t.Errorf("Overriding failed : got %v", base.AutoStackSize)
+	}
+	if base.DefaultStackSize != 64 {
+		t.Errorf("Overriding failed : got %v", base.DefaultStackSize)
+	}
+
+	baseAutoStackSize = true
+	base = &TargetSpec{
+		AutoStackSize:    &baseAutoStackSize,
+		DefaultStackSize: 42,
+	}
+	child = &TargetSpec{
+		AutoStackSize:    nil,
+		DefaultStackSize: 0,
+	}
+	base.overrideProperties(child)
+	if *base.AutoStackSize != true {
+		t.Errorf("Overriding failed : got %v", base.AutoStackSize)
+	}
+	if base.DefaultStackSize != 42 {
+		t.Errorf("Overriding failed : got %v", base.DefaultStackSize)
+	}
+
+}
@@ -4,16 +4,19 @@ package compiler
 // required by the Go programming language.

 import (
+	"fmt"
+	"go/token"
 	"go/types"

+	"golang.org/x/tools/go/ssa"
 	"tinygo.org/x/go-llvm"
 )

-// emitLookupBoundsCheck emits a bounds check before doing a lookup into a
+// createLookupBoundsCheck emits a bounds check before doing a lookup into a
 // slice. This is required by the Go language spec: an index out of bounds must
 // cause a panic.
-func (c *Compiler) emitLookupBoundsCheck(frame *Frame, arrayLen, index llvm.Value, indexType types.Type) {
-	if frame.fn.IsNoBounds() {
+func (b *builder) createLookupBoundsCheck(arrayLen, index llvm.Value, indexType types.Type) {
+	if b.info.nobounds {
 		// The //go:nobounds pragma was added to the function to avoid bounds
 		// checking.
 		return
@@ -23,41 +26,29 @@ func (c *Compiler) emitLookupBoundsCheck(frame *Frame, arrayLen, index llvm.Valu
 		// Sometimes, the index can be e.g. an uint8 or int8, and we have to
 		// correctly extend that type.
 		if indexType.Underlying().(*types.Basic).Info()&types.IsUnsigned == 0 {
-			index = c.builder.CreateZExt(index, arrayLen.Type(), "")
+			index = b.CreateZExt(index, arrayLen.Type(), "")
 		} else {
-			index = c.builder.CreateSExt(index, arrayLen.Type(), "")
+			index = b.CreateSExt(index, arrayLen.Type(), "")
 		}
 	} else if index.Type().IntTypeWidth() > arrayLen.Type().IntTypeWidth() {
 		// The index is bigger than the array length type, so extend it.
-		arrayLen = c.builder.CreateZExt(arrayLen, index.Type(), "")
+		arrayLen = b.CreateZExt(arrayLen, index.Type(), "")
 	}

-	faultBlock := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "lookup.outofbounds")
-	nextBlock := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "lookup.next")
-	frame.blockExits[frame.currentBlock] = nextBlock // adjust outgoing block for phi nodes
-
 	// Now do the bounds check: index >= arrayLen
-	outOfBounds := c.builder.CreateICmp(llvm.IntUGE, index, arrayLen, "")
-	c.builder.CreateCondBr(outOfBounds, faultBlock, nextBlock)
-
-	// Fail: this is a nil pointer, exit with a panic.
-	c.builder.SetInsertPointAtEnd(faultBlock)
-	c.createRuntimeCall("lookupPanic", nil, "")
-	c.builder.CreateUnreachable()
-
-	// Ok: this is a valid pointer.
-	c.builder.SetInsertPointAtEnd(nextBlock)
+	outOfBounds := b.CreateICmp(llvm.IntUGE, index, arrayLen, "")
+	b.createRuntimeAssert(outOfBounds, "lookup", "lookupPanic")
 }

-// emitSliceBoundsCheck emits a bounds check before a slicing operation to make
+// createSliceBoundsCheck emits a bounds check before a slicing operation to make
 // sure it is within bounds.
 //
 // This function is both used for slicing a slice (low and high have their
 // normal meaning) and for creating a new slice, where 'capacity' means the
 // biggest possible slice capacity, 'low' means len and 'high' means cap. The
 // logic is the same in both cases.
-func (c *Compiler) emitSliceBoundsCheck(frame *Frame, capacity, low, high, max llvm.Value, lowType, highType, maxType *types.Basic) {
-	if frame.fn.IsNoBounds() {
+func (b *builder) createSliceBoundsCheck(capacity, low, high, max llvm.Value, lowType, highType, maxType *types.Basic) {
+	if b.info.nobounds {
 		// The //go:nobounds pragma was added to the function to avoid bounds
 		// checking.
 		return
@@ -75,92 +66,164 @@ func (c *Compiler) emitSliceBoundsCheck(frame *Frame, capacity, low, high, max l
 		capacityType = max.Type()
 	}
 	if capacityType != capacity.Type() {
-		capacity = c.builder.CreateZExt(capacity, capacityType, "")
+		capacity = b.CreateZExt(capacity, capacityType, "")
 	}

 	// Extend low and high to be the same size as capacity.
 	if low.Type().IntTypeWidth() < capacityType.IntTypeWidth() {
 		if lowType.Info()&types.IsUnsigned != 0 {
-			low = c.builder.CreateZExt(low, capacityType, "")
+			low = b.CreateZExt(low, capacityType, "")
 		} else {
-			low = c.builder.CreateSExt(low, capacityType, "")
+			low = b.CreateSExt(low, capacityType, "")
 		}
 	}
 	if high.Type().IntTypeWidth() < capacityType.IntTypeWidth() {
 		if highType.Info()&types.IsUnsigned != 0 {
-			high = c.builder.CreateZExt(high, capacityType, "")
+			high = b.CreateZExt(high, capacityType, "")
 		} else {
-			high = c.builder.CreateSExt(high, capacityType, "")
+			high = b.CreateSExt(high, capacityType, "")
 		}
 	}
 	if max.Type().IntTypeWidth() < capacityType.IntTypeWidth() {
 		if maxType.Info()&types.IsUnsigned != 0 {
-			max = c.builder.CreateZExt(max, capacityType, "")
+			max = b.CreateZExt(max, capacityType, "")
 		} else {
-			max = c.builder.CreateSExt(max, capacityType, "")
+			max = b.CreateSExt(max, capacityType, "")
 		}
 	}

-	faultBlock := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "slice.outofbounds")
-	nextBlock := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "slice.next")
-	frame.blockExits[frame.currentBlock] = nextBlock // adjust outgoing block for phi nodes
-
 	// Now do the bounds check: low > high || high > capacity
-	outOfBounds1 := c.builder.CreateICmp(llvm.IntUGT, low, high, "slice.lowhigh")
-	outOfBounds2 := c.builder.CreateICmp(llvm.IntUGT, high, max, "slice.highmax")
-	outOfBounds3 := c.builder.CreateICmp(llvm.IntUGT, max, capacity, "slice.maxcap")
-	outOfBounds := c.builder.CreateOr(outOfBounds1, outOfBounds2, "slice.lowmax")
-	outOfBounds = c.builder.CreateOr(outOfBounds, outOfBounds3, "slice.lowcap")
-	c.builder.CreateCondBr(outOfBounds, faultBlock, nextBlock)
-
-	// Fail: this is a nil pointer, exit with a panic.
-	c.builder.SetInsertPointAtEnd(faultBlock)
-	c.createRuntimeCall("slicePanic", nil, "")
-	c.builder.CreateUnreachable()
-
-	// Ok: this is a valid pointer.
-	c.builder.SetInsertPointAtEnd(nextBlock)
+	outOfBounds1 := b.CreateICmp(llvm.IntUGT, low, high, "slice.lowhigh")
+	outOfBounds2 := b.CreateICmp(llvm.IntUGT, high, max, "slice.highmax")
+	outOfBounds3 := b.CreateICmp(llvm.IntUGT, max, capacity, "slice.maxcap")
+	outOfBounds := b.CreateOr(outOfBounds1, outOfBounds2, "slice.lowmax")
+	outOfBounds = b.CreateOr(outOfBounds, outOfBounds3, "slice.lowcap")
+	b.createRuntimeAssert(outOfBounds, "slice", "slicePanic")
 }

-// emitNilCheck checks whether the given pointer is nil, and panics if it is. It
-// has no effect in well-behaved programs, but makes sure no uncaught nil
+// createChanBoundsCheck creates a bounds check before creating a new channel to
+// check that the value is not too big for runtime.chanMake.
+func (b *builder) createChanBoundsCheck(elementSize uint64, bufSize llvm.Value, bufSizeType *types.Basic, pos token.Pos) {
+	if b.info.nobounds {
+		// The //go:nobounds pragma was added to the function to avoid bounds
+		// checking.
+		return
+	}
+
+	// Check whether the bufSize parameter must be cast to a wider integer for
+	// comparison.
+	if bufSize.Type().IntTypeWidth() < b.uintptrType.IntTypeWidth() {
+		if bufSizeType.Info()&types.IsUnsigned != 0 {
+			// Unsigned, so zero-extend to uint type.
+			bufSizeType = types.Typ[types.Uint]
+			bufSize = b.CreateZExt(bufSize, b.intType, "")
+		} else {
+			// Signed, so sign-extend to int type.
+			bufSizeType = types.Typ[types.Int]
+			bufSize = b.CreateSExt(bufSize, b.intType, "")
+		}
+	}
+
+	// Calculate (^uintptr(0)) >> 1, which is the max value that fits in an
+	// uintptr if uintptrs were signed.
+	maxBufSize := llvm.ConstLShr(llvm.ConstNot(llvm.ConstInt(b.uintptrType, 0, false)), llvm.ConstInt(b.uintptrType, 1, false))
+	if elementSize > maxBufSize.ZExtValue() {
+		b.addError(pos, fmt.Sprintf("channel element type is too big (%v bytes)", elementSize))
+		return
+	}
+	// Avoid divide-by-zero.
+	if elementSize == 0 {
+		elementSize = 1
+	}
+	// Make the maxBufSize actually the maximum allowed value (in number of
+	// elements in the channel buffer).
+	maxBufSize = llvm.ConstUDiv(maxBufSize, llvm.ConstInt(b.uintptrType, elementSize, false))
+
+	// Make sure maxBufSize has the same type as bufSize.
+	if maxBufSize.Type() != bufSize.Type() {
+		maxBufSize = llvm.ConstZExt(maxBufSize, bufSize.Type())
+	}
+
+	// Do the check for a too large (or negative) buffer size.
+	bufSizeTooBig := b.CreateICmp(llvm.IntUGE, bufSize, maxBufSize, "")
+	b.createRuntimeAssert(bufSizeTooBig, "chan", "chanMakePanic")
+}
+
+// createNilCheck checks whether the given pointer is nil, and panics if it is.
+// It has no effect in well-behaved programs, but makes sure no uncaught nil
 // pointer dereferences exist in valid Go code.
-func (c *Compiler) emitNilCheck(frame *Frame, ptr llvm.Value, blockPrefix string) {
+func (b *builder) createNilCheck(inst ssa.Value, ptr llvm.Value, blockPrefix string) {
 	// Check whether we need to emit this check at all.
 	if !ptr.IsAGlobalValue().IsNil() {
 		return
 	}

-	// Check whether this is a nil pointer.
-	faultBlock := c.ctx.AddBasicBlock(frame.fn.LLVMFn, blockPrefix+".nil")
-	nextBlock := c.ctx.AddBasicBlock(frame.fn.LLVMFn, blockPrefix+".next")
-	frame.blockExits[frame.currentBlock] = nextBlock // adjust outgoing block for phi nodes
+	switch inst := inst.(type) {
+	case *ssa.IndexAddr:
+		// This pointer is the result of an index operation into a slice or
+		// array. Such slices/arrays are already bounds checked so the pointer
+		// must be a valid (non-nil) pointer. No nil checking is necessary.
+		return
+	case *ssa.Convert:
+		// This is a pointer that comes from a conversion from unsafe.Pointer.
+		// Don't do nil checking because this is unsafe code and the code should
+		// know what it is doing.
+		// Note: all *ssa.Convert instructions that result in a pointer must
+		// come from unsafe.Pointer. Testing here for unsafe.Pointer to be sure.
+		if inst.X.Type() == types.Typ[types.UnsafePointer] {
+			return
+		}
+	}

 	// Compare against nil.
-	var isnil llvm.Value
-	if ptr.Type().PointerAddressSpace() == 0 {
-		// Do the nil check using the isnil builtin, which marks the parameter
-		// as nocapture.
-		// The reason it has to go through a builtin, is that a regular icmp
-		// instruction may capture the pointer in LLVM semantics, see
-		// https://reviews.llvm.org/D60047 for details. Pointer capturing
-		// unfortunately breaks escape analysis, so we use this trick to let the
-		// functionattr pass know that this pointer doesn't really escape.
-		ptr = c.builder.CreateBitCast(ptr, c.i8ptrType, "")
-		isnil = c.createRuntimeCall("isnil", []llvm.Value{ptr}, "")
-	} else {
-		// Do the nil check using a regular icmp. This can happen with function
-		// pointers on AVR, which don't benefit from escape analysis anyway.
-		nilptr := llvm.ConstPointerNull(ptr.Type())
-		isnil = c.builder.CreateICmp(llvm.IntEQ, ptr, nilptr, "")
-	}
-	c.builder.CreateCondBr(isnil, faultBlock, nextBlock)
+	// We previously used a hack to make sure this wouldn't break escape
+	// analysis, but this is not necessary anymore since
+	// https://reviews.llvm.org/D60047 has been merged.
+	nilptr := llvm.ConstPointerNull(ptr.Type())
+	isnil := b.CreateICmp(llvm.IntEQ, ptr, nilptr, "")

-	// Fail: this is a nil pointer, exit with a panic.
-	c.builder.SetInsertPointAtEnd(faultBlock)
-	c.createRuntimeCall("nilPanic", nil, "")
-	c.builder.CreateUnreachable()
-
-	// Ok: this is a valid pointer.
-	c.builder.SetInsertPointAtEnd(nextBlock)
+	// Emit the nil check in IR.
+	b.createRuntimeAssert(isnil, blockPrefix, "nilPanic")
+}
+
+// createNegativeShiftCheck creates an assertion that panics if the given shift value is negative.
+// This function assumes that the shift value is signed.
+func (b *builder) createNegativeShiftCheck(shift llvm.Value) {
+	if b.info.nobounds {
+		// Function disabled bounds checking - skip shift check.
+		return
+	}
+
+	// isNegative = shift < 0
+	isNegative := b.CreateICmp(llvm.IntSLT, shift, llvm.ConstInt(shift.Type(), 0, false), "")
+	b.createRuntimeAssert(isNegative, "shift", "negativeShiftPanic")
+}
+
+// createRuntimeAssert is a common function to create a new branch on an assert
+// bool, calling an assert func if the assert value is true (1).
+func (b *builder) createRuntimeAssert(assert llvm.Value, blockPrefix, assertFunc string) {
+	// Check whether we can resolve this check at compile time.
+	if !assert.IsAConstantInt().IsNil() {
+		val := assert.ZExtValue()
+		if val == 0 {
+			// Everything is constant so the check does not have to be emitted
+			// in IR. This avoids emitting some redundant IR.
+			return
+		}
+	}
+
+	faultBlock := b.ctx.AddBasicBlock(b.llvmFn, blockPrefix+".throw")
+	nextBlock := b.ctx.AddBasicBlock(b.llvmFn, blockPrefix+".next")
+	b.blockExits[b.currentBlock] = nextBlock // adjust outgoing block for phi nodes
+
+	// Now branch to the out-of-bounds or the regular block.
+	b.CreateCondBr(assert, faultBlock, nextBlock)
+
+	// Fail: the assert triggered so panic.
+	b.SetInsertPointAtEnd(faultBlock)
+	b.createRuntimeCall(assertFunc, nil, "")
+	b.CreateUnreachable()
+
+	// Ok: assert didn't trigger so continue normally.
+	b.SetInsertPointAtEnd(nextBlock)
 }
@@ -0,0 +1,84 @@
+package compiler
+
+import (
+	"strings"
+
+	"golang.org/x/tools/go/ssa"
+	"tinygo.org/x/go-llvm"
+)
+
+// createAtomicOp lowers an atomic library call by lowering it as an LLVM atomic
+// operation. It returns the result of the operation and true if the call could
+// be lowered inline, and false otherwise.
+func (b *builder) createAtomicOp(call *ssa.CallCommon) (llvm.Value, bool) {
+	name := call.Value.(*ssa.Function).Name()
+	switch name {
+	case "AddInt32", "AddInt64", "AddUint32", "AddUint64", "AddUintptr":
+		ptr := b.getValue(call.Args[0])
+		val := b.getValue(call.Args[1])
+		oldVal := b.CreateAtomicRMW(llvm.AtomicRMWBinOpAdd, ptr, val, llvm.AtomicOrderingSequentiallyConsistent, true)
+		// Return the new value, not the original value returned by atomicrmw.
+		return b.CreateAdd(oldVal, val, ""), true
+	case "SwapInt32", "SwapInt64", "SwapUint32", "SwapUint64", "SwapUintptr", "SwapPointer":
+		ptr := b.getValue(call.Args[0])
+		val := b.getValue(call.Args[1])
+		isPointer := val.Type().TypeKind() == llvm.PointerTypeKind
+		if isPointer {
+			// atomicrmw only supports integers, so cast to an integer.
+			val = b.CreatePtrToInt(val, b.uintptrType, "")
+			ptr = b.CreateBitCast(ptr, llvm.PointerType(val.Type(), 0), "")
+		}
+		oldVal := b.CreateAtomicRMW(llvm.AtomicRMWBinOpXchg, ptr, val, llvm.AtomicOrderingSequentiallyConsistent, true)
+		if isPointer {
+			oldVal = b.CreateIntToPtr(oldVal, b.i8ptrType, "")
+		}
+		return oldVal, true
+	case "CompareAndSwapInt32", "CompareAndSwapInt64", "CompareAndSwapUint32", "CompareAndSwapUint64", "CompareAndSwapUintptr", "CompareAndSwapPointer":
+		ptr := b.getValue(call.Args[0])
+		old := b.getValue(call.Args[1])
+		newVal := b.getValue(call.Args[2])
+		if strings.HasSuffix(name, "64") {
+			arch := strings.Split(b.Triple, "-")[0]
+			if strings.HasPrefix(arch, "arm") && strings.HasSuffix(arch, "m") {
+				// Work around a bug in LLVM, at least LLVM 11:
+				// https://reviews.llvm.org/D95891
+				// Check for armv6m, armv7, armv7em, and perhaps others.
+				// See also: https://gcc.gnu.org/onlinedocs/gcc/_005f_005fsync-Builtins.html
+				compareAndSwap := b.mod.NamedFunction("__sync_val_compare_and_swap_8")
+				if compareAndSwap.IsNil() {
+					// Declare the function if it isn't already declared.
+					i64Type := b.ctx.Int64Type()
+					fnType := llvm.FunctionType(i64Type, []llvm.Type{llvm.PointerType(i64Type, 0), i64Type, i64Type}, false)
+					compareAndSwap = llvm.AddFunction(b.mod, "__sync_val_compare_and_swap_8", fnType)
+				}
+				actualOldValue := b.CreateCall(compareAndSwap, []llvm.Value{ptr, old, newVal}, "")
+				// The __sync_val_compare_and_swap_8 function returns the old
+				// value. However, we shouldn't return the old value, we should
+				// return whether the compare/exchange was successful. This is
+				// easily done by comparing the returned (actual) old value with
+				// the expected old value passed to
+				// __sync_val_compare_and_swap_8.
+				swapped := b.CreateICmp(llvm.IntEQ, old, actualOldValue, "")
+				return swapped, true
+			}
+		}
+		tuple := b.CreateAtomicCmpXchg(ptr, old, newVal, llvm.AtomicOrderingSequentiallyConsistent, llvm.AtomicOrderingSequentiallyConsistent, true)
+		swapped := b.CreateExtractValue(tuple, 1, "")
+		return swapped, true
+	case "LoadInt32", "LoadInt64", "LoadUint32", "LoadUint64", "LoadUintptr", "LoadPointer":
+		ptr := b.getValue(call.Args[0])
+		val := b.CreateLoad(ptr, "")
+		val.SetOrdering(llvm.AtomicOrderingSequentiallyConsistent)
+		val.SetAlignment(b.targetData.PrefTypeAlignment(val.Type())) // required
+		return val, true
+	case "StoreInt32", "StoreInt64", "StoreUint32", "StoreUint64", "StoreUintptr", "StorePointer":
+		ptr := b.getValue(call.Args[0])
+		val := b.getValue(call.Args[1])
+		store := b.CreateStore(val, ptr)
+		store.SetOrdering(llvm.AtomicOrderingSequentiallyConsistent)
+		store.SetAlignment(b.targetData.PrefTypeAlignment(val.Type())) // required
+		return store, true
+	default:
+		return llvm.Value{}, false
+	}
+}
@@ -1,7 +1,9 @@
 package compiler

 import (
-	"fmt"
+	"go/types"
+	"strconv"
+
 	"golang.org/x/tools/go/ssa"
 	"tinygo.org/x/go-llvm"
 )
@@ -11,62 +13,80 @@ import (

 // The maximum number of arguments that can be expanded from a single struct. If
 // a struct contains more fields, it is passed as a struct without expanding.
-const MaxFieldsPerParam = 3
+const maxFieldsPerParam = 3

-// Shortcut: create a call to runtime.<fnName> with the given arguments.
-func (c *Compiler) createRuntimeCall(fnName string, args []llvm.Value, name string) llvm.Value {
-	runtimePkg := c.ir.Program.ImportedPackage("runtime")
-	member := runtimePkg.Members[fnName]
-	if member == nil {
-		panic("trying to call runtime." + fnName)
-	}
-	fn := c.ir.GetFunction(member.(*ssa.Function))
-	if fn.LLVMFn.IsNil() {
-		panic(fmt.Errorf("function %s does not appear in LLVM IR", fnName))
-	}
-	if !fn.IsExported() {
-		args = append(args, llvm.Undef(c.i8ptrType))            // unused context parameter
-		args = append(args, llvm.ConstPointerNull(c.i8ptrType)) // coroutine handle
-	}
-	return c.createCall(fn.LLVMFn, args, name)
+// paramInfo contains some information collected about a function parameter,
+// useful while declaring or defining a function.
+type paramInfo struct {
+	llvmType llvm.Type
+	name     string // name, possibly with suffixes for e.g. struct fields
+	flags    paramFlags
 }

-// Create a call to the given function with the arguments possibly expanded.
-func (c *Compiler) createCall(fn llvm.Value, args []llvm.Value, name string) llvm.Value {
+// paramFlags identifies parameter attributes for flags. Most importantly, it
+// determines which parameters are dereferenceable_or_null and which aren't.
+type paramFlags uint8
+
+const (
+	// Parameter may have the deferenceable_or_null attribute. This attribute
+	// cannot be applied to unsafe.Pointer and to the data pointer of slices.
+	paramIsDeferenceableOrNull = 1 << iota
+)
+
+// createCall creates a new call to runtime.<fnName> with the given arguments.
+func (b *builder) createRuntimeCall(fnName string, args []llvm.Value, name string) llvm.Value {
+	fn := b.program.ImportedPackage("runtime").Members[fnName].(*ssa.Function)
+	llvmFn := b.getFunction(fn)
+	if llvmFn.IsNil() {
+		panic("trying to call non-existent function: " + fn.RelString(nil))
+	}
+	args = append(args, llvm.Undef(b.i8ptrType))            // unused context parameter
+	args = append(args, llvm.ConstPointerNull(b.i8ptrType)) // coroutine handle
+	return b.createCall(llvmFn, args, name)
+}
+
+// createCall creates a call to the given function with the arguments possibly
+// expanded.
+func (b *builder) createCall(fn llvm.Value, args []llvm.Value, name string) llvm.Value {
 	expanded := make([]llvm.Value, 0, len(args))
 	for _, arg := range args {
-		fragments := c.expandFormalParam(arg)
+		fragments := b.expandFormalParam(arg)
 		expanded = append(expanded, fragments...)
 	}
-	return c.builder.CreateCall(fn, expanded, name)
+	return b.CreateCall(fn, expanded, name)
 }

 // Expand an argument type to a list that can be used in a function call
 // parameter list.
-func (c *Compiler) expandFormalParamType(t llvm.Type) []llvm.Type {
+func expandFormalParamType(t llvm.Type, name string, goType types.Type) []paramInfo {
 	switch t.TypeKind() {
 	case llvm.StructTypeKind:
-		fields := c.flattenAggregateType(t)
-		if len(fields) <= MaxFieldsPerParam {
-			return fields
+		fieldInfos := flattenAggregateType(t, name, goType)
+		if len(fieldInfos) <= maxFieldsPerParam {
+			return fieldInfos
 		} else {
 			// failed to lower
-			return []llvm.Type{t}
 		}
-	default:
-		// TODO: split small arrays
-		return []llvm.Type{t}
+	}
+	// TODO: split small arrays
+	return []paramInfo{
+		{
+			llvmType: t,
+			name:     name,
+			flags:    getTypeFlags(goType),
+		},
 	}
 }

-// Expand an argument type to a list of offsets from the start of the object.
-// Used together with expandFormalParam to get the offset of each value from the
-// start of the non-expanded value.
-func (c *Compiler) expandFormalParamOffsets(t llvm.Type) []uint64 {
+// expandFormalParamOffsets returns a list of offsets from the start of an
+// object of type t after it would have been split up by expandFormalParam. This
+// is useful for debug information, where it is necessary to know the offset
+// from the start of the combined object.
+func (b *builder) expandFormalParamOffsets(t llvm.Type) []uint64 {
 	switch t.TypeKind() {
 	case llvm.StructTypeKind:
-		fields := c.flattenAggregateTypeOffsets(t)
-		if len(fields) <= MaxFieldsPerParam {
+		fields := b.flattenAggregateTypeOffsets(t)
+		if len(fields) <= maxFieldsPerParam {
 			return fields
 		} else {
 			// failed to lower
@@ -78,14 +98,17 @@ func (c *Compiler) expandFormalParamOffsets(t llvm.Type) []uint64 {
 	}
 }

-// Equivalent of expandFormalParamType for parameter values.
-func (c *Compiler) expandFormalParam(v llvm.Value) []llvm.Value {
+// expandFormalParam splits a formal param value into pieces, so it can be
+// passed directly as part of a function call. For example, it splits up small
+// structs into individual fields. It is the equivalent of expandFormalParamType
+// for parameter values.
+func (b *builder) expandFormalParam(v llvm.Value) []llvm.Value {
 	switch v.Type().TypeKind() {
 	case llvm.StructTypeKind:
-		fieldTypes := c.flattenAggregateType(v.Type())
-		if len(fieldTypes) <= MaxFieldsPerParam {
-			fields := c.flattenAggregate(v)
-			if len(fields) != len(fieldTypes) {
+		fieldInfos := flattenAggregateType(v.Type(), "", nil)
+		if len(fieldInfos) <= maxFieldsPerParam {
+			fields := b.flattenAggregate(v)
+			if len(fields) != len(fieldInfos) {
 				panic("type and value param lowering don't match")
 			}
 			return fields
@@ -101,24 +124,98 @@ func (c *Compiler) expandFormalParam(v llvm.Value) []llvm.Value {

 // Try to flatten a struct type to a list of types. Returns a 1-element slice
 // with the passed in type if this is not possible.
-func (c *Compiler) flattenAggregateType(t llvm.Type) []llvm.Type {
+func flattenAggregateType(t llvm.Type, name string, goType types.Type) []paramInfo {
+	typeFlags := getTypeFlags(goType)
 	switch t.TypeKind() {
 	case llvm.StructTypeKind:
-		fields := make([]llvm.Type, 0, t.StructElementTypesCount())
-		for _, subfield := range t.StructElementTypes() {
-			subfields := c.flattenAggregateType(subfield)
-			fields = append(fields, subfields...)
+		paramInfos := make([]paramInfo, 0, t.StructElementTypesCount())
+		for i, subfield := range t.StructElementTypes() {
+			suffix := strconv.Itoa(i)
+			if goType != nil {
+				// Try to come up with a good suffix for this struct field,
+				// depending on which Go type it's based on.
+				switch goType := goType.Underlying().(type) {
+				case *types.Interface:
+					suffix = []string{"typecode", "value"}[i]
+				case *types.Slice:
+					suffix = []string{"data", "len", "cap"}[i]
+				case *types.Struct:
+					suffix = goType.Field(i).Name()
+				case *types.Basic:
+					switch goType.Kind() {
+					case types.Complex64, types.Complex128:
+						suffix = []string{"r", "i"}[i]
+					case types.String:
+						suffix = []string{"data", "len"}[i]
+					}
+				case *types.Signature:
+					suffix = []string{"context", "funcptr"}[i]
+				}
+			}
+			subInfos := flattenAggregateType(subfield, name+"."+suffix, extractSubfield(goType, i))
+			for i := range subInfos {
+				subInfos[i].flags |= typeFlags
+			}
+			paramInfos = append(paramInfos, subInfos...)
 		}
-		return fields
+		return paramInfos
 	default:
-		return []llvm.Type{t}
+		return []paramInfo{
+			{
+				llvmType: t,
+				name:     name,
+				flags:    typeFlags,
+			},
+		}
 	}
 }

-// Return the offsets from the start of the object if this object type were
-// flattened like in flattenAggregate. Used together with flattenAggregate to
-// know the start indices of each value in the non-flattened object.
-func (c *Compiler) flattenAggregateTypeOffsets(t llvm.Type) []uint64 {
+// getTypeFlags returns the type flags for a given type. It will not recurse
+// into sub-types (such as in structs).
+func getTypeFlags(t types.Type) paramFlags {
+	if t == nil {
+		return 0
+	}
+	switch t.Underlying().(type) {
+	case *types.Pointer:
+		// Pointers in Go must either point to an object or be nil.
+		return paramIsDeferenceableOrNull
+	case *types.Chan, *types.Map:
+		// Channels and maps are implemented as pointers pointing to some
+		// object, and follow the same rules as *types.Pointer.
+		return paramIsDeferenceableOrNull
+	default:
+		return 0
+	}
+}
+
+// extractSubfield extracts a field from a struct, or returns null if this is
+// not a struct and thus no subfield can be obtained.
+func extractSubfield(t types.Type, field int) types.Type {
+	if t == nil {
+		return nil
+	}
+	switch t := t.Underlying().(type) {
+	case *types.Struct:
+		return t.Field(field).Type()
+	case *types.Interface, *types.Slice, *types.Basic, *types.Signature:
+		// These Go types are (sometimes) implemented as LLVM structs but can't
+		// really be split further up in Go (with the possible exception of
+		// complex numbers).
+		return nil
+	default:
+		// This should be unreachable.
+		panic("cannot split subfield: " + t.String())
+	}
+}
+
+// flattenAggregateTypeOffset returns the offsets from the start of an object of
+// type t if this object were flattened like in flattenAggregate. Used together
+// with flattenAggregate to know the start indices of each value in the
+// non-flattened object.
+//
+// Note: this is an implementation detail, use expandFormalParamOffsets instead.
+func (c *compilerContext) flattenAggregateTypeOffsets(t llvm.Type) []uint64 {
 	switch t.TypeKind() {
 	case llvm.StructTypeKind:
 		fields := make([]uint64, 0, t.StructElementTypesCount())
@@ -136,15 +233,15 @@ func (c *Compiler) flattenAggregateTypeOffsets(t llvm.Type) []uint64 {
 	}
 }

-// Break down a struct into its elementary types for argument passing. The value
-// equivalent of flattenAggregateType
-func (c *Compiler) flattenAggregate(v llvm.Value) []llvm.Value {
+// flattenAggregate breaks down a struct into its elementary values for argument
+// passing. It is the value equivalent of flattenAggregateType
+func (b *builder) flattenAggregate(v llvm.Value) []llvm.Value {
 	switch v.Type().TypeKind() {
 	case llvm.StructTypeKind:
 		fields := make([]llvm.Value, 0, v.Type().StructElementTypesCount())
 		for i := range v.Type().StructElementTypes() {
-			subfield := c.builder.CreateExtractValue(v, i, "")
-			subfields := c.flattenAggregate(subfield)
+			subfield := b.CreateExtractValue(v, i, "")
+			subfields := b.flattenAggregate(subfield)
 			fields = append(fields, subfields...)
 		}
 		return fields
@@ -153,25 +250,29 @@ func (c *Compiler) flattenAggregate(v llvm.Value) []llvm.Value {
 	}
 }

-// Collapse a list of fields into its original value.
-func (c *Compiler) collapseFormalParam(t llvm.Type, fields []llvm.Value) llvm.Value {
-	param, remaining := c.collapseFormalParamInternal(t, fields)
+// collapseFormalParam combines an aggregate object back into the original
+// value. This is used to join multiple LLVM parameters into a single Go value
+// in the function entry block.
+func (b *builder) collapseFormalParam(t llvm.Type, fields []llvm.Value) llvm.Value {
+	param, remaining := b.collapseFormalParamInternal(t, fields)
 	if len(remaining) != 0 {
 		panic("failed to expand back all fields")
 	}
 	return param
 }

-// Returns (value, remainingFields). Used by collapseFormalParam.
-func (c *Compiler) collapseFormalParamInternal(t llvm.Type, fields []llvm.Value) (llvm.Value, []llvm.Value) {
+// collapseFormalParamInternal is an implementation detail of
+// collapseFormalParam: it works by recursing until there are no fields left.
+func (b *builder) collapseFormalParamInternal(t llvm.Type, fields []llvm.Value) (llvm.Value, []llvm.Value) {
 	switch t.TypeKind() {
 	case llvm.StructTypeKind:
-		if len(c.flattenAggregateType(t)) <= MaxFieldsPerParam {
+		flattened := flattenAggregateType(t, "", nil)
+		if len(flattened) <= maxFieldsPerParam {
 			value := llvm.ConstNull(t)
 			for i, subtyp := range t.StructElementTypes() {
-				structField, remaining := c.collapseFormalParamInternal(subtyp, fields)
+				structField, remaining := b.collapseFormalParamInternal(subtyp, fields)
 				fields = remaining
-				value = c.builder.CreateInsertValue(value, structField, i, "")
+				value = b.CreateInsertValue(value, structField, i, "")
 			}
 			return value, fields
 		} else {
@@ -11,73 +11,88 @@ import (
 	"tinygo.org/x/go-llvm"
 )

-func (c *Compiler) emitMakeChan(frame *Frame, expr *ssa.MakeChan) llvm.Value {
-	elementSize := c.targetData.TypeAllocSize(c.getLLVMType(expr.Type().(*types.Chan).Elem()))
-	elementSizeValue := llvm.ConstInt(c.uintptrType, elementSize, false)
-	bufSize := c.getValue(frame, expr.Size)
-	return c.createRuntimeCall("chanMake", []llvm.Value{elementSizeValue, bufSize}, "")
+func (b *builder) createMakeChan(expr *ssa.MakeChan) llvm.Value {
+	elementSize := b.targetData.TypeAllocSize(b.getLLVMType(expr.Type().Underlying().(*types.Chan).Elem()))
+	elementSizeValue := llvm.ConstInt(b.uintptrType, elementSize, false)
+	bufSize := b.getValue(expr.Size)
+	b.createChanBoundsCheck(elementSize, bufSize, expr.Size.Type().Underlying().(*types.Basic), expr.Pos())
+	if bufSize.Type().IntTypeWidth() < b.uintptrType.IntTypeWidth() {
+		bufSize = b.CreateZExt(bufSize, b.uintptrType, "")
+	} else if bufSize.Type().IntTypeWidth() > b.uintptrType.IntTypeWidth() {
+		bufSize = b.CreateTrunc(bufSize, b.uintptrType, "")
+	}
+	return b.createRuntimeCall("chanMake", []llvm.Value{elementSizeValue, bufSize}, "")
 }

-// emitChanSend emits a pseudo chan send operation. It is lowered to the actual
-// channel send operation during goroutine lowering.
-func (c *Compiler) emitChanSend(frame *Frame, instr *ssa.Send) {
-	ch := c.getValue(frame, instr.Chan)
-	chanValue := c.getValue(frame, instr.X)
+// createChanSend emits a pseudo chan send operation. It is lowered to the
+// actual channel send operation during goroutine lowering.
+func (b *builder) createChanSend(instr *ssa.Send) {
+	ch := b.getValue(instr.Chan)
+	chanValue := b.getValue(instr.X)

 	// store value-to-send
-	valueType := c.getLLVMType(instr.X.Type())
-	valueAlloca, valueAllocaCast, valueAllocaSize := c.createTemporaryAlloca(valueType, "chan.value")
-	c.builder.CreateStore(chanValue, valueAlloca)
+	valueType := b.getLLVMType(instr.X.Type())
+	valueAlloca, valueAllocaCast, valueAllocaSize := b.createTemporaryAlloca(valueType, "chan.value")
+	b.CreateStore(chanValue, valueAlloca)
+
+	// Allocate blockedlist buffer.
+	channelBlockedList := b.mod.GetTypeByName("runtime.channelBlockedList")
+	channelBlockedListAlloca, channelBlockedListAllocaCast, channelBlockedListAllocaSize := b.createTemporaryAlloca(channelBlockedList, "chan.blockedList")

 	// Do the send.
-	c.createRuntimeCall("chanSend", []llvm.Value{ch, valueAllocaCast}, "")
+	b.createRuntimeCall("chanSend", []llvm.Value{ch, valueAllocaCast, channelBlockedListAlloca}, "")

-	// End the lifetime of the alloca.
+	// End the lifetime of the allocas.
 	// This also works around a bug in CoroSplit, at least in LLVM 8:
 	// https://bugs.llvm.org/show_bug.cgi?id=41742
-	c.emitLifetimeEnd(valueAllocaCast, valueAllocaSize)
+	b.emitLifetimeEnd(channelBlockedListAllocaCast, channelBlockedListAllocaSize)
+	b.emitLifetimeEnd(valueAllocaCast, valueAllocaSize)
 }

-// emitChanRecv emits a pseudo chan receive operation. It is lowered to the
+// createChanRecv emits a pseudo chan receive operation. It is lowered to the
 // actual channel receive operation during goroutine lowering.
-func (c *Compiler) emitChanRecv(frame *Frame, unop *ssa.UnOp) llvm.Value {
-	valueType := c.getLLVMType(unop.X.Type().(*types.Chan).Elem())
-	ch := c.getValue(frame, unop.X)
+func (b *builder) createChanRecv(unop *ssa.UnOp) llvm.Value {
+	valueType := b.getLLVMType(unop.X.Type().Underlying().(*types.Chan).Elem())
+	ch := b.getValue(unop.X)

 	// Allocate memory to receive into.
-	valueAlloca, valueAllocaCast, valueAllocaSize := c.createTemporaryAlloca(valueType, "chan.value")
+	valueAlloca, valueAllocaCast, valueAllocaSize := b.createTemporaryAlloca(valueType, "chan.value")
+
+	// Allocate blockedlist buffer.
+	channelBlockedList := b.mod.GetTypeByName("runtime.channelBlockedList")
+	channelBlockedListAlloca, channelBlockedListAllocaCast, channelBlockedListAllocaSize := b.createTemporaryAlloca(channelBlockedList, "chan.blockedList")

 	// Do the receive.
-	commaOk := c.createRuntimeCall("chanRecv", []llvm.Value{ch, valueAllocaCast}, "")
-	received := c.builder.CreateLoad(valueAlloca, "chan.received")
-	c.emitLifetimeEnd(valueAllocaCast, valueAllocaSize)
+	commaOk := b.createRuntimeCall("chanRecv", []llvm.Value{ch, valueAllocaCast, channelBlockedListAlloca}, "")
+	received := b.CreateLoad(valueAlloca, "chan.received")
+	b.emitLifetimeEnd(channelBlockedListAllocaCast, channelBlockedListAllocaSize)
+	b.emitLifetimeEnd(valueAllocaCast, valueAllocaSize)

 	if unop.CommaOk {
-		tuple := llvm.Undef(c.ctx.StructType([]llvm.Type{valueType, c.ctx.Int1Type()}, false))
-		tuple = c.builder.CreateInsertValue(tuple, received, 0, "")
-		tuple = c.builder.CreateInsertValue(tuple, commaOk, 1, "")
+		tuple := llvm.Undef(b.ctx.StructType([]llvm.Type{valueType, b.ctx.Int1Type()}, false))
+		tuple = b.CreateInsertValue(tuple, received, 0, "")
+		tuple = b.CreateInsertValue(tuple, commaOk, 1, "")
 		return tuple
 	} else {
 		return received
 	}
 }

-// emitChanClose closes the given channel.
-func (c *Compiler) emitChanClose(frame *Frame, param ssa.Value) {
-	ch := c.getValue(frame, param)
-	c.createRuntimeCall("chanClose", []llvm.Value{ch}, "")
+// createChanClose closes the given channel.
+func (b *builder) createChanClose(ch llvm.Value) {
+	b.createRuntimeCall("chanClose", []llvm.Value{ch}, "")
 }

-// emitSelect emits all IR necessary for a select statements. That's a
+// createSelect emits all IR necessary for a select statements. That's a
 // non-trivial amount of code because select is very complex to implement.
-func (c *Compiler) emitSelect(frame *Frame, expr *ssa.Select) llvm.Value {
+func (b *builder) createSelect(expr *ssa.Select) llvm.Value {
 	if len(expr.States) == 0 {
 		// Shortcuts for some simple selects.
-		llvmType := c.getLLVMType(expr.Type())
+		llvmType := b.getLLVMType(expr.Type())
 		if expr.Blocking {
 			// Blocks forever:
 			//     select {}
-			c.createRuntimeCall("deadlock", nil, "")
+			b.createRuntimeCall("deadlock", nil, "")
 			return llvm.Undef(llvmType)
 		} else {
 			// No-op:
@@ -85,7 +100,7 @@ func (c *Compiler) emitSelect(frame *Frame, expr *ssa.Select) llvm.Value {
 			//     default:
 			//     }
 			retval := llvm.Undef(llvmType)
-			retval = c.builder.CreateInsertValue(retval, llvm.ConstInt(c.intType, 0xffffffffffffffff, true), 0, "")
+			retval = b.CreateInsertValue(retval, llvm.ConstInt(b.intType, 0xffffffffffffffff, true), 0, "")
 			return retval // {-1, false}
 		}
 	}
@@ -103,30 +118,30 @@ func (c *Compiler) emitSelect(frame *Frame, expr *ssa.Select) llvm.Value {
 	recvbufAlign := 0
 	hasReceives := false
 	var selectStates []llvm.Value
-	chanSelectStateType := c.getLLVMRuntimeType("chanSelectState")
+	chanSelectStateType := b.getLLVMRuntimeType("chanSelectState")
 	for _, state := range expr.States {
-		ch := c.getValue(frame, state.Chan)
+		ch := b.getValue(state.Chan)
 		selectState := llvm.ConstNull(chanSelectStateType)
-		selectState = c.builder.CreateInsertValue(selectState, ch, 0, "")
+		selectState = b.CreateInsertValue(selectState, ch, 0, "")
 		switch state.Dir {
 		case types.RecvOnly:
 			// Make sure the receive buffer is big enough and has the correct alignment.
-			llvmType := c.getLLVMType(state.Chan.Type().(*types.Chan).Elem())
-			if size := c.targetData.TypeAllocSize(llvmType); size > recvbufSize {
+			llvmType := b.getLLVMType(state.Chan.Type().Underlying().(*types.Chan).Elem())
+			if size := b.targetData.TypeAllocSize(llvmType); size > recvbufSize {
 				recvbufSize = size
 			}
-			if align := c.targetData.ABITypeAlignment(llvmType); align > recvbufAlign {
+			if align := b.targetData.ABITypeAlignment(llvmType); align > recvbufAlign {
 				recvbufAlign = align
 			}
 			hasReceives = true
 		case types.SendOnly:
 			// Store this value in an alloca and put a pointer to this alloca
 			// in the send state.
-			sendValue := c.getValue(frame, state.Send)
-			alloca := llvmutil.CreateEntryBlockAlloca(c.builder, sendValue.Type(), "select.send.value")
-			c.builder.CreateStore(sendValue, alloca)
-			ptr := c.builder.CreateBitCast(alloca, c.i8ptrType, "")
-			selectState = c.builder.CreateInsertValue(selectState, ptr, 1, "")
+			sendValue := b.getValue(state.Send)
+			alloca := llvmutil.CreateEntryBlockAlloca(b.Builder, sendValue.Type(), "select.send.value")
+			b.CreateStore(sendValue, alloca)
+			ptr := b.CreateBitCast(alloca, b.i8ptrType, "")
+			selectState = b.CreateInsertValue(selectState, ptr, 1, "")
 		default:
 			panic("unreachable")
 		}
@@ -134,74 +149,74 @@ func (c *Compiler) emitSelect(frame *Frame, expr *ssa.Select) llvm.Value {
 	}

 	// Create a receive buffer, where the received value will be stored.
-	recvbuf := llvm.Undef(c.i8ptrType)
+	recvbuf := llvm.Undef(b.i8ptrType)
 	if hasReceives {
-		allocaType := llvm.ArrayType(c.ctx.Int8Type(), int(recvbufSize))
-		recvbufAlloca, _, _ := c.createTemporaryAlloca(allocaType, "select.recvbuf.alloca")
+		allocaType := llvm.ArrayType(b.ctx.Int8Type(), int(recvbufSize))
+		recvbufAlloca, _, _ := b.createTemporaryAlloca(allocaType, "select.recvbuf.alloca")
 		recvbufAlloca.SetAlignment(recvbufAlign)
-		recvbuf = c.builder.CreateGEP(recvbufAlloca, []llvm.Value{
-			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
-			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+		recvbuf = b.CreateGEP(recvbufAlloca, []llvm.Value{
+			llvm.ConstInt(b.ctx.Int32Type(), 0, false),
+			llvm.ConstInt(b.ctx.Int32Type(), 0, false),
 		}, "select.recvbuf")
 	}

 	// Create the states slice (allocated on the stack).
 	statesAllocaType := llvm.ArrayType(chanSelectStateType, len(selectStates))
-	statesAlloca, statesI8, statesSize := c.createTemporaryAlloca(statesAllocaType, "select.states.alloca")
+	statesAlloca, statesI8, statesSize := b.createTemporaryAlloca(statesAllocaType, "select.states.alloca")
 	for i, state := range selectStates {
 		// Set each slice element to the appropriate channel.
-		gep := c.builder.CreateGEP(statesAlloca, []llvm.Value{
-			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
-			llvm.ConstInt(c.ctx.Int32Type(), uint64(i), false),
+		gep := b.CreateGEP(statesAlloca, []llvm.Value{
+			llvm.ConstInt(b.ctx.Int32Type(), 0, false),
+			llvm.ConstInt(b.ctx.Int32Type(), uint64(i), false),
 		}, "")
-		c.builder.CreateStore(state, gep)
+		b.CreateStore(state, gep)
 	}
-	statesPtr := c.builder.CreateGEP(statesAlloca, []llvm.Value{
-		llvm.ConstInt(c.ctx.Int32Type(), 0, false),
-		llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+	statesPtr := b.CreateGEP(statesAlloca, []llvm.Value{
+		llvm.ConstInt(b.ctx.Int32Type(), 0, false),
+		llvm.ConstInt(b.ctx.Int32Type(), 0, false),
 	}, "select.states")
-	statesLen := llvm.ConstInt(c.uintptrType, uint64(len(selectStates)), false)
+	statesLen := llvm.ConstInt(b.uintptrType, uint64(len(selectStates)), false)

 	// Do the select in the runtime.
 	var results llvm.Value
 	if expr.Blocking {
 		// Stack-allocate operation structures.
 		// If these were simply created as a slice, they would heap-allocate.
-		chBlockAllocaType := llvm.ArrayType(c.getLLVMRuntimeType("channelBlockedList"), len(selectStates))
-		chBlockAlloca, chBlockAllocaPtr, chBlockSize := c.createTemporaryAlloca(chBlockAllocaType, "select.block.alloca")
-		chBlockLen := llvm.ConstInt(c.uintptrType, uint64(len(selectStates)), false)
-		chBlockPtr := c.builder.CreateGEP(chBlockAlloca, []llvm.Value{
-			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
-			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+		chBlockAllocaType := llvm.ArrayType(b.getLLVMRuntimeType("channelBlockedList"), len(selectStates))
+		chBlockAlloca, chBlockAllocaPtr, chBlockSize := b.createTemporaryAlloca(chBlockAllocaType, "select.block.alloca")
+		chBlockLen := llvm.ConstInt(b.uintptrType, uint64(len(selectStates)), false)
+		chBlockPtr := b.CreateGEP(chBlockAlloca, []llvm.Value{
+			llvm.ConstInt(b.ctx.Int32Type(), 0, false),
+			llvm.ConstInt(b.ctx.Int32Type(), 0, false),
 		}, "select.block")

-		results = c.createRuntimeCall("chanSelect", []llvm.Value{
+		results = b.createRuntimeCall("chanSelect", []llvm.Value{
 			recvbuf,
 			statesPtr, statesLen, statesLen, // []chanSelectState
 			chBlockPtr, chBlockLen, chBlockLen, // []channelBlockList
 		}, "select.result")

 		// Terminate the lifetime of the operation structures.
-		c.emitLifetimeEnd(chBlockAllocaPtr, chBlockSize)
+		b.emitLifetimeEnd(chBlockAllocaPtr, chBlockSize)
 	} else {
-		results = c.createRuntimeCall("tryChanSelect", []llvm.Value{
+		results = b.createRuntimeCall("tryChanSelect", []llvm.Value{
 			recvbuf,
 			statesPtr, statesLen, statesLen, // []chanSelectState
 		}, "select.result")
 	}

 	// Terminate the lifetime of the states alloca.
-	c.emitLifetimeEnd(statesI8, statesSize)
+	b.emitLifetimeEnd(statesI8, statesSize)

 	// The result value does not include all the possible received values,
 	// because we can't load them in advance. Instead, the *ssa.Extract
 	// instruction will treat a *ssa.Select specially and load it there inline.
 	// Store the receive alloca in a sidetable until we hit this extract
 	// instruction.
-	if frame.selectRecvBuf == nil {
-		frame.selectRecvBuf = make(map[*ssa.Select]llvm.Value)
+	if b.selectRecvBuf == nil {
+		b.selectRecvBuf = make(map[*ssa.Select]llvm.Value)
 	}
-	frame.selectRecvBuf[expr] = recvbuf
+	b.selectRecvBuf[expr] = recvbuf

 	return results
 }
@@ -210,28 +225,28 @@ func (c *Compiler) emitSelect(frame *Frame, expr *ssa.Select) llvm.Value {
 // when extracting a value from a select statement (*ssa.Select). Because
 // *ssa.Select cannot load all values in advance, it does this later in the
 // *ssa.Extract expression.
-func (c *Compiler) getChanSelectResult(frame *Frame, expr *ssa.Extract) llvm.Value {
+func (b *builder) getChanSelectResult(expr *ssa.Extract) llvm.Value {
 	if expr.Index == 0 {
 		// index
-		value := c.getValue(frame, expr.Tuple)
-		index := c.builder.CreateExtractValue(value, expr.Index, "")
-		if index.Type().IntTypeWidth() < c.intType.IntTypeWidth() {
-			index = c.builder.CreateSExt(index, c.intType, "")
+		value := b.getValue(expr.Tuple)
+		index := b.CreateExtractValue(value, expr.Index, "")
+		if index.Type().IntTypeWidth() < b.intType.IntTypeWidth() {
+			index = b.CreateSExt(index, b.intType, "")
 		}
 		return index
 	} else if expr.Index == 1 {
 		// comma-ok
-		value := c.getValue(frame, expr.Tuple)
-		return c.builder.CreateExtractValue(value, expr.Index, "")
+		value := b.getValue(expr.Tuple)
+		return b.CreateExtractValue(value, expr.Index, "")
 	} else {
 		// Select statements are (index, ok, ...) where ... is a number of
 		// received values, depending on how many receive statements there
 		// are. They are all combined into one alloca (because only one
 		// receive can proceed at a time) so we'll get that alloca, bitcast
 		// it to the correct type, and dereference it.
-		recvbuf := frame.selectRecvBuf[expr.Tuple.(*ssa.Select)]
-		typ := llvm.PointerType(c.getLLVMType(expr.Type()), 0)
-		ptr := c.builder.CreateBitCast(recvbuf, typ, "")
-		return c.builder.CreateLoad(ptr, "")
+		recvbuf := b.selectRecvBuf[expr.Tuple.(*ssa.Select)]
+		typ := llvm.PointerType(b.getLLVMType(expr.Type()), 0)
+		ptr := b.CreateBitCast(recvbuf, typ, "")
+		return b.CreateLoad(ptr, "")
 	}
 }
@@ -0,0 +1,172 @@
+package compiler
+
+import (
+	"flag"
+	"go/types"
+	"io/ioutil"
+	"regexp"
+	"strconv"
+	"strings"
+	"testing"
+
+	"github.com/tinygo-org/tinygo/compileopts"
+	"github.com/tinygo-org/tinygo/loader"
+	"tinygo.org/x/go-llvm"
+)
+
+// Pass -update to go test to update the output of the test files.
+var flagUpdate = flag.Bool("update", false, "update tests based on test output")
+
+// Basic tests for the compiler. Build some Go files and compare the output with
+// the expected LLVM IR for regression testing.
+func TestCompiler(t *testing.T) {
+	target, err := compileopts.LoadTarget("i686--linux")
+	if err != nil {
+		t.Fatal("failed to load target:", err)
+	}
+	config := &compileopts.Config{
+		Options: &compileopts.Options{},
+		Target:  target,
+	}
+	compilerConfig := &Config{
+		Triple:             config.Triple(),
+		GOOS:               config.GOOS(),
+		GOARCH:             config.GOARCH(),
+		CodeModel:          config.CodeModel(),
+		RelocationModel:    config.RelocationModel(),
+		Scheduler:          config.Scheduler(),
+		FuncImplementation: config.FuncImplementation(),
+		AutomaticStackSize: config.AutomaticStackSize(),
+	}
+	machine, err := NewTargetMachine(compilerConfig)
+	if err != nil {
+		t.Fatal("failed to create target machine:", err)
+	}
+
+	tests := []string{
+		"basic.go",
+		"pointer.go",
+		"slice.go",
+		"float.go",
+	}
+
+	for _, testCase := range tests {
+		t.Run(testCase, func(t *testing.T) {
+			// Load entire program AST into memory.
+			lprogram, err := loader.Load(config, []string{"./testdata/" + testCase}, config.ClangHeaders, types.Config{
+				Sizes: Sizes(machine),
+			})
+			if err != nil {
+				t.Fatal("failed to create target machine:", err)
+			}
+			err = lprogram.Parse()
+			if err != nil {
+				t.Fatalf("could not parse test case %s: %s", testCase, err)
+			}
+
+			// Compile AST to IR.
+			pkg := lprogram.MainPkg()
+			mod, errs := CompilePackage(testCase, pkg, machine, compilerConfig, false)
+			if errs != nil {
+				for _, err := range errs {
+					t.Log("error:", err)
+				}
+				return
+			}
+
+			// Optimize IR a little.
+			funcPasses := llvm.NewFunctionPassManagerForModule(mod)
+			defer funcPasses.Dispose()
+			funcPasses.AddInstructionCombiningPass()
+			funcPasses.InitializeFunc()
+			for fn := mod.FirstFunction(); !fn.IsNil(); fn = llvm.NextFunction(fn) {
+				funcPasses.RunFunc(fn)
+			}
+			funcPasses.FinalizeFunc()
+
+			outfile := "./testdata/" + testCase[:len(testCase)-3] + ".ll"
+
+			// Update test if needed. Do not check the result.
+			if *flagUpdate {
+				err := ioutil.WriteFile(outfile, []byte(mod.String()), 0666)
+				if err != nil {
+					t.Error("failed to write updated output file:", err)
+				}
+				return
+			}
+
+			expected, err := ioutil.ReadFile(outfile)
+			if err != nil {
+				t.Fatal("failed to read golden file:", err)
+			}
+
+			if !fuzzyEqualIR(mod.String(), string(expected)) {
+				t.Errorf("output does not match expected output:\n%s", mod.String())
+			}
+		})
+	}
+}
+
+var alignRegexp = regexp.MustCompile(", align [0-9]+$")
+
+// fuzzyEqualIR returns true if the two LLVM IR strings passed in are roughly
+// equal. That means, only relevant lines are compared (excluding comments
+// etc.).
+func fuzzyEqualIR(s1, s2 string) bool {
+	lines1 := filterIrrelevantIRLines(strings.Split(s1, "\n"))
+	lines2 := filterIrrelevantIRLines(strings.Split(s2, "\n"))
+	if len(lines1) != len(lines2) {
+		return false
+	}
+	for i, line1 := range lines1 {
+		line2 := lines2[i]
+		match1 := alignRegexp.MatchString(line1)
+		match2 := alignRegexp.MatchString(line2)
+		if match1 != match2 {
+			// Only one of the lines has the align keyword. Remove it.
+			// This is a change to make the test work in both LLVM 10 and LLVM
+			// 11 (LLVM 11 appears to automatically add alignment everywhere).
+			line1 = alignRegexp.ReplaceAllString(line1, "")
+			line2 = alignRegexp.ReplaceAllString(line2, "")
+		}
+		if line1 != line2 {
+			return false
+		}
+	}
+
+	return true
+}
+
+// filterIrrelevantIRLines removes lines from the input slice of strings that
+// are not relevant in comparing IR. For example, empty lines and comments are
+// stripped out.
+func filterIrrelevantIRLines(lines []string) []string {
+	var out []string
+	llvmVersion, err := strconv.Atoi(strings.Split(llvm.Version, ".")[0])
+	if err != nil {
+		// Note: this should never happen and if it does, it will always happen
+		// for a particular build because llvm.Version is a constant.
+		panic(err)
+	}
+	for _, line := range lines {
+		line = strings.Split(line, ";")[0]    // strip out comments/info
+		line = strings.TrimRight(line, "\r ") // drop '\r' on Windows and remove trailing spaces from comments
+		if line == "" {
+			continue
+		}
+		if strings.HasPrefix(line, "source_filename = ") {
+			continue
+		}
+		if llvmVersion < 10 && strings.HasPrefix(line, "attributes ") {
+			// Ignore attribute groups. These may change between LLVM versions.
+			// Right now test outputs are for LLVM 10.
+			continue
+		}
+		if llvmVersion < 10 && strings.HasPrefix(line, "target datalayout ") {
+			// Ignore the target layout. This may change between LLVM versions.
+			continue
+		}
+		out = append(out, line)
+	}
+	return out
+}
@@ -14,8 +14,9 @@ package compiler
 //     frames.

 import (
+	"go/types"
+
 	"github.com/tinygo-org/tinygo/compiler/llvmutil"
-	"github.com/tinygo-org/tinygo/ir"
 	"golang.org/x/tools/go/ssa"
 	"tinygo.org/x/go-llvm"
 )
@@ -23,16 +24,18 @@ import (
 // deferInitFunc sets up this function for future deferred calls. It must be
 // called from within the entry block when this function contains deferred
 // calls.
-func (c *Compiler) deferInitFunc(frame *Frame) {
+func (b *builder) deferInitFunc() {
 	// Some setup.
-	frame.deferFuncs = make(map[*ir.Function]int)
-	frame.deferInvokeFuncs = make(map[string]int)
-	frame.deferClosureFuncs = make(map[*ir.Function]int)
+	b.deferFuncs = make(map[*ssa.Function]int)
+	b.deferInvokeFuncs = make(map[string]int)
+	b.deferClosureFuncs = make(map[*ssa.Function]int)
+	b.deferExprFuncs = make(map[ssa.Value]int)
+	b.deferBuiltinFuncs = make(map[ssa.Value]deferBuiltin)

 	// Create defer list pointer.
-	deferType := llvm.PointerType(c.getLLVMRuntimeType("_defer"), 0)
-	frame.deferPtr = c.builder.CreateAlloca(deferType, "deferPtr")
-	c.builder.CreateStore(llvm.ConstPointerNull(deferType), frame.deferPtr)
+	deferType := llvm.PointerType(b.getLLVMRuntimeType("_defer"), 0)
+	b.deferPtr = b.CreateAlloca(deferType, "deferPtr")
+	b.CreateStore(llvm.ConstPointerNull(deferType), b.deferPtr)
 }

 // isInLoop checks if there is a path from a basic block to itself.
@@ -69,53 +72,52 @@ func isInLoop(start *ssa.BasicBlock) bool {
 	return false
 }

-// emitDefer emits a single defer instruction, to be run when this function
+// createDefer emits a single defer instruction, to be run when this function
 // returns.
-func (c *Compiler) emitDefer(frame *Frame, instr *ssa.Defer) {
+func (b *builder) createDefer(instr *ssa.Defer) {
 	// The pointer to the previous defer struct, which we will replace to
 	// make a linked list.
-	next := c.builder.CreateLoad(frame.deferPtr, "defer.next")
+	next := b.CreateLoad(b.deferPtr, "defer.next")

 	var values []llvm.Value
-	valueTypes := []llvm.Type{c.uintptrType, next.Type()}
+	valueTypes := []llvm.Type{b.uintptrType, next.Type()}
 	if instr.Call.IsInvoke() {
 		// Method call on an interface.

 		// Get callback type number.
 		methodName := instr.Call.Method.FullName()
-		if _, ok := frame.deferInvokeFuncs[methodName]; !ok {
-			frame.deferInvokeFuncs[methodName] = len(frame.allDeferFuncs)
-			frame.allDeferFuncs = append(frame.allDeferFuncs, &instr.Call)
+		if _, ok := b.deferInvokeFuncs[methodName]; !ok {
+			b.deferInvokeFuncs[methodName] = len(b.allDeferFuncs)
+			b.allDeferFuncs = append(b.allDeferFuncs, &instr.Call)
 		}
-		callback := llvm.ConstInt(c.uintptrType, uint64(frame.deferInvokeFuncs[methodName]), false)
+		callback := llvm.ConstInt(b.uintptrType, uint64(b.deferInvokeFuncs[methodName]), false)

 		// Collect all values to be put in the struct (starting with
 		// runtime._defer fields, followed by the call parameters).
-		itf := c.getValue(frame, instr.Call.Value) // interface
-		receiverValue := c.builder.CreateExtractValue(itf, 1, "invoke.func.receiver")
-		values = []llvm.Value{callback, next, receiverValue}
-		valueTypes = append(valueTypes, c.i8ptrType)
+		itf := b.getValue(instr.Call.Value) // interface
+		typecode := b.CreateExtractValue(itf, 0, "invoke.func.typecode")
+		receiverValue := b.CreateExtractValue(itf, 1, "invoke.func.receiver")
+		values = []llvm.Value{callback, next, typecode, receiverValue}
+		valueTypes = append(valueTypes, b.uintptrType, b.i8ptrType)
 		for _, arg := range instr.Call.Args {
-			val := c.getValue(frame, arg)
+			val := b.getValue(arg)
 			values = append(values, val)
 			valueTypes = append(valueTypes, val.Type())
 		}

 	} else if callee, ok := instr.Call.Value.(*ssa.Function); ok {
 		// Regular function call.
-		fn := c.ir.GetFunction(callee)
-
-		if _, ok := frame.deferFuncs[fn]; !ok {
-			frame.deferFuncs[fn] = len(frame.allDeferFuncs)
-			frame.allDeferFuncs = append(frame.allDeferFuncs, fn)
+		if _, ok := b.deferFuncs[callee]; !ok {
+			b.deferFuncs[callee] = len(b.allDeferFuncs)
+			b.allDeferFuncs = append(b.allDeferFuncs, callee)
 		}
-		callback := llvm.ConstInt(c.uintptrType, uint64(frame.deferFuncs[fn]), false)
+		callback := llvm.ConstInt(b.uintptrType, uint64(b.deferFuncs[callee]), false)

 		// Collect all values to be put in the struct (starting with
 		// runtime._defer fields).
 		values = []llvm.Value{callback, next}
 		for _, param := range instr.Call.Args {
-			llvmParam := c.getValue(frame, param)
+			llvmParam := b.getValue(param)
 			values = append(values, llvmParam)
 			valueTypes = append(valueTypes, llvmParam.Type())
 		}
@@ -127,65 +129,109 @@ func (c *Compiler) emitDefer(frame *Frame, instr *ssa.Defer) {
 		// pointer.
 		// TODO: ignore this closure entirely and put pointers to the free
 		// variables directly in the defer struct, avoiding a memory allocation.
-		closure := c.getValue(frame, instr.Call.Value)
-		context := c.builder.CreateExtractValue(closure, 0, "")
+		closure := b.getValue(instr.Call.Value)
+		context := b.CreateExtractValue(closure, 0, "")

 		// Get the callback number.
-		fn := c.ir.GetFunction(makeClosure.Fn.(*ssa.Function))
-		if _, ok := frame.deferClosureFuncs[fn]; !ok {
-			frame.deferClosureFuncs[fn] = len(frame.allDeferFuncs)
-			frame.allDeferFuncs = append(frame.allDeferFuncs, makeClosure)
+		fn := makeClosure.Fn.(*ssa.Function)
+		if _, ok := b.deferClosureFuncs[fn]; !ok {
+			b.deferClosureFuncs[fn] = len(b.allDeferFuncs)
+			b.allDeferFuncs = append(b.allDeferFuncs, makeClosure)
 		}
-		callback := llvm.ConstInt(c.uintptrType, uint64(frame.deferClosureFuncs[fn]), false)
+		callback := llvm.ConstInt(b.uintptrType, uint64(b.deferClosureFuncs[fn]), false)

 		// Collect all values to be put in the struct (starting with
 		// runtime._defer fields, followed by all parameters including the
 		// context pointer).
 		values = []llvm.Value{callback, next}
 		for _, param := range instr.Call.Args {
-			llvmParam := c.getValue(frame, param)
+			llvmParam := b.getValue(param)
 			values = append(values, llvmParam)
 			valueTypes = append(valueTypes, llvmParam.Type())
 		}
 		values = append(values, context)
 		valueTypes = append(valueTypes, context.Type())

+	} else if builtin, ok := instr.Call.Value.(*ssa.Builtin); ok {
+		var argTypes []types.Type
+		var argValues []llvm.Value
+		for _, arg := range instr.Call.Args {
+			argTypes = append(argTypes, arg.Type())
+			argValues = append(argValues, b.getValue(arg))
+		}
+
+		if _, ok := b.deferBuiltinFuncs[instr.Call.Value]; !ok {
+			b.deferBuiltinFuncs[instr.Call.Value] = deferBuiltin{
+				callName: builtin.Name(),
+				pos:      builtin.Pos(),
+				argTypes: argTypes,
+				callback: len(b.allDeferFuncs),
+			}
+			b.allDeferFuncs = append(b.allDeferFuncs, instr.Call.Value)
+		}
+		callback := llvm.ConstInt(b.uintptrType, uint64(b.deferBuiltinFuncs[instr.Call.Value].callback), false)
+
+		// Collect all values to be put in the struct (starting with
+		// runtime._defer fields).
+		values = []llvm.Value{callback, next}
+		for _, param := range argValues {
+			values = append(values, param)
+			valueTypes = append(valueTypes, param.Type())
+		}
+
 	} else {
-		c.addError(instr.Pos(), "todo: defer on uncommon function call type")
-		return
+		funcValue := b.getValue(instr.Call.Value)
+
+		if _, ok := b.deferExprFuncs[instr.Call.Value]; !ok {
+			b.deferExprFuncs[instr.Call.Value] = len(b.allDeferFuncs)
+			b.allDeferFuncs = append(b.allDeferFuncs, &instr.Call)
+		}
+
+		callback := llvm.ConstInt(b.uintptrType, uint64(b.deferExprFuncs[instr.Call.Value]), false)
+
+		// Collect all values to be put in the struct (starting with
+		// runtime._defer fields, followed by all parameters including the
+		// context pointer).
+		values = []llvm.Value{callback, next, funcValue}
+		valueTypes = append(valueTypes, funcValue.Type())
+		for _, param := range instr.Call.Args {
+			llvmParam := b.getValue(param)
+			values = append(values, llvmParam)
+			valueTypes = append(valueTypes, llvmParam.Type())
+		}
 	}

 	// Make a struct out of the collected values to put in the defer frame.
-	deferFrameType := c.ctx.StructType(valueTypes, false)
+	deferFrameType := b.ctx.StructType(valueTypes, false)
 	deferFrame := llvm.ConstNull(deferFrameType)
 	for i, value := range values {
-		deferFrame = c.builder.CreateInsertValue(deferFrame, value, i, "")
+		deferFrame = b.CreateInsertValue(deferFrame, value, i, "")
 	}

 	// Put this struct in an allocation.
 	var alloca llvm.Value
 	if !isInLoop(instr.Block()) {
 		// This can safely use a stack allocation.
-		alloca = llvmutil.CreateEntryBlockAlloca(c.builder, deferFrameType, "defer.alloca")
+		alloca = llvmutil.CreateEntryBlockAlloca(b.Builder, deferFrameType, "defer.alloca")
 	} else {
 		// This may be hit a variable number of times, so use a heap allocation.
-		size := c.targetData.TypeAllocSize(deferFrameType)
-		sizeValue := llvm.ConstInt(c.uintptrType, size, false)
-		allocCall := c.createRuntimeCall("alloc", []llvm.Value{sizeValue}, "defer.alloc.call")
-		alloca = c.builder.CreateBitCast(allocCall, llvm.PointerType(deferFrameType, 0), "defer.alloc")
+		size := b.targetData.TypeAllocSize(deferFrameType)
+		sizeValue := llvm.ConstInt(b.uintptrType, size, false)
+		allocCall := b.createRuntimeCall("alloc", []llvm.Value{sizeValue}, "defer.alloc.call")
+		alloca = b.CreateBitCast(allocCall, llvm.PointerType(deferFrameType, 0), "defer.alloc")
 	}
-	if c.NeedsStackObjects() {
-		c.trackPointer(alloca)
+	if b.NeedsStackObjects {
+		b.trackPointer(alloca)
 	}
-	c.builder.CreateStore(deferFrame, alloca)
+	b.CreateStore(deferFrame, alloca)

 	// Push it on top of the linked list by replacing deferPtr.
-	allocaCast := c.builder.CreateBitCast(alloca, next.Type(), "defer.alloca.cast")
-	c.builder.CreateStore(allocaCast, frame.deferPtr)
+	allocaCast := b.CreateBitCast(alloca, next.Type(), "defer.alloca.cast")
+	b.CreateStore(allocaCast, b.deferPtr)
 }

-// emitRunDefers emits code to run all deferred functions.
-func (c *Compiler) emitRunDefers(frame *Frame) {
+// createRunDefers emits code to run all deferred functions.
+func (b *builder) createRunDefers() {
 	// Add a loop like the following:
 	//     for stack != nil {
 	//         _stack := stack
@@ -202,151 +248,208 @@ func (c *Compiler) emitRunDefers(frame *Frame) {
 	//     }

 	// Create loop.
-	loophead := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "rundefers.loophead")
-	loop := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "rundefers.loop")
-	unreachable := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "rundefers.default")
-	end := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "rundefers.end")
-	c.builder.CreateBr(loophead)
+	loophead := b.ctx.AddBasicBlock(b.llvmFn, "rundefers.loophead")
+	loop := b.ctx.AddBasicBlock(b.llvmFn, "rundefers.loop")
+	unreachable := b.ctx.AddBasicBlock(b.llvmFn, "rundefers.default")
+	end := b.ctx.AddBasicBlock(b.llvmFn, "rundefers.end")
+	b.CreateBr(loophead)

 	// Create loop head:
 	//     for stack != nil {
-	c.builder.SetInsertPointAtEnd(loophead)
-	deferData := c.builder.CreateLoad(frame.deferPtr, "")
-	stackIsNil := c.builder.CreateICmp(llvm.IntEQ, deferData, llvm.ConstPointerNull(deferData.Type()), "stackIsNil")
-	c.builder.CreateCondBr(stackIsNil, end, loop)
+	b.SetInsertPointAtEnd(loophead)
+	deferData := b.CreateLoad(b.deferPtr, "")
+	stackIsNil := b.CreateICmp(llvm.IntEQ, deferData, llvm.ConstPointerNull(deferData.Type()), "stackIsNil")
+	b.CreateCondBr(stackIsNil, end, loop)

 	// Create loop body:
 	//     _stack := stack
 	//     stack = stack.next
 	//     switch stack.callback {
-	c.builder.SetInsertPointAtEnd(loop)
-	nextStackGEP := c.builder.CreateInBoundsGEP(deferData, []llvm.Value{
-		llvm.ConstInt(c.ctx.Int32Type(), 0, false),
-		llvm.ConstInt(c.ctx.Int32Type(), 1, false), // .next field
+	b.SetInsertPointAtEnd(loop)
+	nextStackGEP := b.CreateInBoundsGEP(deferData, []llvm.Value{
+		llvm.ConstInt(b.ctx.Int32Type(), 0, false),
+		llvm.ConstInt(b.ctx.Int32Type(), 1, false), // .next field
 	}, "stack.next.gep")
-	nextStack := c.builder.CreateLoad(nextStackGEP, "stack.next")
-	c.builder.CreateStore(nextStack, frame.deferPtr)
-	gep := c.builder.CreateInBoundsGEP(deferData, []llvm.Value{
-		llvm.ConstInt(c.ctx.Int32Type(), 0, false),
-		llvm.ConstInt(c.ctx.Int32Type(), 0, false), // .callback field
+	nextStack := b.CreateLoad(nextStackGEP, "stack.next")
+	b.CreateStore(nextStack, b.deferPtr)
+	gep := b.CreateInBoundsGEP(deferData, []llvm.Value{
+		llvm.ConstInt(b.ctx.Int32Type(), 0, false),
+		llvm.ConstInt(b.ctx.Int32Type(), 0, false), // .callback field
 	}, "callback.gep")
-	callback := c.builder.CreateLoad(gep, "callback")
-	sw := c.builder.CreateSwitch(callback, unreachable, len(frame.allDeferFuncs))
+	callback := b.CreateLoad(gep, "callback")
+	sw := b.CreateSwitch(callback, unreachable, len(b.allDeferFuncs))

-	for i, callback := range frame.allDeferFuncs {
+	for i, callback := range b.allDeferFuncs {
 		// Create switch case, for example:
 		//     case 0:
 		//         // run first deferred call
-		block := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "rundefers.callback")
-		sw.AddCase(llvm.ConstInt(c.uintptrType, uint64(i), false), block)
-		c.builder.SetInsertPointAtEnd(block)
+		block := b.ctx.AddBasicBlock(b.llvmFn, "rundefers.callback")
+		sw.AddCase(llvm.ConstInt(b.uintptrType, uint64(i), false), block)
+		b.SetInsertPointAtEnd(block)
 		switch callback := callback.(type) {
 		case *ssa.CallCommon:
-			// Call on an interface value.
-			if !callback.IsInvoke() {
-				panic("expected an invoke call, not a direct call")
-			}
+			// Call on an value or interface value.

 			// Get the real defer struct type and cast to it.
-			valueTypes := []llvm.Type{c.uintptrType, llvm.PointerType(c.getLLVMRuntimeType("_defer"), 0), c.i8ptrType}
-			for _, arg := range callback.Args {
-				valueTypes = append(valueTypes, c.getLLVMType(arg.Type()))
+			valueTypes := []llvm.Type{b.uintptrType, llvm.PointerType(b.getLLVMRuntimeType("_defer"), 0)}
+
+			if !callback.IsInvoke() {
+				//Expect funcValue to be passed through the defer frame.
+				valueTypes = append(valueTypes, b.getFuncType(callback.Signature()))
+			} else {
+				//Expect typecode
+				valueTypes = append(valueTypes, b.uintptrType, b.i8ptrType)
 			}
-			deferFrameType := c.ctx.StructType(valueTypes, false)
-			deferFramePtr := c.builder.CreateBitCast(deferData, llvm.PointerType(deferFrameType, 0), "deferFrame")
+
+			for _, arg := range callback.Args {
+				valueTypes = append(valueTypes, b.getLLVMType(arg.Type()))
+			}
+
+			deferFrameType := b.ctx.StructType(valueTypes, false)
+			deferFramePtr := b.CreateBitCast(deferData, llvm.PointerType(deferFrameType, 0), "deferFrame")

 			// Extract the params from the struct (including receiver).
 			forwardParams := []llvm.Value{}
-			zero := llvm.ConstInt(c.ctx.Int32Type(), 0, false)
+			zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
 			for i := 2; i < len(valueTypes); i++ {
-				gep := c.builder.CreateInBoundsGEP(deferFramePtr, []llvm.Value{zero, llvm.ConstInt(c.ctx.Int32Type(), uint64(i), false)}, "gep")
-				forwardParam := c.builder.CreateLoad(gep, "param")
+				gep := b.CreateInBoundsGEP(deferFramePtr, []llvm.Value{zero, llvm.ConstInt(b.ctx.Int32Type(), uint64(i), false)}, "gep")
+				forwardParam := b.CreateLoad(gep, "param")
 				forwardParams = append(forwardParams, forwardParam)
 			}

-			// Add the context parameter. An interface call cannot also be a
-			// closure but we have to supply the parameter anyway for platforms
-			// with a strict calling convention.
-			forwardParams = append(forwardParams, llvm.Undef(c.i8ptrType))
+			var fnPtr llvm.Value
+
+			if !callback.IsInvoke() {
+				// Isolate the func value.
+				funcValue := forwardParams[0]
+				forwardParams = forwardParams[1:]
+
+				//Get function pointer and context
+				fp, context := b.decodeFuncValue(funcValue, callback.Signature())
+				fnPtr = fp
+
+				//Pass context
+				forwardParams = append(forwardParams, context)
+			} else {
+				// Isolate the typecode.
+				typecode := forwardParams[0]
+				forwardParams = forwardParams[1:]
+				fnPtr = b.getInvokePtr(callback, typecode)
+
+				// Add the context parameter. An interface call cannot also be a
+				// closure but we have to supply the parameter anyway for platforms
+				// with a strict calling convention.
+				forwardParams = append(forwardParams, llvm.Undef(b.i8ptrType))
+			}

 			// Parent coroutine handle.
-			forwardParams = append(forwardParams, llvm.Undef(c.i8ptrType))
+			forwardParams = append(forwardParams, llvm.Undef(b.i8ptrType))

-			fnPtr, _ := c.getInvokeCall(frame, callback)
-			c.createCall(fnPtr, forwardParams, "")
+			b.createCall(fnPtr, forwardParams, "")

-		case *ir.Function:
+		case *ssa.Function:
 			// Direct call.

 			// Get the real defer struct type and cast to it.
-			valueTypes := []llvm.Type{c.uintptrType, llvm.PointerType(c.getLLVMRuntimeType("_defer"), 0)}
+			valueTypes := []llvm.Type{b.uintptrType, llvm.PointerType(b.getLLVMRuntimeType("_defer"), 0)}
 			for _, param := range callback.Params {
-				valueTypes = append(valueTypes, c.getLLVMType(param.Type()))
+				valueTypes = append(valueTypes, b.getLLVMType(param.Type()))
 			}
-			deferFrameType := c.ctx.StructType(valueTypes, false)
-			deferFramePtr := c.builder.CreateBitCast(deferData, llvm.PointerType(deferFrameType, 0), "deferFrame")
+			deferFrameType := b.ctx.StructType(valueTypes, false)
+			deferFramePtr := b.CreateBitCast(deferData, llvm.PointerType(deferFrameType, 0), "deferFrame")

 			// Extract the params from the struct.
 			forwardParams := []llvm.Value{}
-			zero := llvm.ConstInt(c.ctx.Int32Type(), 0, false)
+			zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
 			for i := range callback.Params {
-				gep := c.builder.CreateInBoundsGEP(deferFramePtr, []llvm.Value{zero, llvm.ConstInt(c.ctx.Int32Type(), uint64(i+2), false)}, "gep")
-				forwardParam := c.builder.CreateLoad(gep, "param")
+				gep := b.CreateInBoundsGEP(deferFramePtr, []llvm.Value{zero, llvm.ConstInt(b.ctx.Int32Type(), uint64(i+2), false)}, "gep")
+				forwardParam := b.CreateLoad(gep, "param")
 				forwardParams = append(forwardParams, forwardParam)
 			}

-			// Add the context parameter. We know it is ignored by the receiving
-			// function, but we have to pass one anyway.
-			forwardParams = append(forwardParams, llvm.Undef(c.i8ptrType))
+			// Plain TinyGo functions add some extra parameters to implement async functionality and function recievers.
+			// These parameters should not be supplied when calling into an external C/ASM function.
+			if !b.getFunctionInfo(callback).exported {
+				// Add the context parameter. We know it is ignored by the receiving
+				// function, but we have to pass one anyway.
+				forwardParams = append(forwardParams, llvm.Undef(b.i8ptrType))

-			// Parent coroutine handle.
-			forwardParams = append(forwardParams, llvm.Undef(c.i8ptrType))
+				// Parent coroutine handle.
+				forwardParams = append(forwardParams, llvm.Undef(b.i8ptrType))
+			}

 			// Call real function.
-			c.createCall(callback.LLVMFn, forwardParams, "")
+			b.createCall(b.getFunction(callback), forwardParams, "")

 		case *ssa.MakeClosure:
 			// Get the real defer struct type and cast to it.
-			fn := c.ir.GetFunction(callback.Fn.(*ssa.Function))
-			valueTypes := []llvm.Type{c.uintptrType, llvm.PointerType(c.getLLVMRuntimeType("_defer"), 0)}
+			fn := callback.Fn.(*ssa.Function)
+			valueTypes := []llvm.Type{b.uintptrType, llvm.PointerType(b.getLLVMRuntimeType("_defer"), 0)}
 			params := fn.Signature.Params()
 			for i := 0; i < params.Len(); i++ {
-				valueTypes = append(valueTypes, c.getLLVMType(params.At(i).Type()))
+				valueTypes = append(valueTypes, b.getLLVMType(params.At(i).Type()))
 			}
-			valueTypes = append(valueTypes, c.i8ptrType) // closure
-			deferFrameType := c.ctx.StructType(valueTypes, false)
-			deferFramePtr := c.builder.CreateBitCast(deferData, llvm.PointerType(deferFrameType, 0), "deferFrame")
+			valueTypes = append(valueTypes, b.i8ptrType) // closure
+			deferFrameType := b.ctx.StructType(valueTypes, false)
+			deferFramePtr := b.CreateBitCast(deferData, llvm.PointerType(deferFrameType, 0), "deferFrame")

 			// Extract the params from the struct.
 			forwardParams := []llvm.Value{}
-			zero := llvm.ConstInt(c.ctx.Int32Type(), 0, false)
+			zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
 			for i := 2; i < len(valueTypes); i++ {
-				gep := c.builder.CreateInBoundsGEP(deferFramePtr, []llvm.Value{zero, llvm.ConstInt(c.ctx.Int32Type(), uint64(i), false)}, "")
-				forwardParam := c.builder.CreateLoad(gep, "param")
+				gep := b.CreateInBoundsGEP(deferFramePtr, []llvm.Value{zero, llvm.ConstInt(b.ctx.Int32Type(), uint64(i), false)}, "")
+				forwardParam := b.CreateLoad(gep, "param")
 				forwardParams = append(forwardParams, forwardParam)
 			}

 			// Parent coroutine handle.
-			forwardParams = append(forwardParams, llvm.Undef(c.i8ptrType))
+			forwardParams = append(forwardParams, llvm.Undef(b.i8ptrType))

 			// Call deferred function.
-			c.createCall(fn.LLVMFn, forwardParams, "")
+			b.createCall(b.getFunction(fn), forwardParams, "")
+		case *ssa.Builtin:
+			db := b.deferBuiltinFuncs[callback]

+			//Get parameter types
+			valueTypes := []llvm.Type{b.uintptrType, llvm.PointerType(b.getLLVMRuntimeType("_defer"), 0)}
+
+			//Get signature from call results
+			params := callback.Type().Underlying().(*types.Signature).Params()
+			for i := 0; i < params.Len(); i++ {
+				valueTypes = append(valueTypes, b.getLLVMType(params.At(i).Type()))
+			}
+
+			deferFrameType := b.ctx.StructType(valueTypes, false)
+			deferFramePtr := b.CreateBitCast(deferData, llvm.PointerType(deferFrameType, 0), "deferFrame")
+
+			// Extract the params from the struct.
+			var argValues []llvm.Value
+			zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
+			for i := 0; i < params.Len(); i++ {
+				gep := b.CreateInBoundsGEP(deferFramePtr, []llvm.Value{zero, llvm.ConstInt(b.ctx.Int32Type(), uint64(i+2), false)}, "gep")
+				forwardParam := b.CreateLoad(gep, "param")
+				argValues = append(argValues, forwardParam)
+			}
+
+			_, err := b.createBuiltin(db.argTypes, argValues, db.callName, db.pos)
+			if err != nil {
+				b.diagnostics = append(b.diagnostics, err)
+			}
 		default:
 			panic("unknown deferred function type")
 		}

 		// Branch back to the start of the loop.
-		c.builder.CreateBr(loophead)
+		b.CreateBr(loophead)
 	}

 	// Create default unreachable block:
 	//     default:
 	//         unreachable
 	//     }
-	c.builder.SetInsertPointAtEnd(unreachable)
-	c.builder.CreateUnreachable()
+	b.SetInsertPointAtEnd(unreachable)
+	b.CreateUnreachable()

 	// End of loop.
-	c.builder.SetInsertPointAtEnd(end)
+	b.SetInsertPointAtEnd(end)
 }
@@ -1,5 +1,7 @@
 package compiler

+// This file contains some utility functions related to error handling.
+
 import (
 	"go/scanner"
 	"go/token"
@@ -9,15 +11,16 @@ import (
 	"tinygo.org/x/go-llvm"
 )

-func (c *Compiler) makeError(pos token.Pos, msg string) types.Error {
+// makeError makes it easy to create an error from a token.Pos with a message.
+func (c *compilerContext) makeError(pos token.Pos, msg string) types.Error {
 	return types.Error{
-		Fset: c.ir.Program.Fset,
+		Fset: c.program.Fset,
 		Pos:  pos,
 		Msg:  msg,
 	}
 }

-func (c *Compiler) addError(pos token.Pos, msg string) {
+func (c *compilerContext) addError(pos token.Pos, msg string) {
 	c.diagnostics = append(c.diagnostics, c.makeError(pos, msg))
 }

@@ -10,49 +10,21 @@ import (
 	"tinygo.org/x/go-llvm"
 )

-type funcValueImplementation int
-
-const (
-	funcValueNone funcValueImplementation = iota
-
-	// A func value is implemented as a pair of pointers:
-	//     {context, function pointer}
-	// where the context may be a pointer to a heap-allocated struct containing
-	// the free variables, or it may be undef if the function being pointed to
-	// doesn't need a context. The function pointer is a regular function
-	// pointer.
-	funcValueDoubleword
-
-	// As funcValueDoubleword, but with the function pointer replaced by a
-	// unique ID per function signature. Function values are called by using a
-	// switch statement and choosing which function to call.
-	funcValueSwitch
-)
-
-// funcImplementation picks an appropriate func value implementation for the
-// target.
-func (c *Compiler) funcImplementation() funcValueImplementation {
-	// Always pick the switch implementation, as it allows the use of blocking
-	// inside a function that is used as a func value.
-	switch c.Scheduler() {
-	case "coroutines":
-		return funcValueSwitch
-	case "tasks":
-		return funcValueDoubleword
-	default:
-		panic("unknown scheduler type")
-	}
+// createFuncValue creates a function value from a raw function pointer with no
+// context.
+func (b *builder) createFuncValue(funcPtr, context llvm.Value, sig *types.Signature) llvm.Value {
+	return b.compilerContext.createFuncValue(b.Builder, funcPtr, context, sig)
 }

 // createFuncValue creates a function value from a raw function pointer with no
 // context.
-func (c *Compiler) createFuncValue(funcPtr, context llvm.Value, sig *types.Signature) llvm.Value {
+func (c *compilerContext) createFuncValue(builder llvm.Builder, funcPtr, context llvm.Value, sig *types.Signature) llvm.Value {
 	var funcValueScalar llvm.Value
-	switch c.funcImplementation() {
-	case funcValueDoubleword:
+	switch c.FuncImplementation {
+	case "doubleword":
 		// Closure is: {context, function pointer}
 		funcValueScalar = funcPtr
-	case funcValueSwitch:
+	case "switch":
 		sigGlobal := c.getTypeCode(sig)
 		funcValueWithSignatureGlobalName := funcPtr.Name() + "$withSignature"
 		funcValueWithSignatureGlobal := c.mod.NamedGlobal(funcValueWithSignatureGlobalName)
@@ -73,35 +45,35 @@ func (c *Compiler) createFuncValue(funcPtr, context llvm.Value, sig *types.Signa
 	}
 	funcValueType := c.getFuncType(sig)
 	funcValue := llvm.Undef(funcValueType)
-	funcValue = c.builder.CreateInsertValue(funcValue, context, 0, "")
-	funcValue = c.builder.CreateInsertValue(funcValue, funcValueScalar, 1, "")
+	funcValue = builder.CreateInsertValue(funcValue, context, 0, "")
+	funcValue = builder.CreateInsertValue(funcValue, funcValueScalar, 1, "")
 	return funcValue
 }

 // extractFuncScalar returns some scalar that can be used in comparisons. It is
 // a cheap operation.
-func (c *Compiler) extractFuncScalar(funcValue llvm.Value) llvm.Value {
-	return c.builder.CreateExtractValue(funcValue, 1, "")
+func (b *builder) extractFuncScalar(funcValue llvm.Value) llvm.Value {
+	return b.CreateExtractValue(funcValue, 1, "")
 }

 // extractFuncContext extracts the context pointer from this function value. It
 // is a cheap operation.
-func (c *Compiler) extractFuncContext(funcValue llvm.Value) llvm.Value {
-	return c.builder.CreateExtractValue(funcValue, 0, "")
+func (b *builder) extractFuncContext(funcValue llvm.Value) llvm.Value {
+	return b.CreateExtractValue(funcValue, 0, "")
 }

 // decodeFuncValue extracts the context and the function pointer from this func
 // value. This may be an expensive operation.
-func (c *Compiler) decodeFuncValue(funcValue llvm.Value, sig *types.Signature) (funcPtr, context llvm.Value) {
-	context = c.builder.CreateExtractValue(funcValue, 0, "")
-	switch c.funcImplementation() {
-	case funcValueDoubleword:
-		funcPtr = c.builder.CreateExtractValue(funcValue, 1, "")
-	case funcValueSwitch:
-		llvmSig := c.getRawFuncType(sig)
-		sigGlobal := c.getTypeCode(sig)
-		funcPtr = c.createRuntimeCall("getFuncPtr", []llvm.Value{funcValue, sigGlobal}, "")
-		funcPtr = c.builder.CreateIntToPtr(funcPtr, llvmSig, "")
+func (b *builder) decodeFuncValue(funcValue llvm.Value, sig *types.Signature) (funcPtr, context llvm.Value) {
+	context = b.CreateExtractValue(funcValue, 0, "")
+	switch b.FuncImplementation {
+	case "doubleword":
+		funcPtr = b.CreateExtractValue(funcValue, 1, "")
+	case "switch":
+		llvmSig := b.getRawFuncType(sig)
+		sigGlobal := b.getTypeCode(sig)
+		funcPtr = b.createRuntimeCall("getFuncPtr", []llvm.Value{funcValue, sigGlobal}, "")
+		funcPtr = b.CreateIntToPtr(funcPtr, llvmSig, "")
 	default:
 		panic("unimplemented func value variant")
 	}
@@ -109,12 +81,12 @@ func (c *Compiler) decodeFuncValue(funcValue llvm.Value, sig *types.Signature) (
 }

 // getFuncType returns the type of a func value given a signature.
-func (c *Compiler) getFuncType(typ *types.Signature) llvm.Type {
-	switch c.funcImplementation() {
-	case funcValueDoubleword:
+func (c *compilerContext) getFuncType(typ *types.Signature) llvm.Type {
+	switch c.FuncImplementation {
+	case "doubleword":
 		rawPtr := c.getRawFuncType(typ)
 		return c.ctx.StructType([]llvm.Type{c.i8ptrType, rawPtr}, false)
-	case funcValueSwitch:
+	case "switch":
 		return c.getLLVMRuntimeType("funcValue")
 	default:
 		panic("unimplemented func value variant")
@@ -122,7 +94,7 @@ func (c *Compiler) getFuncType(typ *types.Signature) llvm.Type {
 }

 // getRawFuncType returns a LLVM function pointer type for a given signature.
-func (c *Compiler) getRawFuncType(typ *types.Signature) llvm.Type {
+func (c *compilerContext) getRawFuncType(typ *types.Signature) llvm.Type {
 	// Get the return type.
 	var returnType llvm.Type
 	switch typ.Results().Len() {
@@ -152,11 +124,15 @@ func (c *Compiler) getRawFuncType(typ *types.Signature) llvm.Type {
 			// The receiver is not an interface, but a i8* type.
 			recv = c.i8ptrType
 		}
-		paramTypes = append(paramTypes, c.expandFormalParamType(recv)...)
+		for _, info := range expandFormalParamType(recv, "", nil) {
+			paramTypes = append(paramTypes, info.llvmType)
+		}
 	}
 	for i := 0; i < typ.Params().Len(); i++ {
 		subType := c.getLLVMType(typ.Params().At(i).Type())
-		paramTypes = append(paramTypes, c.expandFormalParamType(subType)...)
+		for _, info := range expandFormalParamType(subType, "", nil) {
+			paramTypes = append(paramTypes, info.llvmType)
+		}
 	}
 	// All functions take these parameters at the end.
 	paramTypes = append(paramTypes, c.i8ptrType) // context
@@ -168,24 +144,24 @@ func (c *Compiler) getRawFuncType(typ *types.Signature) llvm.Type {

 // parseMakeClosure makes a function value (with context) from the given
 // closure expression.
-func (c *Compiler) parseMakeClosure(frame *Frame, expr *ssa.MakeClosure) (llvm.Value, error) {
+func (b *builder) parseMakeClosure(expr *ssa.MakeClosure) (llvm.Value, error) {
 	if len(expr.Bindings) == 0 {
 		panic("unexpected: MakeClosure without bound variables")
 	}
-	f := c.ir.GetFunction(expr.Fn.(*ssa.Function))
+	f := expr.Fn.(*ssa.Function)

 	// Collect all bound variables.
 	boundVars := make([]llvm.Value, len(expr.Bindings))
 	for i, binding := range expr.Bindings {
 		// The context stores the bound variables.
-		llvmBoundVar := c.getValue(frame, binding)
+		llvmBoundVar := b.getValue(binding)
 		boundVars[i] = llvmBoundVar
 	}

 	// Store the bound variables in a single object, allocating it on the heap
 	// if necessary.
-	context := c.emitPointerPack(boundVars)
+	context := b.emitPointerPack(boundVars)

 	// Create the closure.
-	return c.createFuncValue(f.LLVMFn, context, f.Signature), nil
+	return b.createFuncValue(b.getFunction(f), context, f.Signature), nil
 }
@@ -12,53 +12,59 @@ import (

 // trackExpr inserts pointer tracking intrinsics for the GC if the expression is
 // one of the expressions that need this.
-func (c *Compiler) trackExpr(frame *Frame, expr ssa.Value, value llvm.Value) {
+func (b *builder) trackExpr(expr ssa.Value, value llvm.Value) {
 	// There are uses of this expression, Make sure the pointers
 	// are tracked during GC.
 	switch expr := expr.(type) {
 	case *ssa.Alloc, *ssa.MakeChan, *ssa.MakeMap:
 		// These values are always of pointer type in IR.
-		c.trackPointer(value)
+		b.trackPointer(value)
 	case *ssa.Call, *ssa.Convert, *ssa.MakeClosure, *ssa.MakeInterface, *ssa.MakeSlice, *ssa.Next:
 		if !value.IsNil() {
-			c.trackValue(value)
+			b.trackValue(value)
 		}
 	case *ssa.Select:
-		if alloca, ok := frame.selectRecvBuf[expr]; ok {
+		if alloca, ok := b.selectRecvBuf[expr]; ok {
 			if alloca.IsAUndefValue().IsNil() {
-				c.trackPointer(alloca)
+				b.trackPointer(alloca)
 			}
 		}
 	case *ssa.UnOp:
 		switch expr.Op {
 		case token.MUL:
 			// Pointer dereference.
-			c.trackValue(value)
+			b.trackValue(value)
 		case token.ARROW:
 			// Channel receive operator.
 			// It's not necessary to look at commaOk here, because in that
 			// case it's just an aggregate and trackValue will extract the
 			// pointer in there (if there is one).
-			c.trackValue(value)
+			b.trackValue(value)
+		}
+	case *ssa.BinOp:
+		switch expr.Op {
+		case token.ADD:
+			// String concatenation.
+			b.trackValue(value)
 		}
 	}
 }

 // trackValue locates pointers in a value (possibly an aggregate) and tracks the
 // individual pointers
-func (c *Compiler) trackValue(value llvm.Value) {
+func (b *builder) trackValue(value llvm.Value) {
 	typ := value.Type()
 	switch typ.TypeKind() {
 	case llvm.PointerTypeKind:
-		c.trackPointer(value)
+		b.trackPointer(value)
 	case llvm.StructTypeKind:
 		if !typeHasPointers(typ) {
 			return
 		}
 		numElements := typ.StructElementTypesCount()
 		for i := 0; i < numElements; i++ {
-			subValue := c.builder.CreateExtractValue(value, i, "")
-			c.trackValue(subValue)
+			subValue := b.CreateExtractValue(value, i, "")
+			b.trackValue(subValue)
 		}
 	case llvm.ArrayTypeKind:
 		if !typeHasPointers(typ) {
@@ -66,19 +72,19 @@ func (c *Compiler) trackValue(value llvm.Value) {
 		}
 		numElements := typ.ArrayLength()
 		for i := 0; i < numElements; i++ {
-			subValue := c.builder.CreateExtractValue(value, i, "")
-			c.trackValue(subValue)
+			subValue := b.CreateExtractValue(value, i, "")
+			b.trackValue(subValue)
 		}
 	}
 }

 // trackPointer creates a call to runtime.trackPointer, bitcasting the poitner
 // first if needed. The input value must be of LLVM pointer type.
-func (c *Compiler) trackPointer(value llvm.Value) {
-	if value.Type() != c.i8ptrType {
-		value = c.builder.CreateBitCast(value, c.i8ptrType, "")
+func (b *builder) trackPointer(value llvm.Value) {
+	if value.Type() != b.i8ptrType {
+		value = b.CreateBitCast(value, b.i8ptrType, "")
 	}
-	c.createRuntimeCall("trackPointer", []llvm.Value{value}, "")
+	b.createRuntimeCall("trackPointer", []llvm.Value{value}, "")
 }

 // typeHasPointers returns whether this type is a pointer or contains pointers.
@@ -3,34 +3,49 @@ package compiler
 // This file implements the 'go' keyword to start a new goroutine. See
 // goroutine-lowering.go for more details.

-import "tinygo.org/x/go-llvm"
+import (
+	"go/token"

-// emitStartGoroutine starts a new goroutine with the provided function pointer
+	"github.com/tinygo-org/tinygo/compiler/llvmutil"
+	"golang.org/x/tools/go/ssa"
+	"tinygo.org/x/go-llvm"
+)
+
+// createGoInstruction starts a new goroutine with the provided function pointer
 // and parameters.
 // In general, you should pass all regular parameters plus the context parameter.
 // There is one exception: the task-based scheduler needs to have the function
 // pointer passed in as a parameter too in addition to the context.
 //
 // Because a go statement doesn't return anything, return undef.
-func (c *Compiler) emitStartGoroutine(funcPtr llvm.Value, params []llvm.Value) llvm.Value {
-	switch c.Scheduler() {
-	case "tasks":
-		paramBundle := c.emitPointerPack(params)
-		paramBundle = c.builder.CreatePtrToInt(paramBundle, c.uintptrType, "")
-
-		calleeValue := c.createGoroutineStartWrapper(funcPtr)
-		c.createRuntimeCall("startGoroutine", []llvm.Value{calleeValue, paramBundle}, "")
+func (b *builder) createGoInstruction(funcPtr llvm.Value, params []llvm.Value, prefix string, pos token.Pos) llvm.Value {
+	paramBundle := b.emitPointerPack(params)
+	var callee, stackSize llvm.Value
+	switch b.Scheduler {
+	case "none", "tasks":
+		callee = b.createGoroutineStartWrapper(funcPtr, prefix, pos)
+		if b.AutomaticStackSize {
+			// The stack size is not known until after linking. Call a dummy
+			// function that will be replaced with a load from a special ELF
+			// section that contains the stack size (and is modified after
+			// linking).
+			stackSizeFn := b.getFunction(b.program.ImportedPackage("internal/task").Members["getGoroutineStackSize"].(*ssa.Function))
+			stackSize = b.createCall(stackSizeFn, []llvm.Value{callee, llvm.Undef(b.i8ptrType), llvm.Undef(b.i8ptrType)}, "stacksize")
+		} else {
+			// The stack size is fixed at compile time. By emitting it here as a
+			// constant, it can be optimized.
+			stackSize = llvm.ConstInt(b.uintptrType, b.DefaultStackSize, false)
+		}
 	case "coroutines":
-		// We roundtrip through runtime.makeGoroutine as a signal (to find these
-		// calls) and to break any optimizations LLVM will try to do: they are
-		// invalid if we called this as a regular function to be updated later.
-		calleeValue := c.builder.CreatePtrToInt(funcPtr, c.uintptrType, "")
-		calleeValue = c.createRuntimeCall("makeGoroutine", []llvm.Value{calleeValue}, "")
-		calleeValue = c.builder.CreateIntToPtr(calleeValue, funcPtr.Type(), "")
-		c.createCall(calleeValue, append(params, llvm.ConstPointerNull(c.i8ptrType)), "")
+		callee = b.CreatePtrToInt(funcPtr, b.uintptrType, "")
+		// There is no goroutine stack size: coroutines are used instead of
+		// stacks.
+		stackSize = llvm.Undef(b.uintptrType)
 	default:
 		panic("unreachable")
 	}
+	start := b.getFunction(b.program.ImportedPackage("internal/task").Members["start"].(*ssa.Function))
+	b.createCall(start, []llvm.Value{callee, paramBundle, stackSize, llvm.Undef(b.i8ptrType), llvm.ConstPointerNull(b.i8ptrType)}, "")
 	return llvm.Undef(funcPtr.Type().ElementType().ReturnType())
 }

@@ -52,36 +67,58 @@ func (c *Compiler) emitStartGoroutine(funcPtr llvm.Value, params []llvm.Value) l
 // allows a single (pointer) argument to the newly started goroutine. Also, it
 // ignores the return value because newly started goroutines do not have a
 // return value.
-func (c *Compiler) createGoroutineStartWrapper(fn llvm.Value) llvm.Value {
+func (c *compilerContext) createGoroutineStartWrapper(fn llvm.Value, prefix string, pos token.Pos) llvm.Value {
 	var wrapper llvm.Value

+	builder := c.ctx.NewBuilder()
+	defer builder.Dispose()
+
 	if !fn.IsAFunction().IsNil() {
 		// See whether this wrapper has already been created. If so, return it.
 		name := fn.Name()
 		wrapper = c.mod.NamedFunction(name + "$gowrapper")
 		if !wrapper.IsNil() {
-			return c.builder.CreatePtrToInt(wrapper, c.uintptrType, "")
+			return llvm.ConstPtrToInt(wrapper, c.uintptrType)
 		}

-		// Save the current position in the IR builder.
-		currentBlock := c.builder.GetInsertBlock()
-		defer c.builder.SetInsertPointAtEnd(currentBlock)
-
 		// Create the wrapper.
 		wrapperType := llvm.FunctionType(c.ctx.VoidType(), []llvm.Type{c.i8ptrType}, false)
 		wrapper = llvm.AddFunction(c.mod, name+"$gowrapper", wrapperType)
-		wrapper.SetLinkage(llvm.PrivateLinkage)
+		wrapper.SetLinkage(llvm.InternalLinkage)
 		wrapper.SetUnnamedAddr(true)
+		wrapper.AddAttributeAtIndex(-1, c.ctx.CreateStringAttribute("tinygo-gowrapper", name))
 		entry := c.ctx.AddBasicBlock(wrapper, "entry")
-		c.builder.SetInsertPointAtEnd(entry)
+		builder.SetInsertPointAtEnd(entry)
+
+		if c.Debug {
+			pos := c.program.Fset.Position(pos)
+			diFuncType := c.dibuilder.CreateSubroutineType(llvm.DISubroutineType{
+				File:       c.getDIFile(pos.Filename),
+				Parameters: nil, // do not show parameters in debugger
+				Flags:      0,   // ?
+			})
+			difunc := c.dibuilder.CreateFunction(c.getDIFile(pos.Filename), llvm.DIFunction{
+				Name:         "<goroutine wrapper>",
+				File:         c.getDIFile(pos.Filename),
+				Line:         pos.Line,
+				Type:         diFuncType,
+				LocalToUnit:  true,
+				IsDefinition: true,
+				ScopeLine:    0,
+				Flags:        llvm.FlagPrototyped,
+				Optimized:    true,
+			})
+			wrapper.SetSubprogram(difunc)
+			builder.SetCurrentDebugLocation(uint(pos.Line), uint(pos.Column), difunc, llvm.Metadata{})
+		}

 		// Create the list of params for the call.
 		paramTypes := fn.Type().ElementType().ParamTypes()
-		params := c.emitPointerUnpack(wrapper.Param(0), paramTypes[:len(paramTypes)-1])
+		params := llvmutil.EmitPointerUnpack(builder, c.mod, wrapper.Param(0), paramTypes[:len(paramTypes)-1])
 		params = append(params, llvm.Undef(c.i8ptrType))

 		// Create the call.
-		c.builder.CreateCall(fn, params, "")
+		builder.CreateCall(fn, params, "")

 	} else {
 		// For a function pointer like this:
@@ -101,22 +138,41 @@ func (c *Compiler) createGoroutineStartWrapper(fn llvm.Value) llvm.Value {
 		// With a bit of luck, identical wrapper functions like these can be
 		// merged into one.

-		// Save the current position in the IR builder.
-		currentBlock := c.builder.GetInsertBlock()
-		defer c.builder.SetInsertPointAtEnd(currentBlock)
-
 		// Create the wrapper.
 		wrapperType := llvm.FunctionType(c.ctx.VoidType(), []llvm.Type{c.i8ptrType}, false)
-		wrapper = llvm.AddFunction(c.mod, ".gowrapper", wrapperType)
+		wrapper = llvm.AddFunction(c.mod, prefix+".gowrapper", wrapperType)
 		wrapper.SetLinkage(llvm.InternalLinkage)
 		wrapper.SetUnnamedAddr(true)
+		wrapper.AddAttributeAtIndex(-1, c.ctx.CreateStringAttribute("tinygo-gowrapper", ""))
 		entry := c.ctx.AddBasicBlock(wrapper, "entry")
-		c.builder.SetInsertPointAtEnd(entry)
+		builder.SetInsertPointAtEnd(entry)
+
+		if c.Debug {
+			pos := c.program.Fset.Position(pos)
+			diFuncType := c.dibuilder.CreateSubroutineType(llvm.DISubroutineType{
+				File:       c.getDIFile(pos.Filename),
+				Parameters: nil, // do not show parameters in debugger
+				Flags:      0,   // ?
+			})
+			difunc := c.dibuilder.CreateFunction(c.getDIFile(pos.Filename), llvm.DIFunction{
+				Name:         "<goroutine wrapper>",
+				File:         c.getDIFile(pos.Filename),
+				Line:         pos.Line,
+				Type:         diFuncType,
+				LocalToUnit:  true,
+				IsDefinition: true,
+				ScopeLine:    0,
+				Flags:        llvm.FlagPrototyped,
+				Optimized:    true,
+			})
+			wrapper.SetSubprogram(difunc)
+			builder.SetCurrentDebugLocation(uint(pos.Line), uint(pos.Column), difunc, llvm.Metadata{})
+		}

 		// Get the list of parameters, with the extra parameters at the end.
 		paramTypes := fn.Type().ElementType().ParamTypes()
 		paramTypes[len(paramTypes)-1] = fn.Type() // the last element is the function pointer
-		params := c.emitPointerUnpack(wrapper.Param(0), paramTypes)
+		params := llvmutil.EmitPointerUnpack(builder, c.mod, wrapper.Param(0), paramTypes)

 		// Get the function pointer.
 		fnPtr := params[len(params)-1]
@@ -126,13 +182,13 @@ func (c *Compiler) createGoroutineStartWrapper(fn llvm.Value) llvm.Value {
 		params[len(params)-1] = llvm.Undef(c.i8ptrType)

 		// Create the call.
-		c.builder.CreateCall(fnPtr, params, "")
+		builder.CreateCall(fnPtr, params, "")
 	}

 	// Finish the function. Every basic block must end in a terminator, and
 	// because goroutines never return a value we can simply return void.
-	c.builder.CreateRetVoid()
+	builder.CreateRetVoid()

 	// Return a ptrtoint of the wrapper, not the function itself.
-	return c.builder.CreatePtrToInt(wrapper, c.uintptrType, "")
+	return builder.CreatePtrToInt(wrapper, c.uintptrType, "")
 }
@@ -13,27 +13,6 @@ import (
 	"tinygo.org/x/go-llvm"
 )

-// This is a compiler builtin, which reads the given register by name:
-//
-//     func ReadRegister(name string) uintptr
-//
-// The register name must be a constant, for example "sp".
-func (c *Compiler) emitReadRegister(name string, args []ssa.Value) (llvm.Value, error) {
-	fnType := llvm.FunctionType(c.uintptrType, []llvm.Type{}, false)
-	regname := constant.StringVal(args[0].(*ssa.Const).Value)
-	var asm string
-	switch name {
-	case "device/arm.ReadRegister":
-		asm = "mov $0, " + regname
-	case "device/riscv.ReadRegister":
-		asm = "mv $0, " + regname
-	default:
-		panic("unknown architecture")
-	}
-	target := llvm.InlineAsm(fnType, asm, "=r", false, false, 0)
-	return c.builder.CreateCall(target, nil, ""), nil
-}
-
 // This is a compiler builtin, which emits a piece of inline assembly with no
 // operands or return values. It is useful for trivial instructions, like wfi in
 // ARM or sleep in AVR.
@@ -41,18 +20,18 @@ func (c *Compiler) emitReadRegister(name string, args []ssa.Value) (llvm.Value,
 //     func Asm(asm string)
 //
 // The provided assembly must be a constant.
-func (c *Compiler) emitAsm(args []ssa.Value) (llvm.Value, error) {
+func (b *builder) createInlineAsm(args []ssa.Value) (llvm.Value, error) {
 	// Magic function: insert inline assembly instead of calling it.
-	fnType := llvm.FunctionType(c.ctx.VoidType(), []llvm.Type{}, false)
+	fnType := llvm.FunctionType(b.ctx.VoidType(), []llvm.Type{}, false)
 	asm := constant.StringVal(args[0].(*ssa.Const).Value)
 	target := llvm.InlineAsm(fnType, asm, "", true, false, 0)
-	return c.builder.CreateCall(target, nil, ""), nil
+	return b.CreateCall(target, nil, ""), nil
 }

 // This is a compiler builtin, which allows assembly to be called in a flexible
 // way.
 //
-//     func AsmFull(asm string, regs map[string]interface{})
+//     func AsmFull(asm string, regs map[string]interface{}) uintptr
 //
 // The asm parameter must be a constant string. The regs parameter must be
 // provided immediately. For example:
@@ -63,27 +42,27 @@ func (c *Compiler) emitAsm(args []ssa.Value) (llvm.Value, error) {
 //             "value":  1
 //             "result": &dest,
 //         })
-func (c *Compiler) emitAsmFull(frame *Frame, instr *ssa.CallCommon) (llvm.Value, error) {
+func (b *builder) createInlineAsmFull(instr *ssa.CallCommon) (llvm.Value, error) {
 	asmString := constant.StringVal(instr.Args[0].(*ssa.Const).Value)
 	registers := map[string]llvm.Value{}
-	registerMap := instr.Args[1].(*ssa.MakeMap)
-	for _, r := range *registerMap.Referrers() {
-		switch r := r.(type) {
-		case *ssa.DebugRef:
-			// ignore
-		case *ssa.MapUpdate:
-			if r.Block() != registerMap.Block() {
-				return llvm.Value{}, c.makeError(instr.Pos(), "register value map must be created in the same basic block")
+	if registerMap, ok := instr.Args[1].(*ssa.MakeMap); ok {
+		for _, r := range *registerMap.Referrers() {
+			switch r := r.(type) {
+			case *ssa.DebugRef:
+				// ignore
+			case *ssa.MapUpdate:
+				if r.Block() != registerMap.Block() {
+					return llvm.Value{}, b.makeError(instr.Pos(), "register value map must be created in the same basic block")
+				}
+				key := constant.StringVal(r.Key.(*ssa.Const).Value)
+				registers[key] = b.getValue(r.Value.(*ssa.MakeInterface).X)
+			case *ssa.Call:
+				if r.Common() == instr {
+					break
+				}
+			default:
+				return llvm.Value{}, b.makeError(instr.Pos(), "don't know how to handle argument to inline assembly: "+r.String())
 			}
-			key := constant.StringVal(r.Key.(*ssa.Const).Value)
-			//println("value:", r.Value.(*ssa.MakeInterface).X.String())
-			registers[key] = c.getValue(frame, r.Value.(*ssa.MakeInterface).X)
-		case *ssa.Call:
-			if r.Common() == instr {
-				break
-			}
-		default:
-			return llvm.Value{}, c.makeError(instr.Pos(), "don't know how to handle argument to inline assembly: "+r.String())
 		}
 	}
 	// TODO: handle dollar signs in asm string
@@ -92,13 +71,22 @@ func (c *Compiler) emitAsmFull(frame *Frame, instr *ssa.CallCommon) (llvm.Value,
 	argTypes := []llvm.Type{}
 	args := []llvm.Value{}
 	constraints := []string{}
+	hasOutput := false
+	asmString = regexp.MustCompile("\\{\\}").ReplaceAllStringFunc(asmString, func(s string) string {
+		hasOutput = true
+		return "$0"
+	})
+	if hasOutput {
+		constraints = append(constraints, "=&r")
+		registerNumbers[""] = 0
+	}
 	asmString = regexp.MustCompile("\\{[a-zA-Z]+\\}").ReplaceAllStringFunc(asmString, func(s string) string {
 		// TODO: skip strings like {r4} etc. that look like ARM push/pop
 		// instructions.
 		name := s[1 : len(s)-1]
 		if _, ok := registers[name]; !ok {
 			if err == nil {
-				err = c.makeError(instr.Pos(), "unknown register name: "+name)
+				err = b.makeError(instr.Pos(), "unknown register name: "+name)
 			}
 			return s
 		}
@@ -112,7 +100,7 @@ func (c *Compiler) emitAsmFull(frame *Frame, instr *ssa.CallCommon) (llvm.Value,
 			case llvm.PointerTypeKind:
 				constraints = append(constraints, "*m")
 			default:
-				err = c.makeError(instr.Pos(), "unknown type in inline assembly for value: "+name)
+				err = b.makeError(instr.Pos(), "unknown type in inline assembly for value: "+name)
 				return s
 			}
 		}
@@ -121,9 +109,21 @@ func (c *Compiler) emitAsmFull(frame *Frame, instr *ssa.CallCommon) (llvm.Value,
 	if err != nil {
 		return llvm.Value{}, err
 	}
-	fnType := llvm.FunctionType(c.ctx.VoidType(), argTypes, false)
+	var outputType llvm.Type
+	if hasOutput {
+		outputType = b.uintptrType
+	} else {
+		outputType = b.ctx.VoidType()
+	}
+	fnType := llvm.FunctionType(outputType, argTypes, false)
 	target := llvm.InlineAsm(fnType, asmString, strings.Join(constraints, ","), true, false, 0)
-	return c.builder.CreateCall(target, args, ""), nil
+	result := b.CreateCall(target, args, "")
+	if hasOutput {
+		return result, nil
+	} else {
+		// Make sure we return something valid.
+		return llvm.ConstInt(b.uintptrType, 0, false), nil
+	}
 }

 // This is a compiler builtin which emits an inline SVCall instruction. It can
@@ -137,7 +137,7 @@ func (c *Compiler) emitAsmFull(frame *Frame, instr *ssa.CallCommon) (llvm.Value,
 //
 // The num parameter must be a constant. All other parameters may be any scalar
 // value supported by LLVM inline assembly.
-func (c *Compiler) emitSVCall(frame *Frame, args []ssa.Value) (llvm.Value, error) {
+func (b *builder) emitSVCall(args []ssa.Value) (llvm.Value, error) {
 	num, _ := constant.Uint64Val(args[0].(*ssa.Const).Value)
 	llvmArgs := []llvm.Value{}
 	argTypes := []llvm.Type{}
@@ -150,7 +150,7 @@ func (c *Compiler) emitSVCall(frame *Frame, args []ssa.Value) (llvm.Value, error
 		} else {
 			constraints += ",{r" + strconv.Itoa(i) + "}"
 		}
-		llvmValue := c.getValue(frame, arg)
+		llvmValue := b.getValue(arg)
 		llvmArgs = append(llvmArgs, llvmValue)
 		argTypes = append(argTypes, llvmValue.Type())
 	}
@@ -158,9 +158,47 @@ func (c *Compiler) emitSVCall(frame *Frame, args []ssa.Value) (llvm.Value, error
 	// clobbered. r0 is used as an output register so doesn't have to be
 	// marked as clobbered.
 	constraints += ",~{r1},~{r2},~{r3}"
-	fnType := llvm.FunctionType(c.uintptrType, argTypes, false)
+	fnType := llvm.FunctionType(b.uintptrType, argTypes, false)
 	target := llvm.InlineAsm(fnType, asm, constraints, true, false, 0)
-	return c.builder.CreateCall(target, llvmArgs, ""), nil
+	return b.CreateCall(target, llvmArgs, ""), nil
+}
+
+// This is a compiler builtin which emits an inline SVCall instruction. It can
+// be one of:
+//
+//     func SVCall0(num uintptr) uintptr
+//     func SVCall1(num uintptr, a1 interface{}) uintptr
+//     func SVCall2(num uintptr, a1, a2 interface{}) uintptr
+//     func SVCall3(num uintptr, a1, a2, a3 interface{}) uintptr
+//     func SVCall4(num uintptr, a1, a2, a3, a4 interface{}) uintptr
+//
+// The num parameter must be a constant. All other parameters may be any scalar
+// value supported by LLVM inline assembly.
+// Same as emitSVCall but for AArch64
+func (b *builder) emitSV64Call(args []ssa.Value) (llvm.Value, error) {
+	num, _ := constant.Uint64Val(args[0].(*ssa.Const).Value)
+	llvmArgs := []llvm.Value{}
+	argTypes := []llvm.Type{}
+	asm := "svc #" + strconv.FormatUint(num, 10)
+	constraints := "={x0}"
+	for i, arg := range args[1:] {
+		arg = arg.(*ssa.MakeInterface).X
+		if i == 0 {
+			constraints += ",0"
+		} else {
+			constraints += ",{x" + strconv.Itoa(i) + "}"
+		}
+		llvmValue := b.getValue(arg)
+		llvmArgs = append(llvmArgs, llvmValue)
+		argTypes = append(argTypes, llvmValue.Type())
+	}
+	// Implement the ARM64 calling convention by marking x1-x7 as
+	// clobbered. x0 is used as an output register so doesn't have to be
+	// marked as clobbered.
+	constraints += ",~{x1},~{x2},~{x3},~{x4},~{x5},~{x6},~{x7}"
+	fnType := llvm.FunctionType(b.uintptrType, argTypes, false)
+	target := llvm.InlineAsm(fnType, asm, constraints, true, false, 0)
+	return b.CreateCall(target, llvmArgs, ""), nil
 }

 // This is a compiler builtin which emits CSR instructions. It can be one of:
@@ -172,38 +210,38 @@ func (c *Compiler) emitSVCall(frame *Frame, args []ssa.Value) (llvm.Value, error
 //
 // The csr parameter (method receiver) must be a constant. Other parameter can
 // be any value.
-func (c *Compiler) emitCSROperation(frame *Frame, call *ssa.CallCommon) (llvm.Value, error) {
+func (b *builder) emitCSROperation(call *ssa.CallCommon) (llvm.Value, error) {
 	csrConst, ok := call.Args[0].(*ssa.Const)
 	if !ok {
-		return llvm.Value{}, c.makeError(call.Pos(), "CSR must be constant")
+		return llvm.Value{}, b.makeError(call.Pos(), "CSR must be constant")
 	}
 	csr := csrConst.Uint64()
 	switch name := call.StaticCallee().Name(); name {
 	case "Get":
 		// Note that this instruction may have side effects, and thus must be
 		// marked as such.
-		fnType := llvm.FunctionType(c.uintptrType, nil, false)
+		fnType := llvm.FunctionType(b.uintptrType, nil, false)
 		asm := fmt.Sprintf("csrr $0, %d", csr)
 		target := llvm.InlineAsm(fnType, asm, "=r", true, false, 0)
-		return c.builder.CreateCall(target, nil, ""), nil
+		return b.CreateCall(target, nil, ""), nil
 	case "Set":
-		fnType := llvm.FunctionType(c.ctx.VoidType(), []llvm.Type{c.uintptrType}, false)
+		fnType := llvm.FunctionType(b.ctx.VoidType(), []llvm.Type{b.uintptrType}, false)
 		asm := fmt.Sprintf("csrw %d, $0", csr)
 		target := llvm.InlineAsm(fnType, asm, "r", true, false, 0)
-		return c.builder.CreateCall(target, []llvm.Value{c.getValue(frame, call.Args[1])}, ""), nil
+		return b.CreateCall(target, []llvm.Value{b.getValue(call.Args[1])}, ""), nil
 	case "SetBits":
 		// Note: it may be possible to optimize this to csrrsi in many cases.
-		fnType := llvm.FunctionType(c.uintptrType, []llvm.Type{c.uintptrType}, false)
+		fnType := llvm.FunctionType(b.uintptrType, []llvm.Type{b.uintptrType}, false)
 		asm := fmt.Sprintf("csrrs $0, %d, $1", csr)
 		target := llvm.InlineAsm(fnType, asm, "=r,r", true, false, 0)
-		return c.builder.CreateCall(target, []llvm.Value{c.getValue(frame, call.Args[1])}, ""), nil
+		return b.CreateCall(target, []llvm.Value{b.getValue(call.Args[1])}, ""), nil
 	case "ClearBits":
 		// Note: it may be possible to optimize this to csrrci in many cases.
-		fnType := llvm.FunctionType(c.uintptrType, []llvm.Type{c.uintptrType}, false)
+		fnType := llvm.FunctionType(b.uintptrType, []llvm.Type{b.uintptrType}, false)
 		asm := fmt.Sprintf("csrrc $0, %d, $1", csr)
 		target := llvm.InlineAsm(fnType, asm, "=r,r", true, false, 0)
-		return c.builder.CreateCall(target, []llvm.Value{c.getValue(frame, call.Args[1])}, ""), nil
+		return b.CreateCall(target, []llvm.Value{b.getValue(call.Args[1])}, ""), nil
 	default:
-		return llvm.Value{}, c.makeError(call.Pos(), "unknown CSR operation: "+name)
+		return llvm.Value{}, b.makeError(call.Pos(), "unknown CSR operation: "+name)
 	}
 }
@@ -11,40 +11,39 @@ import (
 	"strconv"
 	"strings"

-	"github.com/tinygo-org/tinygo/ir"
 	"golang.org/x/tools/go/ssa"
 	"tinygo.org/x/go-llvm"
 )

-// parseMakeInterface emits the LLVM IR for the *ssa.MakeInterface instruction.
+// createMakeInterface emits the LLVM IR for the *ssa.MakeInterface instruction.
 // It tries to put the type in the interface value, but if that's not possible,
 // it will do an allocation of the right size and put that in the interface
 // value field.
 //
-// An interface value is a {typecode, value} tuple, or {i16, i8*} to be exact.
-func (c *Compiler) parseMakeInterface(val llvm.Value, typ types.Type, pos token.Pos) llvm.Value {
-	itfValue := c.emitPointerPack([]llvm.Value{val})
-	itfTypeCodeGlobal := c.getTypeCode(typ)
-	itfMethodSetGlobal := c.getTypeMethodSet(typ)
-	itfConcreteTypeGlobal := c.mod.NamedGlobal("typeInInterface:" + itfTypeCodeGlobal.Name())
+// An interface value is a {typecode, value} tuple named runtime._interface.
+func (b *builder) createMakeInterface(val llvm.Value, typ types.Type, pos token.Pos) llvm.Value {
+	itfValue := b.emitPointerPack([]llvm.Value{val})
+	itfTypeCodeGlobal := b.getTypeCode(typ)
+	itfMethodSetGlobal := b.getTypeMethodSet(typ)
+	itfConcreteTypeGlobal := b.mod.NamedGlobal("typeInInterface:" + itfTypeCodeGlobal.Name())
 	if itfConcreteTypeGlobal.IsNil() {
-		typeInInterface := c.getLLVMRuntimeType("typeInInterface")
-		itfConcreteTypeGlobal = llvm.AddGlobal(c.mod, typeInInterface, "typeInInterface:"+itfTypeCodeGlobal.Name())
+		typeInInterface := b.getLLVMRuntimeType("typeInInterface")
+		itfConcreteTypeGlobal = llvm.AddGlobal(b.mod, typeInInterface, "typeInInterface:"+itfTypeCodeGlobal.Name())
 		itfConcreteTypeGlobal.SetInitializer(llvm.ConstNamedStruct(typeInInterface, []llvm.Value{itfTypeCodeGlobal, itfMethodSetGlobal}))
 		itfConcreteTypeGlobal.SetGlobalConstant(true)
 		itfConcreteTypeGlobal.SetLinkage(llvm.PrivateLinkage)
 	}
-	itfTypeCode := c.builder.CreatePtrToInt(itfConcreteTypeGlobal, c.uintptrType, "")
-	itf := llvm.Undef(c.getLLVMRuntimeType("_interface"))
-	itf = c.builder.CreateInsertValue(itf, itfTypeCode, 0, "")
-	itf = c.builder.CreateInsertValue(itf, itfValue, 1, "")
+	itfTypeCode := b.CreatePtrToInt(itfConcreteTypeGlobal, b.uintptrType, "")
+	itf := llvm.Undef(b.getLLVMRuntimeType("_interface"))
+	itf = b.CreateInsertValue(itf, itfTypeCode, 0, "")
+	itf = b.CreateInsertValue(itf, itfValue, 1, "")
 	return itf
 }

 // getTypeCode returns a reference to a type code.
 // It returns a pointer to an external global which should be replaced with the
 // real type in the interface lowering pass.
-func (c *Compiler) getTypeCode(typ types.Type) llvm.Value {
+func (c *compilerContext) getTypeCode(typ types.Type) llvm.Value {
 	globalName := "reflect/types.type:" + getTypeCodeName(typ)
 	global := c.mod.NamedGlobal(globalName)
 	if global.IsNil() {
@@ -91,7 +90,7 @@ func (c *Compiler) getTypeCode(typ types.Type) llvm.Value {
 // makeStructTypeFields creates a new global that stores all type information
 // related to this struct type, and returns the resulting global. This global is
 // actually an array of all the fields in the structs.
-func (c *Compiler) makeStructTypeFields(typ *types.Struct) llvm.Value {
+func (c *compilerContext) makeStructTypeFields(typ *types.Struct) llvm.Value {
 	// The global is an array of runtime.structField structs.
 	runtimeStructField := c.getLLVMRuntimeType("structField")
 	structGlobalType := llvm.ArrayType(runtimeStructField, typ.NumFields())
@@ -228,7 +227,7 @@ func getTypeCodeName(t types.Type) string {

 // getTypeMethodSet returns a reference (GEP) to a global method set. This
 // method set should be unreferenced after the interface lowering pass.
-func (c *Compiler) getTypeMethodSet(typ types.Type) llvm.Value {
+func (c *compilerContext) getTypeMethodSet(typ types.Type) llvm.Value {
 	global := c.mod.NamedGlobal(typ.String() + "$methodset")
 	zero := llvm.ConstInt(c.ctx.Int32Type(), 0, false)
 	if !global.IsNil() {
@@ -236,7 +235,7 @@ func (c *Compiler) getTypeMethodSet(typ types.Type) llvm.Value {
 		return llvm.ConstGEP(global, []llvm.Value{zero, zero})
 	}

-	ms := c.ir.Program.MethodSets.MethodSet(typ)
+	ms := c.program.MethodSets.MethodSet(typ)
 	if ms.Len() == 0 {
 		// no methods, so can leave that one out
 		return llvm.ConstPointerNull(llvm.PointerType(c.getLLVMRuntimeType("interfaceMethodInfo"), 0))
@@ -247,15 +246,16 @@ func (c *Compiler) getTypeMethodSet(typ types.Type) llvm.Value {
 	for i := 0; i < ms.Len(); i++ {
 		method := ms.At(i)
 		signatureGlobal := c.getMethodSignature(method.Obj().(*types.Func))
-		f := c.ir.GetFunction(c.ir.Program.MethodValue(method))
-		if f.LLVMFn.IsNil() {
+		fn := c.program.MethodValue(method)
+		llvmFn := c.getFunction(fn)
+		if llvmFn.IsNil() {
 			// compiler error, so panic
-			panic("cannot find function: " + f.LinkName())
+			panic("cannot find function: " + c.getFunctionInfo(fn).linkName)
 		}
-		fn := c.getInterfaceInvokeWrapper(f)
+		wrapper := c.getInterfaceInvokeWrapper(fn, llvmFn)
 		methodInfo := llvm.ConstNamedStruct(interfaceMethodInfoType, []llvm.Value{
 			signatureGlobal,
-			llvm.ConstPtrToInt(fn, c.uintptrType),
+			llvm.ConstPtrToInt(wrapper, c.uintptrType),
 		})
 		methods[i] = methodInfo
 	}
@@ -271,8 +271,13 @@ func (c *Compiler) getTypeMethodSet(typ types.Type) llvm.Value {
 // getInterfaceMethodSet returns a global variable with the method set of the
 // given named interface type. This method set is used by the interface lowering
 // pass.
-func (c *Compiler) getInterfaceMethodSet(typ *types.Named) llvm.Value {
-	global := c.mod.NamedGlobal(typ.String() + "$interface")
+func (c *compilerContext) getInterfaceMethodSet(typ types.Type) llvm.Value {
+	name := typ.String()
+	if _, ok := typ.(*types.Named); !ok {
+		// Anonymous interface.
+		name = "reflect/types.interface:" + name
+	}
+	global := c.mod.NamedGlobal(name + "$interface")
 	zero := llvm.ConstInt(c.ctx.Int32Type(), 0, false)
 	if !global.IsNil() {
 		// method set already exist, return it
@@ -287,7 +292,7 @@ func (c *Compiler) getInterfaceMethodSet(typ *types.Named) llvm.Value {
 	}

 	value := llvm.ConstArray(c.i8ptrType, methods)
-	global = llvm.AddGlobal(c.mod, value.Type(), typ.String()+"$interface")
+	global = llvm.AddGlobal(c.mod, value.Type(), name+"$interface")
 	global.SetInitializer(value)
 	global.SetGlobalConstant(true)
 	global.SetLinkage(llvm.PrivateLinkage)
@@ -297,8 +302,8 @@ func (c *Compiler) getInterfaceMethodSet(typ *types.Named) llvm.Value {
 // getMethodSignature returns a global variable which is a reference to an
 // external *i8 indicating the indicating the signature of this method. It is
 // used during the interface lowering pass.
-func (c *Compiler) getMethodSignature(method *types.Func) llvm.Value {
-	signature := ir.MethodSignature(method)
+func (c *compilerContext) getMethodSignature(method *types.Func) llvm.Value {
+	signature := methodSignature(method)
 	signatureGlobal := c.mod.NamedGlobal("func " + signature)
 	if signatureGlobal.IsNil() {
 		signatureGlobal = llvm.AddGlobal(c.mod, c.ctx.Int8Type(), "func "+signature)
@@ -307,18 +312,18 @@ func (c *Compiler) getMethodSignature(method *types.Func) llvm.Value {
 	return signatureGlobal
 }

-// parseTypeAssert will emit the code for a typeassert, used in if statements
+// createTypeAssert will emit the code for a typeassert, used in if statements
 // and in type switches (Go SSA does not have type switches, only if/else
 // chains). Note that even though the Go SSA does not contain type switches,
 // LLVM will recognize the pattern and make it a real switch in many cases.
 //
 // Type asserts on concrete types are trivial: just compare type numbers. Type
 // asserts on interfaces are more difficult, see the comments in the function.
-func (c *Compiler) parseTypeAssert(frame *Frame, expr *ssa.TypeAssert) llvm.Value {
-	itf := c.getValue(frame, expr.X)
-	assertedType := c.getLLVMType(expr.AssertedType)
+func (b *builder) createTypeAssert(expr *ssa.TypeAssert) llvm.Value {
+	itf := b.getValue(expr.X)
+	assertedType := b.getLLVMType(expr.AssertedType)

-	actualTypeNum := c.builder.CreateExtractValue(itf, 0, "interface.type")
+	actualTypeNum := b.CreateExtractValue(itf, 0, "interface.type")
 	commaOk := llvm.Value{}
 	if _, ok := expr.AssertedType.Underlying().(*types.Interface); ok {
 		// Type assert on interface type.
@@ -329,15 +334,15 @@ func (c *Compiler) parseTypeAssert(frame *Frame, expr *ssa.TypeAssert) llvm.Valu
 		// the main Go compiler, where the runtime checks whether the type
 		// implements each method of the interface. See:
 		// https://research.swtch.com/interfaces
-		methodSet := c.getInterfaceMethodSet(expr.AssertedType.(*types.Named))
-		commaOk = c.createRuntimeCall("interfaceImplements", []llvm.Value{actualTypeNum, methodSet}, "")
+		methodSet := b.getInterfaceMethodSet(expr.AssertedType)
+		commaOk = b.createRuntimeCall("interfaceImplements", []llvm.Value{actualTypeNum, methodSet}, "")

 	} else {
 		// Type assert on concrete type.
 		// Call runtime.typeAssert, which will be lowered to a simple icmp or
 		// const false in the interface lowering pass.
-		assertedTypeCodeGlobal := c.getTypeCode(expr.AssertedType)
-		commaOk = c.createRuntimeCall("typeAssert", []llvm.Value{actualTypeNum, assertedTypeCodeGlobal}, "typecode")
+		assertedTypeCodeGlobal := b.getTypeCode(expr.AssertedType)
+		commaOk = b.createRuntimeCall("typeAssert", []llvm.Value{actualTypeNum, assertedTypeCodeGlobal}, "typecode")
 	}

 	// Add 2 new basic blocks (that should get optimized away): one for the
@@ -351,15 +356,15 @@ func (c *Compiler) parseTypeAssert(frame *Frame, expr *ssa.TypeAssert) llvm.Valu
 	// typeassert should return a zero value, not an incorrectly casted
 	// value.

-	prevBlock := c.builder.GetInsertBlock()
-	okBlock := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "typeassert.ok")
-	nextBlock := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "typeassert.next")
-	frame.blockExits[frame.currentBlock] = nextBlock // adjust outgoing block for phi nodes
-	c.builder.CreateCondBr(commaOk, okBlock, nextBlock)
+	prevBlock := b.GetInsertBlock()
+	okBlock := b.ctx.AddBasicBlock(b.llvmFn, "typeassert.ok")
+	nextBlock := b.ctx.AddBasicBlock(b.llvmFn, "typeassert.next")
+	b.blockExits[b.currentBlock] = nextBlock // adjust outgoing block for phi nodes
+	b.CreateCondBr(commaOk, okBlock, nextBlock)

 	// Retrieve the value from the interface if the type assert was
 	// successful.
-	c.builder.SetInsertPointAtEnd(okBlock)
+	b.SetInsertPointAtEnd(okBlock)
 	var valueOk llvm.Value
 	if _, ok := expr.AssertedType.Underlying().(*types.Interface); ok {
 		// Type assert on interface type. Easy: just return the same
@@ -368,75 +373,71 @@ func (c *Compiler) parseTypeAssert(frame *Frame, expr *ssa.TypeAssert) llvm.Valu
 	} else {
 		// Type assert on concrete type. Extract the underlying type from
 		// the interface (but only after checking it matches).
-		valuePtr := c.builder.CreateExtractValue(itf, 1, "typeassert.value.ptr")
-		valueOk = c.emitPointerUnpack(valuePtr, []llvm.Type{assertedType})[0]
+		valuePtr := b.CreateExtractValue(itf, 1, "typeassert.value.ptr")
+		valueOk = b.emitPointerUnpack(valuePtr, []llvm.Type{assertedType})[0]
 	}
-	c.builder.CreateBr(nextBlock)
+	b.CreateBr(nextBlock)

 	// Continue after the if statement.
-	c.builder.SetInsertPointAtEnd(nextBlock)
-	phi := c.builder.CreatePHI(assertedType, "typeassert.value")
+	b.SetInsertPointAtEnd(nextBlock)
+	phi := b.CreatePHI(assertedType, "typeassert.value")
 	phi.AddIncoming([]llvm.Value{llvm.ConstNull(assertedType), valueOk}, []llvm.BasicBlock{prevBlock, okBlock})

 	if expr.CommaOk {
-		tuple := c.ctx.ConstStruct([]llvm.Value{llvm.Undef(assertedType), llvm.Undef(c.ctx.Int1Type())}, false) // create empty tuple
-		tuple = c.builder.CreateInsertValue(tuple, phi, 0, "")                                                  // insert value
-		tuple = c.builder.CreateInsertValue(tuple, commaOk, 1, "")                                              // insert 'comma ok' boolean
+		tuple := b.ctx.ConstStruct([]llvm.Value{llvm.Undef(assertedType), llvm.Undef(b.ctx.Int1Type())}, false) // create empty tuple
+		tuple = b.CreateInsertValue(tuple, phi, 0, "")                                                          // insert value
+		tuple = b.CreateInsertValue(tuple, commaOk, 1, "")                                                      // insert 'comma ok' boolean
 		return tuple
 	} else {
 		// This is kind of dirty as the branch above becomes mostly useless,
 		// but hopefully this gets optimized away.
-		c.createRuntimeCall("interfaceTypeAssert", []llvm.Value{commaOk}, "")
+		b.createRuntimeCall("interfaceTypeAssert", []llvm.Value{commaOk}, "")
 		return phi
 	}
 }

-// getInvokeCall creates and returns the function pointer and parameters of an
-// interface call. It can be used in a call or defer instruction.
-func (c *Compiler) getInvokeCall(frame *Frame, instr *ssa.CallCommon) (llvm.Value, []llvm.Value) {
-	// Call an interface method with dynamic dispatch.
-	itf := c.getValue(frame, instr.Value) // interface
-
-	llvmFnType := c.getRawFuncType(instr.Method.Type().(*types.Signature))
-
-	typecode := c.builder.CreateExtractValue(itf, 0, "invoke.typecode")
+// getInvokePtr creates an interface function pointer lookup for the specified invoke instruction, using a specified typecode.
+func (b *builder) getInvokePtr(instr *ssa.CallCommon, typecode llvm.Value) llvm.Value {
+	llvmFnType := b.getRawFuncType(instr.Method.Type().(*types.Signature))
 	values := []llvm.Value{
 		typecode,
-		c.getInterfaceMethodSet(instr.Value.Type().(*types.Named)),
-		c.getMethodSignature(instr.Method),
+		b.getInterfaceMethodSet(instr.Value.Type()),
+		b.getMethodSignature(instr.Method),
 	}
-	fn := c.createRuntimeCall("interfaceMethod", values, "invoke.func")
-	fnCast := c.builder.CreateIntToPtr(fn, llvmFnType, "invoke.func.cast")
-	receiverValue := c.builder.CreateExtractValue(itf, 1, "invoke.func.receiver")
+	fn := b.createRuntimeCall("interfaceMethod", values, "invoke.func")
+	return b.CreateIntToPtr(fn, llvmFnType, "invoke.func.cast")
+}
+
+// getInvokeCall creates and returns the function pointer and parameters of an
+// interface call.
+func (b *builder) getInvokeCall(instr *ssa.CallCommon) (llvm.Value, []llvm.Value) {
+	// Call an interface method with dynamic dispatch.
+	itf := b.getValue(instr.Value) // interface
+
+	typecode := b.CreateExtractValue(itf, 0, "invoke.typecode")
+	fnCast := b.getInvokePtr(instr, typecode)
+	receiverValue := b.CreateExtractValue(itf, 1, "invoke.func.receiver")

 	args := []llvm.Value{receiverValue}
 	for _, arg := range instr.Args {
-		args = append(args, c.getValue(frame, arg))
+		args = append(args, b.getValue(arg))
 	}
 	// Add the context parameter. An interface call never takes a context but we
 	// have to supply the parameter anyway.
-	args = append(args, llvm.Undef(c.i8ptrType))
+	args = append(args, llvm.Undef(b.i8ptrType))
 	// Add the parent goroutine handle.
-	args = append(args, llvm.Undef(c.i8ptrType))
+	args = append(args, llvm.Undef(b.i8ptrType))

 	return fnCast, args
 }

-// interfaceInvokeWrapper keeps some state between getInterfaceInvokeWrapper and
-// createInterfaceInvokeWrapper. The former is called during IR construction
-// itself and the latter is called when finishing up the IR.
-type interfaceInvokeWrapper struct {
-	fn           *ir.Function
-	wrapper      llvm.Value
-	receiverType llvm.Type
-}
-
-// Wrap an interface method function pointer. The wrapper takes in a pointer to
-// the underlying value, dereferences it, and calls the real method. This
-// wrapper is only needed when the interface value actually doesn't fit in a
-// pointer and a pointer to the value must be created.
-func (c *Compiler) getInterfaceInvokeWrapper(f *ir.Function) llvm.Value {
-	wrapperName := f.LinkName() + "$invoke"
+// getInterfaceInvokeWrapper returns a wrapper for the given method so it can be
+// invoked from an interface. The wrapper takes in a pointer to the underlying
+// value, dereferences or unpacks it if necessary, and calls the real method.
+// If the method to wrap has a pointer receiver, no wrapping is necessary and
+// the function is returned directly.
+func (c *compilerContext) getInterfaceInvokeWrapper(fn *ssa.Function, llvmFn llvm.Value) llvm.Value {
+	wrapperName := llvmFn.Name() + "$invoke"
 	wrapper := c.mod.NamedFunction(wrapperName)
 	if !wrapper.IsNil() {
 		// Wrapper already created. Return it directly.
@@ -444,8 +445,11 @@ func (c *Compiler) getInterfaceInvokeWrapper(f *ir.Function) llvm.Value {
 	}

 	// Get the expanded receiver type.
-	receiverType := c.getLLVMType(f.Params[0].Type())
-	expandedReceiverType := c.expandFormalParamType(receiverType)
+	receiverType := c.getLLVMType(fn.Params[0].Type())
+	var expandedReceiverType []llvm.Type
+	for _, info := range expandFormalParamType(receiverType, "", nil) {
+		expandedReceiverType = append(expandedReceiverType, info.llvmType)
+	}

 	// Does this method even need any wrapping?
 	if len(expandedReceiverType) == 1 && receiverType.TypeKind() == llvm.PointerTypeKind {
@@ -453,52 +457,93 @@ func (c *Compiler) getInterfaceInvokeWrapper(f *ir.Function) llvm.Value {
 		// Casting a function signature to a different signature and calling it
 		// with a receiver pointer bitcasted to *i8 (as done in calls on an
 		// interface) is hopefully a safe (defined) operation.
-		return f.LLVMFn
+		return llvmFn
 	}

 	// create wrapper function
-	fnType := f.LLVMFn.Type().ElementType()
+	fnType := llvmFn.Type().ElementType()
 	paramTypes := append([]llvm.Type{c.i8ptrType}, fnType.ParamTypes()[len(expandedReceiverType):]...)
 	wrapFnType := llvm.FunctionType(fnType.ReturnType(), paramTypes, false)
 	wrapper = llvm.AddFunction(c.mod, wrapperName, wrapFnType)
-	if f.LLVMFn.LastParam().Name() == "parentHandle" {
-		wrapper.LastParam().SetName("parentHandle")
-	}
-	c.interfaceInvokeWrappers = append(c.interfaceInvokeWrappers, interfaceInvokeWrapper{
-		fn:           f,
-		wrapper:      wrapper,
-		receiverType: receiverType,
-	})
-	return wrapper
-}
+	wrapper.LastParam().SetName("parentHandle")

-// createInterfaceInvokeWrapper finishes the work of getInterfaceInvokeWrapper,
-// see that function for details.
-func (c *Compiler) createInterfaceInvokeWrapper(state interfaceInvokeWrapper) {
-	wrapper := state.wrapper
-	fn := state.fn
-	receiverType := state.receiverType
 	wrapper.SetLinkage(llvm.InternalLinkage)
 	wrapper.SetUnnamedAddr(true)

+	// Create a new builder just to create this wrapper.
+	b := builder{
+		compilerContext: c,
+		Builder:         c.ctx.NewBuilder(),
+	}
+	defer b.Builder.Dispose()
+
 	// add debug info if needed
-	if c.Debug() {
-		pos := c.ir.Program.Fset.Position(fn.Pos())
+	if c.Debug {
+		pos := c.program.Fset.Position(fn.Pos())
 		difunc := c.attachDebugInfoRaw(fn, wrapper, "$invoke", pos.Filename, pos.Line)
-		c.builder.SetCurrentDebugLocation(uint(pos.Line), uint(pos.Column), difunc, llvm.Metadata{})
+		b.SetCurrentDebugLocation(uint(pos.Line), uint(pos.Column), difunc, llvm.Metadata{})
 	}

 	// set up IR builder
-	block := c.ctx.AddBasicBlock(wrapper, "entry")
-	c.builder.SetInsertPointAtEnd(block)
+	block := b.ctx.AddBasicBlock(wrapper, "entry")
+	b.SetInsertPointAtEnd(block)

-	receiverValue := c.emitPointerUnpack(wrapper.Param(0), []llvm.Type{receiverType})[0]
-	params := append(c.expandFormalParam(receiverValue), wrapper.Params()[1:]...)
-	if fn.LLVMFn.Type().ElementType().ReturnType().TypeKind() == llvm.VoidTypeKind {
-		c.builder.CreateCall(fn.LLVMFn, params, "")
-		c.builder.CreateRetVoid()
+	receiverValue := b.emitPointerUnpack(wrapper.Param(0), []llvm.Type{receiverType})[0]
+	params := append(b.expandFormalParam(receiverValue), wrapper.Params()[1:]...)
+	if llvmFn.Type().ElementType().ReturnType().TypeKind() == llvm.VoidTypeKind {
+		b.CreateCall(llvmFn, params, "")
+		b.CreateRetVoid()
 	} else {
-		ret := c.builder.CreateCall(fn.LLVMFn, params, "ret")
-		c.builder.CreateRet(ret)
+		ret := b.CreateCall(llvmFn, params, "ret")
+		b.CreateRet(ret)
 	}
+
+	return wrapper
+}
+
+// methodSignature creates a readable version of a method signature (including
+// the function name, excluding the receiver name). This string is used
+// internally to match interfaces and to call the correct method on an
+// interface. Examples:
+//
+//     String() string
+//     Read([]byte) (int, error)
+func methodSignature(method *types.Func) string {
+	return method.Name() + signature(method.Type().(*types.Signature))
+}
+
+// Make a readable version of a function (pointer) signature.
+// Examples:
+//
+//     () string
+//     (string, int) (int, error)
+func signature(sig *types.Signature) string {
+	s := ""
+	if sig.Params().Len() == 0 {
+		s += "()"
+	} else {
+		s += "("
+		for i := 0; i < sig.Params().Len(); i++ {
+			if i > 0 {
+				s += ", "
+			}
+			s += sig.Params().At(i).Type().String()
+		}
+		s += ")"
+	}
+	if sig.Results().Len() == 0 {
+		// keep as-is
+	} else if sig.Results().Len() == 1 {
+		s += " " + sig.Results().At(0).Type().String()
+	} else {
+		s += " ("
+		for i := 0; i < sig.Results().Len(); i++ {
+			if i > 0 {
+				s += ", "
+			}
+			s += sig.Results().At(i).Type().String()
+		}
+		s += ")"
+	}
+	return s
 }
@@ -8,62 +8,62 @@ import (
 	"tinygo.org/x/go-llvm"
 )

-// emitInterruptGlobal creates a new runtime/interrupt.Interrupt struct that
+// createInterruptGlobal creates a new runtime/interrupt.Interrupt struct that
 // will be lowered to a real interrupt during interrupt lowering.
 //
 // This two-stage approach allows unused interrupts to be optimized away if
 // necessary.
-func (c *Compiler) emitInterruptGlobal(frame *Frame, instr *ssa.CallCommon) (llvm.Value, error) {
+func (b *builder) createInterruptGlobal(instr *ssa.CallCommon) (llvm.Value, error) {
 	// Get the interrupt number, which must be a compile-time constant.
 	id, ok := instr.Args[0].(*ssa.Const)
 	if !ok {
-		return llvm.Value{}, c.makeError(instr.Pos(), "interrupt ID is not a constant")
+		return llvm.Value{}, b.makeError(instr.Pos(), "interrupt ID is not a constant")
 	}

 	// Get the func value, which also must be a compile time constant.
 	// Note that bound functions are allowed if the function has a pointer
 	// receiver and is a global. This is rather strict but still allows for
 	// idiomatic Go code.
-	funcValue := c.getValue(frame, instr.Args[1])
+	funcValue := b.getValue(instr.Args[1])
 	if funcValue.IsAConstant().IsNil() {
 		// Try to determine the cause of the non-constantness for a nice error
 		// message.
 		switch instr.Args[1].(type) {
 		case *ssa.MakeClosure:
 			// This may also be a bound method.
-			return llvm.Value{}, c.makeError(instr.Pos(), "closures are not supported in interrupt.New")
+			return llvm.Value{}, b.makeError(instr.Pos(), "closures are not supported in interrupt.New")
 		}
 		// Fall back to a generic error.
-		return llvm.Value{}, c.makeError(instr.Pos(), "interrupt function must be constant")
+		return llvm.Value{}, b.makeError(instr.Pos(), "interrupt function must be constant")
 	}

 	// Create a new global of type runtime/interrupt.handle. Globals of this
 	// type are lowered in the interrupt lowering pass.
-	globalType := c.ir.Program.ImportedPackage("runtime/interrupt").Type("handle").Type()
-	globalLLVMType := c.getLLVMType(globalType)
+	globalType := b.program.ImportedPackage("runtime/interrupt").Type("handle").Type()
+	globalLLVMType := b.getLLVMType(globalType)
 	globalName := "runtime/interrupt.$interrupt" + strconv.FormatInt(id.Int64(), 10)
-	if global := c.mod.NamedGlobal(globalName); !global.IsNil() {
-		return llvm.Value{}, c.makeError(instr.Pos(), "interrupt redeclared in this program")
+	if global := b.mod.NamedGlobal(globalName); !global.IsNil() {
+		return llvm.Value{}, b.makeError(instr.Pos(), "interrupt redeclared in this program")
 	}
-	global := llvm.AddGlobal(c.mod, globalLLVMType, globalName)
+	global := llvm.AddGlobal(b.mod, globalLLVMType, globalName)
 	global.SetLinkage(llvm.PrivateLinkage)
 	global.SetGlobalConstant(true)
 	global.SetUnnamedAddr(true)
 	initializer := llvm.ConstNull(globalLLVMType)
 	initializer = llvm.ConstInsertValue(initializer, funcValue, []uint32{0})
-	initializer = llvm.ConstInsertValue(initializer, llvm.ConstInt(c.intType, uint64(id.Int64()), true), []uint32{1, 0})
+	initializer = llvm.ConstInsertValue(initializer, llvm.ConstInt(b.intType, uint64(id.Int64()), true), []uint32{1, 0})
 	global.SetInitializer(initializer)

 	// Add debug info to the interrupt global.
-	if c.Debug() {
-		pos := c.ir.Program.Fset.Position(instr.Pos())
-		diglobal := c.dibuilder.CreateGlobalVariableExpression(c.difiles[pos.Filename], llvm.DIGlobalVariableExpression{
+	if b.Debug {
+		pos := b.program.Fset.Position(instr.Pos())
+		diglobal := b.dibuilder.CreateGlobalVariableExpression(b.getDIFile(pos.Filename), llvm.DIGlobalVariableExpression{
 			Name:        "interrupt" + strconv.FormatInt(id.Int64(), 10),
 			LinkageName: globalName,
-			File:        c.getDIFile(pos.Filename),
+			File:        b.getDIFile(pos.Filename),
 			Line:        pos.Line,
-			Type:        c.getDIType(globalType),
-			Expr:        c.dibuilder.CreateExpression(nil),
+			Type:        b.getDIType(globalType),
+			Expr:        b.dibuilder.CreateExpression(nil),
 			LocalToUnit: false,
 		})
 		global.AddMetadata(0, diglobal)
@@ -71,21 +71,21 @@ func (c *Compiler) emitInterruptGlobal(frame *Frame, instr *ssa.CallCommon) (llv

 	// Create the runtime/interrupt.Interrupt type. It is a struct with a single
 	// member of type int.
-	num := llvm.ConstPtrToInt(global, c.intType)
-	interrupt := llvm.ConstNamedStruct(c.mod.GetTypeByName("runtime/interrupt.Interrupt"), []llvm.Value{num})
+	num := llvm.ConstPtrToInt(global, b.intType)
+	interrupt := llvm.ConstNamedStruct(b.mod.GetTypeByName("runtime/interrupt.Interrupt"), []llvm.Value{num})

 	// Add dummy "use" call for AVR, because interrupts may be used even though
 	// they are never referenced again. This is unlike Cortex-M or the RISC-V
 	// PLIC where each interrupt must be enabled using the interrupt number, and
 	// thus keeps the Interrupt object alive.
 	// This call is removed during interrupt lowering.
-	if strings.HasPrefix(c.Triple(), "avr") {
-		useFn := c.mod.NamedFunction("runtime/interrupt.use")
+	if strings.HasPrefix(b.Triple, "avr") {
+		useFn := b.mod.NamedFunction("runtime/interrupt.use")
 		if useFn.IsNil() {
-			useFnType := llvm.FunctionType(c.ctx.VoidType(), []llvm.Type{interrupt.Type()}, false)
-			useFn = llvm.AddFunction(c.mod, "runtime/interrupt.use", useFnType)
+			useFnType := llvm.FunctionType(b.ctx.VoidType(), []llvm.Type{interrupt.Type()}, false)
+			useFn = llvm.AddFunction(b.mod, "runtime/interrupt.use", useFnType)
 		}
-		c.builder.CreateCall(useFn, []llvm.Value{interrupt}, "")
+		b.CreateCall(useFn, []llvm.Value{interrupt}, "")
 	}

 	return interrupt, nil
@@ -0,0 +1,50 @@
+package compiler
+
+// This file contains helper functions to create calls to LLVM intrinsics.
+
+import (
+	"strconv"
+
+	"golang.org/x/tools/go/ssa"
+	"tinygo.org/x/go-llvm"
+)
+
+// createMemoryCopyCall creates a call to a builtin LLVM memcpy or memmove
+// function, declaring this function if needed. These calls are treated
+// specially by optimization passes possibly resulting in better generated code,
+// and will otherwise be lowered to regular libc memcpy/memmove calls.
+func (b *builder) createMemoryCopyCall(fn *ssa.Function, args []ssa.Value) (llvm.Value, error) {
+	fnName := "llvm." + fn.Name() + ".p0i8.p0i8.i" + strconv.Itoa(b.uintptrType.IntTypeWidth())
+	llvmFn := b.mod.NamedFunction(fnName)
+	if llvmFn.IsNil() {
+		fnType := llvm.FunctionType(b.ctx.VoidType(), []llvm.Type{b.i8ptrType, b.i8ptrType, b.uintptrType, b.ctx.Int1Type()}, false)
+		llvmFn = llvm.AddFunction(b.mod, fnName, fnType)
+	}
+	var params []llvm.Value
+	for _, param := range args {
+		params = append(params, b.getValue(param))
+	}
+	params = append(params, llvm.ConstInt(b.ctx.Int1Type(), 0, false))
+	b.CreateCall(llvmFn, params, "")
+	return llvm.Value{}, nil
+}
+
+// createMemoryZeroCall creates calls to llvm.memset.* to zero a block of
+// memory, declaring the function if needed. These calls will be lowered to
+// regular libc memset calls if they aren't optimized out in a different way.
+func (b *builder) createMemoryZeroCall(args []ssa.Value) (llvm.Value, error) {
+	fnName := "llvm.memset.p0i8.i" + strconv.Itoa(b.uintptrType.IntTypeWidth())
+	llvmFn := b.mod.NamedFunction(fnName)
+	if llvmFn.IsNil() {
+		fnType := llvm.FunctionType(b.ctx.VoidType(), []llvm.Type{b.i8ptrType, b.ctx.Int8Type(), b.uintptrType, b.ctx.Int1Type()}, false)
+		llvmFn = llvm.AddFunction(b.mod, fnName, fnType)
+	}
+	params := []llvm.Value{
+		b.getValue(args[0]),
+		llvm.ConstInt(b.ctx.Int8Type(), 0, false),
+		b.getValue(args[1]),
+		llvm.ConstInt(b.ctx.Int1Type(), 0, false),
+	}
+	b.CreateCall(llvmFn, params, "")
+	return llvm.Value{}, nil
+}
@@ -1,7 +1,8 @@
-package compiler
-
-// This file implements a set of sanity checks for the IR that is generated.
-// It can catch some mistakes that LLVM's verifier cannot.
+// Package ircheck implements a checker for LLVM IR, that goes a bit further
+// than the regular LLVM IR verifier. Note that it checks different things, so
+// this is not a replacement for the LLVM verifier but does catch things that
+// the LLVM verifier doesn't catch.
+package ircheck

 import (
 	"errors"
@@ -10,7 +11,11 @@ import (
 	"tinygo.org/x/go-llvm"
 )

-func (c *Compiler) checkType(t llvm.Type, checked map[llvm.Type]struct{}, specials map[llvm.TypeKind]llvm.Type) error {
+type checker struct {
+	ctx llvm.Context
+}
+
+func (c *checker) checkType(t llvm.Type, checked map[llvm.Type]struct{}, specials map[llvm.TypeKind]llvm.Type) error {
 	// prevent infinite recursion for self-referential types
 	if _, ok := checked[t]; ok {
 		return nil
@@ -81,7 +86,7 @@ func (c *Compiler) checkType(t llvm.Type, checked map[llvm.Type]struct{}, specia
 	return nil
 }

-func (c *Compiler) checkValue(v llvm.Value, types map[llvm.Type]struct{}, specials map[llvm.TypeKind]llvm.Type) error {
+func (c *checker) checkValue(v llvm.Value, types map[llvm.Type]struct{}, specials map[llvm.TypeKind]llvm.Type) error {
 	// check type
 	if err := c.checkType(v.Type(), types, specials); err != nil {
 		return fmt.Errorf("failed to verify type of value: %s", err.Error())
@@ -95,12 +100,30 @@ func (c *Compiler) checkValue(v llvm.Value, types map[llvm.Type]struct{}, specia
 	return nil
 }

-func (c *Compiler) checkInstruction(inst llvm.Value, types map[llvm.Type]struct{}, specials map[llvm.TypeKind]llvm.Type) error {
+func (c *checker) checkInstruction(inst llvm.Value, types map[llvm.Type]struct{}, specials map[llvm.TypeKind]llvm.Type) error {
 	// check value properties
 	if err := c.checkValue(inst, types, specials); err != nil {
 		return errorAt(inst, err.Error())
 	}

+	// The alloca instruction can be present in every basic block. However,
+	// allocas in basic blocks other than the entry basic block have a number of
+	// problems:
+	//   * They are hard to optimize, leading to potential missed optimizations.
+	//   * They may cause stack overflows in loops that would otherwise be
+	//     innocent.
+	//   * They cause extra code to be generated, because it requires the use of
+	//     a frame pointer.
+	//   * Perhaps most importantly, the coroutine lowering pass of LLVM (as of
+	//     LLVM 9) cannot deal with these allocas:
+	//     https://llvm.org/docs/Coroutines.html
+	// Therefore, alloca instructions should be limited to the entry block.
+	if !inst.IsAAllocaInst().IsNil() {
+		if inst.InstructionParent() != inst.InstructionParent().Parent().EntryBasicBlock() {
+			return errorAt(inst, "internal error: non-static alloca")
+		}
+	}
+
 	// check operands
 	for i := 0; i < inst.OperandsCount(); i++ {
 		if err := c.checkValue(inst.Operand(i), types, specials); err != nil {
@@ -111,7 +134,7 @@ func (c *Compiler) checkInstruction(inst llvm.Value, types map[llvm.Type]struct{
 	return nil
 }

-func (c *Compiler) checkBasicBlock(bb llvm.BasicBlock, types map[llvm.Type]struct{}, specials map[llvm.TypeKind]llvm.Type) []error {
+func (c *checker) checkBasicBlock(bb llvm.BasicBlock, types map[llvm.Type]struct{}, specials map[llvm.TypeKind]llvm.Type) []error {
 	// check basic block value and type
 	var errs []error
 	if err := c.checkValue(bb.AsValue(), types, specials); err != nil {
@@ -128,7 +151,7 @@ func (c *Compiler) checkBasicBlock(bb llvm.BasicBlock, types map[llvm.Type]struc
 	return errs
 }

-func (c *Compiler) checkFunction(fn llvm.Value, types map[llvm.Type]struct{}, specials map[llvm.TypeKind]llvm.Type) []error {
+func (c *checker) checkFunction(fn llvm.Value, types map[llvm.Type]struct{}, specials map[llvm.TypeKind]llvm.Type) []error {
 	// check function value and type
 	var errs []error
 	if err := c.checkValue(fn, types, specials); err != nil {
@@ -143,26 +166,25 @@ func (c *Compiler) checkFunction(fn llvm.Value, types map[llvm.Type]struct{}, sp
 	return errs
 }

-func (c *Compiler) checkModule() []error {
+// Module checks the given module and returns a slice of error, if there are
+// any.
+func Module(mod llvm.Module) []error {
 	// check for any context mismatches
 	var errs []error
-	switch {
-	case c.mod.Context() == c.ctx:
-		// this is correct
-	case c.mod.Context() == llvm.GlobalContext():
+	c := checker{
+		ctx: mod.Context(),
+	}
+	if c.ctx == llvm.GlobalContext() {
 		// somewhere we accidentally used the global context instead of a real context
 		errs = append(errs, errors.New("module uses global context"))
-	default:
-		// we used some other context by accident
-		errs = append(errs, fmt.Errorf("module uses context %v instead of the main context %v", c.mod.Context(), c.ctx))
 	}

 	types := map[llvm.Type]struct{}{}
 	specials := map[llvm.TypeKind]llvm.Type{}
-	for fn := c.mod.FirstFunction(); !fn.IsNil(); fn = llvm.NextFunction(fn) {
+	for fn := mod.FirstFunction(); !fn.IsNil(); fn = llvm.NextFunction(fn) {
 		errs = append(errs, c.checkFunction(fn, types, specials)...)
 	}
-	for g := c.mod.FirstGlobal(); !g.IsNil(); g = llvm.NextGlobal(g) {
+	for g := mod.FirstGlobal(); !g.IsNil(); g = llvm.NextGlobal(g) {
 		if err := c.checkValue(g, types, specials); err != nil {
 			errs = append(errs, fmt.Errorf("failed to verify global %s of module: %s", g.Name(), err.Error()))
 		}
@@ -0,0 +1,48 @@
+package ircheck
+
+import (
+	"go/scanner"
+	"go/token"
+	"path/filepath"
+
+	"tinygo.org/x/go-llvm"
+)
+
+// errorAt returns an error value at the location of the instruction.
+// The location information may not be complete as it depends on debug
+// information in the IR.
+func errorAt(inst llvm.Value, msg string) scanner.Error {
+	return scanner.Error{
+		Pos: getPosition(inst),
+		Msg: msg,
+	}
+}
+
+// getPosition returns the position information for the given value, as far as
+// it is available.
+func getPosition(val llvm.Value) token.Position {
+	if !val.IsAInstruction().IsNil() {
+		loc := val.InstructionDebugLoc()
+		if loc.IsNil() {
+			return token.Position{}
+		}
+		file := loc.LocationScope().ScopeFile()
+		return token.Position{
+			Filename: filepath.Join(file.FileDirectory(), file.FileFilename()),
+			Line:     int(loc.LocationLine()),
+			Column:   int(loc.LocationColumn()),
+		}
+	} else if !val.IsAFunction().IsNil() {
+		loc := val.Subprogram()
+		if loc.IsNil() {
+			return token.Position{}
+		}
+		file := loc.ScopeFile()
+		return token.Position{
+			Filename: filepath.Join(file.FileDirectory(), file.FileFilename()),
+			Line:     int(loc.SubprogramLine()),
+		}
+	} else {
+		return token.Position{}
+	}
+}
@@ -24,39 +24,39 @@ func getUses(value llvm.Value) []llvm.Value {
 }

 // createTemporaryAlloca creates a new alloca in the entry block and adds
-// lifetime start infromation in the IR signalling that the alloca won't be used
+// lifetime start information in the IR signalling that the alloca won't be used
 // before this point.
 //
 // This is useful for creating temporary allocas for intrinsics. Don't forget to
 // end the lifetime using emitLifetimeEnd after you're done with it.
-func (c *Compiler) createTemporaryAlloca(t llvm.Type, name string) (alloca, bitcast, size llvm.Value) {
-	return llvmutil.CreateTemporaryAlloca(c.builder, c.mod, t, name)
+func (b *builder) createTemporaryAlloca(t llvm.Type, name string) (alloca, bitcast, size llvm.Value) {
+	return llvmutil.CreateTemporaryAlloca(b.Builder, b.mod, t, name)
 }

 // emitLifetimeEnd signals the end of an (alloca) lifetime by calling the
 // llvm.lifetime.end intrinsic. It is commonly used together with
 // createTemporaryAlloca.
-func (c *Compiler) emitLifetimeEnd(ptr, size llvm.Value) {
-	llvmutil.EmitLifetimeEnd(c.builder, c.mod, ptr, size)
+func (b *builder) emitLifetimeEnd(ptr, size llvm.Value) {
+	llvmutil.EmitLifetimeEnd(b.Builder, b.mod, ptr, size)
 }

 // emitPointerPack packs the list of values into a single pointer value using
 // bitcasts, or else allocates a value on the heap if it cannot be packed in the
 // pointer value directly. It returns the pointer with the packed data.
-func (c *Compiler) emitPointerPack(values []llvm.Value) llvm.Value {
-	return llvmutil.EmitPointerPack(c.builder, c.mod, c.Config, values)
+func (b *builder) emitPointerPack(values []llvm.Value) llvm.Value {
+	return llvmutil.EmitPointerPack(b.Builder, b.mod, b.NeedsStackObjects, values)
 }

 // emitPointerUnpack extracts a list of values packed using emitPointerPack.
-func (c *Compiler) emitPointerUnpack(ptr llvm.Value, valueTypes []llvm.Type) []llvm.Value {
-	return llvmutil.EmitPointerUnpack(c.builder, c.mod, ptr, valueTypes)
+func (b *builder) emitPointerUnpack(ptr llvm.Value, valueTypes []llvm.Type) []llvm.Value {
+	return llvmutil.EmitPointerUnpack(b.Builder, b.mod, ptr, valueTypes)
 }

 // makeGlobalArray creates a new LLVM global with the given name and integers as
 // contents, and returns the global.
 // Note that it is left with the default linkage etc., you should set
 // linkage/constant/etc properties yourself.
-func (c *Compiler) makeGlobalArray(buf []byte, name string, elementType llvm.Type) llvm.Value {
+func (c *compilerContext) makeGlobalArray(buf []byte, name string, elementType llvm.Type) llvm.Value {
 	globalType := llvm.ArrayType(elementType, len(buf))
 	global := llvm.AddGlobal(c.mod, globalType, name)
 	value := llvm.Undef(globalType)
@@ -5,14 +5,14 @@ package llvmutil
 // itself if possible and legal.

 import (
-	"github.com/tinygo-org/tinygo/compileopts"
 	"tinygo.org/x/go-llvm"
 )

 // EmitPointerPack packs the list of values into a single pointer value using
 // bitcasts, or else allocates a value on the heap if it cannot be packed in the
 // pointer value directly. It returns the pointer with the packed data.
-func EmitPointerPack(builder llvm.Builder, mod llvm.Module, config *compileopts.Config, values []llvm.Value) llvm.Value {
+// If the values are all constants, they are be stored in a constant global and deduplicated.
+func EmitPointerPack(builder llvm.Builder, mod llvm.Module, needsStackObjects bool, values []llvm.Value) llvm.Value {
 	ctx := mod.Context()
 	targetData := llvm.NewTargetData(mod.DataLayout())
 	i8ptrType := llvm.PointerType(mod.Context().Int8Type(), 0)
@@ -25,7 +25,6 @@ func EmitPointerPack(builder llvm.Builder, mod llvm.Module, config *compileopts.
 	packedType := ctx.StructType(valueTypes, false)

 	// Allocate memory for the packed data.
-	var packedAlloc, packedHeapAlloc llvm.Value
 	size := targetData.TypeAllocSize(packedType)
 	if size == 0 {
 		return llvm.ConstPointerNull(i8ptrType)
@@ -38,19 +37,70 @@ func EmitPointerPack(builder llvm.Builder, mod llvm.Module, config *compileopts.
 			// Try to keep this cast in SSA form.
 			return builder.CreateIntToPtr(values[0], i8ptrType, "pack.int")
 		}
+
 		// Because packedType is a struct and we have to cast it to a *i8, store
-		// it in an alloca first for bitcasting (store+bitcast+load).
-		packedAlloc, _, _ = CreateTemporaryAlloca(builder, mod, packedType, "")
+		// it in a *i8 alloca first and load the *i8 value from there. This is
+		// effectively a bitcast.
+		packedAlloc, _, _ := CreateTemporaryAlloca(builder, mod, i8ptrType, "")
+
+		if size < targetData.TypeAllocSize(i8ptrType) {
+			// The alloca is bigger than the value that will be stored in it.
+			// To avoid having some bits undefined, zero the alloca first.
+			// Hopefully this will get optimized away.
+			builder.CreateStore(llvm.ConstNull(i8ptrType), packedAlloc)
+		}
+
+		// Store all values in the alloca.
+		packedAllocCast := builder.CreateBitCast(packedAlloc, llvm.PointerType(packedType, 0), "")
+		for i, value := range values {
+			indices := []llvm.Value{
+				llvm.ConstInt(ctx.Int32Type(), 0, false),
+				llvm.ConstInt(ctx.Int32Type(), uint64(i), false),
+			}
+			gep := builder.CreateInBoundsGEP(packedAllocCast, indices, "")
+			builder.CreateStore(value, gep)
+		}
+
+		// Load value (the *i8) from the alloca.
+		result := builder.CreateLoad(packedAlloc, "")
+
+		// End the lifetime of the alloca, to help the optimizer.
+		packedPtr := builder.CreateBitCast(packedAlloc, i8ptrType, "")
+		packedSize := llvm.ConstInt(ctx.Int64Type(), targetData.TypeAllocSize(packedAlloc.Type()), false)
+		EmitLifetimeEnd(builder, mod, packedPtr, packedSize)
+
+		return result
 	} else {
+		// Check if the values are all constants.
+		constant := true
+		for _, v := range values {
+			if !v.IsConstant() {
+				constant = false
+				break
+			}
+		}
+
+		if constant {
+			// The data is known at compile time, so store it in a constant global.
+			// The global address is marked as unnamed, which allows LLVM to merge duplicates.
+			funcName := builder.GetInsertBlock().Parent().Name()
+			global := llvm.AddGlobal(mod, packedType, funcName+"$pack")
+			global.SetInitializer(ctx.ConstStruct(values, false))
+			global.SetGlobalConstant(true)
+			global.SetUnnamedAddr(true)
+			global.SetLinkage(llvm.PrivateLinkage)
+			return llvm.ConstBitCast(global, i8ptrType)
+		}
+
 		// Packed data is bigger than a pointer, so allocate it on the heap.
 		sizeValue := llvm.ConstInt(uintptrType, size, false)
 		alloc := mod.NamedFunction("runtime.alloc")
-		packedHeapAlloc = builder.CreateCall(alloc, []llvm.Value{
+		packedHeapAlloc := builder.CreateCall(alloc, []llvm.Value{
 			sizeValue,
 			llvm.Undef(i8ptrType),            // unused context parameter
 			llvm.ConstPointerNull(i8ptrType), // coroutine handle
 		}, "")
-		if config.NeedsStackObjects() {
+		if needsStackObjects {
 			trackPointer := mod.NamedFunction("runtime.trackPointer")
 			builder.CreateCall(trackPointer, []llvm.Value{
 				packedHeapAlloc,
@@ -58,28 +108,19 @@ func EmitPointerPack(builder llvm.Builder, mod llvm.Module, config *compileopts.
 				llvm.ConstPointerNull(i8ptrType), // coroutine handle
 			}, "")
 		}
-		packedAlloc = builder.CreateBitCast(packedHeapAlloc, llvm.PointerType(packedType, 0), "")
-	}
-	// Store all values in the alloca or heap pointer.
-	for i, value := range values {
-		indices := []llvm.Value{
-			llvm.ConstInt(ctx.Int32Type(), 0, false),
-			llvm.ConstInt(ctx.Int32Type(), uint64(i), false),
-		}
-		gep := builder.CreateInBoundsGEP(packedAlloc, indices, "")
-		builder.CreateStore(value, gep)
-	}
+		packedAlloc := builder.CreateBitCast(packedHeapAlloc, llvm.PointerType(packedType, 0), "")

-	if packedHeapAlloc.IsNil() {
-		// Load value (as *i8) from the alloca.
-		packedAlloc = builder.CreateBitCast(packedAlloc, llvm.PointerType(i8ptrType, 0), "")
-		result := builder.CreateLoad(packedAlloc, "")
-		packedPtr := builder.CreateBitCast(packedAlloc, i8ptrType, "")
-		packedSize := llvm.ConstInt(ctx.Int64Type(), targetData.TypeAllocSize(packedAlloc.Type()), false)
-		EmitLifetimeEnd(builder, mod, packedPtr, packedSize)
-		return result
-	} else {
-		// Get the original heap allocation pointer, which already is an *i8.
+		// Store all values in the heap pointer.
+		for i, value := range values {
+			indices := []llvm.Value{
+				llvm.ConstInt(ctx.Int32Type(), 0, false),
+				llvm.ConstInt(ctx.Int32Type(), uint64(i), false),
+			}
+			gep := builder.CreateInBoundsGEP(packedAlloc, indices, "")
+			builder.CreateStore(value, gep)
+		}
+
+		// Return the original heap allocation pointer, which already is an *i8.
 		return packedHeapAlloc
 	}
 }
@@ -10,133 +10,148 @@ import (
 	"tinygo.org/x/go-llvm"
 )

-// emitMakeMap creates a new map object (runtime.hashmap) by allocating and
+// createMakeMap creates a new map object (runtime.hashmap) by allocating and
 // initializing an appropriately sized object.
-func (c *Compiler) emitMakeMap(frame *Frame, expr *ssa.MakeMap) (llvm.Value, error) {
+func (b *builder) createMakeMap(expr *ssa.MakeMap) (llvm.Value, error) {
 	mapType := expr.Type().Underlying().(*types.Map)
 	keyType := mapType.Key().Underlying()
-	llvmValueType := c.getLLVMType(mapType.Elem().Underlying())
+	llvmValueType := b.getLLVMType(mapType.Elem().Underlying())
 	var llvmKeyType llvm.Type
 	if t, ok := keyType.(*types.Basic); ok && t.Info()&types.IsString != 0 {
 		// String keys.
-		llvmKeyType = c.getLLVMType(keyType)
+		llvmKeyType = b.getLLVMType(keyType)
 	} else if hashmapIsBinaryKey(keyType) {
 		// Trivially comparable keys.
-		llvmKeyType = c.getLLVMType(keyType)
+		llvmKeyType = b.getLLVMType(keyType)
 	} else {
 		// All other keys. Implemented as map[interface{}]valueType for ease of
 		// implementation.
-		llvmKeyType = c.getLLVMRuntimeType("_interface")
+		llvmKeyType = b.getLLVMRuntimeType("_interface")
 	}
-	keySize := c.targetData.TypeAllocSize(llvmKeyType)
-	valueSize := c.targetData.TypeAllocSize(llvmValueType)
-	llvmKeySize := llvm.ConstInt(c.ctx.Int8Type(), keySize, false)
-	llvmValueSize := llvm.ConstInt(c.ctx.Int8Type(), valueSize, false)
-	sizeHint := llvm.ConstInt(c.uintptrType, 8, false)
+	keySize := b.targetData.TypeAllocSize(llvmKeyType)
+	valueSize := b.targetData.TypeAllocSize(llvmValueType)
+	llvmKeySize := llvm.ConstInt(b.ctx.Int8Type(), keySize, false)
+	llvmValueSize := llvm.ConstInt(b.ctx.Int8Type(), valueSize, false)
+	sizeHint := llvm.ConstInt(b.uintptrType, 8, false)
 	if expr.Reserve != nil {
-		sizeHint = c.getValue(frame, expr.Reserve)
+		sizeHint = b.getValue(expr.Reserve)
 		var err error
-		sizeHint, err = c.parseConvert(expr.Reserve.Type(), types.Typ[types.Uintptr], sizeHint, expr.Pos())
+		sizeHint, err = b.createConvert(expr.Reserve.Type(), types.Typ[types.Uintptr], sizeHint, expr.Pos())
 		if err != nil {
 			return llvm.Value{}, err
 		}
 	}
-	hashmap := c.createRuntimeCall("hashmapMake", []llvm.Value{llvmKeySize, llvmValueSize, sizeHint}, "")
+	hashmap := b.createRuntimeCall("hashmapMake", []llvm.Value{llvmKeySize, llvmValueSize, sizeHint}, "")
 	return hashmap, nil
 }

-func (c *Compiler) emitMapLookup(keyType, valueType types.Type, m, key llvm.Value, commaOk bool, pos token.Pos) (llvm.Value, error) {
-	llvmValueType := c.getLLVMType(valueType)
+// createMapLookup returns the value in a map. It calls a runtime function
+// depending on the map key type to load the map value and its comma-ok value.
+func (b *builder) createMapLookup(keyType, valueType types.Type, m, key llvm.Value, commaOk bool, pos token.Pos) (llvm.Value, error) {
+	llvmValueType := b.getLLVMType(valueType)

 	// Allocate the memory for the resulting type. Do not zero this memory: it
 	// will be zeroed by the hashmap get implementation if the key is not
 	// present in the map.
-	mapValueAlloca, mapValuePtr, mapValueSize := c.createTemporaryAlloca(llvmValueType, "hashmap.value")
+	mapValueAlloca, mapValuePtr, mapValueAllocaSize := b.createTemporaryAlloca(llvmValueType, "hashmap.value")
+
+	// We need the map size (with type uintptr) to pass to the hashmap*Get
+	// functions. This is necessary because those *Get functions are valid on
+	// nil maps, and they'll need to zero the value pointer by that number of
+	// bytes.
+	mapValueSize := mapValueAllocaSize
+	if mapValueSize.Type().IntTypeWidth() > b.uintptrType.IntTypeWidth() {
+		mapValueSize = llvm.ConstTrunc(mapValueSize, b.uintptrType)
+	}

 	// Do the lookup. How it is done depends on the key type.
 	var commaOkValue llvm.Value
 	keyType = keyType.Underlying()
 	if t, ok := keyType.(*types.Basic); ok && t.Info()&types.IsString != 0 {
 		// key is a string
-		params := []llvm.Value{m, key, mapValuePtr}
-		commaOkValue = c.createRuntimeCall("hashmapStringGet", params, "")
+		params := []llvm.Value{m, key, mapValuePtr, mapValueSize}
+		commaOkValue = b.createRuntimeCall("hashmapStringGet", params, "")
 	} else if hashmapIsBinaryKey(keyType) {
 		// key can be compared with runtime.memequal
 		// Store the key in an alloca, in the entry block to avoid dynamic stack
 		// growth.
-		mapKeyAlloca, mapKeyPtr, mapKeySize := c.createTemporaryAlloca(key.Type(), "hashmap.key")
-		c.builder.CreateStore(key, mapKeyAlloca)
+		mapKeyAlloca, mapKeyPtr, mapKeySize := b.createTemporaryAlloca(key.Type(), "hashmap.key")
+		b.CreateStore(key, mapKeyAlloca)
 		// Fetch the value from the hashmap.
-		params := []llvm.Value{m, mapKeyPtr, mapValuePtr}
-		commaOkValue = c.createRuntimeCall("hashmapBinaryGet", params, "")
-		c.emitLifetimeEnd(mapKeyPtr, mapKeySize)
+		params := []llvm.Value{m, mapKeyPtr, mapValuePtr, mapValueSize}
+		commaOkValue = b.createRuntimeCall("hashmapBinaryGet", params, "")
+		b.emitLifetimeEnd(mapKeyPtr, mapKeySize)
 	} else {
 		// Not trivially comparable using memcmp. Make it an interface instead.
 		itfKey := key
 		if _, ok := keyType.(*types.Interface); !ok {
 			// Not already an interface, so convert it to an interface now.
-			itfKey = c.parseMakeInterface(key, keyType, pos)
+			itfKey = b.createMakeInterface(key, keyType, pos)
 		}
-		params := []llvm.Value{m, itfKey, mapValuePtr}
-		commaOkValue = c.createRuntimeCall("hashmapInterfaceGet", params, "")
+		params := []llvm.Value{m, itfKey, mapValuePtr, mapValueSize}
+		commaOkValue = b.createRuntimeCall("hashmapInterfaceGet", params, "")
 	}

 	// Load the resulting value from the hashmap. The value is set to the zero
 	// value if the key doesn't exist in the hashmap.
-	mapValue := c.builder.CreateLoad(mapValueAlloca, "")
-	c.emitLifetimeEnd(mapValuePtr, mapValueSize)
+	mapValue := b.CreateLoad(mapValueAlloca, "")
+	b.emitLifetimeEnd(mapValuePtr, mapValueAllocaSize)

 	if commaOk {
-		tuple := llvm.Undef(c.ctx.StructType([]llvm.Type{llvmValueType, c.ctx.Int1Type()}, false))
-		tuple = c.builder.CreateInsertValue(tuple, mapValue, 0, "")
-		tuple = c.builder.CreateInsertValue(tuple, commaOkValue, 1, "")
+		tuple := llvm.Undef(b.ctx.StructType([]llvm.Type{llvmValueType, b.ctx.Int1Type()}, false))
+		tuple = b.CreateInsertValue(tuple, mapValue, 0, "")
+		tuple = b.CreateInsertValue(tuple, commaOkValue, 1, "")
 		return tuple, nil
 	} else {
 		return mapValue, nil
 	}
 }

-func (c *Compiler) emitMapUpdate(keyType types.Type, m, key, value llvm.Value, pos token.Pos) {
-	valueAlloca, valuePtr, valueSize := c.createTemporaryAlloca(value.Type(), "hashmap.value")
-	c.builder.CreateStore(value, valueAlloca)
+// createMapUpdate updates a map key to a given value, by creating an
+// appropriate runtime call.
+func (b *builder) createMapUpdate(keyType types.Type, m, key, value llvm.Value, pos token.Pos) {
+	valueAlloca, valuePtr, valueSize := b.createTemporaryAlloca(value.Type(), "hashmap.value")
+	b.CreateStore(value, valueAlloca)
 	keyType = keyType.Underlying()
 	if t, ok := keyType.(*types.Basic); ok && t.Info()&types.IsString != 0 {
 		// key is a string
 		params := []llvm.Value{m, key, valuePtr}
-		c.createRuntimeCall("hashmapStringSet", params, "")
+		b.createRuntimeCall("hashmapStringSet", params, "")
 	} else if hashmapIsBinaryKey(keyType) {
 		// key can be compared with runtime.memequal
-		keyAlloca, keyPtr, keySize := c.createTemporaryAlloca(key.Type(), "hashmap.key")
-		c.builder.CreateStore(key, keyAlloca)
+		keyAlloca, keyPtr, keySize := b.createTemporaryAlloca(key.Type(), "hashmap.key")
+		b.CreateStore(key, keyAlloca)
 		params := []llvm.Value{m, keyPtr, valuePtr}
-		c.createRuntimeCall("hashmapBinarySet", params, "")
-		c.emitLifetimeEnd(keyPtr, keySize)
+		b.createRuntimeCall("hashmapBinarySet", params, "")
+		b.emitLifetimeEnd(keyPtr, keySize)
 	} else {
 		// Key is not trivially comparable, so compare it as an interface instead.
 		itfKey := key
 		if _, ok := keyType.(*types.Interface); !ok {
 			// Not already an interface, so convert it to an interface first.
-			itfKey = c.parseMakeInterface(key, keyType, pos)
+			itfKey = b.createMakeInterface(key, keyType, pos)
 		}
 		params := []llvm.Value{m, itfKey, valuePtr}
-		c.createRuntimeCall("hashmapInterfaceSet", params, "")
+		b.createRuntimeCall("hashmapInterfaceSet", params, "")
 	}
-	c.emitLifetimeEnd(valuePtr, valueSize)
+	b.emitLifetimeEnd(valuePtr, valueSize)
 }

-func (c *Compiler) emitMapDelete(keyType types.Type, m, key llvm.Value, pos token.Pos) error {
+// createMapDelete deletes a key from a map by calling the appropriate runtime
+// function. It is the implementation of the Go delete() builtin.
+func (b *builder) createMapDelete(keyType types.Type, m, key llvm.Value, pos token.Pos) error {
 	keyType = keyType.Underlying()
 	if t, ok := keyType.(*types.Basic); ok && t.Info()&types.IsString != 0 {
 		// key is a string
 		params := []llvm.Value{m, key}
-		c.createRuntimeCall("hashmapStringDelete", params, "")
+		b.createRuntimeCall("hashmapStringDelete", params, "")
 		return nil
 	} else if hashmapIsBinaryKey(keyType) {
-		keyAlloca, keyPtr, keySize := c.createTemporaryAlloca(key.Type(), "hashmap.key")
-		c.builder.CreateStore(key, keyAlloca)
+		keyAlloca, keyPtr, keySize := b.createTemporaryAlloca(key.Type(), "hashmap.key")
+		b.CreateStore(key, keyAlloca)
 		params := []llvm.Value{m, keyPtr}
-		c.createRuntimeCall("hashmapBinaryDelete", params, "")
-		c.emitLifetimeEnd(keyPtr, keySize)
+		b.createRuntimeCall("hashmapBinaryDelete", params, "")
+		b.emitLifetimeEnd(keyPtr, keySize)
 		return nil
 	} else {
 		// Key is not trivially comparable, so compare it as an interface
@@ -144,10 +159,10 @@ func (c *Compiler) emitMapDelete(keyType types.Type, m, key llvm.Value, pos toke
 		itfKey := key
 		if _, ok := keyType.(*types.Interface); !ok {
 			// Not already an interface, so convert it to an interface first.
-			itfKey = c.parseMakeInterface(key, keyType, pos)
+			itfKey = b.createMakeInterface(key, keyType, pos)
 		}
 		params := []llvm.Value{m, itfKey}
-		c.createRuntimeCall("hashmapInterfaceDelete", params, "")
+		b.createRuntimeCall("hashmapInterfaceDelete", params, "")
 		return nil
 	}
 }
@@ -1,164 +0,0 @@
-package compiler
-
-import (
-	"errors"
-
-	"github.com/tinygo-org/tinygo/transform"
-	"tinygo.org/x/go-llvm"
-)
-
-// Run the LLVM optimizer over the module.
-// The inliner can be disabled (if necessary) by passing 0 to the inlinerThreshold.
-func (c *Compiler) Optimize(optLevel, sizeLevel int, inlinerThreshold uint) []error {
-	builder := llvm.NewPassManagerBuilder()
-	defer builder.Dispose()
-	builder.SetOptLevel(optLevel)
-	builder.SetSizeLevel(sizeLevel)
-	if inlinerThreshold != 0 {
-		builder.UseInlinerWithThreshold(inlinerThreshold)
-	}
-	builder.AddCoroutinePassesToExtensionPoints()
-
-	if c.PanicStrategy() == "trap" {
-		transform.ReplacePanicsWithTrap(c.mod) // -panic=trap
-	}
-
-	// run a check of all of our code
-	if c.VerifyIR() {
-		errs := c.checkModule()
-		if errs != nil {
-			return errs
-		}
-	}
-
-	// Run function passes for each function.
-	funcPasses := llvm.NewFunctionPassManagerForModule(c.mod)
-	defer funcPasses.Dispose()
-	builder.PopulateFunc(funcPasses)
-	funcPasses.InitializeFunc()
-	for fn := c.mod.FirstFunction(); !fn.IsNil(); fn = llvm.NextFunction(fn) {
-		funcPasses.RunFunc(fn)
-	}
-	funcPasses.FinalizeFunc()
-
-	if optLevel > 0 {
-		// Run some preparatory passes for the Go optimizer.
-		goPasses := llvm.NewPassManager()
-		defer goPasses.Dispose()
-		goPasses.AddGlobalDCEPass()
-		goPasses.AddGlobalOptimizerPass()
-		goPasses.AddConstantPropagationPass()
-		goPasses.AddAggressiveDCEPass()
-		goPasses.AddFunctionAttrsPass()
-		goPasses.Run(c.mod)
-
-		// Run Go-specific optimization passes.
-		transform.OptimizeMaps(c.mod)
-		transform.OptimizeStringToBytes(c.mod)
-		transform.OptimizeAllocs(c.mod)
-		transform.LowerInterfaces(c.mod)
-
-		errs := transform.LowerInterrupts(c.mod)
-		if len(errs) > 0 {
-			return errs
-		}
-
-		if c.funcImplementation() == funcValueSwitch {
-			transform.LowerFuncValues(c.mod)
-		}
-
-		// After interfaces are lowered, there are many more opportunities for
-		// interprocedural optimizations. To get them to work, function
-		// attributes have to be updated first.
-		goPasses.Run(c.mod)
-
-		// Run TinyGo-specific interprocedural optimizations.
-		transform.OptimizeAllocs(c.mod)
-		transform.OptimizeStringToBytes(c.mod)
-
-		// Lower runtime.isnil calls to regular nil comparisons.
-		isnil := c.mod.NamedFunction("runtime.isnil")
-		if !isnil.IsNil() {
-			for _, use := range getUses(isnil) {
-				c.builder.SetInsertPointBefore(use)
-				ptr := use.Operand(0)
-				if !ptr.IsABitCastInst().IsNil() {
-					ptr = ptr.Operand(0)
-				}
-				nilptr := llvm.ConstPointerNull(ptr.Type())
-				icmp := c.builder.CreateICmp(llvm.IntEQ, ptr, nilptr, "")
-				use.ReplaceAllUsesWith(icmp)
-				use.EraseFromParentAsInstruction()
-			}
-		}
-
-		err := c.LowerGoroutines()
-		if err != nil {
-			return []error{err}
-		}
-	} else {
-		// Must be run at any optimization level.
-		transform.LowerInterfaces(c.mod)
-		if c.funcImplementation() == funcValueSwitch {
-			transform.LowerFuncValues(c.mod)
-		}
-		err := c.LowerGoroutines()
-		if err != nil {
-			return []error{err}
-		}
-		errs := transform.LowerInterrupts(c.mod)
-		if len(errs) > 0 {
-			return errs
-		}
-	}
-	if c.VerifyIR() {
-		if errs := c.checkModule(); errs != nil {
-			return errs
-		}
-	}
-	if err := c.Verify(); err != nil {
-		return []error{errors.New("optimizations caused a verification failure")}
-	}
-
-	if sizeLevel >= 2 {
-		// Set the "optsize" attribute to make slightly smaller binaries at the
-		// cost of some performance.
-		kind := llvm.AttributeKindID("optsize")
-		attr := c.ctx.CreateEnumAttribute(kind, 0)
-		for fn := c.mod.FirstFunction(); !fn.IsNil(); fn = llvm.NextFunction(fn) {
-			fn.AddFunctionAttr(attr)
-		}
-	}
-
-	// After TinyGo-specific transforms have finished, undo exporting these functions.
-	for _, name := range c.getFunctionsUsedInTransforms() {
-		fn := c.mod.NamedFunction(name)
-		if fn.IsNil() {
-			continue
-		}
-		fn.SetLinkage(llvm.InternalLinkage)
-	}
-
-	// Run function passes again, because without it, llvm.coro.size.i32()
-	// doesn't get lowered.
-	for fn := c.mod.FirstFunction(); !fn.IsNil(); fn = llvm.NextFunction(fn) {
-		funcPasses.RunFunc(fn)
-	}
-	funcPasses.FinalizeFunc()
-
-	// Run module passes.
-	modPasses := llvm.NewPassManager()
-	defer modPasses.Dispose()
-	builder.Populate(modPasses)
-	modPasses.Run(c.mod)
-
-	hasGCPass := transform.AddGlobalsBitmap(c.mod)
-	hasGCPass = transform.MakeGCStackSlots(c.mod) || hasGCPass
-	if hasGCPass {
-		if err := c.Verify(); err != nil {
-			return []error{errors.New("GC pass caused a verification failure")}
-		}
-	}
-
-	return nil
-}
@@ -0,0 +1,164 @@
+package compiler
+
+// This file implements a simple reachability analysis, to reduce compile time.
+// This DCE pass used to be necessary for improving other passes but now it
+// isn't necessary anymore.
+
+import (
+	"errors"
+	"go/types"
+	"sort"
+
+	"github.com/tinygo-org/tinygo/loader"
+	"golang.org/x/tools/go/ssa"
+)
+
+type dceState struct {
+	*compilerContext
+	functions   []*dceFunction
+	functionMap map[*ssa.Function]*dceFunction
+}
+
+type dceFunction struct {
+	*ssa.Function
+	functionInfo
+	flag bool // used by dead code elimination
+}
+
+func (p *dceState) addFunction(ssaFn *ssa.Function) {
+	if _, ok := p.functionMap[ssaFn]; ok {
+		return
+	}
+	f := &dceFunction{Function: ssaFn}
+	f.functionInfo = p.getFunctionInfo(ssaFn)
+	p.functions = append(p.functions, f)
+	p.functionMap[ssaFn] = f
+
+	for _, anon := range ssaFn.AnonFuncs {
+		p.addFunction(anon)
+	}
+}
+
+// simpleDCE returns a list of alive functions in the program. Compiling only
+// these functions makes the compiler faster.
+//
+// This functionality will likely be replaced in the future with build caching.
+func (c *compilerContext) simpleDCE(lprogram *loader.Program) ([]*ssa.Function, error) {
+	mainPkg := c.program.Package(lprogram.MainPkg().Pkg)
+	if mainPkg == nil {
+		panic("could not find main package")
+	}
+	p := &dceState{
+		compilerContext: c,
+		functionMap:     make(map[*ssa.Function]*dceFunction),
+	}
+
+	for _, pkg := range lprogram.Sorted() {
+		pkg := c.program.Package(pkg.Pkg)
+		memberNames := make([]string, 0)
+		for name := range pkg.Members {
+			memberNames = append(memberNames, name)
+		}
+		sort.Strings(memberNames)
+
+		for _, name := range memberNames {
+			member := pkg.Members[name]
+			switch member := member.(type) {
+			case *ssa.Function:
+				p.addFunction(member)
+			case *ssa.Type:
+				methods := getAllMethods(pkg.Prog, member.Type())
+				if !types.IsInterface(member.Type()) {
+					// named type
+					for _, method := range methods {
+						p.addFunction(pkg.Prog.MethodValue(method))
+					}
+				}
+			case *ssa.Global:
+				// Ignore. Globals are not handled here.
+			case *ssa.NamedConst:
+				// Ignore: these are already resolved.
+			default:
+				panic("unknown member type: " + member.String())
+			}
+		}
+	}
+
+	// Initial set of live functions. Include main.main, *.init and runtime.*
+	// functions.
+	main, ok := mainPkg.Members["main"].(*ssa.Function)
+	if !ok {
+		if mainPkg.Members["main"] == nil {
+			return nil, errors.New("function main is undeclared in the main package")
+		} else {
+			return nil, errors.New("cannot declare main - must be func")
+		}
+	}
+	runtimePkg := c.program.ImportedPackage("runtime")
+	mathPkg := c.program.ImportedPackage("math")
+	taskPkg := c.program.ImportedPackage("internal/task")
+	p.functionMap[main].flag = true
+	worklist := []*ssa.Function{main}
+	for _, f := range p.functions {
+		if f.exported || f.Synthetic == "package initializer" || f.Pkg == runtimePkg || f.Pkg == taskPkg || (f.Pkg == mathPkg && f.Pkg != nil) {
+			if f.flag {
+				continue
+			}
+			f.flag = true
+			worklist = append(worklist, f.Function)
+		}
+	}
+
+	// Mark all called functions recursively.
+	for len(worklist) != 0 {
+		f := worklist[len(worklist)-1]
+		worklist = worklist[:len(worklist)-1]
+		for _, block := range f.Blocks {
+			for _, instr := range block.Instrs {
+				if instr, ok := instr.(*ssa.MakeInterface); ok {
+					for _, sel := range getAllMethods(c.program, instr.X.Type()) {
+						fn := c.program.MethodValue(sel)
+						callee := p.functionMap[fn]
+						if callee == nil {
+							// TODO: why is this necessary?
+							p.addFunction(fn)
+							callee = p.functionMap[fn]
+						}
+						if !callee.flag {
+							callee.flag = true
+							worklist = append(worklist, callee.Function)
+						}
+					}
+				}
+				for _, operand := range instr.Operands(nil) {
+					if operand == nil || *operand == nil {
+						continue
+					}
+					switch operand := (*operand).(type) {
+					case *ssa.Function:
+						f := p.functionMap[operand]
+						if f == nil {
+							// FIXME HACK: this function should have been
+							// discovered already. It is not for bound methods.
+							p.addFunction(operand)
+							f = p.functionMap[operand]
+						}
+						if !f.flag {
+							f.flag = true
+							worklist = append(worklist, operand)
+						}
+					}
+				}
+			}
+		}
+	}
+
+	// Return all live functions.
+	liveFunctions := []*ssa.Function{}
+	for _, f := range p.functions {
+		if f.flag {
+			liveFunctions = append(liveFunctions, f.Function)
+		}
+	}
+	return liveFunctions, nil
+}
@@ -10,13 +10,13 @@ import (
 // The original license can be found here:
 //     https://golang.org/LICENSE

-type StdSizes struct {
+type stdSizes struct {
 	IntSize  int64
 	PtrSize  int64
 	MaxAlign int64
 }

-func (s *StdSizes) Alignof(T types.Type) int64 {
+func (s *stdSizes) Alignof(T types.Type) int64 {
 	// For arrays and structs, alignment is defined in terms
 	// of alignment of the elements and fields, respectively.
 	switch t := T.Underlying().(type) {
@@ -61,7 +61,7 @@ func (s *StdSizes) Alignof(T types.Type) int64 {
 	return a
 }

-func (s *StdSizes) Offsetsof(fields []*types.Var) []int64 {
+func (s *stdSizes) Offsetsof(fields []*types.Var) []int64 {
 	offsets := make([]int64, len(fields))
 	var o int64
 	for i, f := range fields {
@@ -89,7 +89,7 @@ var basicSizes = [...]byte{
 	types.Complex128: 16,
 }

-func (s *StdSizes) Sizeof(T types.Type) int64 {
+func (s *stdSizes) Sizeof(T types.Type) int64 {
 	switch t := T.Underlying().(type) {
 	case *types.Basic:
 		k := t.Kind()
@@ -15,6 +15,269 @@ import (
 	"tinygo.org/x/go-llvm"
 )

+// functionInfo contains some information about a function or method. In
+// particular, it contains information obtained from pragmas.
+//
+// The linkName value contains a valid link name, even if //go:linkname is not
+// present.
+type functionInfo struct {
+	module     string     // go:wasm-module
+	importName string     // go:linkname, go:export - The name the developer assigns
+	linkName   string     // go:linkname, go:export - The name that we map for the particular module -> importName
+	exported   bool       // go:export, CGo
+	nobounds   bool       // go:nobounds
+	variadic   bool       // go:variadic (CGo only)
+	inline     inlineType // go:inline
+}
+
+type inlineType int
+
+// How much to inline.
+const (
+	// Default behavior. The compiler decides for itself whether any given
+	// function will be inlined. Whether any function is inlined depends on the
+	// optimization level.
+	inlineDefault inlineType = iota
+
+	// Inline hint, just like the C inline keyword (signalled using
+	// //go:inline). The compiler will be more likely to inline this function,
+	// but it is not a guarantee.
+	inlineHint
+
+	// Don't inline, just like the GCC noinline attribute. Signalled using
+	// //go:noinline.
+	inlineNone
+)
+
+// getFunction returns the LLVM function for the given *ssa.Function, creating
+// it if needed. It can later be filled with compilerContext.createFunction().
+func (c *compilerContext) getFunction(fn *ssa.Function) llvm.Value {
+	info := c.getFunctionInfo(fn)
+	llvmFn := c.mod.NamedFunction(info.linkName)
+	if !llvmFn.IsNil() {
+		return llvmFn
+	}
+
+	var retType llvm.Type
+	if fn.Signature.Results() == nil {
+		retType = c.ctx.VoidType()
+	} else if fn.Signature.Results().Len() == 1 {
+		retType = c.getLLVMType(fn.Signature.Results().At(0).Type())
+	} else {
+		results := make([]llvm.Type, 0, fn.Signature.Results().Len())
+		for i := 0; i < fn.Signature.Results().Len(); i++ {
+			results = append(results, c.getLLVMType(fn.Signature.Results().At(i).Type()))
+		}
+		retType = c.ctx.StructType(results, false)
+	}
+
+	var paramInfos []paramInfo
+	for _, param := range fn.Params {
+		paramType := c.getLLVMType(param.Type())
+		paramFragmentInfos := expandFormalParamType(paramType, param.Name(), param.Type())
+		paramInfos = append(paramInfos, paramFragmentInfos...)
+	}
+
+	// Add an extra parameter as the function context. This context is used in
+	// closures and bound methods, but should be optimized away when not used.
+	if !info.exported {
+		paramInfos = append(paramInfos, paramInfo{llvmType: c.i8ptrType, name: "context", flags: 0})
+		paramInfos = append(paramInfos, paramInfo{llvmType: c.i8ptrType, name: "parentHandle", flags: 0})
+	}
+
+	var paramTypes []llvm.Type
+	for _, info := range paramInfos {
+		paramTypes = append(paramTypes, info.llvmType)
+	}
+
+	fnType := llvm.FunctionType(retType, paramTypes, info.variadic)
+	llvmFn = llvm.AddFunction(c.mod, info.linkName, fnType)
+	if strings.HasPrefix(c.Triple, "wasm") {
+		// C functions without prototypes like this:
+		//   void foo();
+		// are actually variadic functions. However, it appears that it has been
+		// decided in WebAssembly that such prototype-less functions are not
+		// allowed in WebAssembly.
+		// In C, this can only happen when there are zero parameters, hence this
+		// check here. For more information:
+		// https://reviews.llvm.org/D48443
+		// https://github.com/WebAssembly/tool-conventions/issues/16
+		if info.variadic && len(fn.Params) == 0 {
+			attr := c.ctx.CreateStringAttribute("no-prototype", "")
+			llvmFn.AddFunctionAttr(attr)
+		}
+	}
+
+	dereferenceableOrNullKind := llvm.AttributeKindID("dereferenceable_or_null")
+	for i, info := range paramInfos {
+		if info.flags&paramIsDeferenceableOrNull == 0 {
+			continue
+		}
+		if info.llvmType.TypeKind() == llvm.PointerTypeKind {
+			el := info.llvmType.ElementType()
+			size := c.targetData.TypeAllocSize(el)
+			if size == 0 {
+				// dereferenceable_or_null(0) appears to be illegal in LLVM.
+				continue
+			}
+			dereferenceableOrNull := c.ctx.CreateEnumAttribute(dereferenceableOrNullKind, size)
+			llvmFn.AddAttributeAtIndex(i+1, dereferenceableOrNull)
+		}
+	}
+
+	// Set a number of function or parameter attributes, depending on the
+	// function. These functions are runtime functions that are known to have
+	// certain attributes that might not be inferred by the compiler.
+	switch info.linkName {
+	case "abort":
+		// On *nix systems, the "abort" functuion in libc is used to handle fatal panics.
+		// Mark it as noreturn so LLVM can optimize away code.
+		llvmFn.AddFunctionAttr(c.ctx.CreateEnumAttribute(llvm.AttributeKindID("noreturn"), 0))
+	case "runtime.alloc":
+		// Tell the optimizer that runtime.alloc is an allocator, meaning that it
+		// returns values that are never null and never alias to an existing value.
+		for _, attrName := range []string{"noalias", "nonnull"} {
+			llvmFn.AddAttributeAtIndex(0, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(attrName), 0))
+		}
+	case "runtime.trackPointer":
+		// This function is necessary for tracking pointers on the stack in a
+		// portable way (see gc_stack_portable.go). Indicate to the optimizer
+		// that the only thing we'll do is read the pointer.
+		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("nocapture"), 0))
+		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("readonly"), 0))
+	}
+
+	// External/exported functions may not retain pointer values.
+	// https://golang.org/cmd/cgo/#hdr-Passing_pointers
+	if info.exported {
+		// Set the wasm-import-module attribute if the function's module is set.
+		if info.module != "" {
+
+			// We need to add the wasm-import-module and the wasm-import-name
+			wasmImportModuleAttr := c.ctx.CreateStringAttribute("wasm-import-module", info.module)
+			llvmFn.AddFunctionAttr(wasmImportModuleAttr)
+
+			// Add the Wasm Import Name, if we are a named wasm import
+			if info.importName != "" {
+				wasmImportNameAttr := c.ctx.CreateStringAttribute("wasm-import-name", info.importName)
+				llvmFn.AddFunctionAttr(wasmImportNameAttr)
+			}
+		}
+		nocaptureKind := llvm.AttributeKindID("nocapture")
+		nocapture := c.ctx.CreateEnumAttribute(nocaptureKind, 0)
+		for i, typ := range paramTypes {
+			if typ.TypeKind() == llvm.PointerTypeKind {
+				llvmFn.AddAttributeAtIndex(i+1, nocapture)
+			}
+		}
+	}
+
+	return llvmFn
+}
+
+// getFunctionInfo returns information about a function that is not directly
+// present in *ssa.Function, such as the link name and whether it should be
+// exported.
+func (c *compilerContext) getFunctionInfo(f *ssa.Function) functionInfo {
+	info := functionInfo{}
+	if strings.HasPrefix(f.Name(), "C.") {
+		// Created by CGo: such a name cannot be created by regular C code.
+		info.linkName = f.Name()[2:]
+		info.exported = true
+	} else {
+		// Pick the default linkName.
+		info.linkName = f.RelString(nil)
+	}
+	// Check for //go: pragmas, which may change the link name (among others).
+	info.parsePragmas(f)
+	return info
+}
+
+// parsePragmas is used by getFunctionInfo to parse function pragmas such as
+// //export or //go:noinline.
+func (info *functionInfo) parsePragmas(f *ssa.Function) {
+	if f.Syntax() == nil {
+		return
+	}
+	if decl, ok := f.Syntax().(*ast.FuncDecl); ok && decl.Doc != nil {
+
+		// Our importName for a wasm module (if we are compiling to wasm), or llvm link name
+		var importName string
+
+		for _, comment := range decl.Doc.List {
+			text := comment.Text
+			if strings.HasPrefix(text, "//export ") {
+				// Rewrite '//export' to '//go:export' for compatibility with
+				// gc.
+				text = "//go:" + text[2:]
+			}
+			if !strings.HasPrefix(text, "//go:") {
+				continue
+			}
+			parts := strings.Fields(text)
+			switch parts[0] {
+			case "//go:export":
+				if len(parts) != 2 {
+					continue
+				}
+
+				importName = parts[1]
+				info.exported = true
+			case "//go:wasm-module":
+				// Alternative comment for setting the import module.
+				if len(parts) != 2 {
+					continue
+				}
+				info.module = parts[1]
+			case "//go:inline":
+				info.inline = inlineHint
+			case "//go:noinline":
+				info.inline = inlineNone
+			case "//go:linkname":
+				if len(parts) != 3 || parts[1] != f.Name() {
+					continue
+				}
+				// Only enable go:linkname when the package imports "unsafe".
+				// This is a slightly looser requirement than what gc uses: gc
+				// requires the file to import "unsafe", not the package as a
+				// whole.
+				if hasUnsafeImport(f.Pkg.Pkg) {
+					info.linkName = parts[2]
+				}
+			case "//go:nobounds":
+				// Skip bounds checking in this function. Useful for some
+				// runtime functions.
+				// This is somewhat dangerous and thus only imported in packages
+				// that import unsafe.
+				if hasUnsafeImport(f.Pkg.Pkg) {
+					info.nobounds = true
+				}
+			case "//go:variadic":
+				// The //go:variadic pragma is emitted by the CGo preprocessing
+				// pass for C variadic functions. This includes both explicit
+				// (with ...) and implicit (no parameters in signature)
+				// functions.
+				if strings.HasPrefix(f.Name(), "C.") {
+					// This prefix cannot naturally be created, it must have
+					// been created as a result of CGo preprocessing.
+					info.variadic = true
+				}
+			}
+		}
+
+		// Set the importName for our exported function if we have one
+		if importName != "" {
+			if info.module == "" {
+				info.linkName = importName
+			} else {
+				// WebAssembly import
+				info.importName = importName
+			}
+		}
+
+	}
+}
+
 // globalInfo contains some information about a specific global. By default,
 // linkName is equal to .RelString(nil) on a global and extern is false, but for
 // some symbols this is different (due to //go:extern for example).
@@ -26,7 +289,7 @@ type globalInfo struct {

 // loadASTComments loads comments on globals from the AST, for use later in the
 // program. In particular, they are required for //go:extern pragmas on globals.
-func (c *Compiler) loadASTComments(lprogram *loader.Program) {
+func (c *compilerContext) loadASTComments(lprogram *loader.Program) {
 	c.astComments = map[string]*ast.CommentGroup{}
 	for _, pkgInfo := range lprogram.Sorted() {
 		for _, file := range pkgInfo.Files {
@@ -56,7 +319,7 @@ func (c *Compiler) loadASTComments(lprogram *loader.Program) {

 // getGlobal returns a LLVM IR global value for a Go SSA global. It is added to
 // the LLVM IR if it has not been added already.
-func (c *Compiler) getGlobal(g *ssa.Global) llvm.Value {
+func (c *compilerContext) getGlobal(g *ssa.Global) llvm.Value {
 	info := c.getGlobalInfo(g)
 	llvmGlobal := c.mod.NamedGlobal(info.linkName)
 	if llvmGlobal.IsNil() {
@@ -82,11 +345,11 @@ func (c *Compiler) getGlobal(g *ssa.Global) llvm.Value {
 			}
 		}

-		if c.Debug() && !info.extern {
+		if c.Debug && !info.extern {
 			// Add debug info.
 			// TODO: this should be done for every global in the program, not just
 			// the ones that are referenced from some code.
-			pos := c.ir.Program.Fset.Position(g.Pos())
+			pos := c.program.Fset.Position(g.Pos())
 			diglobal := c.dibuilder.CreateGlobalVariableExpression(c.difiles[pos.Filename], llvm.DIGlobalVariableExpression{
 				Name:        g.RelString(nil),
 				LinkageName: info.linkName,
@@ -104,7 +367,7 @@ func (c *Compiler) getGlobal(g *ssa.Global) llvm.Value {
 }

 // getGlobalInfo returns some information about a specific global.
-func (c *Compiler) getGlobalInfo(g *ssa.Global) globalInfo {
+func (c *compilerContext) getGlobalInfo(g *ssa.Global) globalInfo {
 	info := globalInfo{}
 	if strings.HasPrefix(g.Name(), "C.") {
 		// Created by CGo: such a name cannot be created by regular C code.
@@ -145,3 +408,23 @@ func (info *globalInfo) parsePragmas(doc *ast.CommentGroup) {
 		}
 	}
 }
+
+// Get all methods of a type.
+func getAllMethods(prog *ssa.Program, typ types.Type) []*types.Selection {
+	ms := prog.MethodSets.MethodSet(typ)
+	methods := make([]*types.Selection, ms.Len())
+	for i := 0; i < ms.Len(); i++ {
+		methods[i] = ms.At(i)
+	}
+	return methods
+}
+
+// Return true if this package imports "unsafe", false otherwise.
+func hasUnsafeImport(pkg *types.Package) bool {
+	for _, imp := range pkg.Imports() {
+		if imp == types.Unsafe {
+			return true
+		}
+	}
+	return false
+}
@@ -10,14 +10,14 @@ import (
 	"tinygo.org/x/go-llvm"
 )

-// emitSyscall emits an inline system call instruction, depending on the target
-// OS/arch.
-func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value, error) {
-	num := c.getValue(frame, call.Args[0])
+// createSyscall emits an inline system call instruction, depending on the
+// target OS/arch.
+func (b *builder) createSyscall(call *ssa.CallCommon) (llvm.Value, error) {
+	num := b.getValue(call.Args[0])
 	var syscallResult llvm.Value
 	switch {
-	case c.GOARCH() == "amd64":
-		if c.GOOS() == "darwin" {
+	case b.GOARCH == "amd64":
+		if b.GOOS == "darwin" {
 			// Darwin adds this magic number to system call numbers:
 			//
 			// > Syscall classes for 64-bit system call entry.
@@ -28,13 +28,13 @@ func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value,
 			// > All system classes enter the kernel via the syscall instruction.
 			//
 			// Source: https://opensource.apple.com/source/xnu/xnu-792.13.8/osfmk/mach/i386/syscall_sw.h
-			num = c.builder.CreateOr(num, llvm.ConstInt(c.uintptrType, 0x2000000, false), "")
+			num = b.CreateOr(num, llvm.ConstInt(b.uintptrType, 0x2000000, false), "")
 		}
 		// Sources:
 		//   https://stackoverflow.com/a/2538212
 		//   https://en.wikibooks.org/wiki/X86_Assembly/Interfacing_with_Linux#syscall
 		args := []llvm.Value{num}
-		argTypes := []llvm.Type{c.uintptrType}
+		argTypes := []llvm.Type{b.uintptrType}
 		// Constraints will look something like:
 		//   "={rax},0,{rdi},{rsi},{rdx},{r10},{r8},{r9},~{rcx},~{r11}"
 		constraints := "={rax},0"
@@ -50,21 +50,21 @@ func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value,
 				"{r12}",
 				"{r13}",
 			}[i]
-			llvmValue := c.getValue(frame, arg)
+			llvmValue := b.getValue(arg)
 			args = append(args, llvmValue)
 			argTypes = append(argTypes, llvmValue.Type())
 		}
 		constraints += ",~{rcx},~{r11}"
-		fnType := llvm.FunctionType(c.uintptrType, argTypes, false)
+		fnType := llvm.FunctionType(b.uintptrType, argTypes, false)
 		target := llvm.InlineAsm(fnType, "syscall", constraints, true, false, llvm.InlineAsmDialectIntel)
-		syscallResult = c.builder.CreateCall(target, args, "")
-	case c.GOARCH() == "386" && c.GOOS() == "linux":
+		syscallResult = b.CreateCall(target, args, "")
+	case b.GOARCH == "386" && b.GOOS == "linux":
 		// Sources:
 		//   syscall(2) man page
 		//   https://stackoverflow.com/a/2538212
 		//   https://en.wikibooks.org/wiki/X86_Assembly/Interfacing_with_Linux#int_0x80
 		args := []llvm.Value{num}
-		argTypes := []llvm.Type{c.uintptrType}
+		argTypes := []llvm.Type{b.uintptrType}
 		// Constraints will look something like:
 		//   "={eax},0,{ebx},{ecx},{edx},{esi},{edi},{ebp}"
 		constraints := "={eax},0"
@@ -77,14 +77,14 @@ func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value,
 				"{edi}",
 				"{ebp}",
 			}[i]
-			llvmValue := c.getValue(frame, arg)
+			llvmValue := b.getValue(arg)
 			args = append(args, llvmValue)
 			argTypes = append(argTypes, llvmValue.Type())
 		}
-		fnType := llvm.FunctionType(c.uintptrType, argTypes, false)
+		fnType := llvm.FunctionType(b.uintptrType, argTypes, false)
 		target := llvm.InlineAsm(fnType, "int 0x80", constraints, true, false, llvm.InlineAsmDialectIntel)
-		syscallResult = c.builder.CreateCall(target, args, "")
-	case c.GOARCH() == "arm" && c.GOOS() == "linux":
+		syscallResult = b.CreateCall(target, args, "")
+	case b.GOARCH == "arm" && b.GOOS == "linux":
 		// Implement the EABI system call convention for Linux.
 		// Source: syscall(2) man page.
 		args := []llvm.Value{}
@@ -102,21 +102,21 @@ func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value,
 				"{r5}",
 				"{r6}",
 			}[i]
-			llvmValue := c.getValue(frame, arg)
+			llvmValue := b.getValue(arg)
 			args = append(args, llvmValue)
 			argTypes = append(argTypes, llvmValue.Type())
 		}
 		args = append(args, num)
-		argTypes = append(argTypes, c.uintptrType)
+		argTypes = append(argTypes, b.uintptrType)
 		constraints += ",{r7}" // syscall number
 		for i := len(call.Args) - 1; i < 4; i++ {
 			// r0-r3 get clobbered after the syscall returns
 			constraints += ",~{r" + strconv.Itoa(i) + "}"
 		}
-		fnType := llvm.FunctionType(c.uintptrType, argTypes, false)
+		fnType := llvm.FunctionType(b.uintptrType, argTypes, false)
 		target := llvm.InlineAsm(fnType, "svc #0", constraints, true, false, 0)
-		syscallResult = c.builder.CreateCall(target, args, "")
-	case c.GOARCH() == "arm64" && c.GOOS() == "linux":
+		syscallResult = b.CreateCall(target, args, "")
+	case b.GOARCH == "arm64" && b.GOOS == "linux":
 		// Source: syscall(2) man page.
 		args := []llvm.Value{}
 		argTypes := []llvm.Type{}
@@ -132,12 +132,12 @@ func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value,
 				"{x4}",
 				"{x5}",
 			}[i]
-			llvmValue := c.getValue(frame, arg)
+			llvmValue := b.getValue(arg)
 			args = append(args, llvmValue)
 			argTypes = append(argTypes, llvmValue.Type())
 		}
 		args = append(args, num)
-		argTypes = append(argTypes, c.uintptrType)
+		argTypes = append(argTypes, b.uintptrType)
 		constraints += ",{x8}" // syscall number
 		for i := len(call.Args) - 1; i < 8; i++ {
 			// x0-x7 may get clobbered during the syscall following the aarch64
@@ -145,13 +145,13 @@ func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value,
 			constraints += ",~{x" + strconv.Itoa(i) + "}"
 		}
 		constraints += ",~{x16},~{x17}" // scratch registers
-		fnType := llvm.FunctionType(c.uintptrType, argTypes, false)
+		fnType := llvm.FunctionType(b.uintptrType, argTypes, false)
 		target := llvm.InlineAsm(fnType, "svc #0", constraints, true, false, 0)
-		syscallResult = c.builder.CreateCall(target, args, "")
+		syscallResult = b.CreateCall(target, args, "")
 	default:
-		return llvm.Value{}, c.makeError(call.Pos(), "unknown GOOS/GOARCH for syscall: "+c.GOOS()+"/"+c.GOARCH())
+		return llvm.Value{}, b.makeError(call.Pos(), "unknown GOOS/GOARCH for syscall: "+b.GOOS+"/"+b.GOARCH)
 	}
-	switch c.GOOS() {
+	switch b.GOOS {
 	case "linux", "freebsd":
 		// Return values: r0, r1 uintptr, err Errno
 		// Pseudocode:
@@ -160,15 +160,15 @@ func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value,
 		//         err = -syscallResult
 		//     }
 		//     return syscallResult, 0, err
-		zero := llvm.ConstInt(c.uintptrType, 0, false)
-		inrange1 := c.builder.CreateICmp(llvm.IntSLT, syscallResult, llvm.ConstInt(c.uintptrType, 0, false), "")
-		inrange2 := c.builder.CreateICmp(llvm.IntSGT, syscallResult, llvm.ConstInt(c.uintptrType, 0xfffffffffffff000, true), "") // -4096
-		hasError := c.builder.CreateAnd(inrange1, inrange2, "")
-		errResult := c.builder.CreateSelect(hasError, c.builder.CreateSub(zero, syscallResult, ""), zero, "syscallError")
-		retval := llvm.Undef(c.ctx.StructType([]llvm.Type{c.uintptrType, c.uintptrType, c.uintptrType}, false))
-		retval = c.builder.CreateInsertValue(retval, syscallResult, 0, "")
-		retval = c.builder.CreateInsertValue(retval, zero, 1, "")
-		retval = c.builder.CreateInsertValue(retval, errResult, 2, "")
+		zero := llvm.ConstInt(b.uintptrType, 0, false)
+		inrange1 := b.CreateICmp(llvm.IntSLT, syscallResult, llvm.ConstInt(b.uintptrType, 0, false), "")
+		inrange2 := b.CreateICmp(llvm.IntSGT, syscallResult, llvm.ConstInt(b.uintptrType, 0xfffffffffffff000, true), "") // -4096
+		hasError := b.CreateAnd(inrange1, inrange2, "")
+		errResult := b.CreateSelect(hasError, b.CreateSub(zero, syscallResult, ""), zero, "syscallError")
+		retval := llvm.Undef(b.ctx.StructType([]llvm.Type{b.uintptrType, b.uintptrType, b.uintptrType}, false))
+		retval = b.CreateInsertValue(retval, syscallResult, 0, "")
+		retval = b.CreateInsertValue(retval, zero, 1, "")
+		retval = b.CreateInsertValue(retval, errResult, 2, "")
 		return retval, nil
 	case "darwin":
 		// Return values: r0, r1 uintptr, err Errno
@@ -178,15 +178,15 @@ func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value,
 		//         err = syscallResult
 		//     }
 		//     return syscallResult, 0, err
-		zero := llvm.ConstInt(c.uintptrType, 0, false)
-		hasError := c.builder.CreateICmp(llvm.IntNE, syscallResult, llvm.ConstInt(c.uintptrType, 0, false), "")
-		errResult := c.builder.CreateSelect(hasError, syscallResult, zero, "syscallError")
-		retval := llvm.Undef(c.ctx.StructType([]llvm.Type{c.uintptrType, c.uintptrType, c.uintptrType}, false))
-		retval = c.builder.CreateInsertValue(retval, syscallResult, 0, "")
-		retval = c.builder.CreateInsertValue(retval, zero, 1, "")
-		retval = c.builder.CreateInsertValue(retval, errResult, 2, "")
+		zero := llvm.ConstInt(b.uintptrType, 0, false)
+		hasError := b.CreateICmp(llvm.IntNE, syscallResult, llvm.ConstInt(b.uintptrType, 0, false), "")
+		errResult := b.CreateSelect(hasError, syscallResult, zero, "syscallError")
+		retval := llvm.Undef(b.ctx.StructType([]llvm.Type{b.uintptrType, b.uintptrType, b.uintptrType}, false))
+		retval = b.CreateInsertValue(retval, syscallResult, 0, "")
+		retval = b.CreateInsertValue(retval, zero, 1, "")
+		retval = b.CreateInsertValue(retval, errResult, 2, "")
 		return retval, nil
 	default:
-		return llvm.Value{}, c.makeError(call.Pos(), "unknown GOOS/GOARCH for syscall: "+c.GOOS()+"/"+c.GOARCH())
+		return llvm.Value{}, b.makeError(call.Pos(), "unknown GOOS/GOARCH for syscall: "+b.GOOS+"/"+b.GOARCH)
 	}
 }
@@ -0,0 +1,57 @@
+package main
+
+// Basic tests that don't need to be split into a separate file.
+
+func addInt(x, y int) int {
+	return x + y
+}
+
+func equalInt(x, y int) bool {
+	return x == y
+}
+
+func floatEQ(x, y float32) bool {
+	return x == y
+}
+
+func floatNE(x, y float32) bool {
+	return x != y
+}
+
+func floatLower(x, y float32) bool {
+	return x < y
+}
+
+func floatLowerEqual(x, y float32) bool {
+	return x <= y
+}
+
+func floatGreater(x, y float32) bool {
+	return x > y
+}
+
+func floatGreaterEqual(x, y float32) bool {
+	return x >= y
+}
+
+func complexReal(x complex64) float32 {
+	return real(x)
+}
+
+func complexImag(x complex64) float32 {
+	return imag(x)
+}
+
+func complexAdd(x, y complex64) complex64 {
+	return x + y
+}
+
+func complexSub(x, y complex64) complex64 {
+	return x - y
+}
+
+func complexMul(x, y complex64) complex64 {
+	return x * y
+}
+
+// TODO: complexDiv (requires runtime call)
@@ -0,0 +1,98 @@
+; ModuleID = 'basic.go'
+source_filename = "basic.go"
+target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-f64:32:64-f80:32-n8:16:32-S128"
+target triple = "i686--linux"
+
+define internal void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret void
+}
+
+define internal i32 @main.addInt(i32 %x, i32 %y, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = add i32 %x, %y
+  ret i32 %0
+}
+
+define internal i1 @main.equalInt(i32 %x, i32 %y, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = icmp eq i32 %x, %y
+  ret i1 %0
+}
+
+define internal i1 @main.floatEQ(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = fcmp oeq float %x, %y
+  ret i1 %0
+}
+
+define internal i1 @main.floatNE(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = fcmp une float %x, %y
+  ret i1 %0
+}
+
+define internal i1 @main.floatLower(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = fcmp olt float %x, %y
+  ret i1 %0
+}
+
+define internal i1 @main.floatLowerEqual(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = fcmp ole float %x, %y
+  ret i1 %0
+}
+
+define internal i1 @main.floatGreater(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = fcmp ogt float %x, %y
+  ret i1 %0
+}
+
+define internal i1 @main.floatGreaterEqual(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = fcmp oge float %x, %y
+  ret i1 %0
+}
+
+define internal float @main.complexReal(float %x.r, float %x.i, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret float %x.r
+}
+
+define internal float @main.complexImag(float %x.r, float %x.i, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret float %x.i
+}
+
+define internal { float, float } @main.complexAdd(float %x.r, float %x.i, float %y.r, float %y.i, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = fadd float %x.r, %y.r
+  %1 = fadd float %x.i, %y.i
+  %2 = insertvalue { float, float } undef, float %0, 0
+  %3 = insertvalue { float, float } %2, float %1, 1
+  ret { float, float } %3
+}
+
+define internal { float, float } @main.complexSub(float %x.r, float %x.i, float %y.r, float %y.i, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = fsub float %x.r, %y.r
+  %1 = fsub float %x.i, %y.i
+  %2 = insertvalue { float, float } undef, float %0, 0
+  %3 = insertvalue { float, float } %2, float %1, 1
+  ret { float, float } %3
+}
+
+define internal { float, float } @main.complexMul(float %x.r, float %x.i, float %y.r, float %y.i, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = fmul float %x.r, %y.r
+  %1 = fmul float %x.i, %y.i
+  %2 = fsub float %0, %1
+  %3 = fmul float %x.r, %y.i
+  %4 = fmul float %x.i, %y.r
+  %5 = fadd float %3, %4
+  %6 = insertvalue { float, float } undef, float %2, 0
+  %7 = insertvalue { float, float } %6, float %5, 1
+  ret { float, float } %7
+}
@@ -0,0 +1,39 @@
+package main
+
+// Test converting floats to ints.
+
+func f32tou32(v float32) uint32 {
+	return uint32(v)
+}
+
+func maxu32f() float32 {
+	return float32(^uint32(0))
+}
+
+func maxu32tof32() uint32 {
+	f := float32(^uint32(0))
+	return uint32(f)
+}
+
+func inftoi32() (uint32, uint32, int32, int32) {
+	inf := 1.0
+	inf /= 0.0
+
+	return uint32(inf), uint32(-inf), int32(inf), int32(-inf)
+}
+
+func u32tof32tou32(v uint32) uint32 {
+	return uint32(float32(v))
+}
+
+func f32tou32tof32(v float32) float32 {
+	return float32(uint32(v))
+}
+
+func f32tou8(v float32) uint8 {
+	return uint8(v)
+}
+
+func f32toi8(v float32) int8 {
+	return int8(v)
+}
@@ -0,0 +1,80 @@
+; ModuleID = 'float.go'
+source_filename = "float.go"
+target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-f64:32:64-f80:32-n8:16:32-S128"
+target triple = "i686--linux"
+
+define internal void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret void
+}
+
+define internal i32 @main.f32tou32(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %positive = fcmp oge float %v, 0.000000e+00
+  %withinmax = fcmp ole float %v, 0x41EFFFFFC0000000
+  %inbounds = and i1 %positive, %withinmax
+  %saturated = sext i1 %positive to i32
+  %normal = fptoui float %v to i32
+  %0 = select i1 %inbounds, i32 %normal, i32 %saturated
+  ret i32 %0
+}
+
+define internal float @main.maxu32f(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret float 0x41F0000000000000
+}
+
+define internal i32 @main.maxu32tof32(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret i32 -1
+}
+
+define internal { i32, i32, i32, i32 } @main.inftoi32(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret { i32, i32, i32, i32 } { i32 -1, i32 0, i32 2147483647, i32 -2147483648 }
+}
+
+define internal i32 @main.u32tof32tou32(i32 %v, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = uitofp i32 %v to float
+  %withinmax = fcmp ole float %0, 0x41EFFFFFC0000000
+  %normal = fptoui float %0 to i32
+  %1 = select i1 %withinmax, i32 %normal, i32 -1
+  ret i32 %1
+}
+
+define internal float @main.f32tou32tof32(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %positive = fcmp oge float %v, 0.000000e+00
+  %withinmax = fcmp ole float %v, 0x41EFFFFFC0000000
+  %inbounds = and i1 %positive, %withinmax
+  %saturated = sext i1 %positive to i32
+  %normal = fptoui float %v to i32
+  %0 = select i1 %inbounds, i32 %normal, i32 %saturated
+  %1 = uitofp i32 %0 to float
+  ret float %1
+}
+
+define internal i8 @main.f32tou8(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %positive = fcmp oge float %v, 0.000000e+00
+  %withinmax = fcmp ole float %v, 2.550000e+02
+  %inbounds = and i1 %positive, %withinmax
+  %saturated = sext i1 %positive to i8
+  %normal = fptoui float %v to i8
+  %0 = select i1 %inbounds, i8 %normal, i8 %saturated
+  ret i8 %0
+}
+
+define internal i8 @main.f32toi8(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %abovemin = fcmp oge float %v, -1.280000e+02
+  %belowmax = fcmp ole float %v, 1.270000e+02
+  %inbounds = and i1 %abovemin, %belowmax
+  %saturated = select i1 %abovemin, i8 127, i8 -128
+  %isnan = fcmp uno float %v, 0.000000e+00
+  %remapped = select i1 %isnan, i8 0, i8 %saturated
+  %normal = fptosi float %v to i8
+  %0 = select i1 %inbounds, i8 %normal, i8 %remapped
+  ret i8 %0
+}
@@ -0,0 +1,41 @@
+package main
+
+// This file tests various operations on pointers, such as pointer arithmetic
+// and dereferencing pointers.
+
+import "unsafe"
+
+// Dereference pointers.
+
+func pointerDerefZero(x *[0]int) [0]int {
+	return *x // This is a no-op, there is nothing to load.
+}
+
+// Unsafe pointer casts, they are sometimes a no-op.
+
+func pointerCastFromUnsafe(x unsafe.Pointer) *int {
+	return (*int)(x)
+}
+
+func pointerCastToUnsafe(x *int) unsafe.Pointer {
+	return unsafe.Pointer(x)
+}
+
+func pointerCastToUnsafeNoop(x *byte) unsafe.Pointer {
+	return unsafe.Pointer(x)
+}
+
+// The compiler has support for a few special cast+add patterns that are
+// transformed into a single GEP.
+
+func pointerUnsafeGEPFixedOffset(ptr *byte) *byte {
+	return (*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(ptr)) + 10))
+}
+
+func pointerUnsafeGEPByteOffset(ptr *byte, offset uintptr) *byte {
+	return (*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(ptr)) + offset))
+}
+
+func pointerUnsafeGEPIntOffset(ptr *int32, offset uintptr) *int32 {
+	return (*int32)(unsafe.Pointer(uintptr(unsafe.Pointer(ptr)) + offset*4))
+}
@@ -0,0 +1,49 @@
+; ModuleID = 'pointer.go'
+source_filename = "pointer.go"
+target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-f64:32:64-f80:32-n8:16:32-S128"
+target triple = "i686--linux"
+
+define internal void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret void
+}
+
+define internal [0 x i32] @main.pointerDerefZero([0 x i32]* %x, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret [0 x i32] zeroinitializer
+}
+
+define internal i32* @main.pointerCastFromUnsafe(i8* %x, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = bitcast i8* %x to i32*
+  ret i32* %0
+}
+
+define internal i8* @main.pointerCastToUnsafe(i32* dereferenceable_or_null(4) %x, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = bitcast i32* %x to i8*
+  ret i8* %0
+}
+
+define internal i8* @main.pointerCastToUnsafeNoop(i8* dereferenceable_or_null(1) %x, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret i8* %x
+}
+
+define internal i8* @main.pointerUnsafeGEPFixedOffset(i8* dereferenceable_or_null(1) %ptr, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = getelementptr inbounds i8, i8* %ptr, i32 10
+  ret i8* %0
+}
+
+define internal i8* @main.pointerUnsafeGEPByteOffset(i8* dereferenceable_or_null(1) %ptr, i32 %offset, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = getelementptr inbounds i8, i8* %ptr, i32 %offset
+  ret i8* %0
+}
+
+define internal i32* @main.pointerUnsafeGEPIntOffset(i32* dereferenceable_or_null(4) %ptr, i32 %offset, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = getelementptr i32, i32* %ptr, i32 %offset
+  ret i32* %0
+}
@@ -0,0 +1,9 @@
+package main
+
+func sliceLen(ints []int) int {
+	return len(ints)
+}
+
+func sliceCap(ints []int) int {
+	return cap(ints)
+}
@@ -0,0 +1,19 @@
+; ModuleID = 'slice.go'
+source_filename = "slice.go"
+target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-f64:32:64-f80:32-n8:16:32-S128"
+target triple = "i686--linux"
+
+define internal void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret void
+}
+
+define internal i32 @main.sliceLen(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret i32 %ints.len
+}
+
+define internal i32 @main.sliceCap(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret i32 %ints.cap
+}
@@ -8,19 +8,23 @@ import (
 	"tinygo.org/x/go-llvm"
 )

-func (c *Compiler) emitVolatileLoad(frame *Frame, instr *ssa.CallCommon) (llvm.Value, error) {
-	addr := c.getValue(frame, instr.Args[0])
-	c.emitNilCheck(frame, addr, "deref")
-	val := c.builder.CreateLoad(addr, "")
+// createVolatileLoad is the implementation of the intrinsic function
+// runtime/volatile.LoadT().
+func (b *builder) createVolatileLoad(instr *ssa.CallCommon) (llvm.Value, error) {
+	addr := b.getValue(instr.Args[0])
+	b.createNilCheck(instr.Args[0], addr, "deref")
+	val := b.CreateLoad(addr, "")
 	val.SetVolatile(true)
 	return val, nil
 }

-func (c *Compiler) emitVolatileStore(frame *Frame, instr *ssa.CallCommon) (llvm.Value, error) {
-	addr := c.getValue(frame, instr.Args[0])
-	val := c.getValue(frame, instr.Args[1])
-	c.emitNilCheck(frame, addr, "deref")
-	store := c.builder.CreateStore(val, addr)
+// createVolatileStore is the implementation of the intrinsic function
+// runtime/volatile.StoreT().
+func (b *builder) createVolatileStore(instr *ssa.CallCommon) (llvm.Value, error) {
+	addr := b.getValue(instr.Args[0])
+	val := b.getValue(instr.Args[1])
+	b.createNilCheck(instr.Args[0], addr, "deref")
+	store := b.CreateStore(val, addr)
 	store.SetVolatile(true)
 	return llvm.Value{}, nil
 }
@@ -4,11 +4,13 @@ go 1.11

 require (
 	github.com/blakesmith/ar v0.0.0-20150311145944-8bd4349a67f2
-	github.com/creack/goselect v0.1.0 // indirect
+	github.com/chromedp/cdproto v0.0.0-20210113043257-dabd2f2e7693
+	github.com/chromedp/chromedp v0.6.4
 	github.com/google/shlex v0.0.0-20181106134648-c34317bd91bf
-	github.com/marcinbor85/gohex v0.0.0-20180128172054-7a43cd876e46
-	go.bug.st/serial.v1 v0.0.0-20180827123349-5f7892a7bb45
-	golang.org/x/sys v0.0.0-20191010194322-b09406accb47 // indirect
-	golang.org/x/tools v0.0.0-20190227180812-8dcc6e70cdef
-	tinygo.org/x/go-llvm v0.0.0-20200104190746-1ff21df33566
+	github.com/marcinbor85/gohex v0.0.0-20200531091804-343a4b548892
+	github.com/mattn/go-colorable v0.1.8
+	go.bug.st/serial v1.1.2
+	golang.org/x/sys v0.0.0-20210113181707-4bcb84eeeb78
+	golang.org/x/tools v0.0.0-20200216192241-b320d3a0f5a2
+	tinygo.org/x/go-llvm v0.0.0-20210206225315-7fe719483a0f
 )
@@ -1,36 +1,61 @@
 github.com/blakesmith/ar v0.0.0-20150311145944-8bd4349a67f2 h1:oMCHnXa6CCCafdPDbMh/lWRhRByN0VFLvv+g+ayx1SI=
 github.com/blakesmith/ar v0.0.0-20150311145944-8bd4349a67f2/go.mod h1:PkYb9DJNAwrSvRx5DYA+gUcOIgTGVMNkfSCbZM8cWpI=
-github.com/creack/goselect v0.1.0 h1:4QiXIhcpSQF50XGaBsFzesjwX/1qOY5bOveQPmN9CXY=
-github.com/creack/goselect v0.1.0/go.mod h1:gHrIcH/9UZDn2qgeTUeW5K9eZsVYCH6/60J/FHysWyE=
-github.com/golang/protobuf v1.2.0/go.mod h1:6lQm79b+lXiMfvg/cZm0SGofjICqVBUtrP5yJMmIC1U=
+github.com/chromedp/cdproto v0.0.0-20210113043257-dabd2f2e7693 h1:11eq/RkpaotwdF6b1TRMcdgQUPNmyFEJOB7zLvh0O/Y=
+github.com/chromedp/cdproto v0.0.0-20210113043257-dabd2f2e7693/go.mod h1:55pim6Ht4LJKdVLlyFJV/g++HsEA1hQxPbB5JyNdZC0=
+github.com/chromedp/chromedp v0.6.4 h1:Gx7ZkRyrSVmbbDDja/ieNgNGJIvElroPOyeqYQGVDSY=
+github.com/chromedp/chromedp v0.6.4/go.mod h1:vodUdJf5dF/b8n0UBJv6NeM/QK28RjP3j+eM7fq4+84=
+github.com/chromedp/sysutil v1.0.0 h1:+ZxhTpfpZlmchB58ih/LBHX52ky7w2VhQVKQMucy3Ic=
+github.com/chromedp/sysutil v1.0.0/go.mod h1:kgWmDdq8fTzXYcKIBqIYvRRTnYb9aNS9moAV0xufSww=
+github.com/creack/goselect v0.1.1 h1:tiSSgKE1eJtxs1h/VgGQWuXUP0YS4CDIFMp6vaI1ls0=
+github.com/creack/goselect v0.1.1/go.mod h1:a/NhLweNvqIYMuxcMOuWY516Cimucms3DglDzQP3hKY=
+github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/gobwas/httphead v0.1.0 h1:exrUm0f4YX0L7EBwZHuCF4GDp8aJfVeBrlLQrs6NqWU=
+github.com/gobwas/httphead v0.1.0/go.mod h1:O/RXo79gxV8G+RqlR/otEwx4Q36zl9rqC5u12GKvMCM=
+github.com/gobwas/pool v0.2.1 h1:xfeeEhW7pwmX8nuLVlqbzVc7udMDrwetjEv+TZIz1og=
+github.com/gobwas/pool v0.2.1/go.mod h1:q8bcK0KcYlCgd9e7WYLm9LpyS+YeLd8JVDW6WezmKEw=
+github.com/gobwas/ws v1.0.4 h1:5eXU1CZhpQdq5kXbKb+sECH5Ia5KiO6CYzIzdlVx6Bs=
+github.com/gobwas/ws v1.0.4/go.mod h1:szmBTxLgaFppYjEmNtny/v3w89xOydFnnZMcgRRu/EM=
 github.com/google/shlex v0.0.0-20181106134648-c34317bd91bf h1:7+FW5aGwISbqUtkfmIpZJGRgNFg2ioYPvFaUxdqpDsg=
 github.com/google/shlex v0.0.0-20181106134648-c34317bd91bf/go.mod h1:RpwtwJQFrIEPstU94h88MWPXP2ektJZ8cZ0YntAmXiE=
-github.com/marcinbor85/gohex v0.0.0-20180128172054-7a43cd876e46 h1:wXG2bA8fO7Vv7lLk2PihFMTqmbT173Tje39oKzQ50Mo=
-github.com/marcinbor85/gohex v0.0.0-20180128172054-7a43cd876e46/go.mod h1:Pb6XcsXyropB9LNHhnqaknG/vEwYztLkQzVCHv8sQ3M=
-go.bug.st/serial.v1 v0.0.0-20180827123349-5f7892a7bb45 h1:mACY1anK6HNCZtm/DK2Rf2ZPHggVqeB0+7rY9Gl6wyI=
-go.bug.st/serial.v1 v0.0.0-20180827123349-5f7892a7bb45/go.mod h1:dRSl/CVCTf56CkXgJMDOdSwNfo2g1orOGE/gBGdvjZw=
-golang.org/x/net v0.0.0-20180724234803-3673e40ba225/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
-golang.org/x/net v0.0.0-20190213061140-3a22650c66bd/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
-golang.org/x/sync v0.0.0-20181108010431-42b317875d0f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
-golang.org/x/sys v0.0.0-20191010194322-b09406accb47 h1:/XfQ9z7ib8eEJX2hdgFTZJ/ntt0swNk5oYBziWeTCvY=
-golang.org/x/sys v0.0.0-20191010194322-b09406accb47/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+github.com/josharian/intern v1.0.0 h1:vlS4z54oSdjm0bgjRigI+G1HpF+tI+9rE5LLzOg8HmY=
+github.com/josharian/intern v1.0.0/go.mod h1:5DoeVV0s6jJacbCEi61lwdGj/aVlrQvzHFFd8Hwg//Y=
+github.com/mailru/easyjson v0.7.6 h1:8yTIVnZgCoiM1TgqoeTl+LfU5Jg6/xL3QhGQnimLYnA=
+github.com/mailru/easyjson v0.7.6/go.mod h1:xzfreul335JAWq5oZzymOObrkdz5UnU4kGfJJLY9Nlc=
+github.com/marcinbor85/gohex v0.0.0-20200531091804-343a4b548892 h1:6J+qramlHVLmiBOgRiBOnQkno8uprqG6YFFQTt6uYIw=
+github.com/marcinbor85/gohex v0.0.0-20200531091804-343a4b548892/go.mod h1:Pb6XcsXyropB9LNHhnqaknG/vEwYztLkQzVCHv8sQ3M=
+github.com/mattn/go-colorable v0.1.8 h1:c1ghPdyEDarC70ftn0y+A/Ee++9zz8ljHG1b13eJ0s8=
+github.com/mattn/go-colorable v0.1.8/go.mod h1:u6P/XSegPjTcexA+o6vUJrdnUu04hMope9wVRipJSqc=
+github.com/mattn/go-isatty v0.0.12 h1:wuysRhFDzyxgEmMf5xjvJ2M9dZoWAXNNr5LSBS7uHXY=
+github.com/mattn/go-isatty v0.0.12/go.mod h1:cbi8OIDigv2wuxKPP5vlRcQ1OAZbq2CE4Kysco4FUpU=
+github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
+github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
+github.com/stretchr/testify v1.4.0/go.mod h1:j7eGeouHqKxXV5pUuKE4zz7dFj8WfuZ+81PSLYec5m4=
+go.bug.st/serial v1.0.0 h1:ogEPzrllCsnG00EqKRjeYvPRsO7NJW6DqykzkdD6E/k=
+go.bug.st/serial v1.0.0/go.mod h1:rpXPISGjuNjPTRTcMlxi9lN6LoIPxd1ixVjBd8aSk/Q=
+go.bug.st/serial v1.1.2 h1:6xDpbta8KJ+VLRTeM8ghhxXRMLE/Lr8h9iDKwydarAY=
+go.bug.st/serial v1.1.2/go.mod h1:VmYBeyJWp5BnJ0tw2NUJHZdJTGl2ecBGABHlzRK1knY=
+golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACkg1iLfiJU5Ep61QUkGW8qpdssI0+w=
+golang.org/x/crypto v0.0.0-20191011191535-87dc89f01550/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
+golang.org/x/mod v0.1.1-0.20191105210325-c90efee705ee h1:WG0RUwxtNT4qqaXX3DPA8zHFNm/D9xaBpxzHt1WcA/E=
+golang.org/x/mod v0.1.1-0.20191105210325-c90efee705ee/go.mod h1:QqPTAvyqsEbceGzBzNggFXnrqF1CaUcvgkdR5Ot7KZg=
+golang.org/x/net v0.0.0-20190404232315-eb5bcb51f2a3/go.mod h1:t9HGtf8HONx5eT2rtn7q6eTqICYqUVnKs3thJo3Qplg=
+golang.org/x/net v0.0.0-20190620200207-3b0461eec859/go.mod h1:z5CRVTTTmAJ677TzLLGU+0bjPO0LkuOLi4/5GtJWs/s=
+golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
+golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
+golang.org/x/sys v0.0.0-20190412213103-97732733099d/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20191128015809-6d18c012aee9 h1:ZBzSG/7F4eNKz2L3GE9o300RX0Az1Bw5HF7PDraD+qU=
+golang.org/x/sys v0.0.0-20191128015809-6d18c012aee9/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20200116001909-b77594299b42/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20200223170610-d5e6a3e2c0ae/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20200909081042-eff7692f9009/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
+golang.org/x/sys v0.0.0-20210113181707-4bcb84eeeb78 h1:nVuTkr9L6Bq62qpUqKo/RnZCFfzDBL0bYo6w9OJUqZY=
+golang.org/x/sys v0.0.0-20210113181707-4bcb84eeeb78/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
 golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
-golang.org/x/tools v0.0.0-20190227180812-8dcc6e70cdef h1:ymc9FeDom3RIEA3coKokSllBB1hRcMT0tZ1W3Jf9Ids=
-golang.org/x/tools v0.0.0-20190227180812-8dcc6e70cdef/go.mod h1:9Yl7xja0Znq3iFh3HoIrodX9oNMXvdceNzlUR8zjMvY=
-google.golang.org/appengine v1.4.0/go.mod h1:xpcJRLb0r/rnEns0DIKYYv+WjYCduHsrkT7/EB5XEv4=
-tinygo.org/x/go-llvm v0.0.0-20190224120431-7707ae5d1261 h1:rJS2Hga39YAnm7DE4qrPm6Dr/67EOojL0XPzvbEeBiw=
-tinygo.org/x/go-llvm v0.0.0-20190224120431-7707ae5d1261/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
-tinygo.org/x/go-llvm v0.0.0-20190818154551-95bc4ffe1add h1:dFjMH1sLhYADg8UQm7DB56B7e+TfvAmWmEZLhyv3r/w=
-tinygo.org/x/go-llvm v0.0.0-20190818154551-95bc4ffe1add/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
-tinygo.org/x/go-llvm v0.0.0-20191103182207-90b6e4bdc0b9 h1:d6rAX39a3C0pKrY5HcojEGyN8w9ocU0v7X28lC/TRKU=
-tinygo.org/x/go-llvm v0.0.0-20191103182207-90b6e4bdc0b9/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
-tinygo.org/x/go-llvm v0.0.0-20191103200204-37e93e3f04e2 h1:Q5Hv3e5cLMGkiYwYgZL1Zrv6nb/EY+DJpRWrdO6ws6o=
-tinygo.org/x/go-llvm v0.0.0-20191103200204-37e93e3f04e2/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
-tinygo.org/x/go-llvm v0.0.0-20191113125529-bad6d01809e8 h1:9Bfvso+tTVQg16UzOA614NaYA4x8vsRBNtd3eBrXwp0=
-tinygo.org/x/go-llvm v0.0.0-20191113125529-bad6d01809e8/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
-tinygo.org/x/go-llvm v0.0.0-20191124211856-b2db3df3f257 h1:o8VDylrMN7gWemBMu8rEyuogKPhcLTdx5KrUAp9macc=
-tinygo.org/x/go-llvm v0.0.0-20191124211856-b2db3df3f257/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
-tinygo.org/x/go-llvm v0.0.0-20191215173731-ad71f3d24aae h1:s8J5EyxCkHxXB08UI3gk9W9IS/ekizRvSX+PfZxnAB0=
-tinygo.org/x/go-llvm v0.0.0-20191215173731-ad71f3d24aae/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
-tinygo.org/x/go-llvm v0.0.0-20200104190746-1ff21df33566 h1:a4y30bTf7U0zDA75v2PTL+XQ2OzJetj19gK8XwQpUNY=
-tinygo.org/x/go-llvm v0.0.0-20200104190746-1ff21df33566/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
+golang.org/x/tools v0.0.0-20200216192241-b320d3a0f5a2 h1:0sfSpGSa544Fwnbot3Oxq/U6SXqjty6Jy/3wRhVS7ig=
+golang.org/x/tools v0.0.0-20200216192241-b320d3a0f5a2/go.mod h1:TB2adYChydJhpapKDTa4BR/hXlZSLoq2Wpct/0txZ28=
+golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898 h1:/atklqdjdhuosWIl6AIbOeHJjicWYPqR9bpxqxYG2pA=
+golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/yaml.v2 v2.2.2/go.mod h1:hI93XBmqTisBFMUTm0b8Fm+jr3Dg1NNxqwp+5A1VGuI=
+tinygo.org/x/go-llvm v0.0.0-20210206225315-7fe719483a0f h1:FP5Do5omlQ/dLQ3Hfy7oyJo69VS5Hn46rZw004r0lGU=
+tinygo.org/x/go-llvm v0.0.0-20210206225315-7fe719483a0f/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
@@ -0,0 +1,68 @@
+package goenv
+
+import (
+	"errors"
+	"fmt"
+	"io"
+	"io/ioutil"
+	"path/filepath"
+	"regexp"
+	"strings"
+)
+
+// Version of TinyGo.
+// Update this value before release of new version of software.
+const Version = "0.17.0"
+
+// GetGorootVersion returns the major and minor version for a given GOROOT path.
+// If the goroot cannot be determined, (0, 0) is returned.
+func GetGorootVersion(goroot string) (major, minor int, err error) {
+	s, err := GorootVersionString(goroot)
+	if err != nil {
+		return 0, 0, err
+	}
+
+	if s == "" || s[:2] != "go" {
+		return 0, 0, errors.New("could not parse Go version: version does not start with 'go' prefix")
+	}
+
+	parts := strings.Split(s[2:], ".")
+	if len(parts) < 2 {
+		return 0, 0, errors.New("could not parse Go version: version has less than two parts")
+	}
+
+	// Ignore the errors, we don't really handle errors here anyway.
+	var trailing string
+	n, err := fmt.Sscanf(s, "go%d.%d%s", &major, &minor, &trailing)
+	if n == 2 && err == io.EOF {
+		// Means there were no trailing characters (i.e., not an alpha/beta)
+		err = nil
+	}
+	if err != nil {
+		return 0, 0, fmt.Errorf("failed to parse version: %s", err)
+	}
+	return
+}
+
+// GorootVersionString returns the version string as reported by the Go
+// toolchain for the given GOROOT path. It is usually of the form `go1.x.y` but
+// can have some variations (for beta releases, for example).
+func GorootVersionString(goroot string) (string, error) {
+	if data, err := ioutil.ReadFile(filepath.Join(
+		goroot, "src", "runtime", "internal", "sys", "zversion.go")); err == nil {
+
+		r := regexp.MustCompile("const TheVersion = `(.*)`")
+		matches := r.FindSubmatch(data)
+		if len(matches) != 2 {
+			return "", errors.New("Invalid go version output:\n" + string(data))
+		}
+
+		return string(matches[1]), nil
+
+	} else if data, err := ioutil.ReadFile(filepath.Join(goroot, "VERSION")); err == nil {
+		return string(data), nil
+
+	} else {
+		return "", err
+	}
+}
@@ -0,0 +1,5 @@
+# Hooks for Docker Hub
+
+Files in this directory are custom commands to be run during the different Docker Hub build phases.
+
+See https://docs.docker.com/docker-hub/builds/advanced/#custom-build-phase-hooks
@@ -0,0 +1,4 @@
+#!/bin/bash
+# Docker hub does a recursive clone, then checks the branch out,
+# so when a PR adds a submodule (or updates it), it fails.
+git submodule update --init
@@ -6,50 +6,81 @@ possible and only run unknown expressions (e.g. external calls) at runtime. This
 is in practice a partial evaluator of the `runtime.initAll` function, which
 calls each package initializer.

-It works by directly interpreting LLVM IR:
+This package is a rewrite of a previous partial evaluator that worked
+directly on LLVM IR and used the module and LLVM constants as intermediate
+values. This newer version instead uses a mostly Go intermediate form. It
+compiles functions and extracts relevant data first (compiler.go), then
+executes those functions (interpreter.go) in a memory space that can be
+rolled back per function (memory.go). This means that it is not necessary to
+scan functions to see whether they can be run at compile time, which was very
+error prone. Instead it just tries to execute everything and if it hits
+something it cannot interpret (such as a store to memory-mapped I/O) it rolls
+back the execution of that function and runs the function at runtime instead.
+All in all, this design provides several benefits:

-  * Almost all operations work directly on constants, and are implemented using
-    the llvm.Const* set of functions that are evaluated directly.
-  * External function calls and some other operations (inline assembly, volatile
-    load, volatile store) are seen as having limited side effects. Limited in
-    the sense that it is known at compile time which globals it affects, which
-    then are marked 'dirty' (meaning, further operations on it must be done at
-    runtime). These operations are emitted directly in the `runtime.initAll`
-    function. Return values are also considered 'dirty'.
-  * Such 'dirty' objects and local values must be executed at runtime instead of
-    at compile time. This dirtyness propagates further through the IR, for
-    example storing a dirty local value to a global also makes the global dirty,
-    meaning that the global may not be read or written at compile time as it's
-    contents at that point during interpretation is unknown.
-  * There are some heuristics in place to avoid doing too much with dirty
-    values. For example, a branch based on a dirty local marks the whole
-    function itself as having side effect (as if it is an external function).
-    However, all globals it touches are still taken into account and when a call
-    is inserted in `runtime.initAll`, all globals it references are also marked
-    dirty.
-  * Heap allocation (`runtime.alloc`) is emulated by creating new objects. The
-    value in the allocation is the initializer of the global, the zero value is
-    the zero initializer.
-  * Stack allocation (`alloca`) is often emulated using a fake alloca object,
-    until the address of the alloca is taken in which case it is also created as
-    a real `alloca` in `runtime.initAll` and marked dirty. This may be necessary
-    when calling an external function with the given alloca as paramter.
+  * Much better error handling. By being able to revert to runtime execution
+    without the need for scanning functions, this version is able to
+    automatically work around many bugs in the previous implementation.
+  * More correct memory model. This is not inherent to the new design, but the
+    new design also made the memory model easier to reason about.
+  * Faster execution of initialization code. While it is not much faster for
+    normal interpretation (maybe 25% or so) due to the compilation overhead,
+    it should be a whole lot faster for loops as it doesn't have to call into
+    LLVM (via CGo) for every operation.
+
+As mentioned, this partial evaulator comes in three parts: a compiler, an
+interpreter, and a memory manager.
+
+## Compiler
+
+The main task of the compiler is that it extracts all necessary data from
+every instruction in a function so that when this instruction is interpreted,
+no additional CGo calls are necessary. This is not currently done for all
+instructions (`runtime.alloc` is a notable exception), but at least it does
+so for the vast majority of instructions.
+
+## Interpreter
+
+The interpreter runs an instruction just like it would if it were executed
+'for real'. The vast majority of instructions can be executed at compile
+time. As indicated above, some instructions need to be executed at runtime
+instead.
+
+## Memory
+
+Memory is represented as objects (the `object` type) that contains data that
+will eventually be stored in a global and values (the `value` interface) that
+can be worked with while running the interpreter. Values therefore are only
+used locally and are always passed by value (just like most LLVM constants)
+while objects represent the backing storage (like LLVM globals). Some values
+are pointer values, and point to an object.
+
+Importantly, this partial evaluator can roll back the execution of a
+function. This is implemented by creating a new memory view per function
+activation, which makes sure that any change to a global (such as a store
+instruction) is stored in the memory view. It creates a copy of the object
+and stores that in the memory view to be modified. Once the function has
+executed successfully, all these modified objects are then copied into the
+parent function, up to the root function invocation which (on successful
+execution) writes the values back into the LLVM module. This way, function
+invocations can be rolled back without leaving a trace.
+
+Pointer values point to memory objects, but not to a particular memory
+object. Every memory object is given an index, and pointers use that index to
+look up the current active object for the pointer to load from or to copy
+when storing to it.
+
+Rolling back a function should roll back everyting, including the few
+instructions emitted at runtime. This is done by treating instructions much
+like memory objects and removing the created instructions when necessary.

 ## Why is this necessary?

 A partial evaluator is hard to get right, so why go through all the trouble of
 writing one?

-The main reason is that the previous attempt wasn't complete and wasn't sound.
-It simply tried to evaluate Go SSA directly, which was good but more difficult
-than necessary. An IR based interpreter needs to understand fewer instructions
-as the LLVM IR simply has less (complex) instructions than Go SSA. Also, LLVM
-provides some useful tools like easily getting all uses of a function or global,
-which Go SSA does not provide.
-
-But why is it necessary at all? The answer is that globals with initializers are
-much easier to optimize by LLVM than initialization code. Also, there are a few
-other benefits:
+The answer is that globals with initializers are much easier to optimize by
+LLVM than initialization code. Also, there are a few other benefits:

  * Dead globals are trivial to optimize away.
  * Constant globals are easier to detect. Remember that Go does not have global
@@ -60,5 +91,29 @@ other benefits:
  * Constants are much more efficent on microcontrollers, as they can be
    allocated in flash instead of RAM.

+The Go SSA package does not create constant initializers for globals.
+Instead, it emits initialization functions, so if you write the following:
+
+```go
+var foo = []byte{1, 2, 3, 4}
+```
+
+It would generate something like this:
+
+```go
+var foo []byte
+
+func init() {
+    foo = make([]byte, 4)
+    foo[0] = 1
+    foo[1] = 2
+    foo[2] = 3
+    foo[3] = 4
+}
+```
+
+This is of course hugely wasteful, it's much better to create `foo` as a
+global array instead of initializing it at runtime.
+
 For more details, see [this section of the
 documentation](https://tinygo.org/compiler-internals/differences-from-go/).
@@ -0,0 +1,410 @@
+package interp
+
+// This file compiles the LLVM IR to a form that's easy to efficiently
+// interpret.
+
+import (
+	"strings"
+
+	"tinygo.org/x/go-llvm"
+)
+
+// A function is a compiled LLVM function, which means that interpreting it
+// avoids most CGo calls necessary. This is done in a separate step so the
+// result can be cached.
+// Functions are in SSA form, just like the LLVM version if it. The first block
+// (blocks[0]) is the entry block.
+type function struct {
+	llvmFn llvm.Value
+	name   string       // precalculated llvmFn.Name()
+	params []llvm.Value // precalculated llvmFn.Params()
+	blocks []*basicBlock
+	locals map[llvm.Value]int
+}
+
+// basicBlock represents a LLVM basic block and contains a slice of
+// instructions. The last instruction must be a terminator instruction.
+type basicBlock struct {
+	instructions []instruction
+}
+
+// instruction is a precompiled LLVM IR instruction. The operands can be either
+// an already known value (such as literalValue or pointerValue) but can also be
+// the special localValue, which means that the value is a function parameter or
+// is produced by another instruction in the function. In that case, the
+// interpreter will replace the operand with that local value.
+type instruction struct {
+	opcode     llvm.Opcode
+	localIndex int
+	operands   []value
+	llvmInst   llvm.Value
+	name       string
+}
+
+// String returns a nice human-readable version of this instruction.
+func (inst *instruction) String() string {
+	operands := make([]string, len(inst.operands))
+	for i, op := range inst.operands {
+		operands[i] = op.String()
+	}
+
+	name := instructionNameMap[inst.opcode]
+	if name == "" {
+		name = "<unknown op>"
+	}
+	return name + " " + strings.Join(operands, " ")
+}
+
+// compileFunction compiles a given LLVM function to an easier to interpret
+// version of the function. As far as possible, all operands are preprocessed so
+// that the interpreter doesn't have to call into LLVM.
+func (r *runner) compileFunction(llvmFn llvm.Value) *function {
+	fn := &function{
+		llvmFn: llvmFn,
+		name:   llvmFn.Name(),
+		params: llvmFn.Params(),
+		locals: make(map[llvm.Value]int),
+	}
+	if llvmFn.IsDeclaration() {
+		// Nothing to do.
+		return fn
+	}
+
+	for i, param := range fn.params {
+		fn.locals[param] = i
+	}
+
+	// Make a map of all the blocks, to quickly find the block number for a
+	// given branch instruction.
+	blockIndices := make(map[llvm.Value]int)
+	for llvmBB := llvmFn.FirstBasicBlock(); !llvmBB.IsNil(); llvmBB = llvm.NextBasicBlock(llvmBB) {
+		index := len(blockIndices)
+		blockIndices[llvmBB.AsValue()] = index
+	}
+
+	// Compile every block.
+	for llvmBB := llvmFn.FirstBasicBlock(); !llvmBB.IsNil(); llvmBB = llvm.NextBasicBlock(llvmBB) {
+		bb := &basicBlock{}
+		fn.blocks = append(fn.blocks, bb)
+
+		// Compile every instruction in the block.
+		for llvmInst := llvmBB.FirstInstruction(); !llvmInst.IsNil(); llvmInst = llvm.NextInstruction(llvmInst) {
+			// Create instruction skeleton.
+			opcode := llvmInst.InstructionOpcode()
+			inst := instruction{
+				opcode:     opcode,
+				localIndex: len(fn.locals),
+				llvmInst:   llvmInst,
+			}
+			fn.locals[llvmInst] = len(fn.locals)
+
+			// Add operands specific for this instruction.
+			switch opcode {
+			case llvm.Ret:
+				// Return instruction, which can either be a `ret void` (no
+				// return value) or return a value.
+				numOperands := llvmInst.OperandsCount()
+				if numOperands != 0 {
+					inst.operands = []value{
+						r.getValue(llvmInst.Operand(0)),
+					}
+				}
+			case llvm.Br:
+				// Branch instruction. Can be either a conditional branch (with
+				// 3 operands) or unconditional branch (with just one basic
+				// block operand).
+				numOperands := llvmInst.OperandsCount()
+				switch numOperands {
+				case 3:
+					// Conditional jump to one of two blocks. Comparable to an
+					// if/else in procedural languages.
+					thenBB := llvmInst.Operand(2)
+					elseBB := llvmInst.Operand(1)
+					inst.operands = []value{
+						r.getValue(llvmInst.Operand(0)),
+						literalValue{uint32(blockIndices[thenBB])},
+						literalValue{uint32(blockIndices[elseBB])},
+					}
+				case 1:
+					// Unconditional jump to a target basic block. Comparable to
+					// a jump in C and Go.
+					jumpBB := llvmInst.Operand(0)
+					inst.operands = []value{
+						literalValue{uint32(blockIndices[jumpBB])},
+					}
+				default:
+					panic("unknown number of operands")
+				}
+			case llvm.PHI:
+				inst.name = llvmInst.Name()
+				incomingCount := inst.llvmInst.IncomingCount()
+				for i := 0; i < incomingCount; i++ {
+					incomingBB := inst.llvmInst.IncomingBlock(i)
+					incomingValue := inst.llvmInst.IncomingValue(i)
+					inst.operands = append(inst.operands,
+						literalValue{uint32(blockIndices[incomingBB.AsValue()])},
+						r.getValue(incomingValue),
+					)
+				}
+			case llvm.Select:
+				// Select is a special instruction that is much like a ternary
+				// operator. It produces operand 1 or 2 based on the boolean
+				// that is operand 0.
+				inst.name = llvmInst.Name()
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+					r.getValue(llvmInst.Operand(1)),
+					r.getValue(llvmInst.Operand(2)),
+				}
+			case llvm.Call:
+				// Call is a regular function call but could also be a runtime
+				// intrinsic. Some runtime intrinsics are treated specially by
+				// the interpreter, such as runtime.alloc. We don't
+				// differentiate between them here because these calls may also
+				// need to be run at runtime, in which case they should all be
+				// created in the same way.
+				llvmCalledValue := llvmInst.CalledValue()
+				if !llvmCalledValue.IsAFunction().IsNil() {
+					name := llvmCalledValue.Name()
+					if name == "llvm.dbg.value" || strings.HasPrefix(name, "llvm.lifetime.") {
+						// These intrinsics should not be interpreted, they are not
+						// relevant to the execution of this function.
+						continue
+					}
+				}
+				inst.name = llvmInst.Name()
+				numOperands := llvmInst.OperandsCount()
+				inst.operands = append(inst.operands, r.getValue(llvmCalledValue))
+				for i := 0; i < numOperands-1; i++ {
+					inst.operands = append(inst.operands, r.getValue(llvmInst.Operand(i)))
+				}
+			case llvm.Load:
+				// Load instruction. The interpreter will load from the
+				// appropriate memory view.
+				// Also provide the memory size to be loaded, which is necessary
+				// with a lack of type information.
+				inst.name = llvmInst.Name()
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+					literalValue{r.targetData.TypeAllocSize(llvmInst.Type())},
+				}
+			case llvm.Store:
+				// Store instruction. The interpreter will create a new object
+				// in the memory view of the function invocation and store to
+				// that, to make it possible to roll back this store.
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+					r.getValue(llvmInst.Operand(1)),
+				}
+			case llvm.Alloca:
+				// Alloca allocates stack space for local variables.
+				numElements := r.getValue(inst.llvmInst.Operand(0)).(literalValue).value.(uint32)
+				elementSize := r.targetData.TypeAllocSize(inst.llvmInst.Type().ElementType())
+				inst.operands = []value{
+					literalValue{elementSize * uint64(numElements)},
+				}
+			case llvm.GetElementPtr:
+				// GetElementPtr does pointer arithmetic.
+				inst.name = llvmInst.Name()
+				ptr := llvmInst.Operand(0)
+				n := llvmInst.OperandsCount()
+				elementType := ptr.Type().ElementType()
+				// gep: [source ptr, dest value size, pairs of indices...]
+				inst.operands = []value{
+					r.getValue(ptr),
+					literalValue{r.targetData.TypeAllocSize(llvmInst.Type().ElementType())},
+					r.getValue(llvmInst.Operand(1)),
+					literalValue{r.targetData.TypeAllocSize(elementType)},
+				}
+				for i := 2; i < n; i++ {
+					operand := r.getValue(llvmInst.Operand(i))
+					if elementType.TypeKind() == llvm.StructTypeKind {
+						index := operand.(literalValue).value.(uint32)
+						elementOffset := r.targetData.ElementOffset(elementType, int(index))
+						// Encode operands in a special way. The elementOffset
+						// is just the offset in bytes. The elementSize is a
+						// negative number (when cast to a int64) by flipping
+						// all the bits. This allows the interpreter to detect
+						// this is a struct field and that it should not
+						// multiply it with the elementOffset to get the offset.
+						// It is important for the interpreter to know the
+						// struct field index for when the GEP must be done at
+						// runtime.
+						inst.operands = append(inst.operands, literalValue{elementOffset}, literalValue{^uint64(index)})
+						elementType = elementType.StructElementTypes()[index]
+					} else {
+						elementType = elementType.ElementType()
+						elementSize := r.targetData.TypeAllocSize(elementType)
+						elementSizeOperand := literalValue{elementSize}
+						// Add operand * elementSizeOperand bytes to the pointer.
+						inst.operands = append(inst.operands, operand, elementSizeOperand)
+					}
+				}
+			case llvm.BitCast, llvm.IntToPtr, llvm.PtrToInt:
+				// Bitcasts are ususally used to cast a pointer from one type to
+				// another leaving the pointer itself intact.
+				inst.name = llvmInst.Name()
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+				}
+			case llvm.ExtractValue:
+				inst.name = llvmInst.Name()
+				agg := llvmInst.Operand(0)
+				var offset uint64
+				indexingType := agg.Type()
+				for _, index := range inst.llvmInst.Indices() {
+					switch indexingType.TypeKind() {
+					case llvm.StructTypeKind:
+						offset += r.targetData.ElementOffset(indexingType, int(index))
+						indexingType = indexingType.StructElementTypes()[index]
+					default: // ArrayTypeKind
+						indexingType = indexingType.ElementType()
+						elementSize := r.targetData.TypeAllocSize(indexingType)
+						offset += elementSize * uint64(index)
+					}
+				}
+				size := r.targetData.TypeAllocSize(inst.llvmInst.Type())
+				// extractvalue [agg, byteOffset, byteSize]
+				inst.operands = []value{
+					r.getValue(agg),
+					literalValue{offset},
+					literalValue{size},
+				}
+			case llvm.InsertValue:
+				inst.name = llvmInst.Name()
+				agg := llvmInst.Operand(0)
+				var offset uint64
+				indexingType := agg.Type()
+				for _, index := range inst.llvmInst.Indices() {
+					switch indexingType.TypeKind() {
+					case llvm.StructTypeKind:
+						offset += r.targetData.ElementOffset(indexingType, int(index))
+						indexingType = indexingType.StructElementTypes()[index]
+					default: // ArrayTypeKind
+						indexingType = indexingType.ElementType()
+						elementSize := r.targetData.TypeAllocSize(indexingType)
+						offset += elementSize * uint64(index)
+					}
+				}
+				// insertvalue [agg, elt, byteOffset]
+				inst.operands = []value{
+					r.getValue(agg),
+					r.getValue(llvmInst.Operand(1)),
+					literalValue{offset},
+				}
+			case llvm.ICmp:
+				inst.name = llvmInst.Name()
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+					r.getValue(llvmInst.Operand(1)),
+					literalValue{uint8(llvmInst.IntPredicate())},
+				}
+			case llvm.FCmp:
+				inst.name = llvmInst.Name()
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+					r.getValue(llvmInst.Operand(1)),
+					literalValue{uint8(llvmInst.FloatPredicate())},
+				}
+			case llvm.Add, llvm.Sub, llvm.Mul, llvm.UDiv, llvm.SDiv, llvm.URem, llvm.SRem, llvm.Shl, llvm.LShr, llvm.AShr, llvm.And, llvm.Or, llvm.Xor:
+				// Integer binary operations.
+				inst.name = llvmInst.Name()
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+					r.getValue(llvmInst.Operand(1)),
+				}
+			case llvm.SExt, llvm.ZExt, llvm.Trunc:
+				// Extend or shrink an integer size.
+				// No sign extension going on so easy to do.
+				// zext: [value, bitwidth]
+				// trunc: [value, bitwidth]
+				inst.name = llvmInst.Name()
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+					literalValue{uint64(llvmInst.Type().IntTypeWidth())},
+				}
+			case llvm.SIToFP, llvm.UIToFP:
+				// Convert an integer to a floating point instruction.
+				// opcode: [value, bitwidth]
+				inst.name = llvmInst.Name()
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+					literalValue{uint64(r.targetData.TypeAllocSize(llvmInst.Type()) * 8)},
+				}
+			default:
+				// Unknown instruction, which is already set in inst.opcode so
+				// is detectable.
+				// This error is handled when actually trying to interpret this
+				// instruction (to not trigger on code that won't be executed).
+			}
+			bb.instructions = append(bb.instructions, inst)
+		}
+	}
+	return fn
+}
+
+// instructionNameMap maps from instruction opcodes to instruction names. This
+// can be useful for debug logging.
+var instructionNameMap = [...]string{
+	llvm.Ret:         "ret",
+	llvm.Br:          "br",
+	llvm.Switch:      "switch",
+	llvm.IndirectBr:  "indirectbr",
+	llvm.Invoke:      "invoke",
+	llvm.Unreachable: "unreachable",
+
+	// Standard Binary Operators
+	llvm.Add:  "add",
+	llvm.FAdd: "fadd",
+	llvm.Sub:  "sub",
+	llvm.FSub: "fsub",
+	llvm.Mul:  "mul",
+	llvm.FMul: "fmul",
+	llvm.UDiv: "udiv",
+	llvm.SDiv: "sdiv",
+	llvm.FDiv: "fdiv",
+	llvm.URem: "urem",
+	llvm.SRem: "srem",
+	llvm.FRem: "frem",
+
+	// Logical Operators
+	llvm.Shl:  "shl",
+	llvm.LShr: "lshr",
+	llvm.AShr: "ashr",
+	llvm.And:  "and",
+	llvm.Or:   "or",
+	llvm.Xor:  "xor",
+
+	// Memory Operators
+	llvm.Alloca:        "alloca",
+	llvm.Load:          "load",
+	llvm.Store:         "store",
+	llvm.GetElementPtr: "getelementptr",
+
+	// Cast Operators
+	llvm.Trunc:    "trunc",
+	llvm.ZExt:     "zext",
+	llvm.SExt:     "sext",
+	llvm.FPToUI:   "fptoui",
+	llvm.FPToSI:   "fptosi",
+	llvm.UIToFP:   "uitofp",
+	llvm.SIToFP:   "sitofp",
+	llvm.FPTrunc:  "fptrunc",
+	llvm.FPExt:    "fpext",
+	llvm.PtrToInt: "ptrtoint",
+	llvm.IntToPtr: "inttoptr",
+	llvm.BitCast:  "bitcast",
+
+	// Other Operators
+	llvm.ICmp:           "icmp",
+	llvm.FCmp:           "fcmp",
+	llvm.PHI:            "phi",
+	llvm.Call:           "call",
+	llvm.Select:         "select",
+	llvm.VAArg:          "vaarg",
+	llvm.ExtractElement: "extractelement",
+	llvm.InsertElement:  "insertelement",
+	llvm.ShuffleVector:  "shufflevector",
+	llvm.ExtractValue:   "extractvalue",
+	llvm.InsertValue:    "insertvalue",
+}
@@ -11,57 +11,52 @@ import (
 	"tinygo.org/x/go-llvm"
 )

-// errUnreachable is returned when an unreachable instruction is executed. This
-// error should not be visible outside of the interp package.
-var errUnreachable = errors.New("interp: unreachable executed")
+var errLiteralToPointer = errors.New("interp: trying to convert literal value to pointer")

-// Unsupported is the specific error that is returned when an unsupported
-// instruction is hit while trying to interpret all initializers.
-type Unsupported struct {
-	ImportPath string
-	Inst       llvm.Value
-	Pos        token.Position
+// These errors are expected during normal execution and can be recovered from
+// by running the affected function at runtime instead of compile time.
+var (
+	errExpectedPointer        = errors.New("interp: trying to use an integer as a pointer (memory-mapped I/O?)")
+	errUnsupportedInst        = errors.New("interp: unsupported instruction")
+	errUnsupportedRuntimeInst = errors.New("interp: unsupported instruction (to be emitted at runtime)")
+	errMapAlreadyCreated      = errors.New("interp: map already created")
+)
+
+func isRecoverableError(err error) bool {
+	return err == errExpectedPointer || err == errUnsupportedInst || err == errUnsupportedRuntimeInst || err == errMapAlreadyCreated
 }

-func (e Unsupported) Error() string {
-	// TODO: how to return the actual instruction string?
-	// It looks like LLVM provides no function for that...
-	return scanner.Error{
-		Pos: e.Pos,
-		Msg: "interp: unsupported instruction",
-	}.Error()
-}
-
-// unsupportedInstructionError returns a new "unsupported instruction" error for
-// the given instruction. It includes source location information, when
-// available.
-func (e *evalPackage) unsupportedInstructionError(inst llvm.Value) *Unsupported {
-	return &Unsupported{
-		ImportPath: e.packagePath,
-		Inst:       inst,
-		Pos:        getPosition(inst),
-	}
+// ErrorLine is one line in a traceback. The position may be missing.
+type ErrorLine struct {
+	Pos  token.Position
+	Inst llvm.Value
 }

 // Error encapsulates compile-time interpretation errors with an associated
 // import path. The errors may not have a precise location attached.
 type Error struct {
 	ImportPath string
-	Errs       []scanner.Error
+	Inst       llvm.Value
+	Pos        token.Position
+	Err        error
+	Traceback  []ErrorLine
 }

 // Error returns the string of the first error in the list of errors.
-func (e Error) Error() string {
-	return e.Errs[0].Error()
+func (e *Error) Error() string {
+	return e.Pos.String() + ": " + e.Err.Error()
 }

 // errorAt returns an error value for the currently interpreted package at the
 // location of the instruction. The location information may not be complete as
 // it depends on debug information in the IR.
-func (e *evalPackage) errorAt(inst llvm.Value, msg string) Error {
-	return Error{
-		ImportPath: e.packagePath,
-		Errs:       []scanner.Error{errorAt(inst, msg)},
+func (r *runner) errorAt(inst instruction, err error) *Error {
+	pos := getPosition(inst.llvmInst)
+	return &Error{
+		ImportPath: r.pkgName,
+		Pos:        pos,
+		Err:        err,
+		Traceback:  []ErrorLine{{pos, inst.llvmInst}},
 	}
 }

@@ -1,647 +0,0 @@
-package interp
-
-// This file implements the core interpretation routines, interpreting single
-// functions.
-
-import (
-	"strings"
-
-	"tinygo.org/x/go-llvm"
-)
-
-type frame struct {
-	*evalPackage
-	fn     llvm.Value
-	locals map[llvm.Value]Value
-}
-
-// evalBasicBlock evaluates a single basic block, returning the return value (if
-// ending with a ret instruction), a list of outgoing basic blocks (if not
-// ending with a ret instruction), or an error on failure.
-// Most of it works at compile time. Some calls get translated into calls to be
-// executed at runtime: calls to functions with side effects, external calls,
-// and operations on the result of such instructions.
-func (fr *frame) evalBasicBlock(bb, incoming llvm.BasicBlock, indent string) (retval Value, outgoing []llvm.Value, err error) {
-	for inst := bb.FirstInstruction(); !inst.IsNil(); inst = llvm.NextInstruction(inst) {
-		if fr.Debug {
-			print(indent)
-			inst.Dump()
-			println()
-		}
-		switch {
-		case !inst.IsABinaryOperator().IsNil():
-			lhs := fr.getLocal(inst.Operand(0)).(*LocalValue).Underlying
-			rhs := fr.getLocal(inst.Operand(1)).(*LocalValue).Underlying
-
-			switch inst.InstructionOpcode() {
-			// Standard binary operators
-			case llvm.Add:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateAdd(lhs, rhs, "")}
-			case llvm.FAdd:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFAdd(lhs, rhs, "")}
-			case llvm.Sub:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateSub(lhs, rhs, "")}
-			case llvm.FSub:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFSub(lhs, rhs, "")}
-			case llvm.Mul:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateMul(lhs, rhs, "")}
-			case llvm.FMul:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFMul(lhs, rhs, "")}
-			case llvm.UDiv:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateUDiv(lhs, rhs, "")}
-			case llvm.SDiv:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateSDiv(lhs, rhs, "")}
-			case llvm.FDiv:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFDiv(lhs, rhs, "")}
-			case llvm.URem:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateURem(lhs, rhs, "")}
-			case llvm.SRem:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateSRem(lhs, rhs, "")}
-			case llvm.FRem:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFRem(lhs, rhs, "")}
-
-			// Logical operators
-			case llvm.Shl:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateShl(lhs, rhs, "")}
-			case llvm.LShr:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateLShr(lhs, rhs, "")}
-			case llvm.AShr:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateAShr(lhs, rhs, "")}
-			case llvm.And:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateAnd(lhs, rhs, "")}
-			case llvm.Or:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateOr(lhs, rhs, "")}
-			case llvm.Xor:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateXor(lhs, rhs, "")}
-
-			default:
-				return nil, nil, fr.unsupportedInstructionError(inst)
-			}
-
-		// Memory operators
-		case !inst.IsAAllocaInst().IsNil():
-			allocType := inst.Type().ElementType()
-			alloca := llvm.AddGlobal(fr.Mod, allocType, fr.packagePath+"$alloca")
-			alloca.SetInitializer(llvm.ConstNull(allocType))
-			alloca.SetLinkage(llvm.InternalLinkage)
-			fr.locals[inst] = &LocalValue{
-				Underlying: alloca,
-				Eval:       fr.Eval,
-			}
-		case !inst.IsALoadInst().IsNil():
-			operand := fr.getLocal(inst.Operand(0)).(*LocalValue)
-			var value llvm.Value
-			if !operand.IsConstant() || inst.IsVolatile() || (!operand.Underlying.IsAConstantExpr().IsNil() && operand.Underlying.Opcode() == llvm.BitCast) {
-				value = fr.builder.CreateLoad(operand.Value(), inst.Name())
-			} else {
-				value = operand.Load()
-			}
-			if value.Type() != inst.Type() {
-				return nil, nil, fr.errorAt(inst, "interp: load: type does not match")
-			}
-			fr.locals[inst] = fr.getValue(value)
-		case !inst.IsAStoreInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			ptr := fr.getLocal(inst.Operand(1))
-			if inst.IsVolatile() {
-				fr.builder.CreateStore(value.Value(), ptr.Value())
-			} else {
-				ptr.Store(value.Value())
-			}
-		case !inst.IsAGetElementPtrInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			llvmIndices := make([]llvm.Value, inst.OperandsCount()-1)
-			for i := range llvmIndices {
-				llvmIndices[i] = inst.Operand(i + 1)
-			}
-			indices := make([]uint32, len(llvmIndices))
-			for i, llvmIndex := range llvmIndices {
-				operand := fr.getLocal(llvmIndex)
-				if !operand.IsConstant() {
-					// Not a constant operation.
-					// This should be detected by the scanner, but isn't at the
-					// moment.
-					return nil, nil, fr.errorAt(inst, "todo: non-const gep")
-				}
-				indices[i] = uint32(operand.Value().ZExtValue())
-			}
-			result, err := value.GetElementPtr(indices)
-			if err != nil {
-				return nil, nil, fr.errorAt(inst, err.Error())
-			}
-			if result.Type() != inst.Type() {
-				return nil, nil, fr.errorAt(inst, "interp: gep: type does not match")
-			}
-			fr.locals[inst] = result
-
-		// Cast operators
-		case !inst.IsATruncInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateTrunc(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsAZExtInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateZExt(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsASExtInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateSExt(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsAFPToUIInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFPToUI(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsAFPToSIInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFPToSI(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsAUIToFPInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateUIToFP(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsASIToFPInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateSIToFP(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsAFPTruncInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFPTrunc(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsAFPExtInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFPExt(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsAPtrToIntInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreatePtrToInt(value.Value(), inst.Type(), "")}
-		case !inst.IsABitCastInst().IsNil() && inst.Type().TypeKind() == llvm.PointerTypeKind:
-			operand := inst.Operand(0)
-			if !operand.IsACallInst().IsNil() {
-				fn := operand.CalledValue()
-				if !fn.IsAFunction().IsNil() && fn.Name() == "runtime.alloc" {
-					continue // special case: bitcast of alloc
-				}
-			}
-			if _, ok := fr.getLocal(operand).(*MapValue); ok {
-				// Special case for runtime.trackPointer calls.
-				// Note: this might not be entirely sound in some rare cases
-				// where the map is stored in a dirty global.
-				uses := getUses(inst)
-				if len(uses) == 1 {
-					use := uses[0]
-					if !use.IsACallInst().IsNil() && !use.CalledValue().IsAFunction().IsNil() && use.CalledValue().Name() == "runtime.trackPointer" {
-						continue
-					}
-				}
-				// It is not possible in Go to bitcast a map value to a pointer.
-				return nil, nil, fr.errorAt(inst, "unimplemented: bitcast of map")
-			}
-			value := fr.getLocal(operand).(*LocalValue)
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateBitCast(value.Value(), inst.Type(), "")}
-
-		// Other operators
-		case !inst.IsAICmpInst().IsNil():
-			lhs := fr.getLocal(inst.Operand(0)).(*LocalValue).Underlying
-			rhs := fr.getLocal(inst.Operand(1)).(*LocalValue).Underlying
-			predicate := inst.IntPredicate()
-			if predicate == llvm.IntEQ {
-				var lhsZero, rhsZero bool
-				var ok1, ok2 bool
-				if lhs.Type().TypeKind() == llvm.PointerTypeKind {
-					// Unfortunately, the const propagation in the IR builder
-					// doesn't handle pointer compares of inttoptr values. So we
-					// implement it manually here.
-					lhsZero, ok1 = isPointerNil(lhs)
-					rhsZero, ok2 = isPointerNil(rhs)
-				}
-				if lhs.Type().TypeKind() == llvm.IntegerTypeKind {
-					lhsZero, ok1 = isZero(lhs)
-					rhsZero, ok2 = isZero(rhs)
-				}
-				if ok1 && ok2 {
-					if lhsZero && rhsZero {
-						// Both are zero, so this icmp is always evaluated to true.
-						fr.locals[inst] = &LocalValue{fr.Eval, llvm.ConstInt(fr.Mod.Context().Int1Type(), 1, false)}
-						continue
-					}
-					if lhsZero != rhsZero {
-						// Only one of them is zero, so this comparison must return false.
-						fr.locals[inst] = &LocalValue{fr.Eval, llvm.ConstInt(fr.Mod.Context().Int1Type(), 0, false)}
-						continue
-					}
-				}
-			}
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateICmp(predicate, lhs, rhs, "")}
-		case !inst.IsAFCmpInst().IsNil():
-			lhs := fr.getLocal(inst.Operand(0)).(*LocalValue).Underlying
-			rhs := fr.getLocal(inst.Operand(1)).(*LocalValue).Underlying
-			predicate := inst.FloatPredicate()
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFCmp(predicate, lhs, rhs, "")}
-		case !inst.IsAPHINode().IsNil():
-			for i := 0; i < inst.IncomingCount(); i++ {
-				if inst.IncomingBlock(i) == incoming {
-					fr.locals[inst] = fr.getLocal(inst.IncomingValue(i))
-				}
-			}
-		case !inst.IsACallInst().IsNil():
-			callee := inst.CalledValue()
-			switch {
-			case callee.Name() == "runtime.alloc":
-				// heap allocation
-				users := getUses(inst)
-				var resultInst = inst
-				if len(users) == 1 && !users[0].IsABitCastInst().IsNil() {
-					// happens when allocating something other than i8*
-					resultInst = users[0]
-				}
-				size := fr.getLocal(inst.Operand(0)).(*LocalValue).Underlying.ZExtValue()
-				allocType := resultInst.Type().ElementType()
-				typeSize := fr.TargetData.TypeAllocSize(allocType)
-				elementCount := 1
-				if size != typeSize {
-					// allocate an array
-					if size%typeSize != 0 {
-						return nil, nil, fr.unsupportedInstructionError(inst)
-					}
-					elementCount = int(size / typeSize)
-					allocType = llvm.ArrayType(allocType, elementCount)
-				}
-				alloc := llvm.AddGlobal(fr.Mod, allocType, fr.packagePath+"$alloc")
-				alloc.SetInitializer(llvm.ConstNull(allocType))
-				alloc.SetLinkage(llvm.InternalLinkage)
-				result := &LocalValue{
-					Underlying: alloc,
-					Eval:       fr.Eval,
-				}
-				if elementCount == 1 {
-					fr.locals[resultInst] = result
-				} else {
-					result, err := result.GetElementPtr([]uint32{0, 0})
-					if err != nil {
-						return nil, nil, errorAt(inst, err.Error())
-					}
-					fr.locals[resultInst] = result
-				}
-			case callee.Name() == "runtime.hashmapMake":
-				// create a map
-				keySize := inst.Operand(0).ZExtValue()
-				valueSize := inst.Operand(1).ZExtValue()
-				fr.locals[inst] = &MapValue{
-					Eval:      fr.Eval,
-					PkgName:   fr.packagePath,
-					KeySize:   int(keySize),
-					ValueSize: int(valueSize),
-				}
-			case callee.Name() == "runtime.hashmapStringSet":
-				// set a string key in the map
-				keyBuf := fr.getLocal(inst.Operand(1)).(*LocalValue)
-				keyLen := fr.getLocal(inst.Operand(2)).(*LocalValue)
-				valPtr := fr.getLocal(inst.Operand(3)).(*LocalValue)
-				m, ok := fr.getLocal(inst.Operand(0)).(*MapValue)
-				if !ok || !keyBuf.IsConstant() || !keyLen.IsConstant() || !valPtr.IsConstant() {
-					// The mapassign operation could not be done at compile
-					// time. Do it at runtime instead.
-					m := fr.getLocal(inst.Operand(0)).Value()
-					fr.markDirty(m)
-					llvmParams := []llvm.Value{
-						m,                                    // *runtime.hashmap
-						fr.getLocal(inst.Operand(1)).Value(), // key.ptr
-						fr.getLocal(inst.Operand(2)).Value(), // key.len
-						fr.getLocal(inst.Operand(3)).Value(), // value (unsafe.Pointer)
-						fr.getLocal(inst.Operand(4)).Value(), // context
-						fr.getLocal(inst.Operand(5)).Value(), // parentHandle
-					}
-					fr.builder.CreateCall(callee, llvmParams, "")
-					continue
-				}
-				// "key" is a Go string value, which in the TinyGo calling convention is split up
-				// into separate pointer and length parameters.
-				m.PutString(keyBuf, keyLen, valPtr)
-			case callee.Name() == "runtime.hashmapBinarySet":
-				// set a binary (int etc.) key in the map
-				keyBuf := fr.getLocal(inst.Operand(1)).(*LocalValue)
-				valPtr := fr.getLocal(inst.Operand(2)).(*LocalValue)
-				m, ok := fr.getLocal(inst.Operand(0)).(*MapValue)
-				if !ok || !keyBuf.IsConstant() || !valPtr.IsConstant() {
-					// The mapassign operation could not be done at compile
-					// time. Do it at runtime instead.
-					m := fr.getLocal(inst.Operand(0)).Value()
-					fr.markDirty(m)
-					llvmParams := []llvm.Value{
-						m,                                    // *runtime.hashmap
-						fr.getLocal(inst.Operand(1)).Value(), // key
-						fr.getLocal(inst.Operand(2)).Value(), // value
-						fr.getLocal(inst.Operand(3)).Value(), // context
-						fr.getLocal(inst.Operand(4)).Value(), // parentHandle
-					}
-					fr.builder.CreateCall(callee, llvmParams, "")
-					continue
-				}
-				m.PutBinary(keyBuf, valPtr)
-			case callee.Name() == "runtime.stringConcat":
-				// adding two strings together
-				buf1Ptr := fr.getLocal(inst.Operand(0))
-				buf1Len := fr.getLocal(inst.Operand(1))
-				buf2Ptr := fr.getLocal(inst.Operand(2))
-				buf2Len := fr.getLocal(inst.Operand(3))
-				buf1 := getStringBytes(buf1Ptr, buf1Len.Value())
-				buf2 := getStringBytes(buf2Ptr, buf2Len.Value())
-				result := []byte(string(buf1) + string(buf2))
-				vals := make([]llvm.Value, len(result))
-				for i := range vals {
-					vals[i] = llvm.ConstInt(fr.Mod.Context().Int8Type(), uint64(result[i]), false)
-				}
-				globalType := llvm.ArrayType(fr.Mod.Context().Int8Type(), len(result))
-				globalValue := llvm.ConstArray(fr.Mod.Context().Int8Type(), vals)
-				global := llvm.AddGlobal(fr.Mod, globalType, fr.packagePath+"$stringconcat")
-				global.SetInitializer(globalValue)
-				global.SetLinkage(llvm.InternalLinkage)
-				global.SetGlobalConstant(true)
-				global.SetUnnamedAddr(true)
-				stringType := fr.Mod.GetTypeByName("runtime._string")
-				retPtr := llvm.ConstGEP(global, getLLVMIndices(fr.Mod.Context().Int32Type(), []uint32{0, 0}))
-				retLen := llvm.ConstInt(stringType.StructElementTypes()[1], uint64(len(result)), false)
-				ret := llvm.ConstNull(stringType)
-				ret = llvm.ConstInsertValue(ret, retPtr, []uint32{0})
-				ret = llvm.ConstInsertValue(ret, retLen, []uint32{1})
-				fr.locals[inst] = &LocalValue{fr.Eval, ret}
-			case callee.Name() == "runtime.sliceCopy":
-				elementSize := fr.getLocal(inst.Operand(4)).(*LocalValue).Value().ZExtValue()
-				dstArray := fr.getLocal(inst.Operand(0)).(*LocalValue).stripPointerCasts()
-				srcArray := fr.getLocal(inst.Operand(1)).(*LocalValue).stripPointerCasts()
-				dstLen := fr.getLocal(inst.Operand(2)).(*LocalValue)
-				srcLen := fr.getLocal(inst.Operand(3)).(*LocalValue)
-				if elementSize != 1 && dstArray.Type().ElementType().TypeKind() == llvm.ArrayTypeKind && srcArray.Type().ElementType().TypeKind() == llvm.ArrayTypeKind {
-					// Slice data pointers are created by adding a global array
-					// and getting the address of the first element using a GEP.
-					// However, before the compiler can pass it to
-					// runtime.sliceCopy, it has to perform a bitcast to a *i8,
-					// to make it a unsafe.Pointer. Now, when the IR builder
-					// sees a bitcast of a GEP with zero indices, it will make
-					// a bitcast of the original array instead of the GEP,
-					// which breaks our assumptions.
-					// Re-add this GEP, in the hope that it it is then of the correct type...
-					dstArrayValue, err := dstArray.GetElementPtr([]uint32{0, 0})
-					if err != nil {
-						return nil, nil, errorAt(inst, err.Error())
-					}
-					dstArray = dstArrayValue.(*LocalValue)
-					srcArrayValue, err := srcArray.GetElementPtr([]uint32{0, 0})
-					if err != nil {
-						return nil, nil, errorAt(inst, err.Error())
-					}
-					srcArray = srcArrayValue.(*LocalValue)
-				}
-				if fr.Eval.TargetData.TypeAllocSize(dstArray.Type().ElementType()) != elementSize {
-					return nil, nil, fr.errorAt(inst, "interp: slice dst element size does not match pointer type")
-				}
-				if fr.Eval.TargetData.TypeAllocSize(srcArray.Type().ElementType()) != elementSize {
-					return nil, nil, fr.errorAt(inst, "interp: slice src element size does not match pointer type")
-				}
-				if dstArray.Type() != srcArray.Type() {
-					return nil, nil, fr.errorAt(inst, "interp: slice element types don't match")
-				}
-				length := dstLen.Value().SExtValue()
-				if srcLength := srcLen.Value().SExtValue(); srcLength < length {
-					length = srcLength
-				}
-				if length < 0 {
-					return nil, nil, fr.errorAt(inst, "interp: trying to copy a slice with negative length?")
-				}
-				for i := int64(0); i < length; i++ {
-					var err error
-					// *dst = *src
-					dstArray.Store(srcArray.Load())
-					// dst++
-					dstArrayValue, err := dstArray.GetElementPtr([]uint32{1})
-					if err != nil {
-						return nil, nil, errorAt(inst, err.Error())
-					}
-					dstArray = dstArrayValue.(*LocalValue)
-					// src++
-					srcArrayValue, err := srcArray.GetElementPtr([]uint32{1})
-					if err != nil {
-						return nil, nil, errorAt(inst, err.Error())
-					}
-					srcArray = srcArrayValue.(*LocalValue)
-				}
-			case callee.Name() == "runtime.stringToBytes":
-				// convert a string to a []byte
-				bufPtr := fr.getLocal(inst.Operand(0))
-				bufLen := fr.getLocal(inst.Operand(1))
-				result := getStringBytes(bufPtr, bufLen.Value())
-				vals := make([]llvm.Value, len(result))
-				for i := range vals {
-					vals[i] = llvm.ConstInt(fr.Mod.Context().Int8Type(), uint64(result[i]), false)
-				}
-				globalType := llvm.ArrayType(fr.Mod.Context().Int8Type(), len(result))
-				globalValue := llvm.ConstArray(fr.Mod.Context().Int8Type(), vals)
-				global := llvm.AddGlobal(fr.Mod, globalType, fr.packagePath+"$bytes")
-				global.SetInitializer(globalValue)
-				global.SetLinkage(llvm.InternalLinkage)
-				global.SetGlobalConstant(true)
-				global.SetUnnamedAddr(true)
-				sliceType := inst.Type()
-				retPtr := llvm.ConstGEP(global, getLLVMIndices(fr.Mod.Context().Int32Type(), []uint32{0, 0}))
-				retLen := llvm.ConstInt(sliceType.StructElementTypes()[1], uint64(len(result)), false)
-				ret := llvm.ConstNull(sliceType)
-				ret = llvm.ConstInsertValue(ret, retPtr, []uint32{0}) // ptr
-				ret = llvm.ConstInsertValue(ret, retLen, []uint32{1}) // len
-				ret = llvm.ConstInsertValue(ret, retLen, []uint32{2}) // cap
-				fr.locals[inst] = &LocalValue{fr.Eval, ret}
-			case callee.Name() == "runtime.interfaceImplements":
-				typecode := fr.getLocal(inst.Operand(0)).(*LocalValue).Underlying
-				interfaceMethodSet := fr.getLocal(inst.Operand(1)).(*LocalValue).Underlying
-				if typecode.IsAConstantExpr().IsNil() || typecode.Opcode() != llvm.PtrToInt {
-					return nil, nil, fr.errorAt(inst, "interp: expected typecode to be a ptrtoint")
-				}
-				typecode = typecode.Operand(0)
-				if interfaceMethodSet.IsAConstantExpr().IsNil() || interfaceMethodSet.Opcode() != llvm.GetElementPtr {
-					return nil, nil, fr.errorAt(inst, "interp: expected method set in runtime.interfaceImplements to be a constant gep")
-				}
-				interfaceMethodSet = interfaceMethodSet.Operand(0).Initializer()
-				methodSet := llvm.ConstExtractValue(typecode.Initializer(), []uint32{1})
-				if methodSet.IsAConstantExpr().IsNil() || methodSet.Opcode() != llvm.GetElementPtr {
-					return nil, nil, fr.errorAt(inst, "interp: expected method set to be a constant gep")
-				}
-				methodSet = methodSet.Operand(0).Initializer()
-
-				// Make a set of all the methods on the concrete type, for
-				// easier checking in the next step.
-				definedMethods := map[string]struct{}{}
-				for i := 0; i < methodSet.Type().ArrayLength(); i++ {
-					methodInfo := llvm.ConstExtractValue(methodSet, []uint32{uint32(i)})
-					name := llvm.ConstExtractValue(methodInfo, []uint32{0}).Name()
-					definedMethods[name] = struct{}{}
-				}
-				// Check whether all interface methods are also in the list
-				// of defined methods calculated above.
-				implements := uint64(1) // i1 true
-				for i := 0; i < interfaceMethodSet.Type().ArrayLength(); i++ {
-					name := llvm.ConstExtractValue(interfaceMethodSet, []uint32{uint32(i)}).Name()
-					if _, ok := definedMethods[name]; !ok {
-						// There is a method on the interface that is not
-						// implemented by the type.
-						implements = 0 // i1 false
-						break
-					}
-				}
-				fr.locals[inst] = &LocalValue{fr.Eval, llvm.ConstInt(fr.Mod.Context().Int1Type(), implements, false)}
-			case callee.Name() == "runtime.nanotime":
-				fr.locals[inst] = &LocalValue{fr.Eval, llvm.ConstInt(fr.Mod.Context().Int64Type(), 0, false)}
-			case callee.Name() == "llvm.dbg.value":
-				// do nothing
-			case strings.HasPrefix(callee.Name(), "llvm.lifetime."):
-				// do nothing
-			case callee.Name() == "runtime.trackPointer":
-				// do nothing
-			case strings.HasPrefix(callee.Name(), "runtime.print") || callee.Name() == "runtime._panic":
-				// This are all print instructions, which necessarily have side
-				// effects but no results.
-				// TODO: print an error when executing runtime._panic (with the
-				// exact error message it would print at runtime).
-				var params []llvm.Value
-				for i := 0; i < inst.OperandsCount()-1; i++ {
-					operand := fr.getLocal(inst.Operand(i)).Value()
-					fr.markDirty(operand)
-					params = append(params, operand)
-				}
-				// TODO: accurate debug info, including call chain
-				fr.builder.CreateCall(callee, params, inst.Name())
-			case !callee.IsAFunction().IsNil() && callee.IsDeclaration():
-				// external functions
-				var params []llvm.Value
-				for i := 0; i < inst.OperandsCount()-1; i++ {
-					operand := fr.getLocal(inst.Operand(i)).Value()
-					fr.markDirty(operand)
-					params = append(params, operand)
-				}
-				// TODO: accurate debug info, including call chain
-				result := fr.builder.CreateCall(callee, params, inst.Name())
-				if inst.Type().TypeKind() != llvm.VoidTypeKind {
-					fr.markDirty(result)
-					fr.locals[inst] = &LocalValue{fr.Eval, result}
-				}
-			case !callee.IsAFunction().IsNil():
-				// regular function
-				var params []Value
-				dirtyParams := false
-				for i := 0; i < inst.OperandsCount()-1; i++ {
-					local := fr.getLocal(inst.Operand(i))
-					if !local.IsConstant() {
-						dirtyParams = true
-					}
-					params = append(params, local)
-				}
-				var ret Value
-				scanResult, err := fr.hasSideEffects(callee)
-				if err != nil {
-					return nil, nil, err
-				}
-				if scanResult.severity == sideEffectLimited || dirtyParams && scanResult.severity != sideEffectAll {
-					// Side effect is bounded. This means the operation invokes
-					// side effects (like calling an external function) but it
-					// is known at compile time which side effects it invokes.
-					// This means the function can be called at runtime and the
-					// affected globals can be marked dirty at compile time.
-					llvmParams := make([]llvm.Value, len(params))
-					for i, param := range params {
-						llvmParams[i] = param.Value()
-					}
-					result := fr.builder.CreateCall(callee, llvmParams, inst.Name())
-					ret = &LocalValue{fr.Eval, result}
-					// mark all mentioned globals as dirty
-					for global := range scanResult.mentionsGlobals {
-						fr.markDirty(global)
-					}
-				} else {
-					// Side effect is one of:
-					//   * None: no side effects, can be fully interpreted at
-					//     compile time.
-					//   * Unbounded: cannot call at runtime so we'll try to
-					//     interpret anyway and hope for the best.
-					ret, err = fr.function(callee, params, indent+"    ")
-					if err != nil {
-						return nil, nil, err
-					}
-				}
-				if inst.Type().TypeKind() != llvm.VoidTypeKind {
-					fr.locals[inst] = ret
-				}
-			default:
-				// function pointers, etc.
-				return nil, nil, fr.unsupportedInstructionError(inst)
-			}
-		case !inst.IsAExtractValueInst().IsNil():
-			agg := fr.getLocal(inst.Operand(0)).(*LocalValue) // must be constant
-			indices := inst.Indices()
-			if agg.Underlying.IsConstant() {
-				newValue := llvm.ConstExtractValue(agg.Underlying, indices)
-				fr.locals[inst] = fr.getValue(newValue)
-			} else {
-				if len(indices) != 1 {
-					return nil, nil, fr.errorAt(inst, "interp: cannot handle extractvalue with not exactly 1 index")
-				}
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateExtractValue(agg.Underlying, int(indices[0]), inst.Name())}
-			}
-		case !inst.IsAInsertValueInst().IsNil():
-			agg := fr.getLocal(inst.Operand(0)).(*LocalValue) // must be constant
-			val := fr.getLocal(inst.Operand(1))
-			indices := inst.Indices()
-			if agg.IsConstant() && val.IsConstant() {
-				newValue := llvm.ConstInsertValue(agg.Underlying, val.Value(), indices)
-				fr.locals[inst] = &LocalValue{fr.Eval, newValue}
-			} else {
-				if len(indices) != 1 {
-					return nil, nil, fr.errorAt(inst, "interp: cannot handle insertvalue with not exactly 1 index")
-				}
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateInsertValue(agg.Underlying, val.Value(), int(indices[0]), inst.Name())}
-			}
-
-		case !inst.IsAReturnInst().IsNil() && inst.OperandsCount() == 0:
-			return nil, nil, nil // ret void
-		case !inst.IsAReturnInst().IsNil() && inst.OperandsCount() == 1:
-			return fr.getLocal(inst.Operand(0)), nil, nil
-		case !inst.IsABranchInst().IsNil() && inst.OperandsCount() == 3:
-			// conditional branch (if/then/else)
-			cond := fr.getLocal(inst.Operand(0)).Value()
-			if cond.Type() != fr.Mod.Context().Int1Type() {
-				return nil, nil, fr.errorAt(inst, "expected an i1 in a branch instruction")
-			}
-			thenBB := inst.Operand(1)
-			elseBB := inst.Operand(2)
-			if !cond.IsAInstruction().IsNil() {
-				return nil, nil, fr.errorAt(inst, "interp: branch on a non-constant")
-			}
-			if !cond.IsAConstantExpr().IsNil() {
-				// This may happen when the instruction builder could not
-				// const-fold some instructions.
-				return nil, nil, fr.errorAt(inst, "interp: branch on a non-const-propagated constant expression")
-			}
-			switch cond {
-			case llvm.ConstInt(fr.Mod.Context().Int1Type(), 0, false): // false
-				return nil, []llvm.Value{thenBB}, nil // then
-			case llvm.ConstInt(fr.Mod.Context().Int1Type(), 1, false): // true
-				return nil, []llvm.Value{elseBB}, nil // else
-			default:
-				return nil, nil, fr.errorAt(inst, "branch was not true or false")
-			}
-		case !inst.IsABranchInst().IsNil() && inst.OperandsCount() == 1:
-			// unconditional branch (goto)
-			return nil, []llvm.Value{inst.Operand(0)}, nil
-		case !inst.IsAUnreachableInst().IsNil():
-			// Unreachable was reached (e.g. after a call to panic()).
-			// Report this as an error, as it is not supposed to happen.
-			// This is a sentinel error value.
-			return nil, nil, errUnreachable
-
-		default:
-			return nil, nil, fr.unsupportedInstructionError(inst)
-		}
-	}
-
-	panic("interp: reached end of basic block without terminator")
-}
-
-// Get the Value for an operand, which is a constant value of some sort.
-func (fr *frame) getLocal(v llvm.Value) Value {
-	if ret, ok := fr.locals[v]; ok {
-		return ret
-	} else if value := fr.getValue(v); value != nil {
-		return value
-	} else {
-		// This should not happen under normal circumstances.
-		panic("cannot find value")
-	}
-}
@@ -1,59 +1,77 @@
-// Package interp interprets Go package initializers as much as possible. This
-// avoid running them at runtime, improving code size and making other
-// optimizations possible.
+// Package interp is a partial evaluator of code run at package init time. See
+// the README in this package for details.
 package interp

-// This file provides the overarching Eval object with associated (utility)
-// methods.
-
 import (
+	"fmt"
+	"os"
 	"strings"
+	"time"

 	"tinygo.org/x/go-llvm"
 )

-type Eval struct {
-	Mod             llvm.Module
-	TargetData      llvm.TargetData
-	Debug           bool
-	builder         llvm.Builder
-	dirtyGlobals    map[llvm.Value]struct{}
-	sideEffectFuncs map[llvm.Value]*sideEffectResult // cache of side effect scan results
+// Enable extra checks, which should be disabled by default.
+// This may help track down bugs by adding a few more sanity checks.
+const checks = true
+
+// runner contains all state related to one interp run.
+type runner struct {
+	mod           llvm.Module
+	targetData    llvm.TargetData
+	builder       llvm.Builder
+	pointerSize   uint32                   // cached pointer size from the TargetData
+	i8ptrType     llvm.Type                // often used type so created in advance
+	maxAlign      int                      // maximum alignment of an object, alignment of runtime.alloc() result
+	debug         bool                     // log debug messages
+	pkgName       string                   // package name of the currently executing package
+	functionCache map[llvm.Value]*function // cache of compiled functions
+	objects       []object                 // slice of objects in memory
+	globals       map[llvm.Value]int       // map from global to index in objects slice
+	start         time.Time
+	callsExecuted uint64
 }

-// evalPackage encapsulates the Eval type for just a single package. The Eval
-// type keeps state across the whole program, the evalPackage type keeps extra
-// state for the currently interpreted package.
-type evalPackage struct {
-	*Eval
-	packagePath string
-}
-
-// Run evaluates the function with the given name and then eliminates all
-// callers.
+// Run evaluates runtime.initAll function as much as possible at compile time.
+// Set debug to true if it should print output while running.
 func Run(mod llvm.Module, debug bool) error {
-	if debug {
-		println("\ncompile-time evaluation:")
+	r := runner{
+		mod:           mod,
+		targetData:    llvm.NewTargetData(mod.DataLayout()),
+		debug:         debug,
+		functionCache: make(map[llvm.Value]*function),
+		objects:       []object{{}},
+		globals:       make(map[llvm.Value]int),
+		start:         time.Now(),
 	}
+	r.pointerSize = uint32(r.targetData.PointerSize())
+	r.i8ptrType = llvm.PointerType(mod.Context().Int8Type(), 0)
+	r.maxAlign = r.targetData.PrefTypeAlignment(r.i8ptrType) // assume pointers are maximally aligned (this is not always the case)

-	name := "runtime.initAll"
-	e := &Eval{
-		Mod:          mod,
-		TargetData:   llvm.NewTargetData(mod.DataLayout()),
-		Debug:        debug,
-		dirtyGlobals: map[llvm.Value]struct{}{},
-	}
-	e.builder = mod.Context().NewBuilder()
-
-	initAll := mod.NamedFunction(name)
+	initAll := mod.NamedFunction("runtime.initAll")
 	bb := initAll.EntryBasicBlock()
+
+	// Create a builder, to insert instructions that could not be evaluated at
+	// compile time.
+	r.builder = mod.Context().NewBuilder()
+	defer r.builder.Dispose()
+
 	// Create a dummy alloca in the entry block that we can set the insert point
 	// to. This is necessary because otherwise we might be removing the
 	// instruction (init call) that we are removing after successful
 	// interpretation.
-	e.builder.SetInsertPointBefore(bb.FirstInstruction())
-	dummy := e.builder.CreateAlloca(e.Mod.Context().Int8Type(), "dummy")
-	e.builder.SetInsertPointBefore(dummy)
+	r.builder.SetInsertPointBefore(bb.FirstInstruction())
+	dummy := r.builder.CreateAlloca(r.mod.Context().Int8Type(), "dummy")
+	r.builder.SetInsertPointBefore(dummy)
+	defer dummy.EraseFromParentAsInstruction()
+
+	// Get a list if init calls. A runtime.initAll might look something like this:
+	// func initAll() {
+	//     unsafe.init()
+	//     machine.init()
+	//     runtime.init()
+	// }
+	// This function gets a list of these call instructions.
 	var initCalls []llvm.Value
 	for inst := bb.FirstInstruction(); !inst.IsNil(); inst = llvm.NextInstruction(inst) {
 		if inst == dummy {
@@ -63,99 +81,67 @@ func Run(mod llvm.Module, debug bool) error {
 			break // ret void
 		}
 		if inst.IsACallInst().IsNil() || inst.CalledValue().IsAFunction().IsNil() {
-			return errorAt(inst, "interp: expected all instructions in "+name+" to be direct calls")
+			return errorAt(inst, "interp: expected all instructions in "+initAll.Name()+" to be direct calls")
 		}
 		initCalls = append(initCalls, inst)
 	}

-	// Do this in a separate step to avoid corrupting the iterator above.
-	undefPtr := llvm.Undef(llvm.PointerType(mod.Context().Int8Type(), 0))
+	// Run initializers for each package. Once the package initializer is
+	// finished, the call to the package initializer can be removed.
 	for _, call := range initCalls {
 		initName := call.CalledValue().Name()
 		if !strings.HasSuffix(initName, ".init") {
-			return errorAt(call, "interp: expected all instructions in "+name+" to be *.init() calls")
+			return errorAt(call, "interp: expected all instructions in "+initAll.Name()+" to be *.init() calls")
 		}
-		pkgName := initName[:len(initName)-5]
+		r.pkgName = initName[:len(initName)-len(".init")]
 		fn := call.CalledValue()
+		if r.debug {
+			fmt.Fprintln(os.Stderr, "call:", fn.Name())
+		}
+		_, mem, callErr := r.run(r.getFunction(fn), nil, nil, "    ")
+		if callErr != nil {
+			if isRecoverableError(callErr.Err) {
+				if r.debug {
+					fmt.Fprintln(os.Stderr, "not interpretring", r.pkgName, "because of error:", callErr.Err)
+				}
+				mem.revert()
+				continue
+			}
+			return callErr
+		}
 		call.EraseFromParentAsInstruction()
-		evalPkg := evalPackage{
-			Eval:        e,
-			packagePath: pkgName,
+		for index, obj := range mem.objects {
+			r.objects[index] = obj
 		}
-		_, err := evalPkg.function(fn, []Value{&LocalValue{e, undefPtr}, &LocalValue{e, undefPtr}}, "")
-		if err == errUnreachable {
-			break
+	}
+	r.pkgName = ""
+
+	// Update all global variables in the LLVM module.
+	mem := memoryView{r: &r}
+	for _, obj := range r.objects {
+		if obj.llvmGlobal.IsNil() {
+			continue
 		}
-		if err != nil {
-			return err
+		if obj.buffer == nil {
+			continue
 		}
+		initializer := obj.buffer.toLLVMValue(obj.llvmGlobal.Type().ElementType(), &mem)
+		if checks && initializer.Type() != obj.llvmGlobal.Type().ElementType() {
+			panic("initializer type mismatch")
+		}
+		obj.llvmGlobal.SetInitializer(initializer)
 	}

 	return nil
 }

-// function interprets the given function. The params are the function params
-// and the indent is the string indentation to use when dumping all interpreted
-// instructions.
-func (e *evalPackage) function(fn llvm.Value, params []Value, indent string) (Value, error) {
-	fr := frame{
-		evalPackage: e,
-		fn:          fn,
-		locals:      make(map[llvm.Value]Value),
-	}
-	for i, param := range fn.Params() {
-		fr.locals[param] = params[i]
-	}
-
-	bb := fn.EntryBasicBlock()
-	var lastBB llvm.BasicBlock
-	for {
-		retval, outgoing, err := fr.evalBasicBlock(bb, lastBB, indent)
-		if outgoing == nil {
-			// returned something (a value or void, or an error)
-			return retval, err
-		}
-		if len(outgoing) > 1 {
-			panic("unimplemented: multiple outgoing blocks")
-		}
-		next := outgoing[0]
-		if next.IsABasicBlock().IsNil() {
-			panic("did not switch to a basic block")
-		}
-		lastBB = bb
-		bb = next.AsBasicBlock()
-	}
-}
-
-// getValue determines what kind of LLVM value it gets and returns the
-// appropriate Value type.
-func (e *Eval) getValue(v llvm.Value) Value {
-	return &LocalValue{e, v}
-}
-
-// markDirty marks the passed-in LLVM value dirty, recursively. For example,
-// when it encounters a constant GEP on a global, it marks the global dirty.
-func (e *Eval) markDirty(v llvm.Value) {
-	if !v.IsAGlobalVariable().IsNil() {
-		if v.IsGlobalConstant() {
-			return
-		}
-		if _, ok := e.dirtyGlobals[v]; !ok {
-			e.dirtyGlobals[v] = struct{}{}
-			e.sideEffectFuncs = nil // re-calculate all side effects
-		}
-	} else if v.IsConstant() {
-		if v.OperandsCount() >= 2 && !v.Operand(0).IsAGlobalVariable().IsNil() {
-			// looks like a constant getelementptr of a global.
-			// TODO: find a way to make sure it really is: v.Opcode() returns 0.
-			e.markDirty(v.Operand(0))
-			return
-		}
-		return // nothing to mark
-	} else if !v.IsAGetElementPtrInst().IsNil() {
-		panic("interp: todo: GEP")
-	} else {
-		// Not constant and not a global or GEP so doesn't have to be marked
-		// non-constant.
+// getFunction returns the compiled version of the given LLVM function. It
+// compiles the function if necessary and caches the result.
+func (r *runner) getFunction(llvmFn llvm.Value) *function {
+	if fn, ok := r.functionCache[llvmFn]; ok {
+		return fn
 	}
+	fn := r.compileFunction(llvmFn)
+	r.functionCache[llvmFn] = fn
+	return fn
 }
@@ -3,6 +3,7 @@ package interp
 import (
 	"io/ioutil"
 	"os"
+	"regexp"
 	"strings"
 	"testing"

@@ -15,6 +16,7 @@ func TestInterp(t *testing.T) {
 		"slice-copy",
 		"consteval",
 		"map",
+		"interface",
 	} {
 		name := name // make tc local to this closure
 		t.Run(name, func(t *testing.T) {
@@ -40,9 +42,29 @@ func runTest(t *testing.T, pathPrefix string) {
 	// Perform the transform.
 	err = Run(mod, false)
 	if err != nil {
+		if err, match := err.(*Error); match {
+			println(err.Error())
+			if !err.Inst.IsNil() {
+				err.Inst.Dump()
+				println()
+			}
+			if len(err.Traceback) > 0 {
+				println("\ntraceback:")
+				for _, line := range err.Traceback {
+					println(line.Pos.String() + ":")
+					line.Inst.Dump()
+					println()
+				}
+			}
+		}
 		t.Fatal(err)
 	}

+	// To be sure, verify that the module is still valid.
+	if llvm.VerifyModule(mod, llvm.PrintMessageAction) != nil {
+		t.FailNow()
+	}
+
 	// Run some cleanup passes to get easy-to-read outputs.
 	pm := llvm.NewPassManager()
 	defer pm.Dispose()
@@ -65,6 +87,8 @@ func runTest(t *testing.T, pathPrefix string) {
 	}
 }

+var alignRegexp = regexp.MustCompile(", align [0-9]+$")
+
 // fuzzyEqualIR returns true if the two LLVM IR strings passed in are roughly
 // equal. That means, only relevant lines are compared (excluding comments
 // etc.).
@@ -74,8 +98,18 @@ func fuzzyEqualIR(s1, s2 string) bool {
 	if len(lines1) != len(lines2) {
 		return false
 	}
-	for i, line := range lines1 {
-		if line != lines2[i] {
+	for i, line1 := range lines1 {
+		line2 := lines2[i]
+		match1 := alignRegexp.MatchString(line1)
+		match2 := alignRegexp.MatchString(line2)
+		if match1 != match2 {
+			// Only one of the lines has the align keyword. Remove it.
+			// This is a change to make the test work in both LLVM 10 and LLVM
+			// 11 (LLVM 11 appears to automatically add alignment everywhere).
+			line1 = alignRegexp.ReplaceAllString(line1, "")
+			line2 = alignRegexp.ReplaceAllString(line2, "")
+		}
+		if line1 != line2 {
 			return false
 		}
 	}
@@ -0,0 +1,917 @@
+package interp
+
+import (
+	"errors"
+	"fmt"
+	"math"
+	"os"
+	"strings"
+	"time"
+
+	"tinygo.org/x/go-llvm"
+)
+
+func (r *runner) run(fn *function, params []value, parentMem *memoryView, indent string) (value, memoryView, *Error) {
+	mem := memoryView{r: r, parent: parentMem}
+	locals := make([]value, len(fn.locals))
+	r.callsExecuted++
+
+	if time.Since(r.start) > time.Minute {
+		// Running for more than a minute. This should never happen.
+		return nil, mem, r.errorAt(fn.blocks[0].instructions[0], fmt.Errorf("interp: running for more than a minute, timing out (executed calls: %d)", r.callsExecuted))
+	}
+
+	// Parameters are considered a kind of local values.
+	for i, param := range params {
+		locals[i] = param
+	}
+
+	// Start with the first basic block and the first instruction.
+	// Branch instructions may modify both bb and instIndex when branching.
+	bb := fn.blocks[0]
+	currentBB := 0
+	lastBB := -1 // last basic block is undefined, only defined after a branch
+	var operands []value
+	for instIndex := 0; instIndex < len(bb.instructions); instIndex++ {
+		inst := bb.instructions[instIndex]
+		operands = operands[:0]
+		isRuntimeInst := false
+		if inst.opcode != llvm.PHI {
+			for _, v := range inst.operands {
+				if v, ok := v.(localValue); ok {
+					if localVal := locals[fn.locals[v.value]]; localVal == nil {
+						return nil, mem, r.errorAt(inst, errors.New("interp: local not defined"))
+					} else {
+						operands = append(operands, localVal)
+						if _, ok := localVal.(localValue); ok {
+							isRuntimeInst = true
+						}
+						continue
+					}
+				}
+				operands = append(operands, v)
+			}
+		}
+		if isRuntimeInst {
+			err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+			if err != nil {
+				return nil, mem, err
+			}
+			continue
+		}
+		switch inst.opcode {
+		case llvm.Ret:
+			if len(operands) != 0 {
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"ret", operands[0])
+				}
+				// Return instruction has a value to return.
+				return operands[0], mem, nil
+			}
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"ret")
+			}
+			// Return instruction doesn't return anything, it's just 'ret void'.
+			return nil, mem, nil
+		case llvm.Br:
+			switch len(operands) {
+			case 1:
+				// Unconditional branch: [nextBB]
+				lastBB = currentBB
+				currentBB = int(operands[0].(literalValue).value.(uint32))
+				bb = fn.blocks[currentBB]
+				instIndex = -1 // start at 0 the next cycle
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"br", operands, "->", currentBB)
+				}
+			case 3:
+				// Conditional branch: [cond, thenBB, elseBB]
+				lastBB = currentBB
+				switch operands[0].Uint() {
+				case 1: // true -> thenBB
+					currentBB = int(operands[1].(literalValue).value.(uint32))
+				case 0: // false -> elseBB
+					currentBB = int(operands[2].(literalValue).value.(uint32))
+				default:
+					panic("bool should be 0 or 1")
+				}
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"br", operands, "->", currentBB)
+				}
+				bb = fn.blocks[currentBB]
+				instIndex = -1 // start at 0 the next cycle
+			default:
+				panic("unknown operands length")
+			}
+			break // continue with next block
+		case llvm.PHI:
+			var result value
+			for i := 0; i < len(inst.operands); i += 2 {
+				if int(inst.operands[i].(literalValue).value.(uint32)) == lastBB {
+					incoming := inst.operands[i+1]
+					if local, ok := incoming.(localValue); ok {
+						result = locals[fn.locals[local.value]]
+					} else {
+						result = incoming
+					}
+					break
+				}
+			}
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"phi", inst.operands, "->", result)
+			}
+			if result == nil {
+				panic("could not find PHI input")
+			}
+			locals[inst.localIndex] = result
+		case llvm.Select:
+			// Select is much like a ternary operator: it picks a result from
+			// the second and third operand based on the boolean first operand.
+			var result value
+			switch operands[0].Uint() {
+			case 1:
+				result = operands[1]
+			case 0:
+				result = operands[2]
+			default:
+				panic("boolean must be 0 or 1")
+			}
+			locals[inst.localIndex] = result
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"select", operands, "->", result)
+			}
+		case llvm.Call:
+			// A call instruction can either be a regular call or a runtime intrinsic.
+			fnPtr, err := operands[0].asPointer(r)
+			if err != nil {
+				return nil, mem, r.errorAt(inst, err)
+			}
+			callFn := r.getFunction(fnPtr.llvmValue(&mem))
+			switch {
+			case callFn.name == "runtime.trackPointer":
+				// Allocas and such are created as globals, so don't need a
+				// runtime.trackPointer.
+				// Unless the object is allocated at runtime for example, in
+				// which case this call won't even get to this point but will
+				// already be emitted in initAll.
+				continue
+			case callFn.name == "(reflect.Type).Elem" || strings.HasPrefix(callFn.name, "runtime.print") || callFn.name == "runtime._panic" || callFn.name == "runtime.hashmapGet":
+				// These functions should be run at runtime. Specifically:
+				//   * (reflect.Type).Elem is a special function. It should
+				//     eventually be interpreted, but fall back to a runtime call
+				//     for now.
+				//   * Print and panic functions are best emitted directly without
+				//     interpreting them, otherwise we get a ton of putchar (etc.)
+				//     calls.
+				//   * runtime.hashmapGet tries to access the map value directly.
+				//     This is not possible as the map value is treated as a special
+				//     kind of object in this package.
+				err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+				if err != nil {
+					return nil, mem, err
+				}
+			case callFn.name == "runtime.nanotime" && r.pkgName == "time":
+				// The time package contains a call to runtime.nanotime.
+				// This appears to be to work around a limitation in Windows
+				// Server 2008:
+				//   > Monotonic times are reported as offsets from startNano.
+				//   > We initialize startNano to runtimeNano() - 1 so that on systems where
+				//   > monotonic time resolution is fairly low (e.g. Windows 2008
+				//   > which appears to have a default resolution of 15ms),
+				//   > we avoid ever reporting a monotonic time of 0.
+				//   > (Callers may want to use 0 as "time not set".)
+				// Simply let runtime.nanotime return 0 in this case, which
+				// should be fine and avoids a call to runtime.nanotime. It
+				// means that monotonic time in the time package is counted from
+				// time.Time{}.Sub(1), which should be fine.
+				locals[inst.localIndex] = literalValue{uint64(0)}
+			case callFn.name == "runtime.alloc":
+				// Allocate heap memory. At compile time, this is instead done
+				// by creating a global variable.
+
+				// Get the requested memory size to be allocated.
+				size := operands[1].Uint()
+
+				// Create the object.
+				alloc := object{
+					globalName: r.pkgName + "$alloc",
+					buffer:     newRawValue(uint32(size)),
+					size:       uint32(size),
+				}
+				index := len(r.objects)
+				r.objects = append(r.objects, alloc)
+
+				// And create a pointer to this object, for working with it (so
+				// that stores to it copy it, etc).
+				ptr := newPointerValue(r, index, 0)
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"runtime.alloc:", size, "->", ptr)
+				}
+				locals[inst.localIndex] = ptr
+			case callFn.name == "runtime.sliceCopy":
+				// sliceCopy implements the built-in copy function for slices.
+				// It is implemented here so that it can be used even if the
+				// runtime implementation is not available. Doing it this way
+				// may also be faster.
+				// Code:
+				// func sliceCopy(dst, src unsafe.Pointer, dstLen, srcLen uintptr, elemSize uintptr) int {
+				//     n := srcLen
+				//     if n > dstLen {
+				//         n = dstLen
+				//     }
+				//     memmove(dst, src, n*elemSize)
+				//     return int(n)
+				// }
+				dstLen := operands[3].Uint()
+				srcLen := operands[4].Uint()
+				elemSize := operands[5].Uint()
+				n := srcLen
+				if n > dstLen {
+					n = dstLen
+				}
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"copy:", operands[1], operands[2], n)
+				}
+				if n != 0 {
+					// Only try to copy bytes when there are any bytes to copy.
+					// This is not just an optimization. If one of the slices
+					// (or both) are nil, the asPointer method call will fail
+					// even though copying a nil slice is allowed.
+					dst, err := operands[1].asPointer(r)
+					if err != nil {
+						return nil, mem, r.errorAt(inst, err)
+					}
+					src, err := operands[2].asPointer(r)
+					if err != nil {
+						return nil, mem, r.errorAt(inst, err)
+					}
+					nBytes := uint32(n * elemSize)
+					dstObj := mem.getWritable(dst.index())
+					dstBuf := dstObj.buffer.asRawValue(r)
+					srcBuf := mem.get(src.index()).buffer.asRawValue(r)
+					copy(dstBuf.buf[dst.offset():dst.offset()+nBytes], srcBuf.buf[src.offset():])
+					dstObj.buffer = dstBuf
+					mem.put(dst.index(), dstObj)
+				}
+				switch inst.llvmInst.Type().IntTypeWidth() {
+				case 16:
+					locals[inst.localIndex] = literalValue{uint16(n)}
+				case 32:
+					locals[inst.localIndex] = literalValue{uint32(n)}
+				case 64:
+					locals[inst.localIndex] = literalValue{uint64(n)}
+				default:
+					panic("unknown integer type width")
+				}
+			case strings.HasPrefix(callFn.name, "llvm.memcpy.p0i8.p0i8.") || strings.HasPrefix(callFn.name, "llvm.memmove.p0i8.p0i8."):
+				// Copy a block of memory from one pointer to another.
+				dst, err := operands[1].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				src, err := operands[2].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				nBytes := uint32(operands[3].Uint())
+				dstObj := mem.getWritable(dst.index())
+				dstBuf := dstObj.buffer.asRawValue(r)
+				srcBuf := mem.get(src.index()).buffer.asRawValue(r)
+				copy(dstBuf.buf[dst.offset():dst.offset()+nBytes], srcBuf.buf[src.offset():])
+				dstObj.buffer = dstBuf
+				mem.put(dst.index(), dstObj)
+			case callFn.name == "runtime.typeAssert":
+				// This function must be implemented manually as it is normally
+				// implemented by the interface lowering pass.
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"typeassert:", operands[1:])
+				}
+				typeInInterfacePtr, err := operands[1].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				actualType, err := mem.load(typeInInterfacePtr, r.pointerSize).asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				assertedType, err := operands[2].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				result := assertedType.asRawValue(r).equal(actualType.asRawValue(r))
+				if result {
+					locals[inst.localIndex] = literalValue{uint8(1)}
+				} else {
+					locals[inst.localIndex] = literalValue{uint8(0)}
+				}
+			case callFn.name == "runtime.interfaceImplements":
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"interface assert:", operands[1:])
+				}
+
+				// Load various values for the interface implements check below.
+				typeInInterfacePtr, err := operands[1].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				methodSetPtr, err := mem.load(typeInInterfacePtr.addOffset(r.pointerSize), r.pointerSize).asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				methodSet := mem.get(methodSetPtr.index()).llvmGlobal.Initializer()
+				interfaceMethodSetPtr, err := operands[2].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				interfaceMethodSet := mem.get(interfaceMethodSetPtr.index()).llvmGlobal.Initializer()
+
+				// Make a set of all the methods on the concrete type, for
+				// easier checking in the next step.
+				concreteTypeMethods := map[string]struct{}{}
+				for i := 0; i < methodSet.Type().ArrayLength(); i++ {
+					methodInfo := llvm.ConstExtractValue(methodSet, []uint32{uint32(i)})
+					name := llvm.ConstExtractValue(methodInfo, []uint32{0}).Name()
+					concreteTypeMethods[name] = struct{}{}
+				}
+
+				// Check whether all interface methods are also in the list
+				// of defined methods calculated above. This is the interface
+				// assert itself.
+				assertOk := uint8(1) // i1 true
+				for i := 0; i < interfaceMethodSet.Type().ArrayLength(); i++ {
+					name := llvm.ConstExtractValue(interfaceMethodSet, []uint32{uint32(i)}).Name()
+					if _, ok := concreteTypeMethods[name]; !ok {
+						// There is a method on the interface that is not
+						// implemented by the type. The assertion will fail.
+						assertOk = 0 // i1 false
+						break
+					}
+				}
+				// If assertOk is still 1, the assertion succeeded.
+				locals[inst.localIndex] = literalValue{assertOk}
+			case callFn.name == "runtime.hashmapMake":
+				// Create a new map.
+				hashmapPointerType := inst.llvmInst.Type()
+				keySize := uint32(operands[1].Uint())
+				valueSize := uint32(operands[2].Uint())
+				m := newMapValue(r, hashmapPointerType, keySize, valueSize)
+				alloc := object{
+					llvmType:   hashmapPointerType,
+					globalName: r.pkgName + "$map",
+					buffer:     m,
+					size:       m.len(r),
+				}
+				index := len(r.objects)
+				r.objects = append(r.objects, alloc)
+
+				// Create a pointer to this map. Maps are reference types, so
+				// are implemented as pointers.
+				ptr := newPointerValue(r, index, 0)
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"runtime.hashmapMake:", keySize, valueSize, "->", ptr)
+				}
+				locals[inst.localIndex] = ptr
+			case callFn.name == "runtime.hashmapBinarySet":
+				// Do a mapassign operation with a binary key (that is, without
+				// a string key).
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"runtime.hashmapBinarySet:", operands[1:])
+				}
+				mapPtr, err := operands[1].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				m := mem.getWritable(mapPtr.index()).buffer.(*mapValue)
+				keyPtr, err := operands[2].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				valuePtr, err := operands[3].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				err = m.putBinary(&mem, keyPtr, valuePtr)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+			case callFn.name == "runtime.hashmapStringSet":
+				// Do a mapassign operation with a string key.
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"runtime.hashmapBinarySet:", operands[1:])
+				}
+				mapPtr, err := operands[1].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				m := mem.getWritable(mapPtr.index()).buffer.(*mapValue)
+				stringPtr, err := operands[2].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				stringLen := operands[3].Uint()
+				valuePtr, err := operands[4].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				err = m.putString(&mem, stringPtr, stringLen, valuePtr)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+			default:
+				if len(callFn.blocks) == 0 {
+					// Call to a function declaration without a definition
+					// available.
+					err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+					if err != nil {
+						return nil, mem, err
+					}
+					continue
+				}
+				// Call a function with a definition available. Run it as usual,
+				// possibly trying to recover from it if it failed to execute.
+				if r.debug {
+					argStrings := make([]string, len(operands)-1)
+					for i := range argStrings {
+						argStrings[i] = operands[i+1].String()
+					}
+					fmt.Fprintln(os.Stderr, indent+"call:", callFn.name+"("+strings.Join(argStrings, ", ")+")")
+				}
+				retval, callMem, callErr := r.run(callFn, operands[1:], &mem, indent+"    ")
+				if callErr != nil {
+					if isRecoverableError(callErr.Err) {
+						// This error can be recovered by doing the call at
+						// runtime instead of at compile time. But we need to
+						// revert any changes made by the call first.
+						if r.debug {
+							fmt.Fprintln(os.Stderr, indent+"!! revert because of error:", callErr.Err)
+						}
+						callMem.revert()
+						err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+						if err != nil {
+							return nil, mem, err
+						}
+						continue
+					}
+					// Add to the traceback, so that error handling code can see
+					// how this function got called.
+					callErr.Traceback = append(callErr.Traceback, ErrorLine{
+						Pos:  getPosition(inst.llvmInst),
+						Inst: inst.llvmInst,
+					})
+					return nil, mem, callErr
+				}
+				locals[inst.localIndex] = retval
+				mem.extend(callMem)
+			}
+		case llvm.Load:
+			// Load instruction, loading some data from the topmost memory view.
+			ptr, err := operands[0].asPointer(r)
+			if err != nil {
+				return nil, mem, r.errorAt(inst, err)
+			}
+			size := operands[1].(literalValue).value.(uint64)
+			if mem.hasExternalStore(ptr) {
+				// If there could be an external store (for example, because a
+				// pointer to the object was passed to a function that could not
+				// be interpreted at compile time) then the load must be done at
+				// runtime.
+				err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+				if err != nil {
+					return nil, mem, err
+				}
+				continue
+			}
+			result := mem.load(ptr, uint32(size))
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"load:", ptr, "->", result)
+			}
+			locals[inst.localIndex] = result
+		case llvm.Store:
+			// Store instruction. Create a new object in the memory view and
+			// store to that, to make it possible to roll back this store.
+			ptr, err := operands[1].asPointer(r)
+			if err != nil {
+				return nil, mem, r.errorAt(inst, err)
+			}
+			if mem.hasExternalLoadOrStore(ptr) {
+				err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+				if err != nil {
+					return nil, mem, err
+				}
+				continue
+			}
+			val := operands[0]
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"store:", val, ptr)
+			}
+			mem.store(val, ptr)
+		case llvm.Alloca:
+			// Alloca normally allocates some stack memory. In the interpreter,
+			// it allocates a global instead.
+			// This can likely be optimized, as all it really needs is an alloca
+			// in the initAll function and creating a global is wasteful for
+			// this purpose.
+
+			// Create the new object.
+			size := operands[0].(literalValue).value.(uint64)
+			alloca := object{
+				llvmType:   inst.llvmInst.Type(),
+				globalName: r.pkgName + "$alloca",
+				buffer:     newRawValue(uint32(size)),
+				size:       uint32(size),
+			}
+			index := len(r.objects)
+			r.objects = append(r.objects, alloca)
+
+			// Create a pointer to this object (an alloca produces a pointer).
+			ptr := newPointerValue(r, index, 0)
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"alloca:", operands, "->", ptr)
+			}
+			locals[inst.localIndex] = ptr
+		case llvm.GetElementPtr:
+			// GetElementPtr does pointer arithmetic, changing the offset of the
+			// pointer into the underlying object.
+			var offset uint64
+			var gepOperands []uint64
+			for i := 2; i < len(operands); i += 2 {
+				index := operands[i].Uint()
+				elementSize := operands[i+1].Uint()
+				if int64(elementSize) < 0 {
+					// This is a struct field.
+					// The field number is encoded by flipping all the bits.
+					gepOperands = append(gepOperands, ^elementSize)
+					offset += index
+				} else {
+					// This is a normal GEP, probably an array index.
+					gepOperands = append(gepOperands, index)
+					offset += elementSize * index
+				}
+			}
+			ptr, err := operands[0].asPointer(r)
+			if err != nil {
+				return nil, mem, r.errorAt(inst, err)
+			}
+			ptr = ptr.addOffset(uint32(offset))
+			locals[inst.localIndex] = ptr
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"gep:", operands, "->", ptr)
+			}
+		case llvm.BitCast, llvm.IntToPtr, llvm.PtrToInt:
+			// Various bitcast-like instructions that all keep the same bits
+			// while changing the LLVM type.
+			// Because interp doesn't preserve the type, these operations are
+			// identity operations.
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+instructionNameMap[inst.opcode]+":", operands[0])
+			}
+			locals[inst.localIndex] = operands[0]
+		case llvm.ExtractValue:
+			agg := operands[0].asRawValue(r)
+			offset := operands[1].(literalValue).value.(uint64)
+			size := operands[2].(literalValue).value.(uint64)
+			elt := rawValue{
+				buf: agg.buf[offset : offset+size],
+			}
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"extractvalue:", operands, "->", elt)
+			}
+			locals[inst.localIndex] = elt
+		case llvm.InsertValue:
+			agg := operands[0].asRawValue(r)
+			elt := operands[1].asRawValue(r)
+			offset := int(operands[2].(literalValue).value.(uint64))
+			newagg := newRawValue(uint32(len(agg.buf)))
+			copy(newagg.buf, agg.buf)
+			copy(newagg.buf[offset:], elt.buf)
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"insertvalue:", operands, "->", newagg)
+			}
+			locals[inst.localIndex] = newagg
+		case llvm.ICmp:
+			predicate := llvm.IntPredicate(operands[2].(literalValue).value.(uint8))
+			var result bool
+			lhs := operands[0]
+			rhs := operands[1]
+			switch predicate {
+			case llvm.IntEQ, llvm.IntNE:
+				lhsPointer, lhsErr := lhs.asPointer(r)
+				rhsPointer, rhsErr := rhs.asPointer(r)
+				if (lhsErr == nil) != (rhsErr == nil) {
+					// Fast path: only one is a pointer, so they can't be equal.
+					result = false
+				} else if lhsErr == nil {
+					// Both must be nil, so both are pointers.
+					// Compare them directly.
+					result = lhsPointer.equal(rhsPointer)
+				} else {
+					// Fall back to generic comparison.
+					result = lhs.asRawValue(r).equal(rhs.asRawValue(r))
+				}
+				if predicate == llvm.IntNE {
+					result = !result
+				}
+			case llvm.IntUGT:
+				result = lhs.Uint() > rhs.Uint()
+			case llvm.IntUGE:
+				result = lhs.Uint() >= rhs.Uint()
+			case llvm.IntULT:
+				result = lhs.Uint() < rhs.Uint()
+			case llvm.IntULE:
+				result = lhs.Uint() <= rhs.Uint()
+			case llvm.IntSGT:
+				result = lhs.Int() > rhs.Int()
+			case llvm.IntSGE:
+				result = lhs.Int() >= rhs.Int()
+			case llvm.IntSLT:
+				result = lhs.Int() < rhs.Int()
+			case llvm.IntSLE:
+				result = lhs.Int() <= rhs.Int()
+			default:
+				return nil, mem, r.errorAt(inst, errors.New("interp: unsupported icmp"))
+			}
+			if result {
+				locals[inst.localIndex] = literalValue{uint8(1)}
+			} else {
+				locals[inst.localIndex] = literalValue{uint8(0)}
+			}
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"icmp:", operands[0], intPredicateString(predicate), operands[1], "->", result)
+			}
+		case llvm.FCmp:
+			predicate := llvm.FloatPredicate(operands[2].(literalValue).value.(uint8))
+			var result bool
+			var lhs, rhs float64
+			switch operands[0].len(r) {
+			case 8:
+				lhs = math.Float64frombits(operands[0].Uint())
+				rhs = math.Float64frombits(operands[1].Uint())
+			case 4:
+				lhs = float64(math.Float32frombits(uint32(operands[0].Uint())))
+				rhs = float64(math.Float32frombits(uint32(operands[1].Uint())))
+			default:
+				panic("unknown float type")
+			}
+			switch predicate {
+			case llvm.FloatOEQ:
+				result = lhs == rhs
+			case llvm.FloatUNE:
+				result = lhs != rhs
+			case llvm.FloatOGT:
+				result = lhs > rhs
+			case llvm.FloatOGE:
+				result = lhs >= rhs
+			case llvm.FloatOLT:
+				result = lhs < rhs
+			case llvm.FloatOLE:
+				result = lhs <= rhs
+			default:
+				return nil, mem, r.errorAt(inst, errors.New("interp: unsupported fcmp"))
+			}
+			if result {
+				locals[inst.localIndex] = literalValue{uint8(1)}
+			} else {
+				locals[inst.localIndex] = literalValue{uint8(0)}
+			}
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"fcmp:", operands[0], predicate, operands[1], "->", result)
+			}
+		case llvm.Add, llvm.Sub, llvm.Mul, llvm.UDiv, llvm.SDiv, llvm.URem, llvm.SRem, llvm.Shl, llvm.LShr, llvm.AShr, llvm.And, llvm.Or, llvm.Xor:
+			// Integer binary operations.
+			lhs := operands[0]
+			rhs := operands[1]
+			lhsPtr, err := lhs.asPointer(r)
+			if err == nil {
+				// The lhs is a pointer. This sometimes happens for particular
+				// pointer tricks.
+				switch inst.opcode {
+				case llvm.Add:
+					// This likely means this is part of a
+					// unsafe.Pointer(uintptr(ptr) + offset) pattern.
+					lhsPtr = lhsPtr.addOffset(uint32(rhs.Uint()))
+					locals[inst.localIndex] = lhsPtr
+					continue
+				case llvm.Xor:
+					if rhs.Uint() == 0 {
+						// Special workaround for strings.noescape, see
+						// src/strings/builder.go in the Go source tree. This is
+						// the identity operator, so we can return the input.
+						locals[inst.localIndex] = lhs
+						continue
+					}
+				default:
+					// Catch-all for weird operations that should just be done
+					// at runtime.
+					err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+					if err != nil {
+						return nil, mem, err
+					}
+					continue
+				}
+			}
+			var result uint64
+			switch inst.opcode {
+			case llvm.Add:
+				result = lhs.Uint() + rhs.Uint()
+			case llvm.Sub:
+				result = lhs.Uint() - rhs.Uint()
+			case llvm.Mul:
+				result = lhs.Uint() * rhs.Uint()
+			case llvm.UDiv:
+				result = lhs.Uint() / rhs.Uint()
+			case llvm.SDiv:
+				result = uint64(lhs.Int() / rhs.Int())
+			case llvm.URem:
+				result = lhs.Uint() % rhs.Uint()
+			case llvm.SRem:
+				result = uint64(lhs.Int() % rhs.Int())
+			case llvm.Shl:
+				result = lhs.Uint() << rhs.Uint()
+			case llvm.LShr:
+				result = lhs.Uint() >> rhs.Uint()
+			case llvm.AShr:
+				result = uint64(lhs.Int() >> rhs.Uint())
+			case llvm.And:
+				result = lhs.Uint() & rhs.Uint()
+			case llvm.Or:
+				result = lhs.Uint() | rhs.Uint()
+			case llvm.Xor:
+				result = lhs.Uint() ^ rhs.Uint()
+			default:
+				panic("unreachable")
+			}
+			switch lhs.len(r) {
+			case 8:
+				locals[inst.localIndex] = literalValue{result}
+			case 4:
+				locals[inst.localIndex] = literalValue{uint32(result)}
+			case 2:
+				locals[inst.localIndex] = literalValue{uint16(result)}
+			case 1:
+				locals[inst.localIndex] = literalValue{uint8(result)}
+			default:
+				panic("unknown integer size")
+			}
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+instructionNameMap[inst.opcode]+":", lhs, rhs, "->", result)
+			}
+		case llvm.SExt, llvm.ZExt, llvm.Trunc:
+			// Change the size of an integer to a larger or smaller bit width.
+			// We make use of the fact that the Uint() function already
+			// zero-extends the value and that Int() already sign-extends the
+			// value, so we only need to truncate it to the appropriate bit
+			// width. This means we can implement sext, zext and trunc in the
+			// same way, by first {zero,sign}extending all the way up to uint64
+			// and then truncating it as necessary.
+			var value uint64
+			if inst.opcode == llvm.SExt {
+				value = uint64(operands[0].Int())
+			} else {
+				value = operands[0].Uint()
+			}
+			bitwidth := operands[1].Uint()
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+instructionNameMap[inst.opcode]+":", value, bitwidth)
+			}
+			switch bitwidth {
+			case 64:
+				locals[inst.localIndex] = literalValue{value}
+			case 32:
+				locals[inst.localIndex] = literalValue{uint32(value)}
+			case 16:
+				locals[inst.localIndex] = literalValue{uint16(value)}
+			case 8:
+				locals[inst.localIndex] = literalValue{uint8(value)}
+			default:
+				panic("unknown integer size in sext/zext/trunc")
+			}
+		case llvm.SIToFP, llvm.UIToFP:
+			var value float64
+			switch inst.opcode {
+			case llvm.SIToFP:
+				value = float64(operands[0].Int())
+			case llvm.UIToFP:
+				value = float64(operands[0].Uint())
+			}
+			bitwidth := operands[1].Uint()
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+instructionNameMap[inst.opcode]+":", value, bitwidth)
+			}
+			switch bitwidth {
+			case 64:
+				locals[inst.localIndex] = literalValue{math.Float64bits(value)}
+			case 32:
+				locals[inst.localIndex] = literalValue{math.Float32bits(float32(value))}
+			default:
+				panic("unknown integer size in sitofp/uitofp")
+			}
+		default:
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+inst.String())
+			}
+			return nil, mem, r.errorAt(inst, errUnsupportedInst)
+		}
+	}
+	return nil, mem, r.errorAt(bb.instructions[len(bb.instructions)-1], errors.New("interp: reached end of basic block without terminator"))
+}
+
+func (r *runner) runAtRuntime(fn *function, inst instruction, locals []value, mem *memoryView, indent string) *Error {
+	numOperands := inst.llvmInst.OperandsCount()
+	operands := make([]llvm.Value, numOperands)
+	for i := 0; i < numOperands; i++ {
+		operand := inst.llvmInst.Operand(i)
+		if !operand.IsAInstruction().IsNil() || !operand.IsAArgument().IsNil() {
+			operand = locals[fn.locals[operand]].toLLVMValue(operand.Type(), mem)
+		}
+		operands[i] = operand
+	}
+	if r.debug {
+		fmt.Fprintln(os.Stderr, indent+inst.String())
+	}
+	var result llvm.Value
+	switch inst.opcode {
+	case llvm.Call:
+		llvmFn := operands[len(operands)-1]
+		args := operands[:len(operands)-1]
+		for _, arg := range args {
+			if arg.Type().TypeKind() == llvm.PointerTypeKind {
+				mem.markExternalStore(arg)
+			}
+		}
+		result = r.builder.CreateCall(llvmFn, args, inst.name)
+	case llvm.Load:
+		mem.markExternalLoad(operands[0])
+		result = r.builder.CreateLoad(operands[0], inst.name)
+		if inst.llvmInst.IsVolatile() {
+			result.SetVolatile(true)
+		}
+	case llvm.Store:
+		mem.markExternalStore(operands[1])
+		result = r.builder.CreateStore(operands[0], operands[1])
+		if inst.llvmInst.IsVolatile() {
+			result.SetVolatile(true)
+		}
+	case llvm.BitCast:
+		result = r.builder.CreateBitCast(operands[0], inst.llvmInst.Type(), inst.name)
+	case llvm.ExtractValue:
+		indices := inst.llvmInst.Indices()
+		if len(indices) != 1 {
+			panic("expected exactly one index")
+		}
+		result = r.builder.CreateExtractValue(operands[0], int(indices[0]), inst.name)
+	case llvm.InsertValue:
+		indices := inst.llvmInst.Indices()
+		if len(indices) != 1 {
+			panic("expected exactly one index")
+		}
+		result = r.builder.CreateInsertValue(operands[0], operands[1], int(indices[0]), inst.name)
+	case llvm.Add:
+		result = r.builder.CreateAdd(operands[0], operands[1], inst.name)
+	case llvm.Sub:
+		result = r.builder.CreateSub(operands[0], operands[1], inst.name)
+	case llvm.Mul:
+		result = r.builder.CreateMul(operands[0], operands[1], inst.name)
+	case llvm.UDiv:
+		result = r.builder.CreateUDiv(operands[0], operands[1], inst.name)
+	case llvm.SDiv:
+		result = r.builder.CreateSDiv(operands[0], operands[1], inst.name)
+	case llvm.URem:
+		result = r.builder.CreateURem(operands[0], operands[1], inst.name)
+	case llvm.SRem:
+		result = r.builder.CreateSRem(operands[0], operands[1], inst.name)
+	case llvm.ZExt:
+		result = r.builder.CreateZExt(operands[0], inst.llvmInst.Type(), inst.name)
+	default:
+		return r.errorAt(inst, errUnsupportedRuntimeInst)
+	}
+	locals[inst.localIndex] = localValue{result}
+	mem.instructions = append(mem.instructions, result)
+	return nil
+}
+
+func intPredicateString(predicate llvm.IntPredicate) string {
+	switch predicate {
+	case llvm.IntEQ:
+		return "eq"
+	case llvm.IntNE:
+		return "ne"
+	case llvm.IntUGT:
+		return "ugt"
+	case llvm.IntUGE:
+		return "uge"
+	case llvm.IntULT:
+		return "ult"
+	case llvm.IntULE:
+		return "ule"
+	case llvm.IntSGT:
+		return "sgt"
+	case llvm.IntSGE:
+		return "sge"
+	case llvm.IntSLT:
+		return "slt"
+	case llvm.IntSLE:
+		return "sle"
+	default:
+		return "cmp?"
+	}
+}
--- a/Show More
+++ b/Show More