nrf: Added support for writing with nrfjprog on all nrf boards

`tinygo flash -programmer command` to write using nrfjprog
builder, compiler: compile and cache packages in parallel
2026-07-26 14:48:40 +00:00 · 2021-03-22 19:00:21 +09:00 · 2021-03-21 11:51:35 +01:00 · 2021-03-21 11:51:35 +01:00 · 2021-03-21 11:25:10 +01:00 · 2021-03-21 10:00:43 +01:00
594 changed files with 34548 additions and 8484 deletions
@@ -28,6 +28,7 @@ commands:
                qemu-user \
                gcc-avr \
                avr-libc
+            sudo apt-get install --no-install-recommends libc6-dev-i386 lib32gcc-8-dev
  install-node:
    steps:
      - run:
@@ -37,28 +38,54 @@ 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-10-v0
+            - llvm-source-11-v1
      - run:
          name: "Fetch LLVM source"
          command: make llvm-source
      - save_cache:
-          key: llvm-source-10-v0
+          key: llvm-source-11-v1
          paths:
            - llvm-project
  build-wasi-libc:
    steps:
      - restore_cache:
          keys:
-            - wasi-libc-sysroot-v2
+            - wasi-libc-sysroot-v4
      - run:
          name: "Build wasi-libc"
          command: make wasi-libc
      - save_cache:
-          key: wasi-libc-sysroot-v2
+          key: wasi-libc-sysroot-v4
          paths:
            - lib/wasi-libc/sysroot
  test-linux:
@@ -71,6 +98,8 @@ commands:
      - apt-dependencies:
          llvm: "<<parameters.llvm>>"
      - install-node
+      - install-chrome
+      - install-wasmtime
      - restore_cache:
          keys:
            - go-cache-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_PREVIOUS_BUILD_NUM }}
@@ -79,20 +108,21 @@ commands:
      - run: go install -tags=llvm<<parameters.llvm>> .
      - restore_cache:
          keys:
-            - wasi-libc-sysroot-systemclang-v1
+            - wasi-libc-sysroot-systemclang-v3
      - run: make wasi-libc
      - save_cache:
-          key: wasi-libc-sysroot-systemclang-v1
+          key: wasi-libc-sysroot-systemclang-v3
          paths:
            - lib/wasi-libc/sysroot
-      - run: go test -v -tags=llvm<<parameters.llvm>> ./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:
@@ -102,9 +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 \
@@ -112,7 +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 }}
@@ -120,17 +154,14 @@ commands:
      - llvm-source-linux
      - restore_cache:
          keys:
-            - llvm-build-10-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
@@ -138,7 +169,7 @@ commands:
              make ASSERT=1 llvm-build
            fi
      - save_cache:
-          key: llvm-build-10-linux-v0-assert
+          key: llvm-build-11-linux-v2-assert
          paths:
            llvm-build
      - run: make ASSERT=1
@@ -150,7 +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
@@ -161,9 +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 \
@@ -171,7 +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 }}
@@ -179,17 +213,14 @@ commands:
      - llvm-source-linux
      - restore_cache:
          keys:
-            - llvm-build-10-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
@@ -197,25 +228,32 @@ commands:
              make llvm-build
            fi
      - save_cache:
-          key: llvm-build-10-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"
@@ -232,27 +270,29 @@ commands:
      - run:
          name: "Install dependencies"
          command: |
-            curl https://dl.google.com/go/go1.14.darwin-amd64.tar.gz -o go1.14.darwin-amd64.tar.gz
-            sudo tar -C /usr/local -xzf go1.14.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-10-macos-v0
+            - llvm-source-11-macos-v1
      - run:
          name: "Fetch LLVM source"
          command: make llvm-source
      - save_cache:
-          key: llvm-source-10-macos-v0
+          key: llvm-source-11-macos-v1
          paths:
            - llvm-project
      - restore_cache:
          keys:
-            - llvm-build-10-macos-v0
+            - llvm-build-11-macos-v2
      - run:
          name: "Build LLVM"
          command: |
@@ -264,17 +304,17 @@ commands:
              make llvm-build
            fi
      - save_cache:
-          key: llvm-build-10-macos-v0
+          key: llvm-build-11-macos-v2
          paths:
            llvm-build
      - restore_cache:
          keys:
-            - wasi-libc-sysroot-macos-v1
+            - wasi-libc-sysroot-macos-v3
      - run:
          name: "Build wasi-libc"
          command: make wasi-libc
      - save_cache:
-          key: wasi-libc-sysroot-macos-v1
+          key: wasi-libc-sysroot-macos-v3
          paths:
            - lib/wasi-libc/sysroot
      - run:
@@ -294,33 +334,14 @@ 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:
-    docker:
-      - image: circleci/golang:1.11-buster
-    steps:
-      - test-linux:
-          llvm: "9"
-  test-llvm10-go112:
-    docker:
-      - image: circleci/golang:1.12-buster
-    steps:
-      - test-linux:
-          llvm: "10"
  test-llvm10-go113:
    docker:
      - image: circleci/golang:1.13-buster
@@ -333,6 +354,18 @@ jobs:
    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:1.14-stretch
@@ -351,14 +384,13 @@ jobs:



-
 workflows:
  test-all:
    jobs:
-      - test-llvm9-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 +3,26 @@ 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
+
+# Ignore files generated by smoketest
+test.gba
+test.hex
+test.nro
+test.wasm
+wasm.wasm
@@ -9,7 +9,7 @@
 	url = https://github.com/avr-rust/avr-mcu.git
 [submodule "lib/cmsis-svd"]
 	path = lib/cmsis-svd
-	url = https://github.com/posborne/cmsis-svd
+	url = https://github.com/tinygo-org/cmsis-svd
 [submodule "lib/compiler-rt"]
 	path = lib/compiler-rt
 	url = https://github.com/llvm-mirror/compiler-rt.git
@@ -20,3 +20,6 @@
 [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
@@ -10,12 +10,20 @@ 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
 build tools to be built. Go is of course necessary to build TinyGo itself.

-  * Go (1.11+)
+  * Go (1.13+)
  * Standard build tools (gcc/clang)
  * git
  * CMake
@@ -35,11 +43,6 @@ You can also store LLVM outside of the TinyGo root directory by setting the
 `LLVM_BUILDDIR`, `CLANG_SRC` and `LLD_SRC` make variables, but that is not
 covered by this guide.

-TinyGo uses Go modules, so if you clone TinyGo inside your GOPATH (and are using
-Go below 1.13), make sure that Go modules are enabled:
-
-    export GO111MODULE=on
-
 ## Build LLVM, Clang, LLD

 Before starting the build, you may want to set the following environment
@@ -84,6 +87,10 @@ 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 +1,316 @@
+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**
@@ -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 1.14, LLVM 10 and the TinyGo compiler itself.
-FROM golang:1.14 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-10 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-10-dev libclang-10-dev git
+    apt-get install -y llvm-11-dev libclang-11-dev lld-11 git

 COPY . /tinygo

@@ -27,10 +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-10 main" >> /etc/apt/sources.list && \
+RUN cd /tinygo/ && \
    apt-get update && \
-    apt-get install -y libllvm10 lld-10
+    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
@@ -61,7 +61,7 @@ COPY --from=tinygo-base /tinygo/lib /tinygo/lib

 RUN cd /tinygo/ && \
    apt-get update && \
-    apt-get install -y apt-utils make clang-10 && \
+    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.
@@ -73,7 +73,7 @@ COPY --from=tinygo-base /tinygo/lib /tinygo/lib

 RUN cd /tinygo/ && \
    apt-get update && \
-    apt-get install -y apt-utils make clang-10 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) 2018-2020 TinyGo Authors. All rights reserved.
+Copyright (c) 2018-2021 TinyGo Authors. All rights reserved.

 TinyGo includes portions of the Go standard library.
-Copyright (c) 2009-2020 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.
@@ -9,25 +9,10 @@ CLANG_SRC ?= $(LLVM_PROJECTDIR)/clang
 LLD_SRC ?= $(LLVM_PROJECTDIR)/lld

 # Try to autodetect LLVM build tools.
-ifneq (, $(shell command -v llvm-build/bin/clang 2> /dev/null))
-    CLANG ?= $(abspath llvm-build/bin/clang)
-else
-    CLANG ?= clang-10
-endif
-ifneq (, $(shell command -v llvm-build/bin/llvm-ar 2> /dev/null))
-    LLVM_AR ?= $(abspath llvm-build/bin/llvm-ar)
-else ifneq (, $(shell command -v llvm-ar-10 2> /dev/null))
-    LLVM_AR ?= llvm-ar-10
-else
-    LLVM_AR ?= llvm-ar
-endif
-ifneq (, $(shell command -v llvm-build/bin/llvm-nm 2> /dev/null))
-    LLVM_NM ?= $(abspath llvm-build/bin/llvm-nm)
-else ifneq (, $(shell command -v llvm-nm-10 2> /dev/null))
-    LLVM_NM ?= llvm-nm-10
-else
-    LLVM_NM ?= llvm-nm
-endif
+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
@@ -37,7 +22,7 @@ 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 command -v ccache 2> /dev/null))
@@ -51,7 +36,7 @@ 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 frontendopenmp instrumentation interpreter ipo irreader linker lto mc mcjit objcarcopts option profiledata scalaropts support target

@@ -63,64 +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 $(LLD_SRC))/include
+    CGO_CPPFLAGS+=$(shell $(LLVM_BUILDDIR)/bin/llvm-config --cppflags) -I$(abspath $(LLVM_BUILDDIR))/tools/clang/include -I$(abspath $(CLANG_SRC))/include -I$(abspath $(LLD_SRC))/include
    CGO_CXXFLAGS=-std=c++14
-    CGO_LDFLAGS+=$(LIBCLANG_PATH) -std=c++14 -L$(abspath $(LLVM_BUILDDIR)/lib) $(CLANG_LIBS) $(LLD_LIBS) $(shell $(LLVM_BUILDDIR)/bin/llvm-config --ldflags --libs --system-libs $(LLVM_COMPONENTS)) -lstdc++ $(CGO_LDFLAGS_EXTRA)
+    CGO_LDFLAGS+=$(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/
@@ -129,10 +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
@@ -141,48 +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_PROJECTDIR)/README.md:
-	git clone -b release/10.x https://github.com/llvm/llvm-project $(LLVM_PROJECTDIR)
+	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_PROJECTDIR)/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
 	$(TINYGO) build -size short -o test.hex -target=pca10040            examples/adc
@@ -203,7 +223,9 @@ smoketest:
 	@$(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
@@ -214,8 +236,6 @@ smoketest:
 	# test simulated boards on play.tinygo.org
 	$(TINYGO) build             -o test.wasm -tags=arduino              examples/blinky1
 	@$(MD5SUM) test.wasm
-	$(TINYGO) build             -o test.wasm -tags=hifive1-qemu         examples/serial
-	@$(MD5SUM) test.wasm
 	$(TINYGO) build             -o test.wasm -tags=hifive1b             examples/blinky1
 	@$(MD5SUM) test.wasm
 	$(TINYGO) build             -o test.wasm -tags=reelboard            examples/blinky1
@@ -231,6 +251,8 @@ smoketest:
 	@$(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
 	$(TINYGO) build -size short -o test.hex -target=pca10031            examples/blinky1
@@ -257,11 +279,15 @@ smoketest:
 	@$(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
+	$(TINYGO) build -size short -o test.hex -target=clue-alpha          examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.gba -target=gameboy-advance     examples/gba-display
 	@$(MD5SUM) test.gba
@@ -285,17 +311,56 @@ smoketest:
 	@$(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
 	$(TINYGO) build -size short -o test.hex -target=arduino-nano        examples/blinky1
@@ -304,13 +369,35 @@ ifneq ($(AVR), 0)
 	@$(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
+	$(TINYGO) build -size short -o test.hex -target=pca10040 -opt=0     ./testdata/stdlib.go
+	@$(MD5SUM) test.hex

-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
@@ -345,4 +432,13 @@ release: tinygo gen-device wasi-libc
 	./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
@@ -43,27 +43,42 @@ See the [getting started instructions](https://tinygo.org/getting-started/) for

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

-The following 32 microcontroller boards are currently supported:
+The following 55 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 Alpha](https://www.adafruit.com/product/4500)
+* [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 QT Py](https://www.adafruit.com/product/4600)
 * [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)
@@ -72,9 +87,17 @@ The following 32 microcontroller boards are currently supported:
 * [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)
+* [PJRC Teensy 4.0](https://www.pjrc.com/store/teensy40.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/)

@@ -136,4 +159,6 @@ 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/10.0.0/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:
@@ -12,12 +12,13 @@ jobs:
  steps:
    - task: GoTool@0
      inputs:
-        version: '1.14.1'
+        version: '1.16'
    - checkout: self
-    - task: CacheBeta@0
+      fetchDepth: 1
+    - task: Cache@2
      displayName: Cache LLVM source
      inputs:
-        key: llvm-source-10-windows-v0
+        key: llvm-source-11-windows-v1
        path: llvm-project
    - task: Bash@3
      displayName: Download LLVM source
@@ -27,7 +28,7 @@ jobs:
    - task: CacheBeta@0
      displayName: Cache LLVM build
      inputs:
-        key: llvm-build-10-windows-v0
+        key: llvm-build-11-windows-v4
        path: llvm-build
    - task: Bash@3
      displayName: Build LLVM
@@ -36,18 +37,22 @@ jobs:
        script: |
          if [ ! -f llvm-build/lib/liblldELF.a ]
          then
+            # install dependencies
            choco install ninja
+            # 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-v2
+        key: wasi-libc-sysroot-v4
        path: lib/wasi-libc/sysroot
    - task: Bash@3
      displayName: Build wasi-libc
@@ -69,7 +74,7 @@ jobs:
        script: |
          export PATH="$PATH:./llvm-build/bin:/c/Program Files/qemu"
          unset GOROOT
-          make release -j4
+          make build/release -j4
    - publish: $(System.DefaultWorkingDirectory)/build/release/tinygo
      displayName: Publish zip as artifact
      artifact: tinygo
@@ -80,4 +85,4 @@ jobs:
        script: |
          export PATH="$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 +4,19 @@
 package builder

 import (
+	"crypto/sha512"
+	"debug/elf"
+	"encoding/binary"
+	"encoding/hex"
+	"encoding/json"
 	"errors"
 	"fmt"
+	"go/types"
 	"io/ioutil"
 	"os"
 	"path/filepath"
+	"runtime"
+	"sort"
 	"strconv"
 	"strings"

@@ -16,36 +24,554 @@ 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
+}
+
+// packageAction is the struct that is serialized to JSON and hashed, to work as
+// a cache key of compiled packages. It should contain all the information that
+// goes into a compiled package to avoid using stale data.
+//
+// Right now it's still important to include a hash of every import, because a
+// dependency might have a public constant that this package uses and thus this
+// package will need to be recompiled if that constant changes. In the future,
+// the type data should be serialized to disk which can then be used as cache
+// key, avoiding the need for recompiling all dependencies when only the
+// implementation of an imported package changes.
+type packageAction struct {
+	ImportPath      string
+	CompilerVersion int // compiler.Version
+	LLVMVersion     string
+	Config          *compiler.Config
+	CFlags          []string
+	FileHashes      map[string]string // hash of every file that's part of the package
+	Imports         map[string]string // map from imported package to action ID hash
+}
+
 // 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 {
-	// Compile Go code to IR.
-	machine, err := compiler.NewTargetMachine(config)
+func Build(pkgName, outpath string, config *compileopts.Config, action func(BuildResult) error) error {
+	// Create a temporary directory for intermediary files.
+	dir, err := ioutil.TempDir("", "tinygo")
 	if err != nil {
 		return err
 	}
-	mod, extraFiles, errs := compiler.Compile(pkgName, machine, config)
-	if errs != nil {
-		return newMultiError(errs)
+	defer os.RemoveAll(dir)
+
+	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(),
 	}

-	if config.Options.PrintIR {
-		fmt.Println("; Generated LLVM IR:")
-		fmt.Println(mod.String())
-	}
-	if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
-		return errors.New("verification error after IR construction")
+	// 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
 	}

-	err = interp.Run(mod, config.DumpSSA())
+	// 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
+
+	// Create the *ssa.Program. This does not yet build the entire SSA of the
+	// program so it's pretty fast and doesn't need to be parallelized.
+	program := lprogram.LoadSSA()
+
+	// Add jobs to compile each package.
+	// Packages that have a cache hit will not be compiled again.
+	var packageJobs []*compileJob
+	packageBitcodePaths := make(map[string]string)
+	packageActionIDs := make(map[string]string)
+	for _, pkg := range lprogram.Sorted() {
+		pkg := pkg // necessary to avoid a race condition
+
+		// Create a cache key: a hash from the action ID below that contains all
+		// the parameters for the build.
+		actionID := packageAction{
+			ImportPath:      pkg.ImportPath,
+			CompilerVersion: compiler.Version,
+			LLVMVersion:     llvm.Version,
+			Config:          compilerConfig,
+			CFlags:          pkg.CFlags,
+			FileHashes:      make(map[string]string, len(pkg.FileHashes)),
+			Imports:         make(map[string]string, len(pkg.Pkg.Imports())),
+		}
+		for filePath, hash := range pkg.FileHashes {
+			actionID.FileHashes[filePath] = hex.EncodeToString(hash)
+		}
+		for _, imported := range pkg.Pkg.Imports() {
+			hash, ok := packageActionIDs[imported.Path()]
+			if !ok {
+				return fmt.Errorf("package %s imports %s but couldn't find dependency", pkg.ImportPath, imported.Path())
+			}
+			actionID.Imports[imported.Path()] = hash
+		}
+		buf, err := json.Marshal(actionID)
+		if err != nil {
+			panic(err) // shouldn't happen
+		}
+		hash := sha512.Sum512_224(buf)
+		packageActionIDs[pkg.ImportPath] = hex.EncodeToString(hash[:])
+
+		// Determine the path of the bitcode file (which is a serialized version
+		// of a LLVM module).
+		cacheDir := goenv.Get("GOCACHE")
+		if cacheDir == "off" {
+			// Use temporary build directory instead, effectively disabling the
+			// build cache.
+			cacheDir = dir
+		}
+		bitcodePath := filepath.Join(cacheDir, "pkg-"+hex.EncodeToString(hash[:])+".bc")
+		packageBitcodePaths[pkg.ImportPath] = bitcodePath
+
+		// Check whether this package has been compiled before, and if so don't
+		// compile it again.
+		if _, err := os.Stat(bitcodePath); err == nil {
+			// Already cached, don't recreate this package.
+			continue
+		}
+
+		// The package has not yet been compiled, so create a job to do so.
+		job := &compileJob{
+			description: "compile package " + pkg.ImportPath,
+			run: func() error {
+				// Compile AST to IR. The compiler.CompilePackage function will
+				// build the SSA as needed.
+				mod, errs := compiler.CompilePackage(pkg.ImportPath, pkg, program.Package(pkg.Pkg), machine, compilerConfig, config.DumpSSA())
+				if errs != nil {
+					return newMultiError(errs)
+				}
+				if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
+					return errors.New("verification error after compiling package " + pkg.ImportPath)
+				}
+
+				// Serialize the LLVM module as a bitcode file.
+				// Write to a temporary path that is renamed to the destination
+				// file to avoid race conditions with other TinyGo invocatiosn
+				// that might also be compiling this package at the same time.
+				f, err := ioutil.TempFile(filepath.Dir(bitcodePath), filepath.Base(bitcodePath))
+				if err != nil {
+					return err
+				}
+				if runtime.GOOS == "windows" {
+					// Work around a problem on Windows.
+					// For some reason, WriteBitcodeToFile causes TinyGo to
+					// exit with the following message:
+					//   LLVM ERROR: IO failure on output stream: Bad file descriptor
+					buf := llvm.WriteBitcodeToMemoryBuffer(mod)
+					defer buf.Dispose()
+					_, err = f.Write(buf.Bytes())
+				} else {
+					// Otherwise, write bitcode directly to the file (probably
+					// faster).
+					err = llvm.WriteBitcodeToFile(mod, f)
+				}
+				if err != nil {
+					// WriteBitcodeToFile doesn't produce a useful error on its
+					// own, so create a somewhat useful error message here.
+					return fmt.Errorf("failed to write bitcode for package %s to file %s", pkg.ImportPath, bitcodePath)
+				}
+				err = f.Close()
+				if err != nil {
+					return err
+				}
+				return os.Rename(f.Name(), bitcodePath)
+			},
+		}
+		jobs = append(jobs, job)
+		packageJobs = append(packageJobs, job)
+	}
+
+	// Add job that links and optimizes all packages together.
+	var mod llvm.Module
+	var stackSizeLoads []string
+	programJob := &compileJob{
+		description:  "link+optimize packages (LTO)",
+		dependencies: packageJobs,
+		run: func() error {
+			// Load and link all the bitcode files. This does not yet optimize
+			// anything, it only links the bitcode files together.
+			ctx := llvm.NewContext()
+			mod = ctx.NewModule("")
+			for _, pkg := range lprogram.Sorted() {
+				pkgMod, err := ctx.ParseBitcodeFile(packageBitcodePaths[pkg.ImportPath])
+				if err != nil {
+					return fmt.Errorf("failed to load bitcode file: %w", err)
+				}
+				err = llvm.LinkModules(mod, pkgMod)
+				if err != nil {
+					return fmt.Errorf("failed to link module: %w", err)
+				}
+			}
+
+			// Create runtime.initAll function that calls the runtime
+			// initializer of each package.
+			llvmInitFn := mod.NamedFunction("runtime.initAll")
+			llvmInitFn.SetLinkage(llvm.InternalLinkage)
+			llvmInitFn.SetUnnamedAddr(true)
+			llvmInitFn.Param(0).SetName("context")
+			llvmInitFn.Param(1).SetName("parentHandle")
+			block := mod.Context().AddBasicBlock(llvmInitFn, "entry")
+			irbuilder := mod.Context().NewBuilder()
+			defer irbuilder.Dispose()
+			irbuilder.SetInsertPointAtEnd(block)
+			i8ptrType := llvm.PointerType(mod.Context().Int8Type(), 0)
+			for _, pkg := range lprogram.Sorted() {
+				pkgInit := mod.NamedFunction(pkg.Pkg.Path() + ".init")
+				if pkgInit.IsNil() {
+					panic("init not found for " + pkg.Pkg.Path())
+				}
+				irbuilder.CreateCall(pkgInit, []llvm.Value{llvm.Undef(i8ptrType), llvm.Undef(i8ptrType)}, "")
+			}
+			irbuilder.CreateRetVoid()
+
+			// After linking, functions should (as far as possible) be set to
+			// private linkage or internal linkage. The compiler package marks
+			// non-exported functions by setting the visibility to hidden or
+			// (for thunks) to linkonce_odr linkage. Change the linkage here to
+			// internal to benefit much more from interprocedural optimizations.
+			for fn := mod.FirstFunction(); !fn.IsNil(); fn = llvm.NextFunction(fn) {
+				if fn.Visibility() == llvm.HiddenVisibility {
+					fn.SetVisibility(llvm.DefaultVisibility)
+					fn.SetLinkage(llvm.InternalLinkage)
+				} else if fn.Linkage() == llvm.LinkOnceODRLinkage {
+					fn.SetLinkage(llvm.InternalLinkage)
+				}
+			}
+
+			// Do the same for globals.
+			for global := mod.FirstGlobal(); !global.IsNil(); global = llvm.NextGlobal(global) {
+				if global.Visibility() == llvm.HiddenVisibility {
+					global.SetVisibility(llvm.DefaultVisibility)
+					global.SetLinkage(llvm.InternalLinkage)
+				} else if global.Linkage() == llvm.LinkOnceODRLinkage {
+					global.SetLinkage(llvm.InternalLinkage)
+				}
+			}
+
+			if config.Options.PrintIR {
+				fmt.Println("; Generated LLVM IR:")
+				fmt.Println(mod.String())
+			}
+
+			// Run all optimization passes, which are much more effective now
+			// that the optimizer can see the whole program at once.
+			err := optimizeProgram(mod, config)
+			if err != nil {
+				return err
+			}
+
+			// 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 nil
+		},
+	}
+	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.
+
+	// 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,
+	})
+}
+
+// optimizeProgram runs a series of optimizations and transformations that are
+// needed to convert a program to its final form. Some transformations are not
+// optional and must be run as the compiler expects them to run.
+func optimizeProgram(mod llvm.Module, config *compileopts.Config) error {
+	err := interp.Run(mod, config.DumpSSA())
 	if err != nil {
 		return err
 	}
@@ -62,7 +588,7 @@ 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") {
+	if config.WasmAbi() == "js" {
 		err := transform.ExternalInt64AsPtr(mod)
 		if err != nil {
 			return err
@@ -71,7 +597,7 @@ func Build(pkgName, outpath string, config *compileopts.Config, action func(stri

 	// Optimization levels here are roughly the same as Clang, but probably not
 	// exactly.
-	errs = nil
+	var errs []error
 	switch config.Options.Opt {
 	case "none", "0":
 		errs = transform.Optimize(mod, config, 0, 0, 0) // -O0
@@ -84,7 +610,7 @@ func Build(pkgName, outpath string, config *compileopts.Config, action func(stri
 	case "z":
 		errs = transform.Optimize(mod, config, 2, 2, 5) // -Oz, default
 	default:
-		errs = []error{errors.New("unknown optimization level: -opt=" + config.Options.Opt)}
+		return errors.New("unknown optimization level: -opt=" + config.Options.Opt)
 	}
 	if len(errs) > 0 {
 		return newMultiError(errs)
@@ -93,140 +619,249 @@ func Build(pkgName, outpath string, config *compileopts.Config, action func(stri
 		return 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") {
-		transform.NonConstGlobals(mod)
-		if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); 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)
+	}
+
+	return nil
+}
+
+// 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
+}
+
+// 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
+				}
+				if callee := inst.CalledValue(); callee.IsAFunction().IsNil() && callee.IsAInlineAsm().IsNil() {
+					callsIndirectFunction = append(callsIndirectFunction, fn.Name())
+				}
+			}
+		}
+
+		// 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,
 		}
 	}

-	// Generate output.
-	outext := filepath.Ext(outpath)
-	switch outext {
-	case ".o":
-		llvmBuf, err := machine.EmitToMemoryBuffer(mod, llvm.ObjectFile)
-		if err != nil {
-			return err
+	// 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))
 		}
-		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:
-		// Act as a compiler driver.
-
-		// Create a temporary directory for intermediary files.
-		dir, err := ioutil.TempDir("", "tinygo")
-		if err != nil {
-			return err
+		humanName := gowrapperNames[name]
+		if humanName == "" {
+			humanName = name // fallback
 		}
-		defer os.RemoveAll(dir)
-
-		// Write the object file.
-		objfile := filepath.Join(dir, "main.o")
-		llvmBuf, err := machine.EmitToMemoryBuffer(mod, llvm.ObjectFile)
-		if err != nil {
-			return err
+		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
 		}
-		err = ioutil.WriteFile(objfile, llvmBuf.Bytes(), 0666)
-		if err != nil {
-			return err
+		sizes[name] = functionStackSize{
+			stackSize:        stackSize,
+			stackSizeType:    stackSizeType,
+			missingStackSize: missingStackSize,
+			humanName:        humanName,
 		}
+	}

-		// Prepare link command.
-		executable := filepath.Join(dir, "main")
-		tmppath := executable // final file
-		ldflags := append(config.LDFlags(), "-o", executable, objfile)
+	if resetFunction != "" {
+		return append([]string{resetFunction}, gowrappers...), sizes, nil
+	}
+	return gowrappers, sizes, nil
+}

-		// Load builtins library from the cache, possibly compiling it on the
-		// fly.
-		if config.Target.RTLib == "compiler-rt" {
-			librt, err := CompilerRT.Load(config.Triple())
-			if err != nil {
-				return err
+// 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
 			}
-			ldflags = append(ldflags, librt)
-		}

-		// Add libc.
-		if config.Target.Libc == "picolibc" {
-			libc, err := Picolibc.Load(config.Triple())
-			if err != nil {
-				return err
-			}
-			ldflags = append(ldflags, libc)
+			// Finally write the stack size to the binary.
+			binary.LittleEndian.PutUint32(data[i*4:], stackSize)
 		}
+	}

-		// 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)
-		}
+	// Write back the modified stack sizes.
+	_, err = fp.WriteAt(data, int64(section.Offset))
+	if err != nil {
+		return err
+	}

-		// Compile C files in packages.
-		for i, file := range extraFiles {
-			outpath := filepath.Join(dir, "pkg"+strconv.Itoa(i)+"-"+filepath.Base(file)+".o")
-			err := runCCompiler(config.Target.Compiler, append(config.CFlags(), "-c", "-o", outpath, file)...)
-			if err != nil {
-				return &commandError{"failed to build", file, err}
-			}
-			ldflags = append(ldflags, outpath)
-		}
+	return nil
+}

-		// Link the object files together.
-		err = link(config.Target.Linker, ldflags...)
-		if err != nil {
-			return &commandError{"failed to link", executable, err}
+// 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)
 		}
-
-		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)
-			}
-		}
-
-		// 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
-			}
-		}
-		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
@@ -101,7 +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.
@@ -132,13 +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)) {
@@ -154,8 +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)
@@ -183,8 +190,9 @@ bool AssemblerInvocation::CreateFromArgs(AssemblerInvocation &Opts,
  Opts.NoExecStack = Args.hasArg(OPT_mno_exec_stack);
  Opts.FatalWarnings = Args.hasArg(OPT_massembler_fatal_warnings);
  Opts.NoWarn = Args.hasArg(OPT_massembler_no_warn);
-  Opts.RelocationModel = Args.getLastArgValue(OPT_mrelocation_model, "pic");
-  Opts.TargetABI = Args.getLastArgValue(OPT_target_abi);
+  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);
@@ -314,12 +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;

@@ -383,7 +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.
@@ -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;
@@ -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 > 14 {
-		return nil, fmt.Errorf("requires go version 1.11, 1.12, 1.13, or 1.14, got go%d.%d", major, minor)
+	if major != 1 || minor < 13 || minor > 16 {
+		return nil, fmt.Errorf("requires go version 1.13 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
+	}
+}
@@ -1,7 +1,6 @@
 package builder

 import (
-	"io/ioutil"
 	"os"
 	"path/filepath"
 	"strings"
@@ -40,63 +39,112 @@ func (l *Library) sourcePaths(target string) []string {
 }

 // Load the library archive, possibly generating and caching it if needed.
-func (l *Library) Load(target string) (path string, err error) {
+// 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
+		return precompiledPath, nil, nil
 	}

-	outfile := l.name + "-" + target + ".a"
+	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, commands["clang"][0], l.sourcePaths(target)); path != "" || err != nil {
+	if path, err := cacheLoad(outfile, l.sourcePaths(target)); path != "" || err != nil {
 		// Cache hit.
-		return path, err
+		return path, nil, err
 	}
 	// Cache miss, build it now.

-	dirPrefix := "tinygo-" + l.name
-	remapDir := filepath.Join(os.TempDir(), dirPrefix)
-	dir, err := ioutil.TempDir(os.TempDir(), dirPrefix)
+	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 "", err
+		return "", nil, err
 	}
-	defer os.RemoveAll(dir)

 	// 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")
+		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")
+	}

-	// Compile all sources.
+	// 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)
-		// 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.
-		err := runCCompiler("clang", append(args, "-o", objpath, srcpath)...)
-		if err != nil {
-			return "", &commandError{"failed to build", srcpath, err}
-		}
+		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
+			},
+		})
 	}

-	// Put all the object files in a single archive. This archive file will be
-	// used to statically link this library.
-	arpath := filepath.Join(dir, l.name+".a")
-	err = makeArchive(arpath, objs)
-	if err != nil {
-		return "", err
-	}
-
-	// Store this archive in the cache.
-	return cacheStore(arpath, outfile, commands["clang"][0], l.sourcePaths(target))
+	return arpath, job, nil
 }
@@ -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")
 	}
@@ -41,6 +41,8 @@ type cgoPackage struct {
 	elaboratedTypes map[string]*elaboratedTypeInfo
 	enums           map[string]enumInfo
 	anonStructNum   int
+	ldflags         []string
+	visitedFiles    map[string][]byte
 }

 // constantInfo stores some information about a CGo constant found by libclang
@@ -53,9 +55,10 @@ 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).
@@ -154,9 +157,10 @@ typedef unsigned long long  _Cgo_ulonglong;
 // Process extracts `import "C"` statements from the AST, parses the comment
 // with libclang, and modifies the AST to use this information. It returns a
 // 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) {
+// files, the LDFLAGS for this package, and a map of file hashes of the accessed
+// C header 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, []string, map[string][]byte, []error) {
 	p := &cgoPackage{
 		dir:             dir,
 		fset:            fset,
@@ -168,6 +172,7 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 		typedefs:        map[string]*typedefInfo{},
 		elaboratedTypes: map[string]*elaboratedTypeInfo{},
 		enums:           map[string]enumInfo{},
+		visitedFiles:    map[string][]byte{},
 	}

 	// Disable _FORTIFY_SOURCE as it causes problems on macOS.
@@ -183,7 +188,7 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 	// Find the absolute path for this package.
 	packagePath, err := filepath.Abs(fset.File(files[0].Pos()).Name())
 	if err != nil {
-		return nil, []error{
+		return nil, nil, nil, []error{
 			scanner.Error{
 				Pos: fset.Position(files[0].Pos()),
 				Msg: "cgo: cannot find absolute path: " + err.Error(), // TODO: wrap this error
@@ -359,6 +364,19 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 					}
 					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)
@@ -412,7 +430,7 @@ 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.visitedFiles, p.errors
 }

 // makePathsAbsolute converts some common path compiler flags (-I, -L) from
@@ -470,6 +488,16 @@ 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{
@@ -5,6 +5,7 @@ import (
 	"flag"
 	"fmt"
 	"go/ast"
+	"go/build"
 	"go/format"
 	"go/parser"
 	"go/token"
@@ -23,7 +24,7 @@ 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)
+	actual := strings.ReplaceAll(result, "\r\n", "\n")

 	// 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
@@ -41,6 +42,20 @@ func TestCGo(t *testing.T) {
 	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()
@@ -50,7 +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
@@ -98,7 +113,7 @@ func TestCGo(t *testing.T) {
 			if err != nil {
 				t.Fatalf("could not read expected output: %v", err)
 			}
-			expected := strings.Replace(string(expectedBytes), "\r\n", "\n", -1)
+			expected := strings.ReplaceAll(string(expectedBytes), "\r\n", "\n")

 			// Check whether the output is as expected.
 			if expected != actual {
@@ -139,7 +154,7 @@ func formatDiagnostic(err error) string {
 	msg := err.Error()
 	if runtime.GOOS == "windows" {
 		// Fix Windows path slashes.
-		msg = strings.Replace(msg, "testdata\\", "testdata/", -1)
+		msg = strings.ReplaceAll(msg, "testdata\\", "testdata/")
 	}
 	return "//     " + msg + "\n"
 }
@@ -4,6 +4,7 @@ package cgo
 // modification. It does not touch the AST itself.

 import (
+	"crypto/sha512"
 	"fmt"
 	"go/ast"
 	"go/scanner"
@@ -56,6 +57,7 @@ unsigned tinygo_clang_Cursor_isBitField(GoCXCursor c);
 int tinygo_clang_globals_visitor(GoCXCursor c, GoCXCursor parent, CXClientData client_data);
 int tinygo_clang_struct_visitor(GoCXCursor c, GoCXCursor parent, CXClientData client_data);
 int tinygo_clang_enum_visitor(GoCXCursor c, GoCXCursor parent, CXClientData client_data);
+void tinygo_clang_inclusion_visitor(CXFile included_file, CXSourceLocation *inclusion_stack, unsigned include_len, CXClientData client_data);
 */
 import "C"

@@ -114,6 +116,7 @@ func (p *cgoPackage) parseFragment(fragment string, cflags []string, posFilename
 	}
 	defer C.clang_disposeTranslationUnit(unit)

+	// Report parser and type errors.
 	if numDiagnostics := int(C.clang_getNumDiagnostics(unit)); numDiagnostics != 0 {
 		addDiagnostic := func(diagnostic C.CXDiagnostic) {
 			spelling := getString(C.clang_getDiagnosticSpelling(diagnostic))
@@ -134,10 +137,36 @@ func (p *cgoPackage) parseFragment(fragment string, cflags []string, posFilename
 		}
 	}

+	// Extract information required by CGo.
 	ref := storedRefs.Put(p)
 	defer storedRefs.Remove(ref)
 	cursor := C.tinygo_clang_getTranslationUnitCursor(unit)
 	C.tinygo_clang_visitChildren(cursor, C.CXCursorVisitor(C.tinygo_clang_globals_visitor), C.CXClientData(ref))
+
+	// Determine files read during CGo processing, for caching.
+	inclusionCallback := func(includedFile C.CXFile) {
+		// Get full file path.
+		path := getString(C.clang_getFileName(includedFile))
+
+		// Get contents of file (that should be in-memory).
+		size := C.size_t(0)
+		rawData := C.clang_getFileContents(unit, includedFile, &size)
+		if rawData == nil {
+			// Sanity check. This should (hopefully) never trigger.
+			panic("libclang: file contents was not loaded")
+		}
+		data := (*[1 << 24]byte)(unsafe.Pointer(rawData))[:size]
+
+		// Hash the contents if it isn't hashed yet.
+		if _, ok := p.visitedFiles[path]; !ok {
+			// already stored
+			sum := sha512.Sum512_224(data)
+			p.visitedFiles[path] = sum[:]
+		}
+	}
+	inclusionCallbackRef := storedRefs.Put(inclusionCallback)
+	defer storedRefs.Remove(inclusionCallbackRef)
+	C.clang_getInclusions(unit, C.CXInclusionVisitor(C.tinygo_clang_inclusion_visitor), C.CXClientData(inclusionCallbackRef))
 }

 //export tinygo_clang_globals_visitor
@@ -152,12 +181,10 @@ 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 +273,9 @@ 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.
@@ -774,3 +801,9 @@ func tinygo_clang_enum_visitor(c, parent C.GoCXCursor, client_data C.CXClientDat
 	}
 	return C.CXChildVisit_Continue
 }
+
+//export tinygo_clang_inclusion_visitor
+func tinygo_clang_inclusion_visitor(includedFile C.CXFile, inclusionStack *C.CXSourceLocation, includeLen C.unsigned, clientData C.CXClientData) {
+	callback := storedRefs.Get(unsafe.Pointer(clientData)).(func(C.CXFile))
+	callback(includedFile)
+}
@@ -1,14 +1,14 @@
 // +build !byollvm
-// +build !llvm9
+// +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
+#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"
@@ -1,14 +0,0 @@
-// +build !byollvm
-// +build llvm9
-
-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/llvm9/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/llvm9/lib -lclang
-*/
-import "C"
@@ -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

@@ -21,6 +21,13 @@ 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"

@@ -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

@@ -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
@@ -7,7 +7,6 @@ import (
 	"fmt"
 	"path/filepath"
 	"regexp"
-	"strconv"
 	"strings"

 	"github.com/tinygo-org/tinygo/goenv"
@@ -22,29 +21,6 @@ type Config struct {
 	TestConfig     TestConfig
 }

-// FuncValueImplementation is an enum for the particular implementations of Go
-// func values.
-type FuncValueImplementation int
-
-// These constants describe the various possible implementations of Go func
-// values.
-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
-)
-
 // Triple returns the LLVM target triple, like armv6m-none-eabi.
 func (c *Config) Triple() string {
 	return c.Target.Triple
@@ -118,19 +94,19 @@ func (c *Config) NeedsStackObjects() bool {
 	switch c.GC() {
 	case "conservative", "extalloc":
 		for _, tag := range c.BuildTags() {
-			if tag == "baremetal" {
-				return false
+			if tag == "wasm" {
+				return true
 			}
 		}

-		return true
+		return false
 	default:
 		return false
 	}
 }

-// 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
@@ -144,14 +120,24 @@ func (c *Config) Scheduler() string {

 // FuncImplementation picks an appropriate func value implementation for the
 // target.
-func (c *Config) FuncImplementation() FuncValueImplementation {
-	// Always pick the switch implementation, as it allows the use of blocking
-	// inside a function that is used as a func value.
+func (c *Config) FuncImplementation() string {
 	switch c.Scheduler() {
-	case "none", "coroutines":
-		return FuncValueSwitch
 	case "tasks":
-		return FuncValueDoubleword
+		// 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")
 	}
@@ -164,18 +150,31 @@ 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 {
 	cflags := append([]string{}, c.Options.CFlags...)
 	for _, flag := range c.Target.CFlags {
-		cflags = append(cflags, strings.Replace(flag, "{root}", goenv.Get("TINYGOROOT"), -1))
+		cflags = append(cflags, strings.ReplaceAll(flag, "{root}", goenv.Get("TINYGOROOT")))
 	}
 	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
 }

@@ -187,15 +186,9 @@ func (c *Config) LDFlags() []string {
 	// Merge and adjust LDFlags.
 	ldflags := append([]string{}, c.Options.LDFlags...)
 	for _, flag := range c.Target.LDFlags {
-		ldflags = append(ldflags, strings.Replace(flag, "{root}", root, -1))
+		ldflags = append(ldflags, strings.ReplaceAll(flag, "{root}", root))
 	}
 	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)
 	}
@@ -226,6 +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.
@@ -265,6 +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)
 	}
@@ -272,6 +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"

@@ -33,11 +34,13 @@ type TargetSpec struct {
 	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"`
@@ -45,90 +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
-	}
-	if spec2.Libc != "" {
-		spec.Libc = spec2.Libc
-	}
-	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())
+		}
 	}
 }

@@ -177,11 +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
@@ -247,6 +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",
@@ -262,39 +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)
+	}
+
+}
@@ -16,7 +16,7 @@ import (
 // slice. This is required by the Go language spec: an index out of bounds must
 // cause a panic.
 func (b *builder) createLookupBoundsCheck(arrayLen, index llvm.Value, indexType types.Type) {
-	if b.fn.IsNoBounds() {
+	if b.info.nobounds {
 		// The //go:nobounds pragma was added to the function to avoid bounds
 		// checking.
 		return
@@ -48,7 +48,7 @@ func (b *builder) createLookupBoundsCheck(arrayLen, index llvm.Value, indexType
 // biggest possible slice capacity, 'low' means len and 'high' means cap. The
 // logic is the same in both cases.
 func (b *builder) createSliceBoundsCheck(capacity, low, high, max llvm.Value, lowType, highType, maxType *types.Basic) {
-	if b.fn.IsNoBounds() {
+	if b.info.nobounds {
 		// The //go:nobounds pragma was added to the function to avoid bounds
 		// checking.
 		return
@@ -104,7 +104,7 @@ func (b *builder) createSliceBoundsCheck(capacity, low, high, max llvm.Value, lo
 // 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.fn.IsNoBounds() {
+	if b.info.nobounds {
 		// The //go:nobounds pragma was added to the function to avoid bounds
 		// checking.
 		return
@@ -189,7 +189,7 @@ func (b *builder) createNilCheck(inst ssa.Value, ptr llvm.Value, blockPrefix str
 // 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.fn.IsNoBounds() {
+	if b.info.nobounds {
 		// Function disabled bounds checking - skip shift check.
 		return
 	}
@@ -212,8 +212,8 @@ func (b *builder) createRuntimeAssert(assert llvm.Value, blockPrefix, assertFunc
 		}
 	}

-	faultBlock := b.ctx.AddBasicBlock(b.fn.LLVMFn, blockPrefix+".throw")
-	nextBlock := b.ctx.AddBasicBlock(b.fn.LLVMFn, blockPrefix+".next")
+	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.
@@ -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
+	}
+}
@@ -4,6 +4,7 @@ import (
 	"go/types"
 	"strconv"

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

@@ -34,14 +35,14 @@ const (

 // 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 {
-	fullName := "runtime." + fnName
-	fn := b.mod.NamedFunction(fullName)
-	if fn.IsNil() {
-		panic("trying to call non-existent function: " + fullName)
+	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(fn, args, name)
+	return b.createCall(llvmFn, args, name)
 }

 // createCall creates a call to the given function with the arguments possibly
@@ -12,7 +12,7 @@ import (
 )

 func (b *builder) createMakeChan(expr *ssa.MakeChan) llvm.Value {
-	elementSize := b.targetData.TypeAllocSize(b.getLLVMType(expr.Type().(*types.Chan).Elem()))
+	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())
@@ -35,27 +35,37 @@ func (b *builder) createChanSend(instr *ssa.Send) {
 	valueAlloca, valueAllocaCast, valueAllocaSize := b.createTemporaryAlloca(valueType, "chan.value")
 	b.CreateStore(chanValue, valueAlloca)

-	// Do the send.
-	b.createRuntimeCall("chanSend", []llvm.Value{ch, valueAllocaCast}, "")
+	// Allocate blockedlist buffer.
+	channelBlockedList := b.mod.GetTypeByName("runtime.channelBlockedList")
+	channelBlockedListAlloca, channelBlockedListAllocaCast, channelBlockedListAllocaSize := b.createTemporaryAlloca(channelBlockedList, "chan.blockedList")

-	// End the lifetime of the alloca.
+	// Do the send.
+	b.createRuntimeCall("chanSend", []llvm.Value{ch, valueAllocaCast, channelBlockedListAlloca}, "")
+
+	// 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
+	b.emitLifetimeEnd(channelBlockedListAllocaCast, channelBlockedListAllocaSize)
 	b.emitLifetimeEnd(valueAllocaCast, valueAllocaSize)
 }

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

 	// Allocate memory to receive into.
 	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 := b.createRuntimeCall("chanRecv", []llvm.Value{ch, valueAllocaCast}, "")
+	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 {
@@ -69,8 +79,7 @@ func (b *builder) createChanRecv(unop *ssa.UnOp) llvm.Value {
 }

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

@@ -117,7 +126,7 @@ func (b *builder) createSelect(expr *ssa.Select) llvm.Value {
 		switch state.Dir {
 		case types.RecvOnly:
 			// Make sure the receive buffer is big enough and has the correct alignment.
-			llvmType := b.getLLVMType(state.Chan.Type().(*types.Chan).Elem())
+			llvmType := b.getLLVMType(state.Chan.Type().Underlying().(*types.Chan).Elem())
 			if size := b.targetData.TypeAllocSize(llvmType); size > recvbufSize {
 				recvbufSize = size
 			}
@@ -0,0 +1,173 @@
+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.
+			program := lprogram.LoadSSA()
+			pkg := lprogram.MainPkg()
+			mod, errs := CompilePackage(testCase, pkg, program.Package(pkg.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"
 )
@@ -25,9 +26,11 @@ import (
 // calls.
 func (b *builder) deferInitFunc() {
 	// Some setup.
-	b.deferFuncs = make(map[*ir.Function]int)
+	b.deferFuncs = make(map[*ssa.Function]int)
 	b.deferInvokeFuncs = make(map[string]int)
-	b.deferClosureFuncs = make(map[*ir.Function]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(b.getLLVMRuntimeType("_defer"), 0)
@@ -104,13 +107,11 @@ func (b *builder) createDefer(instr *ssa.Defer) {

 	} else if callee, ok := instr.Call.Value.(*ssa.Function); ok {
 		// Regular function call.
-		fn := b.ir.GetFunction(callee)
-
-		if _, ok := b.deferFuncs[fn]; !ok {
-			b.deferFuncs[fn] = len(b.allDeferFuncs)
-			b.allDeferFuncs = append(b.allDeferFuncs, fn)
+		if _, ok := b.deferFuncs[callee]; !ok {
+			b.deferFuncs[callee] = len(b.allDeferFuncs)
+			b.allDeferFuncs = append(b.allDeferFuncs, callee)
 		}
-		callback := llvm.ConstInt(b.uintptrType, uint64(b.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).
@@ -132,7 +133,7 @@ func (b *builder) createDefer(instr *ssa.Defer) {
 		context := b.CreateExtractValue(closure, 0, "")

 		// Get the callback number.
-		fn := b.ir.GetFunction(makeClosure.Fn.(*ssa.Function))
+		fn := makeClosure.Fn.(*ssa.Function)
 		if _, ok := b.deferClosureFuncs[fn]; !ok {
 			b.deferClosureFuncs[fn] = len(b.allDeferFuncs)
 			b.allDeferFuncs = append(b.allDeferFuncs, makeClosure)
@@ -151,9 +152,53 @@ func (b *builder) createDefer(instr *ssa.Defer) {
 		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 {
-		b.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.
@@ -175,7 +220,7 @@ func (b *builder) createDefer(instr *ssa.Defer) {
 		allocCall := b.createRuntimeCall("alloc", []llvm.Value{sizeValue}, "defer.alloc.call")
 		alloca = b.CreateBitCast(allocCall, llvm.PointerType(deferFrameType, 0), "defer.alloc")
 	}
-	if b.NeedsStackObjects() {
+	if b.NeedsStackObjects {
 		b.trackPointer(alloca)
 	}
 	b.CreateStore(deferFrame, alloca)
@@ -203,10 +248,10 @@ func (b *builder) createRunDefers() {
 	//     }

 	// Create loop.
-	loophead := b.ctx.AddBasicBlock(b.fn.LLVMFn, "rundefers.loophead")
-	loop := b.ctx.AddBasicBlock(b.fn.LLVMFn, "rundefers.loop")
-	unreachable := b.ctx.AddBasicBlock(b.fn.LLVMFn, "rundefers.default")
-	end := b.ctx.AddBasicBlock(b.fn.LLVMFn, "rundefers.end")
+	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:
@@ -238,21 +283,28 @@ func (b *builder) createRunDefers() {
 		// Create switch case, for example:
 		//     case 0:
 		//         // run first deferred call
-		block := b.ctx.AddBasicBlock(b.fn.LLVMFn, "rundefers.callback")
+		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{b.uintptrType, llvm.PointerType(b.getLLVMRuntimeType("_defer"), 0), b.uintptrType, b.i8ptrType}
+			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)
+			}
+
 			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")

@@ -265,26 +317,42 @@ func (b *builder) createRunDefers() {
 				forwardParams = append(forwardParams, forwardParam)
 			}

-			// Isolate the typecode.
-			typecode, forwardParams := forwardParams[0], forwardParams[1:]
+			var fnPtr llvm.Value

-			// 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))
+			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(b.i8ptrType))

-			fnPtr := b.getInvokePtr(callback, typecode)
 			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{b.uintptrType, llvm.PointerType(b.getLLVMRuntimeType("_defer"), 0)}
-			for _, param := range callback.Params {
+			for _, param := range getParams(callback.Signature) {
 				valueTypes = append(valueTypes, b.getLLVMType(param.Type()))
 			}
 			deferFrameType := b.ctx.StructType(valueTypes, false)
@@ -293,7 +361,7 @@ func (b *builder) createRunDefers() {
 			// Extract the params from the struct.
 			forwardParams := []llvm.Value{}
 			zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
-			for i := range callback.Params {
+			for i := range getParams(callback.Signature) {
 				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)
@@ -301,7 +369,7 @@ func (b *builder) createRunDefers() {

 			// 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 !callback.IsExported() {
+			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))
@@ -311,11 +379,11 @@ func (b *builder) createRunDefers() {
 			}

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

 		case *ssa.MakeClosure:
 			// Get the real defer struct type and cast to it.
-			fn := b.ir.GetFunction(callback.Fn.(*ssa.Function))
+			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++ {
@@ -338,8 +406,35 @@ func (b *builder) createRunDefers() {
 			forwardParams = append(forwardParams, llvm.Undef(b.i8ptrType))

 			// Call deferred function.
-			b.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")
 		}
@@ -14,7 +14,7 @@ import (
 // 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,
 	}
@@ -6,7 +6,6 @@ package compiler
 import (
 	"go/types"

-	"github.com/tinygo-org/tinygo/compileopts"
 	"golang.org/x/tools/go/ssa"
 	"tinygo.org/x/go-llvm"
 )
@@ -21,11 +20,11 @@ func (b *builder) createFuncValue(funcPtr, context llvm.Value, sig *types.Signat
 // context.
 func (c *compilerContext) createFuncValue(builder llvm.Builder, funcPtr, context llvm.Value, sig *types.Signature) llvm.Value {
 	var funcValueScalar llvm.Value
-	switch c.FuncImplementation() {
-	case compileopts.FuncValueDoubleword:
+	switch c.FuncImplementation {
+	case "doubleword":
 		// Closure is: {context, function pointer}
 		funcValueScalar = funcPtr
-	case compileopts.FuncValueSwitch:
+	case "switch":
 		sigGlobal := c.getTypeCode(sig)
 		funcValueWithSignatureGlobalName := funcPtr.Name() + "$withSignature"
 		funcValueWithSignatureGlobal := c.mod.NamedGlobal(funcValueWithSignatureGlobalName)
@@ -38,7 +37,7 @@ func (c *compilerContext) createFuncValue(builder llvm.Builder, funcPtr, context
 			funcValueWithSignatureGlobal = llvm.AddGlobal(c.mod, funcValueWithSignatureType, funcValueWithSignatureGlobalName)
 			funcValueWithSignatureGlobal.SetInitializer(funcValueWithSignature)
 			funcValueWithSignatureGlobal.SetGlobalConstant(true)
-			funcValueWithSignatureGlobal.SetLinkage(llvm.InternalLinkage)
+			funcValueWithSignatureGlobal.SetLinkage(llvm.LinkOnceODRLinkage)
 		}
 		funcValueScalar = llvm.ConstPtrToInt(funcValueWithSignatureGlobal, c.uintptrType)
 	default:
@@ -67,10 +66,10 @@ func (b *builder) extractFuncContext(funcValue llvm.Value) llvm.Value {
 // value. This may be an expensive operation.
 func (b *builder) decodeFuncValue(funcValue llvm.Value, sig *types.Signature) (funcPtr, context llvm.Value) {
 	context = b.CreateExtractValue(funcValue, 0, "")
-	switch b.FuncImplementation() {
-	case compileopts.FuncValueDoubleword:
+	switch b.FuncImplementation {
+	case "doubleword":
 		funcPtr = b.CreateExtractValue(funcValue, 1, "")
-	case compileopts.FuncValueSwitch:
+	case "switch":
 		llvmSig := b.getRawFuncType(sig)
 		sigGlobal := b.getTypeCode(sig)
 		funcPtr = b.createRuntimeCall("getFuncPtr", []llvm.Value{funcValue, sigGlobal}, "")
@@ -83,11 +82,11 @@ func (b *builder) decodeFuncValue(funcValue llvm.Value, sig *types.Signature) (f

 // getFuncType returns the type of a func value given a signature.
 func (c *compilerContext) getFuncType(typ *types.Signature) llvm.Type {
-	switch c.FuncImplementation() {
-	case compileopts.FuncValueDoubleword:
+	switch c.FuncImplementation {
+	case "doubleword":
 		rawPtr := c.getRawFuncType(typ)
 		return c.ctx.StructType([]llvm.Type{c.i8ptrType, rawPtr}, false)
-	case compileopts.FuncValueSwitch:
+	case "switch":
 		return c.getLLVMRuntimeType("funcValue")
 	default:
 		panic("unimplemented func value variant")
@@ -149,7 +148,7 @@ func (b *builder) parseMakeClosure(expr *ssa.MakeClosure) (llvm.Value, error) {
 	if len(expr.Bindings) == 0 {
 		panic("unexpected: MakeClosure without bound variables")
 	}
-	f := b.ir.GetFunction(expr.Fn.(*ssa.Function))
+	f := expr.Fn.(*ssa.Function)

 	// Collect all bound variables.
 	boundVars := make([]llvm.Value, len(expr.Bindings))
@@ -164,5 +163,5 @@ func (b *builder) parseMakeClosure(expr *ssa.MakeClosure) (llvm.Value, error) {
 	context := b.emitPointerPack(boundVars)

 	// Create the closure.
-	return b.createFuncValue(f.LLVMFn, context, f.Signature), nil
+	return b.createFuncValue(b.getFunction(f), context, f.Signature), nil
 }
@@ -7,6 +7,7 @@ import (
 	"go/token"

 	"github.com/tinygo-org/tinygo/compiler/llvmutil"
+	"golang.org/x/tools/go/ssa"
 	"tinygo.org/x/go-llvm"
 )

@@ -19,16 +20,32 @@ import (
 // Because a go statement doesn't return anything, return undef.
 func (b *builder) createGoInstruction(funcPtr llvm.Value, params []llvm.Value, prefix string, pos token.Pos) llvm.Value {
 	paramBundle := b.emitPointerPack(params)
-	var callee llvm.Value
-	switch b.Scheduler() {
+	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":
 		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")
 	}
-	b.createCall(b.mod.NamedFunction("internal/task.start"), []llvm.Value{callee, paramBundle, llvm.Undef(b.i8ptrType), llvm.ConstPointerNull(b.i8ptrType)}, "")
+	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())
 }

@@ -67,13 +84,14 @@ func (c *compilerContext) createGoroutineStartWrapper(fn llvm.Value, prefix stri
 		// 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.LinkOnceODRLinkage)
 		wrapper.SetUnnamedAddr(true)
+		wrapper.AddAttributeAtIndex(-1, c.ctx.CreateStringAttribute("tinygo-gowrapper", name))
 		entry := c.ctx.AddBasicBlock(wrapper, "entry")
 		builder.SetInsertPointAtEnd(entry)

-		if c.Debug() {
-			pos := c.ir.Program.Fset.Position(pos)
+		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
@@ -123,13 +141,14 @@ func (c *compilerContext) createGoroutineStartWrapper(fn llvm.Value, prefix stri
 		// Create the wrapper.
 		wrapperType := llvm.FunctionType(c.ctx.VoidType(), []llvm.Type{c.i8ptrType}, false)
 		wrapper = llvm.AddFunction(c.mod, prefix+".gowrapper", wrapperType)
-		wrapper.SetLinkage(llvm.InternalLinkage)
+		wrapper.SetLinkage(llvm.LinkOnceODRLinkage)
 		wrapper.SetUnnamedAddr(true)
+		wrapper.AddAttributeAtIndex(-1, c.ctx.CreateStringAttribute("tinygo-gowrapper", ""))
 		entry := c.ctx.AddBasicBlock(wrapper, "entry")
 		builder.SetInsertPointAtEnd(entry)

-		if c.Debug() {
-			pos := c.ir.Program.Fset.Position(pos)
+		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
@@ -163,6 +163,44 @@ func (b *builder) emitSVCall(args []ssa.Value) (llvm.Value, error) {
 	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:
 //
 //     func (csr CSR) Get() uintptr
@@ -11,7 +11,6 @@ import (
 	"strconv"
 	"strings"

-	"github.com/tinygo-org/tinygo/ir"
 	"golang.org/x/tools/go/ssa"
 	"tinygo.org/x/go-llvm"
 )
@@ -32,7 +31,7 @@ func (b *builder) createMakeInterface(val llvm.Value, typ types.Type, pos token.
 		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)
+		itfConcreteTypeGlobal.SetLinkage(llvm.LinkOnceODRLinkage)
 	}
 	itfTypeCode := b.CreatePtrToInt(itfConcreteTypeGlobal, b.uintptrType, "")
 	itf := llvm.Undef(b.getLLVMRuntimeType("_interface"))
@@ -81,7 +80,7 @@ func (c *compilerContext) getTypeCode(typ types.Type) llvm.Value {
 				globalValue = llvm.ConstInsertValue(globalValue, lengthValue, []uint32{1})
 			}
 			global.SetInitializer(globalValue)
-			global.SetLinkage(llvm.PrivateLinkage)
+			global.SetLinkage(llvm.LinkOnceODRLinkage)
 		}
 		global.SetGlobalConstant(true)
 	}
@@ -236,7 +235,7 @@ func (c *compilerContext) 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 *compilerContext) 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
 	}
@@ -264,7 +264,7 @@ func (c *compilerContext) getTypeMethodSet(typ types.Type) llvm.Value {
 	global = llvm.AddGlobal(c.mod, arrayType, typ.String()+"$methodset")
 	global.SetInitializer(value)
 	global.SetGlobalConstant(true)
-	global.SetLinkage(llvm.PrivateLinkage)
+	global.SetLinkage(llvm.LinkOnceODRLinkage)
 	return llvm.ConstGEP(global, []llvm.Value{zero, zero})
 }

@@ -295,7 +295,7 @@ func (c *compilerContext) getInterfaceMethodSet(typ types.Type) llvm.Value {
 	global = llvm.AddGlobal(c.mod, value.Type(), name+"$interface")
 	global.SetInitializer(value)
 	global.SetGlobalConstant(true)
-	global.SetLinkage(llvm.PrivateLinkage)
+	global.SetLinkage(llvm.LinkOnceODRLinkage)
 	return llvm.ConstGEP(global, []llvm.Value{zero, zero})
 }

@@ -303,7 +303,7 @@ func (c *compilerContext) getInterfaceMethodSet(typ types.Type) llvm.Value {
 // external *i8 indicating the indicating the signature of this method. It is
 // used during the interface lowering pass.
 func (c *compilerContext) getMethodSignature(method *types.Func) llvm.Value {
-	signature := ir.MethodSignature(method)
+	signature := methodSignature(method)
 	signatureGlobal := c.mod.NamedGlobal("func " + signature)
 	if signatureGlobal.IsNil() {
 		signatureGlobal = llvm.AddGlobal(c.mod, c.ctx.Int8Type(), "func "+signature)
@@ -357,8 +357,8 @@ func (b *builder) createTypeAssert(expr *ssa.TypeAssert) llvm.Value {
 	// value.

 	prevBlock := b.GetInsertBlock()
-	okBlock := b.ctx.AddBasicBlock(b.fn.LLVMFn, "typeassert.ok")
-	nextBlock := b.ctx.AddBasicBlock(b.fn.LLVMFn, "typeassert.next")
+	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)

@@ -436,8 +436,8 @@ func (b *builder) getInvokeCall(instr *ssa.CallCommon) (llvm.Value, []llvm.Value
 // 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(f *ir.Function) llvm.Value {
-	wrapperName := f.LinkName() + "$invoke"
+func (c *compilerContext) getInterfaceInvokeWrapper(fn *ssa.Function, llvmFn llvm.Value) llvm.Value {
+	wrapperName := llvmFn.Name() + "$invoke"
 	wrapper := c.mod.NamedFunction(wrapperName)
 	if !wrapper.IsNil() {
 		// Wrapper already created. Return it directly.
@@ -445,7 +445,7 @@ func (c *compilerContext) getInterfaceInvokeWrapper(f *ir.Function) llvm.Value {
 	}

 	// Get the expanded receiver type.
-	receiverType := c.getLLVMType(f.Params[0].Type())
+	receiverType := c.getLLVMType(fn.Signature.Recv().Type())
 	var expandedReceiverType []llvm.Type
 	for _, info := range expandFormalParamType(receiverType, "", nil) {
 		expandedReceiverType = append(expandedReceiverType, info.llvmType)
@@ -457,19 +457,17 @@ func (c *compilerContext) 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")
-	}
+	wrapper.LastParam().SetName("parentHandle")

-	wrapper.SetLinkage(llvm.InternalLinkage)
+	wrapper.SetLinkage(llvm.LinkOnceODRLinkage)
 	wrapper.SetUnnamedAddr(true)

 	// Create a new builder just to create this wrapper.
@@ -480,9 +478,9 @@ func (c *compilerContext) getInterfaceInvokeWrapper(f *ir.Function) llvm.Value {
 	defer b.Builder.Dispose()

 	// add debug info if needed
-	if c.Debug() {
-		pos := c.ir.Program.Fset.Position(f.Pos())
-		difunc := c.attachDebugInfoRaw(f, wrapper, "$invoke", pos.Filename, pos.Line)
+	if c.Debug {
+		pos := c.program.Fset.Position(fn.Pos())
+		difunc := c.attachDebugInfoRaw(fn, wrapper, "$invoke", pos.Filename, pos.Line)
 		b.SetCurrentDebugLocation(uint(pos.Line), uint(pos.Column), difunc, llvm.Metadata{})
 	}

@@ -492,13 +490,60 @@ func (c *compilerContext) getInterfaceInvokeWrapper(f *ir.Function) llvm.Value {

 	receiverValue := b.emitPointerUnpack(wrapper.Param(0), []llvm.Type{receiverType})[0]
 	params := append(b.expandFormalParam(receiverValue), wrapper.Params()[1:]...)
-	if f.LLVMFn.Type().ElementType().ReturnType().TypeKind() == llvm.VoidTypeKind {
-		b.CreateCall(f.LLVMFn, params, "")
+	if llvmFn.Type().ElementType().ReturnType().TypeKind() == llvm.VoidTypeKind {
+		b.CreateCall(llvmFn, params, "")
 		b.CreateRetVoid()
 	} else {
-		ret := b.CreateCall(f.LLVMFn, params, "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
+}
@@ -39,14 +39,14 @@ func (b *builder) createInterruptGlobal(instr *ssa.CallCommon) (llvm.Value, erro

 	// Create a new global of type runtime/interrupt.handle. Globals of this
 	// type are lowered in the interrupt lowering pass.
-	globalType := b.ir.Program.ImportedPackage("runtime/interrupt").Type("handle").Type()
+	globalType := b.program.ImportedPackage("runtime/interrupt").Type("handle").Type()
 	globalLLVMType := b.getLLVMType(globalType)
 	globalName := "runtime/interrupt.$interrupt" + strconv.FormatInt(id.Int64(), 10)
 	if global := b.mod.NamedGlobal(globalName); !global.IsNil() {
 		return llvm.Value{}, b.makeError(instr.Pos(), "interrupt redeclared in this program")
 	}
 	global := llvm.AddGlobal(b.mod, globalLLVMType, globalName)
-	global.SetLinkage(llvm.PrivateLinkage)
+	global.SetVisibility(llvm.HiddenVisibility)
 	global.SetGlobalConstant(true)
 	global.SetUnnamedAddr(true)
 	initializer := llvm.ConstNull(globalLLVMType)
@@ -55,8 +55,8 @@ func (b *builder) createInterruptGlobal(instr *ssa.CallCommon) (llvm.Value, erro
 	global.SetInitializer(initializer)

 	// Add debug info to the interrupt global.
-	if b.Debug() {
-		pos := b.ir.Program.Fset.Position(instr.Pos())
+	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,
@@ -79,7 +79,7 @@ func (b *builder) createInterruptGlobal(instr *ssa.CallCommon) (llvm.Value, erro
 	// 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(b.Triple(), "avr") {
+	if strings.HasPrefix(b.Triple, "avr") {
 		useFn := b.mod.NamedFunction("runtime/interrupt.use")
 		if useFn.IsNil() {
 			useFnType := llvm.FunctionType(b.ctx.VoidType(), []llvm.Type{interrupt.Type()}, false)
@@ -44,7 +44,7 @@ func (b *builder) emitLifetimeEnd(ptr, size llvm.Value) {
 // 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 (b *builder) emitPointerPack(values []llvm.Value) llvm.Value {
-	return llvmutil.EmitPointerPack(b.Builder, b.mod, b.Config, values)
+	return llvmutil.EmitPointerPack(b.Builder, b.mod, b.NeedsStackObjects, values)
 }

 // emitPointerUnpack extracts a list of values packed using emitPointerPack.
@@ -5,7 +5,6 @@ package llvmutil
 // itself if possible and legal.

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

@@ -13,7 +12,7 @@ import (
 // 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.
 // If the values are all constants, they are be stored in a constant global and deduplicated.
-func EmitPointerPack(builder llvm.Builder, mod llvm.Module, config *compileopts.Config, values []llvm.Value) llvm.Value {
+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)
@@ -26,7 +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)
@@ -39,9 +37,39 @@ 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
@@ -67,12 +95,12 @@ func EmitPointerPack(builder llvm.Builder, mod llvm.Module, config *compileopts.
 		// 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,
@@ -80,28 +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
 	}
 }
@@ -15,6 +15,298 @@ 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 getParams(fn.Signature) {
+		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)
+			}
+		}
+	}
+
+	// Synthetic functions are functions that do not appear in the source code,
+	// they are artificially constructed. Usually they are wrapper functions
+	// that are not referenced anywhere except in a SSA call instruction so
+	// should be created right away.
+	// The exception is the package initializer, which does appear in the
+	// *ssa.Package members and so shouldn't be created here.
+	if fn.Synthetic != "" && fn.Synthetic != "package initializer" {
+		irbuilder := c.ctx.NewBuilder()
+		b := newBuilder(c, irbuilder, fn)
+		b.createFunction()
+		irbuilder.Dispose()
+		llvmFn.SetLinkage(llvm.LinkOnceODRLinkage)
+		llvmFn.SetUnnamedAddr(true)
+	}
+
+	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
+			}
+		}
+
+	}
+}
+
+// getParams returns the function parameters, including the receiver at the
+// start. This is an alternative to the Params member of *ssa.Function, which is
+// not yet populated when the package has not yet been built.
+func getParams(sig *types.Signature) []*types.Var {
+	params := []*types.Var{}
+	if sig.Recv() != nil {
+		params = append(params, sig.Recv())
+	}
+	for i := 0; i < sig.Params().Len(); i++ {
+		params = append(params, sig.Params().At(i))
+	}
+	return params
+}
+
 // 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).
@@ -26,25 +318,22 @@ 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 *compilerContext) loadASTComments(lprogram *loader.Program) {
-	c.astComments = map[string]*ast.CommentGroup{}
-	for _, pkgInfo := range lprogram.Sorted() {
-		for _, file := range pkgInfo.Files {
-			for _, decl := range file.Decls {
-				switch decl := decl.(type) {
-				case *ast.GenDecl:
-					switch decl.Tok {
-					case token.VAR:
-						if len(decl.Specs) != 1 {
-							continue
-						}
-						for _, spec := range decl.Specs {
-							switch spec := spec.(type) {
-							case *ast.ValueSpec: // decl.Tok == token.VAR
-								for _, name := range spec.Names {
-									id := pkgInfo.Pkg.Path() + "." + name.Name
-									c.astComments[id] = decl.Doc
-								}
+func (c *compilerContext) loadASTComments(pkg *loader.Package) {
+	for _, file := range pkg.Files {
+		for _, decl := range file.Decls {
+			switch decl := decl.(type) {
+			case *ast.GenDecl:
+				switch decl.Tok {
+				case token.VAR:
+					if len(decl.Specs) != 1 {
+						continue
+					}
+					for _, spec := range decl.Specs {
+						switch spec := spec.(type) {
+						case *ast.ValueSpec: // decl.Tok == token.VAR
+							for _, name := range spec.Names {
+								id := pkg.Pkg.Path() + "." + name.Name
+								c.astComments[id] = decl.Doc
 							}
 						}
 					}
@@ -63,10 +352,6 @@ func (c *compilerContext) getGlobal(g *ssa.Global) llvm.Value {
 		typ := g.Type().(*types.Pointer).Elem()
 		llvmType := c.getLLVMType(typ)
 		llvmGlobal = llvm.AddGlobal(c.mod, llvmType, info.linkName)
-		if !info.extern {
-			llvmGlobal.SetInitializer(llvm.ConstNull(llvmType))
-			llvmGlobal.SetLinkage(llvm.InternalLinkage)
-		}

 		// Set alignment from the //go:align comment.
 		var alignInBits uint32
@@ -82,11 +367,9 @@ func (c *compilerContext) 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,
@@ -145,3 +428,23 @@ 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
+}
@@ -16,8 +16,8 @@ func (b *builder) createSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 	num := b.getValue(call.Args[0])
 	var syscallResult llvm.Value
 	switch {
-	case b.GOARCH() == "amd64":
-		if b.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.
@@ -58,7 +58,7 @@ func (b *builder) createSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 		fnType := llvm.FunctionType(b.uintptrType, argTypes, false)
 		target := llvm.InlineAsm(fnType, "syscall", constraints, true, false, llvm.InlineAsmDialectIntel)
 		syscallResult = b.CreateCall(target, args, "")
-	case b.GOARCH() == "386" && b.GOOS() == "linux":
+	case b.GOARCH == "386" && b.GOOS == "linux":
 		// Sources:
 		//   syscall(2) man page
 		//   https://stackoverflow.com/a/2538212
@@ -84,7 +84,7 @@ func (b *builder) createSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 		fnType := llvm.FunctionType(b.uintptrType, argTypes, false)
 		target := llvm.InlineAsm(fnType, "int 0x80", constraints, true, false, llvm.InlineAsmDialectIntel)
 		syscallResult = b.CreateCall(target, args, "")
-	case b.GOARCH() == "arm" && b.GOOS() == "linux":
+	case b.GOARCH == "arm" && b.GOOS == "linux":
 		// Implement the EABI system call convention for Linux.
 		// Source: syscall(2) man page.
 		args := []llvm.Value{}
@@ -116,7 +116,7 @@ func (b *builder) createSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 		fnType := llvm.FunctionType(b.uintptrType, argTypes, false)
 		target := llvm.InlineAsm(fnType, "svc #0", constraints, true, false, 0)
 		syscallResult = b.CreateCall(target, args, "")
-	case b.GOARCH() == "arm64" && b.GOOS() == "linux":
+	case b.GOARCH == "arm64" && b.GOOS == "linux":
 		// Source: syscall(2) man page.
 		args := []llvm.Value{}
 		argTypes := []llvm.Type{}
@@ -149,9 +149,9 @@ func (b *builder) createSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 		target := llvm.InlineAsm(fnType, "svc #0", constraints, true, false, 0)
 		syscallResult = b.CreateCall(target, args, "")
 	default:
-		return llvm.Value{}, b.makeError(call.Pos(), "unknown GOOS/GOARCH for syscall: "+b.GOOS()+"/"+b.GOARCH())
+		return llvm.Value{}, b.makeError(call.Pos(), "unknown GOOS/GOARCH for syscall: "+b.GOOS+"/"+b.GOARCH)
 	}
-	switch b.GOOS() {
+	switch b.GOOS {
 	case "linux", "freebsd":
 		// Return values: r0, r1 uintptr, err Errno
 		// Pseudocode:
@@ -187,6 +187,6 @@ func (b *builder) createSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 		retval = b.CreateInsertValue(retval, errResult, 2, "")
 		return retval, nil
 	default:
-		return llvm.Value{}, b.makeError(call.Pos(), "unknown GOOS/GOARCH for syscall: "+b.GOOS()+"/"+b.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 +1,100 @@
+; 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"
+
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret void
+}
+
+define hidden i32 @main.addInt(i32 %x, i32 %y, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = add i32 %x, %y
+  ret i32 %0
+}
+
+define hidden i1 @main.equalInt(i32 %x, i32 %y, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = icmp eq i32 %x, %y
+  ret i1 %0
+}
+
+define hidden i1 @main.floatEQ(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = fcmp oeq float %x, %y
+  ret i1 %0
+}
+
+define hidden i1 @main.floatNE(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = fcmp une float %x, %y
+  ret i1 %0
+}
+
+define hidden i1 @main.floatLower(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = fcmp olt float %x, %y
+  ret i1 %0
+}
+
+define hidden i1 @main.floatLowerEqual(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = fcmp ole float %x, %y
+  ret i1 %0
+}
+
+define hidden i1 @main.floatGreater(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = fcmp ogt float %x, %y
+  ret i1 %0
+}
+
+define hidden i1 @main.floatGreaterEqual(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = fcmp oge float %x, %y
+  ret i1 %0
+}
+
+define hidden float @main.complexReal(float %x.r, float %x.i, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret float %x.r
+}
+
+define hidden float @main.complexImag(float %x.r, float %x.i, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret float %x.i
+}
+
+define hidden { 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 hidden { 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 hidden { 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 +1,82 @@
+; 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"
+
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret void
+}
+
+define hidden 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 hidden float @main.maxu32f(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret float 0x41F0000000000000
+}
+
+define hidden i32 @main.maxu32tof32(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret i32 -1
+}
+
+define hidden { 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 hidden 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 hidden 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 hidden 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 hidden 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 +1,51 @@
+; 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"
+
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret void
+}
+
+define hidden [0 x i32] @main.pointerDerefZero([0 x i32]* %x, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret [0 x i32] zeroinitializer
+}
+
+define hidden i32* @main.pointerCastFromUnsafe(i8* %x, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = bitcast i8* %x to i32*
+  ret i32* %0
+}
+
+define hidden 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 hidden i8* @main.pointerCastToUnsafeNoop(i8* dereferenceable_or_null(1) %x, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret i8* %x
+}
+
+define hidden 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 hidden 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 hidden 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 +1,21 @@
+; 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"
+
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret void
+}
+
+define hidden i32 @main.sliceLen(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret i32 %ints.len
+}
+
+define hidden i32 @main.sliceCap(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret i32 %ints.cap
+}
@@ -1,13 +1,16 @@
 module github.com/tinygo-org/tinygo

-go 1.11
+go 1.13

 require (
 	github.com/blakesmith/ar v0.0.0-20150311145944-8bd4349a67f2
+	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 v1.0.0
+	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
-	google.golang.org/appengine v1.4.0 // indirect
-	tinygo.org/x/go-llvm v0.0.0-20200401165421-8d120882fc7a
+	tinygo.org/x/go-llvm v0.0.0-20210308112806-9ef958b6bed4
 )
@@ -1,59 +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/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/golang/protobuf v1.2.0/go.mod h1:6lQm79b+lXiMfvg/cZm0SGofjICqVBUtrP5yJMmIC1U=
+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=
+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 v0.0.0-20180724234803-3673e40ba225/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
-golang.org/x/net v0.0.0-20190213061140-3a22650c66bd/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 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 v0.0.0-20181108010431-42b317875d0f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 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=
 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=
-google.golang.org/appengine v1.4.0/go.mod h1:xpcJRLb0r/rnEns0DIKYYv+WjYCduHsrkT7/EB5XEv4=
 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 v0.0.0-20190224120431-7707ae5d1261 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=
-tinygo.org/x/go-llvm v0.0.0-20200226165415-53522ab6713d h1:mtgZh/e8a3wxneQFuLXoQYO//1mvlki02yZ1JCwMKp4=
-tinygo.org/x/go-llvm v0.0.0-20200226165415-53522ab6713d/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
-tinygo.org/x/go-llvm v0.0.0-20200401165421-8d120882fc7a h1:Ugje2Lxuv8CFncHzs5W+hWfJvPsM+W4K0zRvzFbLvoE=
-tinygo.org/x/go-llvm v0.0.0-20200401165421-8d120882fc7a/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
+tinygo.org/x/go-llvm v0.0.0-20210308112806-9ef958b6bed4 h1:CMUHxVTb+UuUePuMf8vkWjZ3gTp9BBK91KrgOCwoNHs=
+tinygo.org/x/go-llvm v0.0.0-20210308112806-9ef958b6bed4/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
@@ -142,10 +142,11 @@ func getHomeDir() string {
 // getGoroot returns an appropriate GOROOT from various sources. If it can't be
 // found, it returns an empty string.
 func getGoroot() string {
+	// An explicitly set GOROOT always has preference.
 	goroot := os.Getenv("GOROOT")
 	if goroot != "" {
-		// An explicitly set GOROOT always has preference.
-		return goroot
+		// Convert to the standard GOROOT being referenced, if it's a TinyGo cache.
+		return getStandardGoroot(goroot)
 	}

 	// Check for the location of the 'go' binary and base GOROOT on that.
@@ -195,3 +196,21 @@ func isGoroot(goroot string) bool {
 	_, err := os.Stat(filepath.Join(goroot, "src", "runtime", "internal", "sys", "zversion.go"))
 	return err == nil
 }
+
+// getStandardGoroot returns the physical path to a real, standard Go GOROOT
+// implied by the given path.
+// If the given path appears to be a TinyGo cached GOROOT, it returns the path
+// referenced by symlinks contained in the cache. Otherwise, it returns the
+// given path as-is.
+func getStandardGoroot(path string) string {
+	// Check if the "bin" subdirectory of our given GOROOT is a symlink, and then
+	// return the _parent_ directory of its destination.
+	if dest, err := os.Readlink(filepath.Join(path, "bin")); nil == err {
+		// Clean the destination to remove any trailing slashes, so that
+		// filepath.Dir will always return the parent.
+		//   (because both "/foo" and "/foo/" are valid symlink destinations,
+		//   but filepath.Dir would return "/" and "/foo", respectively)
+		return filepath.Dir(filepath.Clean(dest))
+	}
+	return path
+}
@@ -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 = "0.18.0-dev"
+
+// 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",
+}
@@ -11,15 +11,17 @@ 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 = &Error{Err: errors.New("interp: unreachable executed")}
+// These errors are expected during normal execution and can be recovered from
+// by running the affected function at runtime instead of compile time.
+var (
+	errIntegerAsPointer       = 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")
+)

-// 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) *Error {
-	return e.errorAt(inst, errors.New("interp: unsupported instruction"))
+func isRecoverableError(err error) bool {
+	return err == errIntegerAsPointer || err == errUnsupportedInst || err == errUnsupportedRuntimeInst || err == errMapAlreadyCreated
 }

 // ErrorLine is one line in a traceback. The position may be missing.
@@ -46,11 +48,13 @@ func (e *Error) Error() string {
 // 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, err error) *Error {
+func (r *runner) errorAt(inst instruction, err error) *Error {
+	pos := getPosition(inst.llvmInst)
 	return &Error{
-		ImportPath: e.packagePath,
-		Pos:        getPosition(inst),
+		ImportPath: r.pkgName,
+		Pos:        pos,
 		Err:        err,
+		Traceback:  []ErrorLine{{pos, inst.llvmInst}},
 	}
 }

@@ -1,680 +0,0 @@
-package interp
-
-// This file implements the core interpretation routines, interpreting single
-// functions.
-
-import (
-	"errors"
-	"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, errors.New("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, errors.New("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)
-			}
-			if result.Type() != inst.Type() {
-				return nil, nil, fr.errorAt(inst, errors.New("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, errors.New("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, fr.errorAt(inst, err)
-					}
-					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, fr.errorAt(inst, err)
-					}
-					dstArray = dstArrayValue.(*LocalValue)
-					srcArrayValue, err := srcArray.GetElementPtr([]uint32{0, 0})
-					if err != nil {
-						return nil, nil, fr.errorAt(inst, err)
-					}
-					srcArray = srcArrayValue.(*LocalValue)
-				}
-				if fr.Eval.TargetData.TypeAllocSize(dstArray.Type().ElementType()) != elementSize {
-					return nil, nil, fr.errorAt(inst, errors.New("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, errors.New("interp: slice src element size does not match pointer type"))
-				}
-				if dstArray.Type() != srcArray.Type() {
-					return nil, nil, fr.errorAt(inst, errors.New("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, errors.New("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, fr.errorAt(inst, err)
-					}
-					dstArray = dstArrayValue.(*LocalValue)
-					// src++
-					srcArrayValue, err := srcArray.GetElementPtr([]uint32{1})
-					if err != nil {
-						return nil, nil, fr.errorAt(inst, err)
-					}
-					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.typeAssert":
-				actualTypeInt := fr.getLocal(inst.Operand(0)).(*LocalValue).Underlying
-				assertedType := fr.getLocal(inst.Operand(1)).(*LocalValue).Underlying
-				if actualTypeInt.IsAConstantExpr().IsNil() || actualTypeInt.Opcode() != llvm.PtrToInt {
-					return nil, nil, fr.errorAt(inst, errors.New("interp: expected typecode in runtime.typeAssert to be a ptrtoint"))
-				}
-				actualType := actualTypeInt.Operand(0)
-				if actualType.IsAConstant().IsNil() || assertedType.IsAConstant().IsNil() {
-					return nil, nil, fr.errorAt(inst, errors.New("interp: unimplemented: type assert with non-constant interface value"))
-				}
-				assertOk := uint64(0)
-				if llvm.ConstExtractValue(actualType.Initializer(), []uint32{0}) == assertedType {
-					assertOk = 1
-				}
-				fr.locals[inst] = &LocalValue{fr.Eval, llvm.ConstInt(fr.Mod.Context().Int1Type(), assertOk, false)}
-			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, errors.New("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, errors.New("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, errors.New("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 {
-						// Record this function call in the backtrace.
-						err.Traceback = append(err.Traceback, ErrorLine{
-							Pos:  getPosition(inst),
-							Inst: inst,
-						})
-						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, errors.New("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, errors.New("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.IsASelectInst().IsNil():
-			// var result T
-			// if cond {
-			//   result = x
-			// } else {
-			//   result = y
-			// }
-			// return result
-			cond := fr.getLocal(inst.Operand(0)).(*LocalValue).Underlying
-			x := fr.getLocal(inst.Operand(1)).(*LocalValue).Underlying
-			y := fr.getLocal(inst.Operand(2)).(*LocalValue).Underlying
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateSelect(cond, x, y, "")}
-
-		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, errors.New("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, errors.New("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, errors.New("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, errors.New("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 interpreting", r.pkgName, "because of error:", callErr.Error())
+				}
+				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"

@@ -41,9 +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()
@@ -66,6 +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.).
@@ -75,8 +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 +1,932 @@
+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 {
+				if err != errIntegerAsPointer {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				// GEP on fixed pointer value (for example, memory-mapped I/O).
+				ptrValue := operands[0].Uint() + offset
+				switch operands[0].len(r) {
+				case 8:
+					locals[inst.localIndex] = literalValue{uint64(ptrValue)}
+				case 4:
+					locals[inst.localIndex] = literalValue{uint32(ptrValue)}
+				case 2:
+					locals[inst.localIndex] = literalValue{uint16(ptrValue)}
+				default:
+					panic("pointer operand is not of a known pointer size")
+				}
+				continue
+			}
+			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?"
+	}
+}
@@ -1,253 +0,0 @@
-package interp
-
-import (
-	"errors"
-	"strings"
-
-	"tinygo.org/x/go-llvm"
-)
-
-type sideEffectSeverity int
-
-func (severity sideEffectSeverity) String() string {
-	switch severity {
-	case sideEffectInProgress:
-		return "in progress"
-	case sideEffectNone:
-		return "none"
-	case sideEffectLimited:
-		return "limited"
-	case sideEffectAll:
-		return "all"
-	default:
-		return "unknown"
-	}
-}
-
-const (
-	sideEffectInProgress sideEffectSeverity = iota // computing side effects is in progress (for recursive functions)
-	sideEffectNone                                 // no side effects at all (pure)
-	sideEffectLimited                              // has side effects, but the effects are known
-	sideEffectAll                                  // has unknown side effects
-)
-
-// sideEffectResult contains the scan results after scanning a function for side
-// effects (recursively).
-type sideEffectResult struct {
-	severity        sideEffectSeverity
-	mentionsGlobals map[llvm.Value]struct{}
-}
-
-// hasSideEffects scans this function and all descendants, recursively. It
-// returns whether this function has side effects and if it does, which globals
-// it mentions anywhere in this function or any called functions.
-func (e *evalPackage) hasSideEffects(fn llvm.Value) (*sideEffectResult, *Error) {
-	name := fn.Name()
-	switch {
-	case name == "runtime.alloc":
-		// Cannot be scanned but can be interpreted.
-		return &sideEffectResult{severity: sideEffectNone}, nil
-	case name == "runtime.nanotime":
-		// Fixed value at compile time.
-		return &sideEffectResult{severity: sideEffectNone}, nil
-	case name == "runtime._panic":
-		return &sideEffectResult{severity: sideEffectLimited}, nil
-	case name == "runtime.typeAssert":
-		return &sideEffectResult{severity: sideEffectNone}, nil
-	case name == "runtime.interfaceImplements":
-		return &sideEffectResult{severity: sideEffectNone}, nil
-	case name == "runtime.sliceCopy":
-		return &sideEffectResult{severity: sideEffectNone}, nil
-	case name == "runtime.trackPointer":
-		return &sideEffectResult{severity: sideEffectNone}, nil
-	case name == "llvm.dbg.value":
-		return &sideEffectResult{severity: sideEffectNone}, nil
-	case strings.HasPrefix(name, "llvm.lifetime."):
-		return &sideEffectResult{severity: sideEffectNone}, nil
-	}
-	if fn.IsDeclaration() {
-		return &sideEffectResult{severity: sideEffectLimited}, nil
-	}
-	if e.sideEffectFuncs == nil {
-		e.sideEffectFuncs = make(map[llvm.Value]*sideEffectResult)
-	}
-	if se, ok := e.sideEffectFuncs[fn]; ok {
-		return se, nil
-	}
-	result := &sideEffectResult{
-		severity:        sideEffectInProgress,
-		mentionsGlobals: map[llvm.Value]struct{}{},
-	}
-	e.sideEffectFuncs[fn] = result
-	dirtyLocals := map[llvm.Value]struct{}{}
-	for bb := fn.EntryBasicBlock(); !bb.IsNil(); bb = llvm.NextBasicBlock(bb) {
-		for inst := bb.FirstInstruction(); !inst.IsNil(); inst = llvm.NextInstruction(inst) {
-			if inst.IsAInstruction().IsNil() {
-				// Should not happen in valid IR.
-				panic("not an instruction")
-			}
-
-			// Check for any globals mentioned anywhere in the function. Assume
-			// any mentioned globals may be read from or written to when
-			// executed, thus must be marked dirty with a call.
-			for i := 0; i < inst.OperandsCount(); i++ {
-				operand := inst.Operand(i)
-				if !operand.IsAGlobalVariable().IsNil() {
-					result.mentionsGlobals[operand] = struct{}{}
-				}
-			}
-
-			switch inst.InstructionOpcode() {
-			case llvm.IndirectBr, llvm.Invoke:
-				// Not emitted by the compiler.
-				return nil, e.errorAt(inst, errors.New("unknown instructions"))
-			case llvm.Call:
-				child := inst.CalledValue()
-				if !child.IsAInlineAsm().IsNil() {
-					// Inline assembly. This most likely has side effects.
-					// Assume they're only limited side effects, similar to
-					// external function calls.
-					result.updateSeverity(sideEffectLimited)
-					continue
-				}
-				if child.IsAFunction().IsNil() {
-					// Indirect call?
-					// In any case, we can't know anything here about what it
-					// affects exactly so mark this function as invoking all
-					// possible side effects.
-					result.updateSeverity(sideEffectAll)
-					continue
-				}
-				if child.IsDeclaration() {
-					// External function call. Assume only limited side effects
-					// (no affected globals, etc.).
-					switch child.Name() {
-					case "runtime.typeAssert":
-						continue // implemented in interp
-					case "runtime.interfaceImplements":
-						continue // implemented in interp
-					}
-					if e.hasLocalSideEffects(dirtyLocals, inst) {
-						result.updateSeverity(sideEffectLimited)
-					}
-					continue
-				}
-				childSideEffects, err := e.hasSideEffects(child)
-				if err != nil {
-					return nil, err
-				}
-				switch childSideEffects.severity {
-				case sideEffectInProgress, sideEffectNone:
-					// no side effects or recursive function - continue scanning
-				case sideEffectLimited:
-					// The return value may be problematic.
-					if e.hasLocalSideEffects(dirtyLocals, inst) {
-						result.updateSeverity(sideEffectLimited)
-					}
-				case sideEffectAll:
-					result.updateSeverity(sideEffectAll)
-				default:
-					panic("unreachable")
-				}
-			case llvm.Load:
-				if inst.IsVolatile() {
-					result.updateSeverity(sideEffectLimited)
-				}
-				if _, ok := e.dirtyGlobals[inst.Operand(0)]; ok {
-					if e.hasLocalSideEffects(dirtyLocals, inst) {
-						result.updateSeverity(sideEffectLimited)
-					}
-				}
-			case llvm.Store:
-				if inst.IsVolatile() {
-					result.updateSeverity(sideEffectLimited)
-				}
-			case llvm.IntToPtr:
-				// Pointer casts are not yet supported.
-				result.updateSeverity(sideEffectLimited)
-			default:
-				// Ignore most instructions.
-				// Check this list for completeness:
-				// https://godoc.org/github.com/llvm-mirror/llvm/bindings/go/llvm#Opcode
-			}
-		}
-	}
-
-	if result.severity == sideEffectInProgress {
-		// No side effect was reported for this function.
-		result.severity = sideEffectNone
-	}
-	return result, nil
-}
-
-// hasLocalSideEffects checks whether the given instruction flows into a branch
-// or return instruction, in which case the whole function must be marked as
-// having side effects and be called at runtime.
-func (e *Eval) hasLocalSideEffects(dirtyLocals map[llvm.Value]struct{}, inst llvm.Value) bool {
-	if _, ok := dirtyLocals[inst]; ok {
-		// It is already known that this local is dirty.
-		return true
-	}
-
-	for use := inst.FirstUse(); !use.IsNil(); use = use.NextUse() {
-		user := use.User()
-		if user.IsAInstruction().IsNil() {
-			// Should not happen in valid IR.
-			panic("user not an instruction")
-		}
-		switch user.InstructionOpcode() {
-		case llvm.Br, llvm.Switch:
-			// A branch on a dirty value makes this function dirty: it cannot be
-			// interpreted at compile time so has to be run at runtime. It is
-			// marked as having side effects for this reason.
-			return true
-		case llvm.Ret:
-			// This function returns a dirty value so it is itself marked as
-			// dirty to make sure it is called at runtime.
-			return true
-		case llvm.Store:
-			ptr := user.Operand(1)
-			if !ptr.IsAGlobalVariable().IsNil() {
-				// Store to a global variable.
-				// Already handled in (*Eval).hasSideEffects.
-				continue
-			}
-			// This store might affect all kinds of values. While it is
-			// certainly possible to traverse through all of them, the easiest
-			// option right now is to just assume the worst and say that this
-			// function has side effects.
-			// TODO: traverse through all stores and mark all relevant allocas /
-			// globals dirty.
-			return true
-		default:
-			// All instructions that take 0 or more operands (1 or more if it
-			// was a use) and produce a result.
-			// For a list:
-			// https://godoc.org/github.com/llvm-mirror/llvm/bindings/go/llvm#Opcode
-			dirtyLocals[user] = struct{}{}
-			if e.hasLocalSideEffects(dirtyLocals, user) {
-				return true
-			}
-		}
-	}
-
-	// No side effects found.
-	return false
-}
-
-// updateSeverity sets r.severity to the max of r.severity and severity,
-// conservatively assuming the worst severity.
-func (r *sideEffectResult) updateSeverity(severity sideEffectSeverity) {
-	if severity > r.severity {
-		r.severity = severity
-	}
-}
-
-// updateSeverity updates the severity with the severity of the child severity,
-// like in a function call. This means it also copies the mentioned globals.
-func (r *sideEffectResult) update(child *sideEffectResult) {
-	r.updateSeverity(child.severity)
-	for global := range child.mentionsGlobals {
-		r.mentionsGlobals[global] = struct{}{}
-	}
-}
@@ -1,95 +0,0 @@
-package interp
-
-import (
-	"os"
-	"sort"
-	"testing"
-
-	"tinygo.org/x/go-llvm"
-)
-
-var scanTestTable = []struct {
-	name            string
-	severity        sideEffectSeverity
-	mentionsGlobals []string
-}{
-	{"returnsConst", sideEffectNone, nil},
-	{"returnsArg", sideEffectNone, nil},
-	{"externalCallOnly", sideEffectNone, nil},
-	{"externalCallAndReturn", sideEffectLimited, nil},
-	{"externalCallBranch", sideEffectLimited, nil},
-	{"readCleanGlobal", sideEffectNone, []string{"cleanGlobalInt"}},
-	{"readDirtyGlobal", sideEffectLimited, []string{"dirtyGlobalInt"}},
-	{"callFunctionPointer", sideEffectAll, []string{"functionPointer"}},
-	{"getDirtyPointer", sideEffectLimited, nil},
-	{"storeToPointer", sideEffectLimited, nil},
-	{"callTypeAssert", sideEffectNone, nil},
-	{"callInterfaceImplements", sideEffectNone, nil},
-}
-
-func TestScan(t *testing.T) {
-	t.Parallel()
-
-	// Read the input IR.
-	path := "testdata/scan.ll"
-	ctx := llvm.NewContext()
-	buf, err := llvm.NewMemoryBufferFromFile(path)
-	os.Stat(path) // make sure this file is tracked by `go test` caching
-	if err != nil {
-		t.Fatalf("could not read file %s: %v", path, err)
-	}
-	mod, err := ctx.ParseIR(buf)
-	if err != nil {
-		t.Fatalf("could not load module:\n%v", err)
-	}
-
-	// Check all to-be-tested functions.
-	for _, tc := range scanTestTable {
-		// Create an eval object, for testing.
-		e := &Eval{
-			Mod:          mod,
-			TargetData:   llvm.NewTargetData(mod.DataLayout()),
-			dirtyGlobals: map[llvm.Value]struct{}{},
-		}
-
-		// Mark some globals dirty, for testing.
-		e.markDirty(mod.NamedGlobal("dirtyGlobalInt"))
-
-		// Scan for side effects.
-		fn := mod.NamedFunction(tc.name)
-		if fn.IsNil() {
-			t.Errorf("scan test: could not find tested function %s in the IR", tc.name)
-			continue
-		}
-		evalPkg := &evalPackage{e, "testdata"}
-		result, err := evalPkg.hasSideEffects(fn)
-		if err != nil {
-			t.Errorf("scan test: failed to scan %s for side effects: %v", fn.Name(), err)
-		}
-
-		// Check whether the result is what we expect.
-		if result.severity != tc.severity {
-			t.Errorf("scan test: function %s should have severity %s but it has %s", tc.name, tc.severity, result.severity)
-		}
-
-		// Check whether the mentioned globals match with what we'd expect.
-		mentionsGlobalNames := make([]string, 0, len(result.mentionsGlobals))
-		for global := range result.mentionsGlobals {
-			mentionsGlobalNames = append(mentionsGlobalNames, global.Name())
-		}
-		sort.Strings(mentionsGlobalNames)
-		globalsMismatch := false
-		if len(result.mentionsGlobals) != len(tc.mentionsGlobals) {
-			globalsMismatch = true
-		} else {
-			for i, globalName := range mentionsGlobalNames {
-				if tc.mentionsGlobals[i] != globalName {
-					globalsMismatch = true
-				}
-			}
-		}
-		if globalsMismatch {
-			t.Errorf("scan test: expected %s to mention globals %v, but it mentions globals %v", tc.name, tc.mentionsGlobals, mentionsGlobalNames)
-		}
-	}
-}
@@ -4,6 +4,10 @@ target triple = "x86_64--linux"
@main.v1 = internal global i64 0
@main.nonConst1 = global [4 x i64] zeroinitializer
@main.nonConst2 = global i64 0
+@main.someArray = global [8 x {i16, i32}] zeroinitializer
+@main.exportedValue = global [1 x i16*] [i16* @main.exposedValue1]
+@main.exposedValue1 = global i16 0
+@main.exposedValue2 = global i16 0

 declare void @runtime.printint64(i64) unnamed_addr

@@ -47,6 +51,20 @@ entry:
  %value2 = load i64, i64* %gep2
  store i64 %value2, i64* @main.nonConst2

+  ; Test that the following GEP works:
+  ;     var someArray
+  ;     modifyExternal(&someArray[3].field1)
+  %gep3 = getelementptr [8 x {i16, i32}], [8 x {i16, i32}]* @main.someArray, i32 0, i32 3, i32 1
+  call void @modifyExternal(i32* %gep3)
+
+  ; Test that marking a value as external also marks all referenced values.
+  call void @modifyExternal(i32* bitcast ([1 x i16*]* @main.exportedValue to i32*))
+  store i16 5, i16* @main.exposedValue1
+
+  ; Test that this even propagates through functions.
+  call void @modifyExternal(i32* bitcast (void ()* @willModifyGlobal to i32*))
+  store i16 7, i16* @main.exposedValue2
+
  ret void
 }

@@ -58,3 +76,14 @@ entry:
 }

 declare i64 @someValue()
+
+declare void @modifyExternal(i32*)
+
+; This function will modify an external value. By passing this function as a
+; function pointer to an external function, @main.exposedValue2 should be
+; marked as external.
+define void @willModifyGlobal() {
+entry:
+  store i16 8, i16* @main.exposedValue2
+  ret void
+}
@@ -3,6 +3,10 @@ target triple = "x86_64--linux"

@main.nonConst1 = local_unnamed_addr global [4 x i64] zeroinitializer
@main.nonConst2 = local_unnamed_addr global i64 0
+@main.someArray = global [8 x { i16, i32 }] zeroinitializer
+@main.exportedValue = global [1 x i16*] [i16* @main.exposedValue1]
+@main.exposedValue1 = global i16 0
+@main.exposedValue2 = local_unnamed_addr global i16 0

 declare void @runtime.printint64(i64) unnamed_addr

@@ -16,6 +20,11 @@ entry:
  store i64 %value1, i64* getelementptr inbounds ([4 x i64], [4 x i64]* @main.nonConst1, i32 0, i32 0)
  %value2 = load i64, i64* getelementptr inbounds ([4 x i64], [4 x i64]* @main.nonConst1, i32 0, i32 0)
  store i64 %value2, i64* @main.nonConst2
+  call void @modifyExternal(i32* getelementptr inbounds ([8 x { i16, i32 }], [8 x { i16, i32 }]* @main.someArray, i32 0, i32 3, i32 1))
+  call void @modifyExternal(i32* bitcast ([1 x i16*]* @main.exportedValue to i32*))
+  store i16 5, i16* @main.exposedValue1
+  call void @modifyExternal(i32* bitcast (void ()* @willModifyGlobal to i32*))
+  store i16 7, i16* @main.exposedValue2
  ret void
 }

@@ -27,3 +36,11 @@ entry:
 }

 declare i64 @someValue() local_unnamed_addr
+
+declare void @modifyExternal(i32*) local_unnamed_addr
+
+define void @willModifyGlobal() {
+entry:
+  store i16 8, i16* @main.exposedValue2
+  ret void
+}
@@ -48,8 +48,7 @@ entry:
 define internal void @main.testNonConstantBinarySet() {
  %hashmap.key = alloca i8
  %hashmap.value = alloca i8
-  ; Create hashmap from global. This breaks the normal hashmapBinarySet
-  ; optimization, to test the fallback.
+  ; Create hashmap from global.
  %map.new = call %runtime.hashmap* @runtime.hashmapMake(i8 1, i8 1, i32 1, i8* undef, i8* null)
  store %runtime.hashmap* %map.new, %runtime.hashmap** @main.binaryMap
  %map = load %runtime.hashmap*, %runtime.hashmap** @main.binaryMap
@@ -64,8 +63,7 @@ define internal void @main.testNonConstantBinarySet() {
 ; operations (with string keys).
 define internal void @main.testNonConstantStringSet() {
  %hashmap.value = alloca i8
-  ; Create hashmap from global. This breaks the normal hashmapStringSet
-  ; optimization, to test the fallback.
+  ; Create hashmap from global.
  %map.new = call %runtime.hashmap* @runtime.hashmapMake(i8 8, i8 1, i32 1, i8* undef, i8* null)
  store %runtime.hashmap* %map.new, %runtime.hashmap** @main.stringMap
  %map = load %runtime.hashmap*, %runtime.hashmap** @main.stringMap
@@ -2,27 +2,19 @@ target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64"
 target triple = "armv6m-none-eabi"

 %runtime.hashmap = type { %runtime.hashmap*, i8*, i32, i8, i8, i8 }
-%runtime._string = type { i8*, i32 }

@main.m = local_unnamed_addr global %runtime.hashmap* @"main$map"
-@main.binaryMap = local_unnamed_addr global %runtime.hashmap* @"main$map.4"
-@main.stringMap = local_unnamed_addr global %runtime.hashmap* @"main$map.6"
+@main.binaryMap = local_unnamed_addr global %runtime.hashmap* @"main$map.1"
+@main.stringMap = local_unnamed_addr global %runtime.hashmap* @"main$map.3"
@main.init.string = internal unnamed_addr constant [7 x i8] c"CONNECT"
-@"main$mapbucket" = internal unnamed_addr global { [8 x i8], i8*, [8 x i8], [8 x %runtime._string] } { [8 x i8] c"\04\00\00\00\00\00\00\00", i8* null, [8 x i8] c"\01\00\00\00\00\00\00\00", [8 x %runtime._string] [%runtime._string { i8* getelementptr inbounds ([7 x i8], [7 x i8]* @main.init.string, i32 0, i32 0), i32 7 }, %runtime._string zeroinitializer, %runtime._string zeroinitializer, %runtime._string zeroinitializer, %runtime._string zeroinitializer, %runtime._string zeroinitializer, %runtime._string zeroinitializer, %runtime._string zeroinitializer] }
-@"main$map" = internal unnamed_addr global %runtime.hashmap { %runtime.hashmap* null, i8* getelementptr inbounds ({ [8 x i8], i8*, [8 x i8], [8 x %runtime._string] }, { [8 x i8], i8*, [8 x i8], [8 x %runtime._string] }* @"main$mapbucket", i32 0, i32 0, i32 0), i32 1, i8 1, i8 8, i8 0 }
-@"main$alloca.2" = internal global i8 1
-@"main$alloca.3" = internal global i8 2
-@"main$map.4" = internal unnamed_addr global %runtime.hashmap { %runtime.hashmap* null, i8* null, i32 0, i8 1, i8 1, i8 0 }
-@"main$alloca.5" = internal global i8 2
-@"main$map.6" = internal unnamed_addr global %runtime.hashmap { %runtime.hashmap* null, i8* null, i32 0, i8 8, i8 1, i8 0 }
-
-declare void @runtime.hashmapBinarySet(%runtime.hashmap*, i8*, i8*, i8*, i8*) local_unnamed_addr
-
-declare void @runtime.hashmapStringSet(%runtime.hashmap*, i8*, i32, i8*, i8*, i8*) local_unnamed_addr
+@"main$map" = internal global %runtime.hashmap { %runtime.hashmap* null, i8* getelementptr inbounds ({ [8 x i8], i8*, { i8, [7 x i8] }, { { [7 x i8]*, [4 x i8] }, [56 x i8] } }, { [8 x i8], i8*, { i8, [7 x i8] }, { { [7 x i8]*, [4 x i8] }, [56 x i8] } }* @"main$mapbucket", i32 0, i32 0, i32 0), i32 1, i8 1, i8 8, i8 0 }
+@"main$mapbucket" = internal unnamed_addr global { [8 x i8], i8*, { i8, [7 x i8] }, { { [7 x i8]*, [4 x i8] }, [56 x i8] } } { [8 x i8] c"\04\00\00\00\00\00\00\00", i8* null, { i8, [7 x i8] } { i8 1, [7 x i8] zeroinitializer }, { { [7 x i8]*, [4 x i8] }, [56 x i8] } { { [7 x i8]*, [4 x i8] } { [7 x i8]* @main.init.string, [4 x i8] c"\07\00\00\00" }, [56 x i8] zeroinitializer } }
+@"main$map.1" = internal global %runtime.hashmap { %runtime.hashmap* null, i8* getelementptr inbounds ({ [8 x i8], i8*, { i8, [7 x i8] }, { i8, [7 x i8] } }, { [8 x i8], i8*, { i8, [7 x i8] }, { i8, [7 x i8] } }* @"main$mapbucket.2", i32 0, i32 0, i32 0), i32 1, i8 1, i8 1, i8 0 }
+@"main$mapbucket.2" = internal unnamed_addr global { [8 x i8], i8*, { i8, [7 x i8] }, { i8, [7 x i8] } } { [8 x i8] c"\04\00\00\00\00\00\00\00", i8* null, { i8, [7 x i8] } { i8 1, [7 x i8] zeroinitializer }, { i8, [7 x i8] } { i8 2, [7 x i8] zeroinitializer } }
+@"main$map.3" = internal global %runtime.hashmap { %runtime.hashmap* null, i8* getelementptr inbounds ({ [8 x i8], i8*, { { [7 x i8]*, [4 x i8] }, [56 x i8] }, { i8, [7 x i8] } }, { [8 x i8], i8*, { { [7 x i8]*, [4 x i8] }, [56 x i8] }, { i8, [7 x i8] } }* @"main$mapbucket.4", i32 0, i32 0, i32 0), i32 1, i8 8, i8 1, i8 0 }
+@"main$mapbucket.4" = internal unnamed_addr global { [8 x i8], i8*, { { [7 x i8]*, [4 x i8] }, [56 x i8] }, { i8, [7 x i8] } } { [8 x i8] c"x\00\00\00\00\00\00\00", i8* null, { { [7 x i8]*, [4 x i8] }, [56 x i8] } { { [7 x i8]*, [4 x i8] } { [7 x i8]* @main.init.string, [4 x i8] c"\07\00\00\00" }, [56 x i8] zeroinitializer }, { i8, [7 x i8] } { i8 2, [7 x i8] zeroinitializer } }

 define void @runtime.initAll() unnamed_addr {
 entry:
-  call void @runtime.hashmapBinarySet(%runtime.hashmap* @"main$map.4", i8* @"main$alloca.2", i8* @"main$alloca.3", i8* undef, i8* null)
-  call void @runtime.hashmapStringSet(%runtime.hashmap* @"main$map.6", i8* getelementptr inbounds ([7 x i8], [7 x i8]* @main.init.string, i32 0, i32 0), i32 7, i8* @"main$alloca.5", i8* undef, i8* null)
  ret void
 }
@@ -1,78 +0,0 @@
-target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
-target triple = "x86_64--linux"
-
-%runtime.typecodeID = type { %runtime.typecodeID*, i64 }
-
-declare i1 @runtime.typeAssert(i64, %runtime.typecodeID*, i8*, i8*)
-declare i1 @runtime.interfaceImplements(i64, i8**)
-
-define i64 @returnsConst() {
-  ret i64 0
-}
-
-define i64 @returnsArg(i64 %arg) {
-  ret i64 %arg
-}
-
-declare i64 @externalCall()
-
-define i64 @externalCallOnly() {
-  %result = call i64 @externalCall()
-  ret i64 0
-}
-
-define i64 @externalCallAndReturn() {
-  %result = call i64 @externalCall()
-  ret i64 %result
-}
-
-define i64 @externalCallBranch() {
-  %result = call i64 @externalCall()
-  %zero = icmp eq i64 %result, 0
-  br i1 %zero, label %if.then, label %if.done
-
-if.then:
-  ret i64 2
-
-if.done:
-  ret i64 4
-}
-
-@cleanGlobalInt = global i64 5
-define i64 @readCleanGlobal() {
-  %global = load i64, i64* @cleanGlobalInt
-  ret i64 %global
-}
-
-@dirtyGlobalInt = global i64 5
-define i64 @readDirtyGlobal() {
-  %global = load i64, i64* @dirtyGlobalInt
-  ret i64 %global
-}
-
-declare i64* @getDirtyPointer()
-
-define void @storeToPointer() {
-  %ptr = call i64* @getDirtyPointer()
-  store i64 3, i64* %ptr
-  ret void
-}
-
-@functionPointer = global i64()* null
-define i64 @callFunctionPointer() {
-  %fp = load i64()*, i64()** @functionPointer
-  %result = call i64 %fp()
-  ret i64 %result
-}
-
-define i1 @callTypeAssert() {
-  ; Note: parameters are not realistic.
-  %ok = call i1 @runtime.typeAssert(i64 0, %runtime.typecodeID* null, i8* undef, i8* null)
-  ret i1 %ok
-}
-
-define i1 @callInterfaceImplements() {
-  ; Note: parameters are not realistic.
-  %ok = call i1 @runtime.interfaceImplements(i64 0, i8** null)
-  ret i1 %ok
-}
@@ -1,6 +1,8 @@
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64--linux"

+@"main$alloc.1" = internal unnamed_addr constant [6 x i8] c"\05\00{\00\00\04"
+
 declare void @runtime.printuint8(i8) local_unnamed_addr

 declare void @runtime.printint16(i16) local_unnamed_addr
@@ -15,6 +17,7 @@ entry:
  call void @runtime.printuint8(i8 3)
  call void @runtime.printuint8(i8 3)
  call void @runtime.printint16(i16 5)
-  call void @runtime.printint16(i16 5)
+  %int16SliceDst.val = load i16, i16* bitcast ([6 x i8]* @"main$alloc.1" to i16*)
+  call void @runtime.printint16(i16 %int16SliceDst.val)
  ret void
 }
@@ -1,121 +0,0 @@
-package interp
-
-import (
-	"tinygo.org/x/go-llvm"
-)
-
-// Return a list of values (actually, instructions) where this value is used as
-// an operand.
-func getUses(value llvm.Value) []llvm.Value {
-	var uses []llvm.Value
-	use := value.FirstUse()
-	for !use.IsNil() {
-		uses = append(uses, use.User())
-		use = use.NextUse()
-	}
-	return uses
-}
-
-// getStringBytes loads the byte slice of a Go string represented as a
-// {ptr, len} pair.
-func getStringBytes(strPtr Value, strLen llvm.Value) []byte {
-	if !strLen.IsConstant() {
-		panic("getStringBytes with a non-constant length")
-	}
-	buf := make([]byte, strLen.ZExtValue())
-	for i := range buf {
-		gep, err := strPtr.GetElementPtr([]uint32{uint32(i)})
-		if err != nil {
-			panic(err) // TODO
-		}
-		c := gep.Load()
-		buf[i] = byte(c.ZExtValue())
-	}
-	return buf
-}
-
-// getLLVMIndices converts an []uint32 into an []llvm.Value, for use in
-// llvm.ConstGEP.
-func getLLVMIndices(int32Type llvm.Type, indices []uint32) []llvm.Value {
-	llvmIndices := make([]llvm.Value, len(indices))
-	for i, index := range indices {
-		llvmIndices[i] = llvm.ConstInt(int32Type, uint64(index), false)
-	}
-	return llvmIndices
-}
-
-// Return true if this type is a scalar value (integer or floating point), false
-// otherwise.
-func isScalar(t llvm.Type) bool {
-	switch t.TypeKind() {
-	case llvm.IntegerTypeKind, llvm.FloatTypeKind, llvm.DoubleTypeKind:
-		return true
-	default:
-		return false
-	}
-}
-
-// isPointerNil returns whether this is a nil pointer or not. The ok value
-// indicates whether the result is certain: if it is false the result boolean is
-// not valid.
-func isPointerNil(v llvm.Value) (result bool, ok bool) {
-	if !v.IsAConstantExpr().IsNil() {
-		switch v.Opcode() {
-		case llvm.IntToPtr:
-			// Whether a constant inttoptr is nil is easy to
-			// determine.
-			result, ok = isZero(v.Operand(0))
-			if ok {
-				return
-			}
-		case llvm.BitCast, llvm.GetElementPtr:
-			// These const instructions are just a kind of wrappers for the
-			// underlying pointer.
-			return isPointerNil(v.Operand(0))
-		}
-	}
-	if !v.IsAConstantPointerNull().IsNil() {
-		// A constant pointer null is always null, of course.
-		return true, true
-	}
-	if !v.IsAGlobalValue().IsNil() {
-		// A global value is never null.
-		return false, true
-	}
-	return false, false // not valid
-}
-
-// isZero returns whether the value in v is the integer zero, and whether that
-// can be known right now.
-func isZero(v llvm.Value) (result bool, ok bool) {
-	if !v.IsAConstantExpr().IsNil() {
-		switch v.Opcode() {
-		case llvm.PtrToInt:
-			return isPointerNil(v.Operand(0))
-		}
-	}
-	if !v.IsAConstantInt().IsNil() {
-		val := v.ZExtValue()
-		return val == 0, true
-	}
-	return false, false // not valid
-}
-
-// unwrap returns the underlying value, with GEPs removed. This can be useful to
-// get the underlying global of a GEP pointer.
-func unwrap(value llvm.Value) llvm.Value {
-	for {
-		if !value.IsAConstantExpr().IsNil() {
-			switch value.Opcode() {
-			case llvm.GetElementPtr:
-				value = value.Operand(0)
-				continue
-			}
-		} else if !value.IsAGetElementPtrInst().IsNil() {
-			value = value.Operand(0)
-			continue
-		}
-		break
-	}
-	return value
-}
@@ -1,419 +0,0 @@
-package interp
-
-// This file provides a litte bit of abstraction around LLVM values.
-
-import (
-	"errors"
-	"strconv"
-
-	"tinygo.org/x/go-llvm"
-)
-
-// A Value is a LLVM value with some extra methods attached for easier
-// interpretation.
-type Value interface {
-	Value() llvm.Value                     // returns a LLVM value
-	Type() llvm.Type                       // equal to Value().Type()
-	IsConstant() bool                      // returns true if this value is a constant value
-	Load() llvm.Value                      // dereference a pointer
-	Store(llvm.Value)                      // store to a pointer
-	GetElementPtr([]uint32) (Value, error) // returns an interior pointer
-	String() string                        // string representation, for debugging
-}
-
-// A type that simply wraps a LLVM constant value.
-type LocalValue struct {
-	Eval       *Eval
-	Underlying llvm.Value
-}
-
-// Value implements Value by returning the constant value itself.
-func (v *LocalValue) Value() llvm.Value {
-	return v.Underlying
-}
-
-func (v *LocalValue) Type() llvm.Type {
-	return v.Underlying.Type()
-}
-
-func (v *LocalValue) IsConstant() bool {
-	if _, ok := v.Eval.dirtyGlobals[unwrap(v.Underlying)]; ok {
-		return false
-	}
-	return v.Underlying.IsConstant()
-}
-
-// Load loads a constant value if this is a constant pointer.
-func (v *LocalValue) Load() llvm.Value {
-	if !v.Underlying.IsAGlobalVariable().IsNil() {
-		return v.Underlying.Initializer()
-	}
-	switch v.Underlying.Opcode() {
-	case llvm.GetElementPtr:
-		indices := v.getConstGEPIndices()
-		if indices[0] != 0 {
-			panic("invalid GEP")
-		}
-		global := v.Eval.getValue(v.Underlying.Operand(0))
-		agg := global.Load()
-		return llvm.ConstExtractValue(agg, indices[1:])
-	case llvm.BitCast:
-		panic("interp: load from a bitcast")
-	default:
-		panic("interp: load from a constant")
-	}
-}
-
-// Store stores to the underlying value if the value type is a pointer type,
-// otherwise it panics.
-func (v *LocalValue) Store(value llvm.Value) {
-	if !v.Underlying.IsAGlobalVariable().IsNil() {
-		if !value.IsConstant() {
-			v.MarkDirty()
-			v.Eval.builder.CreateStore(value, v.Underlying)
-		} else {
-			v.Underlying.SetInitializer(value)
-		}
-		return
-	}
-	if !value.IsConstant() {
-		v.MarkDirty()
-		v.Eval.builder.CreateStore(value, v.Underlying)
-		return
-	}
-	switch v.Underlying.Opcode() {
-	case llvm.GetElementPtr:
-		indices := v.getConstGEPIndices()
-		if indices[0] != 0 {
-			panic("invalid GEP")
-		}
-		global := &LocalValue{v.Eval, v.Underlying.Operand(0)}
-		agg := global.Load()
-		agg = llvm.ConstInsertValue(agg, value, indices[1:])
-		global.Store(agg)
-		return
-	default:
-		panic("interp: store on a constant")
-	}
-}
-
-// GetElementPtr returns a GEP when the underlying value is of pointer type.
-func (v *LocalValue) GetElementPtr(indices []uint32) (Value, error) {
-	if !v.Underlying.IsAGlobalVariable().IsNil() {
-		int32Type := v.Underlying.Type().Context().Int32Type()
-		gep := llvm.ConstGEP(v.Underlying, getLLVMIndices(int32Type, indices))
-		return &LocalValue{v.Eval, gep}, nil
-	}
-	if !v.Underlying.IsAConstantExpr().IsNil() {
-		switch v.Underlying.Opcode() {
-		case llvm.GetElementPtr, llvm.IntToPtr, llvm.BitCast:
-			int32Type := v.Underlying.Type().Context().Int32Type()
-			llvmIndices := getLLVMIndices(int32Type, indices)
-			return &LocalValue{v.Eval, llvm.ConstGEP(v.Underlying, llvmIndices)}, nil
-		}
-	}
-	return nil, errors.New("interp: unknown GEP")
-}
-
-// stripPointerCasts removes all const bitcasts from pointer values, if there
-// are any.
-func (v *LocalValue) stripPointerCasts() *LocalValue {
-	value := v.Underlying
-	for {
-		if !value.IsAConstantExpr().IsNil() {
-			switch value.Opcode() {
-			case llvm.BitCast:
-				value = value.Operand(0)
-				continue
-			}
-		}
-		return &LocalValue{
-			Eval:       v.Eval,
-			Underlying: value,
-		}
-	}
-}
-
-func (v *LocalValue) String() string {
-	isConstant := "false"
-	if v.IsConstant() {
-		isConstant = "true"
-	}
-	return "&LocalValue{Type: " + v.Type().String() + ", IsConstant: " + isConstant + "}"
-}
-
-// getConstGEPIndices returns indices of this constant GEP, if this is a GEP
-// instruction. If it is not, the behavior is undefined.
-func (v *LocalValue) getConstGEPIndices() []uint32 {
-	indices := make([]uint32, v.Underlying.OperandsCount()-1)
-	for i := range indices {
-		operand := v.Underlying.Operand(i + 1)
-		indices[i] = uint32(operand.ZExtValue())
-	}
-	return indices
-}
-
-// MarkDirty marks this global as dirty, meaning that every load from and store
-// to this global (from now on) must be performed at runtime.
-func (v *LocalValue) MarkDirty() {
-	underlying := unwrap(v.Underlying)
-	if underlying.IsAGlobalVariable().IsNil() {
-		panic("trying to mark a non-global as dirty")
-	}
-	if !v.IsConstant() {
-		return // already dirty
-	}
-	v.Eval.dirtyGlobals[underlying] = struct{}{}
-}
-
-// MapValue implements a Go map which is created at compile time and stored as a
-// global variable.
-type MapValue struct {
-	Eval       *Eval
-	PkgName    string
-	Underlying llvm.Value
-	Keys       []Value
-	Values     []Value
-	KeySize    int
-	ValueSize  int
-	KeyType    llvm.Type
-	ValueType  llvm.Type
-}
-
-func (v *MapValue) newBucket() llvm.Value {
-	ctx := v.Eval.Mod.Context()
-	i8ptrType := llvm.PointerType(ctx.Int8Type(), 0)
-	bucketType := ctx.StructType([]llvm.Type{
-		llvm.ArrayType(ctx.Int8Type(), 8), // tophash
-		i8ptrType,                         // next bucket
-		llvm.ArrayType(v.KeyType, 8),      // key type
-		llvm.ArrayType(v.ValueType, 8),    // value type
-	}, false)
-	bucketValue := llvm.ConstNull(bucketType)
-	bucket := llvm.AddGlobal(v.Eval.Mod, bucketType, v.PkgName+"$mapbucket")
-	bucket.SetInitializer(bucketValue)
-	bucket.SetLinkage(llvm.InternalLinkage)
-	bucket.SetUnnamedAddr(true)
-	return bucket
-}
-
-// Value returns a global variable which is a pointer to the actual hashmap.
-func (v *MapValue) Value() llvm.Value {
-	if !v.Underlying.IsNil() {
-		return v.Underlying
-	}
-
-	ctx := v.Eval.Mod.Context()
-	i8ptrType := llvm.PointerType(ctx.Int8Type(), 0)
-
-	var firstBucketGlobal llvm.Value
-	if len(v.Keys) == 0 {
-		// there are no buckets
-		firstBucketGlobal = llvm.ConstPointerNull(i8ptrType)
-	} else {
-		// create initial bucket
-		firstBucketGlobal = v.newBucket()
-	}
-
-	// Insert each key/value pair in the hashmap.
-	bucketGlobal := firstBucketGlobal
-	for i, key := range v.Keys {
-		var keyBuf []byte
-		llvmKey := key.Value()
-		llvmValue := v.Values[i].Value()
-		if key.Type().TypeKind() == llvm.StructTypeKind && key.Type().StructName() == "runtime._string" {
-			keyPtr := llvm.ConstExtractValue(llvmKey, []uint32{0})
-			keyLen := llvm.ConstExtractValue(llvmKey, []uint32{1})
-			keyPtrVal := v.Eval.getValue(keyPtr)
-			keyBuf = getStringBytes(keyPtrVal, keyLen)
-		} else if key.Type().TypeKind() == llvm.IntegerTypeKind {
-			keyBuf = make([]byte, v.Eval.TargetData.TypeAllocSize(key.Type()))
-			n := key.Value().ZExtValue()
-			for i := range keyBuf {
-				keyBuf[i] = byte(n)
-				n >>= 8
-			}
-		} else if key.Type().TypeKind() == llvm.ArrayTypeKind &&
-			key.Type().ElementType().TypeKind() == llvm.IntegerTypeKind &&
-			key.Type().ElementType().IntTypeWidth() == 8 {
-			keyBuf = make([]byte, v.Eval.TargetData.TypeAllocSize(key.Type()))
-			for i := range keyBuf {
-				keyBuf[i] = byte(llvm.ConstExtractValue(llvmKey, []uint32{uint32(i)}).ZExtValue())
-			}
-		} else {
-			panic("interp: map key type not implemented: " + key.Type().String())
-		}
-		hash := v.hash(keyBuf)
-
-		if i%8 == 0 && i != 0 {
-			// Bucket is full, create a new one.
-			newBucketGlobal := v.newBucket()
-			zero := llvm.ConstInt(ctx.Int32Type(), 0, false)
-			newBucketPtr := llvm.ConstInBoundsGEP(newBucketGlobal, []llvm.Value{zero})
-			newBucketPtrCast := llvm.ConstBitCast(newBucketPtr, i8ptrType)
-			// insert pointer into old bucket
-			bucket := bucketGlobal.Initializer()
-			bucket = llvm.ConstInsertValue(bucket, newBucketPtrCast, []uint32{1})
-			bucketGlobal.SetInitializer(bucket)
-			// switch to next bucket
-			bucketGlobal = newBucketGlobal
-		}
-
-		tophashValue := llvm.ConstInt(ctx.Int8Type(), uint64(v.topHash(hash)), false)
-		bucket := bucketGlobal.Initializer()
-		bucket = llvm.ConstInsertValue(bucket, tophashValue, []uint32{0, uint32(i % 8)})
-		bucket = llvm.ConstInsertValue(bucket, llvmKey, []uint32{2, uint32(i % 8)})
-		bucket = llvm.ConstInsertValue(bucket, llvmValue, []uint32{3, uint32(i % 8)})
-		bucketGlobal.SetInitializer(bucket)
-	}
-
-	// Create the hashmap itself.
-	zero := llvm.ConstInt(ctx.Int32Type(), 0, false)
-	bucketPtr := llvm.ConstInBoundsGEP(firstBucketGlobal, []llvm.Value{zero})
-	hashmapType := v.Type()
-	hashmap := llvm.ConstNamedStruct(hashmapType, []llvm.Value{
-		llvm.ConstPointerNull(llvm.PointerType(hashmapType, 0)),                        // next
-		llvm.ConstBitCast(bucketPtr, i8ptrType),                                        // buckets
-		llvm.ConstInt(hashmapType.StructElementTypes()[2], uint64(len(v.Keys)), false), // count
-		llvm.ConstInt(ctx.Int8Type(), uint64(v.KeySize), false),                        // keySize
-		llvm.ConstInt(ctx.Int8Type(), uint64(v.ValueSize), false),                      // valueSize
-		llvm.ConstInt(ctx.Int8Type(), 0, false),                                        // bucketBits
-	})
-
-	// Create a pointer to this hashmap.
-	hashmapPtr := llvm.AddGlobal(v.Eval.Mod, hashmap.Type(), v.PkgName+"$map")
-	hashmapPtr.SetInitializer(hashmap)
-	hashmapPtr.SetLinkage(llvm.InternalLinkage)
-	hashmapPtr.SetUnnamedAddr(true)
-	v.Underlying = llvm.ConstInBoundsGEP(hashmapPtr, []llvm.Value{zero})
-	return v.Underlying
-}
-
-// Type returns type runtime.hashmap, which is the actual hashmap type.
-func (v *MapValue) Type() llvm.Type {
-	return v.Eval.Mod.GetTypeByName("runtime.hashmap")
-}
-
-func (v *MapValue) IsConstant() bool {
-	return true // TODO: dirty maps
-}
-
-// Load panics: maps are of reference type so cannot be dereferenced.
-func (v *MapValue) Load() llvm.Value {
-	panic("interp: load from a map")
-}
-
-// Store panics: maps are of reference type so cannot be stored to.
-func (v *MapValue) Store(value llvm.Value) {
-	panic("interp: store on a map")
-}
-
-// GetElementPtr panics: maps are of reference type so their (interior)
-// addresses cannot be calculated.
-func (v *MapValue) GetElementPtr(indices []uint32) (Value, error) {
-	return nil, errors.New("interp: GEP on a map")
-}
-
-// PutString does a map assign operation, assuming that the map is of type
-// map[string]T.
-func (v *MapValue) PutString(keyBuf, keyLen, valPtr *LocalValue) {
-	if !v.Underlying.IsNil() {
-		panic("map already created")
-	}
-
-	if valPtr.Underlying.Opcode() == llvm.BitCast {
-		valPtr = &LocalValue{v.Eval, valPtr.Underlying.Operand(0)}
-	}
-	value := valPtr.Load()
-	if v.ValueType.IsNil() {
-		v.ValueType = value.Type()
-		if int(v.Eval.TargetData.TypeAllocSize(v.ValueType)) != v.ValueSize {
-			panic("interp: map store value type has the wrong size")
-		}
-	} else {
-		if value.Type() != v.ValueType {
-			panic("interp: map store value type is inconsistent")
-		}
-	}
-
-	keyType := v.Eval.Mod.GetTypeByName("runtime._string")
-	v.KeyType = keyType
-	key := llvm.ConstNull(keyType)
-	key = llvm.ConstInsertValue(key, keyBuf.Value(), []uint32{0})
-	key = llvm.ConstInsertValue(key, keyLen.Value(), []uint32{1})
-
-	// TODO: avoid duplicate keys
-	v.Keys = append(v.Keys, &LocalValue{v.Eval, key})
-	v.Values = append(v.Values, &LocalValue{v.Eval, value})
-}
-
-// PutBinary does a map assign operation.
-func (v *MapValue) PutBinary(keyPtr, valPtr *LocalValue) {
-	if !v.Underlying.IsNil() {
-		panic("map already created")
-	}
-
-	if valPtr.Underlying.Opcode() == llvm.BitCast {
-		valPtr = &LocalValue{v.Eval, valPtr.Underlying.Operand(0)}
-	}
-	value := valPtr.Load()
-	if v.ValueType.IsNil() {
-		v.ValueType = value.Type()
-		if int(v.Eval.TargetData.TypeAllocSize(v.ValueType)) != v.ValueSize {
-			panic("interp: map store value type has the wrong size")
-		}
-	} else {
-		if value.Type() != v.ValueType {
-			panic("interp: map store value type is inconsistent")
-		}
-	}
-
-	if !keyPtr.Underlying.IsAConstantExpr().IsNil() {
-		if keyPtr.Underlying.Opcode() == llvm.BitCast {
-			keyPtr = &LocalValue{v.Eval, keyPtr.Underlying.Operand(0)}
-		} else if keyPtr.Underlying.Opcode() == llvm.GetElementPtr {
-			keyPtr = &LocalValue{v.Eval, keyPtr.Underlying.Operand(0)}
-		}
-	}
-	key := keyPtr.Load()
-	if v.KeyType.IsNil() {
-		v.KeyType = key.Type()
-		if int(v.Eval.TargetData.TypeAllocSize(v.KeyType)) != v.KeySize {
-			panic("interp: map store key type has the wrong size")
-		}
-	} else {
-		if key.Type() != v.KeyType {
-			panic("interp: map store key type is inconsistent")
-		}
-	}
-
-	// TODO: avoid duplicate keys
-	v.Keys = append(v.Keys, &LocalValue{v.Eval, key})
-	v.Values = append(v.Values, &LocalValue{v.Eval, value})
-}
-
-// Get FNV-1a hash of this string.
-//
-// https://en.wikipedia.org/wiki/Fowler%E2%80%93Noll%E2%80%93Vo_hash_function#FNV-1a_hash
-func (v *MapValue) hash(data []byte) uint32 {
-	var result uint32 = 2166136261 // FNV offset basis
-	for _, c := range data {
-		result ^= uint32(c)
-		result *= 16777619 // FNV prime
-	}
-	return result
-}
-
-// Get the topmost 8 bits of the hash, without using a special value (like 0).
-func (v *MapValue) topHash(hash uint32) uint8 {
-	tophash := uint8(hash >> 24)
-	if tophash < 1 {
-		// 0 means empty slot, so make it bigger.
-		tophash += 1
-	}
-	return tophash
-}
-
-func (v *MapValue) String() string {
-	return "&MapValue{KeySize: " + strconv.Itoa(v.KeySize) + ", ValueSize: " + strconv.Itoa(v.ValueSize) + "}"
-}
@@ -1,330 +0,0 @@
-package ir
-
-import (
-	"go/ast"
-	"go/types"
-	"sort"
-	"strings"
-
-	"github.com/tinygo-org/tinygo/loader"
-	"golang.org/x/tools/go/ssa"
-	"tinygo.org/x/go-llvm"
-)
-
-// This file provides a wrapper around go/ssa values and adds extra
-// functionality to them.
-
-// View on all functions, types, and globals in a program, with analysis
-// results.
-type Program struct {
-	Program       *ssa.Program
-	LoaderProgram *loader.Program
-	mainPkg       *ssa.Package
-	Functions     []*Function
-	functionMap   map[*ssa.Function]*Function
-}
-
-// Function or method.
-type Function struct {
-	*ssa.Function
-	LLVMFn   llvm.Value
-	module   string     // go:wasm-module
-	linkName string     // go:linkname, go:export
-	exported bool       // go:export
-	nobounds bool       // go:nobounds
-	flag     bool       // used by dead code elimination
-	inline   InlineType // go:inline
-}
-
-// Interface type that is at some point used in a type assert (to check whether
-// it implements another interface).
-type Interface struct {
-	Num  int
-	Type *types.Interface
-}
-
-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
-)
-
-// Create and initialize a new *Program from a *ssa.Program.
-func NewProgram(lprogram *loader.Program, mainPath string) *Program {
-	program := lprogram.LoadSSA()
-	program.Build()
-
-	// Find the main package, which is a bit difficult when running a .go file
-	// directly.
-	mainPkg := program.ImportedPackage(mainPath)
-	if mainPkg == nil {
-		for _, pkgInfo := range program.AllPackages() {
-			if pkgInfo.Pkg.Name() == "main" {
-				if mainPkg != nil {
-					panic("more than one main package found")
-				}
-				mainPkg = pkgInfo
-			}
-		}
-	}
-	if mainPkg == nil {
-		panic("could not find main package")
-	}
-
-	// Make a list of packages in import order.
-	packageList := []*ssa.Package{}
-	packageSet := map[string]struct{}{}
-	worklist := []string{"runtime", mainPath}
-	for len(worklist) != 0 {
-		pkgPath := worklist[0]
-		var pkg *ssa.Package
-		if pkgPath == mainPath {
-			pkg = mainPkg // necessary for compiling individual .go files
-		} else {
-			pkg = program.ImportedPackage(pkgPath)
-		}
-		if pkg == nil {
-			// Non-SSA package (e.g. cgo).
-			packageSet[pkgPath] = struct{}{}
-			worklist = worklist[1:]
-			continue
-		}
-		if _, ok := packageSet[pkgPath]; ok {
-			// Package already in the final package list.
-			worklist = worklist[1:]
-			continue
-		}
-
-		unsatisfiedImports := make([]string, 0)
-		imports := pkg.Pkg.Imports()
-		for _, pkg := range imports {
-			if _, ok := packageSet[pkg.Path()]; ok {
-				continue
-			}
-			unsatisfiedImports = append(unsatisfiedImports, pkg.Path())
-		}
-		if len(unsatisfiedImports) == 0 {
-			// All dependencies of this package are satisfied, so add this
-			// package to the list.
-			packageList = append(packageList, pkg)
-			packageSet[pkgPath] = struct{}{}
-			worklist = worklist[1:]
-		} else {
-			// Prepend all dependencies to the worklist and reconsider this
-			// package (by not removing it from the worklist). At that point, it
-			// must be possible to add it to packageList.
-			worklist = append(unsatisfiedImports, worklist...)
-		}
-	}
-
-	p := &Program{
-		Program:       program,
-		LoaderProgram: lprogram,
-		mainPkg:       mainPkg,
-		functionMap:   make(map[*ssa.Function]*Function),
-	}
-
-	for _, pkg := range packageList {
-		p.AddPackage(pkg)
-	}
-
-	return p
-}
-
-// Add a package to this Program. All packages need to be added first before any
-// analysis is done for correct results.
-func (p *Program) AddPackage(pkg *ssa.Package) {
-	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())
-		}
-	}
-}
-
-func (p *Program) addFunction(ssaFn *ssa.Function) {
-	if _, ok := p.functionMap[ssaFn]; ok {
-		return
-	}
-	f := &Function{Function: ssaFn}
-	f.parsePragmas()
-	p.Functions = append(p.Functions, f)
-	p.functionMap[ssaFn] = f
-
-	for _, anon := range ssaFn.AnonFuncs {
-		p.addFunction(anon)
-	}
-}
-
-// 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
-}
-
-func (p *Program) GetFunction(ssaFn *ssa.Function) *Function {
-	return p.functionMap[ssaFn]
-}
-
-func (p *Program) MainPkg() *ssa.Package {
-	return p.mainPkg
-}
-
-// Parse compiler directives in the preceding comments.
-func (f *Function) parsePragmas() {
-	if f.Syntax() == nil {
-		return
-	}
-	if decl, ok := f.Syntax().(*ast.FuncDecl); ok && decl.Doc != nil {
-		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
-				}
-				f.linkName = parts[1]
-				f.exported = true
-			case "//go:wasm-module":
-				// Alternative comment for setting the import module.
-				if len(parts) != 2 {
-					continue
-				}
-				f.module = parts[1]
-			case "//go:inline":
-				f.inline = InlineHint
-			case "//go:noinline":
-				f.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) {
-					f.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) {
-					f.nobounds = true
-				}
-			}
-		}
-	}
-}
-
-func (f *Function) IsNoBounds() bool {
-	return f.nobounds
-}
-
-// Return true iff this function is externally visible.
-func (f *Function) IsExported() bool {
-	return f.exported || f.CName() != ""
-}
-
-// Return the inline directive of this function.
-func (f *Function) Inline() InlineType {
-	return f.inline
-}
-
-// Return the module name if not the default.
-func (f *Function) Module() string {
-	return f.module
-}
-
-// Return the link name for this function.
-func (f *Function) LinkName() string {
-	if f.linkName != "" {
-		return f.linkName
-	}
-	if f.Signature.Recv() != nil {
-		// Method on a defined type (which may be a pointer).
-		return f.RelString(nil)
-	} else {
-		// Bare function.
-		if name := f.CName(); name != "" {
-			// Name CGo functions directly.
-			return name
-		} else {
-			return f.RelString(nil)
-		}
-	}
-}
-
-// Return the name of the C function if this is a CGo wrapper. Otherwise, return
-// a zero-length string.
-func (f *Function) CName() string {
-	name := f.Name()
-	if strings.HasPrefix(name, "_Cfunc_") {
-		// emitted by `go tool cgo`
-		return name[len("_Cfunc_"):]
-	}
-	if strings.HasPrefix(name, "C.") {
-		// created by ../loader/cgo.go
-		return name[2:]
-	}
-	return ""
-}
-
-// 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
-}
@@ -1,149 +0,0 @@
-package ir
-
-import (
-	"errors"
-	"go/types"
-
-	"golang.org/x/tools/go/ssa"
-)
-
-// This file implements several optimization passes (analysis + transform) to
-// optimize code in SSA form before it is compiled to LLVM IR. It is based on
-// the IR defined in ir.go.
-
-// Make 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
-}
-
-// Simple pass that removes dead code. This pass makes later analysis passes
-// more useful.
-func (p *Program) SimpleDCE() error {
-	// Unmark all functions.
-	for _, f := range p.Functions {
-		f.flag = false
-	}
-
-	// Initial set of live functions. Include main.main, *.init and runtime.*
-	// functions.
-	main, ok := p.mainPkg.Members["main"].(*ssa.Function)
-	if !ok {
-		if p.mainPkg.Members["main"] == nil {
-			return errors.New("function main is undeclared in the main package")
-		} else {
-			return errors.New("cannot declare main - must be func")
-		}
-	}
-	runtimePkg := p.Program.ImportedPackage("runtime")
-	mathPkg := p.Program.ImportedPackage("math")
-	taskPkg := p.Program.ImportedPackage("internal/task")
-	p.GetFunction(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(p.Program, instr.X.Type()) {
-						fn := p.Program.MethodValue(sel)
-						callee := p.GetFunction(fn)
-						if callee == nil {
-							// TODO: why is this necessary?
-							p.addFunction(fn)
-							callee = p.GetFunction(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.GetFunction(operand)
-						if f == nil {
-							// FIXME HACK: this function should have been
-							// discovered already. It is not for bound methods.
-							p.addFunction(operand)
-							f = p.GetFunction(operand)
-						}
-						if !f.flag {
-							f.flag = true
-							worklist = append(worklist, operand)
-						}
-					}
-				}
-			}
-		}
-	}
-
-	// Remove unmarked functions.
-	livefunctions := []*Function{}
-	for _, f := range p.Functions {
-		if f.flag {
-			livefunctions = append(livefunctions, f)
-		} else {
-			delete(p.functionMap, f.Function)
-		}
-	}
-	p.Functions = livefunctions
-
-	return nil
-}
@@ -1,9 +1,6 @@
 package loader

-import (
-	"go/token"
-	"strings"
-)
+import "go/scanner"

 // Errors contains a list of parser errors or a list of typechecker errors for
 // the given package.
@@ -16,24 +13,13 @@ func (e Errors) Error() string {
 	return "could not compile: " + e.Errs[0].Error()
 }

-// ImportCycleErrors is returned when encountering an import cycle. The list of
-// packages is a list from the root package to the leaf package that imports one
-// of the packages in the list.
-type ImportCycleError struct {
-	Packages        []string
-	ImportPositions []token.Position
+// Error is a regular error but with an added import stack. This is especially
+// useful for debugging import cycle errors.
+type Error struct {
+	ImportStack []string
+	Err         scanner.Error
 }

-func (e *ImportCycleError) Error() string {
-	var msg strings.Builder
-	msg.WriteString("import cycle:\n\t")
-	msg.WriteString(strings.Join(e.Packages, "\n\t"))
-	msg.WriteString("\n at ")
-	for i, pos := range e.ImportPositions {
-		if i > 0 {
-			msg.WriteString(", ")
-		}
-		msg.WriteString(pos.String())
-	}
-	return msg.String()
+func (e Error) Error() string {
+	return e.Err.Error()
 }
@@ -0,0 +1,281 @@
+package loader
+
+// This file constructs a new temporary GOROOT directory by merging both the
+// standard Go GOROOT and the GOROOT from TinyGo using symlinks.
+
+import (
+	"crypto/sha512"
+	"encoding/hex"
+	"errors"
+	"fmt"
+	"io"
+	"io/ioutil"
+	"os"
+	"os/exec"
+	"path"
+	"path/filepath"
+	"runtime"
+	"sync"
+
+	"github.com/tinygo-org/tinygo/compileopts"
+	"github.com/tinygo-org/tinygo/goenv"
+)
+
+var gorootCreateMutex sync.Mutex
+
+// GetCachedGoroot creates a new GOROOT by merging both the standard GOROOT and
+// the GOROOT from TinyGo using lots of symbolic links.
+func GetCachedGoroot(config *compileopts.Config) (string, error) {
+	goroot := goenv.Get("GOROOT")
+	if goroot == "" {
+		return "", errors.New("could not determine GOROOT")
+	}
+	tinygoroot := goenv.Get("TINYGOROOT")
+	if tinygoroot == "" {
+		return "", errors.New("could not determine TINYGOROOT")
+	}
+
+	// Determine the location of the cached GOROOT.
+	version, err := goenv.GorootVersionString(goroot)
+	if err != nil {
+		return "", err
+	}
+	// This hash is really a cache key, that contains (hopefully) enough
+	// information to make collisions unlikely during development.
+	// By including the Go version and TinyGo version, cache collisions should
+	// not happen outside of development.
+	hash := sha512.New512_256()
+	fmt.Fprintln(hash, goroot)
+	fmt.Fprintln(hash, version)
+	fmt.Fprintln(hash, goenv.Version)
+	fmt.Fprintln(hash, tinygoroot)
+	gorootsHash := hash.Sum(nil)
+	gorootsHashHex := hex.EncodeToString(gorootsHash[:])
+	cachedgorootName := "goroot-" + version + "-" + gorootsHashHex
+	cachedgoroot := filepath.Join(goenv.Get("GOCACHE"), cachedgorootName)
+	if needsSyscallPackage(config.BuildTags()) {
+		cachedgoroot += "-syscall"
+	}
+
+	// Do not try to create the cached GOROOT in parallel, that's only a waste
+	// of I/O bandwidth and thus speed. Instead, use a mutex to make sure only
+	// one goroutine does it at a time.
+	// This is not a way to ensure atomicity (a different TinyGo invocation
+	// could be creating the same directory), but instead a way to avoid
+	// creating it many times in parallel when running tests in parallel.
+	gorootCreateMutex.Lock()
+	defer gorootCreateMutex.Unlock()
+
+	if _, err := os.Stat(cachedgoroot); err == nil {
+		return cachedgoroot, nil
+	}
+	err = os.MkdirAll(goenv.Get("GOCACHE"), 0777)
+	if err != nil {
+		return "", err
+	}
+	tmpgoroot, err := ioutil.TempDir(goenv.Get("GOCACHE"), cachedgorootName+".tmp")
+	if err != nil {
+		return "", err
+	}
+
+	// Remove the temporary directory if it wasn't moved to the right place
+	// (for example, when there was an error).
+	defer os.RemoveAll(tmpgoroot)
+
+	for _, name := range []string{"bin", "lib", "pkg"} {
+		err = symlink(filepath.Join(goroot, name), filepath.Join(tmpgoroot, name))
+		if err != nil {
+			return "", err
+		}
+	}
+	err = mergeDirectory(goroot, tinygoroot, tmpgoroot, "", pathsToOverride(needsSyscallPackage(config.BuildTags())))
+	if err != nil {
+		return "", err
+	}
+	err = os.Rename(tmpgoroot, cachedgoroot)
+	if err != nil {
+		if os.IsExist(err) {
+			// Another invocation of TinyGo also seems to have created a GOROOT.
+			// Use that one instead. Our new GOROOT will be automatically
+			// deleted by the defer above.
+			return cachedgoroot, nil
+		}
+		if runtime.GOOS == "windows" && os.IsPermission(err) {
+			// On Windows, a rename with a destination directory that already
+			// exists does not result in an IsExist error, but rather in an
+			// access denied error. To be sure, check for this case by checking
+			// whether the target directory exists.
+			if _, err := os.Stat(cachedgoroot); err == nil {
+				return cachedgoroot, nil
+			}
+		}
+		return "", err
+	}
+	return cachedgoroot, nil
+}
+
+// mergeDirectory merges two roots recursively. The tmpgoroot is the directory
+// that will be created by this call by either symlinking the directory from
+// goroot or tinygoroot, or by creating the directory and merging the contents.
+func mergeDirectory(goroot, tinygoroot, tmpgoroot, importPath string, overrides map[string]bool) error {
+	if mergeSubdirs, ok := overrides[importPath+"/"]; ok {
+		if !mergeSubdirs {
+			// This directory and all subdirectories should come from the TinyGo
+			// root, so simply make a symlink.
+			newname := filepath.Join(tmpgoroot, "src", importPath)
+			oldname := filepath.Join(tinygoroot, "src", importPath)
+			return symlink(oldname, newname)
+		}
+
+		// Merge subdirectories. Start by making the directory to merge.
+		err := os.Mkdir(filepath.Join(tmpgoroot, "src", importPath), 0777)
+		if err != nil {
+			return err
+		}
+
+		// Symlink all files from TinyGo, and symlink directories from TinyGo
+		// that need to be overridden.
+		tinygoEntries, err := ioutil.ReadDir(filepath.Join(tinygoroot, "src", importPath))
+		if err != nil {
+			return err
+		}
+		for _, e := range tinygoEntries {
+			if e.IsDir() {
+				// A directory, so merge this thing.
+				err := mergeDirectory(goroot, tinygoroot, tmpgoroot, path.Join(importPath, e.Name()), overrides)
+				if err != nil {
+					return err
+				}
+			} else {
+				// A file, so symlink this.
+				newname := filepath.Join(tmpgoroot, "src", importPath, e.Name())
+				oldname := filepath.Join(tinygoroot, "src", importPath, e.Name())
+				err := symlink(oldname, newname)
+				if err != nil {
+					return err
+				}
+			}
+		}
+
+		// Symlink all directories from $GOROOT that are not part of the TinyGo
+		// overrides.
+		gorootEntries, err := ioutil.ReadDir(filepath.Join(goroot, "src", importPath))
+		if err != nil {
+			return err
+		}
+		for _, e := range gorootEntries {
+			if !e.IsDir() {
+				// Don't merge in files from Go. Otherwise we'd end up with a
+				// weird syscall package with files from both roots.
+				continue
+			}
+			if _, ok := overrides[path.Join(importPath, e.Name())+"/"]; ok {
+				// Already included above, so don't bother trying to create this
+				// symlink.
+				continue
+			}
+			newname := filepath.Join(tmpgoroot, "src", importPath, e.Name())
+			oldname := filepath.Join(goroot, "src", importPath, e.Name())
+			err := symlink(oldname, newname)
+			if err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+// needsSyscallPackage returns whether the syscall package should be overriden
+// with the TinyGo version. This is the case on some targets.
+func needsSyscallPackage(buildTags []string) bool {
+	for _, tag := range buildTags {
+		if tag == "baremetal" || tag == "darwin" || tag == "nintendoswitch" || tag == "wasi" {
+			return true
+		}
+	}
+	return false
+}
+
+// The boolean indicates whether to merge the subdirs. True means merge, false
+// means use the TinyGo version.
+func pathsToOverride(needsSyscallPackage bool) map[string]bool {
+	paths := map[string]bool{
+		"/":                     true,
+		"device/":               false,
+		"examples/":             false,
+		"internal/":             true,
+		"internal/bytealg/":     false,
+		"internal/reflectlite/": false,
+		"internal/task/":        false,
+		"machine/":              false,
+		"os/":                   true,
+		"reflect/":              false,
+		"runtime/":              false,
+		"sync/":                 true,
+		"testing/":              true,
+	}
+	if needsSyscallPackage {
+		paths["syscall/"] = true // include syscall/js
+	}
+	return paths
+}
+
+// symlink creates a symlink or something similar. On Unix-like systems, it
+// always creates a symlink. On Windows, it tries to create a symlink and if
+// that fails, creates a hardlink or directory junction instead.
+//
+// Note that while Windows 10 does support symlinks and allows them to be
+// created using os.Symlink, it requires developer mode to be enabled.
+// Therefore provide a fallback for when symlinking is not possible.
+// Unfortunately this fallback only works when TinyGo is installed on the same
+// filesystem as the TinyGo cache and the Go installation (which is usually the
+// C drive).
+func symlink(oldname, newname string) error {
+	symlinkErr := os.Symlink(oldname, newname)
+	if runtime.GOOS == "windows" && symlinkErr != nil {
+		// Fallback for when developer mode is disabled.
+		// Note that we return the symlink error even if something else fails
+		// later on. This is because symlinks are the easiest to support
+		// (they're also used on Linux and MacOS) and enabling them is easy:
+		// just enable developer mode.
+		st, err := os.Stat(oldname)
+		if err != nil {
+			return symlinkErr
+		}
+		if st.IsDir() {
+			// Make a directory junction. There may be a way to do this
+			// programmatically, but it involves a lot of magic. Use the mklink
+			// command built into cmd instead (mklink is a builtin, not an
+			// external command).
+			err := exec.Command("cmd", "/k", "mklink", "/J", newname, oldname).Run()
+			if err != nil {
+				return symlinkErr
+			}
+		} else {
+			// Try making a hard link.
+			err := os.Link(oldname, newname)
+			if err != nil {
+				// Making a hardlink failed. Try copying the file as a last
+				// fallback.
+				inf, err := os.Open(oldname)
+				if err != nil {
+					return err
+				}
+				defer inf.Close()
+				outf, err := os.Create(newname)
+				if err != nil {
+					return err
+				}
+				defer outf.Close()
+				_, err = io.Copy(outf, inf)
+				if err != nil {
+					os.Remove(newname)
+					return err
+				}
+				// File was copied.
+			}
+		}
+		return nil // success
+	}
+	return symlinkErr
+}
@@ -0,0 +1,30 @@
+package loader
+
+import (
+	"os"
+	"os/exec"
+	"strings"
+
+	"github.com/tinygo-org/tinygo/compileopts"
+)
+
+// List returns a ready-to-run *exec.Cmd for running the `go list` command with
+// the configuration used for TinyGo.
+func List(config *compileopts.Config, extraArgs, pkgs []string) (*exec.Cmd, error) {
+	goroot, err := GetCachedGoroot(config)
+	if err != nil {
+		return nil, err
+	}
+	args := append([]string{"list"}, extraArgs...)
+	if len(config.BuildTags()) != 0 {
+		args = append(args, "-tags", strings.Join(config.BuildTags(), " "))
+	}
+	args = append(args, pkgs...)
+	cgoEnabled := "0"
+	if config.CgoEnabled() {
+		cgoEnabled = "1"
+	}
+	cmd := exec.Command("go", args...)
+	cmd.Env = append(os.Environ(), "GOROOT="+goroot, "GOOS="+config.GOOS(), "GOARCH="+config.GOARCH(), "CGO_ENABLED="+cgoEnabled)
+	return cmd, nil
+}
@@ -2,226 +2,276 @@ package loader

 import (
 	"bytes"
+	"crypto/sha512"
+	"encoding/json"
 	"errors"
+	"fmt"
 	"go/ast"
-	"go/build"
 	"go/parser"
 	"go/scanner"
 	"go/token"
 	"go/types"
+	"io"
+	"io/ioutil"
 	"os"
+	"os/exec"
 	"path/filepath"
-	"sort"
+	"strconv"
 	"strings"
-	"text/template"

 	"github.com/tinygo-org/tinygo/cgo"
+	"github.com/tinygo-org/tinygo/compileopts"
+	"github.com/tinygo-org/tinygo/goenv"
 )

 // Program holds all packages and some metadata about the program as a whole.
 type Program struct {
-	mainPkg      string
-	Build        *build.Context
-	OverlayBuild *build.Context
-	OverlayPath  func(path string) string
-	Packages     map[string]*Package
-	sorted       []*Package
-	fset         *token.FileSet
-	TypeChecker  types.Config
-	Dir          string // current working directory (for error reporting)
-	TINYGOROOT   string // root of the TinyGo installation or root of the source code
-	CFlags       []string
-	ClangHeaders string
+	config       *compileopts.Config
+	clangHeaders string
+	typeChecker  types.Config
+	goroot       string // synthetic GOROOT
+	workingDir   string
+
+	Packages map[string]*Package
+	sorted   []*Package
+	fset     *token.FileSet
+
+	// Information obtained during parsing.
+	LDFlags []string
+}
+
+// PackageJSON is a subset of the JSON struct returned from `go list`.
+type PackageJSON struct {
+	Dir        string
+	ImportPath string
+	Name       string
+	ForTest    string
+
+	// Source files
+	GoFiles  []string
+	CgoFiles []string
+	CFiles   []string
+
+	// Dependency information
+	Imports   []string
+	ImportMap map[string]string
+
+	// Error information
+	Error *struct {
+		ImportStack []string
+		Pos         string
+		Err         string
+	}
 }

 // Package holds a loaded package, its imports, and its parsed files.
 type Package struct {
-	*Program
-	*build.Package
-	Imports   map[string]*Package
-	Importing bool
-	Files     []*ast.File
-	Pkg       *types.Package
-	types.Info
+	PackageJSON
+
+	program    *Program
+	Files      []*ast.File
+	FileHashes map[string][]byte
+	CFlags     []string // CFlags used during CGo preprocessing (only set if CGo is used)
+	Pkg        *types.Package
+	info       types.Info
 }

-// Import loads the given package relative to srcDir (for the vendor directory).
-// It only loads the current package without recursion.
-func (p *Program) Import(path, srcDir string, pos token.Position) (*Package, error) {
-	if p.Packages == nil {
-		p.Packages = make(map[string]*Package)
-	}
-
-	// Load this package.
-	ctx := p.Build
-	if newPath := p.OverlayPath(path); newPath != "" {
-		ctx = p.OverlayBuild
-		path = newPath
-	}
-	buildPkg, err := ctx.Import(path, srcDir, build.ImportComment)
-	if err != nil {
-		return nil, scanner.Error{
-			Pos: pos,
-			Msg: err.Error(), // TODO: define a new error type that will wrap the inner error
-		}
-	}
-	if existingPkg, ok := p.Packages[buildPkg.ImportPath]; ok {
-		// Already imported, or at least started the import.
-		return existingPkg, nil
-	}
-	p.sorted = nil // invalidate the sorted order of packages
-	pkg := p.newPackage(buildPkg)
-	p.Packages[buildPkg.ImportPath] = pkg
-
-	if p.mainPkg == "" {
-		p.mainPkg = buildPkg.ImportPath
-	}
-
-	return pkg, nil
-}
-
-// ImportFile loads and parses the import statements in the given path and
-// creates a pseudo-package out of it.
-func (p *Program) ImportFile(path string) (*Package, error) {
-	if p.Packages == nil {
-		p.Packages = make(map[string]*Package)
-	}
-	if _, ok := p.Packages[path]; ok {
-		// unlikely
-		return nil, errors.New("loader: cannot import file that is already imported as package: " + path)
-	}
-
-	file, err := p.parseFile(path, parser.ImportsOnly)
+// Load loads the given package with all dependencies (including the runtime
+// package). Call .Parse() afterwards to parse all Go files (including CGo
+// processing, if necessary).
+func Load(config *compileopts.Config, inputPkgs []string, clangHeaders string, typeChecker types.Config) (*Program, error) {
+	goroot, err := GetCachedGoroot(config)
 	if err != nil {
 		return nil, err
 	}
-	buildPkg := &build.Package{
-		Dir:        filepath.Dir(path),
-		ImportPath: path,
-		GoFiles:    []string{filepath.Base(path)},
+	wd, err := os.Getwd()
+	if err != nil {
+		return nil, err
 	}
-	for _, importSpec := range file.Imports {
-		buildPkg.Imports = append(buildPkg.Imports, importSpec.Path.Value[1:len(importSpec.Path.Value)-1])
-	}
-	p.sorted = nil // invalidate the sorted order of packages
-	pkg := p.newPackage(buildPkg)
-	p.Packages[buildPkg.ImportPath] = pkg
-
-	if p.mainPkg == "" {
-		p.mainPkg = buildPkg.ImportPath
+	p := &Program{
+		config:       config,
+		clangHeaders: clangHeaders,
+		typeChecker:  typeChecker,
+		goroot:       goroot,
+		workingDir:   wd,
+		Packages:     make(map[string]*Package),
+		fset:         token.NewFileSet(),
 	}

-	return pkg, nil
+	// List the dependencies of this package, in raw JSON format.
+	extraArgs := []string{"-json", "-deps"}
+	if config.TestConfig.CompileTestBinary {
+		extraArgs = append(extraArgs, "-test")
+	}
+	cmd, err := List(config, extraArgs, inputPkgs)
+	if err != nil {
+		return nil, err
+	}
+	buf := &bytes.Buffer{}
+	cmd.Stdout = buf
+	cmd.Stderr = os.Stderr
+	err = cmd.Run()
+	if err != nil {
+		if exitErr, ok := err.(*exec.ExitError); ok {
+			os.Exit(exitErr.ExitCode())
+		}
+		return nil, fmt.Errorf("failed to run `go list`: %s", err)
+	}
+
+	// Parse the returned json from `go list`.
+	decoder := json.NewDecoder(buf)
+	for {
+		pkg := &Package{
+			program:    p,
+			FileHashes: make(map[string][]byte),
+			info: types.Info{
+				Types:      make(map[ast.Expr]types.TypeAndValue),
+				Defs:       make(map[*ast.Ident]types.Object),
+				Uses:       make(map[*ast.Ident]types.Object),
+				Implicits:  make(map[ast.Node]types.Object),
+				Scopes:     make(map[ast.Node]*types.Scope),
+				Selections: make(map[*ast.SelectorExpr]*types.Selection),
+			},
+		}
+		err := decoder.Decode(&pkg.PackageJSON)
+		if err != nil {
+			if err == io.EOF {
+				break
+			}
+			return nil, err
+		}
+		if pkg.Error != nil {
+			// There was an error while importing (for example, a circular
+			// dependency).
+			pos := token.Position{}
+			fields := strings.Split(pkg.Error.Pos, ":")
+			if len(fields) >= 2 {
+				// There is some file/line/column information.
+				if n, err := strconv.Atoi(fields[len(fields)-2]); err == nil {
+					// Format: filename.go:line:colum
+					pos.Filename = strings.Join(fields[:len(fields)-2], ":")
+					pos.Line = n
+					pos.Column, _ = strconv.Atoi(fields[len(fields)-1])
+				} else {
+					// Format: filename.go:line
+					pos.Filename = strings.Join(fields[:len(fields)-1], ":")
+					pos.Line, _ = strconv.Atoi(fields[len(fields)-1])
+				}
+				pos.Filename = p.getOriginalPath(pos.Filename)
+			}
+			err := scanner.Error{
+				Pos: pos,
+				Msg: pkg.Error.Err,
+			}
+			if len(pkg.Error.ImportStack) != 0 {
+				return nil, Error{
+					ImportStack: pkg.Error.ImportStack,
+					Err:         err,
+				}
+			}
+			return nil, err
+		}
+		if config.TestConfig.CompileTestBinary {
+			// When creating a test binary, `go list` will list two or three
+			// packages used for testing the package. The first is the original
+			// package as if it were built normally, the second is the same
+			// package but with the *_test.go files included. A possible third
+			// may be included for _test packages (such as math_test), used to
+			// test the external API with no access to internal functions.
+			// All packages that are necessary for testing (including the to be
+			// tested package with *_test.go files, but excluding the original
+			// unmodified package) have a suffix added to the import path, for
+			// example the math package has import path "math [math.test]" and
+			// test dependencies such as fmt will have an import path of the
+			// form "fmt [math.test]".
+			// The code below removes this suffix, and if this results in a
+			// duplicate (which happens with the to-be-tested package without
+			// *.test.go files) the previous package is removed from the list of
+			// packages included in this build.
+			// This is necessary because the change in import paths results in
+			// breakage to //go:linkname. Additionally, the duplicated package
+			// slows down the build and so is best removed.
+			if pkg.ForTest != "" && strings.HasSuffix(pkg.ImportPath, " ["+pkg.ForTest+".test]") {
+				newImportPath := pkg.ImportPath[:len(pkg.ImportPath)-len(" ["+pkg.ForTest+".test]")]
+				if _, ok := p.Packages[newImportPath]; ok {
+					// Delete the previous package (that this package overrides).
+					delete(p.Packages, newImportPath)
+					for i, pkg := range p.sorted {
+						if pkg.ImportPath == newImportPath {
+							p.sorted = append(p.sorted[:i], p.sorted[i+1:]...) // remove element from slice
+							break
+						}
+					}
+				}
+				pkg.ImportPath = newImportPath
+			}
+		}
+		p.sorted = append(p.sorted, pkg)
+		p.Packages[pkg.ImportPath] = pkg
+	}
+
+	return p, nil
 }

-// newPackage instantiates a new *Package object with initialized members.
-func (p *Program) newPackage(pkg *build.Package) *Package {
-	return &Package{
-		Program: p,
-		Package: pkg,
-		Imports: make(map[string]*Package, len(pkg.Imports)),
-		Info: types.Info{
-			Types:      make(map[ast.Expr]types.TypeAndValue),
-			Defs:       make(map[*ast.Ident]types.Object),
-			Uses:       make(map[*ast.Ident]types.Object),
-			Implicits:  make(map[ast.Node]types.Object),
-			Scopes:     make(map[ast.Node]*types.Scope),
-			Selections: make(map[*ast.SelectorExpr]*types.Selection),
-		},
+// getOriginalPath looks whether this path is in the generated GOROOT and if so,
+// replaces the path with the original path (in GOROOT or TINYGOROOT). Otherwise
+// the input path is returned.
+func (p *Program) getOriginalPath(path string) string {
+	originalPath := path
+	if strings.HasPrefix(path, p.goroot+string(filepath.Separator)) {
+		// If this file is part of the synthetic GOROOT, try to infer the
+		// original path.
+		relpath := path[len(filepath.Join(p.goroot, "src"))+1:]
+		realgorootPath := filepath.Join(goenv.Get("GOROOT"), "src", relpath)
+		if _, err := os.Stat(realgorootPath); err == nil {
+			originalPath = realgorootPath
+		}
+		maybeInTinyGoRoot := false
+		for prefix := range pathsToOverride(needsSyscallPackage(p.config.BuildTags())) {
+			if !strings.HasPrefix(relpath, prefix) {
+				continue
+			}
+			maybeInTinyGoRoot = true
+		}
+		if maybeInTinyGoRoot {
+			tinygoPath := filepath.Join(goenv.Get("TINYGOROOT"), "src", relpath)
+			if _, err := os.Stat(tinygoPath); err == nil {
+				originalPath = tinygoPath
+			}
+		}
 	}
+	return originalPath
 }

 // Sorted returns a list of all packages, sorted in a way that no packages come
 // before the packages they depend upon.
 func (p *Program) Sorted() []*Package {
-	if p.sorted == nil {
-		p.sort()
-	}
 	return p.sorted
 }

-func (p *Program) sort() {
-	p.sorted = nil
-	packageList := make([]*Package, 0, len(p.Packages))
-	packageSet := make(map[string]struct{}, len(p.Packages))
-	worklist := make([]string, 0, len(p.Packages))
-	for path := range p.Packages {
-		worklist = append(worklist, path)
-	}
-	sort.Strings(worklist)
-	for len(worklist) != 0 {
-		pkgPath := worklist[0]
-		pkg := p.Packages[pkgPath]
-
-		if _, ok := packageSet[pkgPath]; ok {
-			// Package already in the final package list.
-			worklist = worklist[1:]
-			continue
-		}
-
-		unsatisfiedImports := make([]string, 0)
-		for _, pkg := range pkg.Imports {
-			if _, ok := packageSet[pkg.ImportPath]; ok {
-				continue
-			}
-			unsatisfiedImports = append(unsatisfiedImports, pkg.ImportPath)
-		}
-		sort.Strings(unsatisfiedImports)
-		if len(unsatisfiedImports) == 0 {
-			// All dependencies of this package are satisfied, so add this
-			// package to the list.
-			packageList = append(packageList, pkg)
-			packageSet[pkgPath] = struct{}{}
-			worklist = worklist[1:]
-		} else {
-			// Prepend all dependencies to the worklist and reconsider this
-			// package (by not removing it from the worklist). At that point, it
-			// must be possible to add it to packageList.
-			worklist = append(unsatisfiedImports, worklist...)
-		}
-	}
-
-	p.sorted = packageList
+// MainPkg returns the last package in the Sorted() slice. This is the main
+// package of the program.
+func (p *Program) MainPkg() *Package {
+	return p.sorted[len(p.sorted)-1]
 }

-// Parse recursively imports all packages, parses them, and typechecks them.
+// Parse parses all packages and typechecks them.
 //
 // The returned error may be an Errors error, which contains a list of errors.
 //
 // Idempotent.
-func (p *Program) Parse(compileTestBinary bool) error {
-	includeTests := compileTestBinary
-
-	// Load all imports
-	for _, pkg := range p.Sorted() {
-		err := pkg.importRecursively(includeTests)
-		if err != nil {
-			if err, ok := err.(*ImportCycleError); ok {
-				if pkg.ImportPath != err.Packages[0] {
-					err.Packages = append([]string{pkg.ImportPath}, err.Packages...)
-				}
-			}
-			return err
-		}
-	}
-
+func (p *Program) Parse() error {
 	// Parse all packages.
-	for _, pkg := range p.Sorted() {
-		err := pkg.Parse(includeTests)
-		if err != nil {
-			return err
-		}
-	}
-
-	if compileTestBinary {
-		err := p.SwapTestMain()
+	// TODO: do this in parallel.
+	for _, pkg := range p.sorted {
+		err := pkg.Parse()
 		if err != nil {
 			return err
 		}
 	}

 	// Typecheck all packages.
-	for _, pkg := range p.Sorted() {
+	for _, pkg := range p.sorted {
 		err := pkg.Check()
 		if err != nil {
 			return err
@@ -231,122 +281,35 @@ func (p *Program) Parse(compileTestBinary bool) error {
 	return nil
 }

-func (p *Program) SwapTestMain() error {
-	var tests []string
-
-	isTestFunc := func(f *ast.FuncDecl) bool {
-		// TODO: improve signature check
-		if strings.HasPrefix(f.Name.Name, "Test") && f.Name.Name != "TestMain" {
-			return true
-		}
-		return false
-	}
-	mainPkg := p.Packages[p.mainPkg]
-	for _, f := range mainPkg.Files {
-		for i, d := range f.Decls {
-			switch v := d.(type) {
-			case *ast.FuncDecl:
-				if isTestFunc(v) {
-					tests = append(tests, v.Name.Name)
-				}
-				if v.Name.Name == "main" {
-					// Remove main
-					if len(f.Decls) == 1 {
-						f.Decls = make([]ast.Decl, 0)
-					} else {
-						f.Decls[i] = f.Decls[len(f.Decls)-1]
-						f.Decls = f.Decls[:len(f.Decls)-1]
-					}
-				}
-			}
-		}
-	}
-
-	// TODO: Check if they defined a TestMain and call it instead of testing.TestMain
-	const mainBody = `package main
-
-import (
-	"testing"
-)
-
-func main () {
-	m := &testing.M{
-		Tests: []testing.TestToCall{
-{{range .TestFunctions}}
-			{Name: "{{.}}", Func: {{.}}},
-{{end}}
-		},
-	}
-
-	testing.TestMain(m)
-}
-`
-	tmpl := template.Must(template.New("testmain").Parse(mainBody))
-	b := bytes.Buffer{}
-	tmplData := struct {
-		TestFunctions []string
-	}{
-		TestFunctions: tests,
-	}
-
-	err := tmpl.Execute(&b, tmplData)
-	if err != nil {
-		return err
-	}
-	path := filepath.Join(p.mainPkg, "$testmain.go")
-
-	if p.fset == nil {
-		p.fset = token.NewFileSet()
-	}
-
-	newMain, err := parser.ParseFile(p.fset, path, b.Bytes(), parser.AllErrors)
-	if err != nil {
-		return err
-	}
-	mainPkg.Files = append(mainPkg.Files, newMain)
-
-	return nil
-}
-
 // parseFile is a wrapper around parser.ParseFile.
-func (p *Program) parseFile(path string, mode parser.Mode) (*ast.File, error) {
-	if p.fset == nil {
-		p.fset = token.NewFileSet()
-	}
-
-	rd, err := os.Open(path)
+func (p *Package) parseFile(path string, mode parser.Mode) (*ast.File, error) {
+	originalPath := p.program.getOriginalPath(path)
+	data, err := ioutil.ReadFile(path)
 	if err != nil {
 		return nil, err
 	}
-	defer rd.Close()
-	relpath := path
-	if filepath.IsAbs(path) {
-		rp, err := filepath.Rel(p.Dir, path)
-		if err == nil {
-			relpath = rp
-		}
-	}
-	return parser.ParseFile(p.fset, relpath, rd, mode)
+	sum := sha512.Sum512_224(data)
+	p.FileHashes[originalPath] = sum[:]
+	return parser.ParseFile(p.program.fset, originalPath, data, mode)
 }

 // Parse parses and typechecks this package.
 //
 // Idempotent.
-func (p *Package) Parse(includeTests bool) error {
+func (p *Package) Parse() error {
 	if len(p.Files) != 0 {
-		return nil
+		return nil // nothing to do (?)
 	}

 	// Load the AST.
-	// TODO: do this in parallel.
 	if p.ImportPath == "unsafe" {
-		// Special case for the unsafe package. Don't even bother loading
-		// the files.
+		// Special case for the unsafe package, which is defined internally by
+		// the types package.
 		p.Pkg = types.Unsafe
 		return nil
 	}

-	files, err := p.parseFiles(includeTests)
+	files, err := p.parseFiles()
 	if err != nil {
 		return err
 	}
@@ -361,11 +324,11 @@ func (p *Package) Parse(includeTests bool) error {
 // Idempotent.
 func (p *Package) Check() error {
 	if p.Pkg != nil {
-		return nil
+		return nil // already typechecked
 	}

 	var typeErrors []error
-	checker := p.TypeChecker
+	checker := p.program.typeChecker // make a copy, because it will be modified
 	checker.Error = func(err error) {
 		typeErrors = append(typeErrors, err)
 	}
@@ -373,7 +336,16 @@ func (p *Package) Check() error {
 	// Do typechecking of the package.
 	checker.Importer = p

-	typesPkg, err := checker.Check(p.ImportPath, p.fset, p.Files, &p.Info)
+	packageName := p.ImportPath
+	if p.Name == "main" {
+		// The main package normally has a different import path, such as
+		// "command-line-arguments" or "./testdata/cgo". Therefore, use the name
+		// "main" in such a case: this package isn't imported from anywhere.
+		// This is safe as it isn't possible to import a package with the name
+		// "main".
+		packageName = "main"
+	}
+	typesPkg, err := checker.Check(packageName, p.program.fset, p.Files, &p.info)
 	if err != nil {
 		if err, ok := err.(Errors); ok {
 			return err
@@ -385,52 +357,51 @@ func (p *Package) Check() error {
 }

 // parseFiles parses the loaded list of files and returns this list.
-func (p *Package) parseFiles(includeTests bool) ([]*ast.File, error) {
-	// TODO: do this concurrently.
+func (p *Package) parseFiles() ([]*ast.File, error) {
 	var files []*ast.File
 	var fileErrs []error

-	var gofiles []string
-	if includeTests {
-		gofiles = make([]string, 0, len(p.GoFiles)+len(p.TestGoFiles))
-		gofiles = append(gofiles, p.GoFiles...)
-		gofiles = append(gofiles, p.TestGoFiles...)
-	} else {
-		gofiles = p.GoFiles
-	}
-
-	for _, file := range gofiles {
-		f, err := p.parseFile(filepath.Join(p.Package.Dir, file), parser.ParseComments)
-		if err != nil {
-			fileErrs = append(fileErrs, err)
-			continue
+	// Parse all files (incuding CgoFiles).
+	parseFile := func(file string) {
+		if !filepath.IsAbs(file) {
+			file = filepath.Join(p.Dir, file)
 		}
+		f, err := p.parseFile(file, parser.ParseComments)
 		if err != nil {
 			fileErrs = append(fileErrs, err)
-			continue
+			return
 		}
 		files = append(files, f)
 	}
+	for _, file := range p.GoFiles {
+		parseFile(file)
+	}
 	for _, file := range p.CgoFiles {
-		path := filepath.Join(p.Package.Dir, file)
-		f, err := p.parseFile(path, parser.ParseComments)
-		if err != nil {
-			fileErrs = append(fileErrs, err)
-			continue
-		}
-		files = append(files, f)
+		parseFile(file)
 	}
+
+	// Do CGo processing.
 	if len(p.CgoFiles) != 0 {
-		cflags := append(p.CFlags, "-I"+p.Package.Dir)
-		if p.ClangHeaders != "" {
-			cflags = append(cflags, "-Xclang", "-internal-isystem", "-Xclang", p.ClangHeaders)
+		var cflags []string
+		cflags = append(cflags, p.program.config.CFlags()...)
+		cflags = append(cflags, "-I"+p.Dir)
+		if p.program.clangHeaders != "" {
+			cflags = append(cflags, "-Xclang", "-internal-isystem", "-Xclang", p.program.clangHeaders)
+		}
+		p.CFlags = cflags
+		generated, ldflags, accessedFiles, errs := cgo.Process(files, p.program.workingDir, p.program.fset, cflags)
+		for path, hash := range accessedFiles {
+			p.FileHashes[path] = hash
 		}
-		generated, errs := cgo.Process(files, p.Program.Dir, p.fset, cflags)
 		if errs != nil {
 			fileErrs = append(fileErrs, errs...)
 		}
 		files = append(files, generated)
+		p.program.LDFlags = append(p.program.LDFlags, ldflags...)
 	}
+
+	// Only return an error after CGo processing, so that errors in parsing and
+	// CGo can be reported together.
 	if len(fileErrs) != 0 {
 		return nil, Errors{p, fileErrs}
 	}
@@ -444,59 +415,12 @@ func (p *Package) Import(to string) (*types.Package, error) {
 	if to == "unsafe" {
 		return types.Unsafe, nil
 	}
-	if _, ok := p.Imports[to]; ok {
-		return p.Imports[to].Pkg, nil
+	if newTo, ok := p.ImportMap[to]; ok && !strings.HasSuffix(newTo, ".test]") {
+		to = newTo
+	}
+	if imported, ok := p.program.Packages[to]; ok {
+		return imported.Pkg, nil
 	} else {
 		return nil, errors.New("package not imported: " + to)
 	}
 }
-
-// importRecursively calls Program.Import() on all imported packages, and calls
-// importRecursively() on the imported packages as well.
-//
-// Idempotent.
-func (p *Package) importRecursively(includeTests bool) error {
-	p.Importing = true
-
-	imports := p.Package.Imports
-	if includeTests {
-		imports = append(imports, p.Package.TestImports...)
-	}
-
-	for _, to := range imports {
-		if to == "C" {
-			// Do CGo processing in a later stage.
-			continue
-		}
-		if _, ok := p.Imports[to]; ok {
-			continue
-		}
-		// Find error location.
-		var pos token.Position
-		if len(p.Package.ImportPos[to]) > 0 {
-			pos = p.Package.ImportPos[to][0]
-		} else {
-			pos = token.Position{Filename: p.Package.ImportPath}
-		}
-		importedPkg, err := p.Program.Import(to, p.Package.Dir, pos)
-		if err != nil {
-			if err, ok := err.(*ImportCycleError); ok {
-				err.Packages = append([]string{p.ImportPath}, err.Packages...)
-			}
-			return err
-		}
-		if importedPkg.Importing {
-			return &ImportCycleError{[]string{p.ImportPath, importedPkg.ImportPath}, p.ImportPos[to]}
-		}
-		err = importedPkg.importRecursively(false)
-		if err != nil {
-			if err, ok := err.(*ImportCycleError); ok {
-				err.Packages = append([]string{p.ImportPath}, err.Packages...)
-			}
-			return err
-		}
-		p.Imports[to] = importedPkg
-	}
-	p.Importing = false
-	return nil
-}
@@ -10,9 +10,17 @@ import (
 func (p *Program) LoadSSA() *ssa.Program {
 	prog := ssa.NewProgram(p.fset, ssa.SanityCheckFunctions|ssa.BareInits|ssa.GlobalDebug)

-	for _, pkg := range p.Sorted() {
-		prog.CreatePackage(pkg.Pkg, pkg.Files, &pkg.Info, true)
+	for _, pkg := range p.sorted {
+		prog.CreatePackage(pkg.Pkg, pkg.Files, &pkg.info, true)
 	}

 	return prog
 }
+
+// LoadSSA constructs the SSA form of this package.
+//
+// The program must already be parsed and type-checked with the .Parse() method.
+func (p *Package) LoadSSA() *ssa.Package {
+	prog := ssa.NewProgram(p.program.fset, ssa.SanityCheckFunctions|ssa.BareInits|ssa.GlobalDebug)
+	return prog.CreatePackage(p.Pkg, p.Files, &p.info, true)
+}
@@ -8,26 +8,34 @@ import (
 	"go/scanner"
 	"go/types"
 	"io"
+	"io/ioutil"
 	"os"
 	"os/exec"
 	"os/signal"
 	"path/filepath"
 	"runtime"
-	"strconv"
 	"strings"
-	"syscall"
+	"sync/atomic"
 	"time"

+	"github.com/mattn/go-colorable"
 	"github.com/tinygo-org/tinygo/builder"
 	"github.com/tinygo-org/tinygo/compileopts"
 	"github.com/tinygo-org/tinygo/goenv"
 	"github.com/tinygo-org/tinygo/interp"
 	"github.com/tinygo-org/tinygo/loader"
+	"github.com/tinygo-org/tinygo/transform"
 	"tinygo.org/x/go-llvm"

 	"go.bug.st/serial"
 )

+var (
+	// This variable is set at build time using -ldflags parameters.
+	// See: https://stackoverflow.com/a/11355611
+	gitSha1 string
+)
+
 // commandError is an error type to wrap os/exec.Command errors. This provides
 // some more information regarding what went wrong while running a command.
 type commandError struct {
@@ -58,33 +66,36 @@ func moveFile(src, dst string) error {
 	return os.Remove(src)
 }

-// copyFile copies the given file from src to dst. It copies first to a .tmp
-// file which is then moved over a possibly already existing file at the
-// destination.
+// 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)
+	source, err := os.Open(src)
 	if err != nil {
 		return err
 	}
-	defer inf.Close()
-	outpath := dst + ".tmp"
-	outf, err := os.Create(outpath)
+	defer source.Close()
+
+	st, err := source.Stat()
 	if err != nil {
 		return err
 	}

-	_, err = io.Copy(outf, inf)
-	if err != nil {
-		os.Remove(outpath)
-		return err
-	}
-
-	err = outf.Close()
+	destination, err := os.OpenFile(dst, os.O_RDWR|os.O_CREATE|os.O_TRUNC, st.Mode())
 	if err != nil {
 		return err
 	}
+	defer destination.Close()

-	return os.Rename(dst+".tmp", dst)
+	_, err = io.Copy(destination, source)
+	return err
+}
+
+// executeCommand is a simple wrapper to exec.Cmd
+func executeCommand(options *compileopts.Options, name string, arg ...string) *exec.Cmd {
+	if options.PrintCommands {
+		fmt.Printf("%s %s\n ", name, strings.Join(arg, " "))
+	}
+	return exec.Command(name, arg...)
 }

 // Build compiles and links the given package and writes it to outpath.
@@ -94,10 +105,10 @@ func Build(pkgName, outpath string, options *compileopts.Options) error {
 		return err
 	}

-	return builder.Build(pkgName, outpath, config, func(tmppath string) error {
-		if err := os.Rename(tmppath, outpath); err != nil {
+	return builder.Build(pkgName, outpath, config, func(result builder.BuildResult) error {
+		if err := os.Rename(result.Binary, outpath); err != nil {
 			// Moving failed. Do a file copy.
-			inf, err := os.Open(tmppath)
+			inf, err := os.Open(result.Binary)
 			if err != nil {
 				return err
 			}
@@ -123,35 +134,68 @@ func Build(pkgName, outpath string, options *compileopts.Options) error {
 }

 // Test runs the tests in the given package.
-func Test(pkgName string, options *compileopts.Options) error {
+func Test(pkgName string, options *compileopts.Options, testCompileOnly bool, outpath string) error {
+	options.TestConfig.CompileTestBinary = true
 	config, err := builder.NewConfig(options)
 	if err != nil {
 		return err
 	}

-	// Add test build tag. This is incorrect: `go test` only looks at the
-	// _test.go file suffix but does not add the test build tag in the process.
-	// However, it's a simple fix right now.
-	// For details: https://github.com/golang/go/issues/21360
-	config.Target.BuildTags = append(config.Target.BuildTags, "test")
-
-	options.TestConfig.CompileTestBinary = true
-	return builder.Build(pkgName, ".elf", config, func(tmppath string) error {
-		cmd := exec.Command(tmppath)
-		cmd.Stdout = os.Stdout
-		cmd.Stderr = os.Stderr
-		err := cmd.Run()
-		if err != nil {
-			// Propagate the exit code
-			if err, ok := err.(*exec.ExitError); ok {
-				if status, ok := err.Sys().(syscall.WaitStatus); ok {
-					os.Exit(status.ExitStatus())
-				}
-				os.Exit(1)
+	return builder.Build(pkgName, outpath, config, func(result builder.BuildResult) error {
+		if testCompileOnly || outpath != "" {
+			// Write test binary to the specified file name.
+			if outpath == "" {
+				// No -o path was given, so create one now.
+				// This matches the behavior of go test.
+				outpath = filepath.Base(result.MainDir) + ".test"
+			}
+			copyFile(result.Binary, outpath)
+		}
+		if testCompileOnly {
+			// Do not run the test.
+			return nil
+		}
+		if len(config.Target.Emulator) == 0 {
+			// Run directly.
+			cmd := executeCommand(config.Options, result.Binary)
+			cmd.Stdout = os.Stdout
+			cmd.Stderr = os.Stderr
+			cmd.Dir = result.MainDir
+			err := cmd.Run()
+			if err != nil {
+				// Propagate the exit code
+				if err, ok := err.(*exec.ExitError); ok {
+					os.Exit(err.ExitCode())
+				}
+				return &commandError{"failed to run compiled binary", result.Binary, err}
+			}
+			return nil
+		} else {
+			// Run in an emulator.
+			args := append(config.Target.Emulator[1:], result.Binary)
+			cmd := executeCommand(config.Options, config.Target.Emulator[0], args...)
+			buf := &bytes.Buffer{}
+			w := io.MultiWriter(os.Stdout, buf)
+			cmd.Stdout = w
+			cmd.Stderr = os.Stderr
+			err := cmd.Run()
+			if err != nil {
+				if err, ok := err.(*exec.ExitError); !ok || !err.Exited() {
+					// Workaround for QEMU which always exits with an error.
+					return &commandError{"failed to run emulator with", result.Binary, err}
+				}
+			}
+			testOutput := string(buf.Bytes())
+			if testOutput == "PASS\n" || strings.HasSuffix(testOutput, "\nPASS\n") {
+				// Test passed.
+				return nil
+			} else {
+				// Test failed, either by ending with the word "FAIL" or with a
+				// panic of some sort.
+				os.Exit(1)
+				return nil // unreachable
 			}
-			return &commandError{"failed to run compiled binary", tmppath, err}
 		}
-		return nil
 	})
 }

@@ -193,9 +237,9 @@ func Flash(pkgName, port string, options *compileopts.Options) error {
 		return errors.New("unknown flash method: " + flashMethod)
 	}

-	return builder.Build(pkgName, fileExt, config, func(tmppath string) error {
+	return builder.Build(pkgName, fileExt, config, func(result builder.BuildResult) error {
 		// do we need port reset to put MCU into bootloader mode?
-		if config.Target.PortReset == "true" {
+		if config.Target.PortReset == "true" && flashMethod != "openocd" {
 			if port == "" {
 				var err error
 				port, err = getDefaultPort()
@@ -206,7 +250,7 @@ func Flash(pkgName, port string, options *compileopts.Options) error {

 			err := touchSerialPortAt1200bps(port)
 			if err != nil {
-				return &commandError{"failed to reset port", tmppath, err}
+				return &commandError{"failed to reset port", result.Binary, err}
 			}
 			// give the target MCU a chance to restart into bootloader
 			time.Sleep(3 * time.Second)
@@ -218,7 +262,7 @@ func Flash(pkgName, port string, options *compileopts.Options) error {
 			// Create the command.
 			flashCmd := config.Target.FlashCommand
 			fileToken := "{" + fileExt[1:] + "}"
-			flashCmd = strings.Replace(flashCmd, fileToken, tmppath, -1)
+			flashCmd = strings.ReplaceAll(flashCmd, fileToken, result.Binary)

 			if port == "" && strings.Contains(flashCmd, "{port}") {
 				var err error
@@ -228,7 +272,7 @@ func Flash(pkgName, port string, options *compileopts.Options) error {
 				}
 			}

-			flashCmd = strings.Replace(flashCmd, "{port}", port, -1)
+			flashCmd = strings.ReplaceAll(flashCmd, "{port}", port)

 			// Execute the command.
 			var cmd *exec.Cmd
@@ -238,9 +282,9 @@ func Flash(pkgName, port string, options *compileopts.Options) error {
 				if len(command) < 2 {
 					return errors.New("invalid flash command")
 				}
-				cmd = exec.Command(command[0], command[1:]...)
+				cmd = executeCommand(config.Options, command[0], command[1:]...)
 			default:
-				cmd = exec.Command("/bin/sh", "-c", flashCmd)
+				cmd = executeCommand(config.Options, "/bin/sh", "-c", flashCmd)
 			}

 			cmd.Stdout = os.Stdout
@@ -248,21 +292,21 @@ func Flash(pkgName, port string, options *compileopts.Options) error {
 			cmd.Dir = goenv.Get("TINYGOROOT")
 			err := cmd.Run()
 			if err != nil {
-				return &commandError{"failed to flash", tmppath, err}
+				return &commandError{"failed to flash", result.Binary, err}
 			}
 			return nil
 		case "msd":
 			switch fileExt {
 			case ".uf2":
-				err := flashUF2UsingMSD(config.Target.FlashVolume, tmppath)
+				err := flashUF2UsingMSD(config.Target.FlashVolume, result.Binary, config.Options)
 				if err != nil {
-					return &commandError{"failed to flash", tmppath, err}
+					return &commandError{"failed to flash", result.Binary, err}
 				}
 				return nil
 			case ".hex":
-				err := flashHexUsingMSD(config.Target.FlashVolume, tmppath)
+				err := flashHexUsingMSD(config.Target.FlashVolume, result.Binary, config.Options)
 				if err != nil {
-					return &commandError{"failed to flash", tmppath, err}
+					return &commandError{"failed to flash", result.Binary, err}
 				}
 				return nil
 			default:
@@ -273,13 +317,13 @@ func Flash(pkgName, port string, options *compileopts.Options) error {
 			if err != nil {
 				return err
 			}
-			args = append(args, "-c", "program "+tmppath+" reset exit")
-			cmd := exec.Command("openocd", args...)
+			args = append(args, "-c", "program "+filepath.ToSlash(result.Binary)+" reset exit")
+			cmd := executeCommand(config.Options, "openocd", args...)
 			cmd.Stdout = os.Stdout
 			cmd.Stderr = os.Stderr
 			err = cmd.Run()
 			if err != nil {
-				return &commandError{"failed to flash", tmppath, err}
+				return &commandError{"failed to flash", result.Binary, err}
 			}
 			return nil
 		default:
@@ -304,22 +348,28 @@ func FlashGDB(pkgName string, ocdOutput bool, options *compileopts.Options) erro
 		return errors.New("gdb not configured in the target specification")
 	}

-	return builder.Build(pkgName, "", config, func(tmppath string) error {
+	return builder.Build(pkgName, "", config, func(result builder.BuildResult) error {
 		// Find a good way to run GDB.
 		gdbInterface, openocdInterface := config.Programmer()
 		switch gdbInterface {
 		case "msd", "command", "":
 			if len(config.Target.Emulator) != 0 {
-				// Assume QEMU as an emulator.
 				if config.Target.Emulator[0] == "mgba" {
 					gdbInterface = "mgba"
-				} else {
+				} else if config.Target.Emulator[0] == "simavr" {
+					gdbInterface = "simavr"
+				} else if strings.HasPrefix(config.Target.Emulator[0], "qemu-system-") {
 					gdbInterface = "qemu"
+				} else {
+					// Assume QEMU as an emulator.
+					gdbInterface = "qemu-user"
 				}
 			} else if openocdInterface != "" && config.Target.OpenOCDTarget != "" {
 				gdbInterface = "openocd"
 			} else if config.Target.JLinkDevice != "" {
 				gdbInterface = "jlink"
+			} else {
+				gdbInterface = "native"
 			}
 		}

@@ -337,11 +387,11 @@ func FlashGDB(pkgName string, ocdOutput bool, options *compileopts.Options) erro
 			if err != nil {
 				return err
 			}
-			daemon = exec.Command("openocd", args...)
+			daemon = executeCommand(config.Options, "openocd", args...)
 			if ocdOutput {
 				// Make it clear which output is from the daemon.
 				w := &ColorWriter{
-					Out:    os.Stderr,
+					Out:    colorable.NewColorableStderr(),
 					Prefix: "openocd: ",
 					Color:  TermColorYellow,
 				}
@@ -352,11 +402,11 @@ func FlashGDB(pkgName string, ocdOutput bool, options *compileopts.Options) erro
 			gdbCommands = append(gdbCommands, "target remote :2331", "load", "monitor reset halt")

 			// We need a separate debugging daemon for on-chip debugging.
-			daemon = exec.Command("JLinkGDBServer", "-device", config.Target.JLinkDevice)
+			daemon = executeCommand(config.Options, "JLinkGDBServer", "-device", config.Target.JLinkDevice)
 			if ocdOutput {
 				// Make it clear which output is from the daemon.
 				w := &ColorWriter{
-					Out:    os.Stderr,
+					Out:    colorable.NewColorableStderr(),
 					Prefix: "jlink: ",
 					Color:  TermColorYellow,
 				}
@@ -367,16 +417,32 @@ func FlashGDB(pkgName string, ocdOutput bool, options *compileopts.Options) erro
 			gdbCommands = append(gdbCommands, "target remote :1234")

 			// Run in an emulator.
-			args := append(config.Target.Emulator[1:], tmppath, "-s", "-S")
-			daemon = exec.Command(config.Target.Emulator[0], args...)
+			args := append(config.Target.Emulator[1:], result.Binary, "-s", "-S")
+			daemon = executeCommand(config.Options, config.Target.Emulator[0], args...)
+			daemon.Stdout = os.Stdout
+			daemon.Stderr = os.Stderr
+		case "qemu-user":
+			gdbCommands = append(gdbCommands, "target remote :1234")
+
+			// Run in an emulator.
+			args := append(config.Target.Emulator[1:], "-g", "1234", result.Binary)
+			daemon = executeCommand(config.Options, config.Target.Emulator[0], args...)
 			daemon.Stdout = os.Stdout
 			daemon.Stderr = os.Stderr
 		case "mgba":
 			gdbCommands = append(gdbCommands, "target remote :2345")

 			// Run in an emulator.
-			args := append(config.Target.Emulator[1:], tmppath, "-g")
-			daemon = exec.Command(config.Target.Emulator[0], args...)
+			args := append(config.Target.Emulator[1:], result.Binary, "-g")
+			daemon = executeCommand(config.Options, config.Target.Emulator[0], args...)
+			daemon.Stdout = os.Stdout
+			daemon.Stderr = os.Stderr
+		case "simavr":
+			gdbCommands = append(gdbCommands, "target remote :1234")
+
+			// Run in an emulator.
+			args := append(config.Target.Emulator[1:], "-g", result.Binary)
+			daemon = executeCommand(config.Options, config.Target.Emulator[0], args...)
 			daemon.Stdout = os.Stdout
 			daemon.Stderr = os.Stderr
 		case "msd":
@@ -397,8 +463,15 @@ func FlashGDB(pkgName string, ocdOutput bool, options *compileopts.Options) erro
 			}
 			defer func() {
 				daemon.Process.Signal(os.Interrupt)
-				// Maybe we should send a .Kill() after x seconds?
+				var stopped uint32
+				go func() {
+					time.Sleep(time.Millisecond * 100)
+					if atomic.LoadUint32(&stopped) == 0 {
+						daemon.Process.Kill()
+					}
+				}()
 				daemon.Wait()
+				atomic.StoreUint32(&stopped, 1)
 			}()
 		}

@@ -413,17 +486,17 @@ func FlashGDB(pkgName string, ocdOutput bool, options *compileopts.Options) erro
 		// Construct and execute a gdb command.
 		// By default: gdb -ex run <binary>
 		// Exit GDB with Ctrl-D.
-		params := []string{tmppath}
+		params := []string{result.Binary}
 		for _, cmd := range gdbCommands {
 			params = append(params, "-ex", cmd)
 		}
-		cmd := exec.Command(config.Target.GDB, params...)
+		cmd := executeCommand(config.Options, config.Target.GDB, params...)
 		cmd.Stdin = os.Stdin
 		cmd.Stdout = os.Stdout
 		cmd.Stderr = os.Stderr
 		err := cmd.Run()
 		if err != nil {
-			return &commandError{"failed to run gdb with", tmppath, err}
+			return &commandError{"failed to run gdb with", result.Binary, err}
 		}
 		return nil
 	})
@@ -439,10 +512,10 @@ func Run(pkgName string, options *compileopts.Options) error {
 		return err
 	}

-	return builder.Build(pkgName, ".elf", config, func(tmppath string) error {
+	return builder.Build(pkgName, ".elf", config, func(result builder.BuildResult) error {
 		if len(config.Target.Emulator) == 0 {
 			// Run directly.
-			cmd := exec.Command(tmppath)
+			cmd := executeCommand(config.Options, result.Binary)
 			cmd.Stdout = os.Stdout
 			cmd.Stderr = os.Stderr
 			err := cmd.Run()
@@ -451,13 +524,13 @@ func Run(pkgName string, options *compileopts.Options) error {
 					// Workaround for QEMU which always exits with an error.
 					return nil
 				}
-				return &commandError{"failed to run compiled binary", tmppath, err}
+				return &commandError{"failed to run compiled binary", result.Binary, err}
 			}
 			return nil
 		} else {
 			// Run in an emulator.
-			args := append(config.Target.Emulator[1:], tmppath)
-			cmd := exec.Command(config.Target.Emulator[0], args...)
+			args := append(config.Target.Emulator[1:], result.Binary)
+			cmd := executeCommand(config.Options, config.Target.Emulator[0], args...)
 			cmd.Stdout = os.Stdout
 			cmd.Stderr = os.Stderr
 			err := cmd.Run()
@@ -466,7 +539,7 @@ func Run(pkgName string, options *compileopts.Options) error {
 					// Workaround for QEMU which always exits with an error.
 					return nil
 				}
-				return &commandError{"failed to run emulator with", tmppath, err}
+				return &commandError{"failed to run emulator with", result.Binary, err}
 			}
 			return nil
 		}
@@ -479,6 +552,13 @@ func touchSerialPortAt1200bps(port string) (err error) {
 		// Open port
 		p, e := serial.Open(port, &serial.Mode{BaudRate: 1200})
 		if e != nil {
+			if runtime.GOOS == `windows` {
+				se, ok := e.(*serial.PortError)
+				if ok && se.Code() == serial.InvalidSerialPort {
+					// InvalidSerialPort error occurs when transitioning to boot
+					return nil
+				}
+			}
 			time.Sleep(1 * time.Second)
 			err = e
 			continue
@@ -491,7 +571,9 @@ func touchSerialPortAt1200bps(port string) (err error) {
 	return fmt.Errorf("opening port: %s", err)
 }

-func flashUF2UsingMSD(volume, tmppath string) error {
+const maxMSDRetries = 10
+
+func flashUF2UsingMSD(volume, tmppath string, options *compileopts.Options) error {
 	// find standard UF2 info path
 	var infoPath string
 	switch runtime.GOOS {
@@ -500,25 +582,22 @@ func flashUF2UsingMSD(volume, tmppath string) error {
 	case "darwin":
 		infoPath = "/Volumes/" + volume + "/INFO_UF2.TXT"
 	case "windows":
-		path, err := windowsFindUSBDrive(volume)
+		path, err := windowsFindUSBDrive(volume, options)
 		if err != nil {
 			return err
 		}
 		infoPath = path + "/INFO_UF2.TXT"
 	}

-	d, err := filepath.Glob(infoPath)
+	d, err := locateDevice(volume, infoPath)
 	if err != nil {
 		return err
 	}
-	if d == nil {
-		return errors.New("unable to locate UF2 device: " + volume)
-	}

-	return moveFile(tmppath, filepath.Dir(d[0])+"/flash.uf2")
+	return moveFile(tmppath, filepath.Dir(d)+"/flash.uf2")
 }

-func flashHexUsingMSD(volume, tmppath string) error {
+func flashHexUsingMSD(volume, tmppath string, options *compileopts.Options) error {
 	// find expected volume path
 	var destPath string
 	switch runtime.GOOS {
@@ -527,26 +606,42 @@ func flashHexUsingMSD(volume, tmppath string) error {
 	case "darwin":
 		destPath = "/Volumes/" + volume
 	case "windows":
-		path, err := windowsFindUSBDrive(volume)
+		path, err := windowsFindUSBDrive(volume, options)
 		if err != nil {
 			return err
 		}
 		destPath = path + "/"
 	}

-	d, err := filepath.Glob(destPath)
+	d, err := locateDevice(volume, destPath)
 	if err != nil {
 		return err
 	}
-	if d == nil {
-		return errors.New("unable to locate device: " + volume)
-	}

-	return moveFile(tmppath, d[0]+"/flash.hex")
+	return moveFile(tmppath, d+"/flash.hex")
 }

-func windowsFindUSBDrive(volume string) (string, error) {
-	cmd := exec.Command("wmic",
+func locateDevice(volume, path string) (string, error) {
+	var d []string
+	var err error
+	for i := 0; i < maxMSDRetries; i++ {
+		d, err = filepath.Glob(path)
+		if err != nil {
+			return "", err
+		}
+		if d != nil {
+			break
+		}
+		time.Sleep(500 * time.Millisecond)
+	}
+	if d == nil {
+		return "", errors.New("unable to locate device: " + volume)
+	}
+	return d[0], nil
+}
+
+func windowsFindUSBDrive(volume string, options *compileopts.Options) (string, error) {
+	cmd := executeCommand(options, "wmic",
 		"PATH", "Win32_LogicalDisk", "WHERE", "VolumeName = '"+volume+"'",
 		"get", "DeviceID,VolumeName,FileSystem,DriveType")

@@ -568,30 +663,6 @@ func windowsFindUSBDrive(volume string) (string, error) {
 	return "", errors.New("unable to locate a USB device to be flashed")
 }

-// parseSize converts a human-readable size (with k/m/g suffix) into a plain
-// number.
-func parseSize(s string) (int64, error) {
-	s = strings.ToLower(strings.TrimSpace(s))
-	if len(s) == 0 {
-		return 0, errors.New("no size provided")
-	}
-	multiply := int64(1)
-	switch s[len(s)-1] {
-	case 'k':
-		multiply = 1 << 10
-	case 'm':
-		multiply = 1 << 20
-	case 'g':
-		multiply = 1 << 30
-	}
-	if multiply != 1 {
-		s = s[:len(s)-1]
-	}
-	n, err := strconv.ParseInt(s, 0, 64)
-	n *= multiply
-	return n, err
-}
-
 // getDefaultPort returns the default serial port depending on the operating system.
 func getDefaultPort() (port string, err error) {
 	var portPath string
@@ -603,29 +674,18 @@ func getDefaultPort() (port string, err error) {
 	case "freebsd":
 		portPath = "/dev/cuaU*"
 	case "windows":
-		cmd := exec.Command("wmic",
-			"PATH", "Win32_SerialPort", "WHERE", "Caption LIKE 'USB Serial%'", "GET", "DeviceID")
-
-		var out bytes.Buffer
-		cmd.Stdout = &out
-		err := cmd.Run()
+		ports, err := serial.GetPortsList()
 		if err != nil {
 			return "", err
 		}

-		if out.String() == "No Instance(s) Available." {
+		if len(ports) == 0 {
 			return "", errors.New("no serial ports available")
+		} else if len(ports) > 1 {
+			return "", errors.New("multiple serial ports available - use -port flag")
 		}

-		for _, line := range strings.Split(out.String(), "\n") {
-			words := strings.Fields(line)
-			if len(words) == 1 {
-				if strings.Contains(words[0], "COM") {
-					return words[0], nil
-				}
-			}
-		}
-		return "", errors.New("unable to locate a serial port")
+		return ports[0], nil
 	default:
 		return "", errors.New("unable to search for a default USB device to be flashed on this OS")
 	}
@@ -643,7 +703,7 @@ func getDefaultPort() (port string, err error) {

 func usage() {
 	fmt.Fprintln(os.Stderr, "TinyGo is a Go compiler for small places.")
-	fmt.Fprintln(os.Stderr, "version:", version)
+	fmt.Fprintln(os.Stderr, "version:", goenv.Version)
 	fmt.Fprintf(os.Stderr, "usage: %s command [-printir] [-target=<target>] -o <output> <input>\n", os.Args[0])
 	fmt.Fprintln(os.Stderr, "\ncommands:")
 	fmt.Fprintln(os.Stderr, "  build: compile packages and dependencies")
@@ -652,12 +712,27 @@ func usage() {
 	fmt.Fprintln(os.Stderr, "  flash: compile and flash to the device")
 	fmt.Fprintln(os.Stderr, "  gdb:   run/flash and immediately enter GDB")
 	fmt.Fprintln(os.Stderr, "  env:   list environment variables used during build")
+	fmt.Fprintln(os.Stderr, "  list:  run go list using the TinyGo root")
 	fmt.Fprintln(os.Stderr, "  clean: empty cache directory ("+goenv.Get("GOCACHE")+")")
 	fmt.Fprintln(os.Stderr, "  help:  print this help text")
 	fmt.Fprintln(os.Stderr, "\nflags:")
 	flag.PrintDefaults()
 }

+// try to make the path relative to the current working directory. If any error
+// occurs, this error is ignored and the absolute path is returned instead.
+func tryToMakePathRelative(dir string) string {
+	wd, err := os.Getwd()
+	if err != nil {
+		return dir
+	}
+	relpath, err := filepath.Rel(wd, dir)
+	if err != nil {
+		return dir
+	}
+	return relpath
+}
+
 // printCompilerError prints compiler errors using the provided logger function
 // (similar to fmt.Println).
 //
@@ -665,8 +740,24 @@ func usage() {
 // to limitations in the LLVM bindings.
 func printCompilerError(logln func(...interface{}), err error) {
 	switch err := err.(type) {
-	case types.Error, scanner.Error:
+	case types.Error:
+		printCompilerError(logln, scanner.Error{
+			Pos: err.Fset.Position(err.Pos),
+			Msg: err.Msg,
+		})
+	case scanner.Error:
+		if !strings.HasPrefix(err.Pos.Filename, filepath.Join(goenv.Get("GOROOT"), "src")) && !strings.HasPrefix(err.Pos.Filename, filepath.Join(goenv.Get("TINYGOROOT"), "src")) {
+			// This file is not from the standard library (either the GOROOT or
+			// the TINYGOROOT). Make the path relative, for easier reading.
+			// Ignore any errors in the process (falling back to the absolute
+			// path).
+			err.Pos.Filename = tryToMakePathRelative(err.Pos.Filename)
+		}
 		logln(err)
+	case scanner.ErrorList:
+		for _, scannerErr := range err {
+			printCompilerError(logln, *scannerErr)
+		}
 	case *interp.Error:
 		logln("#", err.ImportPath)
 		logln(err.Error())
@@ -682,14 +773,29 @@ func printCompilerError(logln func(...interface{}), err error) {
 				logln()
 			}
 		}
+	case transform.CoroutinesError:
+		logln(err.Pos.String() + ": " + err.Msg)
+		logln("\ntraceback:")
+		for _, line := range err.Traceback {
+			logln(line.Name)
+			if line.Position.IsValid() {
+				logln("\t" + line.Position.String())
+			}
+		}
 	case loader.Errors:
 		logln("#", err.Pkg.ImportPath)
 		for _, err := range err.Errs {
-			logln(err)
+			printCompilerError(logln, err)
+		}
+	case loader.Error:
+		logln(err.Err.Error())
+		logln("package", err.ImportStack[0])
+		for _, pkgPath := range err.ImportStack[1:] {
+			logln("\timports", pkgPath)
 		}
 	case *builder.MultiError:
 		for _, err := range err.Errs {
-			logln(err)
+			printCompilerError(logln, err)
 		}
 	default:
 		logln("error:", err)
@@ -706,26 +812,6 @@ func handleCompilerError(err error) {
 }

 func main() {
-	outpath := flag.String("o", "", "output filename")
-	opt := flag.String("opt", "z", "optimization level: 0, 1, 2, s, z")
-	gc := flag.String("gc", "", "garbage collector to use (none, leaking, extalloc, conservative)")
-	panicStrategy := flag.String("panic", "print", "panic strategy (print, trap)")
-	scheduler := flag.String("scheduler", "", "which scheduler to use (coroutines, tasks)")
-	printIR := flag.Bool("printir", false, "print LLVM IR")
-	dumpSSA := flag.Bool("dumpssa", false, "dump internal Go SSA")
-	verifyIR := flag.Bool("verifyir", false, "run extra verification steps on LLVM IR")
-	tags := flag.String("tags", "", "a space-separated list of extra build tags")
-	target := flag.String("target", "", "LLVM target | .json file with TargetSpec")
-	printSize := flag.String("size", "", "print sizes (none, short, full)")
-	nodebug := flag.Bool("no-debug", false, "disable DWARF debug symbol generation")
-	ocdOutput := flag.Bool("ocd-output", false, "print OCD daemon output during debug")
-	port := flag.String("port", "", "flash port")
-	programmer := flag.String("programmer", "", "which hardware programmer to use")
-	cFlags := flag.String("cflags", "", "additional cflags for compiler")
-	ldFlags := flag.String("ldflags", "", "additional ldflags for linker")
-	wasmAbi := flag.String("wasm-abi", "js", "WebAssembly ABI conventions: js (no i64 params) or generic")
-	heapSize := flag.String("heap-size", "1M", "default heap size in bytes (only supported by WebAssembly)")
-
 	if len(os.Args) < 2 {
 		fmt.Fprintln(os.Stderr, "No command-line arguments supplied.")
 		usage()
@@ -733,6 +819,40 @@ func main() {
 	}
 	command := os.Args[1]

+	opt := flag.String("opt", "z", "optimization level: 0, 1, 2, s, z")
+	gc := flag.String("gc", "", "garbage collector to use (none, leaking, extalloc, conservative)")
+	panicStrategy := flag.String("panic", "print", "panic strategy (print, trap)")
+	scheduler := flag.String("scheduler", "", "which scheduler to use (none, coroutines, tasks)")
+	printIR := flag.Bool("printir", false, "print LLVM IR")
+	dumpSSA := flag.Bool("dumpssa", false, "dump internal Go SSA")
+	verifyIR := flag.Bool("verifyir", false, "run extra verification steps on LLVM IR")
+	tags := flag.String("tags", "", "a space-separated list of extra build tags")
+	target := flag.String("target", "", "LLVM target | .json file with TargetSpec")
+	printSize := flag.String("size", "", "print sizes (none, short, full)")
+	printStacks := flag.Bool("print-stacks", false, "print stack sizes of goroutines")
+	printCommands := flag.Bool("x", false, "Print commands")
+	nodebug := flag.Bool("no-debug", false, "disable DWARF debug symbol generation")
+	ocdOutput := flag.Bool("ocd-output", false, "print OCD daemon output during debug")
+	port := flag.String("port", "", "flash port")
+	programmer := flag.String("programmer", "", "which hardware programmer to use")
+	cFlags := flag.String("cflags", "", "additional cflags for compiler")
+	ldFlags := flag.String("ldflags", "", "additional ldflags for linker")
+	wasmAbi := flag.String("wasm-abi", "", "WebAssembly ABI conventions: js (no i64 params) or generic")
+
+	var flagJSON, flagDeps *bool
+	if command == "help" || command == "list" {
+		flagJSON = flag.Bool("json", false, "print data in JSON format")
+		flagDeps = flag.Bool("deps", false, "")
+	}
+	var outpath string
+	if command == "help" || command == "build" || command == "build-library" || command == "test" {
+		flag.StringVar(&outpath, "o", "", "output filename")
+	}
+	var testCompileOnlyFlag *bool
+	if command == "help" || command == "test" {
+		testCompileOnlyFlag = flag.Bool("c", false, "compile the test binary but do not run it")
+	}
+
 	// Early command processing, before commands are interpreted by the Go flag
 	// library.
 	switch command {
@@ -757,6 +877,8 @@ func main() {
 		VerifyIR:      *verifyIR,
 		Debug:         !*nodebug,
 		PrintSizes:    *printSize,
+		PrintStacks:   *printStacks,
+		PrintCommands: *printCommands,
 		Tags:          *tags,
 		WasmAbi:       *wasmAbi,
 		Programmer:    *programmer,
@@ -770,44 +892,39 @@ func main() {
 		options.LDFlags = strings.Split(*ldFlags, " ")
 	}

-	if *panicStrategy != "print" && *panicStrategy != "trap" {
-		fmt.Fprintln(os.Stderr, "Panic strategy must be either print or trap.")
-		usage()
-		os.Exit(1)
-	}
-
-	var err error
-	if options.HeapSize, err = parseSize(*heapSize); err != nil {
-		fmt.Fprintln(os.Stderr, "Could not read heap size:", *heapSize)
-		usage()
-		os.Exit(1)
-	}
-
 	os.Setenv("CC", "clang -target="+*target)

+	err := options.Verify()
+	if err != nil {
+		fmt.Fprintln(os.Stderr, err.Error())
+		usage()
+		os.Exit(1)
+	}
+
 	switch command {
 	case "build":
-		if *outpath == "" {
+		if outpath == "" {
 			fmt.Fprintln(os.Stderr, "No output filename supplied (-o).")
 			usage()
 			os.Exit(1)
 		}
 		pkgName := "."
 		if flag.NArg() == 1 {
-			pkgName = flag.Arg(0)
+			pkgName = filepath.ToSlash(flag.Arg(0))
 		} else if flag.NArg() > 1 {
 			fmt.Fprintln(os.Stderr, "build only accepts a single positional argument: package name, but multiple were specified")
 			usage()
 			os.Exit(1)
 		}
-		if options.Target == "" && filepath.Ext(*outpath) == ".wasm" {
+		if options.Target == "" && filepath.Ext(outpath) == ".wasm" {
 			options.Target = "wasm"
 		}
-		err := Build(pkgName, *outpath, options)
+
+		err := Build(pkgName, outpath, options)
 		handleCompilerError(err)
 	case "build-library":
 		// Note: this command is only meant to be used while making a release!
-		if *outpath == "" {
+		if outpath == "" {
 			fmt.Fprintln(os.Stderr, "No output filename supplied (-o).")
 			usage()
 			os.Exit(1)
@@ -830,17 +947,21 @@ func main() {
 			fmt.Fprintf(os.Stderr, "Unknown library: %s\n", name)
 			os.Exit(1)
 		}
-		path, err := lib.Load(*target)
-		handleCompilerError(err)
-		copyFile(path, *outpath)
-	case "flash", "gdb":
-		if *outpath != "" {
-			fmt.Fprintln(os.Stderr, "Output cannot be specified with the flash command.")
-			usage()
-			os.Exit(1)
+		tmpdir, err := ioutil.TempDir("", "tinygo*")
+		if err != nil {
+			handleCompilerError(err)
 		}
+		defer os.RemoveAll(tmpdir)
+		path, err := lib.Load(*target, tmpdir)
+		handleCompilerError(err)
+		err = copyFile(path, outpath)
+		if err != nil {
+			handleCompilerError(err)
+		}
+	case "flash", "gdb":
+		pkgName := filepath.ToSlash(flag.Arg(0))
 		if command == "flash" {
-			err := Flash(flag.Arg(0), *port, options)
+			err := Flash(pkgName, *port, options)
 			handleCompilerError(err)
 		} else {
 			if !options.Debug {
@@ -848,7 +969,7 @@ func main() {
 				usage()
 				os.Exit(1)
 			}
-			err := FlashGDB(flag.Arg(0), *ocdOutput, options)
+			err := FlashGDB(pkgName, *ocdOutput, options)
 			handleCompilerError(err)
 		}
 	case "run":
@@ -857,19 +978,49 @@ func main() {
 			usage()
 			os.Exit(1)
 		}
-		err := Run(flag.Arg(0), options)
+		pkgName := filepath.ToSlash(flag.Arg(0))
+		err := Run(pkgName, options)
 		handleCompilerError(err)
 	case "test":
 		pkgName := "."
 		if flag.NArg() == 1 {
-			pkgName = flag.Arg(0)
+			pkgName = filepath.ToSlash(flag.Arg(0))
 		} else if flag.NArg() > 1 {
 			fmt.Fprintln(os.Stderr, "test only accepts a single positional argument: package name, but multiple were specified")
 			usage()
 			os.Exit(1)
 		}
-		err := Test(pkgName, options)
+		err := Test(pkgName, options, *testCompileOnlyFlag, outpath)
 		handleCompilerError(err)
+	case "targets":
+		dir := filepath.Join(goenv.Get("TINYGOROOT"), "targets")
+		entries, err := ioutil.ReadDir(dir)
+		if err != nil {
+			fmt.Fprintln(os.Stderr, "could not list targets:", err)
+			os.Exit(1)
+			return
+		}
+		for _, entry := range entries {
+			if !entry.Mode().IsRegular() || !strings.HasSuffix(entry.Name(), ".json") {
+				// Only inspect JSON files.
+				continue
+			}
+			path := filepath.Join(dir, entry.Name())
+			spec, err := compileopts.LoadTarget(path)
+			if err != nil {
+				fmt.Fprintln(os.Stderr, "could not list target:", err)
+				os.Exit(1)
+				return
+			}
+			if spec.FlashMethod == "" && spec.FlashCommand == "" && spec.Emulator == nil {
+				// This doesn't look like a regular target file, but rather like
+				// a parent target (such as targets/cortex-m.json).
+				continue
+			}
+			name := entry.Name()
+			name = name[:len(name)-5]
+			fmt.Println(name)
+		}
 	case "info":
 		if flag.NArg() == 1 {
 			options.Target = flag.Arg(0)
@@ -889,12 +1040,47 @@ func main() {
 			fmt.Fprintln(os.Stderr, err)
 			os.Exit(1)
 		}
+		cachedGOROOT, err := loader.GetCachedGoroot(config)
+		if err != nil {
+			fmt.Fprintln(os.Stderr, err)
+			os.Exit(1)
+		}
 		fmt.Printf("LLVM triple:       %s\n", config.Triple())
 		fmt.Printf("GOOS:              %s\n", config.GOOS())
 		fmt.Printf("GOARCH:            %s\n", config.GOARCH())
 		fmt.Printf("build tags:        %s\n", strings.Join(config.BuildTags(), " "))
 		fmt.Printf("garbage collector: %s\n", config.GC())
 		fmt.Printf("scheduler:         %s\n", config.Scheduler())
+		fmt.Printf("cached GOROOT:     %s\n", cachedGOROOT)
+	case "list":
+		config, err := builder.NewConfig(options)
+		if err != nil {
+			fmt.Fprintln(os.Stderr, err)
+			usage()
+			os.Exit(1)
+		}
+		var extraArgs []string
+		if *flagJSON {
+			extraArgs = append(extraArgs, "-json")
+		}
+		if *flagDeps {
+			extraArgs = append(extraArgs, "-deps")
+		}
+		cmd, err := loader.List(config, extraArgs, flag.Args())
+		if err != nil {
+			fmt.Fprintln(os.Stderr, "failed to run `go list`:", err)
+			os.Exit(1)
+		}
+		cmd.Stdout = os.Stdout
+		cmd.Stderr = os.Stderr
+		err = cmd.Run()
+		if err != nil {
+			if exitErr, ok := err.(*exec.ExitError); ok {
+				os.Exit(exitErr.ExitCode())
+			}
+			fmt.Fprintln(os.Stderr, "failed to run `go list`:", err)
+			os.Exit(1)
+		}
 	case "clean":
 		// remove cache directory
 		err := os.RemoveAll(goenv.Get("GOCACHE"))
@@ -906,9 +1092,13 @@ func main() {
 		usage()
 	case "version":
 		goversion := "<unknown>"
-		if s, err := builder.GorootVersionString(goenv.Get("GOROOT")); err == nil {
+		if s, err := goenv.GorootVersionString(goenv.Get("GOROOT")); err == nil {
 			goversion = s
 		}
+		version := goenv.Version
+		if strings.HasSuffix(goenv.Version, "-dev") && gitSha1 != "" {
+			version += "-" + gitSha1
+		}
 		fmt.Printf("tinygo version %s %s/%s (using go version %s and LLVM version %s)\n", version, runtime.GOOS, runtime.GOARCH, goversion, llvm.Version)
 	case "env":
 		if flag.NArg() == 0 {
@@ -6,6 +6,7 @@ package main
 import (
 	"bufio"
 	"bytes"
+	"flag"
 	"fmt"
 	"io/ioutil"
 	"os"
@@ -25,6 +26,8 @@ import (

 const TESTDATA = "testdata"

+var testTarget = flag.String("target", "", "override test target")
+
 func TestCompiler(t *testing.T) {
 	matches, err := filepath.Glob(filepath.Join(TESTDATA, "*.go"))
 	if err != nil {
@@ -44,9 +47,20 @@ func TestCompiler(t *testing.T) {

 	sort.Strings(matches)

+	if *testTarget != "" {
+		// This makes it possible to run one specific test (instead of all),
+		// which is especially useful to quickly check whether some changes
+		// affect a particular target architecture.
+		runPlatTests(*testTarget, matches, t)
+		return
+	}
+
 	if runtime.GOOS != "windows" {
 		t.Run("Host", func(t *testing.T) {
 			runPlatTests("", matches, t)
+			if runtime.GOOS == "darwin" {
+				runTest("testdata/libc/env.go", "", t, []string{"ENV1=VALUE1", "ENV2=VALUE2"}...)
+			}
 		})
 	}

@@ -67,13 +81,16 @@ func TestCompiler(t *testing.T) {
 	}

 	if runtime.GOOS == "linux" {
+		t.Run("X86Linux", func(t *testing.T) {
+			runPlatTests("i386--linux-gnu", matches, t)
+		})
 		t.Run("ARMLinux", func(t *testing.T) {
 			runPlatTests("arm--linux-gnueabihf", matches, t)
 		})
 		t.Run("ARM64Linux", func(t *testing.T) {
 			runPlatTests("aarch64--linux-gnu", matches, t)
 		})
-		goVersion, err := builder.GorootVersionString(goenv.Get("GOROOT"))
+		goVersion, err := goenv.GorootVersionString(goenv.Get("GOROOT"))
 		if err != nil {
 			t.Error("could not get Go version:", err)
 			return
@@ -87,6 +104,11 @@ func TestCompiler(t *testing.T) {
 				runPlatTests("wasm", matches, t)
 			})
 		}
+
+		t.Run("WASI", func(t *testing.T) {
+			runPlatTests("wasi", matches, t)
+			runTest("testdata/libc/env.go", "wasi", t, []string{"ENV1=VALUE1", "ENV2=VALUE2"}...)
+		})
 	}
 }

@@ -95,10 +117,8 @@ func runPlatTests(target string, matches []string, t *testing.T) {

 	for _, path := range matches {
 		path := path // redefine to avoid race condition
-
 		t.Run(filepath.Base(path), func(t *testing.T) {
 			t.Parallel()
-
 			runTest(path, target, t)
 		})
 	}
@@ -116,7 +136,7 @@ func runBuild(src, out string, opts *compileopts.Options) error {
 	return Build(src, out, opts)
 }

-func runTest(path, target string, t *testing.T) {
+func runTest(path, target string, t *testing.T, environmentVars ...string) {
 	// Get the expected output for this test.
 	txtpath := path[:len(path)-3] + ".txt"
 	if path[len(path)-1] == os.PathSeparator {
@@ -146,10 +166,11 @@ func runTest(path, target string, t *testing.T) {
 		PrintIR:    false,
 		DumpSSA:    false,
 		VerifyIR:   true,
-		Debug:      false,
+		Debug:      true,
 		PrintSizes: "",
-		WasmAbi:    "js",
+		WasmAbi:    "",
 	}
+
 	binary := filepath.Join(tmpdir, "test")
 	err = runBuild("./"+path, binary, config)
 	if err != nil {
@@ -164,16 +185,26 @@ func runTest(path, target string, t *testing.T) {
 	ranTooLong := false
 	if target == "" {
 		cmd = exec.Command(binary)
+		cmd.Env = append(cmd.Env, environmentVars...)
 	} else {
 		spec, err := compileopts.LoadTarget(target)
 		if err != nil {
 			t.Fatal("failed to load target spec:", err)
 		}
 		if len(spec.Emulator) == 0 {
-			t.Fatal("no emulator available for target:", target)
+			cmd = exec.Command(binary)
+		} else {
+			args := append(spec.Emulator[1:], binary)
+			cmd = exec.Command(spec.Emulator[0], args...)
+		}
+
+		if len(spec.Emulator) != 0 && spec.Emulator[0] == "wasmtime" {
+			for _, v := range environmentVars {
+				cmd.Args = append(cmd.Args, "--env", v)
+			}
+		} else {
+			cmd.Env = append(cmd.Env, environmentVars...)
 		}
-		args := append(spec.Emulator[1:], binary)
-		cmd = exec.Command(spec.Emulator[0], args...)
 	}
 	stdout := &bytes.Buffer{}
 	cmd.Stdout = stdout
--- a/Show More
+++ b/Show More