main: version 0.8.0

This should speed up testing.

.circleci/config.yml

@@ -18,14 +18,12 @@ commands:
             wget -O - https://apt.llvm.org/llvm-snapshot.gpg.key|sudo apt-key add -
             sudo apt-get update
             sudo apt-get install \
-                llvm \
                 python3 \
                 llvm<<parameters.llvm>>-dev \
                 clang<<parameters.llvm>> \
                 libclang<<parameters.llvm>>-dev \
                 lld<<parameters.llvm>> \
                 gcc-arm-linux-gnueabihf \
-                binutils-arm-none-eabi \
                 libc6-dev-armel-cross \
                 gcc-aarch64-linux-gnu \
                 libc6-dev-arm64-cross \
@@ -42,61 +40,229 @@ 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
-  dep:
+  llvm-source-linux:
     steps:
+      - restore_cache:
+          keys:
+            - llvm-source-8-v5
       - run:
-          name: "Install Go dependencies"
-          command: |
-            curl https://raw.githubusercontent.com/golang/dep/master/install.sh | sh
-            dep ensure --vendor-only
-  smoketest:
-    steps:
-      - run: tinygo build -size short -o test.elf -target=pca10040            examples/blinky1
-      - run: tinygo build -size short -o test.elf -target=pca10040            examples/blinky2
-      - run: tinygo build -size short -o blinky2                              examples/blinky2
-      - run: tinygo build -size short -o test.elf -target=pca10040            examples/test
-      - run: tinygo build -size short -o test.elf -target=microbit            examples/echo
-      - run: tinygo build -size short -o test.elf -target=nrf52840-mdk        examples/blinky1
-      - run: tinygo build -size short -o test.elf -target=pca10031            examples/blinky1
-      - run: tinygo build -size short -o test.elf -target=bluepill            examples/blinky1
-      - run: tinygo build -size short -o test.elf -target=arduino             examples/blinky1
-      - run: tinygo build -size short -o test.elf -target=digispark           examples/blinky1
-      - run: tinygo build -size short -o test.elf -target=reelboard           examples/blinky1
-      - run: tinygo build -size short -o test.elf -target=reelboard           examples/blinky2
-      - run: tinygo build -size short -o test.elf -target=pca10056            examples/blinky1
-      - run: tinygo build -size short -o test.elf -target=pca10056            examples/blinky2
-      - run: tinygo build -size short -o test.elf -target=itsybitsy-m0        examples/blinky1
-      - run: tinygo build -size short -o test.elf -target=circuitplay-express examples/blinky1
-
-
-jobs:
-  test-llvm7-go111:
-    docker:
-      - image: circleci/golang:1.11
-
-    working_directory: /go/src/github.com/tinygo-org/tinygo
+          name: "Fetch LLVM source"
+          command: make llvm-source
+      - save_cache:
+          key: llvm-source-8-v5
+          paths:
+            - llvm-project
+  test-linux:
     steps:
       - checkout
       - submodules
       - apt-dependencies:
-          llvm: "-7"
+          llvm: "-8"
       - install-node
       - restore_cache:
           keys:
-            - go-cache-{{ checksum "Gopkg.lock" }}-{{ .Environment.CIRCLE_PREVIOUS_BUILD_NUM }}
-            - go-cache-{{ checksum "Gopkg.lock" }}
-      - dep
+            - go-cache-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_PREVIOUS_BUILD_NUM }}
+            - go-cache-v2-{{ checksum "go.mod" }}
+      - llvm-source-linux
       - run: go install .
-      - run: make test
+      - run: go test -v ./transform .
       - run: make gen-device -j4
-      - smoketest
+      - run: make smoketest RISCV=0
       - save_cache:
-          key: go-cache-{{ checksum "Gopkg.lock" }}-{{ .Environment.CIRCLE_BUILD_NUM }}
+          key: go-cache-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_BUILD_NUM }}
           paths:
             - ~/.cache/go-build
             - ~/.cache/tinygo
+            - /go/pkg/mod
+      - run: make fmt-check
+  build-linux:
+    steps:
+      - checkout
+      - submodules
+      - run:
+          name: "Install apt dependencies"
+          command: |
+            sudo apt-get install \
+                python3 \
+                gcc-arm-linux-gnueabihf \
+                binutils-arm-none-eabi \
+                libc6-dev-armel-cross \
+                gcc-aarch64-linux-gnu \
+                libc6-dev-arm64-cross \
+                qemu-system-arm \
+                qemu-user \
+                gcc-avr \
+                avr-libc
+      - install-node
+      - restore_cache:
+          keys:
+            - go-cache-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_PREVIOUS_BUILD_NUM }}
+            - go-cache-v2-{{ checksum "go.mod" }}
+      - llvm-source-linux
+      - restore_cache:
+          keys:
+            - llvm-build-8-linux-v7
+      - 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++
+              # hack ninja to use less jobs
+              echo -e '#!/bin/sh\n/usr/bin/ninja -j3 "$@"' > /go/bin/ninja
+              chmod +x /go/bin/ninja
+              # build!
+              make llvm-build
+            fi
+      - save_cache:
+          key: llvm-build-8-linux-v7
+          paths:
+            llvm-build
+      - run:
+          name: "Create LLVM symlinks"
+          command: |
+            ln -s $PWD/llvm-build/bin/clang-8 /go/bin/clang-8
+            ln -s $PWD/llvm-build/bin/ld.lld  /go/bin/ld.lld-8
+            ln -s $PWD/llvm-build/bin/wasm-ld /go/bin/wasm-ld-8
+      - run:
+          name: "Test TinyGo"
+          command: make test
+      - run:
+          name: "Build TinyGo release"
+          command: |
+            make release -j3
+            cp -p build/release.tar.gz /tmp/tinygo.linux-amd64.tar.gz
+      - store_artifacts:
+          path: /tmp/tinygo.linux-amd64.tar.gz
+      - 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"
+          command: |
+            mkdir -p ~/lib
+            tar -C ~/lib -xf /tmp/tinygo.linux-amd64.tar.gz
+            ln -s ~/lib/tinygo/bin/tinygo /go/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-linux-ubuntu14.tar.gz
+            sudo tar -C /usr/local --strip-components=1 -xf riscv64-unknown-elf-gcc-8.2.0-2019.05.3-x86_64-linux-ubuntu14.tar.gz
+      - run: make smoketest
+  build-macos:
+    steps:
+      - checkout
+      - submodules
+      - run:
+          name: "Install dependencies"
+          command: |
+            curl https://dl.google.com/go/go1.12.5.darwin-amd64.tar.gz -o go1.12.5.darwin-amd64.tar.gz
+            sudo tar -C /usr/local -xzf go1.12.5.darwin-amd64.tar.gz
+            ln -s /usr/local/go/bin/go /usr/local/bin/go
+            HOMEBREW_NO_AUTO_UPDATE=1 brew install qemu
+      - 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-8-macos-v5
+      - run:
+          name: "Fetch LLVM source"
+          command: make llvm-source
+      - save_cache:
+          key: llvm-source-8-macos-v5
+          paths:
+            - llvm-project
+      - restore_cache:
+          keys:
+            - llvm-build-8-macos-v6
+      - run:
+          name: "Build LLVM"
+          command: |
+            if [ ! -f llvm-build/lib/liblldELF.a ]
+            then
+              # install dependencies
+              HOMEBREW_NO_AUTO_UPDATE=1 brew install cmake ninja
+              # build!
+              make llvm-build
+            fi
+      - save_cache:
+          key: llvm-build-8-macos-v6
+          paths:
+            llvm-build
+      - run:
+          name: "Create LLVM symlinks"
+          command: |
+            ln -s $PWD/llvm-build/bin/clang-8 /usr/local/bin/clang-8
+      - run:
+          name: "Test TinyGo"
+          command: make test
+      - run:
+          name: "Build TinyGo release"
+          command: |
+            make release -j3
+            cp -p build/release.tar.gz /tmp/tinygo.darwin-amd64.tar.gz
+      - store_artifacts:
+          path: /tmp/tinygo.darwin-amd64.tar.gz
+      - run:
+          name: "Extract release tarball"
+          command: |
+            mkdir -p ~/lib
+            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-llvm8-go111:
+    docker:
+      - image: circleci/golang:1.11-stretch
+    steps:
+      - test-linux
+  test-llvm8-go112:
+    docker:
+      - image: circleci/golang:1.12-stretch
+    steps:
+      - test-linux
+  build-linux:
+    docker:
+      - image: circleci/golang:1.12-stretch
+    steps:
+      - build-linux
+  build-macos:
+    macos:
+      xcode: "10.1.0"
+    steps:
+      - build-macos
+
+
+
 
 workflows:
   test-all:
     jobs:
-      - test-llvm7-go111
+      - test-llvm8-go111
+      - test-llvm8-go112
+      - build-linux
+      - build-macos

.gitignore

@@ -5,8 +5,12 @@ src/device/avr/*.ld
 src/device/avr/*.s
 src/device/nrf/*.go
 src/device/nrf/*.s
-src/device/stm32/*.go
-src/device/stm32/*.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
 vendor
+llvm
+llvm-build

.gitmodules

@@ -9,8 +9,8 @@
 	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
-	branch = release_70
+	branch = release_80

.travis.yml

@@ -1,40 +0,0 @@
-language: go
-
-matrix:
-  include:
-  - os: osx
-    go: "1.11"
-    env: PATH="/usr/local/opt/llvm/bin:$PATH"
-    before_install:
-    - mkdir -p /Users/travis/gopath/bin
-
-addons:
-  homebrew:
-    update: true
-    taps: ArmMbed/homebrew-formulae
-    packages:
-    - llvm@7
-    - qemu
-    - arm-none-eabi-gcc
-
-install:
-  - curl https://raw.githubusercontent.com/golang/dep/master/install.sh | sh
-  - dep ensure --vendor-only
-
-script:
-  - go install github.com/tinygo-org/tinygo
-  - go test -v .
-  - make gen-device
-  - tinygo build -size short -o blinky1.nrf.elf    -target=pca10040     examples/blinky1
-  - tinygo build -size short -o blinky2.nrf.elf    -target=pca10040     examples/blinky2
-  - tinygo build             -o blinky2                                 examples/blinky2 # TODO: re-enable -size flag with MachO support
-  - tinygo build -size short -o test.nrf.elf       -target=pca10040     examples/test
-  - tinygo build -size short -o blinky1.nrf51.elf  -target=microbit     examples/echo
-  - tinygo build -size short -o test.nrf.elf       -target=nrf52840-mdk examples/blinky1
-  - tinygo build -size short -o blinky1.nrf51d.elf -target=pca10031     examples/blinky1
-  - tinygo build -size short -o blinky1.stm32.elf  -target=bluepill     examples/blinky1
-  - tinygo build -size short -o blinky1.reel.elf   -target=reelboard    examples/blinky1
-  - tinygo build -size short -o blinky2.reel.elf   -target=reelboard    examples/blinky2
-  - tinygo build -size short -o blinky1.pca10056.elf    -target=pca10056     examples/blinky1
-  - tinygo build -size short -o blinky2.pca10056.elf    -target=pca10056     examples/blinky2
-  - tinygo build -size short -o blinky1.samd21.elf    -target=itsybitsy-m0     examples/blinky1

BUILDING.md

@@ -2,9 +2,9 @@
 
 TinyGo depends on LLVM and libclang, which are both big C++ libraries. It can
 also optionally use a built-in lld to ease cross compiling. There are two ways
-these can be linked: dynamically and statically. The default is dynamic linking
-because it is fast and works almost out of the box on Debian-based systems with
-the right libraries installed.
+these can be linked: dynamically and statically. An install with `go install` is
+dynamic linking because it is fast and works almost out of the box on
+Debian-based systems with the right packages installed.
 
 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
@@ -16,83 +16,54 @@ 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+)
-  * [dep](https://golang.github.io/dep/)
   * Standard build tools (gcc/clang)
-  * git or subversion
+  * git
   * CMake
-  * [Ninja](https://ninja-build.org/) or make (preferably Ninja)
+  * [Ninja](https://ninja-build.org/)
 
 The rest of this guide assumes you're running Linux, but it should be equivalent
 on a different system like Mac.
 
 ## Download the source
 
-The first step is to get the source code. Place it in some directory, assuming
-`$HOME/src` here, but you can pick a different one of course:
+The first step is to download the TinyGo sources (use `--recursive` if you clone
+the git repository). Then, inside the directory, download the LLVM source:
 
-    git clone -b release_70 https://github.com/llvm-mirror/llvm.git $HOME/src/llvm
-    git clone -b release_70 https://github.com/llvm-mirror/clang.git $HOME/src/llvm/tools/clang
-    git clone -b release_70 https://github.com/llvm-mirror/lld.git $HOME/src/llvm/tools/lld
-    go get -d github.com/tinygo-org/tinygo
-    cd $HOME/go/src/github.com/tinygo-org/tinygo
-    dep ensure -vendor-only # download dependencies
+    make llvm-source
 
-Note that Clang and LLD must be placed inside the tools subdirectory of LLVM to
-be automatically built with the rest of the system.
+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
 
-Building LLVM is quite easy compared to some other software packages. However,
-the default configuration is _not_ optimized for distribution. It is optimized
-for development, meaning that binaries produce accurate error messages at the
-cost of huge binaries and slow compiles.
-
-Before configuring, you may want to set the following environment variables to
-speed up the build. Most Linux distributions ship with GCC as the default
-compiler, but Clang is significantly faster and uses much less memory while
-producing binaries that are about as fast.
+Before starting the build, you may want to set the following environment
+variables to speed up the build. Most Linux distributions ship with GCC as the
+default compiler, but Clang is significantly faster and uses much less memory
+while producing binaries that are about as fast.
 
     export CC=clang
     export CXX=clang++
 
-Make a build directory. LLVM requires out-of-tree builds:
+The Makefile includes a default configuration that is good for most users. It
+builds a release version of LLVM (optimized, no asserts) and includes all
+targets supported by TinyGo:
 
-    mkdir $HOME/src/llvm-build
-    cd $HOME/src/llvm-build
-
-Configure LLVM with CMake:
-
-    cmake -G Ninja ../llvm "-DLLVM_TARGETS_TO_BUILD=X86;ARM;AArch64" "-DLLVM_EXPERIMENTAL_TARGETS_TO_BUILD=AVR;WebAssembly" -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=OFF -DLIBCLANG_BUILD_STATIC=ON
-
-You can also choose a different build system than Ninja, but Ninja is fast.
-
-There are various options you can tune here, but the options given above are
-preferable for releases. Here is what they do:
-
-  * `LLVM_TARGETS_TO_BUILD` and `LLVM_EXPERIMENTAL_TARGETS_TO_BUILD`: the
-    targets that are natively supported by the LLVM code generators. The targets
-    listed here are the ones supported by TinyGo. Note that LLVM is a cross
-    compiler by default, unlike some other compilers.
-  * `CMAKE_BUILD_TYPE`: the default is Debug, which produces large inefficient
-    binaries that are easy to debug. We want small and fast binaries.
-  * `LLVM_ENABLE_ASSERTIONS`: the default is ON, which greatly slows down LLVM
-    and is only really useful during development. Disable them here.
-  * `LIBCLANG_BUILD_STATIC`: unlike LLVM, libclang is built as a shared library
-    by default. We want a static library for easy distribution.
-
-Now build it:
-
-    ninja # or make, if you choose make in the previous step
+    make llvm-build
 
 This can take over an hour depending on the speed of your system.
 
 ## Build TinyGo
 
-Now that you have a working version of LLVM, build TinyGo using it. You need to
-specify the directories to the LLVM build directory and to the Clang and LLD source.
+The last step of course is to build TinyGo itself. This can again be done with
+make:
 
-    cd $HOME/go/src/github.com/tinygo-org/tinygo
-    make static LLVM_BUILDDIR=$HOME/src/llvm-build CLANG_SRC=$HOME/src/llvm/tools/clang LLD_SRC=$HOME/src/llvm/tools/lld
+    make
 
 ## Verify TinyGo
 
@@ -109,14 +80,12 @@ The result should not contain libclang or libLLVM.
 
 ## Make a release tarball
 
-Now that we have a working static build, it's time to make a release tarball.
-This is just a slight change from the command to build TinyGo:
+Now that we have a working static build, it's time to make a release tarball:
 
-    cd $HOME/go/src/github.com/tinygo-org/tinygo
-    make release LLVM_BUILDDIR=$HOME/src/llvm-build CLANG_SRC=$HOME/src/llvm/tools/clang LLD_SRC=$HOME/src/llvm/tools/lld
+    make release
 
 The release tarball is stored in build/release.tar.gz, and can be extracted with
-the following command:
+the following command (for example in ~/lib):
 
     tar -xvf path/to/release.tar.gz
 

CHANGELOG.md

@@ -1,3 +1,183 @@
+0.8.0
+---
+* **command line**
+  - fix parsing of beta Go versions
+  - check the major/minor installed version of Go before compiling
+  - validate `-target` flag better to not panic on an invalid target
+* **compiler**
+  - implement full slice expression: `s[:2:4]`
+  - fix a crash when storing a linked list in an interface
+  - fix comparing struct types by making type IDs more unique
+  - fix some bugs in IR generation
+  - add support for linked lists in reflect data
+  - implement `[]rune` to string conversion
+  - implement support for `go` on func values
+* **standard library**
+  - `reflect`: add support for named types
+  - `reflect`: add support for `t.Bits()`
+  - `reflect`: add basic support for `t.AssignableTo()`
+  - `reflect`: implement `t.Align()`
+  - `reflect`: add support for struct types
+  - `reflect`: fix bug in `v.IsNil` and `v.Pointer` for addressable values
+  - `reflect`: implement support for array types
+  - `reflect`: implement `t.Comparable()`
+  - `runtime`: implement stack-based scheduler
+  - `runtime`: fix bug in the sleep queue of the scheduler
+  - `runtime`: implement `memcpy` for Cortex-M
+  - `testing`: implement stub `testing.B` struct
+  - `testing`: add common test logging methods such as Errorf/Fatalf/Printf
+* **targets**
+  - `386`: add support for linux/386 syscalls
+  - `atsamd21`: make SPI pins configurable so that multiple SPI ports can be
+    used
+  - `atsamd21`: correct issue with invalid first reading coming from ADC
+  - `atsamd21`: add support for reset-to-bootloader using 1200baud over USB-CDC
+  - `atsamd21`: make pin selection more flexible for peripherals
+  - `atsamd21`: fix minimum delay in `time.Sleep`
+  - `atsamd51`: fix minimum delay in `time.Sleep`
+  - `nrf`: improve SPI write-only speed, by making use of double buffering
+  - `stm32f103`: fix SPI frequency selection
+  - `stm32f103`: add machine.Pin.Get method for reading GPIO values
+  - `stm32f103`: allow board specific UART usage
+  - `nucleo-f103rb`: add support for NUCLEO-F103RB board
+  - `itsybitsy-m4`: add support for this board with a SAMD51 family chip
+  - `cortex-m`: add support for `arm.SystemReset()`
+  - `gameboy-advance`: add initial support for the GameBoy Advance
+  - `wasm`: add `//go:wasm-module` magic comment to set the wasm module name
+  - `wasm`: add syscall/js.valueSetIndex support
+  - `wasm`: add syscall/js.valueInvoke support
+
+0.7.1
+---
+* **targets**
+  - `atsamd21`: add support for the `-port` flag in the flash subcommand
+
+0.7.0
+---
+* **command line**
+  - try more locations to find Clang built-in headers
+  - add support for `tinygo test`
+  - build current directory if no package is specified
+  - support custom .json target spec with `-target` flag
+  - use zversion.go to detect version of GOROOT version
+  - make initial heap size configurable for some targets (currently WebAssembly
+    only)
+* **cgo**
+  - add support for bitfields using generated getters and setters
+  - add support for anonymous structs
+* **compiler**
+  - show an error instead of panicking on duplicate function definitions
+  - allow packages like github.com/tinygo-org/tinygo/src/\* by aliasing it
+  - remove `//go:volatile` support  
+    It has been replaced with the runtime/volatile package.
+  - allow poiners in map keys
+  - support non-constant syscall numbers
+  - implement non-blocking selects
+  - add support for the `-tags` flag
+  - add support for `string` to `[]rune` conversion
+  - implement a portable conservative garbage collector (with support for wasm)
+  - add the `//go:noinline` pragma
+* **standard library**
+  - `os`: add `os.Exit` and `syscall.Exit`
+  - `os`: add several stubs
+  - `runtime`: fix heap corruption in conservative GC
+  - `runtime`: add support for math intrinsics where supported, massively
+    speeding up some benchmarks
+  - `testing`: add basic support for testing
+* **targets**
+  - add support for a generic target that calls `__tinygo_*` functions for
+    peripheral access
+  - `arduino-nano33`: add support for this board
+  - `hifive1`: add support for this RISC-V board
+  - `reelboard`: add e-paper pins
+  - `reelboard`: add `PowerSupplyActive` to enable voltage for on-board devices
+  - `wasm`: put the stack at the start of linear memory, to detect stack
+    overflows
+
+0.6.0
+---
+* **command line**
+  - some portability improvements
+  - make `$GOROOT` more robust and configurable
+  - check for Clang at the Homebrew install location as fallback
+* **compiler driver**
+  - support multiple variations of LLVM commands, for non-Debian distributions
+* **compiler**
+  - improve code quality in multiple ways
+  - make panic configurable, adding trap on panic
+  - refactor many internal parts of the compiler
+  - print all errors encountered during compilation
+  - implement calling function values of a named type
+  - implement returning values from blocking functions
+  - allow larger-than-int values to be sent across a channel
+  - implement complex arithmetic
+  - improve hashmap support
+  - add debuginfo for function arguments
+  - insert nil checks on stores (increasing code size)
+  - implement volatile operations as compiler builtins
+  - add `//go:inline` pragma
+  - add build tags for the Go stdlib version
+* **cgo**
+  - implement `char`, `enum` and `void*` types
+  - support `#include` for builtin headers
+  - improve typedef/struct/enum support
+  - only include symbols that are necessary, for broader support
+  - mark external function args as `nocapture`
+  - implement support for some `#define` constants
+  - implement support for multiple CGo files in a single package
+- **standard library**
+  - `machine`: remove microbit matrix (moved to drivers repository)
+  - `machine`: refactor pins to use `Pin` type instead of `GPIO`
+  - `runtime`: print more interface types on panic, including `error`
+* **targets**
+  - `arm`: print an error on HardFault (including stack overflows)
+  - `atsamd21`: fix a bug in the ADC peripheral
+  - `atsamd21`: add support for I2S
+  - `feather-m0`: add support for this board
+  - `nrf51`: fix a bug in I2C
+  - `stm32f103xx`: fix a bug in I2C
+  - `syscall`: implement `Exit` on unix
+  - `trinket-m0`: add support for this board
+  - `wasm`: make _main_ example smaller
+  - `wasm`: don't cache wasm file in the server, for ease of debugging
+  - `wasm`: work around bug #41508 that caused a deadlock while linking
+  - `wasm`: add support for `js.FuncOf`
+
+0.5.0
+---
+- **compiler driver**
+  - use `wasm-ld` instead of `wasm-ld-8` on macOS
+  - drop dependency on `llvm-ar`
+  - fix linker script includes when running outside `TINYGOROOT`
+- **compiler**
+  - switch to LLVM 8
+  - add support for the Go 1.12 standard library (Go 1.11 is still supported)
+  - work around lack of escape analysis due to nil checks
+  - implement casting named structs and pointers to them
+  - fix int casting to use the source signedness
+  - fix some bugs around `make([]T, …)` with uncommon index types
+  - some other optimizations
+  - support interface asserts in interp for "math/rand" support
+  - resolve all func value targets at compile time (wasm-only at the moment)
+- **cgo**
+  - improve diagnostics
+  - implement C `struct`, `union`, and arrays
+  - fix CGo-related crash in libclang
+  - implement `C.struct_` types
+- **targets**
+  - all baremetal: pretend to be linux/arm instead of js/wasm
+  - `avr`: improve `uintptr` support
+  - `cortexm`: implement memmove intrinsic generated by LLVM
+  - `cortexm`: use the lld linker instead of `arm-none-eabi-ld`
+  - `darwin`: use custom syscall package that links to libSystem.dylib
+  - `microbit`: add blink example
+  - `samd21`: support I2C1
+  - `samd21`: machine/atsamd21: correct pad/pin handling when using both UART
+     and USBCDC interfaces at same time
+  - `stm32f4discovery`: add support for this board
+  - `wasm`: support async func values
+  - `wasm`: improve documentation and add extra example
+
 0.4.1
 ---
 - **compiler**

CONTRIBUTING.md

@@ -16,12 +16,16 @@ Please open a Github issue with your problem, and we will be happy to assist.
 
 We probably have not implemented it yet. Please take a look at our [Roadmap](https://github.com/tinygo-org/tinygo/wiki/Roadmap). Your pull request adding the functionality to TinyGo would be greatly appreciated.
 
+Please open a Github issue. We want to help, and also make sure that there is no duplications of efforts. Sometimes what you need is already being worked on by someone else.
+
 A long tail of small (and large) language features haven't been implemented yet. In almost all cases, the compiler will show a `todo:` error from `compiler/compiler.go` when you try to use it. You can try implementing it, or open a bug report with a small code sample that fails to compile.
 
 ### Some specific hardware you want to use does not appear to be in TinyGo
 
 As above, we probably have not implemented it yet. Your contribution adding the hardware support to TinyGo would be greatly appreciated.
 
+Please start by opening a Github issue. We want to help you to help us to help you.
+
 Lots of targets/boards are still unsupported. Adding an architecture often requires a few compiler changes, but if the architecture is supported you can try implementing support for a new chip or board in `src/runtime`. For details, see [this wiki entry on adding archs/chips/boards](https://github.com/tinygo-org/tinygo/wiki/Adding-a-new-board).
 
 Microcontrollers have lots of peripherals (I2C, SPI, ADC, etc.) and many don't have an implementation yet in the `machine` package. Adding support for new peripherals is very useful.

CONTRIBUTORS

@@ -0,0 +1,17 @@
+# This is the official list of TinyGo authors for copyright purposes.
+#
+# This file is not actively maintained.
+# To be included, send a change adding the individual or
+# company who owns a contribution's copyright.
+#
+# Names should be added to this file as one of
+#     Organization's name
+#     Individual's name <submission email address>
+#     Individual's name <submission email address> <email2> <emailN>
+#
+# Please keep the list sorted.
+
+Ayke van Laethem <aykevanlaethem@gmail.com>
+Daniel Esteban <conejo@conejo.me>
+Loon, LLC.
+Ron Evans <ron@hybridgroup.com>

Dockerfile

@@ -1,49 +1,49 @@
-# TinyGo base stage just installs LLVM 7 and the TinyGo compiler itself.
-FROM golang:latest AS tinygo-base
+# TinyGo base stage installs Go 1.12, LLVM 8 and the TinyGo compiler itself.
+FROM golang:1.12 AS tinygo-base
 
 RUN wget -O- https://apt.llvm.org/llvm-snapshot.gpg.key| apt-key add - && \
-    echo "deb http://apt.llvm.org/stretch/ llvm-toolchain-stretch-7 main" >> /etc/apt/sources.list && \
+    echo "deb http://apt.llvm.org/buster/ llvm-toolchain-buster-8 main" >> /etc/apt/sources.list && \
     apt-get update && \
-    apt-get install -y llvm-7-dev libclang-7-dev git
+    apt-get install -y llvm-8-dev libclang-8-dev git
 
-RUN wget -O- https://raw.githubusercontent.com/golang/dep/master/install.sh | sh
+COPY . /tinygo
 
-COPY . /go/src/github.com/tinygo-org/tinygo
+# remove submodules directories and re-init them to fix any hard-coded paths
+# after copying the tinygo directory in the previous step.
+RUN cd /tinygo/ && \
+    rm -rf ./lib/* && \
+    git submodule update --init --recursive --force
 
-RUN cd /go/src/github.com/tinygo-org/tinygo/ && \
-    git submodule update --init
-
-RUN cd /go/src/github.com/tinygo-org/tinygo/ && \
-    dep ensure --vendor-only && \
-    go install /go/src/github.com/tinygo-org/tinygo/
+RUN cd /tinygo/ && \
+    go install /tinygo/
 
 # tinygo-wasm stage installs the needed dependencies to compile TinyGo programs for WASM.
 FROM tinygo-base AS tinygo-wasm
 
 COPY --from=tinygo-base /go/bin/tinygo /go/bin/tinygo
-COPY --from=tinygo-base /go/src/github.com/tinygo-org/tinygo/src /go/src/github.com/tinygo-org/tinygo/src
-COPY --from=tinygo-base /go/src/github.com/tinygo-org/tinygo/targets /go/src/github.com/tinygo-org/tinygo/targets
+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/stretch/ llvm-toolchain-stretch-7 main" >> /etc/apt/sources.list && \
+    echo "deb http://apt.llvm.org/buster/ llvm-toolchain-buster-8 main" >> /etc/apt/sources.list && \
     apt-get update && \
-    apt-get install -y libllvm7 lld-7
+    apt-get install -y libllvm8 lld-8
 
 # tinygo-avr stage installs the needed dependencies to compile TinyGo programs for AVR microcontrollers.
 FROM tinygo-base AS tinygo-avr
 
 COPY --from=tinygo-base /go/bin/tinygo /go/bin/tinygo
-COPY --from=tinygo-base /go/src/github.com/tinygo-org/tinygo/src /go/src/github.com/tinygo-org/tinygo/src
-COPY --from=tinygo-base /go/src/github.com/tinygo-org/tinygo/targets /go/src/github.com/tinygo-org/tinygo/targets
-COPY --from=tinygo-base /go/src/github.com/tinygo-org/tinygo/Makefile /go/src/github.com/tinygo-org/tinygo/
-COPY --from=tinygo-base /go/src/github.com/tinygo-org/tinygo/tools /go/src/github.com/tinygo-org/tinygo/tools
-COPY --from=tinygo-base /go/src/github.com/tinygo-org/tinygo/lib /go/src/github.com/tinygo-org/tinygo/lib
+COPY --from=tinygo-base /tinygo/src /tinygo/src
+COPY --from=tinygo-base /tinygo/targets /tinygo/targets
+COPY --from=tinygo-base /tinygo/Makefile /tinygo/
+COPY --from=tinygo-base /tinygo/tools /tinygo/tools
+COPY --from=tinygo-base /tinygo/lib /tinygo/lib
 
-RUN cd /go/src/github.com/tinygo-org/tinygo/ && \
+RUN cd /tinygo/ && \
     apt-get update && \
     apt-get install -y apt-utils python3 make binutils-avr gcc-avr avr-libc && \
     make gen-device-avr && \
-    apt-get remove -y python3 make && \
+    apt-get remove -y python3 && \
     apt-get autoremove -y && \
     apt-get clean
 
@@ -51,32 +51,32 @@ RUN cd /go/src/github.com/tinygo-org/tinygo/ && \
 FROM tinygo-base AS tinygo-arm
 
 COPY --from=tinygo-base /go/bin/tinygo /go/bin/tinygo
-COPY --from=tinygo-base /go/src/github.com/tinygo-org/tinygo/src /go/src/github.com/tinygo-org/tinygo/src
-COPY --from=tinygo-base /go/src/github.com/tinygo-org/tinygo/targets /go/src/github.com/tinygo-org/tinygo/targets
-COPY --from=tinygo-base /go/src/github.com/tinygo-org/tinygo/Makefile /go/src/github.com/tinygo-org/tinygo/
-COPY --from=tinygo-base /go/src/github.com/tinygo-org/tinygo/tools /go/src/github.com/tinygo-org/tinygo/tools
-COPY --from=tinygo-base /go/src/github.com/tinygo-org/tinygo/lib /go/src/github.com/tinygo-org/tinygo/lib
+COPY --from=tinygo-base /tinygo/src /tinygo/src
+COPY --from=tinygo-base /tinygo/targets /tinygo/targets
+COPY --from=tinygo-base /tinygo/Makefile /tinygo/
+COPY --from=tinygo-base /tinygo/tools /tinygo/tools
+COPY --from=tinygo-base /tinygo/lib /tinygo/lib
 
-RUN cd /go/src/github.com/tinygo-org/tinygo/ && \
+RUN cd /tinygo/ && \
     apt-get update && \
-    apt-get install -y apt-utils python3 make binutils-arm-none-eabi clang-7 && \
+    apt-get install -y apt-utils python3 make clang-8 && \
     make gen-device-nrf && make gen-device-stm32 && \
-    apt-get remove -y python3 make && \
+    apt-get remove -y python3 && \
     apt-get autoremove -y && \
     apt-get clean
 
 # tinygo-all stage installs the needed dependencies to compile TinyGo programs for all platforms.
 FROM tinygo-wasm AS tinygo-all
 
-COPY --from=tinygo-base /go/src/github.com/tinygo-org/tinygo/Makefile /go/src/github.com/tinygo-org/tinygo/
-COPY --from=tinygo-base /go/src/github.com/tinygo-org/tinygo/tools /go/src/github.com/tinygo-org/tinygo/tools
-COPY --from=tinygo-base /go/src/github.com/tinygo-org/tinygo/lib /go/src/github.com/tinygo-org/tinygo/lib
+COPY --from=tinygo-base /tinygo/Makefile /tinygo/
+COPY --from=tinygo-base /tinygo/tools /tinygo/tools
+COPY --from=tinygo-base /tinygo/lib /tinygo/lib
 
-RUN cd /go/src/github.com/tinygo-org/tinygo/ && \
+RUN cd /tinygo/ && \
     apt-get update && \
-    apt-get install -y apt-utils python3 make binutils-arm-none-eabi clang-7 binutils-avr gcc-avr avr-libc && \
+    apt-get install -y apt-utils python3 make clang-8 binutils-avr gcc-avr avr-libc && \
     make gen-device && \
-    apt-get remove -y python3 make && \
+    apt-get remove -y python3 && \
     apt-get autoremove -y && \
     apt-get clean
 

Gopkg.lock

@@ -1,42 +0,0 @@
-# This file is autogenerated, do not edit; changes may be undone by the next 'dep ensure'.
-
-
-[[projects]]
-  branch = "master"
-  digest = "1:00b45e06c7843541372fc17d982242bd6adfc2fc382b6f2e9ef9ce53d87a50b9"
-  name = "github.com/marcinbor85/gohex"
-  packages = ["."]
-  pruneopts = "UT"
-  revision = "7a43cd876e46e0f6ddc553f10f91731a78e6e949"
-
-[[projects]]
-  branch = "master"
-  digest = "1:ba70784a3deee74c0ca3c87bcac3c2f93d3b2d27d8f237b768c358b45ba47da8"
-  name = "golang.org/x/tools"
-  packages = [
-    "go/ast/astutil",
-    "go/ssa",
-    "go/types/typeutil",
-  ]
-  pruneopts = "UT"
-  revision = "3744606dbb67b99c60d3f11cb10bd3f9e6dad472"
-
-[[projects]]
-  branch = "master"
-  digest = "1:a6a25fd8906c74978f1ed811bc9fd3422da8093be863b458874b02a782b6ae3e"
-  name = "tinygo.org/x/go-llvm"
-  packages = ["."]
-  pruneopts = "UT"
-  revision = "d5f730401f5069618b275a5241c6417eb0c38a65"
-
-[solve-meta]
-  analyzer-name = "dep"
-  analyzer-version = 1
-  input-imports = [
-    "github.com/marcinbor85/gohex",
-    "golang.org/x/tools/go/ast/astutil",
-    "golang.org/x/tools/go/ssa",
-    "tinygo.org/x/go-llvm",
-  ]
-  solver-name = "gps-cdcl"
-  solver-version = 1

Gopkg.toml

@@ -1,11 +0,0 @@
-[[constraint]]
-  branch = "master"
-  name = "tinygo.org/x/go-llvm"
-
-[[constraint]]
-  branch = "master"
-  name = "golang.org/x/tools"
-
-[prune]
-  go-tests = true
-  unused-packages = true

LICENSE

@@ -1,5 +1,8 @@
 Copyright (c) 2018-2019 TinyGo Authors. All rights reserved.
 
+TinyGo includes portions of the Go standard library.
+Copyright (c) 2009-2019 The Go Authors. All rights reserved.
+
 Redistribution and use in source and binary forms, with or without
 modification, are permitted provided that the following conditions are
 met:

Makefile

@@ -3,92 +3,45 @@
 all: tinygo
 tinygo: build/tinygo
 
-.PHONY: all tinygo static run-test run-blinky run-blinky2 clean fmt gen-device gen-device-nrf gen-device-avr
+# Default build and source directories, as created by `make llvm-build`.
+LLVM_BUILDDIR ?= llvm-build
+CLANG_SRC ?= llvm-project/clang
+LLD_SRC ?= llvm-project/lld
 
-TARGET ?= unix
-
-ifeq ($(TARGET),unix)
-# Regular *nix system.
-
-else ifeq ($(TARGET),pca10040)
-# PCA10040: nRF52832 development board
-OBJCOPY = arm-none-eabi-objcopy
-TGOFLAGS += -target $(TARGET)
-
-else ifeq ($(TARGET),microbit)
-# BBC micro:bit
-OBJCOPY = arm-none-eabi-objcopy
-TGOFLAGS += -target $(TARGET)
-
-else ifeq ($(TARGET),reelboard)
-# reel board
-OBJCOPY = arm-none-eabi-objcopy
-TGOFLAGS += -target $(TARGET)
-
-else ifeq ($(TARGET),bluepill)
-# "blue pill" development board
-# See: https://wiki.stm32duino.com/index.php?title=Blue_Pill
-OBJCOPY = arm-none-eabi-objcopy
-TGOFLAGS += -target $(TARGET)
-
-else ifeq ($(TARGET),arduino)
-OBJCOPY = avr-objcopy
-TGOFLAGS += -target $(TARGET)
-
-else
-$(error Unknown target)
-
-endif
+.PHONY: all tinygo build/tinygo test $(LLVM_BUILDDIR) llvm-source clean fmt gen-device gen-device-nrf gen-device-avr
 
 LLVM_COMPONENTS = all-targets analysis asmparser asmprinter bitreader bitwriter codegen core coroutines debuginfodwarf executionengine instrumentation interpreter ipo irreader linker lto mc mcjit objcarcopts option profiledata scalaropts support target
 
-CLANG_LIBS = -Wl,--start-group $(abspath $(LLVM_BUILDDIR))/lib/libclang.a -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 -lclangToolingRefactor -Wl,--end-group -lstdc++
+UNAME_S := $(shell uname -s)
+ifeq ($(UNAME_S),Linux)
+    START_GROUP = -Wl,--start-group
+    END_GROUP = -Wl,--end-group
+endif
 
-LLD_LIBS = -Wl,--start-group -llldCOFF -llldCommon -llldCore -llldDriver -llldELF -llldMachO -llldMinGW -llldReaderWriter -llldWasm -llldYAML -Wl,--end-group
+CLANG_LIBS = $(START_GROUP) $(abspath $(LLVM_BUILDDIR))/lib/libclang.a -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++
+
+LLD_LIBS = $(START_GROUP) -llldCOFF -llldCommon -llldCore -llldDriver -llldELF -llldMachO -llldMinGW -llldReaderWriter -llldWasm -llldYAML $(END_GROUP)
 
 
 # For static linking.
-CGO_CPPFLAGS=$(shell $(LLVM_BUILDDIR)/bin/llvm-config --cppflags) -I$(abspath $(CLANG_SRC))/include -I$(abspath $(LLD_SRC))/include
-CGO_CXXFLAGS=-std=c++11
-CGO_LDFLAGS=-L$(LLVM_BUILDDIR)/lib $(CLANG_LIBS) $(LLD_LIBS) $(shell $(LLVM_BUILDDIR)/bin/llvm-config --ldflags --libs --system-libs $(LLVM_COMPONENTS))
-
-
-
-run-test: build/test
-	./build/test
-
-run-blinky: run-blinky2
-run-blinky2: build/blinky2
-	./build/blinky2
-
-ifeq ($(TARGET),pca10040)
-flash-%: build/%.hex
-	nrfjprog -f nrf52 --sectorerase --program $< --reset
-else ifeq ($(TARGET),microbit)
-flash-%: build/%.hex
-	openocd -f interface/cmsis-dap.cfg -f target/nrf51.cfg -c 'program $< reset exit'
-else ifeq ($(TARGET),reelboard)
-flash-%: build/%.hex
-	openocd -f interface/cmsis-dap.cfg -f target/nrf51.cfg -c 'program $< reset exit'
-else ifeq ($(TARGET),arduino)
-flash-%: build/%.hex
-	avrdude -c arduino -p atmega328p -P /dev/ttyACM0 -U flash:w:$<
-else ifeq ($(TARGET),bluepill)
-flash-%: build/%.hex
-	openocd -f interface/stlink-v2.cfg -f target/stm32f1x.cfg -c 'program $< reset exit'
+ifneq ("$(wildcard $(LLVM_BUILDDIR)/bin/llvm-config)","")
+    CGO_CPPFLAGS=$(shell $(LLVM_BUILDDIR)/bin/llvm-config --cppflags) -I$(abspath $(CLANG_SRC))/include -I$(abspath $(LLD_SRC))/include
+    CGO_CXXFLAGS=-std=c++11
+    CGO_LDFLAGS=-L$(LLVM_BUILDDIR)/lib $(CLANG_LIBS) $(LLD_LIBS) $(shell $(LLVM_BUILDDIR)/bin/llvm-config --ldflags --libs --system-libs $(LLVM_COMPONENTS))
 endif
 
+
 clean:
 	@rm -rf build
 
+FMT_PATHS = ./*.go cgo compiler interp ir loader src/device/arm src/examples src/machine src/os src/reflect src/runtime src/sync src/syscall
 fmt:
-	@go fmt . ./compiler ./interp ./loader ./ir ./src/device/arm ./src/examples/* ./src/machine ./src/os ./src/runtime ./src/sync
-	@go fmt ./testdata/*.go
+	@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
 
-test:
-	@go test -v .
 
-gen-device: gen-device-avr gen-device-nrf gen-device-sam gen-device-stm32
+gen-device: gen-device-avr gen-device-nrf gen-device-sam gen-device-sifive gen-device-stm32
 
 gen-device-avr:
 	./tools/gen-device-avr.py lib/avr/packs/atmega src/device/avr/
@@ -103,27 +56,99 @@ gen-device-sam:
 	./tools/gen-device-svd.py lib/cmsis-svd/data/Atmel/ src/device/sam/ --source=https://github.com/posborne/cmsis-svd/tree/master/data/Atmel
 	go fmt ./src/device/sam
 
+gen-device-sifive:
+	./tools/gen-device-svd.py lib/cmsis-svd/data/SiFive-Community/ src/device/sifive/ --source=https://github.com/AdaCore/svd2ada/tree/master/CMSIS-SVD/SiFive-Community
+	go fmt ./src/device/sifive
+
 gen-device-stm32:
 	./tools/gen-device-svd.py lib/cmsis-svd/data/STMicro/ src/device/stm32/ --source=https://github.com/posborne/cmsis-svd/tree/master/data/STMicro
 	go fmt ./src/device/stm32
 
+
+# Get LLVM sources.
+llvm-project/README.md:
+	git clone -b release/8.x https://github.com/llvm/llvm-project
+llvm-source: llvm-project/README.md
+
+# Configure LLVM.
+TINYGO_SOURCE_DIR=$(shell pwd)
+$(LLVM_BUILDDIR)/build.ninja: llvm-source
+	mkdir -p $(LLVM_BUILDDIR); cd $(LLVM_BUILDDIR); cmake -G Ninja $(TINYGO_SOURCE_DIR)/llvm-project/llvm "-DLLVM_TARGETS_TO_BUILD=X86;ARM;AArch64;WebAssembly" "-DLLVM_EXPERIMENTAL_TARGETS_TO_BUILD=AVR;RISCV" -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=OFF -DLIBCLANG_BUILD_STATIC=ON -DLLVM_ENABLE_TERMINFO=OFF -DLLVM_ENABLE_ZLIB=OFF -DLLVM_ENABLE_PROJECTS="clang;lld" -DLLVM_TOOL_CLANG_TOOLS_EXTRA_BUILD=OFF
+
+# Build LLVM.
+$(LLVM_BUILDDIR): $(LLVM_BUILDDIR)/build.ninja
+	cd $(LLVM_BUILDDIR); ninja
+
+
 # Build the Go compiler.
 build/tinygo:
-	@mkdir -p build
-	go build -o build/tinygo .
-
-static:
+	@if [ ! -f "$(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 -tags byollvm .
 
-release: static gen-device
+test:
+	CGO_CPPFLAGS="$(CGO_CPPFLAGS)" CGO_CXXFLAGS="$(CGO_CXXFLAGS)" CGO_LDFLAGS="$(CGO_LDFLAGS)" go test -v -tags byollvm ./transform .
+
+tinygo-test:
+	cd tests/tinygotest && tinygo test
+
+.PHONY: smoketest
+smoketest:
+	# test all examples
+	tinygo build -size short -o test.elf -target=pca10040            examples/blinky1
+	tinygo build -size short -o test.elf -target=pca10040            examples/adc
+	tinygo build -size short -o test.elf -target=pca10040            examples/blinkm
+	tinygo build -size short -o test.elf -target=pca10040            examples/blinky2
+	tinygo build -size short -o test.elf -target=pca10040            examples/button
+	tinygo build -size short -o test.elf -target=pca10040            examples/button2
+	tinygo build -size short -o test.elf -target=pca10040            examples/echo
+	tinygo build -size short -o test.elf -target=circuitplay-express examples/i2s
+	tinygo build -size short -o test.elf -target=pca10040            examples/mcp3008
+	tinygo build -size short -o test.elf -target=microbit            examples/microbit-blink
+	tinygo build -size short -o test.elf -target=pca10040            examples/pwm
+	tinygo build -size short -o test.elf -target=pca10040            examples/serial
+	tinygo build -size short -o test.elf -target=pca10040            examples/test
+	# test all targets/boards
+	tinygo build             -o test.wasm -tags=pca10040             examples/blinky2
+	tinygo build -size short -o test.elf -target=microbit            examples/echo
+	tinygo build -size short -o test.elf -target=nrf52840-mdk        examples/blinky1
+	tinygo build -size short -o test.elf -target=pca10031            examples/blinky1
+	tinygo build -size short -o test.elf -target=bluepill            examples/blinky1
+	tinygo build -size short -o test.elf -target=reelboard           examples/blinky1
+	tinygo build -size short -o test.elf -target=reelboard           examples/blinky2
+	tinygo build -size short -o test.elf -target=pca10056            examples/blinky1
+	tinygo build -size short -o test.elf -target=pca10056            examples/blinky2
+	tinygo build -size short -o test.elf -target=itsybitsy-m0        examples/blinky1
+	tinygo build -size short -o test.elf -target=feather-m0          examples/blinky1
+	tinygo build -size short -o test.elf -target=trinket-m0          examples/blinky1
+	tinygo build -size short -o test.elf -target=circuitplay-express examples/blinky1
+	tinygo build -size short -o test.elf -target=stm32f4disco        examples/blinky1
+	tinygo build -size short -o test.elf -target=stm32f4disco        examples/blinky2
+	tinygo build -size short -o test.elf -target=circuitplay-express examples/i2s
+	tinygo build -size short -o test.elf -target=gameboy-advance     examples/gba-display
+	tinygo build -size short -o test.elf -target=itsybitsy-m4        examples/blinky1
+	tinygo build -size short -o test.elf -target=nucleo-f103rb       examples/blinky1
+ifneq ($(AVR), 0)
+	tinygo build -size short -o test.elf -target=arduino             examples/blinky1
+	tinygo build -size short -o test.elf -target=digispark           examples/blinky1
+endif
+ifneq ($(RISCV), 0)
+	tinygo build -size short -o test.elf -target=hifive1b            examples/blinky1
+endif
+	tinygo build             -o wasm.wasm -target=wasm               examples/wasm/export
+	tinygo build             -o wasm.wasm -target=wasm               examples/wasm/main
+
+release: build/tinygo gen-device
 	@mkdir -p build/release/tinygo/bin
+	@mkdir -p build/release/tinygo/lib/clang/include
 	@mkdir -p build/release/tinygo/lib/CMSIS/CMSIS
 	@mkdir -p build/release/tinygo/lib/compiler-rt/lib
 	@mkdir -p build/release/tinygo/lib/nrfx
 	@mkdir -p build/release/tinygo/pkg/armv6m-none-eabi
 	@mkdir -p build/release/tinygo/pkg/armv7m-none-eabi
 	@mkdir -p build/release/tinygo/pkg/armv7em-none-eabi
+	@echo copying source files
 	@cp -p  build/tinygo                 build/release/tinygo/bin
+	@cp -p $(abspath $(CLANG_SRC))/lib/Headers/*.h build/release/tinygo/lib/clang/include
 	@cp -rp lib/CMSIS/CMSIS/Include      build/release/tinygo/lib/CMSIS/CMSIS
 	@cp -rp lib/CMSIS/README.md          build/release/tinygo/lib/CMSIS
 	@cp -rp lib/compiler-rt/lib/builtins build/release/tinygo/lib/compiler-rt/lib
@@ -136,15 +161,3 @@ release: static gen-device
 	./build/tinygo build-builtins -target=armv7m-none-eabi  -o build/release/tinygo/pkg/armv7m-none-eabi/compiler-rt.a
 	./build/tinygo build-builtins -target=armv7em-none-eabi -o build/release/tinygo/pkg/armv7em-none-eabi/compiler-rt.a
 	tar -czf build/release.tar.gz -C build/release tinygo
-
-# Binary that can run on the host.
-build/%: src/examples/% src/examples/%/*.go build/tinygo src/runtime/*.go
-	./build/tinygo build $(TGOFLAGS) -size=short -o $@ $(subst src/,,$<)
-
-# ELF file that can run on a microcontroller.
-build/%.elf: src/examples/% src/examples/%/*.go build/tinygo src/runtime/*.go
-	./build/tinygo build $(TGOFLAGS) -size=short -o $@ $(subst src/,,$<)
-
-# Convert executable to Intel hex file (for flashing).
-build/%.hex: build/%.elf
-	$(OBJCOPY) -O ihex $^ $@

README.md

@@ -1,9 +1,8 @@
 # TinyGo - Go compiler for small places
 
-[![Travis CI](https://travis-ci.com/tinygo-org/tinygo.svg?branch=dev)](https://travis-ci.com/tinygo-org/tinygo)
 [![CircleCI](https://circleci.com/gh/tinygo-org/tinygo/tree/dev.svg?style=svg)](https://circleci.com/gh/tinygo-org/tinygo/tree/dev)
 
-TinyGo is a Go compiler intended for use in small places such as microcontrollers, WebAssembly (WASM), and command-line tools.
+TinyGo is a Go compiler intended for use in small places such as microcontrollers, WebAssembly (Wasm), and command-line tools.
 
 It reuses libraries used by the [Go language tools](https://golang.org/pkg/go/) alongside [LLVM](http://llvm.org) to provide an alternative way to compile programs written in the Go programming language.
 
@@ -18,8 +17,8 @@ import (
 )
 
 func main() {
-    led := machine.GPIO{machine.LED}
-    led.Configure(machine.GPIOConfig{Mode: machine.GPIO_OUTPUT})
+    led := machine.LED
+    led.Configure(machine.PinConfig{Mode: machine.PinOutput})
     for {
         led.Low()
         time.Sleep(time.Millisecond * 1000)
@@ -44,13 +43,20 @@ See the [getting started instructions](https://tinygo.org/getting-started/) for
 
 You can compile TinyGo programs for microcontrollers, WebAssembly and Linux.
 
-The following microcontroller boards are currently supported:
+The following 18 microcontroller boards are currently supported:
 
 * [Adafruit Circuit Playground Express](https://www.adafruit.com/product/3333)
+* [Adafruit Feather M0](https://www.adafruit.com/product/2772)
 * [Adafruit ItsyBitsy M0](https://www.adafruit.com/product/3727)
+* [Adafruit ItsyBitsy M4](https://www.adafruit.com/product/3800)
+* [Adafruit Trinket M0](https://www.adafruit.com/product/3500)
+* [Arduino Nano33 IoT](https://store.arduino.cc/nano-33-iot)
 * [Arduino Uno](https://store.arduino.cc/arduino-uno-rev3)
-* [BBC:Microbit](https://microbit.org/)
+* [BBC micro:bit](https://microbit.org/)
+* [Game Boy Advance](https://en.wikipedia.org/wiki/Game_Boy_Advance)
 * [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 STM32F407 "Discovery"](https://www.st.com/en/evaluation-tools/stm32f4discovery.html)
 * [Digispark](http://digistump.com/products/1)
 * [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)

buildcache.go

@@ -9,9 +9,11 @@ import (
 
 // Get the cache directory, usually ~/.cache/tinygo
 func cacheDir() string {
-	home := getHomeDir()
-	dir := filepath.Join(home, ".cache", "tinygo")
-	return dir
+	dir, err := os.UserCacheDir()
+	if err != nil {
+		panic("could not find cache dir: " + err.Error())
+	}
+	return filepath.Join(dir, "tinygo")
 }
 
 // Return the newest timestamp of all the file paths passed in. Used to check

builtins.go

@@ -1,11 +1,15 @@
 package main
 
 import (
+	"errors"
+	"io"
 	"io/ioutil"
 	"os"
-	"os/exec"
 	"path/filepath"
 	"strings"
+	"time"
+
+	"github.com/blakesmith/ar"
 )
 
 // These are the GENERIC_SOURCES according to CMakeList.txt.
@@ -187,13 +191,13 @@ func loadBuiltins(target string) (path string, err error) {
 		srcs[i] = filepath.Join(builtinsDir, name)
 	}
 
-	if path, err := cacheLoad(outfile, commands["clang"], srcs); path != "" || err != nil {
+	if path, err := cacheLoad(outfile, commands["clang"][0], srcs); path != "" || err != nil {
 		return path, err
 	}
 
 	var cachepath string
 	err = compileBuiltins(target, func(path string) error {
-		path, err := cacheStore(path, outfile, commands["clang"], srcs)
+		path, err := cacheStore(path, outfile, commands["clang"][0], srcs)
 		cachepath = path
 		return err
 	})
@@ -235,28 +239,56 @@ func compileBuiltins(target string, callback func(path string) error) error {
 		// 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.
-		cmd := exec.Command(commands["clang"], "-c", "-Oz", "-g", "-Werror", "-Wall", "-std=c11", "-fshort-enums", "-nostdlibinc", "-ffunction-sections", "-fdata-sections", "--target="+target, "-fdebug-prefix-map="+dir+"="+remapDir, "-o", objpath, srcpath)
-		cmd.Stdout = os.Stdout
-		cmd.Stderr = os.Stderr
-		cmd.Dir = dir
-		err = cmd.Run()
+		err := execCommand(commands["clang"], "-c", "-Oz", "-g", "-Werror", "-Wall", "-std=c11", "-fshort-enums", "-nostdlibinc", "-ffunction-sections", "-fdata-sections", "--target="+target, "-fdebug-prefix-map="+dir+"="+remapDir, "-o", objpath, srcpath)
 		if err != nil {
 			return &commandError{"failed to build", srcpath, err}
 		}
 	}
 
 	// Put all builtins in an archive to link as a static library.
+	// Note: this does not create a symbol index, but ld.lld doesn't seem to
+	// care.
 	arpath := filepath.Join(dir, "librt.a")
-	cmd := exec.Command(commands["ar"], append([]string{"cr", arpath}, objs...)...)
-	cmd.Stdout = os.Stdout
-	cmd.Stderr = os.Stderr
-	cmd.Dir = dir
-	err = cmd.Run()
+	arfile, err := os.Create(arpath)
 	if err != nil {
-		return &commandError{"failed to make static library", arpath, err}
+		return err
+	}
+	defer arfile.Close()
+	arwriter := ar.NewWriter(arfile)
+	err = arwriter.WriteGlobalHeader()
+	if err != nil {
+		return &os.PathError{"write ar header", arpath, err}
+	}
+	for _, objpath := range objs {
+		name := filepath.Base(objpath)
+		objfile, err := os.Open(objpath)
+		if err != nil {
+			return err
+		}
+		defer objfile.Close()
+		st, err := objfile.Stat()
+		if err != nil {
+			return err
+		}
+		arwriter.WriteHeader(&ar.Header{
+			Name:    name,
+			ModTime: time.Unix(0, 0),
+			Uid:     0,
+			Gid:     0,
+			Mode:    0644,
+			Size:    st.Size(),
+		})
+		n, err := io.Copy(arwriter, objfile)
+		if err != nil {
+			return err
+		}
+		if n != st.Size() {
+			return errors.New("file modified during ar creation: " + arpath)
+		}
 	}
 
 	// Give the caller the resulting file. The callback must copy the file,
 	// because after it returns the temporary directory will be removed.
+	arfile.Close()
 	return callback(arpath)
 }

cgo/cgo.go

@@ -0,0 +1,997 @@
+// Package cgo implements CGo by modifying a loaded AST. It does this by parsing
+// the `import "C"` statements found in the source code with libclang and
+// generating stub function and global declarations.
+//
+// There are a few advantages to modifying the AST directly instead of doing CGo
+// as a preprocessing step, with the main advantage being that debug information
+// is kept intact as much as possible.
+package cgo
+
+// This file extracts the `import "C"` statement from the source and modifies
+// the AST for CGo. It does not use libclang directly: see libclang.go for the C
+// source file parsing.
+
+import (
+	"go/ast"
+	"go/token"
+	"sort"
+	"strconv"
+	"strings"
+
+	"golang.org/x/tools/go/ast/astutil"
+)
+
+// cgoPackage holds all CGo-related information of a package.
+type cgoPackage struct {
+	generated       *ast.File
+	generatedPos    token.Pos
+	errors          []error
+	dir             string
+	fset            *token.FileSet
+	tokenFiles      map[string]*token.File
+	missingSymbols  map[string]struct{}
+	constants       map[string]constantInfo
+	functions       map[string]*functionInfo
+	globals         map[string]globalInfo
+	typedefs        map[string]*typedefInfo
+	elaboratedTypes map[string]*elaboratedTypeInfo
+	enums           map[string]enumInfo
+	anonStructNum   int
+}
+
+// constantInfo stores some information about a CGo constant found by libclang
+// and declared in the Go AST.
+type constantInfo struct {
+	expr *ast.BasicLit
+	pos  token.Pos
+}
+
+// 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
+}
+
+// paramInfo is a parameter of a CGo function (see functionInfo).
+type paramInfo struct {
+	name     string
+	typeExpr ast.Expr
+}
+
+// typedefInfo contains information about a single typedef in C.
+type typedefInfo struct {
+	typeExpr ast.Expr
+	pos      token.Pos
+}
+
+// elaboratedTypeInfo contains some information about an elaborated type
+// (struct, union) found in the C AST.
+type elaboratedTypeInfo struct {
+	typeExpr  *ast.StructType
+	pos       token.Pos
+	bitfields []bitfieldInfo
+}
+
+// bitfieldInfo contains information about a single bitfield in a struct. It
+// keeps information about the start, end, and the special (renamed) base field
+// of this bitfield.
+type bitfieldInfo struct {
+	field    *ast.Field
+	name     string
+	pos      token.Pos
+	startBit int64
+	endBit   int64 // may be 0 meaning "until the end of the field"
+}
+
+// enumInfo contains information about an enum in the C.
+type enumInfo struct {
+	typeExpr ast.Expr
+	pos      token.Pos
+}
+
+// globalInfo contains information about a declared global variable in C.
+type globalInfo struct {
+	typeExpr ast.Expr
+	pos      token.Pos
+}
+
+// cgoAliases list type aliases between Go and C, for types that are equivalent
+// in both languages. See addTypeAliases.
+var cgoAliases = map[string]string{
+	"C.int8_t":    "int8",
+	"C.int16_t":   "int16",
+	"C.int32_t":   "int32",
+	"C.int64_t":   "int64",
+	"C.uint8_t":   "uint8",
+	"C.uint16_t":  "uint16",
+	"C.uint32_t":  "uint32",
+	"C.uint64_t":  "uint64",
+	"C.uintptr_t": "uintptr",
+}
+
+// builtinAliases are handled specially because they only exist on the Go side
+// of CGo, not on the CGo side (they're prefixed with "_Cgo_" there).
+var builtinAliases = map[string]struct{}{
+	"char":      struct{}{},
+	"schar":     struct{}{},
+	"uchar":     struct{}{},
+	"short":     struct{}{},
+	"ushort":    struct{}{},
+	"int":       struct{}{},
+	"uint":      struct{}{},
+	"long":      struct{}{},
+	"ulong":     struct{}{},
+	"longlong":  struct{}{},
+	"ulonglong": struct{}{},
+}
+
+// cgoTypes lists some C types with ambiguous sizes that must be retrieved
+// somehow from C. This is done by adding some typedefs to get the size of each
+// type.
+const cgoTypes = `
+typedef char                _Cgo_char;
+typedef signed char         _Cgo_schar;
+typedef unsigned char       _Cgo_uchar;
+typedef short               _Cgo_short;
+typedef unsigned short      _Cgo_ushort;
+typedef int                 _Cgo_int;
+typedef unsigned int        _Cgo_uint;
+typedef long                _Cgo_long;
+typedef unsigned long       _Cgo_ulong;
+typedef long long           _Cgo_longlong;
+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) {
+	p := &cgoPackage{
+		dir:             dir,
+		fset:            fset,
+		tokenFiles:      map[string]*token.File{},
+		missingSymbols:  map[string]struct{}{},
+		constants:       map[string]constantInfo{},
+		functions:       map[string]*functionInfo{},
+		globals:         map[string]globalInfo{},
+		typedefs:        map[string]*typedefInfo{},
+		elaboratedTypes: map[string]*elaboratedTypeInfo{},
+		enums:           map[string]enumInfo{},
+	}
+
+	// Add a new location for the following file.
+	generatedTokenPos := p.fset.AddFile(dir+"/!cgo.go", -1, 0)
+	generatedTokenPos.SetLines([]int{0})
+	p.generatedPos = generatedTokenPos.Pos(0)
+
+	// Construct a new in-memory AST for CGo declarations of this package.
+	unsafeImport := &ast.ImportSpec{
+		Path: &ast.BasicLit{
+			ValuePos: p.generatedPos,
+			Kind:     token.STRING,
+			Value:    "\"unsafe\"",
+		},
+		EndPos: p.generatedPos,
+	}
+	p.generated = &ast.File{
+		Package: p.generatedPos,
+		Name: &ast.Ident{
+			NamePos: p.generatedPos,
+			Name:    files[0].Name.Name,
+		},
+		Decls: []ast.Decl{
+			&ast.GenDecl{
+				TokPos: p.generatedPos,
+				Tok:    token.IMPORT,
+				Specs: []ast.Spec{
+					unsafeImport,
+				},
+			},
+		},
+		Imports: []*ast.ImportSpec{unsafeImport},
+	}
+
+	// Find all C.* symbols.
+	for _, f := range files {
+		astutil.Apply(f, p.findMissingCGoNames, nil)
+	}
+	for name := range builtinAliases {
+		p.missingSymbols["_Cgo_"+name] = struct{}{}
+	}
+
+	// Find `import "C"` statements in the file.
+	for _, f := range files {
+		for i := 0; i < len(f.Decls); i++ {
+			decl := f.Decls[i]
+			genDecl, ok := decl.(*ast.GenDecl)
+			if !ok {
+				continue
+			}
+			if len(genDecl.Specs) != 1 {
+				continue
+			}
+			spec, ok := genDecl.Specs[0].(*ast.ImportSpec)
+			if !ok {
+				continue
+			}
+			path, err := strconv.Unquote(spec.Path.Value)
+			if err != nil {
+				panic("could not parse import path: " + err.Error())
+			}
+			if path != "C" {
+				continue
+			}
+			cgoComment := genDecl.Doc.Text()
+
+			pos := genDecl.Pos()
+			if genDecl.Doc != nil {
+				pos = genDecl.Doc.Pos()
+			}
+			position := fset.PositionFor(pos, true)
+			p.parseFragment(cgoComment+cgoTypes, cflags, position.Filename, position.Line)
+
+			// Remove this import declaration.
+			f.Decls = append(f.Decls[:i], f.Decls[i+1:]...)
+			i--
+		}
+
+		// Print the AST, for debugging.
+		//ast.Print(fset, f)
+	}
+
+	// Declare functions found by libclang.
+	p.addFuncDecls()
+
+	// Declare stub function pointer values found by libclang.
+	p.addFuncPtrDecls()
+
+	// Declare globals found by libclang.
+	p.addConstDecls()
+
+	// Declare globals found by libclang.
+	p.addVarDecls()
+
+	// Forward C types to Go types (like C.uint32_t -> uint32).
+	p.addTypeAliases()
+
+	// Add type declarations for C types, declared using typedef in C.
+	p.addTypedefs()
+
+	// Add elaborated types for C structs and unions.
+	p.addElaboratedTypes()
+
+	// Add enum types and enum constants for C enums.
+	p.addEnumTypes()
+
+	// Patch the AST to use the declared types and functions.
+	for _, f := range files {
+		astutil.Apply(f, p.walker, nil)
+	}
+
+	// Print the newly generated in-memory AST, for debugging.
+	//ast.Print(fset, p.generated)
+
+	return p.generated, p.errors
+}
+
+// addFuncDecls adds the C function declarations found by libclang in the
+// comment above the `import "C"` statement.
+func (p *cgoPackage) addFuncDecls() {
+	names := make([]string, 0, len(p.functions))
+	for name := range p.functions {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	for _, name := range names {
+		fn := p.functions[name]
+		obj := &ast.Object{
+			Kind: ast.Fun,
+			Name: "C." + name,
+		}
+		args := make([]*ast.Field, len(fn.args))
+		decl := &ast.FuncDecl{
+			Name: &ast.Ident{
+				NamePos: fn.pos,
+				Name:    "C." + name,
+				Obj:     obj,
+			},
+			Type: &ast.FuncType{
+				Func: fn.pos,
+				Params: &ast.FieldList{
+					Opening: fn.pos,
+					List:    args,
+					Closing: fn.pos,
+				},
+				Results: fn.results,
+			},
+		}
+		obj.Decl = decl
+		for i, arg := range fn.args {
+			args[i] = &ast.Field{
+				Names: []*ast.Ident{
+					&ast.Ident{
+						NamePos: fn.pos,
+						Name:    arg.name,
+						Obj: &ast.Object{
+							Kind: ast.Var,
+							Name: arg.name,
+							Decl: decl,
+						},
+					},
+				},
+				Type: arg.typeExpr,
+			}
+		}
+		p.generated.Decls = append(p.generated.Decls, decl)
+	}
+}
+
+// addFuncPtrDecls creates stub declarations of function pointer values. These
+// values will later be replaced with the real values in the compiler.
+// It adds code like the following to the AST:
+//
+//     var (
+//         C.add unsafe.Pointer
+//         C.mul unsafe.Pointer
+//         // ...
+//     )
+func (p *cgoPackage) addFuncPtrDecls() {
+	if len(p.functions) == 0 {
+		return
+	}
+	gen := &ast.GenDecl{
+		TokPos: token.NoPos,
+		Tok:    token.VAR,
+		Lparen: token.NoPos,
+		Rparen: token.NoPos,
+	}
+	names := make([]string, 0, len(p.functions))
+	for name := range p.functions {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	for _, name := range names {
+		fn := p.functions[name]
+		obj := &ast.Object{
+			Kind: ast.Typ,
+			Name: "C." + name + "$funcaddr",
+		}
+		valueSpec := &ast.ValueSpec{
+			Names: []*ast.Ident{&ast.Ident{
+				NamePos: fn.pos,
+				Name:    "C." + name + "$funcaddr",
+				Obj:     obj,
+			}},
+			Type: &ast.SelectorExpr{
+				X: &ast.Ident{
+					NamePos: fn.pos,
+					Name:    "unsafe",
+				},
+				Sel: &ast.Ident{
+					NamePos: fn.pos,
+					Name:    "Pointer",
+				},
+			},
+		}
+		obj.Decl = valueSpec
+		gen.Specs = append(gen.Specs, valueSpec)
+	}
+	p.generated.Decls = append(p.generated.Decls, gen)
+}
+
+// addConstDecls declares external C constants in the Go source.
+// It adds code like the following to the AST:
+//
+//     const (
+//         C.CONST_INT = 5
+//         C.CONST_FLOAT = 5.8
+//         // ...
+//     )
+func (p *cgoPackage) addConstDecls() {
+	if len(p.constants) == 0 {
+		return
+	}
+	gen := &ast.GenDecl{
+		TokPos: token.NoPos,
+		Tok:    token.CONST,
+		Lparen: token.NoPos,
+		Rparen: token.NoPos,
+	}
+	names := make([]string, 0, len(p.constants))
+	for name := range p.constants {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	for _, name := range names {
+		constVal := p.constants[name]
+		obj := &ast.Object{
+			Kind: ast.Con,
+			Name: "C." + name,
+		}
+		valueSpec := &ast.ValueSpec{
+			Names: []*ast.Ident{&ast.Ident{
+				NamePos: constVal.pos,
+				Name:    "C." + name,
+				Obj:     obj,
+			}},
+			Values: []ast.Expr{constVal.expr},
+		}
+		obj.Decl = valueSpec
+		gen.Specs = append(gen.Specs, valueSpec)
+	}
+	p.generated.Decls = append(p.generated.Decls, gen)
+}
+
+// addVarDecls declares external C globals in the Go source.
+// It adds code like the following to the AST:
+//
+//     var (
+//         C.globalInt  int
+//         C.globalBool bool
+//         // ...
+//     )
+func (p *cgoPackage) addVarDecls() {
+	if len(p.globals) == 0 {
+		return
+	}
+	gen := &ast.GenDecl{
+		TokPos: token.NoPos,
+		Tok:    token.VAR,
+		Lparen: token.NoPos,
+		Rparen: token.NoPos,
+	}
+	names := make([]string, 0, len(p.globals))
+	for name := range p.globals {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	for _, name := range names {
+		global := p.globals[name]
+		obj := &ast.Object{
+			Kind: ast.Var,
+			Name: "C." + name,
+		}
+		valueSpec := &ast.ValueSpec{
+			Names: []*ast.Ident{&ast.Ident{
+				NamePos: global.pos,
+				Name:    "C." + name,
+				Obj:     obj,
+			}},
+			Type: global.typeExpr,
+		}
+		obj.Decl = valueSpec
+		gen.Specs = append(gen.Specs, valueSpec)
+	}
+	p.generated.Decls = append(p.generated.Decls, gen)
+}
+
+// addTypeAliases aliases some built-in Go types with their equivalent C types.
+// It adds code like the following to the AST:
+//
+//     type (
+//         C.int8_t  = int8
+//         C.int16_t = int16
+//         // ...
+//     )
+func (p *cgoPackage) addTypeAliases() {
+	aliasKeys := make([]string, 0, len(cgoAliases))
+	for key := range cgoAliases {
+		aliasKeys = append(aliasKeys, key)
+	}
+	sort.Strings(aliasKeys)
+	gen := &ast.GenDecl{
+		TokPos: token.NoPos,
+		Tok:    token.TYPE,
+		Lparen: token.NoPos,
+		Rparen: token.NoPos,
+	}
+	for _, typeName := range aliasKeys {
+		goTypeName := cgoAliases[typeName]
+		obj := &ast.Object{
+			Kind: ast.Typ,
+			Name: typeName,
+		}
+		typeSpec := &ast.TypeSpec{
+			Name: &ast.Ident{
+				NamePos: token.NoPos,
+				Name:    typeName,
+				Obj:     obj,
+			},
+			Assign: p.generatedPos,
+			Type: &ast.Ident{
+				NamePos: token.NoPos,
+				Name:    goTypeName,
+			},
+		}
+		obj.Decl = typeSpec
+		gen.Specs = append(gen.Specs, typeSpec)
+	}
+	p.generated.Decls = append(p.generated.Decls, gen)
+}
+
+func (p *cgoPackage) addTypedefs() {
+	if len(p.typedefs) == 0 {
+		return
+	}
+	gen := &ast.GenDecl{
+		TokPos: token.NoPos,
+		Tok:    token.TYPE,
+	}
+	names := make([]string, 0, len(p.typedefs))
+	for name := range p.typedefs {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	for _, name := range names {
+		typedef := p.typedefs[name]
+		typeName := "C." + name
+		isAlias := true
+		if strings.HasPrefix(name, "_Cgo_") {
+			typeName = "C." + name[len("_Cgo_"):]
+			isAlias = false // C.short etc. should not be aliased to the equivalent Go type (not portable)
+		}
+		if _, ok := cgoAliases[typeName]; ok {
+			// This is a type that also exists in Go (defined in stdint.h).
+			continue
+		}
+		obj := &ast.Object{
+			Kind: ast.Typ,
+			Name: typeName,
+		}
+		typeSpec := &ast.TypeSpec{
+			Name: &ast.Ident{
+				NamePos: typedef.pos,
+				Name:    typeName,
+				Obj:     obj,
+			},
+			Type: typedef.typeExpr,
+		}
+		if isAlias {
+			typeSpec.Assign = typedef.pos
+		}
+		obj.Decl = typeSpec
+		gen.Specs = append(gen.Specs, typeSpec)
+	}
+	p.generated.Decls = append(p.generated.Decls, gen)
+}
+
+// addElaboratedTypes adds C elaborated types as aliases. These are the "struct
+// foo" or "union foo" types, often used in a typedef.
+//
+// See also:
+// https://en.cppreference.com/w/cpp/language/elaborated_type_specifier
+func (p *cgoPackage) addElaboratedTypes() {
+	if len(p.elaboratedTypes) == 0 {
+		return
+	}
+	gen := &ast.GenDecl{
+		TokPos: token.NoPos,
+		Tok:    token.TYPE,
+	}
+	names := make([]string, 0, len(p.elaboratedTypes))
+	for name := range p.elaboratedTypes {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	for _, name := range names {
+		typ := p.elaboratedTypes[name]
+		typeName := "C." + name
+		obj := &ast.Object{
+			Kind: ast.Typ,
+			Name: typeName,
+		}
+		typeSpec := &ast.TypeSpec{
+			Name: &ast.Ident{
+				NamePos: typ.pos,
+				Name:    typeName,
+				Obj:     obj,
+			},
+			Type: typ.typeExpr,
+		}
+		obj.Decl = typeSpec
+		gen.Specs = append(gen.Specs, typeSpec)
+		// If this struct has bitfields, create getters for them.
+		for _, bitfield := range typ.bitfields {
+			p.createBitfieldGetter(bitfield, typeName)
+			p.createBitfieldSetter(bitfield, typeName)
+		}
+	}
+	p.generated.Decls = append(p.generated.Decls, gen)
+}
+
+// createBitfieldGetter creates a bitfield getter function like the following:
+//
+//     func (s *C.struct_foo) bitfield_b() byte {
+//         return (s.__bitfield_1 >> 5) & 0x1
+//     }
+func (p *cgoPackage) createBitfieldGetter(bitfield bitfieldInfo, typeName string) {
+	// The value to return from the getter.
+	// Not complete: this is just an expression to get the complete field.
+	var result ast.Expr = &ast.SelectorExpr{
+		X: &ast.Ident{
+			NamePos: bitfield.pos,
+			Name:    "s",
+			Obj:     nil,
+		},
+		Sel: &ast.Ident{
+			NamePos: bitfield.pos,
+			Name:    bitfield.field.Names[0].Name,
+		},
+	}
+	if bitfield.startBit != 0 {
+		// Shift to the right by .startBit so that fields that come before are
+		// shifted off.
+		result = &ast.BinaryExpr{
+			X:     result,
+			OpPos: bitfield.pos,
+			Op:    token.SHR,
+			Y: &ast.BasicLit{
+				ValuePos: bitfield.pos,
+				Kind:     token.INT,
+				Value:    strconv.FormatInt(bitfield.startBit, 10),
+			},
+		}
+	}
+	if bitfield.endBit != 0 {
+		// Mask off the high bits so that fields that come after this field are
+		// masked off.
+		and := (uint64(1) << uint64(bitfield.endBit-bitfield.startBit)) - 1
+		result = &ast.BinaryExpr{
+			X:     result,
+			OpPos: bitfield.pos,
+			Op:    token.AND,
+			Y: &ast.BasicLit{
+				ValuePos: bitfield.pos,
+				Kind:     token.INT,
+				Value:    "0x" + strconv.FormatUint(and, 16),
+			},
+		}
+	}
+
+	// Create the getter function.
+	getter := &ast.FuncDecl{
+		Recv: &ast.FieldList{
+			Opening: bitfield.pos,
+			List: []*ast.Field{
+				&ast.Field{
+					Names: []*ast.Ident{
+						&ast.Ident{
+							NamePos: bitfield.pos,
+							Name:    "s",
+							Obj: &ast.Object{
+								Kind: ast.Var,
+								Name: "s",
+								Decl: nil,
+							},
+						},
+					},
+					Type: &ast.StarExpr{
+						Star: bitfield.pos,
+						X: &ast.Ident{
+							NamePos: bitfield.pos,
+							Name:    typeName,
+							Obj:     nil,
+						},
+					},
+				},
+			},
+			Closing: bitfield.pos,
+		},
+		Name: &ast.Ident{
+			NamePos: bitfield.pos,
+			Name:    "bitfield_" + bitfield.name,
+		},
+		Type: &ast.FuncType{
+			Func: bitfield.pos,
+			Params: &ast.FieldList{
+				Opening: bitfield.pos,
+				Closing: bitfield.pos,
+			},
+			Results: &ast.FieldList{
+				List: []*ast.Field{
+					&ast.Field{
+						Type: bitfield.field.Type,
+					},
+				},
+			},
+		},
+		Body: &ast.BlockStmt{
+			Lbrace: bitfield.pos,
+			List: []ast.Stmt{
+				&ast.ReturnStmt{
+					Return: bitfield.pos,
+					Results: []ast.Expr{
+						result,
+					},
+				},
+			},
+			Rbrace: bitfield.pos,
+		},
+	}
+	p.generated.Decls = append(p.generated.Decls, getter)
+}
+
+// createBitfieldSetter creates a bitfield setter function like the following:
+//
+//     func (s *C.struct_foo) set_bitfield_b(value byte) {
+//         s.__bitfield_1 = s.__bitfield_1 ^ 0x60 | ((value & 1) << 5)
+//     }
+//
+// Or the following:
+//
+//     func (s *C.struct_foo) set_bitfield_c(value byte) {
+//         s.__bitfield_1 = s.__bitfield_1 & 0x3f | (value << 6)
+//     }
+func (p *cgoPackage) createBitfieldSetter(bitfield bitfieldInfo, typeName string) {
+	// The full field with all bitfields.
+	var field ast.Expr = &ast.SelectorExpr{
+		X: &ast.Ident{
+			NamePos: bitfield.pos,
+			Name:    "s",
+			Obj:     nil,
+		},
+		Sel: &ast.Ident{
+			NamePos: bitfield.pos,
+			Name:    bitfield.field.Names[0].Name,
+		},
+	}
+	// The value to insert into the field.
+	var valueToInsert ast.Expr = &ast.Ident{
+		NamePos: bitfield.pos,
+		Name:    "value",
+	}
+
+	if bitfield.endBit != 0 {
+		// Make sure the value is in range with a mask.
+		valueToInsert = &ast.BinaryExpr{
+			X:     valueToInsert,
+			OpPos: bitfield.pos,
+			Op:    token.AND,
+			Y: &ast.BasicLit{
+				ValuePos: bitfield.pos,
+				Kind:     token.INT,
+				Value:    "0x" + strconv.FormatUint((uint64(1)<<uint64(bitfield.endBit-bitfield.startBit))-1, 16),
+			},
+		}
+		// Create a mask for the AND NOT operation.
+		mask := ((uint64(1) << uint64(bitfield.endBit-bitfield.startBit)) - 1) << uint64(bitfield.startBit)
+		// Zero the bits in the field that will soon be inserted.
+		field = &ast.BinaryExpr{
+			X:     field,
+			OpPos: bitfield.pos,
+			Op:    token.AND_NOT,
+			Y: &ast.BasicLit{
+				ValuePos: bitfield.pos,
+				Kind:     token.INT,
+				Value:    "0x" + strconv.FormatUint(mask, 16),
+			},
+		}
+	} else { // bitfield.endBit == 0
+		// We don't know exactly how many high bits should be zeroed. So we do
+		// something different: keep the low bits with a mask and OR the new
+		// value with it.
+		mask := (uint64(1) << uint64(bitfield.startBit)) - 1
+		// Extract the lower bits.
+		field = &ast.BinaryExpr{
+			X:     field,
+			OpPos: bitfield.pos,
+			Op:    token.AND,
+			Y: &ast.BasicLit{
+				ValuePos: bitfield.pos,
+				Kind:     token.INT,
+				Value:    "0x" + strconv.FormatUint(mask, 16),
+			},
+		}
+	}
+
+	// Bitwise OR with the new value (after the new value has been shifted).
+	field = &ast.BinaryExpr{
+		X:     field,
+		OpPos: bitfield.pos,
+		Op:    token.OR,
+		Y: &ast.BinaryExpr{
+			X:     valueToInsert,
+			OpPos: bitfield.pos,
+			Op:    token.SHL,
+			Y: &ast.BasicLit{
+				ValuePos: bitfield.pos,
+				Kind:     token.INT,
+				Value:    strconv.FormatInt(bitfield.startBit, 10),
+			},
+		},
+	}
+
+	// Create the setter function.
+	setter := &ast.FuncDecl{
+		Recv: &ast.FieldList{
+			Opening: bitfield.pos,
+			List: []*ast.Field{
+				&ast.Field{
+					Names: []*ast.Ident{
+						&ast.Ident{
+							NamePos: bitfield.pos,
+							Name:    "s",
+							Obj: &ast.Object{
+								Kind: ast.Var,
+								Name: "s",
+								Decl: nil,
+							},
+						},
+					},
+					Type: &ast.StarExpr{
+						Star: bitfield.pos,
+						X: &ast.Ident{
+							NamePos: bitfield.pos,
+							Name:    typeName,
+							Obj:     nil,
+						},
+					},
+				},
+			},
+			Closing: bitfield.pos,
+		},
+		Name: &ast.Ident{
+			NamePos: bitfield.pos,
+			Name:    "set_bitfield_" + bitfield.name,
+		},
+		Type: &ast.FuncType{
+			Func: bitfield.pos,
+			Params: &ast.FieldList{
+				Opening: bitfield.pos,
+				List: []*ast.Field{
+					&ast.Field{
+						Names: []*ast.Ident{
+							&ast.Ident{
+								NamePos: bitfield.pos,
+								Name:    "value",
+								Obj:     nil,
+							},
+						},
+						Type: bitfield.field.Type,
+					},
+				},
+				Closing: bitfield.pos,
+			},
+		},
+		Body: &ast.BlockStmt{
+			Lbrace: bitfield.pos,
+			List: []ast.Stmt{
+				&ast.AssignStmt{
+					Lhs: []ast.Expr{
+						&ast.SelectorExpr{
+							X: &ast.Ident{
+								NamePos: bitfield.pos,
+								Name:    "s",
+								Obj:     nil,
+							},
+							Sel: &ast.Ident{
+								NamePos: bitfield.pos,
+								Name:    bitfield.field.Names[0].Name,
+							},
+						},
+					},
+					TokPos: bitfield.pos,
+					Tok:    token.ASSIGN,
+					Rhs: []ast.Expr{
+						field,
+					},
+				},
+			},
+			Rbrace: bitfield.pos,
+		},
+	}
+	p.generated.Decls = append(p.generated.Decls, setter)
+}
+
+// addEnumTypes adds C enums to the AST. For example, the following C code:
+//
+//     enum option {
+//         optionA,
+//         optionB = 5,
+//     };
+//
+// is translated to the following Go code equivalent:
+//
+//     type C.enum_option int32
+//
+// The constants are treated just like macros so are inserted into the AST by
+// addConstDecls.
+// See also: https://en.cppreference.com/w/c/language/enum
+func (p *cgoPackage) addEnumTypes() {
+	if len(p.enums) == 0 {
+		return
+	}
+	gen := &ast.GenDecl{
+		TokPos: token.NoPos,
+		Tok:    token.TYPE,
+	}
+	names := make([]string, 0, len(p.enums))
+	for name := range p.enums {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	for _, name := range names {
+		typ := p.enums[name]
+		typeName := "C.enum_" + name
+		obj := &ast.Object{
+			Kind: ast.Typ,
+			Name: typeName,
+		}
+		typeSpec := &ast.TypeSpec{
+			Name: &ast.Ident{
+				NamePos: typ.pos,
+				Name:    typeName,
+				Obj:     obj,
+			},
+			Type: typ.typeExpr,
+		}
+		obj.Decl = typeSpec
+		gen.Specs = append(gen.Specs, typeSpec)
+	}
+	p.generated.Decls = append(p.generated.Decls, gen)
+}
+
+// findMissingCGoNames traverses the AST and finds all C.something names. Only
+// these symbols are extracted from the parsed C AST and converted to the Go
+// equivalent.
+func (p *cgoPackage) findMissingCGoNames(cursor *astutil.Cursor) bool {
+	switch node := cursor.Node().(type) {
+	case *ast.SelectorExpr:
+		x, ok := node.X.(*ast.Ident)
+		if !ok {
+			return true
+		}
+		if x.Name == "C" {
+			name := node.Sel.Name
+			if _, ok := builtinAliases[name]; ok {
+				name = "_Cgo_" + name
+			}
+			p.missingSymbols[name] = struct{}{}
+		}
+	}
+	return true
+}
+
+// walker replaces all "C".<something> expressions to literal "C.<something>"
+// expressions. Such expressions are impossible to write in Go (a dot cannot be
+// used in the middle of a name) so in practice all C identifiers live in a
+// separate namespace (no _Cgo_ hacks like in gc).
+func (p *cgoPackage) walker(cursor *astutil.Cursor) bool {
+	switch node := cursor.Node().(type) {
+	case *ast.CallExpr:
+		fun, ok := node.Fun.(*ast.SelectorExpr)
+		if !ok {
+			return true
+		}
+		x, ok := fun.X.(*ast.Ident)
+		if !ok {
+			return true
+		}
+		if _, ok := p.functions[fun.Sel.Name]; ok && x.Name == "C" {
+			node.Fun = &ast.Ident{
+				NamePos: x.NamePos,
+				Name:    "C." + fun.Sel.Name,
+			}
+		}
+	case *ast.SelectorExpr:
+		x, ok := node.X.(*ast.Ident)
+		if !ok {
+			return true
+		}
+		if x.Name == "C" {
+			name := "C." + node.Sel.Name
+			if _, ok := p.functions[node.Sel.Name]; ok {
+				name += "$funcaddr"
+			}
+			cursor.Replace(&ast.Ident{
+				NamePos: x.NamePos,
+				Name:    name,
+			})
+		}
+	}
+	return true
+}

cgo/libclang.go

@@ -0,0 +1,763 @@
+package cgo
+
+// This file parses a fragment of C with libclang and stores the result for AST
+// modification. It does not touch the AST itself.
+
+import (
+	"fmt"
+	"go/ast"
+	"go/scanner"
+	"go/token"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"unsafe"
+)
+
+/*
+#include <clang-c/Index.h> // if this fails, install libclang-8-dev
+#include <stdlib.h>
+#include <stdint.h>
+
+// This struct should be ABI-compatible on all platforms (uintptr_t has the same
+// alignment etc. as void*) but does not include void* pointers that are not
+// always real pointers.
+// The Go garbage collector assumes that all non-nil pointer-typed integers are
+// actually pointers. This is not always true, as data[1] often contains 0x1,
+// which is clearly not a valid pointer. Usually the GC won't catch this issue,
+// but occasionally it will leading to a crash with a vague error message.
+typedef struct {
+	enum CXCursorKind kind;
+	int xdata;
+	uintptr_t data[3];
+} GoCXCursor;
+
+// Forwarding functions. They are implemented in libclang_stubs.c and forward to
+// the real functions without doing anything else, thus they are entirely
+// compatible with the versions without tinygo_ prefix. The only difference is
+// the CXCursor type, which has been replaced with GoCXCursor.
+GoCXCursor tinygo_clang_getTranslationUnitCursor(CXTranslationUnit tu);
+unsigned tinygo_clang_visitChildren(GoCXCursor parent, CXCursorVisitor visitor, CXClientData client_data);
+CXString tinygo_clang_getCursorSpelling(GoCXCursor c);
+enum CXCursorKind tinygo_clang_getCursorKind(GoCXCursor c);
+CXType tinygo_clang_getCursorType(GoCXCursor c);
+GoCXCursor tinygo_clang_getTypeDeclaration(CXType t);
+CXType tinygo_clang_getTypedefDeclUnderlyingType(GoCXCursor c);
+CXType tinygo_clang_getCursorResultType(GoCXCursor c);
+int tinygo_clang_Cursor_getNumArguments(GoCXCursor c);
+GoCXCursor tinygo_clang_Cursor_getArgument(GoCXCursor c, unsigned i);
+CXSourceLocation tinygo_clang_getCursorLocation(GoCXCursor c);
+CXSourceRange tinygo_clang_getCursorExtent(GoCXCursor c);
+CXTranslationUnit tinygo_clang_Cursor_getTranslationUnit(GoCXCursor c);
+long long tinygo_clang_getEnumConstantDeclValue(GoCXCursor c);
+CXType tinygo_clang_getEnumDeclIntegerType(GoCXCursor c);
+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);
+*/
+import "C"
+
+// storedRefs stores references to types, used for clang_visitChildren.
+var storedRefs refMap
+
+var diagnosticSeverity = [...]string{
+	C.CXDiagnostic_Ignored: "ignored",
+	C.CXDiagnostic_Note:    "note",
+	C.CXDiagnostic_Warning: "warning",
+	C.CXDiagnostic_Error:   "error",
+	C.CXDiagnostic_Fatal:   "fatal",
+}
+
+func (p *cgoPackage) parseFragment(fragment string, cflags []string, posFilename string, posLine int) {
+	index := C.clang_createIndex(0, 0)
+	defer C.clang_disposeIndex(index)
+
+	// pretend to be a .c file
+	filenameC := C.CString(posFilename + "!cgo.c")
+	defer C.free(unsafe.Pointer(filenameC))
+
+	// fix up error locations
+	fragment = fmt.Sprintf("# %d %#v\n", posLine+1, posFilename) + fragment
+
+	fragmentC := C.CString(fragment)
+	defer C.free(unsafe.Pointer(fragmentC))
+
+	unsavedFile := C.struct_CXUnsavedFile{
+		Filename: filenameC,
+		Length:   C.ulong(len(fragment)),
+		Contents: fragmentC,
+	}
+
+	// convert Go slice of strings to C array of strings.
+	cmdargsC := C.malloc(C.size_t(len(cflags)) * C.size_t(unsafe.Sizeof(uintptr(0))))
+	defer C.free(cmdargsC)
+	cmdargs := (*[1 << 16]*C.char)(cmdargsC)
+	for i, cflag := range cflags {
+		s := C.CString(cflag)
+		cmdargs[i] = s
+		defer C.free(unsafe.Pointer(s))
+	}
+
+	var unit C.CXTranslationUnit
+	errCode := C.clang_parseTranslationUnit2(
+		index,
+		filenameC,
+		(**C.char)(cmdargsC), C.int(len(cflags)), // command line args
+		&unsavedFile, 1, // unsaved files
+		C.CXTranslationUnit_DetailedPreprocessingRecord,
+		&unit)
+	if errCode != 0 {
+		panic("loader: failed to parse source with libclang")
+	}
+	defer C.clang_disposeTranslationUnit(unit)
+
+	if numDiagnostics := int(C.clang_getNumDiagnostics(unit)); numDiagnostics != 0 {
+		addDiagnostic := func(diagnostic C.CXDiagnostic) {
+			spelling := getString(C.clang_getDiagnosticSpelling(diagnostic))
+			severity := diagnosticSeverity[C.clang_getDiagnosticSeverity(diagnostic)]
+			location := C.clang_getDiagnosticLocation(diagnostic)
+			var libclangFilename C.CXString
+			var line C.unsigned
+			var column C.unsigned
+			C.clang_getPresumedLocation(location, &libclangFilename, &line, &column)
+			filename := getString(libclangFilename)
+			if filepath.IsAbs(filename) {
+				// Relative paths for readability, like other Go parser errors.
+				relpath, err := filepath.Rel(p.dir, filename)
+				if err == nil {
+					filename = relpath
+				}
+			}
+			p.errors = append(p.errors, &scanner.Error{
+				Pos: token.Position{
+					Filename: filename,
+					Offset:   0, // not provided by clang_getPresumedLocation
+					Line:     int(line),
+					Column:   int(column),
+				},
+				Msg: severity + ": " + spelling,
+			})
+		}
+		for i := 0; i < numDiagnostics; i++ {
+			diagnostic := C.clang_getDiagnostic(unit, C.uint(i))
+			addDiagnostic(diagnostic)
+
+			// Child diagnostics (like notes on redefinitions).
+			diagnostics := C.clang_getChildDiagnostics(diagnostic)
+			for j := 0; j < int(C.clang_getNumDiagnosticsInSet(diagnostics)); j++ {
+				addDiagnostic(C.clang_getDiagnosticInSet(diagnostics, C.uint(j)))
+			}
+		}
+		return
+	}
+
+	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))
+}
+
+//export tinygo_clang_globals_visitor
+func tinygo_clang_globals_visitor(c, parent C.GoCXCursor, client_data C.CXClientData) C.int {
+	p := storedRefs.Get(unsafe.Pointer(client_data)).(*cgoPackage)
+	kind := C.tinygo_clang_getCursorKind(c)
+	pos := p.getCursorPosition(c)
+	switch kind {
+	case C.CXCursor_FunctionDecl:
+		name := getString(C.tinygo_clang_getCursorSpelling(c))
+		if _, required := p.missingSymbols[name]; !required {
+			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,
+		}
+		p.functions[name] = fn
+		for i := 0; i < numArgs; i++ {
+			arg := C.tinygo_clang_Cursor_getArgument(c, C.uint(i))
+			argName := getString(C.tinygo_clang_getCursorSpelling(arg))
+			argType := C.clang_getArgType(cursorType, C.uint(i))
+			if argName == "" {
+				argName = "$" + strconv.Itoa(i)
+			}
+			fn.args = append(fn.args, paramInfo{
+				name:     argName,
+				typeExpr: p.makeASTType(argType, pos),
+			})
+		}
+		resultType := C.tinygo_clang_getCursorResultType(c)
+		if resultType.kind != C.CXType_Void {
+			fn.results = &ast.FieldList{
+				List: []*ast.Field{
+					&ast.Field{
+						Type: p.makeASTType(resultType, pos),
+					},
+				},
+			}
+		}
+	case C.CXCursor_StructDecl:
+		typ := C.tinygo_clang_getCursorType(c)
+		name := getString(C.tinygo_clang_getCursorSpelling(c))
+		if _, required := p.missingSymbols["struct_"+name]; !required {
+			return C.CXChildVisit_Continue
+		}
+		p.makeASTType(typ, pos)
+	case C.CXCursor_TypedefDecl:
+		typedefType := C.tinygo_clang_getCursorType(c)
+		name := getString(C.clang_getTypedefName(typedefType))
+		if _, required := p.missingSymbols[name]; !required {
+			return C.CXChildVisit_Continue
+		}
+		p.makeASTType(typedefType, pos)
+	case C.CXCursor_VarDecl:
+		name := getString(C.tinygo_clang_getCursorSpelling(c))
+		if _, required := p.missingSymbols[name]; !required {
+			return C.CXChildVisit_Continue
+		}
+		cursorType := C.tinygo_clang_getCursorType(c)
+		p.globals[name] = globalInfo{
+			typeExpr: p.makeASTType(cursorType, pos),
+			pos:      pos,
+		}
+	case C.CXCursor_MacroDefinition:
+		name := getString(C.tinygo_clang_getCursorSpelling(c))
+		if _, required := p.missingSymbols[name]; !required {
+			return C.CXChildVisit_Continue
+		}
+		sourceRange := C.tinygo_clang_getCursorExtent(c)
+		start := C.clang_getRangeStart(sourceRange)
+		end := C.clang_getRangeEnd(sourceRange)
+		var file, endFile C.CXFile
+		var startOffset, endOffset C.unsigned
+		C.clang_getExpansionLocation(start, &file, nil, nil, &startOffset)
+		if file == nil {
+			panic("could not find file where macro is defined")
+		}
+		C.clang_getExpansionLocation(end, &endFile, nil, nil, &endOffset)
+		if file != endFile {
+			panic("expected start and end location of a #define to be in the same file")
+		}
+		if startOffset > endOffset {
+			panic("startOffset > endOffset")
+		}
+
+		// read file contents and extract the relevant byte range
+		tu := C.tinygo_clang_Cursor_getTranslationUnit(c)
+		var size C.size_t
+		sourcePtr := C.clang_getFileContents(tu, file, &size)
+		if endOffset >= C.uint(size) {
+			panic("endOffset lies after end of file")
+		}
+		source := string(((*[1 << 28]byte)(unsafe.Pointer(sourcePtr)))[startOffset:endOffset:endOffset])
+		if !strings.HasPrefix(source, name) {
+			panic(fmt.Sprintf("expected #define value to start with %#v, got %#v", name, source))
+		}
+		value := strings.TrimSpace(source[len(name):])
+		for len(value) != 0 && value[0] == '(' && value[len(value)-1] == ')' {
+			value = strings.TrimSpace(value[1 : len(value)-1])
+		}
+		if len(value) == 0 {
+			// Pretend it doesn't exist at all.
+			return C.CXChildVisit_Continue
+		}
+		// For information about integer literals:
+		// https://en.cppreference.com/w/cpp/language/integer_literal
+		if value[0] == '"' {
+			// string constant
+			p.constants[name] = constantInfo{&ast.BasicLit{pos, token.STRING, value}, pos}
+			return C.CXChildVisit_Continue
+		}
+		if value[0] == '\'' {
+			// char constant
+			p.constants[name] = constantInfo{&ast.BasicLit{pos, token.CHAR, value}, pos}
+			return C.CXChildVisit_Continue
+		}
+		// assume it's a number (int or float)
+		value = strings.Replace(value, "'", "", -1) // remove ' chars
+		value = strings.TrimRight(value, "lu")      // remove llu suffixes etc.
+		// find the first non-number
+		nonnum := byte(0)
+		for i := 0; i < len(value); i++ {
+			if value[i] < '0' || value[i] > '9' {
+				nonnum = value[i]
+				break
+			}
+		}
+		// determine number type based on the first non-number
+		switch nonnum {
+		case 0:
+			// no non-number found, must be an integer
+			p.constants[name] = constantInfo{&ast.BasicLit{pos, token.INT, value}, pos}
+		case 'x', 'X':
+			// hex integer constant
+			// TODO: may also be a floating point number per C++17.
+			p.constants[name] = constantInfo{&ast.BasicLit{pos, token.INT, value}, pos}
+		case '.', 'e':
+			// float constant
+			value = strings.TrimRight(value, "fFlL")
+			p.constants[name] = constantInfo{&ast.BasicLit{pos, token.FLOAT, value}, pos}
+		default:
+			// unknown type, ignore
+		}
+	}
+	return C.CXChildVisit_Continue
+}
+
+func getString(clangString C.CXString) (s string) {
+	rawString := C.clang_getCString(clangString)
+	s = C.GoString(rawString)
+	C.clang_disposeString(clangString)
+	return
+}
+
+// getCursorPosition returns a usable token.Pos from a libclang cursor. If the
+// file for this cursor has not been seen before, it is read from libclang
+// (which already has the file in memory) and added to the token.FileSet.
+func (p *cgoPackage) getCursorPosition(cursor C.GoCXCursor) token.Pos {
+	location := C.tinygo_clang_getCursorLocation(cursor)
+	var file C.CXFile
+	var line C.unsigned
+	var column C.unsigned
+	var offset C.unsigned
+	C.clang_getExpansionLocation(location, &file, &line, &column, &offset)
+	if line == 0 || file == nil {
+		// Invalid token.
+		return token.NoPos
+	}
+	filename := getString(C.clang_getFileName(file))
+	if _, ok := p.tokenFiles[filename]; !ok {
+		// File has not been seen before in this package, add line information
+		// now by reading the file from libclang.
+		tu := C.tinygo_clang_Cursor_getTranslationUnit(cursor)
+		var size C.size_t
+		sourcePtr := C.clang_getFileContents(tu, file, &size)
+		source := ((*[1 << 28]byte)(unsafe.Pointer(sourcePtr)))[:size:size]
+		lines := []int{0}
+		for i := 0; i < len(source)-1; i++ {
+			if source[i] == '\n' {
+				lines = append(lines, i+1)
+			}
+		}
+		f := p.fset.AddFile(filename, -1, int(size))
+		f.SetLines(lines)
+		p.tokenFiles[filename] = f
+	}
+	return p.tokenFiles[filename].Pos(int(offset))
+}
+
+// makeASTType return the ast.Expr for the given libclang type. In other words,
+// it converts a libclang type to a type in the Go AST.
+func (p *cgoPackage) makeASTType(typ C.CXType, pos token.Pos) ast.Expr {
+	var typeName string
+	switch typ.kind {
+	case C.CXType_Char_S, C.CXType_Char_U:
+		typeName = "C.char"
+	case C.CXType_SChar:
+		typeName = "C.schar"
+	case C.CXType_UChar:
+		typeName = "C.uchar"
+	case C.CXType_Short:
+		typeName = "C.short"
+	case C.CXType_UShort:
+		typeName = "C.ushort"
+	case C.CXType_Int:
+		typeName = "C.int"
+	case C.CXType_UInt:
+		typeName = "C.uint"
+	case C.CXType_Long:
+		typeName = "C.long"
+	case C.CXType_ULong:
+		typeName = "C.ulong"
+	case C.CXType_LongLong:
+		typeName = "C.longlong"
+	case C.CXType_ULongLong:
+		typeName = "C.ulonglong"
+	case C.CXType_Bool:
+		typeName = "bool"
+	case C.CXType_Float, C.CXType_Double, C.CXType_LongDouble:
+		switch C.clang_Type_getSizeOf(typ) {
+		case 4:
+			typeName = "float32"
+		case 8:
+			typeName = "float64"
+		default:
+			// Don't do anything, rely on the fallback code to show a somewhat
+			// sensible error message like "undeclared name: C.long double".
+		}
+	case C.CXType_Complex:
+		switch C.clang_Type_getSizeOf(typ) {
+		case 8:
+			typeName = "complex64"
+		case 16:
+			typeName = "complex128"
+		}
+	case C.CXType_Pointer:
+		pointeeType := C.clang_getPointeeType(typ)
+		if pointeeType.kind == C.CXType_Void {
+			// void* type is translated to Go as unsafe.Pointer
+			return &ast.SelectorExpr{
+				X: &ast.Ident{
+					NamePos: pos,
+					Name:    "unsafe",
+				},
+				Sel: &ast.Ident{
+					NamePos: pos,
+					Name:    "Pointer",
+				},
+			}
+		}
+		return &ast.StarExpr{
+			Star: pos,
+			X:    p.makeASTType(pointeeType, pos),
+		}
+	case C.CXType_ConstantArray:
+		return &ast.ArrayType{
+			Lbrack: pos,
+			Len: &ast.BasicLit{
+				ValuePos: pos,
+				Kind:     token.INT,
+				Value:    strconv.FormatInt(int64(C.clang_getArraySize(typ)), 10),
+			},
+			Elt: p.makeASTType(C.clang_getElementType(typ), pos),
+		}
+	case C.CXType_FunctionProto:
+		// Be compatible with gc, which uses the *[0]byte type for function
+		// pointer types.
+		// Return type [0]byte because this is a function type, not a pointer to
+		// this function type.
+		return &ast.ArrayType{
+			Lbrack: pos,
+			Len: &ast.BasicLit{
+				ValuePos: pos,
+				Kind:     token.INT,
+				Value:    "0",
+			},
+			Elt: &ast.Ident{
+				NamePos: pos,
+				Name:    "byte",
+			},
+		}
+	case C.CXType_Typedef:
+		name := getString(C.clang_getTypedefName(typ))
+		if _, ok := p.typedefs[name]; !ok {
+			p.typedefs[name] = nil // don't recurse
+			c := C.tinygo_clang_getTypeDeclaration(typ)
+			underlyingType := C.tinygo_clang_getTypedefDeclUnderlyingType(c)
+			expr := p.makeASTType(underlyingType, pos)
+			if strings.HasPrefix(name, "_Cgo_") {
+				expr := expr.(*ast.Ident)
+				typeSize := C.clang_Type_getSizeOf(underlyingType)
+				switch expr.Name {
+				case "C.char":
+					if typeSize != 1 {
+						// This happens for some very special purpose architectures
+						// (DSPs etc.) that are not currently targeted.
+						// https://www.embecosm.com/2017/04/18/non-8-bit-char-support-in-clang-and-llvm/
+						panic("unknown char width")
+					}
+					switch underlyingType.kind {
+					case C.CXType_Char_S:
+						expr.Name = "int8"
+					case C.CXType_Char_U:
+						expr.Name = "uint8"
+					}
+				case "C.schar", "C.short", "C.int", "C.long", "C.longlong":
+					switch typeSize {
+					case 1:
+						expr.Name = "int8"
+					case 2:
+						expr.Name = "int16"
+					case 4:
+						expr.Name = "int32"
+					case 8:
+						expr.Name = "int64"
+					}
+				case "C.uchar", "C.ushort", "C.uint", "C.ulong", "C.ulonglong":
+					switch typeSize {
+					case 1:
+						expr.Name = "uint8"
+					case 2:
+						expr.Name = "uint16"
+					case 4:
+						expr.Name = "uint32"
+					case 8:
+						expr.Name = "uint64"
+					}
+				}
+			}
+			p.typedefs[name] = &typedefInfo{
+				typeExpr: expr,
+				pos:      pos,
+			}
+		}
+		return &ast.Ident{
+			NamePos: pos,
+			Name:    "C." + name,
+		}
+	case C.CXType_Elaborated:
+		underlying := C.clang_Type_getNamedType(typ)
+		switch underlying.kind {
+		case C.CXType_Record:
+			return p.makeASTType(underlying, pos)
+		case C.CXType_Enum:
+			return p.makeASTType(underlying, pos)
+		default:
+			panic("unknown elaborated type")
+		}
+	case C.CXType_Record:
+		cursor := C.tinygo_clang_getTypeDeclaration(typ)
+		name := getString(C.tinygo_clang_getCursorSpelling(cursor))
+		if name == "" {
+			// Anonymous record, probably inside a typedef.
+			typeExpr, bitfieldList := p.makeASTRecordType(cursor, pos)
+			if bitfieldList != nil {
+				// This struct has bitfields, so we have to declare it as a
+				// named type (for bitfield getters/setters to work).
+				p.anonStructNum++
+				cgoName := "struct_" + strconv.Itoa(p.anonStructNum)
+				p.elaboratedTypes[cgoName] = &elaboratedTypeInfo{
+					typeExpr:  typeExpr,
+					pos:       pos,
+					bitfields: bitfieldList,
+				}
+				return &ast.Ident{
+					NamePos: pos,
+					Name:    "C." + cgoName,
+				}
+			}
+			return typeExpr
+		} else {
+			var cgoName string
+			switch C.tinygo_clang_getCursorKind(cursor) {
+			case C.CXCursor_StructDecl:
+				cgoName = "struct_" + name
+			case C.CXCursor_UnionDecl:
+				cgoName = "union_" + name
+			default:
+				panic("unknown record declaration")
+			}
+			if _, ok := p.elaboratedTypes[cgoName]; !ok {
+				p.elaboratedTypes[cgoName] = nil // predeclare (to avoid endless recursion)
+				typeExpr, bitfieldList := p.makeASTRecordType(cursor, pos)
+				p.elaboratedTypes[cgoName] = &elaboratedTypeInfo{
+					typeExpr:  typeExpr,
+					pos:       pos,
+					bitfields: bitfieldList,
+				}
+			}
+			return &ast.Ident{
+				NamePos: pos,
+				Name:    "C." + cgoName,
+			}
+		}
+	case C.CXType_Enum:
+		cursor := C.tinygo_clang_getTypeDeclaration(typ)
+		name := getString(C.tinygo_clang_getCursorSpelling(cursor))
+		underlying := C.tinygo_clang_getEnumDeclIntegerType(cursor)
+		if name == "" {
+			// anonymous enum
+			ref := storedRefs.Put(p)
+			defer storedRefs.Remove(ref)
+			C.tinygo_clang_visitChildren(cursor, C.CXCursorVisitor(C.tinygo_clang_enum_visitor), C.CXClientData(ref))
+			return p.makeASTType(underlying, pos)
+		} else {
+			// named enum
+			if _, ok := p.enums[name]; !ok {
+				ref := storedRefs.Put(p)
+				defer storedRefs.Remove(ref)
+				C.tinygo_clang_visitChildren(cursor, C.CXCursorVisitor(C.tinygo_clang_enum_visitor), C.CXClientData(ref))
+				p.enums[name] = enumInfo{
+					typeExpr: p.makeASTType(underlying, pos),
+					pos:      pos,
+				}
+			}
+			return &ast.Ident{
+				NamePos: pos,
+				Name:    "C.enum_" + name,
+			}
+		}
+	}
+	if typeName == "" {
+		// Report this as an error.
+		spelling := getString(C.clang_getTypeSpelling(typ))
+		p.errors = append(p.errors, scanner.Error{
+			Pos: p.fset.PositionFor(pos, true),
+			Msg: fmt.Sprintf("unknown C type: %v (libclang type kind %d)", spelling, typ.kind),
+		})
+		// Fallback, probably incorrect but at least the error points to an odd
+		// type name.
+		typeName = "C." + spelling
+	}
+	return &ast.Ident{
+		NamePos: pos,
+		Name:    typeName,
+	}
+}
+
+// makeASTRecordType parses a C record (struct or union) and translates it into
+// a Go struct type. Unions are implemented by setting the first field to a
+// zero-lengt "C union" field, which cannot be written in Go directly.
+func (p *cgoPackage) makeASTRecordType(cursor C.GoCXCursor, pos token.Pos) (*ast.StructType, []bitfieldInfo) {
+	fieldList := &ast.FieldList{
+		Opening: pos,
+		Closing: pos,
+	}
+	var bitfieldList []bitfieldInfo
+	inBitfield := false
+	bitfieldNum := 0
+	ref := storedRefs.Put(struct {
+		fieldList    *ast.FieldList
+		pkg          *cgoPackage
+		inBitfield   *bool
+		bitfieldNum  *int
+		bitfieldList *[]bitfieldInfo
+	}{fieldList, p, &inBitfield, &bitfieldNum, &bitfieldList})
+	defer storedRefs.Remove(ref)
+	C.tinygo_clang_visitChildren(cursor, C.CXCursorVisitor(C.tinygo_clang_struct_visitor), C.CXClientData(ref))
+	switch C.tinygo_clang_getCursorKind(cursor) {
+	case C.CXCursor_StructDecl:
+		return &ast.StructType{
+			Struct: pos,
+			Fields: fieldList,
+		}, bitfieldList
+	case C.CXCursor_UnionDecl:
+		if bitfieldList != nil {
+			// This is valid C... but please don't do this.
+			p.errors = append(p.errors, scanner.Error{
+				Pos: p.fset.PositionFor(pos, true),
+				Msg: fmt.Sprintf("bitfield in a union is not supported"),
+			})
+		}
+		if len(fieldList.List) > 1 {
+			// Insert a special field at the front (of zero width) as a
+			// marker that this is struct is actually a union. This is done
+			// by giving the field a name that cannot be expressed directly
+			// in Go.
+			// Other parts of the compiler look at the first element in a
+			// struct (of size > 2) to know whether this is a union.
+			// Note that we don't have to insert it for single-element
+			// unions as they're basically equivalent to a struct.
+			unionMarker := &ast.Field{
+				Type: &ast.StructType{
+					Struct: pos,
+				},
+			}
+			unionMarker.Names = []*ast.Ident{
+				&ast.Ident{
+					NamePos: pos,
+					Name:    "C union",
+					Obj: &ast.Object{
+						Kind: ast.Var,
+						Name: "C union",
+						Decl: unionMarker,
+					},
+				},
+			}
+			fieldList.List = append([]*ast.Field{unionMarker}, fieldList.List...)
+		}
+		return &ast.StructType{
+			Struct: pos,
+			Fields: fieldList,
+		}, bitfieldList
+	default:
+		panic("unknown record declaration")
+	}
+}
+
+//export tinygo_clang_struct_visitor
+func tinygo_clang_struct_visitor(c, parent C.GoCXCursor, client_data C.CXClientData) C.int {
+	passed := storedRefs.Get(unsafe.Pointer(client_data)).(struct {
+		fieldList    *ast.FieldList
+		pkg          *cgoPackage
+		inBitfield   *bool
+		bitfieldNum  *int
+		bitfieldList *[]bitfieldInfo
+	})
+	fieldList := passed.fieldList
+	p := passed.pkg
+	inBitfield := passed.inBitfield
+	bitfieldNum := passed.bitfieldNum
+	bitfieldList := passed.bitfieldList
+	if C.tinygo_clang_getCursorKind(c) != C.CXCursor_FieldDecl {
+		panic("expected field inside cursor")
+	}
+	name := getString(C.tinygo_clang_getCursorSpelling(c))
+	if name == "" {
+		// Assume this is a bitfield of 0 bits.
+		// Warning: this is not necessarily true!
+		return C.CXChildVisit_Continue
+	}
+	typ := C.tinygo_clang_getCursorType(c)
+	pos := p.getCursorPosition(c)
+	field := &ast.Field{
+		Type: p.makeASTType(typ, p.getCursorPosition(c)),
+	}
+	offsetof := int64(C.clang_Type_getOffsetOf(C.tinygo_clang_getCursorType(parent), C.CString(name)))
+	alignOf := int64(C.clang_Type_getAlignOf(typ) * 8)
+	bitfieldOffset := offsetof % alignOf
+	if bitfieldOffset != 0 {
+		if C.tinygo_clang_Cursor_isBitField(c) != 1 {
+			panic("expected a bitfield")
+		}
+		if !*inBitfield {
+			*bitfieldNum++
+		}
+		bitfieldName := "__bitfield_" + strconv.Itoa(*bitfieldNum)
+		prevField := fieldList.List[len(fieldList.List)-1]
+		if !*inBitfield {
+			// The previous element also was a bitfield, but wasn't noticed
+			// then. Add it now.
+			*inBitfield = true
+			*bitfieldList = append(*bitfieldList, bitfieldInfo{
+				field:    prevField,
+				name:     prevField.Names[0].Name,
+				startBit: 0,
+				pos:      prevField.Names[0].NamePos,
+			})
+			prevField.Names[0].Name = bitfieldName
+			prevField.Names[0].Obj.Name = bitfieldName
+		}
+		prevBitfield := &(*bitfieldList)[len(*bitfieldList)-1]
+		prevBitfield.endBit = bitfieldOffset
+		*bitfieldList = append(*bitfieldList, bitfieldInfo{
+			field:    prevField,
+			name:     name,
+			startBit: bitfieldOffset,
+			pos:      pos,
+		})
+		return C.CXChildVisit_Continue
+	}
+	*inBitfield = false
+	field.Names = []*ast.Ident{
+		&ast.Ident{
+			NamePos: pos,
+			Name:    name,
+			Obj: &ast.Object{
+				Kind: ast.Var,
+				Name: name,
+				Decl: field,
+			},
+		},
+	}
+	fieldList.List = append(fieldList.List, field)
+	return C.CXChildVisit_Continue
+}
+
+//export tinygo_clang_enum_visitor
+func tinygo_clang_enum_visitor(c, parent C.GoCXCursor, client_data C.CXClientData) C.int {
+	p := storedRefs.Get(unsafe.Pointer(client_data)).(*cgoPackage)
+	name := getString(C.tinygo_clang_getCursorSpelling(c))
+	pos := p.getCursorPosition(c)
+	value := C.tinygo_clang_getEnumConstantDeclValue(c)
+	p.constants[name] = constantInfo{
+		expr: &ast.BasicLit{pos, token.INT, strconv.FormatInt(int64(value), 10)},
+		pos:  pos,
+	}
+	return C.CXChildVisit_Continue
+}

cgo/libclang_config.go

@@ -1,11 +1,11 @@
 // +build !byollvm
 
-package loader
+package cgo
 
 /*
-#cgo linux  CFLAGS: -I/usr/lib/llvm-7/include
+#cgo linux  CFLAGS: -I/usr/lib/llvm-8/include
 #cgo darwin CFLAGS: -I/usr/local/opt/llvm/include
-#cgo linux  LDFLAGS: -L/usr/lib/llvm-7/lib -lclang
+#cgo linux  LDFLAGS: -L/usr/lib/llvm-8/lib -lclang
 #cgo darwin LDFLAGS: -L/usr/local/opt/llvm/lib -lclang -lffi
 */
 import "C"

cgo/libclang_stubs.c

@@ -0,0 +1,70 @@
+
+// This file implements some small trampoline functions. The signatures
+// are slightly different from the ones defined in libclang.go, but they
+// should be ABI compatible.
+
+#include <clang-c/Index.h> // if this fails, install libclang-8-dev
+
+CXCursor tinygo_clang_getTranslationUnitCursor(CXTranslationUnit tu) {
+	return clang_getTranslationUnitCursor(tu);
+}
+
+unsigned tinygo_clang_visitChildren(CXCursor parent, CXCursorVisitor visitor, CXClientData client_data) {
+	return clang_visitChildren(parent, visitor, client_data);
+}
+
+CXString tinygo_clang_getCursorSpelling(CXCursor c) {
+	return clang_getCursorSpelling(c);
+}
+
+enum CXCursorKind tinygo_clang_getCursorKind(CXCursor c) {
+	return clang_getCursorKind(c);
+}
+
+CXType tinygo_clang_getCursorType(CXCursor c) {
+	return clang_getCursorType(c);
+}
+
+CXCursor tinygo_clang_getTypeDeclaration(CXType t) {
+	return clang_getTypeDeclaration(t);
+}
+
+CXType tinygo_clang_getTypedefDeclUnderlyingType(CXCursor c) {
+	return clang_getTypedefDeclUnderlyingType(c);
+}
+
+CXType tinygo_clang_getCursorResultType(CXCursor c) {
+	return clang_getCursorResultType(c);
+}
+
+int tinygo_clang_Cursor_getNumArguments(CXCursor c) {
+	return clang_Cursor_getNumArguments(c);
+}
+
+CXCursor tinygo_clang_Cursor_getArgument(CXCursor c, unsigned i) {
+	return clang_Cursor_getArgument(c, i);
+}
+
+CXSourceLocation tinygo_clang_getCursorLocation(CXCursor c) {
+	return clang_getCursorLocation(c);
+}
+
+CXSourceRange tinygo_clang_getCursorExtent(CXCursor c) {
+	return clang_getCursorExtent(c);
+}
+
+CXTranslationUnit tinygo_clang_Cursor_getTranslationUnit(CXCursor c) {
+	return clang_Cursor_getTranslationUnit(c);
+}
+
+long long tinygo_clang_getEnumConstantDeclValue(CXCursor c) {
+	return clang_getEnumConstantDeclValue(c);
+}
+
+CXType tinygo_clang_getEnumDeclIntegerType(CXCursor c) {
+	return clang_getEnumDeclIntegerType(c);
+}
+
+unsigned tinygo_clang_Cursor_isBitField(CXCursor c) {
+	return clang_Cursor_isBitField(c);
+}

cgo/sync.go

@@ -0,0 +1,46 @@
+package cgo
+
+import (
+	"sync"
+	"unsafe"
+)
+
+// #include <stdlib.h>
+import "C"
+
+// refMap is a convenient way to store opaque references that can be passed to
+// C. It is useful if an API uses function pointers and you cannot pass a Go
+// pointer but only a C pointer.
+type refMap struct {
+	refs map[unsafe.Pointer]interface{}
+	lock sync.Mutex
+}
+
+// Put stores a value in the map. It can later be retrieved using Get. It must
+// be removed using Remove to avoid memory leaks.
+func (m *refMap) Put(v interface{}) unsafe.Pointer {
+	m.lock.Lock()
+	defer m.lock.Unlock()
+	if m.refs == nil {
+		m.refs = make(map[unsafe.Pointer]interface{}, 1)
+	}
+	ref := C.malloc(1)
+	m.refs[ref] = v
+	return ref
+}
+
+// Get returns a stored value previously inserted with Put. Use the same
+// reference as you got from Put.
+func (m *refMap) Get(ref unsafe.Pointer) interface{} {
+	m.lock.Lock()
+	defer m.lock.Unlock()
+	return m.refs[ref]
+}
+
+// Remove deletes a single reference from the map.
+func (m *refMap) Remove(ref unsafe.Pointer) {
+	m.lock.Lock()
+	defer m.lock.Unlock()
+	delete(m.refs, ref)
+	C.free(ref)
+}

commands.go

@@ -1,11 +1,45 @@
-// +build !darwin
-
 package main
 
-// commands used by the compilation process might have different file names on Linux than those used on macOS.
-var commands = map[string]string{
-	"ar":      "llvm-ar-7",
-	"clang":   "clang-7",
-	"ld.lld":  "ld.lld-7",
-	"wasm-ld": "wasm-ld-7",
+import (
+	"errors"
+	"os"
+	"os/exec"
+	"runtime"
+	"strings"
+)
+
+// Commands used by the compilation process might have different file names
+// across operating systems and distributions.
+var commands = map[string][]string{
+	"clang":   {"clang-8"},
+	"ld.lld":  {"ld.lld-8", "ld.lld"},
+	"wasm-ld": {"wasm-ld-8", "wasm-ld"},
+}
+
+func init() {
+	// Add the path to a Homebrew-installed LLVM 8 for ease of use (no need to
+	// manually set $PATH).
+	if runtime.GOOS == "darwin" {
+		commands["clang"] = append(commands["clang"], "/usr/local/opt/llvm/bin/clang-8")
+		commands["ld.lld"] = append(commands["ld.lld"], "/usr/local/opt/llvm/bin/ld.lld")
+		commands["wasm-ld"] = append(commands["wasm-ld"], "/usr/local/opt/llvm/bin/wasm-ld")
+	}
+}
+
+func execCommand(cmdNames []string, args ...string) error {
+	for _, cmdName := range cmdNames {
+		cmd := exec.Command(cmdName, args...)
+		cmd.Stdout = os.Stdout
+		cmd.Stderr = os.Stderr
+		err := cmd.Run()
+		if err != nil {
+			if err, ok := err.(*exec.Error); ok && err.Err == exec.ErrNotFound {
+				// this command was not found, try the next
+				continue
+			}
+			return err
+		}
+		return nil
+	}
+	return errors.New("none of these commands were found in your $PATH: " + strings.Join(cmdNames, " "))
 }

commands_macos.go

@@ -1,11 +0,0 @@
-// +build darwin
-
-package main
-
-// commands used by the compilation process might have different file names on macOS than those used on Linux.
-var commands = map[string]string{
-	"ar":      "llvm-ar",
-	"clang":   "clang-7",
-	"ld.lld":  "ld.lld-7",
-	"wasm-ld": "wasm-ld-7",
-}

compiler/asserts.go

@@ -42,7 +42,7 @@ func (c *Compiler) emitLookupBoundsCheck(frame *Frame, arrayLen, index llvm.Valu
 
 	// Fail: this is a nil pointer, exit with a panic.
 	c.builder.SetInsertPointAtEnd(faultBlock)
-	c.createRuntimeCall("lookuppanic", nil, "")
+	c.createRuntimeCall("lookupPanic", nil, "")
 	c.builder.CreateUnreachable()
 
 	// Ok: this is a valid pointer.
@@ -51,7 +51,12 @@ func (c *Compiler) emitLookupBoundsCheck(frame *Frame, arrayLen, index llvm.Valu
 
 // emitSliceBoundsCheck emits a bounds check before a slicing operation to make
 // sure it is within bounds.
-func (c *Compiler) emitSliceBoundsCheck(frame *Frame, capacity, low, high llvm.Value, lowType, highType *types.Basic) {
+//
+// This function is both used for slicing a slice (low and high have their
+// normal meaning) and for creating a new slice, where 'capacity' means the
+// biggest possible slice capacity, 'low' means len and 'high' means cap. The
+// logic is the same in both cases.
+func (c *Compiler) emitSliceBoundsCheck(frame *Frame, capacity, low, high, max llvm.Value, lowType, highType, maxType *types.Basic) {
 	if frame.fn.IsNoBounds() {
 		// The //go:nobounds pragma was added to the function to avoid bounds
 		// checking.
@@ -66,6 +71,9 @@ func (c *Compiler) emitSliceBoundsCheck(frame *Frame, capacity, low, high llvm.V
 	if high.Type().IntTypeWidth() > capacityType.IntTypeWidth() {
 		capacityType = high.Type()
 	}
+	if max.Type().IntTypeWidth() > capacityType.IntTypeWidth() {
+		capacityType = max.Type()
+	}
 	if capacityType != capacity.Type() {
 		capacity = c.builder.CreateZExt(capacity, capacityType, "")
 	}
@@ -85,6 +93,13 @@ func (c *Compiler) emitSliceBoundsCheck(frame *Frame, capacity, low, high llvm.V
 			high = c.builder.CreateSExt(high, capacityType, "")
 		}
 	}
+	if max.Type().IntTypeWidth() < capacityType.IntTypeWidth() {
+		if maxType.Info()&types.IsUnsigned != 0 {
+			max = c.builder.CreateZExt(max, capacityType, "")
+		} else {
+			max = c.builder.CreateSExt(max, capacityType, "")
+		}
+	}
 
 	faultBlock := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "slice.outofbounds")
 	nextBlock := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "slice.next")
@@ -92,13 +107,15 @@ func (c *Compiler) emitSliceBoundsCheck(frame *Frame, capacity, low, high llvm.V
 
 	// Now do the bounds check: low > high || high > capacity
 	outOfBounds1 := c.builder.CreateICmp(llvm.IntUGT, low, high, "slice.lowhigh")
-	outOfBounds2 := c.builder.CreateICmp(llvm.IntUGT, high, capacity, "slice.highcap")
-	outOfBounds := c.builder.CreateOr(outOfBounds1, outOfBounds2, "slice.outofbounds")
+	outOfBounds2 := c.builder.CreateICmp(llvm.IntUGT, high, max, "slice.highmax")
+	outOfBounds3 := c.builder.CreateICmp(llvm.IntUGT, max, capacity, "slice.maxcap")
+	outOfBounds := c.builder.CreateOr(outOfBounds1, outOfBounds2, "slice.lowmax")
+	outOfBounds = c.builder.CreateOr(outOfBounds, outOfBounds3, "slice.lowcap")
 	c.builder.CreateCondBr(outOfBounds, faultBlock, nextBlock)
 
 	// Fail: this is a nil pointer, exit with a panic.
 	c.builder.SetInsertPointAtEnd(faultBlock)
-	c.createRuntimeCall("slicepanic", nil, "")
+	c.createRuntimeCall("slicePanic", nil, "")
 	c.builder.CreateUnreachable()
 
 	// Ok: this is a valid pointer.
@@ -109,19 +126,39 @@ func (c *Compiler) emitSliceBoundsCheck(frame *Frame, capacity, low, high llvm.V
 // has no effect in well-behaved programs, but makes sure no uncaught nil
 // pointer dereferences exist in valid Go code.
 func (c *Compiler) emitNilCheck(frame *Frame, ptr llvm.Value, blockPrefix string) {
+	// Check whether we need to emit this check at all.
+	if !ptr.IsAGlobalValue().IsNil() {
+		return
+	}
+
 	// Check whether this is a nil pointer.
 	faultBlock := c.ctx.AddBasicBlock(frame.fn.LLVMFn, blockPrefix+".nil")
 	nextBlock := c.ctx.AddBasicBlock(frame.fn.LLVMFn, blockPrefix+".next")
 	frame.blockExits[frame.currentBlock] = nextBlock // adjust outgoing block for phi nodes
 
 	// Compare against nil.
-	nilptr := llvm.ConstPointerNull(ptr.Type())
-	isnil := c.builder.CreateICmp(llvm.IntEQ, ptr, nilptr, "")
+	var isnil llvm.Value
+	if ptr.Type().PointerAddressSpace() == 0 {
+		// Do the nil check using the isnil builtin, which marks the parameter
+		// as nocapture.
+		// The reason it has to go through a builtin, is that a regular icmp
+		// instruction may capture the pointer in LLVM semantics, see
+		// https://reviews.llvm.org/D60047 for details. Pointer capturing
+		// unfortunately breaks escape analysis, so we use this trick to let the
+		// functionattr pass know that this pointer doesn't really escape.
+		ptr = c.builder.CreateBitCast(ptr, c.i8ptrType, "")
+		isnil = c.createRuntimeCall("isnil", []llvm.Value{ptr}, "")
+	} else {
+		// Do the nil check using a regular icmp. This can happen with function
+		// pointers on AVR, which don't benefit from escape analysis anyway.
+		nilptr := llvm.ConstPointerNull(ptr.Type())
+		isnil = c.builder.CreateICmp(llvm.IntEQ, ptr, nilptr, "")
+	}
 	c.builder.CreateCondBr(isnil, faultBlock, nextBlock)
 
 	// Fail: this is a nil pointer, exit with a panic.
 	c.builder.SetInsertPointAtEnd(faultBlock)
-	c.createRuntimeCall("nilpanic", nil, "")
+	c.createRuntimeCall("nilPanic", nil, "")
 	c.builder.CreateUnreachable()
 
 	// Ok: this is a valid pointer.

compiler/calls.go

@@ -38,7 +38,7 @@ func (c *Compiler) createCall(fn llvm.Value, args []llvm.Value, name string) llv
 }
 
 // Expand an argument type to a list that can be used in a function call
-// paramter list.
+// parameter list.
 func (c *Compiler) expandFormalParamType(t llvm.Type) []llvm.Type {
 	switch t.TypeKind() {
 	case llvm.StructTypeKind:
@@ -55,6 +55,25 @@ func (c *Compiler) expandFormalParamType(t llvm.Type) []llvm.Type {
 	}
 }
 
+// Expand an argument type to a list of offsets from the start of the object.
+// Used together with expandFormalParam to get the offset of each value from the
+// start of the non-expanded value.
+func (c *Compiler) expandFormalParamOffsets(t llvm.Type) []uint64 {
+	switch t.TypeKind() {
+	case llvm.StructTypeKind:
+		fields := c.flattenAggregateTypeOffsets(t)
+		if len(fields) <= MaxFieldsPerParam {
+			return fields
+		} else {
+			// failed to lower
+			return []uint64{0}
+		}
+	default:
+		// TODO: split small arrays
+		return []uint64{0}
+	}
+}
+
 // Equivalent of expandFormalParamType for parameter values.
 func (c *Compiler) expandFormalParam(v llvm.Value) []llvm.Value {
 	switch v.Type().TypeKind() {
@@ -92,6 +111,27 @@ func (c *Compiler) flattenAggregateType(t llvm.Type) []llvm.Type {
 	}
 }
 
+// Return the offsets from the start of the object if this object type were
+// flattened like in flattenAggregate. Used together with flattenAggregate to
+// know the start indices of each value in the non-flattened object.
+func (c *Compiler) flattenAggregateTypeOffsets(t llvm.Type) []uint64 {
+	switch t.TypeKind() {
+	case llvm.StructTypeKind:
+		fields := make([]uint64, 0, t.StructElementTypesCount())
+		for fieldIndex, field := range t.StructElementTypes() {
+			suboffsets := c.flattenAggregateTypeOffsets(field)
+			offset := c.targetData.ElementOffset(t, fieldIndex)
+			for i := range suboffsets {
+				suboffsets[i] += offset
+			}
+			fields = append(fields, suboffsets...)
+		}
+		return fields
+	default:
+		return []uint64{0}
+	}
+}
+
 // Break down a struct into its elementary types for argument passing. The value
 // equivalent of flattenAggregateType
 func (c *Compiler) flattenAggregate(v llvm.Value) []llvm.Value {
@@ -123,10 +163,7 @@ func (c *Compiler) collapseFormalParamInternal(t llvm.Type, fields []llvm.Value)
 	switch t.TypeKind() {
 	case llvm.StructTypeKind:
 		if len(c.flattenAggregateType(t)) <= MaxFieldsPerParam {
-			value, err := c.getZeroValue(t)
-			if err != nil {
-				panic("could not get zero value of struct: " + err.Error())
-			}
+			value := llvm.ConstNull(t)
 			for i, subtyp := range t.StructElementTypes() {
 				structField, remaining := c.collapseFormalParamInternal(subtyp, fields)
 				fields = remaining

compiler/channel.go

@@ -4,6 +4,7 @@ package compiler
 // or pseudo-operations that are lowered during goroutine lowering.
 
 import (
+	"fmt"
 	"go/types"
 
 	"golang.org/x/tools/go/ssa"
@@ -12,86 +13,226 @@ import (
 
 // emitMakeChan returns a new channel value for the given channel type.
 func (c *Compiler) emitMakeChan(expr *ssa.MakeChan) (llvm.Value, error) {
-	valueType, err := c.getLLVMType(expr.Type().(*types.Chan).Elem())
-	if err != nil {
-		return llvm.Value{}, err
-	}
-	if c.targetData.TypeAllocSize(valueType) > c.targetData.TypeAllocSize(c.intType) {
-		// Values bigger than int overflow the data part of the coroutine.
-		// TODO: make the coroutine data part big enough to hold these bigger
-		// values.
-		return llvm.Value{}, c.makeError(expr.Pos(), "todo: channel with values bigger than int")
-	}
-	chanType := c.mod.GetTypeByName("runtime.channel")
-	size := c.targetData.TypeAllocSize(chanType)
+	chanType := c.getLLVMType(expr.Type())
+	size := c.targetData.TypeAllocSize(chanType.ElementType())
 	sizeValue := llvm.ConstInt(c.uintptrType, size, false)
 	ptr := c.createRuntimeCall("alloc", []llvm.Value{sizeValue}, "chan.alloc")
-	ptr = c.builder.CreateBitCast(ptr, llvm.PointerType(chanType, 0), "chan")
+	ptr = c.builder.CreateBitCast(ptr, chanType, "chan")
+	// Set the elementSize field
+	elementSizePtr := c.builder.CreateGEP(ptr, []llvm.Value{
+		llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+		llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+	}, "")
+	elementSize := c.targetData.TypeAllocSize(c.getLLVMType(expr.Type().(*types.Chan).Elem()))
+	if elementSize > 0xffff {
+		return ptr, c.makeError(expr.Pos(), fmt.Sprintf("element size is %d bytes, which is bigger than the maximum of %d bytes", elementSize, 0xffff))
+	}
+	elementSizeValue := llvm.ConstInt(c.ctx.Int16Type(), elementSize, false)
+	c.builder.CreateStore(elementSizeValue, elementSizePtr)
 	return ptr, nil
 }
 
 // emitChanSend emits a pseudo chan send operation. It is lowered to the actual
 // channel send operation during goroutine lowering.
-func (c *Compiler) emitChanSend(frame *Frame, instr *ssa.Send) error {
-	valueType, err := c.getLLVMType(instr.Chan.Type().(*types.Chan).Elem())
-	if err != nil {
-		return err
-	}
-	ch, err := c.parseExpr(frame, instr.Chan)
-	if err != nil {
-		return err
-	}
-	chanValue, err := c.parseExpr(frame, instr.X)
-	if err != nil {
-		return err
-	}
-	valueSize := llvm.ConstInt(c.uintptrType, c.targetData.TypeAllocSize(chanValue.Type()), false)
-	valueAlloca := c.builder.CreateAlloca(valueType, "chan.value")
+func (c *Compiler) emitChanSend(frame *Frame, instr *ssa.Send) {
+	ch := c.getValue(frame, instr.Chan)
+	chanValue := c.getValue(frame, instr.X)
+
+	// store value-to-send
+	valueType := c.getLLVMType(instr.X.Type())
+	valueAlloca, valueAllocaCast, valueAllocaSize := c.createTemporaryAlloca(valueType, "chan.value")
 	c.builder.CreateStore(chanValue, valueAlloca)
-	valueAllocaCast := c.builder.CreateBitCast(valueAlloca, c.i8ptrType, "chan.value.i8ptr")
-	c.createRuntimeCall("chanSendStub", []llvm.Value{llvm.Undef(c.i8ptrType), ch, valueAllocaCast, valueSize}, "")
-	return nil
+
+	// Do the send.
+	coroutine := c.createRuntimeCall("getCoroutine", nil, "")
+	c.createRuntimeCall("chanSend", []llvm.Value{coroutine, ch, valueAllocaCast}, "")
+
+	// End the lifetime of the alloca.
+	// This also works around a bug in CoroSplit, at least in LLVM 8:
+	// https://bugs.llvm.org/show_bug.cgi?id=41742
+	c.emitLifetimeEnd(valueAllocaCast, valueAllocaSize)
 }
 
 // emitChanRecv emits a pseudo chan receive operation. It is lowered to the
 // actual channel receive operation during goroutine lowering.
-func (c *Compiler) emitChanRecv(frame *Frame, unop *ssa.UnOp) (llvm.Value, error) {
-	valueType, err := c.getLLVMType(unop.X.Type().(*types.Chan).Elem())
-	if err != nil {
-		return llvm.Value{}, err
-	}
-	valueSize := llvm.ConstInt(c.uintptrType, c.targetData.TypeAllocSize(valueType), false)
-	ch, err := c.parseExpr(frame, unop.X)
-	if err != nil {
-		return llvm.Value{}, err
-	}
-	valueAlloca := c.builder.CreateAlloca(valueType, "chan.value")
-	valueAllocaCast := c.builder.CreateBitCast(valueAlloca, c.i8ptrType, "chan.value.i8ptr")
-	valueOk := c.builder.CreateAlloca(c.ctx.Int1Type(), "chan.comma-ok.alloca")
-	c.createRuntimeCall("chanRecvStub", []llvm.Value{llvm.Undef(c.i8ptrType), ch, valueAllocaCast, valueOk, valueSize}, "")
+func (c *Compiler) emitChanRecv(frame *Frame, unop *ssa.UnOp) llvm.Value {
+	valueType := c.getLLVMType(unop.X.Type().(*types.Chan).Elem())
+	ch := c.getValue(frame, unop.X)
+
+	// Allocate memory to receive into.
+	valueAlloca, valueAllocaCast, valueAllocaSize := c.createTemporaryAlloca(valueType, "chan.value")
+
+	// Do the receive.
+	coroutine := c.createRuntimeCall("getCoroutine", nil, "")
+	c.createRuntimeCall("chanRecv", []llvm.Value{coroutine, ch, valueAllocaCast}, "")
 	received := c.builder.CreateLoad(valueAlloca, "chan.received")
+	c.emitLifetimeEnd(valueAllocaCast, valueAllocaSize)
+
 	if unop.CommaOk {
-		commaOk := c.builder.CreateLoad(valueOk, "chan.comma-ok")
+		commaOk := c.createRuntimeCall("getTaskStateData", []llvm.Value{coroutine}, "chan.commaOk.wide")
+		commaOk = c.builder.CreateTrunc(commaOk, c.ctx.Int1Type(), "chan.commaOk")
 		tuple := llvm.Undef(c.ctx.StructType([]llvm.Type{valueType, c.ctx.Int1Type()}, false))
 		tuple = c.builder.CreateInsertValue(tuple, received, 0, "")
 		tuple = c.builder.CreateInsertValue(tuple, commaOk, 1, "")
-		return tuple, nil
+		return tuple
 	} else {
-		return received, nil
+		return received
 	}
 }
 
 // emitChanClose closes the given channel.
-func (c *Compiler) emitChanClose(frame *Frame, param ssa.Value) error {
-	valueType, err := c.getLLVMType(param.Type().(*types.Chan).Elem())
-	valueSize := llvm.ConstInt(c.uintptrType, c.targetData.TypeAllocSize(valueType), false)
-	if err != nil {
-		return err
-	}
-	ch, err := c.parseExpr(frame, param)
-	if err != nil {
-		return err
-	}
-	c.createRuntimeCall("chanClose", []llvm.Value{ch, valueSize}, "")
-	return nil
+func (c *Compiler) emitChanClose(frame *Frame, param ssa.Value) {
+	ch := c.getValue(frame, param)
+	c.createRuntimeCall("chanClose", []llvm.Value{ch}, "")
+}
+
+// emitSelect emits all IR necessary for a select statements. That's a
+// non-trivial amount of code because select is very complex to implement.
+func (c *Compiler) emitSelect(frame *Frame, expr *ssa.Select) llvm.Value {
+	if len(expr.States) == 0 {
+		// Shortcuts for some simple selects.
+		llvmType := c.getLLVMType(expr.Type())
+		if expr.Blocking {
+			// Blocks forever:
+			//     select {}
+			c.createRuntimeCall("deadlock", nil, "")
+			return llvm.Undef(llvmType)
+		} else {
+			// No-op:
+			//     select {
+			//     default:
+			//     }
+			retval := llvm.Undef(llvmType)
+			retval = c.builder.CreateInsertValue(retval, llvm.ConstInt(c.intType, 0xffffffffffffffff, true), 0, "")
+			return retval // {-1, false}
+		}
+	}
+
+	// This code create a (stack-allocated) slice containing all the select
+	// cases and then calls runtime.chanSelect to perform the actual select
+	// statement.
+	// Simple selects (blocking and with just one case) are already transformed
+	// into regular chan operations during SSA construction so we don't have to
+	// optimize such small selects.
+
+	// Go through all the cases. Create the selectStates slice and and
+	// determine the receive buffer size and alignment.
+	recvbufSize := uint64(0)
+	recvbufAlign := 0
+	hasReceives := false
+	var selectStates []llvm.Value
+	chanSelectStateType := c.getLLVMRuntimeType("chanSelectState")
+	for _, state := range expr.States {
+		ch := c.getValue(frame, state.Chan)
+		selectState := llvm.ConstNull(chanSelectStateType)
+		selectState = c.builder.CreateInsertValue(selectState, ch, 0, "")
+		switch state.Dir {
+		case types.RecvOnly:
+			// Make sure the receive buffer is big enough and has the correct alignment.
+			llvmType := c.getLLVMType(state.Chan.Type().(*types.Chan).Elem())
+			if size := c.targetData.TypeAllocSize(llvmType); size > recvbufSize {
+				recvbufSize = size
+			}
+			if align := c.targetData.ABITypeAlignment(llvmType); align > recvbufAlign {
+				recvbufAlign = align
+			}
+			hasReceives = true
+		case types.SendOnly:
+			// Store this value in an alloca and put a pointer to this alloca
+			// in the send state.
+			sendValue := c.getValue(frame, state.Send)
+			alloca := c.createEntryBlockAlloca(sendValue.Type(), "select.send.value")
+			c.builder.CreateStore(sendValue, alloca)
+			ptr := c.builder.CreateBitCast(alloca, c.i8ptrType, "")
+			selectState = c.builder.CreateInsertValue(selectState, ptr, 1, "")
+		default:
+			panic("unreachable")
+		}
+		selectStates = append(selectStates, selectState)
+	}
+
+	// Create a receive buffer, where the received value will be stored.
+	recvbuf := llvm.Undef(c.i8ptrType)
+	if hasReceives {
+		allocaType := llvm.ArrayType(c.ctx.Int8Type(), int(recvbufSize))
+		recvbufAlloca := c.builder.CreateAlloca(allocaType, "select.recvbuf.alloca")
+		recvbufAlloca.SetAlignment(recvbufAlign)
+		recvbuf = c.builder.CreateGEP(recvbufAlloca, []llvm.Value{
+			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+		}, "select.recvbuf")
+	}
+
+	// Create the states slice (allocated on the stack).
+	statesAllocaType := llvm.ArrayType(chanSelectStateType, len(selectStates))
+	statesAlloca := c.builder.CreateAlloca(statesAllocaType, "select.states.alloca")
+	for i, state := range selectStates {
+		// Set each slice element to the appropriate channel.
+		gep := c.builder.CreateGEP(statesAlloca, []llvm.Value{
+			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+			llvm.ConstInt(c.ctx.Int32Type(), uint64(i), false),
+		}, "")
+		c.builder.CreateStore(state, gep)
+	}
+	statesPtr := c.builder.CreateGEP(statesAlloca, []llvm.Value{
+		llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+		llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+	}, "select.states")
+	statesLen := llvm.ConstInt(c.uintptrType, uint64(len(selectStates)), false)
+
+	// Convert the 'blocking' flag on this select into a LLVM value.
+	blockingInt := uint64(0)
+	if expr.Blocking {
+		blockingInt = 1
+	}
+	blockingValue := llvm.ConstInt(c.ctx.Int1Type(), blockingInt, false)
+
+	// Do the select in the runtime.
+	results := c.createRuntimeCall("chanSelect", []llvm.Value{
+		recvbuf,
+		statesPtr, statesLen, statesLen, // []chanSelectState
+		blockingValue,
+	}, "")
+
+	// The result value does not include all the possible received values,
+	// because we can't load them in advance. Instead, the *ssa.Extract
+	// instruction will treat a *ssa.Select specially and load it there inline.
+	// Store the receive alloca in a sidetable until we hit this extract
+	// instruction.
+	if frame.selectRecvBuf == nil {
+		frame.selectRecvBuf = make(map[*ssa.Select]llvm.Value)
+	}
+	frame.selectRecvBuf[expr] = recvbuf
+
+	return results
+}
+
+// getChanSelectResult returns the special values from a *ssa.Extract expression
+// when extracting a value from a select statement (*ssa.Select). Because
+// *ssa.Select cannot load all values in advance, it does this later in the
+// *ssa.Extract expression.
+func (c *Compiler) getChanSelectResult(frame *Frame, expr *ssa.Extract) llvm.Value {
+	if expr.Index == 0 {
+		// index
+		value := c.getValue(frame, expr.Tuple)
+		index := c.builder.CreateExtractValue(value, expr.Index, "")
+		if index.Type().IntTypeWidth() < c.intType.IntTypeWidth() {
+			index = c.builder.CreateSExt(index, c.intType, "")
+		}
+		return index
+	} else if expr.Index == 1 {
+		// comma-ok
+		value := c.getValue(frame, expr.Tuple)
+		return c.builder.CreateExtractValue(value, expr.Index, "")
+	} else {
+		// Select statements are (index, ok, ...) where ... is a number of
+		// received values, depending on how many receive statements there
+		// are. They are all combined into one alloca (because only one
+		// receive can proceed at a time) so we'll get that alloca, bitcast
+		// it to the correct type, and dereference it.
+		recvbuf := frame.selectRecvBuf[expr.Tuple.(*ssa.Select)]
+		typ := llvm.PointerType(c.getLLVMType(expr.Type()), 0)
+		ptr := c.builder.CreateBitCast(recvbuf, typ, "")
+		return c.builder.CreateLoad(ptr, "")
+	}
 }

compiler/check.go

@@ -0,0 +1,128 @@
+package compiler
+
+// This file implements a set of sanity checks for the IR that is generated.
+// It can catch some mistakes that LLVM's verifier cannot.
+
+import (
+	"errors"
+	"fmt"
+
+	"tinygo.org/x/go-llvm"
+)
+
+func (c *Compiler) checkType(t llvm.Type, checked map[llvm.Type]struct{}, specials map[llvm.TypeKind]llvm.Type) {
+	if t.IsNil() {
+		panic(t)
+	}
+
+	// prevent infinite recursion for self-referential types
+	if _, ok := checked[t]; ok {
+		return
+	}
+	checked[t] = struct{}{}
+
+	// check for any context mismatches
+	switch {
+	case t.Context() == c.ctx:
+		// this is correct
+	case t.Context() == llvm.GlobalContext():
+		// somewhere we accidentally used the global context instead of a real context
+		panic(fmt.Errorf("type %q uses global context", t.String()))
+	default:
+		// we used some other context by accident
+		panic(fmt.Errorf("type %q uses context %v instead of the main context %v", t.Context(), c.ctx))
+	}
+
+	// if this is a composite type, check the components of the type
+	switch t.TypeKind() {
+	case llvm.VoidTypeKind, llvm.LabelTypeKind, llvm.TokenTypeKind, llvm.MetadataTypeKind:
+		// there should only be one of any of these
+		if s, ok := specials[t.TypeKind()]; !ok {
+			specials[t.TypeKind()] = t
+		} else if s != t {
+			panic(fmt.Errorf("duplicate special type %q: %v and %v", t.TypeKind().String(), t, s))
+		}
+	case llvm.FloatTypeKind, llvm.DoubleTypeKind, llvm.X86_FP80TypeKind, llvm.FP128TypeKind, llvm.PPC_FP128TypeKind:
+		// floating point numbers are primitives - nothing to recurse
+	case llvm.IntegerTypeKind:
+		// integers are primitives - nothing to recurse
+	case llvm.FunctionTypeKind:
+		// check arguments and return(s)
+		for _, v := range t.ParamTypes() {
+			c.checkType(v, checked, specials)
+		}
+		c.checkType(t.ReturnType(), checked, specials)
+	case llvm.StructTypeKind:
+		// check all elements
+		for _, v := range t.StructElementTypes() {
+			c.checkType(v, checked, specials)
+		}
+	case llvm.ArrayTypeKind:
+		// check element type
+		c.checkType(t.ElementType(), checked, specials)
+	case llvm.PointerTypeKind:
+		// check underlying type
+		c.checkType(t.ElementType(), checked, specials)
+	case llvm.VectorTypeKind:
+		// check element type
+		c.checkType(t.ElementType(), checked, specials)
+	}
+}
+
+func (c *Compiler) checkValue(v llvm.Value, types map[llvm.Type]struct{}, specials map[llvm.TypeKind]llvm.Type) {
+	// check type
+	c.checkType(v.Type(), types, specials)
+}
+
+func (c *Compiler) checkInstruction(inst llvm.Value, types map[llvm.Type]struct{}, specials map[llvm.TypeKind]llvm.Type) {
+	// check value properties
+	c.checkValue(inst, types, specials)
+
+	// check operands
+	for i := 0; i < inst.OperandsCount(); i++ {
+		c.checkValue(inst.Operand(i), types, specials)
+	}
+}
+
+func (c *Compiler) checkBasicBlock(bb llvm.BasicBlock, types map[llvm.Type]struct{}, specials map[llvm.TypeKind]llvm.Type) {
+	// check basic block value and type
+	c.checkValue(bb.AsValue(), types, specials)
+
+	// check instructions
+	for inst := bb.FirstInstruction(); !inst.IsNil(); inst = llvm.NextInstruction(inst) {
+		c.checkInstruction(inst, types, specials)
+	}
+}
+
+func (c *Compiler) checkFunction(fn llvm.Value, types map[llvm.Type]struct{}, specials map[llvm.TypeKind]llvm.Type) {
+	// check function value and type
+	c.checkValue(fn, types, specials)
+
+	// check basic blocks
+	for bb := fn.FirstBasicBlock(); !bb.IsNil(); bb = llvm.NextBasicBlock(bb) {
+		c.checkBasicBlock(bb, types, specials)
+	}
+}
+
+func (c *Compiler) checkModule() {
+	// check for any context mismatches
+	switch {
+	case c.mod.Context() == c.ctx:
+		// this is correct
+	case c.mod.Context() == llvm.GlobalContext():
+		// somewhere we accidentally used the global context instead of a real context
+		panic(errors.New("module uses global context"))
+	default:
+		// we used some other context by accident
+		panic(fmt.Errorf("module uses context %v instead of the main context %v", c.mod.Context(), c.ctx))
+	}
+
+	types := map[llvm.Type]struct{}{}
+	specials := map[llvm.TypeKind]llvm.Type{}
+	for fn := c.mod.FirstFunction(); !fn.IsNil(); fn = llvm.NextFunction(fn) {
+		c.checkFunction(fn, types, specials)
+	}
+	for g := c.mod.FirstGlobal(); !g.IsNil(); g = llvm.NextGlobal(g) {
+		c.checkValue(g, types, specials)
+	}
+}

compiler/defer.go

@@ -29,14 +29,14 @@ func (c *Compiler) deferInitFunc(frame *Frame) {
 	frame.deferClosureFuncs = make(map[*ir.Function]int)
 
 	// Create defer list pointer.
-	deferType := llvm.PointerType(c.mod.GetTypeByName("runtime._defer"), 0)
+	deferType := llvm.PointerType(c.getLLVMRuntimeType("_defer"), 0)
 	frame.deferPtr = c.builder.CreateAlloca(deferType, "deferPtr")
 	c.builder.CreateStore(llvm.ConstPointerNull(deferType), frame.deferPtr)
 }
 
 // emitDefer emits a single defer instruction, to be run when this function
 // returns.
-func (c *Compiler) emitDefer(frame *Frame, instr *ssa.Defer) error {
+func (c *Compiler) emitDefer(frame *Frame, instr *ssa.Defer) {
 	// The pointer to the previous defer struct, which we will replace to
 	// make a linked list.
 	next := c.builder.CreateLoad(frame.deferPtr, "defer.next")
@@ -56,18 +56,12 @@ func (c *Compiler) emitDefer(frame *Frame, instr *ssa.Defer) error {
 
 		// Collect all values to be put in the struct (starting with
 		// runtime._defer fields, followed by the call parameters).
-		itf, err := c.parseExpr(frame, instr.Call.Value) // interface
-		if err != nil {
-			return err
-		}
+		itf := c.getValue(frame, instr.Call.Value) // interface
 		receiverValue := c.builder.CreateExtractValue(itf, 1, "invoke.func.receiver")
 		values = []llvm.Value{callback, next, receiverValue}
 		valueTypes = append(valueTypes, c.i8ptrType)
 		for _, arg := range instr.Call.Args {
-			val, err := c.parseExpr(frame, arg)
-			if err != nil {
-				return err
-			}
+			val := c.getValue(frame, arg)
 			values = append(values, val)
 			valueTypes = append(valueTypes, val.Type())
 		}
@@ -86,10 +80,7 @@ func (c *Compiler) emitDefer(frame *Frame, instr *ssa.Defer) error {
 		// runtime._defer fields).
 		values = []llvm.Value{callback, next}
 		for _, param := range instr.Call.Args {
-			llvmParam, err := c.parseExpr(frame, param)
-			if err != nil {
-				return err
-			}
+			llvmParam := c.getValue(frame, param)
 			values = append(values, llvmParam)
 			valueTypes = append(valueTypes, llvmParam.Type())
 		}
@@ -101,10 +92,7 @@ func (c *Compiler) emitDefer(frame *Frame, instr *ssa.Defer) error {
 		// pointer.
 		// TODO: ignore this closure entirely and put pointers to the free
 		// variables directly in the defer struct, avoiding a memory allocation.
-		closure, err := c.parseExpr(frame, instr.Call.Value)
-		if err != nil {
-			return err
-		}
+		closure := c.getValue(frame, instr.Call.Value)
 		context := c.builder.CreateExtractValue(closure, 0, "")
 
 		// Get the callback number.
@@ -120,10 +108,7 @@ func (c *Compiler) emitDefer(frame *Frame, instr *ssa.Defer) error {
 		// context pointer).
 		values = []llvm.Value{callback, next}
 		for _, param := range instr.Call.Args {
-			llvmParam, err := c.parseExpr(frame, param)
-			if err != nil {
-				return err
-			}
+			llvmParam := c.getValue(frame, param)
 			values = append(values, llvmParam)
 			valueTypes = append(valueTypes, llvmParam.Type())
 		}
@@ -131,15 +116,13 @@ func (c *Compiler) emitDefer(frame *Frame, instr *ssa.Defer) error {
 		valueTypes = append(valueTypes, context.Type())
 
 	} else {
-		return c.makeError(instr.Pos(), "todo: defer on uncommon function call type")
+		c.addError(instr.Pos(), "todo: defer on uncommon function call type")
+		return
 	}
 
 	// Make a struct out of the collected values to put in the defer frame.
 	deferFrameType := c.ctx.StructType(valueTypes, false)
-	deferFrame, err := c.getZeroValue(deferFrameType)
-	if err != nil {
-		return err
-	}
+	deferFrame := llvm.ConstNull(deferFrameType)
 	for i, value := range values {
 		deferFrame = c.builder.CreateInsertValue(deferFrame, value, i, "")
 	}
@@ -147,15 +130,17 @@ func (c *Compiler) emitDefer(frame *Frame, instr *ssa.Defer) error {
 	// Put this struct in an alloca.
 	alloca := c.builder.CreateAlloca(deferFrameType, "defer.alloca")
 	c.builder.CreateStore(deferFrame, alloca)
+	if c.needsStackObjects() {
+		c.trackPointer(alloca)
+	}
 
 	// Push it on top of the linked list by replacing deferPtr.
 	allocaCast := c.builder.CreateBitCast(alloca, next.Type(), "defer.alloca.cast")
 	c.builder.CreateStore(allocaCast, frame.deferPtr)
-	return nil
 }
 
 // emitRunDefers emits code to run all deferred functions.
-func (c *Compiler) emitRunDefers(frame *Frame) error {
+func (c *Compiler) emitRunDefers(frame *Frame) {
 	// Add a loop like the following:
 	//     for stack != nil {
 	//         _stack := stack
@@ -172,10 +157,10 @@ func (c *Compiler) emitRunDefers(frame *Frame) error {
 	//     }
 
 	// Create loop.
-	loophead := llvm.AddBasicBlock(frame.fn.LLVMFn, "rundefers.loophead")
-	loop := llvm.AddBasicBlock(frame.fn.LLVMFn, "rundefers.loop")
-	unreachable := llvm.AddBasicBlock(frame.fn.LLVMFn, "rundefers.default")
-	end := llvm.AddBasicBlock(frame.fn.LLVMFn, "rundefers.end")
+	loophead := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "rundefers.loophead")
+	loop := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "rundefers.loop")
+	unreachable := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "rundefers.default")
+	end := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "rundefers.end")
 	c.builder.CreateBr(loophead)
 
 	// Create loop head:
@@ -190,13 +175,13 @@ func (c *Compiler) emitRunDefers(frame *Frame) error {
 	//     stack = stack.next
 	//     switch stack.callback {
 	c.builder.SetInsertPointAtEnd(loop)
-	nextStackGEP := c.builder.CreateGEP(deferData, []llvm.Value{
+	nextStackGEP := c.builder.CreateInBoundsGEP(deferData, []llvm.Value{
 		llvm.ConstInt(c.ctx.Int32Type(), 0, false),
 		llvm.ConstInt(c.ctx.Int32Type(), 1, false), // .next field
 	}, "stack.next.gep")
 	nextStack := c.builder.CreateLoad(nextStackGEP, "stack.next")
 	c.builder.CreateStore(nextStack, frame.deferPtr)
-	gep := c.builder.CreateGEP(deferData, []llvm.Value{
+	gep := c.builder.CreateInBoundsGEP(deferData, []llvm.Value{
 		llvm.ConstInt(c.ctx.Int32Type(), 0, false),
 		llvm.ConstInt(c.ctx.Int32Type(), 0, false), // .callback field
 	}, "callback.gep")
@@ -207,7 +192,7 @@ func (c *Compiler) emitRunDefers(frame *Frame) error {
 		// Create switch case, for example:
 		//     case 0:
 		//         // run first deferred call
-		block := llvm.AddBasicBlock(frame.fn.LLVMFn, "rundefers.callback")
+		block := c.ctx.AddBasicBlock(frame.fn.LLVMFn, "rundefers.callback")
 		sw.AddCase(llvm.ConstInt(c.uintptrType, uint64(i), false), block)
 		c.builder.SetInsertPointAtEnd(block)
 		switch callback := callback.(type) {
@@ -218,13 +203,9 @@ func (c *Compiler) emitRunDefers(frame *Frame) error {
 			}
 
 			// Get the real defer struct type and cast to it.
-			valueTypes := []llvm.Type{c.uintptrType, llvm.PointerType(c.mod.GetTypeByName("runtime._defer"), 0), c.i8ptrType}
+			valueTypes := []llvm.Type{c.uintptrType, llvm.PointerType(c.getLLVMRuntimeType("_defer"), 0), c.i8ptrType}
 			for _, arg := range callback.Args {
-				llvmType, err := c.getLLVMType(arg.Type())
-				if err != nil {
-					return err
-				}
-				valueTypes = append(valueTypes, llvmType)
+				valueTypes = append(valueTypes, c.getLLVMType(arg.Type()))
 			}
 			deferFrameType := c.ctx.StructType(valueTypes, false)
 			deferFramePtr := c.builder.CreateBitCast(deferData, llvm.PointerType(deferFrameType, 0), "deferFrame")
@@ -233,7 +214,7 @@ func (c *Compiler) emitRunDefers(frame *Frame) error {
 			forwardParams := []llvm.Value{}
 			zero := llvm.ConstInt(c.ctx.Int32Type(), 0, false)
 			for i := 2; i < len(valueTypes); i++ {
-				gep := c.builder.CreateGEP(deferFramePtr, []llvm.Value{zero, llvm.ConstInt(c.ctx.Int32Type(), uint64(i), false)}, "gep")
+				gep := c.builder.CreateInBoundsGEP(deferFramePtr, []llvm.Value{zero, llvm.ConstInt(c.ctx.Int32Type(), uint64(i), false)}, "gep")
 				forwardParam := c.builder.CreateLoad(gep, "param")
 				forwardParams = append(forwardParams, forwardParam)
 			}
@@ -246,23 +227,16 @@ func (c *Compiler) emitRunDefers(frame *Frame) error {
 			// Parent coroutine handle.
 			forwardParams = append(forwardParams, llvm.Undef(c.i8ptrType))
 
-			fnPtr, _, err := c.getInvokeCall(frame, callback)
-			if err != nil {
-				return err
-			}
+			fnPtr, _ := c.getInvokeCall(frame, callback)
 			c.createCall(fnPtr, forwardParams, "")
 
 		case *ir.Function:
 			// Direct call.
 
 			// Get the real defer struct type and cast to it.
-			valueTypes := []llvm.Type{c.uintptrType, llvm.PointerType(c.mod.GetTypeByName("runtime._defer"), 0)}
+			valueTypes := []llvm.Type{c.uintptrType, llvm.PointerType(c.getLLVMRuntimeType("_defer"), 0)}
 			for _, param := range callback.Params {
-				llvmType, err := c.getLLVMType(param.Type())
-				if err != nil {
-					return err
-				}
-				valueTypes = append(valueTypes, llvmType)
+				valueTypes = append(valueTypes, c.getLLVMType(param.Type()))
 			}
 			deferFrameType := c.ctx.StructType(valueTypes, false)
 			deferFramePtr := c.builder.CreateBitCast(deferData, llvm.PointerType(deferFrameType, 0), "deferFrame")
@@ -271,7 +245,7 @@ func (c *Compiler) emitRunDefers(frame *Frame) error {
 			forwardParams := []llvm.Value{}
 			zero := llvm.ConstInt(c.ctx.Int32Type(), 0, false)
 			for i := range callback.Params {
-				gep := c.builder.CreateGEP(deferFramePtr, []llvm.Value{zero, llvm.ConstInt(c.ctx.Int32Type(), uint64(i+2), false)}, "gep")
+				gep := c.builder.CreateInBoundsGEP(deferFramePtr, []llvm.Value{zero, llvm.ConstInt(c.ctx.Int32Type(), uint64(i+2), false)}, "gep")
 				forwardParam := c.builder.CreateLoad(gep, "param")
 				forwardParams = append(forwardParams, forwardParam)
 			}
@@ -289,14 +263,10 @@ func (c *Compiler) emitRunDefers(frame *Frame) error {
 		case *ssa.MakeClosure:
 			// Get the real defer struct type and cast to it.
 			fn := c.ir.GetFunction(callback.Fn.(*ssa.Function))
-			valueTypes := []llvm.Type{c.uintptrType, llvm.PointerType(c.mod.GetTypeByName("runtime._defer"), 0)}
+			valueTypes := []llvm.Type{c.uintptrType, llvm.PointerType(c.getLLVMRuntimeType("_defer"), 0)}
 			params := fn.Signature.Params()
 			for i := 0; i < params.Len(); i++ {
-				llvmType, err := c.getLLVMType(params.At(i).Type())
-				if err != nil {
-					return err
-				}
-				valueTypes = append(valueTypes, llvmType)
+				valueTypes = append(valueTypes, c.getLLVMType(params.At(i).Type()))
 			}
 			valueTypes = append(valueTypes, c.i8ptrType) // closure
 			deferFrameType := c.ctx.StructType(valueTypes, false)
@@ -306,7 +276,7 @@ func (c *Compiler) emitRunDefers(frame *Frame) error {
 			forwardParams := []llvm.Value{}
 			zero := llvm.ConstInt(c.ctx.Int32Type(), 0, false)
 			for i := 2; i < len(valueTypes); i++ {
-				gep := c.builder.CreateGEP(deferFramePtr, []llvm.Value{zero, llvm.ConstInt(c.ctx.Int32Type(), uint64(i), false)}, "")
+				gep := c.builder.CreateInBoundsGEP(deferFramePtr, []llvm.Value{zero, llvm.ConstInt(c.ctx.Int32Type(), uint64(i), false)}, "")
 				forwardParam := c.builder.CreateLoad(gep, "param")
 				forwardParams = append(forwardParams, forwardParam)
 			}
@@ -334,5 +304,4 @@ func (c *Compiler) emitRunDefers(frame *Frame) error {
 
 	// End of loop.
 	c.builder.SetInsertPointAtEnd(end)
-	return nil
 }

compiler/errors.go

@@ -12,3 +12,7 @@ func (c *Compiler) makeError(pos token.Pos, msg string) types.Error {
 		Msg:  msg,
 	}
 }
+
+func (c *Compiler) addError(pos token.Pos, msg string) {
+	c.diagnostics = append(c.diagnostics, c.makeError(pos, msg))
+}

compiler/func-lowering.go

@@ -0,0 +1,285 @@
+package compiler
+
+// This file lowers func values into their final form. This is necessary for
+// funcValueSwitch, which needs full program analysis.
+
+import (
+	"sort"
+	"strconv"
+
+	"tinygo.org/x/go-llvm"
+)
+
+// funcSignatureInfo keeps information about a single signature and its uses.
+type funcSignatureInfo struct {
+	sig                     llvm.Value   // *uint8 to identify the signature
+	funcValueWithSignatures []llvm.Value // slice of runtime.funcValueWithSignature
+}
+
+// funcWithUses keeps information about a single function used as func value and
+// the assigned function ID. More commonly used functions are assigned a lower
+// ID.
+type funcWithUses struct {
+	funcPtr  llvm.Value
+	useCount int // how often this function is used in a func value
+	id       int // assigned ID
+}
+
+// Slice to sort functions by their use counts, or else their name if they're
+// used equally often.
+type funcWithUsesList []*funcWithUses
+
+func (l funcWithUsesList) Len() int { return len(l) }
+func (l funcWithUsesList) Less(i, j int) bool {
+	if l[i].useCount != l[j].useCount {
+		// return the reverse: we want the highest use counts sorted first
+		return l[i].useCount > l[j].useCount
+	}
+	iName := l[i].funcPtr.Name()
+	jName := l[j].funcPtr.Name()
+	return iName < jName
+}
+func (l funcWithUsesList) Swap(i, j int) {
+	l[i], l[j] = l[j], l[i]
+}
+
+// LowerFuncValue lowers the runtime.funcValueWithSignature type and
+// runtime.getFuncPtr function to their final form.
+func (c *Compiler) LowerFuncValues() {
+	if c.funcImplementation() != funcValueSwitch {
+		return
+	}
+
+	// Find all func values used in the program with their signatures.
+	funcValueWithSignaturePtr := llvm.PointerType(c.getLLVMRuntimeType("funcValueWithSignature"), 0)
+	signatures := map[string]*funcSignatureInfo{}
+	for global := c.mod.FirstGlobal(); !global.IsNil(); global = llvm.NextGlobal(global) {
+		if global.Type() != funcValueWithSignaturePtr {
+			continue
+		}
+		sig := llvm.ConstExtractValue(global.Initializer(), []uint32{1})
+		name := sig.Name()
+		if info, ok := signatures[name]; ok {
+			info.funcValueWithSignatures = append(info.funcValueWithSignatures, global)
+		} else {
+			signatures[name] = &funcSignatureInfo{
+				sig:                     sig,
+				funcValueWithSignatures: []llvm.Value{global},
+			}
+		}
+	}
+
+	// Sort the signatures, for deterministic execution.
+	names := make([]string, 0, len(signatures))
+	for name := range signatures {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+
+	for _, name := range names {
+		info := signatures[name]
+		functions := make(funcWithUsesList, len(info.funcValueWithSignatures))
+		for i, use := range info.funcValueWithSignatures {
+			var useCount int
+			for _, use2 := range getUses(use) {
+				useCount += len(getUses(use2))
+			}
+			functions[i] = &funcWithUses{
+				funcPtr:  llvm.ConstExtractValue(use.Initializer(), []uint32{0}).Operand(0),
+				useCount: useCount,
+			}
+		}
+		sort.Sort(functions)
+
+		for i, fn := range functions {
+			fn.id = i + 1
+			for _, ptrtoint := range getUses(fn.funcPtr) {
+				if ptrtoint.IsAConstantExpr().IsNil() || ptrtoint.Opcode() != llvm.PtrToInt {
+					continue
+				}
+				for _, funcValueWithSignatureConstant := range getUses(ptrtoint) {
+					for _, funcValueWithSignatureGlobal := range getUses(funcValueWithSignatureConstant) {
+						for _, use := range getUses(funcValueWithSignatureGlobal) {
+							if ptrtoint.IsAConstantExpr().IsNil() || ptrtoint.Opcode() != llvm.PtrToInt {
+								panic("expected const ptrtoint")
+							}
+							use.ReplaceAllUsesWith(llvm.ConstInt(c.uintptrType, uint64(fn.id), false))
+						}
+					}
+				}
+			}
+		}
+
+		for _, getFuncPtrCall := range getUses(info.sig) {
+			if getFuncPtrCall.IsACallInst().IsNil() {
+				continue
+			}
+			if getFuncPtrCall.CalledValue().Name() != "runtime.getFuncPtr" {
+				panic("expected all call uses to be runtime.getFuncPtr")
+			}
+			funcID := getFuncPtrCall.Operand(1)
+			switch len(functions) {
+			case 0:
+				// There are no functions used in a func value that implement
+				// this signature. The only possible value is a nil value.
+				for _, inttoptr := range getUses(getFuncPtrCall) {
+					if inttoptr.IsAIntToPtrInst().IsNil() {
+						panic("expected inttoptr")
+					}
+					nilptr := llvm.ConstPointerNull(inttoptr.Type())
+					inttoptr.ReplaceAllUsesWith(nilptr)
+					inttoptr.EraseFromParentAsInstruction()
+				}
+				getFuncPtrCall.EraseFromParentAsInstruction()
+			case 1:
+				// There is exactly one function with this signature that is
+				// used in a func value. The func value itself can be either nil
+				// or this one function.
+				c.builder.SetInsertPointBefore(getFuncPtrCall)
+				zero := llvm.ConstInt(c.uintptrType, 0, false)
+				isnil := c.builder.CreateICmp(llvm.IntEQ, funcID, zero, "")
+				funcPtrNil := llvm.ConstPointerNull(functions[0].funcPtr.Type())
+				funcPtr := c.builder.CreateSelect(isnil, funcPtrNil, functions[0].funcPtr, "")
+				for _, inttoptr := range getUses(getFuncPtrCall) {
+					if inttoptr.IsAIntToPtrInst().IsNil() {
+						panic("expected inttoptr")
+					}
+					inttoptr.ReplaceAllUsesWith(funcPtr)
+					inttoptr.EraseFromParentAsInstruction()
+				}
+				getFuncPtrCall.EraseFromParentAsInstruction()
+			default:
+				// There are multiple functions used in a func value that
+				// implement this signature.
+				// What we'll do is transform the following:
+				//     rawPtr := runtime.getFuncPtr(func.ptr)
+				//     if rawPtr == nil {
+				//         runtime.nilPanic()
+				//     }
+				//     result := rawPtr(...args, func.context)
+				// into this:
+				//     if false {
+				//         runtime.nilPanic()
+				//     }
+				//     var result // Phi
+				//     switch fn.id {
+				//     case 0:
+				//         runtime.nilPanic()
+				//     case 1:
+				//         result = call first implementation...
+				//     case 2:
+				//         result = call second implementation...
+				//     default:
+				//         unreachable
+				//     }
+
+				// Remove some casts, checks, and the old call which we're going
+				// to replace.
+				for _, callIntPtr := range getUses(getFuncPtrCall) {
+					if !callIntPtr.IsACallInst().IsNil() && callIntPtr.CalledValue().Name() == "runtime.makeGoroutine" {
+						for _, inttoptr := range getUses(callIntPtr) {
+							if inttoptr.IsAIntToPtrInst().IsNil() {
+								panic("expected a inttoptr")
+							}
+							for _, use := range getUses(inttoptr) {
+								c.addFuncLoweringSwitch(funcID, use, c.emitStartGoroutine, functions)
+								use.EraseFromParentAsInstruction()
+							}
+							inttoptr.EraseFromParentAsInstruction()
+						}
+						callIntPtr.EraseFromParentAsInstruction()
+						continue
+					}
+					if callIntPtr.IsAIntToPtrInst().IsNil() {
+						panic("expected inttoptr")
+					}
+					for _, ptrUse := range getUses(callIntPtr) {
+						if !ptrUse.IsABitCastInst().IsNil() {
+							for _, bitcastUse := range getUses(ptrUse) {
+								if bitcastUse.IsACallInst().IsNil() || bitcastUse.CalledValue().IsAFunction().IsNil() {
+									panic("expected a call instruction")
+								}
+								switch bitcastUse.CalledValue().Name() {
+								case "runtime.isnil":
+									bitcastUse.ReplaceAllUsesWith(llvm.ConstInt(c.ctx.Int1Type(), 0, false))
+									bitcastUse.EraseFromParentAsInstruction()
+								default:
+									panic("expected a call to runtime.isnil")
+								}
+							}
+						} else if !ptrUse.IsACallInst().IsNil() && ptrUse.CalledValue() == callIntPtr {
+							c.addFuncLoweringSwitch(funcID, ptrUse, func(funcPtr llvm.Value, params []llvm.Value) llvm.Value {
+								return c.builder.CreateCall(funcPtr, params, "")
+							}, functions)
+						} else {
+							panic("unexpected getFuncPtrCall")
+						}
+						ptrUse.EraseFromParentAsInstruction()
+					}
+					callIntPtr.EraseFromParentAsInstruction()
+				}
+				getFuncPtrCall.EraseFromParentAsInstruction()
+			}
+		}
+	}
+}
+
+// addFuncLoweringSwitch creates a new switch on a function ID and inserts calls
+// to the newly created direct calls. The funcID is the number to switch on,
+// call is the call instruction to replace, and createCall is the callback that
+// actually creates the new call. By changing createCall to something other than
+// c.builder.CreateCall, instead of calling a function it can start a new
+// goroutine for example.
+func (c *Compiler) addFuncLoweringSwitch(funcID, call llvm.Value, createCall func(funcPtr llvm.Value, params []llvm.Value) llvm.Value, functions funcWithUsesList) {
+	// The block that cannot be reached with correct funcValues (to help the
+	// optimizer).
+	c.builder.SetInsertPointBefore(call)
+	defaultBlock := c.ctx.AddBasicBlock(call.InstructionParent().Parent(), "func.default")
+	c.builder.SetInsertPointAtEnd(defaultBlock)
+	c.builder.CreateUnreachable()
+
+	// Create the switch.
+	c.builder.SetInsertPointBefore(call)
+	sw := c.builder.CreateSwitch(funcID, defaultBlock, len(functions)+1)
+
+	// Split right after the switch. We will need to insert a few basic blocks
+	// in this gap.
+	nextBlock := c.splitBasicBlock(sw, llvm.NextBasicBlock(sw.InstructionParent()), "func.next")
+
+	// The 0 case, which is actually a nil check.
+	nilBlock := c.ctx.InsertBasicBlock(nextBlock, "func.nil")
+	c.builder.SetInsertPointAtEnd(nilBlock)
+	c.createRuntimeCall("nilPanic", nil, "")
+	c.builder.CreateUnreachable()
+	sw.AddCase(llvm.ConstInt(c.uintptrType, 0, false), nilBlock)
+
+	// Gather the list of parameters for every call we're going to make.
+	callParams := make([]llvm.Value, call.OperandsCount()-1)
+	for i := range callParams {
+		callParams[i] = call.Operand(i)
+	}
+
+	// If the call produces a value, we need to get it using a PHI
+	// node.
+	phiBlocks := make([]llvm.BasicBlock, len(functions))
+	phiValues := make([]llvm.Value, len(functions))
+	for i, fn := range functions {
+		// Insert a switch case.
+		bb := c.ctx.InsertBasicBlock(nextBlock, "func.call"+strconv.Itoa(fn.id))
+		c.builder.SetInsertPointAtEnd(bb)
+		result := createCall(fn.funcPtr, callParams)
+		c.builder.CreateBr(nextBlock)
+		sw.AddCase(llvm.ConstInt(c.uintptrType, uint64(fn.id), false), bb)
+		phiBlocks[i] = bb
+		phiValues[i] = result
+	}
+	// Create the PHI node so that the call result flows into the
+	// next block (after the split). This is only necessary when the
+	// call produced a value.
+	if call.Type().TypeKind() != llvm.VoidTypeKind {
+		c.builder.SetInsertPointBefore(nextBlock.FirstInstruction())
+		phi := c.builder.CreatePHI(call.Type(), "")
+		phi.AddIncoming(phiValues, phiBlocks)
+		call.ReplaceAllUsesWith(phi)
+	}
+}

compiler/func.go

@@ -0,0 +1,191 @@
+package compiler
+
+// This file implements function values and closures. It may need some lowering
+// in a later step, see func-lowering.go.
+
+import (
+	"go/types"
+
+	"golang.org/x/tools/go/ssa"
+	"tinygo.org/x/go-llvm"
+)
+
+type funcValueImplementation int
+
+const (
+	funcValueNone funcValueImplementation = iota
+
+	// A func value is implemented as a pair of pointers:
+	//     {context, function pointer}
+	// where the context may be a pointer to a heap-allocated struct containing
+	// the free variables, or it may be undef if the function being pointed to
+	// doesn't need a context. The function pointer is a regular function
+	// pointer.
+	funcValueDoubleword
+
+	// As funcValueDoubleword, but with the function pointer replaced by a
+	// unique ID per function signature. Function values are called by using a
+	// switch statement and choosing which function to call.
+	funcValueSwitch
+)
+
+// funcImplementation picks an appropriate func value implementation for the
+// target.
+func (c *Compiler) funcImplementation() funcValueImplementation {
+	// Always pick the switch implementation, as it allows the use of blocking
+	// inside a function that is used as a func value.
+	switch c.selectScheduler() {
+	case "coroutines":
+		return funcValueSwitch
+	case "tasks":
+		return funcValueDoubleword
+	default:
+		panic("unknown scheduler type")
+	}
+}
+
+// createFuncValue creates a function value from a raw function pointer with no
+// context.
+func (c *Compiler) createFuncValue(funcPtr, context llvm.Value, sig *types.Signature) llvm.Value {
+	var funcValueScalar llvm.Value
+	switch c.funcImplementation() {
+	case funcValueDoubleword:
+		// Closure is: {context, function pointer}
+		funcValueScalar = funcPtr
+	case funcValueSwitch:
+		sigGlobal := c.getTypeCode(sig)
+		funcValueWithSignatureGlobalName := funcPtr.Name() + "$withSignature"
+		funcValueWithSignatureGlobal := c.mod.NamedGlobal(funcValueWithSignatureGlobalName)
+		if funcValueWithSignatureGlobal.IsNil() {
+			funcValueWithSignatureType := c.getLLVMRuntimeType("funcValueWithSignature")
+			funcValueWithSignature := llvm.ConstNamedStruct(funcValueWithSignatureType, []llvm.Value{
+				llvm.ConstPtrToInt(funcPtr, c.uintptrType),
+				sigGlobal,
+			})
+			funcValueWithSignatureGlobal = llvm.AddGlobal(c.mod, funcValueWithSignatureType, funcValueWithSignatureGlobalName)
+			funcValueWithSignatureGlobal.SetInitializer(funcValueWithSignature)
+			funcValueWithSignatureGlobal.SetGlobalConstant(true)
+			funcValueWithSignatureGlobal.SetLinkage(llvm.InternalLinkage)
+		}
+		funcValueScalar = llvm.ConstPtrToInt(funcValueWithSignatureGlobal, c.uintptrType)
+	default:
+		panic("unimplemented func value variant")
+	}
+	funcValueType := c.getFuncType(sig)
+	funcValue := llvm.Undef(funcValueType)
+	funcValue = c.builder.CreateInsertValue(funcValue, context, 0, "")
+	funcValue = c.builder.CreateInsertValue(funcValue, funcValueScalar, 1, "")
+	return funcValue
+}
+
+// extractFuncScalar returns some scalar that can be used in comparisons. It is
+// a cheap operation.
+func (c *Compiler) extractFuncScalar(funcValue llvm.Value) llvm.Value {
+	return c.builder.CreateExtractValue(funcValue, 1, "")
+}
+
+// extractFuncContext extracts the context pointer from this function value. It
+// is a cheap operation.
+func (c *Compiler) extractFuncContext(funcValue llvm.Value) llvm.Value {
+	return c.builder.CreateExtractValue(funcValue, 0, "")
+}
+
+// decodeFuncValue extracts the context and the function pointer from this func
+// value. This may be an expensive operation.
+func (c *Compiler) decodeFuncValue(funcValue llvm.Value, sig *types.Signature) (funcPtr, context llvm.Value) {
+	context = c.builder.CreateExtractValue(funcValue, 0, "")
+	switch c.funcImplementation() {
+	case funcValueDoubleword:
+		funcPtr = c.builder.CreateExtractValue(funcValue, 1, "")
+	case funcValueSwitch:
+		llvmSig := c.getRawFuncType(sig)
+		sigGlobal := c.getTypeCode(sig)
+		funcPtr = c.createRuntimeCall("getFuncPtr", []llvm.Value{funcValue, sigGlobal}, "")
+		funcPtr = c.builder.CreateIntToPtr(funcPtr, llvmSig, "")
+	default:
+		panic("unimplemented func value variant")
+	}
+	return
+}
+
+// getFuncType returns the type of a func value given a signature.
+func (c *Compiler) getFuncType(typ *types.Signature) llvm.Type {
+	switch c.funcImplementation() {
+	case funcValueDoubleword:
+		rawPtr := c.getRawFuncType(typ)
+		return c.ctx.StructType([]llvm.Type{c.i8ptrType, rawPtr}, false)
+	case funcValueSwitch:
+		return c.getLLVMRuntimeType("funcValue")
+	default:
+		panic("unimplemented func value variant")
+	}
+}
+
+// getRawFuncType returns a LLVM function pointer type for a given signature.
+func (c *Compiler) getRawFuncType(typ *types.Signature) llvm.Type {
+	// Get the return type.
+	var returnType llvm.Type
+	switch typ.Results().Len() {
+	case 0:
+		// No return values.
+		returnType = c.ctx.VoidType()
+	case 1:
+		// Just one return value.
+		returnType = c.getLLVMType(typ.Results().At(0).Type())
+	default:
+		// Multiple return values. Put them together in a struct.
+		// This appears to be the common way to handle multiple return values in
+		// LLVM.
+		members := make([]llvm.Type, typ.Results().Len())
+		for i := 0; i < typ.Results().Len(); i++ {
+			members[i] = c.getLLVMType(typ.Results().At(i).Type())
+		}
+		returnType = c.ctx.StructType(members, false)
+	}
+
+	// Get the parameter types.
+	var paramTypes []llvm.Type
+	if typ.Recv() != nil {
+		recv := c.getLLVMType(typ.Recv().Type())
+		if recv.StructName() == "runtime._interface" {
+			// This is a call on an interface, not a concrete type.
+			// The receiver is not an interface, but a i8* type.
+			recv = c.i8ptrType
+		}
+		paramTypes = append(paramTypes, c.expandFormalParamType(recv)...)
+	}
+	for i := 0; i < typ.Params().Len(); i++ {
+		subType := c.getLLVMType(typ.Params().At(i).Type())
+		paramTypes = append(paramTypes, c.expandFormalParamType(subType)...)
+	}
+	// All functions take these parameters at the end.
+	paramTypes = append(paramTypes, c.i8ptrType) // context
+	paramTypes = append(paramTypes, c.i8ptrType) // parent coroutine
+
+	// Make a func type out of the signature.
+	return llvm.PointerType(llvm.FunctionType(returnType, paramTypes, false), c.funcPtrAddrSpace)
+}
+
+// parseMakeClosure makes a function value (with context) from the given
+// closure expression.
+func (c *Compiler) parseMakeClosure(frame *Frame, expr *ssa.MakeClosure) (llvm.Value, error) {
+	if len(expr.Bindings) == 0 {
+		panic("unexpected: MakeClosure without bound variables")
+	}
+	f := c.ir.GetFunction(expr.Fn.(*ssa.Function))
+
+	// Collect all bound variables.
+	boundVars := make([]llvm.Value, len(expr.Bindings))
+	for i, binding := range expr.Bindings {
+		// The context stores the bound variables.
+		llvmBoundVar := c.getValue(frame, binding)
+		boundVars[i] = llvmBoundVar
+	}
+
+	// Store the bound variables in a single object, allocating it on the heap
+	// if necessary.
+	context := c.emitPointerPack(boundVars)
+
+	// Create the closure.
+	return c.createFuncValue(f.LLVMFn, context, f.Signature), nil
+}

compiler/gc.go

@@ -0,0 +1,471 @@
+package compiler
+
+// This file provides IR transformations necessary for precise and portable
+// garbage collectors.
+
+import (
+	"go/token"
+	"math/big"
+
+	"golang.org/x/tools/go/ssa"
+	"tinygo.org/x/go-llvm"
+)
+
+// needsStackObjects returns true if the compiler should insert stack objects
+// that can be traced by the garbage collector.
+func (c *Compiler) needsStackObjects() bool {
+	if c.selectGC() != "conservative" {
+		return false
+	}
+	for _, tag := range c.BuildTags {
+		if tag == "baremetal" {
+			return false
+		}
+	}
+
+	return true
+}
+
+// trackExpr inserts pointer tracking intrinsics for the GC if the expression is
+// one of the expressions that need this.
+func (c *Compiler) trackExpr(frame *Frame, expr ssa.Value, value llvm.Value) {
+	// There are uses of this expression, Make sure the pointers
+	// are tracked during GC.
+	switch expr := expr.(type) {
+	case *ssa.Alloc, *ssa.MakeChan, *ssa.MakeMap:
+		// These values are always of pointer type in IR.
+		c.trackPointer(value)
+	case *ssa.Call, *ssa.Convert, *ssa.MakeClosure, *ssa.MakeInterface, *ssa.MakeSlice, *ssa.Next:
+		if !value.IsNil() {
+			c.trackValue(value)
+		}
+	case *ssa.Select:
+		if alloca, ok := frame.selectRecvBuf[expr]; ok {
+			if alloca.IsAUndefValue().IsNil() {
+				c.trackPointer(alloca)
+			}
+		}
+	case *ssa.UnOp:
+		switch expr.Op {
+		case token.MUL:
+			// Pointer dereference.
+			c.trackValue(value)
+		case token.ARROW:
+			// Channel receive operator.
+			// It's not necessary to look at commaOk here, because in that
+			// case it's just an aggregate and trackValue will extract the
+			// pointer in there (if there is one).
+			c.trackValue(value)
+		}
+	}
+}
+
+// trackValue locates pointers in a value (possibly an aggregate) and tracks the
+// individual pointers
+func (c *Compiler) trackValue(value llvm.Value) {
+	typ := value.Type()
+	switch typ.TypeKind() {
+	case llvm.PointerTypeKind:
+		c.trackPointer(value)
+	case llvm.StructTypeKind:
+		if !typeHasPointers(typ) {
+			return
+		}
+		numElements := typ.StructElementTypesCount()
+		for i := 0; i < numElements; i++ {
+			subValue := c.builder.CreateExtractValue(value, i, "")
+			c.trackValue(subValue)
+		}
+	case llvm.ArrayTypeKind:
+		if !typeHasPointers(typ) {
+			return
+		}
+		numElements := typ.ArrayLength()
+		for i := 0; i < numElements; i++ {
+			subValue := c.builder.CreateExtractValue(value, i, "")
+			c.trackValue(subValue)
+		}
+	}
+}
+
+// trackPointer creates a call to runtime.trackPointer, bitcasting the poitner
+// first if needed. The input value must be of LLVM pointer type.
+func (c *Compiler) trackPointer(value llvm.Value) {
+	if value.Type() != c.i8ptrType {
+		value = c.builder.CreateBitCast(value, c.i8ptrType, "")
+	}
+	c.createRuntimeCall("trackPointer", []llvm.Value{value}, "")
+}
+
+// typeHasPointers returns whether this type is a pointer or contains pointers.
+// If the type is an aggregate type, it will check whether there is a pointer
+// inside.
+func typeHasPointers(t llvm.Type) bool {
+	switch t.TypeKind() {
+	case llvm.PointerTypeKind:
+		return true
+	case llvm.StructTypeKind:
+		for _, subType := range t.StructElementTypes() {
+			if typeHasPointers(subType) {
+				return true
+			}
+		}
+		return false
+	case llvm.ArrayTypeKind:
+		if typeHasPointers(t.ElementType()) {
+			return true
+		}
+		return false
+	default:
+		return false
+	}
+}
+
+// makeGCStackSlots converts all calls to runtime.trackPointer to explicit
+// stores to stack slots that are scannable by the GC.
+func (c *Compiler) makeGCStackSlots() bool {
+	// Check whether there are allocations at all.
+	alloc := c.mod.NamedFunction("runtime.alloc")
+	if alloc.IsNil() {
+		// Nothing to. Make sure all remaining bits and pieces for stack
+		// chains are neutralized.
+		for _, call := range getUses(c.mod.NamedFunction("runtime.trackPointer")) {
+			call.EraseFromParentAsInstruction()
+		}
+		stackChainStart := c.mod.NamedGlobal("runtime.stackChainStart")
+		if !stackChainStart.IsNil() {
+			stackChainStart.SetInitializer(llvm.ConstNull(stackChainStart.Type().ElementType()))
+			stackChainStart.SetGlobalConstant(true)
+		}
+	}
+
+	trackPointer := c.mod.NamedFunction("runtime.trackPointer")
+	if trackPointer.IsNil() || trackPointer.FirstUse().IsNil() {
+		return false // nothing to do
+	}
+
+	// Look at *all* functions to see whether they are free of function pointer
+	// calls.
+	// This takes less than 5ms for ~100kB of WebAssembly but would perhaps be
+	// faster when written in C++ (to avoid the CGo overhead).
+	funcsWithFPCall := map[llvm.Value]struct{}{}
+	n := 0
+	for fn := c.mod.FirstFunction(); !fn.IsNil(); fn = llvm.NextFunction(fn) {
+		n++
+		if _, ok := funcsWithFPCall[fn]; ok {
+			continue // already found
+		}
+		done := false
+		for bb := fn.FirstBasicBlock(); !bb.IsNil() && !done; bb = llvm.NextBasicBlock(bb) {
+			for call := bb.FirstInstruction(); !call.IsNil() && !done; call = llvm.NextInstruction(call) {
+				if call.IsACallInst().IsNil() {
+					continue // only looking at calls
+				}
+				called := call.CalledValue()
+				if !called.IsAFunction().IsNil() {
+					continue // only looking for function pointers
+				}
+				funcsWithFPCall[fn] = struct{}{}
+				markParentFunctions(funcsWithFPCall, fn)
+				done = true
+			}
+		}
+	}
+
+	// Determine which functions need stack objects. Many leaf functions don't
+	// need it: it only causes overhead for them.
+	// Actually, in one test it was only able to eliminate stack object from 12%
+	// of functions that had a call to runtime.trackPointer (8 out of 68
+	// functions), so this optimization is not as big as it may seem.
+	allocatingFunctions := map[llvm.Value]struct{}{} // set of allocating functions
+
+	// Work from runtime.alloc and trace all parents to check which functions do
+	// a heap allocation (and thus which functions do not).
+	markParentFunctions(allocatingFunctions, alloc)
+
+	// Also trace all functions that call a function pointer.
+	for fn := range funcsWithFPCall {
+		// Assume that functions that call a function pointer do a heap
+		// allocation as a conservative guess because the called function might
+		// do a heap allocation.
+		allocatingFunctions[fn] = struct{}{}
+		markParentFunctions(allocatingFunctions, fn)
+	}
+
+	// Collect some variables used below in the loop.
+	stackChainStart := c.mod.NamedGlobal("runtime.stackChainStart")
+	if stackChainStart.IsNil() {
+		panic("stack chain start not found!")
+	}
+	stackChainStartType := stackChainStart.Type().ElementType()
+	stackChainStart.SetInitializer(llvm.ConstNull(stackChainStartType))
+
+	// Iterate until runtime.trackPointer has no uses left.
+	for use := trackPointer.FirstUse(); !use.IsNil(); use = trackPointer.FirstUse() {
+		// Pick the first use of runtime.trackPointer.
+		call := use.User()
+		if call.IsACallInst().IsNil() {
+			panic("expected runtime.trackPointer use to be a call")
+		}
+
+		// Pick the parent function.
+		fn := call.InstructionParent().Parent()
+
+		if _, ok := allocatingFunctions[fn]; !ok {
+			// This function nor any of the functions it calls (recursively)
+			// allocate anything from the heap, so it will not trigger a garbage
+			// collection cycle. Thus, it does not need to track local pointer
+			// values.
+			// This is a useful optimization but not as big as you might guess,
+			// as described above (it avoids stack objects for ~12% of
+			// functions).
+			call.EraseFromParentAsInstruction()
+			continue
+		}
+
+		// Find all calls to runtime.trackPointer in this function.
+		var calls []llvm.Value
+		var returns []llvm.Value
+		for bb := fn.FirstBasicBlock(); !bb.IsNil(); bb = llvm.NextBasicBlock(bb) {
+			for inst := bb.FirstInstruction(); !inst.IsNil(); inst = llvm.NextInstruction(inst) {
+				switch inst.InstructionOpcode() {
+				case llvm.Call:
+					if inst.CalledValue() == trackPointer {
+						calls = append(calls, inst)
+					}
+				case llvm.Ret:
+					returns = append(returns, inst)
+				}
+			}
+		}
+
+		// Determine what to do with each call.
+		var allocas, pointers []llvm.Value
+		for _, call := range calls {
+			ptr := call.Operand(0)
+			call.EraseFromParentAsInstruction()
+			if ptr.IsAInstruction().IsNil() {
+				continue
+			}
+
+			// Some trivial optimizations.
+			if ptr.IsAInstruction().IsNil() {
+				continue
+			}
+			switch ptr.InstructionOpcode() {
+			case llvm.PHI, llvm.GetElementPtr:
+				// These values do not create new values: the values already
+				// existed locally in this function so must have been tracked
+				// already.
+				continue
+			case llvm.ExtractValue, llvm.BitCast:
+				// These instructions do not create new values, but their
+				// original value may not be tracked. So keep tracking them for
+				// now.
+				// With more analysis, it should be possible to optimize a
+				// significant chunk of these away.
+			case llvm.Call, llvm.Load, llvm.IntToPtr:
+				// These create new values so must be stored locally. But
+				// perhaps some of these can be fused when they actually refer
+				// to the same value.
+			default:
+				// Ambiguous. These instructions are uncommon, but perhaps could
+				// be optimized if needed.
+			}
+
+			if !ptr.IsAAllocaInst().IsNil() {
+				if typeHasPointers(ptr.Type().ElementType()) {
+					allocas = append(allocas, ptr)
+				}
+			} else {
+				pointers = append(pointers, ptr)
+			}
+		}
+
+		if len(allocas) == 0 && len(pointers) == 0 {
+			// This function does not need to keep track of stack pointers.
+			continue
+		}
+
+		// Determine the type of the required stack slot.
+		fields := []llvm.Type{
+			stackChainStartType, // Pointer to parent frame.
+			c.uintptrType,       // Number of elements in this frame.
+		}
+		for _, alloca := range allocas {
+			fields = append(fields, alloca.Type().ElementType())
+		}
+		for _, ptr := range pointers {
+			fields = append(fields, ptr.Type())
+		}
+		stackObjectType := c.ctx.StructType(fields, false)
+
+		// Create the stack object at the function entry.
+		c.builder.SetInsertPointBefore(fn.EntryBasicBlock().FirstInstruction())
+		stackObject := c.builder.CreateAlloca(stackObjectType, "gc.stackobject")
+		initialStackObject := llvm.ConstNull(stackObjectType)
+		numSlots := (c.targetData.TypeAllocSize(stackObjectType) - c.targetData.TypeAllocSize(c.i8ptrType)*2) / uint64(c.targetData.ABITypeAlignment(c.uintptrType))
+		numSlotsValue := llvm.ConstInt(c.uintptrType, numSlots, false)
+		initialStackObject = llvm.ConstInsertValue(initialStackObject, numSlotsValue, []uint32{1})
+		c.builder.CreateStore(initialStackObject, stackObject)
+
+		// Update stack start.
+		parent := c.builder.CreateLoad(stackChainStart, "")
+		gep := c.builder.CreateGEP(stackObject, []llvm.Value{
+			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+		}, "")
+		c.builder.CreateStore(parent, gep)
+		stackObjectCast := c.builder.CreateBitCast(stackObject, stackChainStartType, "")
+		c.builder.CreateStore(stackObjectCast, stackChainStart)
+
+		// Replace all independent allocas with GEPs in the stack object.
+		for i, alloca := range allocas {
+			gep := c.builder.CreateGEP(stackObject, []llvm.Value{
+				llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+				llvm.ConstInt(c.ctx.Int32Type(), uint64(2+i), false),
+			}, "")
+			alloca.ReplaceAllUsesWith(gep)
+			alloca.EraseFromParentAsInstruction()
+		}
+
+		// Do a store to the stack object after each new pointer that is created.
+		for i, ptr := range pointers {
+			c.builder.SetInsertPointBefore(llvm.NextInstruction(ptr))
+			gep := c.builder.CreateGEP(stackObject, []llvm.Value{
+				llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+				llvm.ConstInt(c.ctx.Int32Type(), uint64(2+len(allocas)+i), false),
+			}, "")
+			c.builder.CreateStore(ptr, gep)
+		}
+
+		// Make sure this stack object is popped from the linked list of stack
+		// objects at return.
+		for _, ret := range returns {
+			c.builder.SetInsertPointBefore(ret)
+			c.builder.CreateStore(parent, stackChainStart)
+		}
+	}
+
+	return true
+}
+
+func (c *Compiler) addGlobalsBitmap() bool {
+	if c.mod.NamedGlobal("runtime.trackedGlobalsStart").IsNil() {
+		return false // nothing to do: no GC in use
+	}
+
+	var trackedGlobals []llvm.Value
+	var trackedGlobalTypes []llvm.Type
+	for global := c.mod.FirstGlobal(); !global.IsNil(); global = llvm.NextGlobal(global) {
+		if global.IsDeclaration() {
+			continue
+		}
+		typ := global.Type().ElementType()
+		ptrs := c.getPointerBitmap(typ, global.Name())
+		if ptrs.BitLen() == 0 {
+			continue
+		}
+		trackedGlobals = append(trackedGlobals, global)
+		trackedGlobalTypes = append(trackedGlobalTypes, typ)
+	}
+
+	globalsBundleType := c.ctx.StructType(trackedGlobalTypes, false)
+	globalsBundle := llvm.AddGlobal(c.mod, globalsBundleType, "tinygo.trackedGlobals")
+	globalsBundle.SetLinkage(llvm.InternalLinkage)
+	globalsBundle.SetUnnamedAddr(true)
+	initializer := llvm.Undef(globalsBundleType)
+	for i, global := range trackedGlobals {
+		initializer = llvm.ConstInsertValue(initializer, global.Initializer(), []uint32{uint32(i)})
+		gep := llvm.ConstGEP(globalsBundle, []llvm.Value{
+			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+			llvm.ConstInt(c.ctx.Int32Type(), uint64(i), false),
+		})
+		global.ReplaceAllUsesWith(gep)
+		global.EraseFromParentAsGlobal()
+	}
+	globalsBundle.SetInitializer(initializer)
+
+	trackedGlobalsStart := llvm.ConstPtrToInt(globalsBundle, c.uintptrType)
+	c.mod.NamedGlobal("runtime.trackedGlobalsStart").SetInitializer(trackedGlobalsStart)
+
+	alignment := c.targetData.PrefTypeAlignment(c.i8ptrType)
+	trackedGlobalsLength := llvm.ConstInt(c.uintptrType, c.targetData.TypeAllocSize(globalsBundleType)/uint64(alignment), false)
+	c.mod.NamedGlobal("runtime.trackedGlobalsLength").SetInitializer(trackedGlobalsLength)
+
+	bitmapBytes := c.getPointerBitmap(globalsBundleType, "globals bundle").Bytes()
+	bitmapValues := make([]llvm.Value, len(bitmapBytes))
+	for i, b := range bitmapBytes {
+		bitmapValues[len(bitmapBytes)-i-1] = llvm.ConstInt(c.ctx.Int8Type(), uint64(b), false)
+	}
+	bitmapArray := llvm.ConstArray(c.ctx.Int8Type(), bitmapValues)
+	bitmapNew := llvm.AddGlobal(c.mod, bitmapArray.Type(), "runtime.trackedGlobalsBitmap.tmp")
+	bitmapOld := c.mod.NamedGlobal("runtime.trackedGlobalsBitmap")
+	bitmapOld.ReplaceAllUsesWith(llvm.ConstBitCast(bitmapNew, bitmapOld.Type()))
+	bitmapNew.SetInitializer(bitmapArray)
+	bitmapNew.SetName("runtime.trackedGlobalsBitmap")
+
+	return true // the IR was changed
+}
+
+func (c *Compiler) getPointerBitmap(typ llvm.Type, name string) *big.Int {
+	alignment := c.targetData.PrefTypeAlignment(c.i8ptrType)
+	switch typ.TypeKind() {
+	case llvm.IntegerTypeKind, llvm.FloatTypeKind, llvm.DoubleTypeKind:
+		return big.NewInt(0)
+	case llvm.PointerTypeKind:
+		return big.NewInt(1)
+	case llvm.StructTypeKind:
+		ptrs := big.NewInt(0)
+		for i, subtyp := range typ.StructElementTypes() {
+			subptrs := c.getPointerBitmap(subtyp, name)
+			if subptrs.BitLen() == 0 {
+				continue
+			}
+			offset := c.targetData.ElementOffset(typ, i)
+			if offset%uint64(alignment) != 0 {
+				panic("precise GC: global contains unaligned pointer: " + name)
+			}
+			subptrs.Lsh(subptrs, uint(offset)/uint(alignment))
+			ptrs.Or(ptrs, subptrs)
+		}
+		return ptrs
+	case llvm.ArrayTypeKind:
+		subtyp := typ.ElementType()
+		subptrs := c.getPointerBitmap(subtyp, name)
+		ptrs := big.NewInt(0)
+		if subptrs.BitLen() == 0 {
+			return ptrs
+		}
+		elementSize := c.targetData.TypeAllocSize(subtyp)
+		for i := 0; i < typ.ArrayLength(); i++ {
+			ptrs.Lsh(ptrs, uint(elementSize)/uint(alignment))
+			ptrs.Or(ptrs, subptrs)
+		}
+		return ptrs
+	default:
+		panic("unknown type kind of global: " + name)
+	}
+}
+
+// markParentFunctions traverses all parent function calls (recursively) and
+// adds them to the set of marked functions. It only considers function calls:
+// any other uses of such a function is ignored.
+func markParentFunctions(marked map[llvm.Value]struct{}, fn llvm.Value) {
+	worklist := []llvm.Value{fn}
+	for len(worklist) != 0 {
+		fn := worklist[len(worklist)-1]
+		worklist = worklist[:len(worklist)-1]
+		for _, use := range getUses(fn) {
+			if use.IsACallInst().IsNil() || use.CalledValue() != fn {
+				// Not the parent function.
+				continue
+			}
+			parent := use.InstructionParent().Parent()
+			if _, ok := marked[parent]; !ok {
+				marked[parent] = struct{}{}
+				worklist = append(worklist, parent)
+			}
+		}
+	}
+}

compiler/goroutine-lowering.go

@@ -1,5 +1,16 @@
 package compiler
 
+// This file implements lowering for the goroutine scheduler. There are two
+// scheduler implementations, one based on tasks (like RTOSes and the main Go
+// runtime) and one based on a coroutine compiler transformation. The task based
+// implementation requires very little work from the compiler but is not very
+// portable (in particular, it is very hard if not impossible to support on
+// WebAssembly). The coroutine based one requires a lot of work by the compiler
+// to implement, but can run virtually anywhere with a single scheduler
+// implementation.
+//
+// The below description is for the coroutine based scheduler.
+//
 // This file lowers goroutine pseudo-functions into coroutines scheduled by a
 // scheduler at runtime. It uses coroutine support in LLVM for this
 // transformation: https://llvm.org/docs/Coroutines.html
@@ -10,8 +21,8 @@ package compiler
 //         go foo()
 //         time.Sleep(2 * time.Second)
 //         println("some other operation")
-//         bar()
-//         println("done")
+//         i := bar()
+//         println("done", *i)
 //     }
 //
 //     func foo() {
@@ -21,9 +32,10 @@ package compiler
 //         }
 //     }
 //
-//     func bar() {
+//     func bar() *int {
 //         time.Sleep(time.Second)
 //         println("blocking operation completed)
+//         return new(int)
 //     }
 //
 // It is transformed by the IR generator in compiler.go into the following
@@ -34,8 +46,8 @@ package compiler
 //         fn()
 //         time.Sleep(2 * time.Second)
 //         println("some other operation")
-//         bar() // imagine an 'await' keyword in front of this call
-//         println("done")
+//         i := bar() // imagine an 'await' keyword in front of this call
+//         println("done", *i)
 //     }
 //
 //     func foo() {
@@ -45,9 +57,10 @@ package compiler
 //         }
 //     }
 //
-//     func bar() {
+//     func bar() *int {
 //         time.Sleep(time.Second)
 //         println("blocking operation completed)
+//         return new(int)
 //     }
 //
 // The pass in this file transforms this code even further, to the following
@@ -59,9 +72,11 @@ package compiler
 //         runtime.sleepTask(hdl, 2 * time.Second) // ask the scheduler to re-activate this coroutine at the right time
 //         llvm.suspend(hdl)                       // suspend point
 //         println("some other operation")
+//         var i *int                              // allocate space on the stack for the return value
+//         runtime.setTaskStatePtr(hdl, &i)        // store return value alloca in our coroutine promise
 //         bar(hdl)                                // await, pass a continuation (hdl) to bar
 //         llvm.suspend(hdl)                       // suspend point, wait for the callee to re-activate
-//         println("done")
+//         println("done", *i)
 //         runtime.activateTask(parent)            // re-activate the parent (nop, there is no parent)
 //     }
 //
@@ -102,10 +117,69 @@ type asyncFunc struct {
 	unreachableBlock llvm.BasicBlock
 }
 
-// LowerGoroutines is a pass called during optimization that transforms the IR
-// into one where all blocking functions are turned into goroutines and blocking
-// calls into await calls.
+// LowerGoroutines performs some IR transformations necessary to support
+// goroutines. It does something different based on whether it uses the
+// coroutine or the tasks implementation of goroutines, and whether goroutines
+// are necessary at all.
 func (c *Compiler) LowerGoroutines() error {
+	switch c.selectScheduler() {
+	case "coroutines":
+		return c.lowerCoroutines()
+	case "tasks":
+		return c.lowerTasks()
+	default:
+		panic("unknown scheduler type")
+	}
+}
+
+// lowerTasks starts the main goroutine and then runs the scheduler.
+// This is enough compiler-level transformation for the task-based scheduler.
+func (c *Compiler) lowerTasks() error {
+	uses := getUses(c.mod.NamedFunction("runtime.callMain"))
+	if len(uses) != 1 || uses[0].IsACallInst().IsNil() {
+		panic("expected exactly 1 call of runtime.callMain, check the entry point")
+	}
+	mainCall := uses[0]
+
+	realMain := c.mod.NamedFunction(c.ir.MainPkg().Pkg.Path() + ".main")
+	if len(getUses(c.mod.NamedFunction("runtime.startGoroutine"))) != 0 {
+		// Program needs a scheduler. Start main.main as a goroutine and start
+		// the scheduler.
+		realMainWrapper := c.createGoroutineStartWrapper(realMain)
+		c.builder.SetInsertPointBefore(mainCall)
+		zero := llvm.ConstInt(c.uintptrType, 0, false)
+		c.createRuntimeCall("startGoroutine", []llvm.Value{realMainWrapper, zero}, "")
+		c.createRuntimeCall("scheduler", nil, "")
+		sleep := c.mod.NamedFunction("time.Sleep")
+		if !sleep.IsNil() {
+			sleep.ReplaceAllUsesWith(c.mod.NamedFunction("runtime.sleepCurrentTask"))
+		}
+	} else {
+		// Program doesn't need a scheduler. Call main.main directly.
+		c.builder.SetInsertPointBefore(mainCall)
+		params := []llvm.Value{
+			llvm.Undef(c.i8ptrType), // unused context parameter
+			llvm.Undef(c.i8ptrType), // unused coroutine handle
+		}
+		c.createCall(realMain, params, "")
+		// runtime.Goexit isn't needed so let it be optimized away by
+		// globalopt.
+		c.mod.NamedFunction("runtime.Goexit").SetLinkage(llvm.InternalLinkage)
+	}
+	mainCall.EraseFromParentAsInstruction()
+
+	// main.main was set to external linkage during IR construction. Set it to
+	// internal linkage to enable interprocedural optimizations.
+	realMain.SetLinkage(llvm.InternalLinkage)
+
+	return nil
+}
+
+// lowerCoroutines transforms the IR into one where all blocking functions are
+// turned into goroutines and blocking calls into await calls. It also makes
+// sure that the first coroutine is started and the coroutine scheduler will be
+// run.
+func (c *Compiler) lowerCoroutines() error {
 	needsScheduler, err := c.markAsyncFunctions()
 	if err != nil {
 		return err
@@ -140,13 +214,6 @@ func (c *Compiler) LowerGoroutines() error {
 	// main.main was set to external linkage during IR construction. Set it to
 	// internal linkage to enable interprocedural optimizations.
 	realMain.SetLinkage(llvm.InternalLinkage)
-	c.mod.NamedFunction("runtime.alloc").SetLinkage(llvm.InternalLinkage)
-	c.mod.NamedFunction("runtime.free").SetLinkage(llvm.InternalLinkage)
-	c.mod.NamedFunction("runtime.chanSend").SetLinkage(llvm.InternalLinkage)
-	c.mod.NamedFunction("runtime.chanRecv").SetLinkage(llvm.InternalLinkage)
-	c.mod.NamedFunction("runtime.sleepTask").SetLinkage(llvm.InternalLinkage)
-	c.mod.NamedFunction("runtime.activateTask").SetLinkage(llvm.InternalLinkage)
-	c.mod.NamedFunction("runtime.scheduler").SetLinkage(llvm.InternalLinkage)
 
 	return nil
 }
@@ -170,13 +237,17 @@ func (c *Compiler) markAsyncFunctions() (needsScheduler bool, err error) {
 	if !sleep.IsNil() {
 		worklist = append(worklist, sleep)
 	}
-	chanSendStub := c.mod.NamedFunction("runtime.chanSendStub")
-	if !chanSendStub.IsNil() {
-		worklist = append(worklist, chanSendStub)
+	deadlock := c.mod.NamedFunction("runtime.deadlock")
+	if !deadlock.IsNil() {
+		worklist = append(worklist, deadlock)
 	}
-	chanRecvStub := c.mod.NamedFunction("runtime.chanRecvStub")
-	if !chanRecvStub.IsNil() {
-		worklist = append(worklist, chanRecvStub)
+	chanSend := c.mod.NamedFunction("runtime.chanSend")
+	if !chanSend.IsNil() {
+		worklist = append(worklist, chanSend)
+	}
+	chanRecv := c.mod.NamedFunction("runtime.chanRecv")
+	if !chanRecv.IsNil() {
+		worklist = append(worklist, chanRecv)
 	}
 
 	if len(worklist) == 0 {
@@ -204,17 +275,25 @@ func (c *Compiler) markAsyncFunctions() (needsScheduler bool, err error) {
 
 		// Add all callees to the worklist.
 		for _, use := range getUses(f) {
-			if use.IsConstant() && use.Opcode() == llvm.BitCast {
-				bitcastUses := getUses(use)
-				for _, call := range bitcastUses {
+			if use.IsConstant() && use.Opcode() == llvm.PtrToInt {
+				for _, call := range getUses(use) {
 					if call.IsACallInst().IsNil() || call.CalledValue().Name() != "runtime.makeGoroutine" {
-						return false, errors.New("async function " + f.Name() + " incorrectly used in bitcast, expected runtime.makeGoroutine")
+						return false, errors.New("async function " + f.Name() + " incorrectly used in ptrtoint, expected runtime.makeGoroutine")
 					}
 				}
 				// This is a go statement. Do not mark the parent as async, as
 				// starting a goroutine is not a blocking operation.
 				continue
 			}
+			if use.IsConstant() && use.Opcode() == llvm.BitCast {
+				// Not sure why this const bitcast is here but as long as it
+				// has no uses it can be ignored, I guess?
+				// I think it was created for the runtime.isnil check but
+				// somehow wasn't removed when all these checks are removed.
+				if len(getUses(use)) == 0 {
+					continue
+				}
+			}
 			if use.IsACallInst().IsNil() {
 				// Not a call instruction. Maybe a store to a global? In any
 				// case, this requires support for async calls across function
@@ -243,12 +322,12 @@ func (c *Compiler) markAsyncFunctions() (needsScheduler bool, err error) {
 		// goroutine is not async (does not do any blocking operation), no
 		// scheduler is necessary as it can be called directly.
 		for _, use := range getUses(makeGoroutine) {
-			// Input param must be const bitcast of function.
-			bitcast := use.Operand(0)
-			if !bitcast.IsConstant() || bitcast.Opcode() != llvm.BitCast {
-				panic("expected const bitcast operand of runtime.makeGoroutine")
+			// Input param must be const ptrtoint of function.
+			ptrtoint := use.Operand(0)
+			if !ptrtoint.IsConstant() || ptrtoint.Opcode() != llvm.PtrToInt {
+				panic("expected const ptrtoint operand of runtime.makeGoroutine")
 			}
-			goroutine := bitcast.Operand(0)
+			goroutine := ptrtoint.Operand(0)
 			if _, ok := asyncFuncs[goroutine]; ok {
 				needsScheduler = true
 				break
@@ -274,9 +353,6 @@ func (c *Compiler) markAsyncFunctions() (needsScheduler bool, err error) {
 	coroBeginType := llvm.FunctionType(c.i8ptrType, []llvm.Type{c.ctx.TokenType(), c.i8ptrType}, false)
 	coroBeginFunc := llvm.AddFunction(c.mod, "llvm.coro.begin", coroBeginType)
 
-	coroPromiseType := llvm.FunctionType(c.i8ptrType, []llvm.Type{c.i8ptrType, c.ctx.Int32Type(), c.ctx.Int1Type()}, false)
-	coroPromiseFunc := llvm.AddFunction(c.mod, "llvm.coro.promise", coroPromiseType)
-
 	coroSuspendType := llvm.FunctionType(c.ctx.Int8Type(), []llvm.Type{c.ctx.TokenType(), c.ctx.Int1Type()}, false)
 	coroSuspendFunc := llvm.AddFunction(c.mod, "llvm.coro.suspend", coroSuspendType)
 
@@ -288,7 +364,7 @@ func (c *Compiler) markAsyncFunctions() (needsScheduler bool, err error) {
 
 	// Transform all async functions into coroutines.
 	for _, f := range asyncList {
-		if f == sleep || f == chanSendStub || f == chanRecvStub {
+		if f == sleep || f == deadlock || f == chanSend || f == chanRecv {
 			continue
 		}
 
@@ -305,7 +381,7 @@ func (c *Compiler) markAsyncFunctions() (needsScheduler bool, err error) {
 			for inst := bb.FirstInstruction(); !inst.IsNil(); inst = llvm.NextInstruction(inst) {
 				if !inst.IsACallInst().IsNil() {
 					callee := inst.CalledValue()
-					if _, ok := asyncFuncs[callee]; !ok || callee == sleep || callee == chanSendStub || callee == chanRecvStub {
+					if _, ok := asyncFuncs[callee]; !ok || callee == sleep || callee == deadlock || callee == chanSend || callee == chanRecv {
 						continue
 					}
 					asyncCalls = append(asyncCalls, inst)
@@ -317,7 +393,7 @@ func (c *Compiler) markAsyncFunctions() (needsScheduler bool, err error) {
 
 		// Coroutine setup.
 		c.builder.SetInsertPointBefore(f.EntryBasicBlock().FirstInstruction())
-		taskState := c.builder.CreateAlloca(c.mod.GetTypeByName("runtime.taskState"), "task.state")
+		taskState := c.builder.CreateAlloca(c.getLLVMRuntimeType("taskState"), "task.state")
 		stateI8 := c.builder.CreateBitCast(taskState, c.i8ptrType, "task.state.i8")
 		id := c.builder.CreateCall(coroIdFunc, []llvm.Value{
 			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
@@ -332,6 +408,9 @@ func (c *Compiler) markAsyncFunctions() (needsScheduler bool, err error) {
 			size = c.builder.CreateZExt(size, c.uintptrType, "task.size.uintptr")
 		}
 		data := c.createRuntimeCall("alloc", []llvm.Value{size}, "task.data")
+		if c.needsStackObjects() {
+			c.trackPointer(data)
+		}
 		frame.taskHandle = c.builder.CreateCall(coroBeginFunc, []llvm.Value{id, data}, "task.handle")
 
 		// Modify async calls so this function suspends right after the child
@@ -343,10 +422,18 @@ func (c *Compiler) markAsyncFunctions() (needsScheduler bool, err error) {
 			// Split this basic block.
 			await := c.splitBasicBlock(inst, llvm.NextBasicBlock(c.builder.GetInsertBlock()), "task.await")
 
-			// Set task state to TASK_STATE_CALL.
-			c.builder.SetInsertPointAtEnd(inst.InstructionParent())
+			// Allocate space for the return value.
+			var retvalAlloca llvm.Value
+			if inst.Type().TypeKind() != llvm.VoidTypeKind {
+				c.builder.SetInsertPointBefore(inst.InstructionParent().Parent().EntryBasicBlock().FirstInstruction())
+				retvalAlloca = c.builder.CreateAlloca(inst.Type(), "coro.retvalAlloca")
+				c.builder.SetInsertPointBefore(inst)
+				data := c.builder.CreateBitCast(retvalAlloca, c.i8ptrType, "")
+				c.createRuntimeCall("setTaskStatePtr", []llvm.Value{frame.taskHandle, data}, "")
+			}
 
 			// Suspend.
+			c.builder.SetInsertPointAtEnd(inst.InstructionParent())
 			continuePoint := c.builder.CreateCall(coroSuspendFunc, []llvm.Value{
 				llvm.ConstNull(c.ctx.TokenType()),
 				llvm.ConstInt(c.ctx.Int1Type(), 0, false),
@@ -354,44 +441,76 @@ func (c *Compiler) markAsyncFunctions() (needsScheduler bool, err error) {
 			sw := c.builder.CreateSwitch(continuePoint, frame.suspendBlock, 2)
 			sw.AddCase(llvm.ConstInt(c.ctx.Int8Type(), 0, false), await)
 			sw.AddCase(llvm.ConstInt(c.ctx.Int8Type(), 1, false), frame.cleanupBlock)
+
+			if inst.Type().TypeKind() != llvm.VoidTypeKind {
+				// Load the return value from the alloca. The callee has
+				// written the return value to it.
+				c.builder.SetInsertPointBefore(await.FirstInstruction())
+				retval := c.builder.CreateLoad(retvalAlloca, "coro.retval")
+				inst.ReplaceAllUsesWith(retval)
+			}
 		}
 
 		// Replace return instructions with suspend points that should
 		// reactivate the parent coroutine.
 		for _, inst := range returns {
-			if inst.OperandsCount() == 0 {
-				// These properties were added by the functionattrs pass.
-				// Remove them, because now we start using the parameter.
-				// https://llvm.org/docs/Passes.html#functionattrs-deduce-function-attributes
-				for _, kind := range []string{"nocapture", "readnone"} {
-					kindID := llvm.AttributeKindID(kind)
-					f.RemoveEnumAttributeAtIndex(f.ParamsCount(), kindID)
-				}
-
-				// Reactivate the parent coroutine. This adds it back to
-				// the run queue, so it is started again by the
-				// scheduler when possible (possibly right after the
-				// following suspend).
-				c.builder.SetInsertPointBefore(inst)
-
-				parentHandle := f.LastParam()
-				c.createRuntimeCall("activateTask", []llvm.Value{parentHandle}, "")
-
-				// Suspend this coroutine.
-				// It would look like this is unnecessary, but if this
-				// suspend point is left out, it leads to undefined
-				// behavior somehow (with the unreachable instruction).
-				continuePoint := c.builder.CreateCall(coroSuspendFunc, []llvm.Value{
-					llvm.ConstNull(c.ctx.TokenType()),
-					llvm.ConstInt(c.ctx.Int1Type(), 1, false),
-				}, "ret")
-				sw := c.builder.CreateSwitch(continuePoint, frame.suspendBlock, 2)
-				sw.AddCase(llvm.ConstInt(c.ctx.Int8Type(), 0, false), frame.unreachableBlock)
-				sw.AddCase(llvm.ConstInt(c.ctx.Int8Type(), 1, false), frame.cleanupBlock)
-				inst.EraseFromParentAsInstruction()
-			} else {
-				panic("todo: return value from coroutine")
+			// These properties were added by the functionattrs pass. Remove
+			// them, because now we start using the parameter.
+			// https://llvm.org/docs/Passes.html#functionattrs-deduce-function-attributes
+			for _, kind := range []string{"nocapture", "readnone"} {
+				kindID := llvm.AttributeKindID(kind)
+				f.RemoveEnumAttributeAtIndex(f.ParamsCount(), kindID)
 			}
+
+			c.builder.SetInsertPointBefore(inst)
+
+			var parentHandle llvm.Value
+			if f.Linkage() == llvm.ExternalLinkage {
+				// Exported function.
+				// Note that getTaskStatePtr will panic if it is called with
+				// a nil pointer, so blocking exported functions that try to
+				// return anything will not work.
+				parentHandle = llvm.ConstPointerNull(c.i8ptrType)
+			} else {
+				parentHandle = f.LastParam()
+				if parentHandle.IsNil() || parentHandle.Name() != "parentHandle" {
+					// sanity check
+					panic("trying to make exported function async: " + f.Name())
+				}
+			}
+
+			// Store return values.
+			switch inst.OperandsCount() {
+			case 0:
+				// Nothing to return.
+			case 1:
+				// Return this value by writing to the pointer stored in the
+				// parent handle. The parent coroutine has made an alloca that
+				// we can write to to store our return value.
+				returnValuePtr := c.createRuntimeCall("getTaskStatePtr", []llvm.Value{parentHandle}, "coro.parentData")
+				alloca := c.builder.CreateBitCast(returnValuePtr, llvm.PointerType(inst.Operand(0).Type(), 0), "coro.parentAlloca")
+				c.builder.CreateStore(inst.Operand(0), alloca)
+			default:
+				panic("unreachable")
+			}
+
+			// Reactivate the parent coroutine. This adds it back to the run
+			// queue, so it is started again by the scheduler when possible
+			// (possibly right after the following suspend).
+			c.createRuntimeCall("activateTask", []llvm.Value{parentHandle}, "")
+
+			// Suspend this coroutine.
+			// It would look like this is unnecessary, but if this
+			// suspend point is left out, it leads to undefined
+			// behavior somehow (with the unreachable instruction).
+			continuePoint := c.builder.CreateCall(coroSuspendFunc, []llvm.Value{
+				llvm.ConstNull(c.ctx.TokenType()),
+				llvm.ConstInt(c.ctx.Int1Type(), 0, false),
+			}, "ret")
+			sw := c.builder.CreateSwitch(continuePoint, frame.suspendBlock, 2)
+			sw.AddCase(llvm.ConstInt(c.ctx.Int8Type(), 0, false), frame.unreachableBlock)
+			sw.AddCase(llvm.ConstInt(c.ctx.Int8Type(), 1, false), frame.cleanupBlock)
+			inst.EraseFromParentAsInstruction()
 		}
 
 		// Coroutine cleanup. Free resources associated with this coroutine.
@@ -416,6 +535,14 @@ func (c *Compiler) markAsyncFunctions() (needsScheduler bool, err error) {
 		c.builder.CreateUnreachable()
 	}
 
+	// Replace calls to runtime.getCoroutineCall with the coroutine of this
+	// frame.
+	for _, getCoroutineCall := range getUses(c.mod.NamedFunction("runtime.getCoroutine")) {
+		frame := asyncFuncs[getCoroutineCall.InstructionParent().Parent()]
+		getCoroutineCall.ReplaceAllUsesWith(frame.taskHandle)
+		getCoroutineCall.EraseFromParentAsInstruction()
+	}
+
 	// Transform calls to time.Sleep() into coroutine suspend points.
 	for _, sleepCall := range getUses(sleep) {
 		// sleepCall must be a call instruction.
@@ -439,37 +566,31 @@ func (c *Compiler) markAsyncFunctions() (needsScheduler bool, err error) {
 		sleepCall.EraseFromParentAsInstruction()
 	}
 
-	// Transform calls to runtime.chanSendStub into channel send operations.
-	for _, sendOp := range getUses(chanSendStub) {
+	// Transform calls to runtime.deadlock into coroutine suspends (without
+	// resume).
+	for _, deadlockCall := range getUses(deadlock) {
+		// deadlockCall must be a call instruction.
+		frame := asyncFuncs[deadlockCall.InstructionParent().Parent()]
+
+		// Exit coroutine.
+		c.builder.SetInsertPointBefore(deadlockCall)
+		continuePoint := c.builder.CreateCall(coroSuspendFunc, []llvm.Value{
+			llvm.ConstNull(c.ctx.TokenType()),
+			llvm.ConstInt(c.ctx.Int1Type(), 0, false),
+		}, "")
+		c.splitBasicBlock(deadlockCall, llvm.NextBasicBlock(c.builder.GetInsertBlock()), "task.wakeup.dead")
+		c.builder.SetInsertPointBefore(deadlockCall)
+		sw := c.builder.CreateSwitch(continuePoint, frame.suspendBlock, 2)
+		sw.AddCase(llvm.ConstInt(c.ctx.Int8Type(), 0, false), frame.unreachableBlock)
+		sw.AddCase(llvm.ConstInt(c.ctx.Int8Type(), 1, false), frame.cleanupBlock)
+		deadlockCall.EraseFromParentAsInstruction()
+	}
+
+	// Transform calls to runtime.chanSend into channel send operations.
+	for _, sendOp := range getUses(chanSend) {
 		// sendOp must be a call instruction.
 		frame := asyncFuncs[sendOp.InstructionParent().Parent()]
 
-		// Send the value over the channel, or block.
-		sendOp.SetOperand(0, frame.taskHandle)
-		sendOp.SetOperand(sendOp.OperandsCount()-1, c.mod.NamedFunction("runtime.chanSend"))
-
-		// Use taskState.data to store the value to send:
-		//     *(*valueType)(&coroutine.promise().data) = valueToSend
-		//     runtime.chanSend(coroutine, ch)
-		bitcast := sendOp.Operand(2)
-		valueAlloca := bitcast.Operand(0)
-		c.builder.SetInsertPointBefore(valueAlloca)
-		promiseType := c.mod.GetTypeByName("runtime.taskState")
-		promiseRaw := c.builder.CreateCall(coroPromiseFunc, []llvm.Value{
-			frame.taskHandle,
-			llvm.ConstInt(c.ctx.Int32Type(), uint64(c.targetData.PrefTypeAlignment(promiseType)), false),
-			llvm.ConstInt(c.ctx.Int1Type(), 0, false),
-		}, "task.promise.raw")
-		promise := c.builder.CreateBitCast(promiseRaw, llvm.PointerType(promiseType, 0), "task.promise")
-		dataPtr := c.builder.CreateGEP(promise, []llvm.Value{
-			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
-			llvm.ConstInt(c.ctx.Int32Type(), 2, false),
-		}, "task.promise.data")
-		sendOp.SetOperand(2, llvm.Undef(c.i8ptrType))
-		valueAlloca.ReplaceAllUsesWith(c.builder.CreateBitCast(dataPtr, valueAlloca.Type(), ""))
-		bitcast.EraseFromParentAsInstruction()
-		valueAlloca.EraseFromParentAsInstruction()
-
 		// Yield to scheduler.
 		c.builder.SetInsertPointBefore(llvm.NextInstruction(sendOp))
 		continuePoint := c.builder.CreateCall(coroSuspendFunc, []llvm.Value{
@@ -482,21 +603,11 @@ func (c *Compiler) markAsyncFunctions() (needsScheduler bool, err error) {
 		sw.AddCase(llvm.ConstInt(c.ctx.Int8Type(), 1, false), frame.cleanupBlock)
 	}
 
-	// Transform calls to runtime.chanRecvStub into channel receive operations.
-	for _, recvOp := range getUses(chanRecvStub) {
+	// Transform calls to runtime.chanRecv into channel receive operations.
+	for _, recvOp := range getUses(chanRecv) {
 		// recvOp must be a call instruction.
 		frame := asyncFuncs[recvOp.InstructionParent().Parent()]
 
-		bitcast := recvOp.Operand(2)
-		commaOk := recvOp.Operand(3)
-		valueAlloca := bitcast.Operand(0)
-
-		// Receive the value over the channel, or block.
-		recvOp.SetOperand(0, frame.taskHandle)
-		recvOp.SetOperand(recvOp.OperandsCount()-1, c.mod.NamedFunction("runtime.chanRecv"))
-		recvOp.SetOperand(2, llvm.Undef(c.i8ptrType))
-		bitcast.EraseFromParentAsInstruction()
-
 		// Yield to scheduler.
 		c.builder.SetInsertPointBefore(llvm.NextInstruction(recvOp))
 		continuePoint := c.builder.CreateCall(coroSuspendFunc, []llvm.Value{
@@ -508,32 +619,6 @@ func (c *Compiler) markAsyncFunctions() (needsScheduler bool, err error) {
 		c.builder.SetInsertPointAtEnd(recvOp.InstructionParent())
 		sw.AddCase(llvm.ConstInt(c.ctx.Int8Type(), 0, false), wakeup)
 		sw.AddCase(llvm.ConstInt(c.ctx.Int8Type(), 1, false), frame.cleanupBlock)
-
-		// The value to receive is stored in taskState.data:
-		//     runtime.chanRecv(coroutine, ch)
-		//     promise := coroutine.promise()
-		//     valueReceived := *(*valueType)(&promise.data)
-		//     ok := promise.commaOk
-		c.builder.SetInsertPointBefore(wakeup.FirstInstruction())
-		promiseType := c.mod.GetTypeByName("runtime.taskState")
-		promiseRaw := c.builder.CreateCall(coroPromiseFunc, []llvm.Value{
-			frame.taskHandle,
-			llvm.ConstInt(c.ctx.Int32Type(), uint64(c.targetData.PrefTypeAlignment(promiseType)), false),
-			llvm.ConstInt(c.ctx.Int1Type(), 0, false),
-		}, "task.promise.raw")
-		promise := c.builder.CreateBitCast(promiseRaw, llvm.PointerType(promiseType, 0), "task.promise")
-		dataPtr := c.builder.CreateGEP(promise, []llvm.Value{
-			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
-			llvm.ConstInt(c.ctx.Int32Type(), 2, false),
-		}, "task.promise.data")
-		valueAlloca.ReplaceAllUsesWith(c.builder.CreateBitCast(dataPtr, valueAlloca.Type(), ""))
-		valueAlloca.EraseFromParentAsInstruction()
-		commaOkPtr := c.builder.CreateGEP(promise, []llvm.Value{
-			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
-			llvm.ConstInt(c.ctx.Int32Type(), 1, false),
-		}, "task.promise.comma-ok")
-		commaOk.ReplaceAllUsesWith(commaOkPtr)
-		recvOp.SetOperand(3, llvm.Undef(commaOk.Type()))
 	}
 
 	return true, c.lowerMakeGoroutineCalls()
@@ -558,14 +643,14 @@ func (c *Compiler) lowerMakeGoroutineCalls() error {
 
 	makeGoroutine := c.mod.NamedFunction("runtime.makeGoroutine")
 	for _, goroutine := range getUses(makeGoroutine) {
-		bitcastIn := goroutine.Operand(0)
-		origFunc := bitcastIn.Operand(0)
+		ptrtointIn := goroutine.Operand(0)
+		origFunc := ptrtointIn.Operand(0)
 		uses := getUses(goroutine)
-		if len(uses) != 1 || uses[0].IsABitCastInst().IsNil() {
-			return errors.New("expected exactly 1 bitcast use of runtime.makeGoroutine")
+		if len(uses) != 1 || uses[0].IsAIntToPtrInst().IsNil() {
+			return errors.New("expected exactly 1 inttoptr use of runtime.makeGoroutine")
 		}
-		bitcastOut := uses[0]
-		uses = getUses(bitcastOut)
+		inttoptrOut := uses[0]
+		uses = getUses(inttoptrOut)
 		if len(uses) != 1 || uses[0].IsACallInst().IsNil() {
 			return errors.New("expected exactly 1 call use of runtime.makeGoroutine bitcast")
 		}
@@ -580,7 +665,7 @@ func (c *Compiler) lowerMakeGoroutineCalls() error {
 		c.builder.SetInsertPointBefore(realCall)
 		c.builder.CreateCall(origFunc, params, "")
 		realCall.EraseFromParentAsInstruction()
-		bitcastOut.EraseFromParentAsInstruction()
+		inttoptrOut.EraseFromParentAsInstruction()
 		goroutine.EraseFromParentAsInstruction()
 	}
 

compiler/goroutine.go

@@ -0,0 +1,135 @@
+package compiler
+
+// This file implements the 'go' keyword to start a new goroutine. See
+// goroutine-lowering.go for more details.
+
+import "tinygo.org/x/go-llvm"
+
+// emitStartGoroutine starts a new goroutine with the provided function pointer
+// and parameters.
+//
+// Because a go statement doesn't return anything, return undef.
+func (c *Compiler) emitStartGoroutine(funcPtr llvm.Value, params []llvm.Value) llvm.Value {
+	switch c.selectScheduler() {
+	case "tasks":
+		paramBundle := c.emitPointerPack(params)
+		paramBundle = c.builder.CreatePtrToInt(paramBundle, c.uintptrType, "")
+
+		calleeValue := c.createGoroutineStartWrapper(funcPtr)
+		c.createRuntimeCall("startGoroutine", []llvm.Value{calleeValue, paramBundle}, "")
+	case "coroutines":
+		// We roundtrip through runtime.makeGoroutine as a signal (to find these
+		// calls) and to break any optimizations LLVM will try to do: they are
+		// invalid if we called this as a regular function to be updated later.
+		calleeValue := c.builder.CreatePtrToInt(funcPtr, c.uintptrType, "")
+		calleeValue = c.createRuntimeCall("makeGoroutine", []llvm.Value{calleeValue}, "")
+		calleeValue = c.builder.CreateIntToPtr(calleeValue, funcPtr.Type(), "")
+		c.createCall(calleeValue, params, "")
+	default:
+		panic("unreachable")
+	}
+	return llvm.Undef(funcPtr.Type().ElementType().ReturnType())
+}
+
+// createGoroutineStartWrapper creates a wrapper for the task-based
+// implementation of goroutines. For example, to call a function like this:
+//
+//     func add(x, y int) int { ... }
+//
+// It creates a wrapper like this:
+//
+//     func add$gowrapper(ptr *unsafe.Pointer) {
+//         args := (*struct{
+//             x, y int
+//         })(ptr)
+//         add(args.x, args.y)
+//     }
+//
+// This is useful because the task-based goroutine start implementation only
+// allows a single (pointer) argument to the newly started goroutine. Also, it
+// ignores the return value because newly started goroutines do not have a
+// return value.
+func (c *Compiler) createGoroutineStartWrapper(fn llvm.Value) llvm.Value {
+	var wrapper llvm.Value
+
+	if !fn.IsAFunction().IsNil() {
+		// See whether this wrapper has already been created. If so, return it.
+		name := fn.Name()
+		wrapper = c.mod.NamedFunction(name + "$gowrapper")
+		if !wrapper.IsNil() {
+			return c.builder.CreatePtrToInt(wrapper, c.uintptrType, "")
+		}
+
+		// Save the current position in the IR builder.
+		currentBlock := c.builder.GetInsertBlock()
+		defer c.builder.SetInsertPointAtEnd(currentBlock)
+
+		// Create the wrapper.
+		wrapperType := llvm.FunctionType(c.ctx.VoidType(), []llvm.Type{c.i8ptrType}, false)
+		wrapper = llvm.AddFunction(c.mod, name+"$gowrapper", wrapperType)
+		wrapper.SetLinkage(llvm.PrivateLinkage)
+		wrapper.SetUnnamedAddr(true)
+		entry := c.ctx.AddBasicBlock(wrapper, "entry")
+		c.builder.SetInsertPointAtEnd(entry)
+
+		// Create the list of params for the call.
+		paramTypes := fn.Type().ElementType().ParamTypes()
+		params := c.emitPointerUnpack(wrapper.Param(0), paramTypes[:len(paramTypes)-2])
+		params = append(params, llvm.Undef(c.i8ptrType), llvm.ConstPointerNull(c.i8ptrType))
+
+		// Create the call.
+		c.builder.CreateCall(fn, params, "")
+
+	} else {
+		// For a function pointer like this:
+		//
+		//     var funcPtr func(x, y int) int
+		//
+		// A wrapper like the following is created:
+		//
+		//     func .gowrapper(ptr *unsafe.Pointer) {
+		//         args := (*struct{
+		//             x, y int
+		//             fn   func(x, y int) int
+		//         })(ptr)
+		//         args.fn(x, y)
+		//     }
+		//
+		// With a bit of luck, identical wrapper functions like these can be
+		// merged into one.
+
+		// Save the current position in the IR builder.
+		currentBlock := c.builder.GetInsertBlock()
+		defer c.builder.SetInsertPointAtEnd(currentBlock)
+
+		// Create the wrapper.
+		wrapperType := llvm.FunctionType(c.ctx.VoidType(), []llvm.Type{c.i8ptrType}, false)
+		wrapper = llvm.AddFunction(c.mod, ".gowrapper", wrapperType)
+		wrapper.SetLinkage(llvm.InternalLinkage)
+		wrapper.SetUnnamedAddr(true)
+		entry := c.ctx.AddBasicBlock(wrapper, "entry")
+		c.builder.SetInsertPointAtEnd(entry)
+
+		// Get the list of parameters, with the extra parameters at the end.
+		paramTypes := fn.Type().ElementType().ParamTypes()
+		paramTypes[len(paramTypes)-1] = fn.Type() // the last element is the function pointer
+		params := c.emitPointerUnpack(wrapper.Param(0), paramTypes)
+
+		// Get the function pointer.
+		fnPtr := params[len(params)-1]
+
+		// Ignore the last param, which isn't used anymore.
+		// TODO: avoid this extra "parent handle" parameter in most functions.
+		params[len(params)-1] = llvm.Undef(c.i8ptrType)
+
+		// Create the call.
+		c.builder.CreateCall(fnPtr, params, "")
+	}
+
+	// Finish the function. Every basic block must end in a terminator, and
+	// because goroutines never return a value we can simply return void.
+	c.builder.CreateRetVoid()
+
+	// Return a ptrtoint of the wrapper, not the function itself.
+	return c.builder.CreatePtrToInt(wrapper, c.uintptrType, "")
+}

compiler/inlineasm.go

@@ -0,0 +1,164 @@
+package compiler
+
+// This file implements inline asm support by calling special functions.
+
+import (
+	"fmt"
+	"go/constant"
+	"regexp"
+	"strconv"
+	"strings"
+
+	"golang.org/x/tools/go/ssa"
+	"tinygo.org/x/go-llvm"
+)
+
+// This is a compiler builtin, which reads the given register by name:
+//
+//     func ReadRegister(name string) uintptr
+//
+// The register name must be a constant, for example "sp".
+func (c *Compiler) emitReadRegister(name string, args []ssa.Value) (llvm.Value, error) {
+	fnType := llvm.FunctionType(c.uintptrType, []llvm.Type{}, false)
+	regname := constant.StringVal(args[0].(*ssa.Const).Value)
+	var asm string
+	switch name {
+	case "device/arm.ReadRegister":
+		asm = "mov $0, " + regname
+	case "device/riscv.ReadRegister":
+		asm = "mv $0, " + regname
+	default:
+		panic("unknown architecture")
+	}
+	target := llvm.InlineAsm(fnType, asm, "=r", false, false, 0)
+	return c.builder.CreateCall(target, nil, ""), nil
+}
+
+// This is a compiler builtin, which emits a piece of inline assembly with no
+// operands or return values. It is useful for trivial instructions, like wfi in
+// ARM or sleep in AVR.
+//
+//     func Asm(asm string)
+//
+// The provided assembly must be a constant.
+func (c *Compiler) emitAsm(args []ssa.Value) (llvm.Value, error) {
+	// Magic function: insert inline assembly instead of calling it.
+	fnType := llvm.FunctionType(c.ctx.VoidType(), []llvm.Type{}, false)
+	asm := constant.StringVal(args[0].(*ssa.Const).Value)
+	target := llvm.InlineAsm(fnType, asm, "", true, false, 0)
+	return c.builder.CreateCall(target, nil, ""), nil
+}
+
+// This is a compiler builtin, which allows assembly to be called in a flexible
+// way.
+//
+//     func AsmFull(asm string, regs map[string]interface{})
+//
+// The asm parameter must be a constant string. The regs parameter must be
+// provided immediately. For example:
+//
+//     arm.AsmFull(
+//         "str {value}, {result}",
+//         map[string]interface{}{
+//             "value":  1
+//             "result": &dest,
+//         })
+func (c *Compiler) emitAsmFull(frame *Frame, instr *ssa.CallCommon) (llvm.Value, error) {
+	asmString := constant.StringVal(instr.Args[0].(*ssa.Const).Value)
+	registers := map[string]llvm.Value{}
+	registerMap := instr.Args[1].(*ssa.MakeMap)
+	for _, r := range *registerMap.Referrers() {
+		switch r := r.(type) {
+		case *ssa.DebugRef:
+			// ignore
+		case *ssa.MapUpdate:
+			if r.Block() != registerMap.Block() {
+				return llvm.Value{}, c.makeError(instr.Pos(), "register value map must be created in the same basic block")
+			}
+			key := constant.StringVal(r.Key.(*ssa.Const).Value)
+			//println("value:", r.Value.(*ssa.MakeInterface).X.String())
+			registers[key] = c.getValue(frame, r.Value.(*ssa.MakeInterface).X)
+		case *ssa.Call:
+			if r.Common() == instr {
+				break
+			}
+		default:
+			return llvm.Value{}, c.makeError(instr.Pos(), "don't know how to handle argument to inline assembly: "+r.String())
+		}
+	}
+	// TODO: handle dollar signs in asm string
+	registerNumbers := map[string]int{}
+	var err error
+	argTypes := []llvm.Type{}
+	args := []llvm.Value{}
+	constraints := []string{}
+	asmString = regexp.MustCompile("\\{[a-zA-Z]+\\}").ReplaceAllStringFunc(asmString, func(s string) string {
+		// TODO: skip strings like {r4} etc. that look like ARM push/pop
+		// instructions.
+		name := s[1 : len(s)-1]
+		if _, ok := registers[name]; !ok {
+			if err == nil {
+				err = c.makeError(instr.Pos(), "unknown register name: "+name)
+			}
+			return s
+		}
+		if _, ok := registerNumbers[name]; !ok {
+			registerNumbers[name] = len(registerNumbers)
+			argTypes = append(argTypes, registers[name].Type())
+			args = append(args, registers[name])
+			switch registers[name].Type().TypeKind() {
+			case llvm.IntegerTypeKind:
+				constraints = append(constraints, "r")
+			case llvm.PointerTypeKind:
+				constraints = append(constraints, "*m")
+			default:
+				err = c.makeError(instr.Pos(), "unknown type in inline assembly for value: "+name)
+				return s
+			}
+		}
+		return fmt.Sprintf("${%v}", registerNumbers[name])
+	})
+	if err != nil {
+		return llvm.Value{}, err
+	}
+	fnType := llvm.FunctionType(c.ctx.VoidType(), argTypes, false)
+	target := llvm.InlineAsm(fnType, asmString, strings.Join(constraints, ","), true, false, 0)
+	return c.builder.CreateCall(target, args, ""), 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.
+func (c *Compiler) emitSVCall(frame *Frame, 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 := "={r0}"
+	for i, arg := range args[1:] {
+		arg = arg.(*ssa.MakeInterface).X
+		if i == 0 {
+			constraints += ",0"
+		} else {
+			constraints += ",{r" + strconv.Itoa(i) + "}"
+		}
+		llvmValue := c.getValue(frame, arg)
+		llvmArgs = append(llvmArgs, llvmValue)
+		argTypes = append(argTypes, llvmValue.Type())
+	}
+	// Implement the ARM calling convention by marking r1-r3 as
+	// clobbered. r0 is used as an output register so doesn't have to be
+	// marked as clobbered.
+	constraints += ",~{r1},~{r2},~{r3}"
+	fnType := llvm.FunctionType(c.uintptrType, argTypes, false)
+	target := llvm.InlineAsm(fnType, asm, constraints, true, false, 0)
+	return c.builder.CreateCall(target, llvmArgs, ""), nil
+}

compiler/interface-lowering.go

@@ -4,7 +4,6 @@ package compiler
 // form, optimizing them in the process.
 //
 // During SSA construction, the following pseudo-calls are created:
-//     runtime.makeInterface(typecode, methodSet)
 //     runtime.typeAssert(typecode, assertedType)
 //     runtime.interfaceImplements(typecode, interfaceMethodSet)
 //     runtime.interfaceMethod(typecode, interfaceMethodSet, signature)
@@ -14,16 +13,13 @@ package compiler
 //
 // This pass lowers the above functions to their final form:
 //
-// makeInterface:
-//     Replaced with a constant typecode.
-//
 // typeAssert:
 //     Replaced with an icmp instruction so it can be directly used in a type
 //     switch. This is very easy to optimize for LLVM: it will often translate a
 //     type switch into a regular switch statement.
 //     When this type assert is not possible (the type is never used in an
-//     interface with makeInterface), this call is replaced with a constant
-//     false to optimize the type assert away completely.
+//     interface), this call is replaced with a constant false to optimize the
+//     type assert away completely.
 //
 // interfaceImplements:
 //     This call is translated into a call that checks whether the underlying
@@ -166,25 +162,36 @@ func (c *Compiler) LowerInterfaces() {
 
 // run runs the pass itself.
 func (p *lowerInterfacesPass) run() {
-	// Count per type how often it is put in an interface. Also, collect all
-	// methods this type has (if it is named).
-	makeInterface := p.mod.NamedFunction("runtime.makeInterface")
-	makeInterfaceUses := getUses(makeInterface)
-	for _, use := range makeInterfaceUses {
-		typecode := use.Operand(0)
-		name := typecode.Name()
-		if t, ok := p.types[name]; !ok {
-			// This is the first time this type has been seen, add it to the
-			// list of types.
-			t = p.addType(typecode)
-			p.addTypeMethods(t, use.Operand(1))
-		} else {
-			p.addTypeMethods(t, use.Operand(1))
-		}
+	// Collect all type codes.
+	typecodeIDPtr := llvm.PointerType(p.getLLVMRuntimeType("typecodeID"), 0)
+	typeInInterfacePtr := llvm.PointerType(p.getLLVMRuntimeType("typeInInterface"), 0)
+	var typesInInterfaces []llvm.Value
+	for global := p.mod.FirstGlobal(); !global.IsNil(); global = llvm.NextGlobal(global) {
+		switch global.Type() {
+		case typecodeIDPtr:
+			// Retrieve Go type information based on an opaque global variable.
+			// Only the name of the global is relevant, the object itself is
+			// discarded afterwards.
+			name := global.Name()
+			t := &typeInfo{
+				name:     name,
+				typecode: global,
+			}
+			p.types[name] = t
+		case typeInInterfacePtr:
+			// Count per type how often it is put in an interface. Also, collect
+			// all methods this type has (if it is named).
+			typesInInterfaces = append(typesInInterfaces, global)
+			initializer := global.Initializer()
+			typecode := llvm.ConstExtractValue(initializer, []uint32{0})
+			methodSet := llvm.ConstExtractValue(initializer, []uint32{1})
+			t := p.types[typecode.Name()]
+			p.addTypeMethods(t, methodSet)
 
-		// Count the number of MakeInterface instructions, for sorting the
-		// typecodes later.
-		p.types[name].countMakeInterfaces++
+			// Count the number of MakeInterface instructions, for sorting the
+			// typecodes later.
+			t.countMakeInterfaces += len(getUses(global))
+		}
 	}
 
 	// Count per type how often it is type asserted on (e.g. in a switch
@@ -194,9 +201,6 @@ func (p *lowerInterfacesPass) run() {
 	for _, use := range typeAssertUses {
 		typecode := use.Operand(1)
 		name := typecode.Name()
-		if _, ok := p.types[name]; !ok {
-			p.addType(typecode)
-		}
 		p.types[name].countTypeAsserts++
 	}
 
@@ -286,16 +290,6 @@ func (p *lowerInterfacesPass) run() {
 		typecode := use.Operand(0)
 		signature := p.signatures[use.Operand(2).Name()]
 
-		// If the interface was created in the same function, we can insert a
-		// direct call. This may not happen often but it is an easy
-		// optimization so let's do it anyway.
-		if !typecode.IsACallInst().IsNil() && typecode.CalledValue() == makeInterface {
-			name := typecode.Operand(0).Name()
-			typ := p.types[name]
-			p.replaceInvokeWithCall(use, typ, signature)
-			continue
-		}
-
 		methodSet := use.Operand(1).Operand(0) // global variable
 		itf := p.interfaces[methodSet.Name()]
 		if len(itf.types) == 0 {
@@ -356,20 +350,6 @@ func (p *lowerInterfacesPass) run() {
 	// types, if possible.
 	for _, use := range interfaceImplementsUses {
 		actualType := use.Operand(0)
-		if !actualType.IsACallInst().IsNil() && actualType.CalledValue() == makeInterface {
-			// Type assert is in the same function that creates the interface
-			// value. This means the underlying type is already known so match
-			// on that.
-			// This may not happen often but it is an easy optimization.
-			name := actualType.Operand(0).Name()
-			typ := p.types[name]
-			p.builder.SetInsertPointBefore(use)
-			assertedType := p.builder.CreatePtrToInt(typ.typecode, p.uintptrType, "typeassert.typecode")
-			commaOk := p.builder.CreateICmp(llvm.IntEQ, assertedType, actualType, "typeassert.ok")
-			use.ReplaceAllUsesWith(commaOk)
-			use.EraseFromParentAsInstruction()
-			continue
-		}
 
 		methodSet := use.Operand(1).Operand(0) // global variable
 		itf := p.interfaces[methodSet.Name()]
@@ -416,12 +396,14 @@ func (p *lowerInterfacesPass) run() {
 	// Assign a type code for each type.
 	p.assignTypeCodes(typeSlice)
 
-	// Replace each call to runtime.makeInterface with the constant type code.
-	for _, use := range makeInterfaceUses {
-		global := use.Operand(0)
-		t := p.types[global.Name()]
-		use.ReplaceAllUsesWith(llvm.ConstPtrToInt(t.typecode, p.uintptrType))
-		use.EraseFromParentAsInstruction()
+	// Replace each use of a runtime.typeInInterface with the constant type
+	// code.
+	for _, global := range typesInInterfaces {
+		for _, use := range getUses(global) {
+			t := p.types[llvm.ConstExtractValue(global.Initializer(), []uint32{0}).Name()]
+			typecode := llvm.ConstInt(p.uintptrType, t.num, false)
+			use.ReplaceAllUsesWith(typecode)
+		}
 	}
 
 	// Replace each type assert with an actual type comparison or (if the type
@@ -458,12 +440,18 @@ func (p *lowerInterfacesPass) run() {
 	// numbers.
 	for _, typ := range p.types {
 		for _, use := range getUses(typ.typecode) {
-			if use.IsConstant() && use.Opcode() == llvm.PtrToInt {
+			if !use.IsAConstantExpr().IsNil() && use.Opcode() == llvm.PtrToInt {
 				use.ReplaceAllUsesWith(llvm.ConstInt(p.uintptrType, typ.num, false))
 			}
 		}
 	}
 
+	// Remove stray runtime.typeInInterface globals. Required for the following
+	// cleanup.
+	for _, global := range typesInInterfaces {
+		global.EraseFromParentAsGlobal()
+	}
+
 	// Remove method sets of types. Unnecessary, but cleans up the IR for
 	// inspection.
 	for _, typ := range p.types {
@@ -474,19 +462,6 @@ func (p *lowerInterfacesPass) run() {
 	}
 }
 
-// addType retrieves Go type information based on a i16 global variable.
-// Only the name of the i16 is relevant, the object itself is const-propagated
-// and discared afterwards.
-func (p *lowerInterfacesPass) addType(typecode llvm.Value) *typeInfo {
-	name := typecode.Name()
-	t := &typeInfo{
-		name:     name,
-		typecode: typecode,
-	}
-	p.types[name] = t
-	return t
-}
-
 // addTypeMethods reads the method set of the given type info struct. It
 // retrieves the signatures and the references to the method functions
 // themselves for later type<->interface matching.
@@ -552,11 +527,40 @@ func (p *lowerInterfacesPass) replaceInvokeWithCall(use llvm.Value, typ *typeInf
 	}
 	inttoptr := inttoptrs[0]
 	function := typ.getMethod(signature).function
-	if inttoptr.Type() != function.Type() {
-		p.builder.SetInsertPointBefore(use)
-		function = p.builder.CreateBitCast(function, inttoptr.Type(), "")
+	if inttoptr.Type() == function.Type() {
+		// Easy case: the types are the same. Simply replace the inttoptr
+		// result (which is directly called) with the actual function.
+		inttoptr.ReplaceAllUsesWith(function)
+	} else {
+		// Harder case: the type is not actually the same. Go through each call
+		// (of which there should be only one), extract the receiver params for
+		// this call and replace the call with a direct call to the target
+		// function.
+		for _, call := range getUses(inttoptr) {
+			if call.IsACallInst().IsNil() || call.CalledValue() != inttoptr {
+				panic("expected the inttoptr to be called as a method, this is not a method call")
+			}
+			operands := make([]llvm.Value, call.OperandsCount()-1)
+			for i := range operands {
+				operands[i] = call.Operand(i)
+			}
+			paramTypes := function.Type().ElementType().ParamTypes()
+			receiverParamTypes := paramTypes[:len(paramTypes)-(len(operands)-1)]
+			methodParamTypes := paramTypes[len(paramTypes)-(len(operands)-1):]
+			for i, methodParamType := range methodParamTypes {
+				if methodParamType != operands[i+1].Type() {
+					panic("expected method call param type and function param type to be the same")
+				}
+			}
+			p.builder.SetInsertPointBefore(call)
+			receiverParams := p.emitPointerUnpack(operands[0], receiverParamTypes)
+			result := p.builder.CreateCall(function, append(receiverParams, operands[1:]...), "")
+			if result.Type().TypeKind() != llvm.VoidTypeKind {
+				call.ReplaceAllUsesWith(result)
+			}
+			call.EraseFromParentAsInstruction()
+		}
 	}
-	inttoptr.ReplaceAllUsesWith(function)
 	inttoptr.EraseFromParentAsInstruction()
 	use.EraseFromParentAsInstruction()
 }
@@ -599,9 +603,9 @@ func (p *lowerInterfacesPass) createInterfaceImplementsFunc(itf *interfaceInfo)
 	// TODO: debug info
 
 	// Create all used basic blocks.
-	entry := llvm.AddBasicBlock(fn, "entry")
-	thenBlock := llvm.AddBasicBlock(fn, "then")
-	elseBlock := llvm.AddBasicBlock(fn, "else")
+	entry := p.ctx.AddBasicBlock(fn, "entry")
+	thenBlock := p.ctx.AddBasicBlock(fn, "then")
+	elseBlock := p.ctx.AddBasicBlock(fn, "else")
 
 	// Add all possible types as cases.
 	p.builder.SetInsertPointAtEnd(entry)
@@ -657,11 +661,11 @@ func (p *lowerInterfacesPass) createInterfaceMethodFunc(itf *interfaceInfo, sign
 	// TODO: debug info
 
 	// Create entry block.
-	entry := llvm.AddBasicBlock(fn, "entry")
+	entry := p.ctx.AddBasicBlock(fn, "entry")
 
 	// Create default block and make it unreachable (which it is, because all
 	// possible types are checked).
-	defaultBlock := llvm.AddBasicBlock(fn, "default")
+	defaultBlock := p.ctx.AddBasicBlock(fn, "default")
 	p.builder.SetInsertPointAtEnd(defaultBlock)
 	p.builder.CreateUnreachable()
 
@@ -680,7 +684,7 @@ func (p *lowerInterfacesPass) createInterfaceMethodFunc(itf *interfaceInfo, sign
 
 	// Define all possible functions that can be called.
 	for _, typ := range itf.types {
-		bb := llvm.AddBasicBlock(fn, typ.name)
+		bb := p.ctx.AddBasicBlock(fn, typ.name)
 		sw.AddCase(llvm.ConstInt(p.uintptrType, typ.num, false), bb)
 
 		// The function we will redirect to when the interface has this type.

compiler/interface.go

@@ -22,75 +22,123 @@ import (
 // value field.
 //
 // An interface value is a {typecode, value} tuple, or {i16, i8*} to be exact.
-func (c *Compiler) parseMakeInterface(val llvm.Value, typ types.Type, pos token.Pos) (llvm.Value, error) {
-	var itfValue llvm.Value
-	size := c.targetData.TypeAllocSize(val.Type())
-	if size > c.targetData.TypeAllocSize(c.i8ptrType) {
-		// Allocate on the heap and put a pointer in the interface.
-		// TODO: escape analysis.
-		sizeValue := llvm.ConstInt(c.uintptrType, size, false)
-		alloc := c.createRuntimeCall("alloc", []llvm.Value{sizeValue}, "makeinterface.alloc")
-		itfValueCast := c.builder.CreateBitCast(alloc, llvm.PointerType(val.Type(), 0), "makeinterface.cast.value")
-		c.builder.CreateStore(val, itfValueCast)
-		itfValue = c.builder.CreateBitCast(itfValueCast, c.i8ptrType, "makeinterface.cast.i8ptr")
-	} else if size == 0 {
-		itfValue = llvm.ConstPointerNull(c.i8ptrType)
-	} else {
-		// Directly place the value in the interface.
-		switch val.Type().TypeKind() {
-		case llvm.IntegerTypeKind:
-			itfValue = c.builder.CreateIntToPtr(val, c.i8ptrType, "makeinterface.cast.int")
-		case llvm.PointerTypeKind:
-			itfValue = c.builder.CreateBitCast(val, c.i8ptrType, "makeinterface.cast.ptr")
-		case llvm.StructTypeKind, llvm.FloatTypeKind, llvm.DoubleTypeKind:
-			// A bitcast would be useful here, but bitcast doesn't allow
-			// aggregate types. So we'll bitcast it using an alloca.
-			// Hopefully this will get optimized away.
-			mem := c.builder.CreateAlloca(c.i8ptrType, "makeinterface.cast.struct")
-			memStructPtr := c.builder.CreateBitCast(mem, llvm.PointerType(val.Type(), 0), "makeinterface.cast.struct.cast")
-			c.builder.CreateStore(val, memStructPtr)
-			itfValue = c.builder.CreateLoad(mem, "makeinterface.cast.load")
-		default:
-			return llvm.Value{}, c.makeError(pos, "todo: makeinterface: cast small type to i8*")
-		}
-	}
+func (c *Compiler) parseMakeInterface(val llvm.Value, typ types.Type, pos token.Pos) llvm.Value {
+	itfValue := c.emitPointerPack([]llvm.Value{val})
 	itfTypeCodeGlobal := c.getTypeCode(typ)
-	itfMethodSetGlobal, err := c.getTypeMethodSet(typ)
-	if err != nil {
-		return llvm.Value{}, nil
+	itfMethodSetGlobal := c.getTypeMethodSet(typ)
+	itfConcreteTypeGlobal := c.mod.NamedGlobal("typeInInterface:" + itfTypeCodeGlobal.Name())
+	if itfConcreteTypeGlobal.IsNil() {
+		typeInInterface := c.getLLVMRuntimeType("typeInInterface")
+		itfConcreteTypeGlobal = llvm.AddGlobal(c.mod, typeInInterface, "typeInInterface:"+itfTypeCodeGlobal.Name())
+		itfConcreteTypeGlobal.SetInitializer(llvm.ConstNamedStruct(typeInInterface, []llvm.Value{itfTypeCodeGlobal, itfMethodSetGlobal}))
+		itfConcreteTypeGlobal.SetGlobalConstant(true)
+		itfConcreteTypeGlobal.SetLinkage(llvm.PrivateLinkage)
 	}
-	itfTypeCode := c.createRuntimeCall("makeInterface", []llvm.Value{itfTypeCodeGlobal, itfMethodSetGlobal}, "makeinterface.typecode")
-	itf := llvm.Undef(c.mod.GetTypeByName("runtime._interface"))
+	itfTypeCode := c.builder.CreatePtrToInt(itfConcreteTypeGlobal, c.uintptrType, "")
+	itf := llvm.Undef(c.getLLVMRuntimeType("_interface"))
 	itf = c.builder.CreateInsertValue(itf, itfTypeCode, 0, "")
 	itf = c.builder.CreateInsertValue(itf, itfValue, 1, "")
-	return itf, nil
+	return itf
 }
 
 // getTypeCode returns a reference to a type code.
 // It returns a pointer to an external global which should be replaced with the
 // real type in the interface lowering pass.
 func (c *Compiler) getTypeCode(typ types.Type) llvm.Value {
-	globalName := "type:" + getTypeCodeName(typ)
+	globalName := "reflect/types.type:" + getTypeCodeName(typ)
 	global := c.mod.NamedGlobal(globalName)
 	if global.IsNil() {
-		global = llvm.AddGlobal(c.mod, c.ctx.Int8Type(), globalName)
+		// Create a new typecode global.
+		global = llvm.AddGlobal(c.mod, c.getLLVMRuntimeType("typecodeID"), globalName)
+		// Some type classes contain more information for underlying types or
+		// element types. Store it directly in the typecode global to make
+		// reflect lowering simpler.
+		var references llvm.Value
+		var length int64
+		switch typ := typ.(type) {
+		case *types.Named:
+			references = c.getTypeCode(typ.Underlying())
+		case *types.Chan:
+			references = c.getTypeCode(typ.Elem())
+		case *types.Pointer:
+			references = c.getTypeCode(typ.Elem())
+		case *types.Slice:
+			references = c.getTypeCode(typ.Elem())
+		case *types.Array:
+			references = c.getTypeCode(typ.Elem())
+			length = typ.Len()
+		case *types.Struct:
+			// Take a pointer to the typecodeID of the first field (if it exists).
+			structGlobal := c.makeStructTypeFields(typ)
+			references = llvm.ConstBitCast(structGlobal, global.Type())
+		}
+		if !references.IsNil() {
+			// Set the 'references' field of the runtime.typecodeID struct.
+			globalValue := llvm.ConstNull(global.Type().ElementType())
+			globalValue = llvm.ConstInsertValue(globalValue, references, []uint32{0})
+			if length != 0 {
+				lengthValue := llvm.ConstInt(c.uintptrType, uint64(length), false)
+				globalValue = llvm.ConstInsertValue(globalValue, lengthValue, []uint32{1})
+			}
+			global.SetInitializer(globalValue)
+			global.SetLinkage(llvm.PrivateLinkage)
+		}
 		global.SetGlobalConstant(true)
 	}
 	return global
 }
 
+// makeStructTypeFields creates a new global that stores all type information
+// related to this struct type, and returns the resulting global. This global is
+// actually an array of all the fields in the structs.
+func (c *Compiler) makeStructTypeFields(typ *types.Struct) llvm.Value {
+	// The global is an array of runtime.structField structs.
+	runtimeStructField := c.getLLVMRuntimeType("structField")
+	structGlobalType := llvm.ArrayType(runtimeStructField, typ.NumFields())
+	structGlobal := llvm.AddGlobal(c.mod, structGlobalType, "reflect/types.structFields")
+	structGlobalValue := llvm.ConstNull(structGlobalType)
+	for i := 0; i < typ.NumFields(); i++ {
+		fieldGlobalValue := llvm.ConstNull(runtimeStructField)
+		fieldGlobalValue = llvm.ConstInsertValue(fieldGlobalValue, c.getTypeCode(typ.Field(i).Type()), []uint32{0})
+		fieldName := c.makeGlobalArray([]byte(typ.Field(i).Name()), "reflect/types.structFieldName", c.ctx.Int8Type())
+		fieldName.SetLinkage(llvm.PrivateLinkage)
+		fieldName.SetUnnamedAddr(true)
+		fieldName = llvm.ConstGEP(fieldName, []llvm.Value{
+			llvm.ConstInt(llvm.Int32Type(), 0, false),
+			llvm.ConstInt(llvm.Int32Type(), 0, false),
+		})
+		fieldGlobalValue = llvm.ConstInsertValue(fieldGlobalValue, fieldName, []uint32{1})
+		if typ.Tag(i) != "" {
+			fieldTag := c.makeGlobalArray([]byte(typ.Tag(i)), "reflect/types.structFieldTag", c.ctx.Int8Type())
+			fieldTag.SetLinkage(llvm.PrivateLinkage)
+			fieldTag.SetUnnamedAddr(true)
+			fieldTag = llvm.ConstGEP(fieldTag, []llvm.Value{
+				llvm.ConstInt(llvm.Int32Type(), 0, false),
+				llvm.ConstInt(llvm.Int32Type(), 0, false),
+			})
+			fieldGlobalValue = llvm.ConstInsertValue(fieldGlobalValue, fieldTag, []uint32{2})
+		}
+		if typ.Field(i).Embedded() {
+			fieldEmbedded := llvm.ConstInt(c.ctx.Int1Type(), 1, false)
+			fieldGlobalValue = llvm.ConstInsertValue(fieldGlobalValue, fieldEmbedded, []uint32{3})
+		}
+		structGlobalValue = llvm.ConstInsertValue(structGlobalValue, fieldGlobalValue, []uint32{uint32(i)})
+	}
+	structGlobal.SetInitializer(structGlobalValue)
+	structGlobal.SetUnnamedAddr(true)
+	structGlobal.SetLinkage(llvm.PrivateLinkage)
+	return structGlobal
+}
+
 // getTypeCodeName returns a name for this type that can be used in the
 // interface lowering pass to assign type codes as expected by the reflect
 // package. See getTypeCodeNum.
 func getTypeCodeName(t types.Type) string {
-	name := ""
-	if named, ok := t.(*types.Named); ok {
-		name = "~" + named.String() + ":"
-		t = t.Underlying()
-	}
 	switch t := t.(type) {
+	case *types.Named:
+		return "named:" + t.String()
 	case *types.Array:
-		return "array:" + name + strconv.FormatInt(t.Len(), 10) + ":" + getTypeCodeName(t.Elem())
+		return "array:" + strconv.FormatInt(t.Len(), 10) + ":" + getTypeCodeName(t.Elem())
 	case *types.Basic:
 		var kind string
 		switch t.Kind() {
@@ -133,21 +181,21 @@ func getTypeCodeName(t types.Type) string {
 		default:
 			panic("unknown basic type: " + t.Name())
 		}
-		return "basic:" + name + kind
+		return "basic:" + kind
 	case *types.Chan:
-		return "chan:" + name + getTypeCodeName(t.Elem())
+		return "chan:" + getTypeCodeName(t.Elem())
 	case *types.Interface:
 		methods := make([]string, t.NumMethods())
 		for i := 0; i < t.NumMethods(); i++ {
 			methods[i] = getTypeCodeName(t.Method(i).Type())
 		}
-		return "interface:" + name + "{" + strings.Join(methods, ",") + "}"
+		return "interface:" + "{" + strings.Join(methods, ",") + "}"
 	case *types.Map:
 		keyType := getTypeCodeName(t.Key())
 		elemType := getTypeCodeName(t.Elem())
-		return "map:" + name + "{" + keyType + "," + elemType + "}"
+		return "map:" + "{" + keyType + "," + elemType + "}"
 	case *types.Pointer:
-		return "pointer:" + name + getTypeCodeName(t.Elem())
+		return "pointer:" + getTypeCodeName(t.Elem())
 	case *types.Signature:
 		params := make([]string, t.Params().Len())
 		for i := 0; i < t.Params().Len(); i++ {
@@ -157,15 +205,26 @@ func getTypeCodeName(t types.Type) string {
 		for i := 0; i < t.Results().Len(); i++ {
 			results[i] = getTypeCodeName(t.Results().At(i).Type())
 		}
-		return "func:" + name + "{" + strings.Join(params, ",") + "}{" + strings.Join(results, ",") + "}"
+		return "func:" + "{" + strings.Join(params, ",") + "}{" + strings.Join(results, ",") + "}"
 	case *types.Slice:
-		return "slice:" + name + getTypeCodeName(t.Elem())
+		return "slice:" + getTypeCodeName(t.Elem())
 	case *types.Struct:
 		elems := make([]string, t.NumFields())
-		for i := 0; i < t.NumFields(); i++ {
-			elems[i] = getTypeCodeName(t.Field(i).Type())
+		if t.NumFields() > 2 && t.Field(0).Name() == "C union" {
+			// TODO: report this as a normal error instead of panicking.
+			panic("cgo unions are not allowed in interfaces")
 		}
-		return "struct:" + name + "{" + strings.Join(elems, ",") + "}"
+		for i := 0; i < t.NumFields(); i++ {
+			embedded := ""
+			if t.Field(i).Embedded() {
+				embedded = "#"
+			}
+			elems[i] = embedded + t.Field(i).Name() + ":" + getTypeCodeName(t.Field(i).Type())
+			if t.Tag(i) != "" {
+				elems[i] += "`" + t.Tag(i) + "`"
+			}
+		}
+		return "struct:" + "{" + strings.Join(elems, ",") + "}"
 	default:
 		panic("unknown type: " + t.String())
 	}
@@ -173,22 +232,22 @@ func getTypeCodeName(t types.Type) string {
 
 // getTypeMethodSet returns a reference (GEP) to a global method set. This
 // method set should be unreferenced after the interface lowering pass.
-func (c *Compiler) getTypeMethodSet(typ types.Type) (llvm.Value, error) {
+func (c *Compiler) getTypeMethodSet(typ types.Type) llvm.Value {
 	global := c.mod.NamedGlobal(typ.String() + "$methodset")
 	zero := llvm.ConstInt(c.ctx.Int32Type(), 0, false)
 	if !global.IsNil() {
 		// the method set already exists
-		return llvm.ConstGEP(global, []llvm.Value{zero, zero}), nil
+		return llvm.ConstGEP(global, []llvm.Value{zero, zero})
 	}
 
 	ms := c.ir.Program.MethodSets.MethodSet(typ)
 	if ms.Len() == 0 {
 		// no methods, so can leave that one out
-		return llvm.ConstPointerNull(llvm.PointerType(c.mod.GetTypeByName("runtime.interfaceMethodInfo"), 0)), nil
+		return llvm.ConstPointerNull(llvm.PointerType(c.getLLVMRuntimeType("interfaceMethodInfo"), 0))
 	}
 
 	methods := make([]llvm.Value, ms.Len())
-	interfaceMethodInfoType := c.mod.GetTypeByName("runtime.interfaceMethodInfo")
+	interfaceMethodInfoType := c.getLLVMRuntimeType("interfaceMethodInfo")
 	for i := 0; i < ms.Len(); i++ {
 		method := ms.At(i)
 		signatureGlobal := c.getMethodSignature(method.Obj().(*types.Func))
@@ -197,10 +256,7 @@ func (c *Compiler) getTypeMethodSet(typ types.Type) (llvm.Value, error) {
 			// compiler error, so panic
 			panic("cannot find function: " + f.LinkName())
 		}
-		fn, err := c.getInterfaceInvokeWrapper(f)
-		if err != nil {
-			return llvm.Value{}, err
-		}
+		fn := c.getInterfaceInvokeWrapper(f)
 		methodInfo := llvm.ConstNamedStruct(interfaceMethodInfoType, []llvm.Value{
 			signatureGlobal,
 			llvm.ConstPtrToInt(fn, c.uintptrType),
@@ -213,7 +269,7 @@ func (c *Compiler) getTypeMethodSet(typ types.Type) (llvm.Value, error) {
 	global.SetInitializer(value)
 	global.SetGlobalConstant(true)
 	global.SetLinkage(llvm.PrivateLinkage)
-	return llvm.ConstGEP(global, []llvm.Value{zero, zero}), nil
+	return llvm.ConstGEP(global, []llvm.Value{zero, zero})
 }
 
 // getInterfaceMethodSet returns a global variable with the method set of the
@@ -262,19 +318,9 @@ func (c *Compiler) getMethodSignature(method *types.Func) llvm.Value {
 //
 // Type asserts on concrete types are trivial: just compare type numbers. Type
 // asserts on interfaces are more difficult, see the comments in the function.
-func (c *Compiler) parseTypeAssert(frame *Frame, expr *ssa.TypeAssert) (llvm.Value, error) {
-	itf, err := c.parseExpr(frame, expr.X)
-	if err != nil {
-		return llvm.Value{}, err
-	}
-	assertedType, err := c.getLLVMType(expr.AssertedType)
-	if err != nil {
-		return llvm.Value{}, err
-	}
-	valueNil, err := c.getZeroValue(assertedType)
-	if err != nil {
-		return llvm.Value{}, err
-	}
+func (c *Compiler) parseTypeAssert(frame *Frame, expr *ssa.TypeAssert) llvm.Value {
+	itf := c.getValue(frame, expr.X)
+	assertedType := c.getLLVMType(expr.AssertedType)
 
 	actualTypeNum := c.builder.CreateExtractValue(itf, 0, "interface.type")
 	commaOk := llvm.Value{}
@@ -327,71 +373,35 @@ func (c *Compiler) parseTypeAssert(frame *Frame, expr *ssa.TypeAssert) (llvm.Val
 		// Type assert on concrete type. Extract the underlying type from
 		// the interface (but only after checking it matches).
 		valuePtr := c.builder.CreateExtractValue(itf, 1, "typeassert.value.ptr")
-		size := c.targetData.TypeAllocSize(assertedType)
-		if size > c.targetData.TypeAllocSize(c.i8ptrType) {
-			// Value was stored in an allocated buffer, load it from there.
-			valuePtrCast := c.builder.CreateBitCast(valuePtr, llvm.PointerType(assertedType, 0), "")
-			valueOk = c.builder.CreateLoad(valuePtrCast, "typeassert.value.ok")
-		} else if size == 0 {
-			valueOk, err = c.getZeroValue(assertedType)
-			if err != nil {
-				return llvm.Value{}, err
-			}
-		} else {
-			// Value was stored directly in the interface.
-			switch assertedType.TypeKind() {
-			case llvm.IntegerTypeKind:
-				valueOk = c.builder.CreatePtrToInt(valuePtr, assertedType, "typeassert.value.ok")
-			case llvm.PointerTypeKind:
-				valueOk = c.builder.CreateBitCast(valuePtr, assertedType, "typeassert.value.ok")
-			default: // struct, float, etc.
-				// A bitcast would be useful here, but bitcast doesn't allow
-				// aggregate types. So we'll bitcast it using an alloca.
-				// Hopefully this will get optimized away.
-				mem := c.builder.CreateAlloca(c.i8ptrType, "")
-				c.builder.CreateStore(valuePtr, mem)
-				memCast := c.builder.CreateBitCast(mem, llvm.PointerType(assertedType, 0), "")
-				valueOk = c.builder.CreateLoad(memCast, "typeassert.value.ok")
-			}
-		}
+		valueOk = c.emitPointerUnpack(valuePtr, []llvm.Type{assertedType})[0]
 	}
 	c.builder.CreateBr(nextBlock)
 
 	// Continue after the if statement.
 	c.builder.SetInsertPointAtEnd(nextBlock)
 	phi := c.builder.CreatePHI(assertedType, "typeassert.value")
-	phi.AddIncoming([]llvm.Value{valueNil, valueOk}, []llvm.BasicBlock{prevBlock, okBlock})
+	phi.AddIncoming([]llvm.Value{llvm.ConstNull(assertedType), valueOk}, []llvm.BasicBlock{prevBlock, okBlock})
 
 	if expr.CommaOk {
 		tuple := c.ctx.ConstStruct([]llvm.Value{llvm.Undef(assertedType), llvm.Undef(c.ctx.Int1Type())}, false) // create empty tuple
 		tuple = c.builder.CreateInsertValue(tuple, phi, 0, "")                                                  // insert value
 		tuple = c.builder.CreateInsertValue(tuple, commaOk, 1, "")                                              // insert 'comma ok' boolean
-		return tuple, nil
+		return tuple
 	} else {
 		// This is kind of dirty as the branch above becomes mostly useless,
 		// but hopefully this gets optimized away.
 		c.createRuntimeCall("interfaceTypeAssert", []llvm.Value{commaOk}, "")
-		return phi, nil
+		return phi
 	}
 }
 
 // getInvokeCall creates and returns the function pointer and parameters of an
 // interface call. It can be used in a call or defer instruction.
-func (c *Compiler) getInvokeCall(frame *Frame, instr *ssa.CallCommon) (llvm.Value, []llvm.Value, error) {
+func (c *Compiler) getInvokeCall(frame *Frame, instr *ssa.CallCommon) (llvm.Value, []llvm.Value) {
 	// Call an interface method with dynamic dispatch.
-	itf, err := c.parseExpr(frame, instr.Value) // interface
-	if err != nil {
-		return llvm.Value{}, nil, err
-	}
+	itf := c.getValue(frame, instr.Value) // interface
 
-	llvmFnType, err := c.getLLVMType(instr.Method.Type())
-	if err != nil {
-		return llvm.Value{}, nil, err
-	}
-	// getLLVMType() has created a closure type for us, but we don't actually
-	// want a closure type as an interface call can never be a closure call. So
-	// extract the function pointer type from the closure.
-	llvmFnType = llvmFnType.Subtypes()[1]
+	llvmFnType := c.getRawFuncType(instr.Method.Type().(*types.Signature))
 
 	typecode := c.builder.CreateExtractValue(itf, 0, "invoke.typecode")
 	values := []llvm.Value{
@@ -405,11 +415,7 @@ func (c *Compiler) getInvokeCall(frame *Frame, instr *ssa.CallCommon) (llvm.Valu
 
 	args := []llvm.Value{receiverValue}
 	for _, arg := range instr.Args {
-		val, err := c.parseExpr(frame, arg)
-		if err != nil {
-			return llvm.Value{}, nil, err
-		}
-		args = append(args, val)
+		args = append(args, c.getValue(frame, arg))
 	}
 	// Add the context parameter. An interface call never takes a context but we
 	// have to supply the parameter anyway.
@@ -417,7 +423,7 @@ func (c *Compiler) getInvokeCall(frame *Frame, instr *ssa.CallCommon) (llvm.Valu
 	// Add the parent goroutine handle.
 	args = append(args, llvm.Undef(c.i8ptrType))
 
-	return fnCast, args, nil
+	return fnCast, args
 }
 
 // interfaceInvokeWrapper keeps some state between getInterfaceInvokeWrapper and
@@ -433,19 +439,16 @@ type interfaceInvokeWrapper struct {
 // the underlying value, dereferences it, and calls the real method. This
 // wrapper is only needed when the interface value actually doesn't fit in a
 // pointer and a pointer to the value must be created.
-func (c *Compiler) getInterfaceInvokeWrapper(f *ir.Function) (llvm.Value, error) {
+func (c *Compiler) getInterfaceInvokeWrapper(f *ir.Function) llvm.Value {
 	wrapperName := f.LinkName() + "$invoke"
 	wrapper := c.mod.NamedFunction(wrapperName)
 	if !wrapper.IsNil() {
 		// Wrapper already created. Return it directly.
-		return wrapper, nil
+		return wrapper
 	}
 
 	// Get the expanded receiver type.
-	receiverType, err := c.getLLVMType(f.Params[0].Type())
-	if err != nil {
-		return llvm.Value{}, err
-	}
+	receiverType := c.getLLVMType(f.Params[0].Type())
 	expandedReceiverType := c.expandFormalParamType(receiverType)
 
 	// Does this method even need any wrapping?
@@ -454,7 +457,7 @@ func (c *Compiler) getInterfaceInvokeWrapper(f *ir.Function) (llvm.Value, error)
 		// 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, nil
+		return f.LLVMFn
 	}
 
 	// create wrapper function
@@ -467,12 +470,12 @@ func (c *Compiler) getInterfaceInvokeWrapper(f *ir.Function) (llvm.Value, error)
 		wrapper:      wrapper,
 		receiverType: receiverType,
 	})
-	return wrapper, nil
+	return wrapper
 }
 
 // createInterfaceInvokeWrapper finishes the work of getInterfaceInvokeWrapper,
 // see that function for details.
-func (c *Compiler) createInterfaceInvokeWrapper(state interfaceInvokeWrapper) error {
+func (c *Compiler) createInterfaceInvokeWrapper(state interfaceInvokeWrapper) {
 	wrapper := state.wrapper
 	fn := state.fn
 	receiverType := state.receiverType
@@ -482,10 +485,7 @@ func (c *Compiler) createInterfaceInvokeWrapper(state interfaceInvokeWrapper) er
 	// add debug info if needed
 	if c.Debug {
 		pos := c.ir.Program.Fset.Position(fn.Pos())
-		difunc, err := c.attachDebugInfoRaw(fn, wrapper, "$invoke", pos.Filename, pos.Line)
-		if err != nil {
-			return err
-		}
+		difunc := c.attachDebugInfoRaw(fn, wrapper, "$invoke", pos.Filename, pos.Line)
 		c.builder.SetCurrentDebugLocation(uint(pos.Line), uint(pos.Column), difunc, llvm.Metadata{})
 	}
 
@@ -493,28 +493,7 @@ func (c *Compiler) createInterfaceInvokeWrapper(state interfaceInvokeWrapper) er
 	block := c.ctx.AddBasicBlock(wrapper, "entry")
 	c.builder.SetInsertPointAtEnd(block)
 
-	var receiverPtr llvm.Value
-	if c.targetData.TypeAllocSize(receiverType) > c.targetData.TypeAllocSize(c.i8ptrType) {
-		// The receiver is passed in using a pointer. We have to load it here
-		// and pass it by value to the real function.
-
-		// Load the underlying value.
-		receiverPtrType := llvm.PointerType(receiverType, 0)
-		receiverPtr = c.builder.CreateBitCast(wrapper.Param(0), receiverPtrType, "receiver.ptr")
-	} else {
-		// The value is stored in the interface, but it is of type struct which
-		// is expanded to multiple parameters (e.g. {i8, i8}). So we have to
-		// receive the struct as parameter, expand it, and pass it on to the
-		// real function.
-
-		// Cast the passed-in i8* to the struct value (using an alloca) and
-		// extract its values.
-		alloca := c.builder.CreateAlloca(c.i8ptrType, "receiver.alloca")
-		c.builder.CreateStore(wrapper.Param(0), alloca)
-		receiverPtr = c.builder.CreateBitCast(alloca, llvm.PointerType(receiverType, 0), "receiver.ptr")
-	}
-
-	receiverValue := c.builder.CreateLoad(receiverPtr, "receiver")
+	receiverValue := c.emitPointerUnpack(wrapper.Param(0), []llvm.Type{receiverType})[0]
 	params := append(c.expandFormalParam(receiverValue), wrapper.Params()[1:]...)
 	if fn.LLVMFn.Type().ElementType().ReturnType().TypeKind() == llvm.VoidTypeKind {
 		c.builder.CreateCall(fn.LLVMFn, params, "")
@@ -523,6 +502,4 @@ func (c *Compiler) createInterfaceInvokeWrapper(state interfaceInvokeWrapper) er
 		ret := c.builder.CreateCall(fn.LLVMFn, params, "ret")
 		c.builder.CreateRet(ret)
 	}
-
-	return nil
 }

compiler/llvm.go

@@ -1,6 +1,8 @@
 package compiler
 
 import (
+	"reflect"
+
 	"tinygo.org/x/go-llvm"
 )
 
@@ -22,6 +24,65 @@ func getUses(value llvm.Value) []llvm.Value {
 	return uses
 }
 
+// createEntryBlockAlloca creates a new alloca in the entry block, even though
+// the IR builder is located elsewhere. It assumes that the insert point is
+// at the end of the current block.
+func (c *Compiler) createEntryBlockAlloca(t llvm.Type, name string) llvm.Value {
+	currentBlock := c.builder.GetInsertBlock()
+	entryBlock := currentBlock.Parent().EntryBasicBlock()
+	if entryBlock.FirstInstruction().IsNil() {
+		c.builder.SetInsertPointAtEnd(entryBlock)
+	} else {
+		c.builder.SetInsertPointBefore(entryBlock.FirstInstruction())
+	}
+	alloca := c.builder.CreateAlloca(t, name)
+	c.builder.SetInsertPointAtEnd(currentBlock)
+	return alloca
+}
+
+// createTemporaryAlloca creates a new alloca in the entry block and adds
+// lifetime start infromation in the IR signalling that the alloca won't be used
+// before this point.
+//
+// This is useful for creating temporary allocas for intrinsics. Don't forget to
+// end the lifetime using emitLifetimeEnd after you're done with it.
+func (c *Compiler) createTemporaryAlloca(t llvm.Type, name string) (alloca, bitcast, size llvm.Value) {
+	alloca = c.createEntryBlockAlloca(t, name)
+	bitcast = c.builder.CreateBitCast(alloca, c.i8ptrType, name+".bitcast")
+	size = llvm.ConstInt(c.ctx.Int64Type(), c.targetData.TypeAllocSize(t), false)
+	c.builder.CreateCall(c.getLifetimeStartFunc(), []llvm.Value{size, bitcast}, "")
+	return
+}
+
+// emitLifetimeEnd signals the end of an (alloca) lifetime by calling the
+// llvm.lifetime.end intrinsic. It is commonly used together with
+// createTemporaryAlloca.
+func (c *Compiler) emitLifetimeEnd(ptr, size llvm.Value) {
+	c.builder.CreateCall(c.getLifetimeEndFunc(), []llvm.Value{size, ptr}, "")
+}
+
+// getLifetimeStartFunc returns the llvm.lifetime.start intrinsic and creates it
+// first if it doesn't exist yet.
+func (c *Compiler) getLifetimeStartFunc() llvm.Value {
+	fn := c.mod.NamedFunction("llvm.lifetime.start.p0i8")
+	if fn.IsNil() {
+		fnType := llvm.FunctionType(c.ctx.VoidType(), []llvm.Type{c.ctx.Int64Type(), c.i8ptrType}, false)
+		fn = llvm.AddFunction(c.mod, "llvm.lifetime.start.p0i8", fnType)
+	}
+	return fn
+}
+
+// getLifetimeEndFunc returns the llvm.lifetime.end intrinsic and creates it
+// first if it doesn't exist yet.
+func (c *Compiler) getLifetimeEndFunc() llvm.Value {
+	fn := c.mod.NamedFunction("llvm.lifetime.end.p0i8")
+	if fn.IsNil() {
+		fnType := llvm.FunctionType(c.ctx.VoidType(), []llvm.Type{c.ctx.Int64Type(), c.i8ptrType}, false)
+		fn = llvm.AddFunction(c.mod, "llvm.lifetime.end.p0i8", fnType)
+	}
+	return fn
+}
+
 // splitBasicBlock splits a LLVM basic block into two parts. All instructions
 // after afterInst are moved into a new basic block (created right after the
 // current one) with the given name.
@@ -93,3 +154,48 @@ func (c *Compiler) splitBasicBlock(afterInst llvm.Value, insertAfter llvm.BasicB
 
 	return newBlock
 }
+
+// makeGlobalArray creates a new LLVM global with the given name and integers as
+// contents, and returns the global.
+// Note that it is left with the default linkage etc., you should set
+// linkage/constant/etc properties yourself.
+func (c *Compiler) makeGlobalArray(bufItf interface{}, name string, elementType llvm.Type) llvm.Value {
+	buf := reflect.ValueOf(bufItf)
+	globalType := llvm.ArrayType(elementType, buf.Len())
+	global := llvm.AddGlobal(c.mod, globalType, name)
+	value := llvm.Undef(globalType)
+	for i := 0; i < buf.Len(); i++ {
+		ch := buf.Index(i).Uint()
+		value = llvm.ConstInsertValue(value, llvm.ConstInt(elementType, ch, false), []uint32{uint32(i)})
+	}
+	global.SetInitializer(value)
+	return global
+}
+
+// getGlobalBytes returns the slice contained in the array of the provided
+// global. It can recover the bytes originally created using makeGlobalArray, if
+// makeGlobalArray was given a byte slice.
+func getGlobalBytes(global llvm.Value) []byte {
+	value := global.Initializer()
+	buf := make([]byte, value.Type().ArrayLength())
+	for i := range buf {
+		buf[i] = byte(llvm.ConstExtractValue(value, []uint32{uint32(i)}).ZExtValue())
+	}
+	return buf
+}
+
+// replaceGlobalByteWithArray replaces a global integer type in the module with
+// an integer array, using a GEP to make the types match. It is a convenience
+// function used for creating reflection sidetables, for example.
+func (c *Compiler) replaceGlobalIntWithArray(name string, buf interface{}) llvm.Value {
+	oldGlobal := c.mod.NamedGlobal(name)
+	global := c.makeGlobalArray(buf, name+".tmp", oldGlobal.Type().ElementType())
+	gep := llvm.ConstGEP(global, []llvm.Value{
+		llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+		llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+	})
+	oldGlobal.ReplaceAllUsesWith(gep)
+	oldGlobal.EraseFromParentAsGlobal()
+	global.SetName(name)
+	return global
+}

compiler/map.go

@@ -10,12 +10,14 @@ import (
 )
 
 func (c *Compiler) emitMapLookup(keyType, valueType types.Type, m, key llvm.Value, commaOk bool, pos token.Pos) (llvm.Value, error) {
-	llvmValueType, err := c.getLLVMType(valueType)
-	if err != nil {
-		return llvm.Value{}, err
-	}
-	mapValueAlloca := c.builder.CreateAlloca(llvmValueType, "hashmap.value")
-	mapValuePtr := c.builder.CreateBitCast(mapValueAlloca, c.i8ptrType, "hashmap.valueptr")
+	llvmValueType := c.getLLVMType(valueType)
+
+	// Allocate the memory for the resulting type. Do not zero this memory: it
+	// will be zeroed by the hashmap get implementation if the key is not
+	// present in the map.
+	mapValueAlloca, mapValuePtr, mapValueSize := c.createTemporaryAlloca(llvmValueType, "hashmap.value")
+
+	// Do the lookup. How it is done depends on the key type.
 	var commaOkValue llvm.Value
 	if t, ok := keyType.(*types.Basic); ok && t.Info()&types.IsString != 0 {
 		// key is a string
@@ -23,15 +25,24 @@ func (c *Compiler) emitMapLookup(keyType, valueType types.Type, m, key llvm.Valu
 		commaOkValue = c.createRuntimeCall("hashmapStringGet", params, "")
 	} else if hashmapIsBinaryKey(keyType) {
 		// key can be compared with runtime.memequal
-		keyAlloca := c.builder.CreateAlloca(key.Type(), "hashmap.key")
-		c.builder.CreateStore(key, keyAlloca)
-		keyPtr := c.builder.CreateBitCast(keyAlloca, c.i8ptrType, "hashmap.keyptr")
-		params := []llvm.Value{m, keyPtr, mapValuePtr}
+		// Store the key in an alloca, in the entry block to avoid dynamic stack
+		// growth.
+		mapKeyAlloca, mapKeyPtr, mapKeySize := c.createTemporaryAlloca(key.Type(), "hashmap.key")
+		c.builder.CreateStore(key, mapKeyAlloca)
+		// Fetch the value from the hashmap.
+		params := []llvm.Value{m, mapKeyPtr, mapValuePtr}
 		commaOkValue = c.createRuntimeCall("hashmapBinaryGet", params, "")
+		c.emitLifetimeEnd(mapKeyPtr, mapKeySize)
 	} else {
-		return llvm.Value{}, c.makeError(pos, "only strings, bools, ints or structs of bools/ints are supported as map keys, but got: "+keyType.String())
+		// Not trivially comparable using memcmp.
+		return llvm.Value{}, c.makeError(pos, "only strings, bools, ints, pointers or structs of bools/ints are supported as map keys, but got: "+keyType.String())
 	}
+
+	// Load the resulting value from the hashmap. The value is set to the zero
+	// value if the key doesn't exist in the hashmap.
 	mapValue := c.builder.CreateLoad(mapValueAlloca, "")
+	c.emitLifetimeEnd(mapValuePtr, mapValueSize)
+
 	if commaOk {
 		tuple := llvm.Undef(c.ctx.StructType([]llvm.Type{llvmValueType, c.ctx.Int1Type()}, false))
 		tuple = c.builder.CreateInsertValue(tuple, mapValue, 0, "")
@@ -42,27 +53,25 @@ func (c *Compiler) emitMapLookup(keyType, valueType types.Type, m, key llvm.Valu
 	}
 }
 
-func (c *Compiler) emitMapUpdate(keyType types.Type, m, key, value llvm.Value, pos token.Pos) error {
-	valueAlloca := c.builder.CreateAlloca(value.Type(), "hashmap.value")
+func (c *Compiler) emitMapUpdate(keyType types.Type, m, key, value llvm.Value, pos token.Pos) {
+	valueAlloca, valuePtr, valueSize := c.createTemporaryAlloca(value.Type(), "hashmap.value")
 	c.builder.CreateStore(value, valueAlloca)
-	valuePtr := c.builder.CreateBitCast(valueAlloca, c.i8ptrType, "hashmap.valueptr")
 	keyType = keyType.Underlying()
 	if t, ok := keyType.(*types.Basic); ok && t.Info()&types.IsString != 0 {
 		// key is a string
 		params := []llvm.Value{m, key, valuePtr}
 		c.createRuntimeCall("hashmapStringSet", params, "")
-		return nil
 	} else if hashmapIsBinaryKey(keyType) {
 		// key can be compared with runtime.memequal
-		keyAlloca := c.builder.CreateAlloca(key.Type(), "hashmap.key")
+		keyAlloca, keyPtr, keySize := c.createTemporaryAlloca(key.Type(), "hashmap.key")
 		c.builder.CreateStore(key, keyAlloca)
-		keyPtr := c.builder.CreateBitCast(keyAlloca, c.i8ptrType, "hashmap.keyptr")
 		params := []llvm.Value{m, keyPtr, valuePtr}
 		c.createRuntimeCall("hashmapBinarySet", params, "")
-		return nil
+		c.emitLifetimeEnd(keyPtr, keySize)
 	} else {
-		return c.makeError(pos, "only strings, bools, ints or structs of bools/ints are supported as map keys, but got: "+keyType.String())
+		c.addError(pos, "only strings, bools, ints, pointers or structs of bools/ints are supported as map keys, but got: "+keyType.String())
 	}
+	c.emitLifetimeEnd(valuePtr, valueSize)
 }
 
 func (c *Compiler) emitMapDelete(keyType types.Type, m, key llvm.Value, pos token.Pos) error {
@@ -73,14 +82,14 @@ func (c *Compiler) emitMapDelete(keyType types.Type, m, key llvm.Value, pos toke
 		c.createRuntimeCall("hashmapStringDelete", params, "")
 		return nil
 	} else if hashmapIsBinaryKey(keyType) {
-		keyAlloca := c.builder.CreateAlloca(key.Type(), "hashmap.key")
+		keyAlloca, keyPtr, keySize := c.createTemporaryAlloca(key.Type(), "hashmap.key")
 		c.builder.CreateStore(key, keyAlloca)
-		keyPtr := c.builder.CreateBitCast(keyAlloca, c.i8ptrType, "hashmap.keyptr")
 		params := []llvm.Value{m, keyPtr}
 		c.createRuntimeCall("hashmapBinaryDelete", params, "")
+		c.emitLifetimeEnd(keyPtr, keySize)
 		return nil
 	} else {
-		return c.makeError(pos, "only strings, bools, ints or structs of bools/ints are supported as map keys, but got: "+keyType.String())
+		return c.makeError(pos, "only strings, bools, ints, pointers or structs of bools/ints are supported as map keys, but got: "+keyType.String())
 	}
 }
 
@@ -112,6 +121,8 @@ func hashmapIsBinaryKey(keyType types.Type) bool {
 	switch keyType := keyType.(type) {
 	case *types.Basic:
 		return keyType.Info()&(types.IsBoolean|types.IsInteger) != 0
+	case *types.Pointer:
+		return true
 	case *types.Struct:
 		for i := 0; i < keyType.NumFields(); i++ {
 			fieldType := keyType.Field(i).Type().Underlying()

compiler/optimizer.go

@@ -3,6 +3,7 @@ package compiler
 import (
 	"errors"
 
+	"github.com/tinygo-org/tinygo/transform"
 	"tinygo.org/x/go-llvm"
 )
 
@@ -18,6 +19,15 @@ func (c *Compiler) Optimize(optLevel, sizeLevel int, inlinerThreshold uint) erro
 	}
 	builder.AddCoroutinePassesToExtensionPoints()
 
+	if c.PanicStrategy == "trap" {
+		c.replacePanicsWithTrap() // -panic=trap
+	}
+
+	// run a check of all of our code
+	if c.VerifyIR {
+		c.checkModule()
+	}
+
 	// Run function passes for each function.
 	funcPasses := llvm.NewFunctionPassManagerForModule(c.mod)
 	defer funcPasses.Dispose()
@@ -39,10 +49,11 @@ func (c *Compiler) Optimize(optLevel, sizeLevel int, inlinerThreshold uint) erro
 		goPasses.Run(c.mod)
 
 		// Run Go-specific optimization passes.
-		c.OptimizeMaps()
+		transform.OptimizeMaps(c.mod)
 		c.OptimizeStringToBytes()
-		c.OptimizeAllocs()
+		transform.OptimizeAllocs(c.mod)
 		c.LowerInterfaces()
+		c.LowerFuncValues()
 
 		// After interfaces are lowered, there are many more opportunities for
 		// interprocedural optimizations. To get them to work, function
@@ -50,9 +61,25 @@ func (c *Compiler) Optimize(optLevel, sizeLevel int, inlinerThreshold uint) erro
 		goPasses.Run(c.mod)
 
 		// Run TinyGo-specific interprocedural optimizations.
-		c.OptimizeAllocs()
+		transform.OptimizeAllocs(c.mod)
 		c.OptimizeStringToBytes()
 
+		// Lower runtime.isnil calls to regular nil comparisons.
+		isnil := c.mod.NamedFunction("runtime.isnil")
+		if !isnil.IsNil() {
+			for _, use := range getUses(isnil) {
+				c.builder.SetInsertPointBefore(use)
+				ptr := use.Operand(0)
+				if !ptr.IsABitCastInst().IsNil() {
+					ptr = ptr.Operand(0)
+				}
+				nilptr := llvm.ConstPointerNull(ptr.Type())
+				icmp := c.builder.CreateICmp(llvm.IntEQ, ptr, nilptr, "")
+				use.ReplaceAllUsesWith(icmp)
+				use.EraseFromParentAsInstruction()
+			}
+		}
+
 		err := c.LowerGoroutines()
 		if err != nil {
 			return err
@@ -60,6 +87,7 @@ func (c *Compiler) Optimize(optLevel, sizeLevel int, inlinerThreshold uint) erro
 	} else {
 		// Must be run at any optimization level.
 		c.LowerInterfaces()
+		c.LowerFuncValues()
 		err := c.LowerGoroutines()
 		if err != nil {
 			return err
@@ -79,6 +107,15 @@ func (c *Compiler) Optimize(optLevel, sizeLevel int, inlinerThreshold uint) erro
 		}
 	}
 
+	// After TinyGo-specific transforms have finished, undo exporting these functions.
+	for _, name := range functionsUsedInTransforms {
+		fn := c.mod.NamedFunction(name)
+		if fn.IsNil() {
+			continue
+		}
+		fn.SetLinkage(llvm.InternalLinkage)
+	}
+
 	// Run function passes again, because without it, llvm.coro.size.i32()
 	// doesn't get lowered.
 	for fn := c.mod.FirstFunction(); !fn.IsNil(); fn = llvm.NextFunction(fn) {
@@ -92,47 +129,32 @@ func (c *Compiler) Optimize(optLevel, sizeLevel int, inlinerThreshold uint) erro
 	builder.Populate(modPasses)
 	modPasses.Run(c.mod)
 
+	hasGCPass := c.addGlobalsBitmap()
+	hasGCPass = c.makeGCStackSlots() || hasGCPass
+	if hasGCPass {
+		if err := c.Verify(); err != nil {
+			return errors.New("GC pass caused a verification failure")
+		}
+	}
+
 	return nil
 }
 
-// Eliminate created but not used maps.
-//
-// In the future, this should statically allocate created but never modified
-// maps. This has not yet been implemented, however.
-func (c *Compiler) OptimizeMaps() {
-	hashmapMake := c.mod.NamedFunction("runtime.hashmapMake")
-	if hashmapMake.IsNil() {
-		// nothing to optimize
-		return
-	}
-
-	hashmapBinarySet := c.mod.NamedFunction("runtime.hashmapBinarySet")
-	hashmapStringSet := c.mod.NamedFunction("runtime.hashmapStringSet")
-
-	for _, makeInst := range getUses(hashmapMake) {
-		updateInsts := []llvm.Value{}
-		unknownUses := false // are there any uses other than setting a value?
-
-		for _, use := range getUses(makeInst) {
-			if use := use.IsACallInst(); !use.IsNil() {
-				switch use.CalledValue() {
-				case hashmapBinarySet, hashmapStringSet:
-					updateInsts = append(updateInsts, use)
-				default:
-					unknownUses = true
-				}
-			} else {
-				unknownUses = true
-			}
+// Replace panic calls with calls to llvm.trap, to reduce code size. This is the
+// -panic=trap intrinsic.
+func (c *Compiler) replacePanicsWithTrap() {
+	trap := c.mod.NamedFunction("llvm.trap")
+	for _, name := range []string{"runtime._panic", "runtime.runtimePanic"} {
+		fn := c.mod.NamedFunction(name)
+		if fn.IsNil() {
+			continue
 		}
-
-		if !unknownUses {
-			// This map can be entirely removed, as it is only created but never
-			// used.
-			for _, inst := range updateInsts {
-				inst.EraseFromParentAsInstruction()
+		for _, use := range getUses(fn) {
+			if use.IsACallInst().IsNil() || use.CalledValue() != fn {
+				panic("expected use of a panic function to be a call")
 			}
-			makeInst.EraseFromParentAsInstruction()
+			c.builder.SetInsertPointBefore(use)
+			c.builder.CreateCall(trap, nil, "")
 		}
 	}
 }
@@ -187,110 +209,6 @@ func (c *Compiler) OptimizeStringToBytes() {
 	}
 }
 
-// Basic escape analysis: translate runtime.alloc calls into alloca
-// instructions.
-func (c *Compiler) OptimizeAllocs() {
-	allocator := c.mod.NamedFunction("runtime.alloc")
-	if allocator.IsNil() {
-		// nothing to optimize
-		return
-	}
-
-	heapallocs := getUses(allocator)
-	for _, heapalloc := range heapallocs {
-		nilValue := llvm.Value{}
-		if heapalloc.Operand(0).IsAConstant() == nilValue {
-			// Do not allocate variable length arrays on the stack.
-			continue
-		}
-		size := heapalloc.Operand(0).ZExtValue()
-		if size > 256 {
-			// The maximum value for a stack allocation.
-			// TODO: tune this, this is just a random value.
-			continue
-		}
-
-		// In general the pattern is:
-		//     %0 = call i8* @runtime.alloc(i32 %size)
-		//     %1 = bitcast i8* %0 to type*
-		//     (use %1 only)
-		// But the bitcast might sometimes be dropped when allocating an *i8.
-		// The 'bitcast' variable below is thus usually a bitcast of the
-		// heapalloc but not always.
-		bitcast := heapalloc // instruction that creates the value
-		if uses := getUses(heapalloc); len(uses) == 1 && uses[0].IsABitCastInst() != nilValue {
-			// getting only bitcast use
-			bitcast = uses[0]
-		}
-		if !c.doesEscape(bitcast) {
-			// Insert alloca in the entry block. Do it here so that mem2reg can
-			// promote it to a SSA value.
-			fn := bitcast.InstructionParent().Parent()
-			c.builder.SetInsertPointBefore(fn.EntryBasicBlock().FirstInstruction())
-			alignment := c.targetData.ABITypeAlignment(c.i8ptrType)
-			sizeInWords := (size + uint64(alignment) - 1) / uint64(alignment)
-			allocaType := llvm.ArrayType(c.ctx.IntType(alignment*8), int(sizeInWords))
-			alloca := c.builder.CreateAlloca(allocaType, "stackalloc.alloca")
-			zero, _ := c.getZeroValue(alloca.Type().ElementType())
-			c.builder.CreateStore(zero, alloca)
-			stackalloc := c.builder.CreateBitCast(alloca, bitcast.Type(), "stackalloc")
-			bitcast.ReplaceAllUsesWith(stackalloc)
-			if heapalloc != bitcast {
-				bitcast.EraseFromParentAsInstruction()
-			}
-			heapalloc.EraseFromParentAsInstruction()
-		}
-	}
-}
-
-// Very basic escape analysis.
-func (c *Compiler) doesEscape(value llvm.Value) bool {
-	uses := getUses(value)
-	for _, use := range uses {
-		nilValue := llvm.Value{}
-		if use.IsAGetElementPtrInst() != nilValue {
-			if c.doesEscape(use) {
-				return true
-			}
-		} else if use.IsABitCastInst() != nilValue {
-			// A bitcast escapes if the casted-to value escapes.
-			if c.doesEscape(use) {
-				return true
-			}
-		} else if use.IsALoadInst() != nilValue {
-			// Load does not escape.
-		} else if use.IsAStoreInst() != nilValue {
-			// Store only escapes when the value is stored to, not when the
-			// value is stored into another value.
-			if use.Operand(0) == value {
-				return true
-			}
-		} else if use.IsACallInst() != nilValue {
-			// Call only escapes when the (pointer) parameter is not marked
-			// "nocapture". This flag means that the parameter does not escape
-			// the give function.
-			if use.CalledValue().IsAFunction() != nilValue {
-				if use.CalledValue().IsDeclaration() {
-					// Kind of dirty: assume external functions don't let
-					// pointers escape.
-					// TODO: introduce //go:noescape that sets the 'nocapture'
-					// flag on each input parameter.
-					continue
-				}
-			}
-			if !c.hasFlag(use, value, "nocapture") {
-				return true
-			}
-		} else {
-			// Unknown instruction, might escape.
-			return true
-		}
-	}
-
-	// does not escape
-	return false
-}
-
 // Check whether the given value (which is of pointer type) is never stored to.
 func (c *Compiler) isReadOnly(value llvm.Value) bool {
 	uses := getUses(value)

compiler/reflect.go

@@ -1,10 +1,43 @@
 package compiler
 
+// This file has some compiler support for run-time reflection using the reflect
+// package. In particular, it encodes type information in type codes in such a
+// way that the reflect package can decode the type from this information.
+// Where needed, it also adds some side tables for looking up more information
+// about a type, when that information cannot be stored directly in the type
+// code.
+//
+// Go has 26 different type kinds.
+//
+// Type kinds are subdivided in basic types (see the list of basicTypes below)
+// that are mostly numeric literals and non-basic (or "complex") types that are
+// more difficult to encode. These non-basic types come in two forms:
+//   * Prefix types (pointer, slice, interface, channel): these just add
+//     something to an existing type. For example, a pointer like *int just adds
+//     the fact that it's a pointer to an existing type (int).
+//     These are encoded efficiently by adding a prefix to a type code.
+//   * Types with multiple fields (struct, array, func, map). All of these have
+//     multiple fields contained within. Most obviously structs can contain many
+//     types as fields. Also arrays contain not just the element type but also
+//     the length parameter which can be any arbitrary number and thus may not
+//     fit in a type code.
+//     These types are encoded using side tables.
+//
+// This distinction is also important for how named types are encoded. At the
+// moment, named basic type just get a unique number assigned while named
+// non-basic types have their underlying type stored in a sidetable.
+
 import (
+	"encoding/binary"
+	"go/ast"
 	"math/big"
 	"strings"
+
+	"tinygo.org/x/go-llvm"
 )
 
+// A list of basic types and their numbers. This list should be kept in sync
+// with the list of Kind constants of type.go in the reflect package.
 var basicTypes = map[string]int64{
 	"bool":       1,
 	"int":        2,
@@ -26,6 +59,72 @@ var basicTypes = map[string]int64{
 	"unsafeptr":  18,
 }
 
+// A list of non-basic types. Adding 19 to this number will give the Kind as
+// used in src/reflect/types.go, and it must be kept in sync with that list.
+var nonBasicTypes = map[string]int64{
+	"chan":      0,
+	"interface": 1,
+	"pointer":   2,
+	"slice":     3,
+	"array":     4,
+	"func":      5,
+	"map":       6,
+	"struct":    7,
+}
+
+// typeCodeAssignmentState keeps some global state around for type code
+// assignments, used to assign one unique type code to each Go type.
+type typeCodeAssignmentState struct {
+	// An integer that's incremented each time it's used to give unique IDs to
+	// type codes that are not yet fully supported otherwise by the reflect
+	// package (or are simply unused in the compiled program).
+	fallbackIndex int
+
+	// This is the length of an uintptr. Only used occasionally to know whether
+	// a given number can be encoded as a varint.
+	uintptrLen int
+
+	// Map of named types to their type code. It is important that named types
+	// get unique IDs for each type.
+	namedBasicTypes    map[string]int
+	namedNonBasicTypes map[string]int
+
+	// Map of array types to their type code.
+	arrayTypes               map[string]int
+	arrayTypesSidetable      []byte
+	needsArrayTypesSidetable bool
+
+	// Map of struct types to their type code.
+	structTypes               map[string]int
+	structTypesSidetable      []byte
+	needsStructNamesSidetable bool
+
+	// Map of struct names and tags to their name string.
+	structNames               map[string]int
+	structNamesSidetable      []byte
+	needsStructTypesSidetable bool
+
+	// This byte array is stored in reflect.namedNonBasicTypesSidetable and is
+	// used at runtime to get details about a named non-basic type.
+	// Entries are varints (see makeVarint below and readVarint in
+	// reflect/sidetables.go for the encoding): one varint per entry. The
+	// integers in namedNonBasicTypes are indices into this array. Because these
+	// are varints, most type codes are really small (just one byte).
+	//
+	// Note that this byte buffer is not created when it is not needed
+	// (reflect.namedNonBasicTypesSidetable has no uses), see
+	// needsNamedTypesSidetable.
+	namedNonBasicTypesSidetable []uint64
+
+	// This indicates whether namedNonBasicTypesSidetable needs to be created at
+	// all. If it is false, namedNonBasicTypesSidetable will contain simple
+	// monotonically increasing numbers.
+	needsNamedNonBasicTypesSidetable bool
+}
+
+// assignTypeCodes is used to assign a type code to each type in the program
+// that is ever stored in an interface. It tries to use the smallest possible
+// numbers to make the code that works with interfaces as small as possible.
 func (c *Compiler) assignTypeCodes(typeSlice typeInfoSlice) {
 	fn := c.mod.NamedFunction("reflect.ValueOf")
 	if fn.IsNil() {
@@ -38,13 +137,21 @@ func (c *Compiler) assignTypeCodes(typeSlice typeInfoSlice) {
 	}
 
 	// Assign typecodes the way the reflect package expects.
-	fallbackIndex := 1
-	namedTypes := make(map[string]int)
+	state := typeCodeAssignmentState{
+		fallbackIndex:                    1,
+		uintptrLen:                       c.uintptrType.IntTypeWidth(),
+		namedBasicTypes:                  make(map[string]int),
+		namedNonBasicTypes:               make(map[string]int),
+		arrayTypes:                       make(map[string]int),
+		structTypes:                      make(map[string]int),
+		structNames:                      make(map[string]int),
+		needsNamedNonBasicTypesSidetable: len(getUses(c.mod.NamedGlobal("reflect.namedNonBasicTypesSidetable"))) != 0,
+		needsStructTypesSidetable:        len(getUses(c.mod.NamedGlobal("reflect.structTypesSidetable"))) != 0,
+		needsStructNamesSidetable:        len(getUses(c.mod.NamedGlobal("reflect.structNamesSidetable"))) != 0,
+		needsArrayTypesSidetable:         len(getUses(c.mod.NamedGlobal("reflect.arrayTypesSidetable"))) != 0,
+	}
 	for _, t := range typeSlice {
-		if t.name[:5] != "type:" {
-			panic("expected type name to start with 'type:'")
-		}
-		num := c.getTypeCodeNum(t.name[5:], &fallbackIndex, namedTypes)
+		num := state.getTypeCodeNum(t.typecode)
 		if num.BitLen() > c.uintptrType.IntTypeWidth() || !num.IsUint64() {
 			// TODO: support this in some way, using a side table for example.
 			// That's less efficient but better than not working at all.
@@ -54,25 +161,41 @@ func (c *Compiler) assignTypeCodes(typeSlice typeInfoSlice) {
 		}
 		t.num = num.Uint64()
 	}
+
+	// Only create this sidetable when it is necessary.
+	if state.needsNamedNonBasicTypesSidetable {
+		global := c.replaceGlobalIntWithArray("reflect.namedNonBasicTypesSidetable", state.namedNonBasicTypesSidetable)
+		global.SetLinkage(llvm.InternalLinkage)
+		global.SetUnnamedAddr(true)
+	}
+	if state.needsArrayTypesSidetable {
+		global := c.replaceGlobalIntWithArray("reflect.arrayTypesSidetable", state.arrayTypesSidetable)
+		global.SetLinkage(llvm.InternalLinkage)
+		global.SetUnnamedAddr(true)
+	}
+	if state.needsStructTypesSidetable {
+		global := c.replaceGlobalIntWithArray("reflect.structTypesSidetable", state.structTypesSidetable)
+		global.SetLinkage(llvm.InternalLinkage)
+		global.SetUnnamedAddr(true)
+	}
+	if state.needsStructNamesSidetable {
+		global := c.replaceGlobalIntWithArray("reflect.structNamesSidetable", state.structNamesSidetable)
+		global.SetLinkage(llvm.InternalLinkage)
+		global.SetUnnamedAddr(true)
+	}
 }
 
 // getTypeCodeNum returns the typecode for a given type as expected by the
 // reflect package. Also see getTypeCodeName, which serializes types to a string
 // based on a types.Type value for this function.
-func (c *Compiler) getTypeCodeNum(id string, fallbackIndex *int, namedTypes map[string]int) *big.Int {
+func (state *typeCodeAssignmentState) getTypeCodeNum(typecode llvm.Value) *big.Int {
 	// Note: see src/reflect/type.go for bit allocations.
-	// A type can be named or unnamed. Example of both:
-	//     basic:~foo:uint64
-	//     basic:uint64
-	// Extract the class (basic, slice, pointer, etc.), the name, and the
-	// contents of this type ID string. Allocate bits based on that, as
-	// src/runtime/types.go expects.
-	class := id[:strings.IndexByte(id, ':')]
-	value := id[len(class)+1:]
+	class, value := getClassAndValueFromTypeCode(typecode)
 	name := ""
-	if value[0] == '~' {
-		name = value[1:strings.IndexByte(value, ':')]
-		value = value[len(name)+2:]
+	if class == "named" {
+		name = value
+		typecode = llvm.ConstExtractValue(typecode.Initializer(), []uint32{0})
+		class, value = getClassAndValueFromTypeCode(typecode)
 	}
 	if class == "basic" {
 		// Basic types follow the following bit pattern:
@@ -81,76 +204,286 @@ func (c *Compiler) getTypeCodeNum(id string, fallbackIndex *int, namedTypes map[
 		// upper bits are used to indicate the named type.
 		num, ok := basicTypes[value]
 		if !ok {
-			panic("invalid basic type: " + id)
+			panic("invalid basic type: " + value)
 		}
 		if name != "" {
 			// This type is named, set the upper bits to the name ID.
-			num |= int64(getNamedTypeNum(namedTypes, name)) << 5
+			num |= int64(state.getBasicNamedTypeNum(name)) << 5
 		}
 		return big.NewInt(num << 1)
 	} else {
-		// Complex types use the following bit pattern:
+		// Non-baisc types use the following bit pattern:
 		//    ...nxxx1
-		// where xxx indicates the complex type (any non-basic type). The upper
-		// bits contain whatever the type contains. Types that wrap a single
-		// other type (channel, interface, pointer, slice) just contain the bits
-		// of the wrapped type. Other types (like struct) have a different
-		// method of encoding the contents of the type.
-		var num *big.Int
+		// where xxx indicates the non-basic type. The upper bits contain
+		// whatever the type contains. Types that wrap a single other type
+		// (channel, interface, pointer, slice) just contain the bits of the
+		// wrapped type. Other types (like struct) need more fields and thus
+		// cannot be encoded as a simple prefix.
 		var classNumber int64
-		switch class {
-		case "chan":
-			num = c.getTypeCodeNum(value, fallbackIndex, namedTypes)
-			classNumber = 0
-		case "interface":
-			num = big.NewInt(int64(*fallbackIndex))
-			*fallbackIndex++
-			classNumber = 1
-		case "pointer":
-			num = c.getTypeCodeNum(value, fallbackIndex, namedTypes)
-			classNumber = 2
-		case "slice":
-			num = c.getTypeCodeNum(value, fallbackIndex, namedTypes)
-			classNumber = 3
-		case "array":
-			num = big.NewInt(int64(*fallbackIndex))
-			*fallbackIndex++
-			classNumber = 4
-		case "func":
-			num = big.NewInt(int64(*fallbackIndex))
-			*fallbackIndex++
-			classNumber = 5
-		case "map":
-			num = big.NewInt(int64(*fallbackIndex))
-			*fallbackIndex++
-			classNumber = 6
-		case "struct":
-			num = big.NewInt(int64(*fallbackIndex))
-			*fallbackIndex++
-			classNumber = 7
-		default:
-			panic("unknown type kind: " + id)
-		}
-		if name == "" {
-			num.Lsh(num, 5).Or(num, big.NewInt((classNumber<<1)+1))
+		if n, ok := nonBasicTypes[class]; ok {
+			classNumber = n
 		} else {
-			// TODO: store num in a sidetable
-			num = big.NewInt(int64(getNamedTypeNum(namedTypes, name))<<1 | 1)
-			num.Lsh(num, 4).Or(num, big.NewInt((classNumber<<1)+1))
+			panic("unknown type kind: " + class)
 		}
+		var num *big.Int
+		lowBits := (classNumber << 1) + 1 // the 5 low bits of the typecode
+		if name == "" {
+			num = state.getNonBasicTypeCode(class, typecode)
+		} else {
+			// We must return a named type here. But first check whether it
+			// has already been defined.
+			if index, ok := state.namedNonBasicTypes[name]; ok {
+				num := big.NewInt(int64(index))
+				num.Lsh(num, 5).Or(num, big.NewInt((classNumber<<1)+1+(1<<4)))
+				return num
+			}
+			lowBits |= 1 << 4 // set the 'n' bit (see above)
+			if !state.needsNamedNonBasicTypesSidetable {
+				// Use simple small integers in this case, to make these numbers
+				// smaller.
+				index := len(state.namedNonBasicTypes) + 1
+				state.namedNonBasicTypes[name] = index
+				num = big.NewInt(int64(index))
+			} else {
+				// We need to store full type information.
+				// First allocate a number in the named non-basic type
+				// sidetable.
+				index := len(state.namedNonBasicTypesSidetable)
+				state.namedNonBasicTypesSidetable = append(state.namedNonBasicTypesSidetable, 0)
+				state.namedNonBasicTypes[name] = index
+				// Get the typecode of the underlying type (which could be the
+				// element type in the case of pointers, for example).
+				num = state.getNonBasicTypeCode(class, typecode)
+				if num.BitLen() > state.uintptrLen || !num.IsUint64() {
+					panic("cannot store value in sidetable")
+				}
+				// Now update the side table with the number we just
+				// determined. We need this multi-step approach to avoid stack
+				// overflow due to adding types recursively in the case of
+				// linked lists (a pointer which points to a struct that
+				// contains that same pointer).
+				state.namedNonBasicTypesSidetable[index] = num.Uint64()
+				num = big.NewInt(int64(index))
+			}
+		}
+		// Concatenate the 'num' and 'lowBits' bitstrings.
+		num.Lsh(num, 5).Or(num, big.NewInt(lowBits))
 		return num
 	}
 }
 
-// getNamedTypeNum returns an appropriate (unique) number for the given named
-// type. If the name already has a number that number is returned, else a new
-// number is returned. The number is always non-zero.
-func getNamedTypeNum(namedTypes map[string]int, name string) int {
-	if num, ok := namedTypes[name]; ok {
-		return num
-	} else {
-		num = len(namedTypes) + 1
-		namedTypes[name] = num
+// getNonBasicTypeCode is used by getTypeCodeNum. It returns the upper bits of
+// the type code used there in the type code.
+func (state *typeCodeAssignmentState) getNonBasicTypeCode(class string, typecode llvm.Value) *big.Int {
+	switch class {
+	case "chan", "pointer", "slice":
+		// Prefix-style type kinds. The upper bits contain the element type.
+		sub := llvm.ConstExtractValue(typecode.Initializer(), []uint32{0})
+		return state.getTypeCodeNum(sub)
+	case "array":
+		// An array is basically a pair of (typecode, length) stored in a
+		// sidetable.
+		return big.NewInt(int64(state.getArrayTypeNum(typecode)))
+	case "struct":
+		// More complicated type kind. The upper bits contain the index to the
+		// struct type in the struct types sidetable.
+		return big.NewInt(int64(state.getStructTypeNum(typecode)))
+	default:
+		// Type has not yet been implemented, so fall back by using a unique
+		// number.
+		num := big.NewInt(int64(state.fallbackIndex))
+		state.fallbackIndex++
 		return num
 	}
 }
+
+// getClassAndValueFromTypeCode takes a typecode (a llvm.Value of type
+// runtime.typecodeID), looks at the name, and extracts the typecode class and
+// value from it. For example, for a typecode with the following name:
+//     reflect/types.type:pointer:named:reflect.ValueError
+// It extracts:
+//     class = "pointer"
+//     value = "named:reflect.ValueError"
+func getClassAndValueFromTypeCode(typecode llvm.Value) (class, value string) {
+	typecodeName := typecode.Name()
+	const prefix = "reflect/types.type:"
+	if !strings.HasPrefix(typecodeName, prefix) {
+		panic("unexpected typecode name: " + typecodeName)
+	}
+	id := typecodeName[len(prefix):]
+	class = id[:strings.IndexByte(id, ':')]
+	value = id[len(class)+1:]
+	return
+}
+
+// getBasicNamedTypeNum returns an appropriate (unique) number for the given
+// named type. If the name already has a number that number is returned, else a
+// new number is returned. The number is always non-zero.
+func (state *typeCodeAssignmentState) getBasicNamedTypeNum(name string) int {
+	if num, ok := state.namedBasicTypes[name]; ok {
+		return num
+	}
+	num := len(state.namedBasicTypes) + 1
+	state.namedBasicTypes[name] = num
+	return num
+}
+
+// getArrayTypeNum returns the array type number, which is an index into the
+// reflect.arrayTypesSidetable or a unique number for this type if this table is
+// not used.
+func (state *typeCodeAssignmentState) getArrayTypeNum(typecode llvm.Value) int {
+	name := typecode.Name()
+	if num, ok := state.arrayTypes[name]; ok {
+		// This array type already has an entry in the sidetable. Don't store
+		// it twice.
+		return num
+	}
+
+	if !state.needsArrayTypesSidetable {
+		// We don't need array sidetables, so we can just assign monotonically
+		// increasing numbers to each array type.
+		num := len(state.arrayTypes)
+		state.arrayTypes[name] = num
+		return num
+	}
+
+	elemTypeCode := llvm.ConstExtractValue(typecode.Initializer(), []uint32{0})
+	elemTypeNum := state.getTypeCodeNum(elemTypeCode)
+	if elemTypeNum.BitLen() > state.uintptrLen || !elemTypeNum.IsUint64() {
+		// TODO: make this a regular error
+		panic("array element type has a type code that is too big")
+	}
+
+	// The array side table is a sequence of {element type, array length}.
+	arrayLength := llvm.ConstExtractValue(typecode.Initializer(), []uint32{1}).ZExtValue()
+	buf := makeVarint(elemTypeNum.Uint64())
+	buf = append(buf, makeVarint(arrayLength)...)
+
+	index := len(state.arrayTypesSidetable)
+	state.arrayTypes[name] = index
+	state.arrayTypesSidetable = append(state.arrayTypesSidetable, buf...)
+	return index
+}
+
+// getStructTypeNum returns the struct type number, which is an index into
+// reflect.structTypesSidetable or an unique number for every struct if this
+// sidetable is not needed in the to-be-compiled program.
+func (state *typeCodeAssignmentState) getStructTypeNum(typecode llvm.Value) int {
+	name := typecode.Name()
+	if num, ok := state.structTypes[name]; ok {
+		// This struct already has an assigned type code.
+		return num
+	}
+
+	if !state.needsStructTypesSidetable {
+		// We don't need struct sidetables, so we can just assign monotonically
+		// increasing numbers to each struct type.
+		num := len(state.structTypes)
+		state.structTypes[name] = num
+		return num
+	}
+
+	// Get the fields this struct type contains.
+	// The struct number will be the start index of
+	structTypeGlobal := llvm.ConstExtractValue(typecode.Initializer(), []uint32{0}).Operand(0).Initializer()
+	numFields := structTypeGlobal.Type().ArrayLength()
+
+	// The first data that is stored in the struct sidetable is the number of
+	// fields this struct contains. This is usually just a single byte because
+	// most structs don't contain that many fields, but make it a varint just
+	// to be sure.
+	buf := makeVarint(uint64(numFields))
+
+	// Iterate over every field in the struct.
+	// Every field is stored sequentially in the struct sidetable. Fields can
+	// be retrieved from this list of fields at runtime by iterating over all
+	// of them until the right field has been found.
+	// Perhaps adding some index would speed things up, but it would also make
+	// the sidetable bigger.
+	for i := 0; i < numFields; i++ {
+		// Collect some information about this field.
+		field := llvm.ConstExtractValue(structTypeGlobal, []uint32{uint32(i)})
+
+		nameGlobal := llvm.ConstExtractValue(field, []uint32{1})
+		if nameGlobal == llvm.ConstPointerNull(nameGlobal.Type()) {
+			panic("compiler: no name for this struct field")
+		}
+		fieldNameBytes := getGlobalBytes(nameGlobal.Operand(0))
+		fieldNameNumber := state.getStructNameNumber(fieldNameBytes)
+
+		// See whether this struct field has an associated tag, and if so,
+		// store that tag in the tags sidetable.
+		tagGlobal := llvm.ConstExtractValue(field, []uint32{2})
+		hasTag := false
+		tagNumber := 0
+		if tagGlobal != llvm.ConstPointerNull(tagGlobal.Type()) {
+			hasTag = true
+			tagBytes := getGlobalBytes(tagGlobal.Operand(0))
+			tagNumber = state.getStructNameNumber(tagBytes)
+		}
+
+		// The 'embedded' or 'anonymous' flag for this field.
+		embedded := llvm.ConstExtractValue(field, []uint32{3}).ZExtValue() != 0
+
+		// The first byte in the struct types sidetable is a flags byte with
+		// two bits in it.
+		flagsByte := byte(0)
+		if embedded {
+			flagsByte |= 1
+		}
+		if hasTag {
+			flagsByte |= 2
+		}
+		if ast.IsExported(string(fieldNameBytes)) {
+			flagsByte |= 4
+		}
+		buf = append(buf, flagsByte)
+
+		// Get the type number and add it to the buffer.
+		// All fields have a type, so include it directly here.
+		typeNum := state.getTypeCodeNum(llvm.ConstExtractValue(field, []uint32{0}))
+		if typeNum.BitLen() > state.uintptrLen || !typeNum.IsUint64() {
+			// TODO: make this a regular error
+			panic("struct field has a type code that is too big")
+		}
+		buf = append(buf, makeVarint(typeNum.Uint64())...)
+
+		// Add the name.
+		buf = append(buf, makeVarint(uint64(fieldNameNumber))...)
+
+		// Add the tag, if there is one.
+		if hasTag {
+			buf = append(buf, makeVarint(uint64(tagNumber))...)
+		}
+	}
+
+	num := len(state.structTypesSidetable)
+	state.structTypes[name] = num
+	state.structTypesSidetable = append(state.structTypesSidetable, buf...)
+	return num
+}
+
+// getStructNameNumber stores this string (name or tag) onto the struct names
+// sidetable. The format is a varint of the length of the struct, followed by
+// the raw bytes of the name. Multiple identical strings are stored under the
+// same name for space efficiency.
+func (state *typeCodeAssignmentState) getStructNameNumber(nameBytes []byte) int {
+	name := string(nameBytes)
+	if n, ok := state.structNames[name]; ok {
+		// This name was used before, re-use it now (for space efficiency).
+		return n
+	}
+	// This name is not yet in the names sidetable. Add it now.
+	n := len(state.structNamesSidetable)
+	state.structNames[name] = n
+	state.structNamesSidetable = append(state.structNamesSidetable, makeVarint(uint64(len(nameBytes)))...)
+	state.structNamesSidetable = append(state.structNamesSidetable, nameBytes...)
+	return n
+}
+
+// makeVarint is a small helper function that returns the bytes of the number in
+// varint encoding.
+func makeVarint(n uint64) []byte {
+	buf := make([]byte, binary.MaxVarintLen64)
+	return buf[:binary.PutUvarint(buf, n)]
+}

compiler/sizes.go

@@ -63,6 +63,12 @@ func (s *StdSizes) Alignof(T types.Type) int64 {
 
 func (s *StdSizes) Offsetsof(fields []*types.Var) []int64 {
 	offsets := make([]int64, len(fields))
+	if len(fields) > 1 && fields[0].Name() == "C union" {
+		// This struct contains the magic "C union" field which indicates that
+		// this is actually a union from CGo.
+		// All fields in the union start at 0 so return that.
+		return offsets // all fields are still set to 0
+	}
 	var o int64
 	for i, f := range fields {
 		a := s.Alignof(f.Type())
@@ -107,6 +113,9 @@ func (s *StdSizes) Sizeof(T types.Type) int64 {
 		if k == types.Uintptr {
 			return s.PtrSize
 		}
+		if k == types.UnsafePointer {
+			return s.PtrSize
+		}
 		panic("unknown basic type: " + t.String())
 	case *types.Array:
 		n := t.Len()
@@ -125,11 +134,38 @@ func (s *StdSizes) Sizeof(T types.Type) int64 {
 			return 0
 		}
 		fields := make([]*types.Var, t.NumFields())
+		maxAlign := int64(1)
 		for i := range fields {
-			fields[i] = t.Field(i)
+			field := t.Field(i)
+			fields[i] = field
+			al := s.Alignof(field.Type())
+			if al > maxAlign {
+				maxAlign = al
+			}
+		}
+		if fields[0].Name() == "C union" {
+			// Magic field that indicates this is a CGo union and not a struct.
+			// The size is the biggest element, aligned to the element with the
+			// biggest alignment. This is not necessarily the same, for example
+			// in the following union:
+			//     union { int32_t l; int16_t s[3] }
+			maxSize := int64(0)
+			for _, field := range fields[1:] {
+				si := s.Sizeof(field.Type())
+				if si > maxSize {
+					maxSize = si
+				}
+			}
+			return align(maxSize, maxAlign)
+		} else {
+			// This is a regular struct.
+			// Pick the size that fits this struct and add some alignment. Some
+			// structs have some extra padding at the end which should also be
+			// taken care of:
+			//     struct { int32 n; byte b }
+			offsets := s.Offsetsof(fields)
+			return align(offsets[n-1]+s.Sizeof(fields[n-1].Type()), maxAlign)
 		}
-		offsets := s.Offsetsof(fields)
-		return offsets[n-1] + s.Sizeof(fields[n-1].Type())
 	case *types.Interface:
 		return s.PtrSize * 2
 	case *types.Pointer:

compiler/symbol.go

@@ -0,0 +1,107 @@
+package compiler
+
+// This file manages symbols, that is, functions and globals. It reads their
+// pragmas, determines the link name, etc.
+
+import (
+	"go/ast"
+	"go/token"
+	"go/types"
+	"strings"
+
+	"github.com/tinygo-org/tinygo/loader"
+	"golang.org/x/tools/go/ssa"
+	"tinygo.org/x/go-llvm"
+)
+
+// 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).
+type globalInfo struct {
+	linkName string // go:extern
+	extern   bool   // go:extern
+}
+
+// loadASTComments loads comments on globals from the AST, for use later in the
+// program. In particular, they are required for //go:extern pragmas on globals.
+func (c *Compiler) loadASTComments(lprogram *loader.Program) {
+	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
+								}
+							}
+						}
+					}
+				}
+			}
+		}
+	}
+}
+
+// getGlobal returns a LLVM IR global value for a Go SSA global. It is added to
+// the LLVM IR if it has not been added already.
+func (c *Compiler) getGlobal(g *ssa.Global) llvm.Value {
+	info := c.getGlobalInfo(g)
+	llvmGlobal := c.mod.NamedGlobal(info.linkName)
+	if llvmGlobal.IsNil() {
+		llvmType := c.getLLVMType(g.Type().(*types.Pointer).Elem())
+		llvmGlobal = llvm.AddGlobal(c.mod, llvmType, info.linkName)
+		if !info.extern {
+			llvmGlobal.SetInitializer(llvm.ConstNull(llvmType))
+			llvmGlobal.SetLinkage(llvm.InternalLinkage)
+		}
+	}
+	return llvmGlobal
+}
+
+// getGlobalInfo returns some information about a specific global.
+func (c *Compiler) getGlobalInfo(g *ssa.Global) globalInfo {
+	info := globalInfo{}
+	if strings.HasPrefix(g.Name(), "C.") {
+		// Created by CGo: such a name cannot be created by regular C code.
+		info.linkName = g.Name()[2:]
+		info.extern = true
+	} else {
+		// Pick the default linkName.
+		info.linkName = g.RelString(nil)
+		// Check for //go: pragmas, which may change the link name (among
+		// others).
+		doc := c.astComments[info.linkName]
+		if doc != nil {
+			info.parsePragmas(doc)
+		}
+	}
+	return info
+}
+
+// Parse //go: pragma comments from the source. In particular, it parses the
+// //go:extern pragma on globals.
+func (info *globalInfo) parsePragmas(doc *ast.CommentGroup) {
+	for _, comment := range doc.List {
+		if !strings.HasPrefix(comment.Text, "//go:") {
+			continue
+		}
+		parts := strings.Fields(comment.Text)
+		switch parts[0] {
+		case "//go:extern":
+			info.extern = true
+			if len(parts) == 2 {
+				info.linkName = parts[1]
+			}
+		}
+	}
+}

compiler/syscall.go

@@ -4,7 +4,6 @@ package compiler
 // compiler builtins.
 
 import (
-	"go/constant"
 	"strconv"
 
 	"golang.org/x/tools/go/ssa"
@@ -14,7 +13,7 @@ import (
 // emitSyscall emits an inline system call instruction, depending on the target
 // OS/arch.
 func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value, error) {
-	num, _ := constant.Uint64Val(call.Args[0].(*ssa.Const).Value)
+	num := c.getValue(frame, call.Args[0])
 	var syscallResult llvm.Value
 	switch {
 	case c.GOARCH == "amd64":
@@ -29,12 +28,12 @@ func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value,
 			// > All system classes enter the kernel via the syscall instruction.
 			//
 			// Source: https://opensource.apple.com/source/xnu/xnu-792.13.8/osfmk/mach/i386/syscall_sw.h
-			num += 0x2000000
+			num = c.builder.CreateOr(num, llvm.ConstInt(c.uintptrType, 0x2000000, false), "")
 		}
 		// Sources:
 		//   https://stackoverflow.com/a/2538212
 		//   https://en.wikibooks.org/wiki/X86_Assembly/Interfacing_with_Linux#syscall
-		args := []llvm.Value{llvm.ConstInt(c.uintptrType, num, false)}
+		args := []llvm.Value{num}
 		argTypes := []llvm.Type{c.uintptrType}
 		// Constraints will look something like:
 		//   "={rax},0,{rdi},{rsi},{rdx},{r10},{r8},{r9},~{rcx},~{r11}"
@@ -51,10 +50,7 @@ func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value,
 				"{r12}",
 				"{r13}",
 			}[i]
-			llvmValue, err := c.parseExpr(frame, arg)
-			if err != nil {
-				return llvm.Value{}, err
-			}
+			llvmValue := c.getValue(frame, arg)
 			args = append(args, llvmValue)
 			argTypes = append(argTypes, llvmValue.Type())
 		}
@@ -62,6 +58,32 @@ func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value,
 		fnType := llvm.FunctionType(c.uintptrType, argTypes, false)
 		target := llvm.InlineAsm(fnType, "syscall", constraints, true, false, llvm.InlineAsmDialectIntel)
 		syscallResult = c.builder.CreateCall(target, args, "")
+	case c.GOARCH == "386" && c.GOOS == "linux":
+		// Sources:
+		//   syscall(2) man page
+		//   https://stackoverflow.com/a/2538212
+		//   https://en.wikibooks.org/wiki/X86_Assembly/Interfacing_with_Linux#int_0x80
+		args := []llvm.Value{num}
+		argTypes := []llvm.Type{c.uintptrType}
+		// Constraints will look something like:
+		//   "={eax},0,{ebx},{ecx},{edx},{esi},{edi},{ebp}"
+		constraints := "={eax},0"
+		for i, arg := range call.Args[1:] {
+			constraints += "," + [...]string{
+				"{ebx}",
+				"{ecx}",
+				"{edx}",
+				"{esi}",
+				"{edi}",
+				"{ebp}",
+			}[i]
+			llvmValue := c.getValue(frame, arg)
+			args = append(args, llvmValue)
+			argTypes = append(argTypes, llvmValue.Type())
+		}
+		fnType := llvm.FunctionType(c.uintptrType, argTypes, false)
+		target := llvm.InlineAsm(fnType, "int 0x80", constraints, true, false, llvm.InlineAsmDialectIntel)
+		syscallResult = c.builder.CreateCall(target, args, "")
 	case c.GOARCH == "arm" && c.GOOS == "linux":
 		// Implement the EABI system call convention for Linux.
 		// Source: syscall(2) man page.
@@ -80,14 +102,11 @@ func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value,
 				"{r5}",
 				"{r6}",
 			}[i]
-			llvmValue, err := c.parseExpr(frame, arg)
-			if err != nil {
-				return llvm.Value{}, err
-			}
+			llvmValue := c.getValue(frame, arg)
 			args = append(args, llvmValue)
 			argTypes = append(argTypes, llvmValue.Type())
 		}
-		args = append(args, llvm.ConstInt(c.uintptrType, num, false))
+		args = append(args, num)
 		argTypes = append(argTypes, c.uintptrType)
 		constraints += ",{r7}" // syscall number
 		for i := len(call.Args) - 1; i < 4; i++ {
@@ -113,14 +132,11 @@ func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value,
 				"{x4}",
 				"{x5}",
 			}[i]
-			llvmValue, err := c.parseExpr(frame, arg)
-			if err != nil {
-				return llvm.Value{}, err
-			}
+			llvmValue := c.getValue(frame, arg)
 			args = append(args, llvmValue)
 			argTypes = append(argTypes, llvmValue.Type())
 		}
-		args = append(args, llvm.ConstInt(c.uintptrType, num, false))
+		args = append(args, num)
 		argTypes = append(argTypes, c.uintptrType)
 		constraints += ",{x8}" // syscall number
 		for i := len(call.Args) - 1; i < 8; i++ {
@@ -149,7 +165,7 @@ func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value,
 		inrange2 := c.builder.CreateICmp(llvm.IntSGT, syscallResult, llvm.ConstInt(c.uintptrType, 0xfffffffffffff000, true), "") // -4096
 		hasError := c.builder.CreateAnd(inrange1, inrange2, "")
 		errResult := c.builder.CreateSelect(hasError, c.builder.CreateSub(zero, syscallResult, ""), zero, "syscallError")
-		retval := llvm.Undef(llvm.StructType([]llvm.Type{c.uintptrType, c.uintptrType, c.uintptrType}, false))
+		retval := llvm.Undef(c.ctx.StructType([]llvm.Type{c.uintptrType, c.uintptrType, c.uintptrType}, false))
 		retval = c.builder.CreateInsertValue(retval, syscallResult, 0, "")
 		retval = c.builder.CreateInsertValue(retval, zero, 1, "")
 		retval = c.builder.CreateInsertValue(retval, errResult, 2, "")
@@ -165,7 +181,7 @@ func (c *Compiler) emitSyscall(frame *Frame, call *ssa.CallCommon) (llvm.Value,
 		zero := llvm.ConstInt(c.uintptrType, 0, false)
 		hasError := c.builder.CreateICmp(llvm.IntNE, syscallResult, llvm.ConstInt(c.uintptrType, 0, false), "")
 		errResult := c.builder.CreateSelect(hasError, syscallResult, zero, "syscallError")
-		retval := llvm.Undef(llvm.StructType([]llvm.Type{c.uintptrType, c.uintptrType, c.uintptrType}, false))
+		retval := llvm.Undef(c.ctx.StructType([]llvm.Type{c.uintptrType, c.uintptrType, c.uintptrType}, false))
 		retval = c.builder.CreateInsertValue(retval, syscallResult, 0, "")
 		retval = c.builder.CreateInsertValue(retval, zero, 1, "")
 		retval = c.builder.CreateInsertValue(retval, errResult, 2, "")

compiler/volatile.go

@@ -0,0 +1,26 @@
+package compiler
+
+// This file implements volatile loads/stores in runtime/volatile.LoadT and
+// runtime/volatile.StoreT as compiler builtins.
+
+import (
+	"golang.org/x/tools/go/ssa"
+	"tinygo.org/x/go-llvm"
+)
+
+func (c *Compiler) emitVolatileLoad(frame *Frame, instr *ssa.CallCommon) (llvm.Value, error) {
+	addr := c.getValue(frame, instr.Args[0])
+	c.emitNilCheck(frame, addr, "deref")
+	val := c.builder.CreateLoad(addr, "")
+	val.SetVolatile(true)
+	return val, nil
+}
+
+func (c *Compiler) emitVolatileStore(frame *Frame, instr *ssa.CallCommon) (llvm.Value, error) {
+	addr := c.getValue(frame, instr.Args[0])
+	val := c.getValue(frame, instr.Args[1])
+	c.emitNilCheck(frame, addr, "deref")
+	store := c.builder.CreateStore(val, addr)
+	store.SetVolatile(true)
+	return llvm.Value{}, nil
+}

compiler/wordpack.go

@@ -0,0 +1,120 @@
+package compiler
+
+// This file contains utility functions to pack and unpack sets of values. It
+// can take in a list of values and tries to store it efficiently in the pointer
+// itself if possible and legal.
+
+import (
+	"tinygo.org/x/go-llvm"
+)
+
+// emitPointerPack packs the list of values into a single pointer value using
+// bitcasts, or else allocates a value on the heap if it cannot be packed in the
+// pointer value directly. It returns the pointer with the packed data.
+func (c *Compiler) emitPointerPack(values []llvm.Value) llvm.Value {
+	valueTypes := make([]llvm.Type, len(values))
+	for i, value := range values {
+		valueTypes[i] = value.Type()
+	}
+	packedType := c.ctx.StructType(valueTypes, false)
+
+	// Allocate memory for the packed data.
+	var packedAlloc, packedHeapAlloc llvm.Value
+	size := c.targetData.TypeAllocSize(packedType)
+	if size == 0 {
+		return llvm.ConstPointerNull(c.i8ptrType)
+	} else if len(values) == 1 && values[0].Type().TypeKind() == llvm.PointerTypeKind {
+		return c.builder.CreateBitCast(values[0], c.i8ptrType, "pack.ptr")
+	} else if size <= c.targetData.TypeAllocSize(c.i8ptrType) {
+		// Packed data fits in a pointer, so store it directly inside the
+		// pointer.
+		if len(values) == 1 && values[0].Type().TypeKind() == llvm.IntegerTypeKind {
+			// Try to keep this cast in SSA form.
+			return c.builder.CreateIntToPtr(values[0], c.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, _, _ = c.createTemporaryAlloca(packedType, "")
+	} else {
+		// Packed data is bigger than a pointer, so allocate it on the heap.
+		sizeValue := llvm.ConstInt(c.uintptrType, size, false)
+		packedHeapAlloc = c.createRuntimeCall("alloc", []llvm.Value{sizeValue}, "")
+		if c.needsStackObjects() {
+			c.trackPointer(packedHeapAlloc)
+		}
+		packedAlloc = c.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(c.ctx.Int32Type(), 0, false),
+			llvm.ConstInt(c.ctx.Int32Type(), uint64(i), false),
+		}
+		gep := c.builder.CreateInBoundsGEP(packedAlloc, indices, "")
+		c.builder.CreateStore(value, gep)
+	}
+
+	if packedHeapAlloc.IsNil() {
+		// Load value (as *i8) from the alloca.
+		packedAlloc = c.builder.CreateBitCast(packedAlloc, llvm.PointerType(c.i8ptrType, 0), "")
+		result := c.builder.CreateLoad(packedAlloc, "")
+		packedPtr := c.builder.CreateBitCast(packedAlloc, c.i8ptrType, "")
+		packedSize := llvm.ConstInt(c.ctx.Int64Type(), c.targetData.TypeAllocSize(packedAlloc.Type()), false)
+		c.emitLifetimeEnd(packedPtr, packedSize)
+		return result
+	} else {
+		// Get the original heap allocation pointer, which already is an *i8.
+		return packedHeapAlloc
+	}
+}
+
+// emitPointerUnpack extracts a list of values packed using emitPointerPack.
+func (c *Compiler) emitPointerUnpack(ptr llvm.Value, valueTypes []llvm.Type) []llvm.Value {
+	packedType := c.ctx.StructType(valueTypes, false)
+
+	// Get a correctly-typed pointer to the packed data.
+	var packedAlloc, packedRawAlloc llvm.Value
+	size := c.targetData.TypeAllocSize(packedType)
+	if size == 0 {
+		// No data to unpack.
+	} else if len(valueTypes) == 1 && valueTypes[0].TypeKind() == llvm.PointerTypeKind {
+		// A single pointer is always stored directly.
+		return []llvm.Value{c.builder.CreateBitCast(ptr, valueTypes[0], "unpack.ptr")}
+	} else if size <= c.targetData.TypeAllocSize(c.i8ptrType) {
+		// Packed data stored directly in pointer.
+		if len(valueTypes) == 1 && valueTypes[0].TypeKind() == llvm.IntegerTypeKind {
+			// Keep this cast in SSA form.
+			return []llvm.Value{c.builder.CreatePtrToInt(ptr, valueTypes[0], "unpack.int")}
+		}
+		// Fallback: load it using an alloca.
+		packedRawAlloc, _, _ = c.createTemporaryAlloca(llvm.PointerType(c.i8ptrType, 0), "unpack.raw.alloc")
+		packedRawValue := c.builder.CreateBitCast(ptr, llvm.PointerType(c.i8ptrType, 0), "unpack.raw.value")
+		c.builder.CreateStore(packedRawValue, packedRawAlloc)
+		packedAlloc = c.builder.CreateBitCast(packedRawAlloc, llvm.PointerType(packedType, 0), "unpack.alloc")
+	} else {
+		// Packed data stored on the heap. Bitcast the passed-in pointer to the
+		// correct pointer type.
+		packedAlloc = c.builder.CreateBitCast(ptr, llvm.PointerType(packedType, 0), "unpack.raw.ptr")
+	}
+	// Load each value from the packed data.
+	values := make([]llvm.Value, len(valueTypes))
+	for i, valueType := range valueTypes {
+		if c.targetData.TypeAllocSize(valueType) == 0 {
+			// This value has length zero, so there's nothing to load.
+			values[i] = llvm.ConstNull(valueType)
+			continue
+		}
+		indices := []llvm.Value{
+			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+			llvm.ConstInt(c.ctx.Int32Type(), uint64(i), false),
+		}
+		gep := c.builder.CreateInBoundsGEP(packedAlloc, indices, "")
+		values[i] = c.builder.CreateLoad(gep, "")
+	}
+	if !packedRawAlloc.IsNil() {
+		allocPtr := c.builder.CreateBitCast(packedRawAlloc, c.i8ptrType, "")
+		allocSize := llvm.ConstInt(c.ctx.Int64Type(), c.targetData.TypeAllocSize(c.uintptrType), false)
+		c.emitLifetimeEnd(allocPtr, allocSize)
+	}
+	return values
+}

go.mod

@@ -0,0 +1,10 @@
+module github.com/tinygo-org/tinygo
+
+go 1.11
+
+require (
+	github.com/blakesmith/ar v0.0.0-20150311145944-8bd4349a67f2
+	github.com/marcinbor85/gohex v0.0.0-20180128172054-7a43cd876e46
+	golang.org/x/tools v0.0.0-20190227180812-8dcc6e70cdef
+	tinygo.org/x/go-llvm v0.0.0-20190818154551-95bc4ffe1add
+)

go.sum

@@ -0,0 +1,16 @@
+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/golang/protobuf v1.2.0/go.mod h1:6lQm79b+lXiMfvg/cZm0SGofjICqVBUtrP5yJMmIC1U=
+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=
+golang.org/x/net v0.0.0-20180724234803-3673e40ba225/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
+golang.org/x/net v0.0.0-20190213061140-3a22650c66bd/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
+golang.org/x/sync v0.0.0-20181108010431-42b317875d0f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
+golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
+golang.org/x/tools v0.0.0-20190227180812-8dcc6e70cdef h1:ymc9FeDom3RIEA3coKokSllBB1hRcMT0tZ1W3Jf9Ids=
+golang.org/x/tools v0.0.0-20190227180812-8dcc6e70cdef/go.mod h1:9Yl7xja0Znq3iFh3HoIrodX9oNMXvdceNzlUR8zjMvY=
+google.golang.org/appengine v1.4.0/go.mod h1:xpcJRLb0r/rnEns0DIKYYv+WjYCduHsrkT7/EB5XEv4=
+tinygo.org/x/go-llvm v0.0.0-20190224120431-7707ae5d1261 h1:rJS2Hga39YAnm7DE4qrPm6Dr/67EOojL0XPzvbEeBiw=
+tinygo.org/x/go-llvm v0.0.0-20190224120431-7707ae5d1261/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
+tinygo.org/x/go-llvm v0.0.0-20190818154551-95bc4ffe1add h1:dFjMH1sLhYADg8UQm7DB56B7e+TfvAmWmEZLhyv3r/w=
+tinygo.org/x/go-llvm v0.0.0-20190818154551-95bc4ffe1add/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=

interp/frame.go

@@ -86,7 +86,7 @@ func (fr *frame) evalBasicBlock(bb, incoming llvm.BasicBlock, indent string) (re
 		case !inst.IsAAllocaInst().IsNil():
 			allocType := inst.Type().ElementType()
 			alloca := llvm.AddGlobal(fr.Mod, allocType, fr.pkgName+"$alloca")
-			alloca.SetInitializer(getZeroValue(allocType))
+			alloca.SetInitializer(llvm.ConstNull(allocType))
 			alloca.SetLinkage(llvm.InternalLinkage)
 			fr.locals[inst] = &LocalValue{
 				Underlying: alloca,
@@ -95,7 +95,7 @@ func (fr *frame) evalBasicBlock(bb, incoming llvm.BasicBlock, indent string) (re
 		case !inst.IsALoadInst().IsNil():
 			operand := fr.getLocal(inst.Operand(0)).(*LocalValue)
 			var value llvm.Value
-			if !operand.IsConstant() || inst.IsVolatile() || operand.Underlying.Opcode() == llvm.BitCast {
+			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()
@@ -176,6 +176,20 @@ func (fr *frame) evalBasicBlock(bb, incoming llvm.BasicBlock, indent string) (re
 					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.
+				panic("unimplemented: bitcast of map")
+			}
 			value := fr.getLocal(operand).(*LocalValue)
 			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateBitCast(value.Value(), inst.Type(), "")}
 
@@ -184,6 +198,25 @@ func (fr *frame) evalBasicBlock(bb, incoming llvm.BasicBlock, indent string) (re
 			lhs := fr.getLocal(inst.Operand(0)).(*LocalValue).Underlying
 			rhs := fr.getLocal(inst.Operand(1)).(*LocalValue).Underlying
 			predicate := inst.IntPredicate()
+			if predicate == llvm.IntEQ && 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.
+				lhsNil, ok1 := isPointerNil(lhs)
+				rhsNil, ok2 := isPointerNil(rhs)
+				if ok1 && ok2 {
+					if lhsNil && rhsNil {
+						// Both are nil, so this icmp is always evaluated to true.
+						fr.locals[inst] = &LocalValue{fr.Eval, llvm.ConstInt(fr.Mod.Context().Int1Type(), 1, false)}
+						continue
+					}
+					if lhsNil != rhsNil {
+						// Only one of them is nil, 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
@@ -220,7 +253,7 @@ func (fr *frame) evalBasicBlock(bb, incoming llvm.BasicBlock, indent string) (re
 					allocType = llvm.ArrayType(allocType, elementCount)
 				}
 				alloc := llvm.AddGlobal(fr.Mod, allocType, fr.pkgName+"$alloc")
-				alloc.SetInitializer(getZeroValue(allocType))
+				alloc.SetInitializer(llvm.ConstNull(allocType))
 				alloc.SetLinkage(llvm.InternalLinkage)
 				result := &LocalValue{
 					Underlying: alloc,
@@ -279,7 +312,7 @@ func (fr *frame) evalBasicBlock(bb, incoming llvm.BasicBlock, indent string) (re
 				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 := getZeroValue(stringType)
+				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}
@@ -302,14 +335,55 @@ func (fr *frame) evalBasicBlock(bb, incoming llvm.BasicBlock, indent string) (re
 				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 := getZeroValue(sliceType)
+				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.makeInterface":
-				uintptrType := callee.Type().Context().IntType(fr.TargetData.PointerSize() * 8)
-				fr.locals[inst] = &LocalValue{fr.Eval, llvm.ConstPtrToInt(inst.Operand(0), uintptrType)}
+			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 {
+					panic("interp: expected typecode to be a ptrtoint")
+				}
+				typecode = typecode.Operand(0)
+				if interfaceMethodSet.IsAConstantExpr().IsNil() || interfaceMethodSet.Opcode() != llvm.GetElementPtr {
+					panic("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 {
+					panic("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 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.
@@ -422,17 +496,21 @@ func (fr *frame) evalBasicBlock(bb, incoming llvm.BasicBlock, indent string) (re
 			}
 			thenBB := inst.Operand(1)
 			elseBB := inst.Operand(2)
-			if !cond.IsConstant() {
+			if !cond.IsAInstruction().IsNil() {
 				return nil, nil, errors.New("interp: branch on a non-constant")
-			} else {
-				switch cond.ZExtValue() {
-				case 0: // false
-					return nil, []llvm.Value{thenBB}, nil // then
-				case 1: // true
-					return nil, []llvm.Value{elseBB}, nil // else
-				default:
-					panic("branch was not true or false")
-				}
+			}
+			if !cond.IsAConstantExpr().IsNil() {
+				// This may happen when the instruction builder could not
+				// const-fold some instructions.
+				return nil, nil, 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:
+				panic("branch was not true or false")
 			}
 		case !inst.IsABranchInst().IsNil() && inst.OperandsCount() == 1:
 			// unconditional branch (goto)

interp/interp.go

@@ -18,7 +18,6 @@ type Eval struct {
 	TargetData      llvm.TargetData
 	Debug           bool
 	builder         llvm.Builder
-	dibuilder       *llvm.DIBuilder
 	dirtyGlobals    map[llvm.Value]struct{}
 	sideEffectFuncs map[llvm.Value]*sideEffectResult // cache of side effect scan results
 }
@@ -38,7 +37,6 @@ func Run(mod llvm.Module, targetData llvm.TargetData, debug bool) error {
 		dirtyGlobals: map[llvm.Value]struct{}{},
 	}
 	e.builder = mod.Context().NewBuilder()
-	e.dibuilder = llvm.NewDIBuilder(mod)
 
 	initAll := mod.NamedFunction(name)
 	bb := initAll.EntryBasicBlock()
@@ -49,7 +47,6 @@ func Run(mod llvm.Module, targetData llvm.TargetData, debug bool) error {
 	e.builder.SetInsertPointBefore(bb.FirstInstruction())
 	dummy := e.builder.CreateAlloca(e.Mod.Context().Int8Type(), "dummy")
 	e.builder.SetInsertPointBefore(dummy)
-	e.builder.SetInstDebugLocation(bb.FirstInstruction())
 	var initCalls []llvm.Value
 	for inst := bb.FirstInstruction(); !inst.IsNil(); inst = llvm.NextInstruction(inst) {
 		if inst == dummy {

interp/scan.go

@@ -28,8 +28,17 @@ func (e *Eval) hasSideEffects(fn llvm.Value) *sideEffectResult {
 	case "runtime.alloc":
 		// Cannot be scanned but can be interpreted.
 		return &sideEffectResult{severity: sideEffectNone}
+	case "runtime.nanotime":
+		// Fixed value at compile time.
+		return &sideEffectResult{severity: sideEffectNone}
 	case "runtime._panic":
 		return &sideEffectResult{severity: sideEffectLimited}
+	case "runtime.interfaceImplements":
+		return &sideEffectResult{severity: sideEffectNone}
+	case "runtime.trackPointer":
+		return &sideEffectResult{severity: sideEffectNone}
+	case "llvm.dbg.value":
+		return &sideEffectResult{severity: sideEffectNone}
 	}
 	if e.sideEffectFuncs == nil {
 		e.sideEffectFuncs = make(map[llvm.Value]*sideEffectResult)
@@ -81,11 +90,6 @@ func (e *Eval) hasSideEffects(fn llvm.Value) *sideEffectResult {
 					continue
 				}
 				if child.IsDeclaration() {
-					switch child.Name() {
-					case "runtime.makeInterface":
-						// Can be interpreted so does not have side effects.
-						continue
-					}
 					// External function call. Assume only limited side effects
 					// (no affected globals, etc.).
 					if e.hasLocalSideEffects(dirtyLocals, inst) {
@@ -107,7 +111,16 @@ func (e *Eval) hasSideEffects(fn llvm.Value) *sideEffectResult {
 				default:
 					panic("unreachable")
 				}
-			case llvm.Load, llvm.Store:
+			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)
 				}

interp/utils.go

@@ -16,50 +16,6 @@ func getUses(value llvm.Value) []llvm.Value {
 	return uses
 }
 
-// Return a zero LLVM value for any LLVM type. Setting this value as an
-// initializer has the same effect as setting 'zeroinitializer' on a value.
-// Sadly, I haven't found a way to do it directly with the Go API but this works
-// just fine.
-func getZeroValue(typ llvm.Type) llvm.Value {
-	switch typ.TypeKind() {
-	case llvm.ArrayTypeKind:
-		subTyp := typ.ElementType()
-		subVal := getZeroValue(subTyp)
-		vals := make([]llvm.Value, typ.ArrayLength())
-		for i := range vals {
-			vals[i] = subVal
-		}
-		return llvm.ConstArray(subTyp, vals)
-	case llvm.FloatTypeKind, llvm.DoubleTypeKind:
-		return llvm.ConstFloat(typ, 0.0)
-	case llvm.IntegerTypeKind:
-		return llvm.ConstInt(typ, 0, false)
-	case llvm.PointerTypeKind:
-		return llvm.ConstPointerNull(typ)
-	case llvm.StructTypeKind:
-		types := typ.StructElementTypes()
-		vals := make([]llvm.Value, len(types))
-		for i, subTyp := range types {
-			val := getZeroValue(subTyp)
-			vals[i] = val
-		}
-		if typ.StructName() != "" {
-			return llvm.ConstNamedStruct(typ, vals)
-		} else {
-			return typ.Context().ConstStruct(vals, false)
-		}
-	case llvm.VectorTypeKind:
-		zero := getZeroValue(typ.ElementType())
-		vals := make([]llvm.Value, typ.VectorSize())
-		for i := range vals {
-			vals[i] = zero
-		}
-		return llvm.ConstVector(vals, false)
-	default:
-		panic("interp: unknown LLVM type: " + typ.String())
-	}
-}
-
 // getStringBytes loads the byte slice of a Go string represented as a
 // {ptr, len} pair.
 func getStringBytes(strPtr Value, strLen llvm.Value) []byte {
@@ -94,3 +50,29 @@ func isScalar(t llvm.Type) bool {
 		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.
+			operand := v.Operand(0)
+			if operand.IsConstant() {
+				return operand.ZExtValue() == 0, true
+			}
+		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
+	}
+	return false, false // not valid
+}

interp/values.go

@@ -162,7 +162,7 @@ func (v *MapValue) newBucket() llvm.Value {
 		llvm.ArrayType(v.KeyType, 8),      // key type
 		llvm.ArrayType(v.ValueType, 8),    // value type
 	}, false)
-	bucketValue := getZeroValue(bucketType)
+	bucketValue := llvm.ConstNull(bucketType)
 	bucket := llvm.AddGlobal(v.Eval.Mod, bucketType, v.PkgName+"$mapbucket")
 	bucket.SetInitializer(bucketValue)
 	bucket.SetLinkage(llvm.InternalLinkage)
@@ -311,7 +311,7 @@ func (v *MapValue) PutString(keyBuf, keyLen, valPtr *LocalValue) {
 
 	keyType := v.Eval.Mod.GetTypeByName("runtime._string")
 	v.KeyType = keyType
-	key := getZeroValue(keyType)
+	key := llvm.ConstNull(keyType)
 	key = llvm.ConstInsertValue(key, keyBuf.Value(), []uint32{0})
 	key = llvm.ConstInsertValue(key, keyLen.Value(), []uint32{1})
 

ir/ir.go

@@ -2,7 +2,6 @@ package ir
 
 import (
 	"go/ast"
-	"go/token"
 	"go/types"
 	"sort"
 	"strings"
@@ -23,43 +22,19 @@ type Program struct {
 	mainPkg       *ssa.Package
 	Functions     []*Function
 	functionMap   map[*ssa.Function]*Function
-	Globals       []*Global
-	globalMap     map[*ssa.Global]*Global
-	comments      map[string]*ast.CommentGroup
-	NamedTypes    []*NamedType
 }
 
 // Function or method.
 type Function struct {
 	*ssa.Function
 	LLVMFn    llvm.Value
-	linkName  string // go:linkname, go:export, go:interrupt
-	exported  bool   // go:export
-	nobounds  bool   // go:nobounds
-	flag      bool   // used by dead code elimination
-	interrupt bool   // go:interrupt
-}
-
-// Global variable, possibly constant.
-type Global struct {
-	*ssa.Global
-	program    *Program
-	LLVMGlobal llvm.Value
-	linkName   string // go:extern
-	extern     bool   // go:extern
-}
-
-// Type with a name and possibly methods.
-type NamedType struct {
-	*ssa.Type
-	LLVMType llvm.Type
-}
-
-// Type that is at some point put in an interface.
-type TypeWithMethods struct {
-	t       types.Type
-	Num     int
-	Methods map[string]*types.Selection
+	module    string     // go:wasm-module
+	linkName  string     // go:linkname, go:export, go:interrupt
+	exported  bool       // go:export
+	nobounds  bool       // go:nobounds
+	flag      bool       // used by dead code elimination
+	interrupt bool       // go:interrupt
+	inline    InlineType // go:inline
 }
 
 // Interface type that is at some point used in a type assert (to check whether
@@ -69,37 +44,27 @@ type Interface struct {
 	Type *types.Interface
 }
 
-// Create and intialize a new *Program from a *ssa.Program.
-func NewProgram(lprogram *loader.Program, mainPath string) *Program {
-	comments := 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.TYPE, token.VAR:
-						if len(decl.Specs) != 1 {
-							continue
-						}
-						for _, spec := range decl.Specs {
-							switch spec := spec.(type) {
-							case *ast.TypeSpec: // decl.Tok == token.TYPE
-								id := pkgInfo.Pkg.Path() + "." + spec.Name.Name
-								comments[id] = decl.Doc
-							case *ast.ValueSpec: // decl.Tok == token.VAR
-								for _, name := range spec.Names {
-									id := pkgInfo.Pkg.Path() + "." + name.Name
-									comments[id] = decl.Doc
-								}
-							}
-						}
-					}
-				}
-			}
-		}
-	}
+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()
 
@@ -171,8 +136,6 @@ func NewProgram(lprogram *loader.Program, mainPath string) *Program {
 		LoaderProgram: lprogram,
 		mainPkg:       mainPkg,
 		functionMap:   make(map[*ssa.Function]*Function),
-		globalMap:     make(map[*ssa.Global]*Global),
-		comments:      comments,
 	}
 
 	for _, pkg := range packageList {
@@ -197,8 +160,6 @@ func (p *Program) AddPackage(pkg *ssa.Package) {
 		case *ssa.Function:
 			p.addFunction(member)
 		case *ssa.Type:
-			t := &NamedType{Type: member}
-			p.NamedTypes = append(p.NamedTypes, t)
 			methods := getAllMethods(pkg.Prog, member.Type())
 			if !types.IsInterface(member.Type()) {
 				// named type
@@ -207,13 +168,7 @@ func (p *Program) AddPackage(pkg *ssa.Package) {
 				}
 			}
 		case *ssa.Global:
-			g := &Global{program: p, Global: member}
-			doc := p.comments[g.RelString(nil)]
-			if doc != nil {
-				g.parsePragmas(doc)
-			}
-			p.Globals = append(p.Globals, g)
-			p.globalMap[member] = g
+			// Ignore. Globals are not handled here.
 		case *ssa.NamedConst:
 			// Ignore: these are already resolved.
 		default:
@@ -247,10 +202,6 @@ func (p *Program) GetFunction(ssaFn *ssa.Function) *Function {
 	return p.functionMap[ssaFn]
 }
 
-func (p *Program) GetGlobal(ssaGlobal *ssa.Global) *Global {
-	return p.globalMap[ssaGlobal]
-}
-
 func (p *Program) MainPkg() *ssa.Package {
 	return p.mainPkg
 }
@@ -279,6 +230,16 @@ func (f *Function) parsePragmas() {
 				}
 				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:interrupt":
 				if len(parts) != 2 {
 					continue
@@ -332,6 +293,16 @@ func (f *Function) IsInterrupt() bool {
 	return f.interrupt
 }
 
+// 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 != "" {
@@ -366,72 +337,6 @@ func (f *Function) CName() string {
 	return ""
 }
 
-// Parse //go: pragma comments from the source.
-func (g *Global) parsePragmas(doc *ast.CommentGroup) {
-	for _, comment := range doc.List {
-		if !strings.HasPrefix(comment.Text, "//go:") {
-			continue
-		}
-		parts := strings.Fields(comment.Text)
-		switch parts[0] {
-		case "//go:extern":
-			g.extern = true
-			if len(parts) == 2 {
-				g.linkName = parts[1]
-			}
-		}
-	}
-}
-
-// Return the link name for this global.
-func (g *Global) LinkName() string {
-	if g.linkName != "" {
-		return g.linkName
-	}
-	if name := g.CName(); name != "" {
-		return name
-	}
-	return g.RelString(nil)
-}
-
-func (g *Global) IsExtern() bool {
-	return g.extern || g.CName() != ""
-}
-
-// Return the name of the C global if this is a CGo wrapper. Otherwise, return a
-// zero-length string.
-func (g *Global) CName() string {
-	name := g.Name()
-	if strings.HasPrefix(name, "C.") {
-		// created by ../loader/cgo.go
-		return name[2:]
-	}
-	return ""
-}
-
-// Return true if this named type is annotated with the //go:volatile pragma,
-// for volatile loads and stores.
-func (p *Program) IsVolatile(t types.Type) bool {
-	if t, ok := t.(*types.Named); !ok {
-		return false
-	} else {
-		if t.Obj().Pkg() == nil {
-			return false
-		}
-		id := t.Obj().Pkg().Path() + "." + t.Obj().Name()
-		doc := p.comments[id]
-		if doc == nil {
-			return false
-		}
-		for _, line := range doc.List {
-			if strings.TrimSpace(line.Text) == "//go:volatile" {
-				return true
-			}
-		}
-		return false
-	}
-}
-
 // Get all methods of a type.
 func getAllMethods(prog *ssa.Program, typ types.Type) []*types.Selection {
 	ms := prog.MethodSets.MethodSet(typ)

linker-builtin.go

@@ -22,9 +22,9 @@ import "C"
 // Link invokes a linker with the given name and flags.
 //
 // This version uses the built-in linker when trying to use lld.
-func Link(dir, linker string, flags ...string) error {
+func Link(linker string, flags ...string) error {
 	switch linker {
-	case "ld.lld", commands["ld.lld"]:
+	case "ld.lld":
 		flags = append([]string{"tinygo:" + linker}, flags...)
 		var cflag *C.char
 		buf := C.calloc(C.size_t(len(flags)), C.size_t(unsafe.Sizeof(cflag)))
@@ -39,7 +39,7 @@ func Link(dir, linker string, flags ...string) error {
 			return errors.New("failed to link using built-in ld.lld")
 		}
 		return nil
-	case "wasm-ld", commands["wasm-ld"]:
+	case "wasm-ld":
 		flags = append([]string{"tinygo:" + linker}, flags...)
 		var cflag *C.char
 		buf := C.calloc(C.size_t(len(flags)), C.size_t(unsafe.Sizeof(cflag)))
@@ -57,10 +57,13 @@ func Link(dir, linker string, flags ...string) error {
 		return nil
 	default:
 		// Fall back to external command.
+		if cmdNames, ok := commands[linker]; ok {
+			return execCommand(cmdNames, flags...)
+		}
 		cmd := exec.Command(linker, flags...)
 		cmd.Stdout = os.Stdout
 		cmd.Stderr = os.Stderr
-		cmd.Dir = dir
+		cmd.Dir = sourceDir()
 		return cmd.Run()
 	}
 }

linker-external.go

@@ -13,10 +13,13 @@ import (
 // Link invokes a linker with the given name and arguments.
 //
 // This version always runs the linker as an external command.
-func Link(dir, linker string, flags ...string) error {
+func Link(linker string, flags ...string) error {
+	if cmdNames, ok := commands[linker]; ok {
+		return execCommand(cmdNames, flags...)
+	}
 	cmd := exec.Command(linker, flags...)
 	cmd.Stdout = os.Stdout
 	cmd.Stderr = os.Stderr
-	cmd.Dir = dir
+	cmd.Dir = sourceDir()
 	return cmd.Run()
 }

loader/cgo.go

@@ -1,414 +0,0 @@
-package loader
-
-// This file extracts the `import "C"` statement from the source and modifies
-// the AST for Cgo. It does not use libclang directly (see libclang.go).
-
-import (
-	"go/ast"
-	"go/token"
-	"sort"
-	"strconv"
-	"strings"
-
-	"golang.org/x/tools/go/ast/astutil"
-)
-
-// fileInfo holds all Cgo-related information of a given *ast.File.
-type fileInfo struct {
-	*ast.File
-	filename   string
-	functions  map[string]*functionInfo
-	globals    map[string]*globalInfo
-	typedefs   map[string]*typedefInfo
-	importCPos token.Pos
-}
-
-// functionInfo stores some information about a Cgo function found by libclang
-// and declared in the AST.
-type functionInfo struct {
-	args    []paramInfo
-	results *ast.FieldList
-}
-
-// paramInfo is a parameter of a Cgo function (see functionInfo).
-type paramInfo struct {
-	name     string
-	typeExpr ast.Expr
-}
-
-// typedefInfo contains information about a single typedef in C.
-type typedefInfo struct {
-	typeExpr ast.Expr
-}
-
-// globalInfo contains information about a declared global variable in C.
-type globalInfo struct {
-	typeExpr ast.Expr
-}
-
-// cgoAliases list type aliases between Go and C, for types that are equivalent
-// in both languages. See addTypeAliases.
-var cgoAliases = map[string]string{
-	"C.int8_t":    "int8",
-	"C.int16_t":   "int16",
-	"C.int32_t":   "int32",
-	"C.int64_t":   "int64",
-	"C.uint8_t":   "uint8",
-	"C.uint16_t":  "uint16",
-	"C.uint32_t":  "uint32",
-	"C.uint64_t":  "uint64",
-	"C.uintptr_t": "uintptr",
-}
-
-// cgoTypes lists some C types with ambiguous sizes that must be retrieved
-// somehow from C. This is done by adding some typedefs to get the size of each
-// type.
-const cgoTypes = `
-typedef signed char         _Cgo_schar;
-typedef unsigned char       _Cgo_uchar;
-typedef short               _Cgo_short;
-typedef unsigned short      _Cgo_ushort;
-typedef int                 _Cgo_int;
-typedef unsigned int        _Cgo_uint;
-typedef long                _Cgo_long;
-typedef unsigned long       _Cgo_ulong;
-typedef long long           _Cgo_longlong;
-typedef unsigned long long  _Cgo_ulonglong;
-`
-
-// processCgo extracts the `import "C"` statement from the AST, parses the
-// comment with libclang, and modifies the AST to use this information.
-func (p *Package) processCgo(filename string, f *ast.File, cflags []string) error {
-	info := &fileInfo{
-		File:      f,
-		filename:  filename,
-		functions: map[string]*functionInfo{},
-		globals:   map[string]*globalInfo{},
-		typedefs:  map[string]*typedefInfo{},
-	}
-
-	// Find `import "C"` statements in the file.
-	for i := 0; i < len(f.Decls); i++ {
-		decl := f.Decls[i]
-		genDecl, ok := decl.(*ast.GenDecl)
-		if !ok {
-			continue
-		}
-		if len(genDecl.Specs) != 1 {
-			continue
-		}
-		spec, ok := genDecl.Specs[0].(*ast.ImportSpec)
-		if !ok {
-			continue
-		}
-		path, err := strconv.Unquote(spec.Path.Value)
-		if err != nil {
-			panic("could not parse import path: " + err.Error())
-		}
-		if path != "C" {
-			continue
-		}
-		cgoComment := genDecl.Doc.Text()
-
-		// Stored for later use by generated functions, to use a somewhat sane
-		// source location.
-		info.importCPos = spec.Path.ValuePos
-
-		err = info.parseFragment(cgoComment+cgoTypes, cflags)
-		if err != nil {
-			return err
-		}
-
-		// Remove this import declaration.
-		f.Decls = append(f.Decls[:i], f.Decls[i+1:]...)
-		i--
-	}
-
-	// Print the AST, for debugging.
-	//ast.Print(p.fset, f)
-
-	// Declare functions found by libclang.
-	info.addFuncDecls()
-
-	// Declare stub function pointer values found by libclang.
-	info.addFuncPtrDecls()
-
-	// Declare globals found by libclang.
-	info.addVarDecls()
-
-	// Forward C types to Go types (like C.uint32_t -> uint32).
-	info.addTypeAliases()
-
-	// Add type declarations for C types, declared using typeef in C.
-	info.addTypedefs()
-
-	// Patch the AST to use the declared types and functions.
-	f = astutil.Apply(f, info.walker, nil).(*ast.File)
-
-	return nil
-}
-
-// addFuncDecls adds the C function declarations found by libclang in the
-// comment above the `import "C"` statement.
-func (info *fileInfo) addFuncDecls() {
-	// TODO: replace all uses of importCPos with the real locations from
-	// libclang.
-	names := make([]string, 0, len(info.functions))
-	for name := range info.functions {
-		names = append(names, name)
-	}
-	sort.Strings(names)
-	for _, name := range names {
-		fn := info.functions[name]
-		obj := &ast.Object{
-			Kind: ast.Fun,
-			Name: "C." + name,
-		}
-		args := make([]*ast.Field, len(fn.args))
-		decl := &ast.FuncDecl{
-			Name: &ast.Ident{
-				NamePos: info.importCPos,
-				Name:    "C." + name,
-				Obj:     obj,
-			},
-			Type: &ast.FuncType{
-				Func: info.importCPos,
-				Params: &ast.FieldList{
-					Opening: info.importCPos,
-					List:    args,
-					Closing: info.importCPos,
-				},
-				Results: fn.results,
-			},
-		}
-		obj.Decl = decl
-		for i, arg := range fn.args {
-			args[i] = &ast.Field{
-				Names: []*ast.Ident{
-					&ast.Ident{
-						NamePos: info.importCPos,
-						Name:    arg.name,
-						Obj: &ast.Object{
-							Kind: ast.Var,
-							Name: arg.name,
-							Decl: decl,
-						},
-					},
-				},
-				Type: arg.typeExpr,
-			}
-		}
-		info.Decls = append(info.Decls, decl)
-	}
-}
-
-// addFuncPtrDecls creates stub declarations of function pointer values. These
-// values will later be replaced with the real values in the compiler.
-// It adds code like the following to the AST:
-//
-//     var (
-//         C.add unsafe.Pointer
-//         C.mul unsafe.Pointer
-//         // ...
-//     )
-func (info *fileInfo) addFuncPtrDecls() {
-	gen := &ast.GenDecl{
-		TokPos: info.importCPos,
-		Tok:    token.VAR,
-		Lparen: info.importCPos,
-		Rparen: info.importCPos,
-	}
-	names := make([]string, 0, len(info.functions))
-	for name := range info.functions {
-		names = append(names, name)
-	}
-	sort.Strings(names)
-	for _, name := range names {
-		obj := &ast.Object{
-			Kind: ast.Typ,
-			Name: "C." + name + "$funcaddr",
-		}
-		valueSpec := &ast.ValueSpec{
-			Names: []*ast.Ident{&ast.Ident{
-				NamePos: info.importCPos,
-				Name:    "C." + name + "$funcaddr",
-				Obj:     obj,
-			}},
-			Type: &ast.SelectorExpr{
-				X: &ast.Ident{
-					NamePos: info.importCPos,
-					Name:    "unsafe",
-				},
-				Sel: &ast.Ident{
-					NamePos: info.importCPos,
-					Name:    "Pointer",
-				},
-			},
-		}
-		obj.Decl = valueSpec
-		gen.Specs = append(gen.Specs, valueSpec)
-	}
-	info.Decls = append(info.Decls, gen)
-}
-
-// addVarDecls declares external C globals in the Go source.
-// It adds code like the following to the AST:
-//
-//     var (
-//         C.globalInt  int
-//         C.globalBool bool
-//         // ...
-//     )
-func (info *fileInfo) addVarDecls() {
-	gen := &ast.GenDecl{
-		TokPos: info.importCPos,
-		Tok:    token.VAR,
-		Lparen: info.importCPos,
-		Rparen: info.importCPos,
-	}
-	names := make([]string, 0, len(info.globals))
-	for name := range info.globals {
-		names = append(names, name)
-	}
-	sort.Strings(names)
-	for _, name := range names {
-		global := info.globals[name]
-		obj := &ast.Object{
-			Kind: ast.Typ,
-			Name: "C." + name,
-		}
-		valueSpec := &ast.ValueSpec{
-			Names: []*ast.Ident{&ast.Ident{
-				NamePos: info.importCPos,
-				Name:    "C." + name,
-				Obj:     obj,
-			}},
-			Type: global.typeExpr,
-		}
-		obj.Decl = valueSpec
-		gen.Specs = append(gen.Specs, valueSpec)
-	}
-	info.Decls = append(info.Decls, gen)
-}
-
-// addTypeAliases aliases some built-in Go types with their equivalent C types.
-// It adds code like the following to the AST:
-//
-//     type (
-//         C.int8_t  = int8
-//         C.int16_t = int16
-//         // ...
-//     )
-func (info *fileInfo) addTypeAliases() {
-	aliasKeys := make([]string, 0, len(cgoAliases))
-	for key := range cgoAliases {
-		aliasKeys = append(aliasKeys, key)
-	}
-	sort.Strings(aliasKeys)
-	gen := &ast.GenDecl{
-		TokPos: info.importCPos,
-		Tok:    token.TYPE,
-		Lparen: info.importCPos,
-		Rparen: info.importCPos,
-	}
-	for _, typeName := range aliasKeys {
-		goTypeName := cgoAliases[typeName]
-		obj := &ast.Object{
-			Kind: ast.Typ,
-			Name: typeName,
-		}
-		typeSpec := &ast.TypeSpec{
-			Name: &ast.Ident{
-				NamePos: info.importCPos,
-				Name:    typeName,
-				Obj:     obj,
-			},
-			Assign: info.importCPos,
-			Type: &ast.Ident{
-				NamePos: info.importCPos,
-				Name:    goTypeName,
-			},
-		}
-		obj.Decl = typeSpec
-		gen.Specs = append(gen.Specs, typeSpec)
-	}
-	info.Decls = append(info.Decls, gen)
-}
-
-func (info *fileInfo) addTypedefs() {
-	gen := &ast.GenDecl{
-		TokPos: info.importCPos,
-		Tok:    token.TYPE,
-	}
-	names := make([]string, 0, len(info.typedefs))
-	for name := range info.typedefs {
-		names = append(names, name)
-	}
-	sort.Strings(names)
-	for _, name := range names {
-		typedef := info.typedefs[name]
-		typeName := "C." + name
-		if strings.HasPrefix(name, "_Cgo_") {
-			typeName = "C." + name[len("_Cgo_"):]
-		}
-		if _, ok := cgoAliases[typeName]; ok {
-			// This is a type that also exists in Go (defined in stdint.h).
-			continue
-		}
-		obj := &ast.Object{
-			Kind: ast.Typ,
-			Name: typeName,
-		}
-		typeSpec := &ast.TypeSpec{
-			Name: &ast.Ident{
-				NamePos: info.importCPos,
-				Name:    typeName,
-				Obj:     obj,
-			},
-			Type: typedef.typeExpr,
-		}
-		obj.Decl = typeSpec
-		gen.Specs = append(gen.Specs, typeSpec)
-	}
-	info.Decls = append(info.Decls, gen)
-}
-
-// walker replaces all "C".<something> expressions to literal "C.<something>"
-// expressions. Such expressions are impossible to write in Go (a dot cannot be
-// used in the middle of a name) so in practice all C identifiers live in a
-// separate namespace (no _Cgo_ hacks like in gc).
-func (info *fileInfo) walker(cursor *astutil.Cursor) bool {
-	switch node := cursor.Node().(type) {
-	case *ast.CallExpr:
-		fun, ok := node.Fun.(*ast.SelectorExpr)
-		if !ok {
-			return true
-		}
-		x, ok := fun.X.(*ast.Ident)
-		if !ok {
-			return true
-		}
-		if _, ok := info.functions[fun.Sel.Name]; ok && x.Name == "C" {
-			node.Fun = &ast.Ident{
-				NamePos: x.NamePos,
-				Name:    "C." + fun.Sel.Name,
-			}
-		}
-	case *ast.SelectorExpr:
-		x, ok := node.X.(*ast.Ident)
-		if !ok {
-			return true
-		}
-		if x.Name == "C" {
-			name := "C." + node.Sel.Name
-			if _, ok := info.functions[node.Sel.Name]; ok {
-				name += "$funcaddr"
-			}
-			cursor.Replace(&ast.Ident{
-				NamePos: x.NamePos,
-				Name:    name,
-			})
-		}
-	}
-	return true
-}

loader/libclang.go

@@ -1,250 +0,0 @@
-package loader
-
-// This file parses a fragment of C with libclang and stores the result for AST
-// modification. It does not touch the AST itself.
-
-import (
-	"errors"
-	"go/ast"
-	"go/token"
-	"strconv"
-	"strings"
-	"unsafe"
-)
-
-/*
-#include <clang-c/Index.h> // if this fails, install libclang-7-dev
-#include <stdlib.h>
-
-int tinygo_clang_visitor(CXCursor c, CXCursor parent, CXClientData client_data);
-*/
-import "C"
-
-var globalFileInfo *fileInfo
-
-func (info *fileInfo) parseFragment(fragment string, cflags []string) error {
-	index := C.clang_createIndex(0, 1)
-	defer C.clang_disposeIndex(index)
-
-	filenameC := C.CString("cgo-fake.c")
-	defer C.free(unsafe.Pointer(filenameC))
-
-	fragmentC := C.CString(fragment)
-	defer C.free(unsafe.Pointer(fragmentC))
-
-	unsavedFile := C.struct_CXUnsavedFile{
-		Filename: filenameC,
-		Length:   C.ulong(len(fragment)),
-		Contents: fragmentC,
-	}
-
-	// convert Go slice of strings to C array of strings.
-	cmdargsC := C.malloc(C.size_t(len(cflags)) * C.size_t(unsafe.Sizeof(uintptr(0))))
-	defer C.free(cmdargsC)
-	cmdargs := (*[1 << 16]*C.char)(cmdargsC)
-	for i, cflag := range cflags {
-		s := C.CString(cflag)
-		cmdargs[i] = s
-		defer C.free(unsafe.Pointer(s))
-	}
-
-	var unit C.CXTranslationUnit
-	errCode := C.clang_parseTranslationUnit2(
-		index,
-		filenameC,
-		(**C.char)(cmdargsC), C.int(len(cflags)), // command line args
-		&unsavedFile, 1, // unsaved files
-		C.CXTranslationUnit_None,
-		&unit)
-	if errCode != 0 {
-		panic("loader: failed to parse source with libclang")
-	}
-	defer C.clang_disposeTranslationUnit(unit)
-
-	if C.clang_getNumDiagnostics(unit) != 0 {
-		return errors.New("cgo: libclang cannot parse fragment")
-	}
-
-	if globalFileInfo != nil {
-		// There is a race condition here but that doesn't really matter as it
-		// is a sanity check anyway.
-		panic("libclang.go cannot be used concurrently yet")
-	}
-	globalFileInfo = info
-	defer func() {
-		globalFileInfo = nil
-	}()
-
-	cursor := C.clang_getTranslationUnitCursor(unit)
-	C.clang_visitChildren(cursor, (*[0]byte)(unsafe.Pointer(C.tinygo_clang_visitor)), C.CXClientData(uintptr(0)))
-
-	return nil
-}
-
-//export tinygo_clang_visitor
-func tinygo_clang_visitor(c, parent C.CXCursor, client_data C.CXClientData) C.int {
-	info := globalFileInfo
-	kind := C.clang_getCursorKind(c)
-	switch kind {
-	case C.CXCursor_FunctionDecl:
-		name := getString(C.clang_getCursorSpelling(c))
-		cursorType := C.clang_getCursorType(c)
-		if C.clang_isFunctionTypeVariadic(cursorType) != 0 {
-			return C.CXChildVisit_Continue // not supported
-		}
-		numArgs := int(C.clang_Cursor_getNumArguments(c))
-		fn := &functionInfo{}
-		info.functions[name] = fn
-		for i := 0; i < numArgs; i++ {
-			arg := C.clang_Cursor_getArgument(c, C.uint(i))
-			argName := getString(C.clang_getCursorSpelling(arg))
-			argType := C.clang_getArgType(cursorType, C.uint(i))
-			if argName == "" {
-				argName = "$" + strconv.Itoa(i)
-			}
-			fn.args = append(fn.args, paramInfo{
-				name:     argName,
-				typeExpr: info.makeASTType(argType),
-			})
-		}
-		resultType := C.clang_getCursorResultType(c)
-		if resultType.kind != C.CXType_Void {
-			fn.results = &ast.FieldList{
-				List: []*ast.Field{
-					&ast.Field{
-						Type: info.makeASTType(resultType),
-					},
-				},
-			}
-		}
-	case C.CXCursor_TypedefDecl:
-		typedefType := C.clang_getCursorType(c)
-		name := getString(C.clang_getTypedefName(typedefType))
-		underlyingType := C.clang_getTypedefDeclUnderlyingType(c)
-		expr := info.makeASTType(underlyingType)
-		if strings.HasPrefix(name, "_Cgo_") {
-			expr := expr.(*ast.Ident)
-			typeSize := C.clang_Type_getSizeOf(underlyingType)
-			switch expr.Name {
-			// TODO: plain char (may be signed or unsigned)
-			case "C.schar", "C.short", "C.int", "C.long", "C.longlong":
-				switch typeSize {
-				case 1:
-					expr.Name = "int8"
-				case 2:
-					expr.Name = "int16"
-				case 4:
-					expr.Name = "int32"
-				case 8:
-					expr.Name = "int64"
-				}
-			case "C.uchar", "C.ushort", "C.uint", "C.ulong", "C.ulonglong":
-				switch typeSize {
-				case 1:
-					expr.Name = "uint8"
-				case 2:
-					expr.Name = "uint16"
-				case 4:
-					expr.Name = "uint32"
-				case 8:
-					expr.Name = "uint64"
-				}
-			}
-		}
-		info.typedefs[name] = &typedefInfo{
-			typeExpr: expr,
-		}
-	case C.CXCursor_VarDecl:
-		name := getString(C.clang_getCursorSpelling(c))
-		cursorType := C.clang_getCursorType(c)
-		info.globals[name] = &globalInfo{
-			typeExpr: info.makeASTType(cursorType),
-		}
-	}
-	return C.CXChildVisit_Continue
-}
-
-func getString(clangString C.CXString) (s string) {
-	rawString := C.clang_getCString(clangString)
-	s = C.GoString(rawString)
-	C.clang_disposeString(clangString)
-	return
-}
-
-// makeASTType return the ast.Expr for the given libclang type. In other words,
-// it converts a libclang type to a type in the Go AST.
-func (info *fileInfo) makeASTType(typ C.CXType) ast.Expr {
-	var typeName string
-	switch typ.kind {
-	case C.CXType_SChar:
-		typeName = "C.schar"
-	case C.CXType_UChar:
-		typeName = "C.uchar"
-	case C.CXType_Short:
-		typeName = "C.short"
-	case C.CXType_UShort:
-		typeName = "C.ushort"
-	case C.CXType_Int:
-		typeName = "C.int"
-	case C.CXType_UInt:
-		typeName = "C.uint"
-	case C.CXType_Long:
-		typeName = "C.long"
-	case C.CXType_ULong:
-		typeName = "C.ulong"
-	case C.CXType_LongLong:
-		typeName = "C.longlong"
-	case C.CXType_ULongLong:
-		typeName = "C.ulonglong"
-	case C.CXType_Bool:
-		typeName = "bool"
-	case C.CXType_Float, C.CXType_Double, C.CXType_LongDouble:
-		switch C.clang_Type_getSizeOf(typ) {
-		case 4:
-			typeName = "float32"
-		case 8:
-			typeName = "float64"
-		default:
-			// Don't do anything, rely on the fallback code to show a somewhat
-			// sensible error message like "undeclared name: C.long double".
-		}
-	case C.CXType_Complex:
-		switch C.clang_Type_getSizeOf(typ) {
-		case 8:
-			typeName = "complex64"
-		case 16:
-			typeName = "complex128"
-		}
-	case C.CXType_Pointer:
-		return &ast.StarExpr{
-			Star: info.importCPos,
-			X:    info.makeASTType(C.clang_getPointeeType(typ)),
-		}
-	case C.CXType_FunctionProto:
-		// Be compatible with gc, which uses the *[0]byte type for function
-		// pointer types.
-		// Return type [0]byte because this is a function type, not a pointer to
-		// this function type.
-		return &ast.ArrayType{
-			Lbrack: info.importCPos,
-			Len: &ast.BasicLit{
-				ValuePos: info.importCPos,
-				Kind:     token.INT,
-				Value:    "0",
-			},
-			Elt: &ast.Ident{
-				NamePos: info.importCPos,
-				Name:    "byte",
-			},
-		}
-	}
-	if typeName == "" {
-		// Fallback, probably incorrect but at least the error points to an odd
-		// type name.
-		typeName = "C." + getString(C.clang_getTypeSpelling(typ))
-	}
-	return &ast.Ident{
-		NamePos: info.importCPos,
-		Name:    typeName,
-	}
-}

loader/loader.go

@@ -1,6 +1,7 @@
 package loader
 
 import (
+	"bytes"
 	"errors"
 	"go/ast"
 	"go/build"
@@ -10,17 +11,26 @@ import (
 	"os"
 	"path/filepath"
 	"sort"
+	"strings"
+	"text/template"
+
+	"github.com/tinygo-org/tinygo/cgo"
 )
 
 // Program holds all packages and some metadata about the program as a whole.
 type Program struct {
-	Build       *build.Context
-	Packages    map[string]*Package
-	sorted      []*Package
-	fset        *token.FileSet
-	TypeChecker types.Config
-	Dir         string // current working directory (for error reporting)
-	CFlags      []string
+	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
 }
 
 // Package holds a loaded package, its imports, and its parsed files.
@@ -42,7 +52,12 @@ func (p *Program) Import(path, srcDir string) (*Package, error) {
 	}
 
 	// Load this package.
-	buildPkg, err := p.Build.Import(path, srcDir, build.ImportComment)
+	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, err
 	}
@@ -53,6 +68,11 @@ func (p *Program) Import(path, srcDir string) (*Package, error) {
 	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
 }
 
@@ -82,6 +102,11 @@ func (p *Program) ImportFile(path string) (*Package, error) {
 	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
 }
 
@@ -160,10 +185,12 @@ func (p *Program) sort() {
 // The returned error may be an Errors error, which contains a list of errors.
 //
 // Idempotent.
-func (p *Program) Parse() error {
+func (p *Program) Parse(compileTestBinary bool) error {
+	includeTests := compileTestBinary
+
 	// Load all imports
 	for _, pkg := range p.Sorted() {
-		err := pkg.importRecursively()
+		err := pkg.importRecursively(includeTests)
 		if err != nil {
 			if err, ok := err.(*ImportCycleError); ok {
 				if pkg.ImportPath != err.Packages[0] {
@@ -176,7 +203,14 @@ func (p *Program) Parse() error {
 
 	// Parse all packages.
 	for _, pkg := range p.Sorted() {
-		err := pkg.Parse()
+		err := pkg.Parse(includeTests)
+		if err != nil {
+			return err
+		}
+	}
+
+	if compileTestBinary {
+		err := p.SwapTestMain()
 		if err != nil {
 			return err
 		}
@@ -193,6 +227,83 @@ func (p *Program) Parse() 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 {
@@ -217,7 +328,7 @@ func (p *Program) parseFile(path string, mode parser.Mode) (*ast.File, error) {
 // Parse parses and typechecks this package.
 //
 // Idempotent.
-func (p *Package) Parse() error {
+func (p *Package) Parse(includeTests bool) error {
 	if len(p.Files) != 0 {
 		return nil
 	}
@@ -231,7 +342,7 @@ func (p *Package) Parse() error {
 		return nil
 	}
 
-	files, err := p.parseFiles()
+	files, err := p.parseFiles(includeTests)
 	if err != nil {
 		return err
 	}
@@ -270,11 +381,21 @@ func (p *Package) Check() error {
 }
 
 // parseFiles parses the loaded list of files and returns this list.
-func (p *Package) parseFiles() ([]*ast.File, error) {
+func (p *Package) parseFiles(includeTests bool) ([]*ast.File, error) {
 	// TODO: do this concurrently.
 	var files []*ast.File
 	var fileErrs []error
-	for _, file := range p.GoFiles {
+
+	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)
@@ -293,16 +414,23 @@ func (p *Package) parseFiles() ([]*ast.File, error) {
 			fileErrs = append(fileErrs, err)
 			continue
 		}
-		err = p.processCgo(path, f, append(p.CFlags, "-I"+p.Package.Dir))
-		if err != nil {
-			fileErrs = append(fileErrs, err)
-			continue
-		}
 		files = append(files, f)
 	}
+	if len(p.CgoFiles) != 0 {
+		cflags := append(p.CFlags, "-I"+p.Package.Dir)
+		if p.ClangHeaders != "" {
+			cflags = append(cflags, "-I"+p.ClangHeaders)
+		}
+		generated, errs := cgo.Process(files, p.Program.Dir, p.fset, cflags)
+		if errs != nil {
+			fileErrs = append(fileErrs, errs...)
+		}
+		files = append(files, generated)
+	}
 	if len(fileErrs) != 0 {
 		return nil, Errors{p, fileErrs}
 	}
+
 	return files, nil
 }
 
@@ -323,11 +451,17 @@ func (p *Package) Import(to string) (*types.Package, error) {
 // importRecursively() on the imported packages as well.
 //
 // Idempotent.
-func (p *Package) importRecursively() error {
+func (p *Package) importRecursively(includeTests bool) error {
 	p.Importing = true
-	for _, to := range p.Package.Imports {
+
+	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.
+			// Do CGo processing in a later stage.
 			continue
 		}
 		if _, ok := p.Imports[to]; ok {
@@ -343,7 +477,7 @@ func (p *Package) importRecursively() error {
 		if importedPkg.Importing {
 			return &ImportCycleError{[]string{p.ImportPath, importedPkg.ImportPath}, p.ImportPos[to]}
 		}
-		err = importedPkg.importRecursively()
+		err = importedPkg.importRecursively(false)
 		if err != nil {
 			if err, ok := err.(*ImportCycleError); ok {
 				err.Packages = append([]string{p.ImportPath}, err.Packages...)

main.go

@@ -33,16 +33,32 @@ func (e *commandError) Error() string {
 	return e.Msg + " " + e.File + ": " + e.Err.Error()
 }
 
+// multiError is a list of multiple errors (actually: diagnostics) returned
+// during LLVM IR generation.
+type multiError struct {
+	Errs []error
+}
+
+func (e *multiError) Error() string {
+	return e.Errs[0].Error()
+}
+
 type BuildConfig struct {
-	opt        string
-	gc         string
-	printIR    bool
-	dumpSSA    bool
-	debug      bool
-	printSizes string
-	cFlags     []string
-	ldFlags    []string
-	wasmAbi    string
+	opt           string
+	gc            string
+	panicStrategy string
+	scheduler     string
+	printIR       bool
+	dumpSSA       bool
+	verifyIR      bool
+	debug         bool
+	printSizes    string
+	cFlags        []string
+	ldFlags       []string
+	tags          string
+	wasmAbi       string
+	heapSize      int64
+	testConfig    compiler.TestConfig
 }
 
 // Helper function for Compiler object.
@@ -51,23 +67,62 @@ func Compile(pkgName, outpath string, spec *TargetSpec, config *BuildConfig, act
 		config.gc = spec.GC
 	}
 
-	// Append command line passed CFlags and LDFlags
-	spec.CFlags = append(spec.CFlags, config.cFlags...)
-	spec.LDFlags = append(spec.LDFlags, config.ldFlags...)
+	root := sourceDir()
 
+	// Merge and adjust CFlags.
+	cflags := append([]string{}, config.cFlags...)
+	for _, flag := range spec.CFlags {
+		cflags = append(cflags, strings.Replace(flag, "{root}", root, -1))
+	}
+
+	// Merge and adjust LDFlags.
+	ldflags := append([]string{}, config.ldFlags...)
+	for _, flag := range spec.LDFlags {
+		ldflags = append(ldflags, strings.Replace(flag, "{root}", root, -1))
+	}
+
+	goroot := getGoroot()
+	if goroot == "" {
+		return errors.New("cannot locate $GOROOT, please set it manually")
+	}
+	tags := spec.BuildTags
+	major, minor, err := getGorootVersion(goroot)
+	if err != nil {
+		return fmt.Errorf("could not read version from GOROOT (%v): %v", goroot, err)
+	}
+	if major != 1 || (minor != 11 && minor != 12) {
+		return fmt.Errorf("requires go version 1.11 or 1.12, got go%d.%d", major, minor)
+	}
+	for i := 1; i <= minor; i++ {
+		tags = append(tags, fmt.Sprintf("go1.%d", i))
+	}
+	if extraTags := strings.Fields(config.tags); len(extraTags) != 0 {
+		tags = append(tags, extraTags...)
+	}
+	scheduler := spec.Scheduler
+	if config.scheduler != "" {
+		scheduler = config.scheduler
+	}
 	compilerConfig := compiler.Config{
-		Triple:    spec.Triple,
-		CPU:       spec.CPU,
-		GOOS:      spec.GOOS,
-		GOARCH:    spec.GOARCH,
-		GC:        config.gc,
-		CFlags:    spec.CFlags,
-		LDFlags:   spec.LDFlags,
-		Debug:     config.debug,
-		DumpSSA:   config.dumpSSA,
-		RootDir:   sourceDir(),
-		GOPATH:    getGopath(),
-		BuildTags: spec.BuildTags,
+		Triple:        spec.Triple,
+		CPU:           spec.CPU,
+		Features:      spec.Features,
+		GOOS:          spec.GOOS,
+		GOARCH:        spec.GOARCH,
+		GC:            config.gc,
+		PanicStrategy: config.panicStrategy,
+		Scheduler:     scheduler,
+		CFlags:        cflags,
+		LDFlags:       ldflags,
+		ClangHeaders:  getClangHeaderPath(root),
+		Debug:         config.debug,
+		DumpSSA:       config.dumpSSA,
+		VerifyIR:      config.verifyIR,
+		TINYGOROOT:    root,
+		GOROOT:        goroot,
+		GOPATH:        getGopath(),
+		BuildTags:     tags,
+		TestConfig:    config.testConfig,
 	}
 	c, err := compiler.NewCompiler(pkgName, compilerConfig)
 	if err != nil {
@@ -75,12 +130,15 @@ func Compile(pkgName, outpath string, spec *TargetSpec, config *BuildConfig, act
 	}
 
 	// Compile Go code to IR.
-	err = c.Compile(pkgName)
-	if err != nil {
-		return err
+	errs := c.Compile(pkgName)
+	if len(errs) != 0 {
+		if len(errs) == 1 {
+			return errs[0]
+		}
+		return &multiError{errs}
 	}
 	if config.printIR {
-		fmt.Println("Generated LLVM IR:")
+		fmt.Println("; Generated LLVM IR:")
 		fmt.Println(c.IR())
 	}
 	if err := c.Verify(); err != nil {
@@ -98,9 +156,6 @@ func Compile(pkgName, outpath string, spec *TargetSpec, config *BuildConfig, act
 	if spec.GOOS != "darwin" {
 		c.ApplyFunctionSections() // -ffunction-sections
 	}
-	if err := c.Verify(); err != nil {
-		return errors.New("verification error after applying function sections")
-	}
 
 	// Browsers cannot handle external functions that have type i64 because it
 	// cannot be represented exactly in JavaScript (JS only has doubles). To
@@ -112,9 +167,6 @@ func Compile(pkgName, outpath string, spec *TargetSpec, config *BuildConfig, act
 		if err != nil {
 			return err
 		}
-		if err := c.Verify(); err != nil {
-			return errors.New("verification error after running the wasm i64 hack")
-		}
 	}
 
 	// Optimization levels here are roughly the same as Clang, but probably not
@@ -191,19 +243,26 @@ func Compile(pkgName, outpath string, spec *TargetSpec, config *BuildConfig, act
 		// Prepare link command.
 		executable := filepath.Join(dir, "main")
 		tmppath := executable // final file
-		ldflags := append(spec.LDFlags, "-o", executable, objfile)
+		ldflags = append(ldflags, "-o", executable, objfile, "-L", root)
 		if spec.RTLib == "compiler-rt" {
 			ldflags = append(ldflags, librt)
 		}
+		if spec.GOARCH == "wasm" {
+			// Round heap size to next multiple of 65536 (the WebAssembly page
+			// size).
+			heapSize := (config.heapSize + (65536 - 1)) &^ (65536 - 1)
+			ldflags = append(ldflags, "--initial-memory="+strconv.FormatInt(heapSize, 10))
+		}
 
 		// Compile extra files.
 		for i, path := range spec.ExtraFiles {
+			abspath := filepath.Join(root, path)
 			outpath := filepath.Join(dir, "extra-"+strconv.Itoa(i)+"-"+filepath.Base(path)+".o")
-			cmd := exec.Command(spec.Compiler, append(spec.CFlags, "-c", "-o", outpath, path)...)
-			cmd.Stdout = os.Stdout
-			cmd.Stderr = os.Stderr
-			cmd.Dir = sourceDir()
-			err := cmd.Run()
+			cmdNames := []string{spec.Compiler}
+			if names, ok := commands[spec.Compiler]; ok {
+				cmdNames = names
+			}
+			err := execCommand(cmdNames, append(cflags, "-c", "-o", outpath, abspath)...)
 			if err != nil {
 				return &commandError{"failed to build", path, err}
 			}
@@ -215,11 +274,11 @@ func Compile(pkgName, outpath string, spec *TargetSpec, config *BuildConfig, act
 			for _, file := range pkg.CFiles {
 				path := filepath.Join(pkg.Package.Dir, file)
 				outpath := filepath.Join(dir, "pkg"+strconv.Itoa(i)+"-"+file+".o")
-				cmd := exec.Command(spec.Compiler, append(spec.CFlags, "-c", "-o", outpath, path)...)
-				cmd.Stdout = os.Stdout
-				cmd.Stderr = os.Stderr
-				cmd.Dir = sourceDir()
-				err := cmd.Run()
+				cmdNames := []string{spec.Compiler}
+				if names, ok := commands[spec.Compiler]; ok {
+					cmdNames = names
+				}
+				err := execCommand(cmdNames, append(cflags, "-c", "-o", outpath, path)...)
 				if err != nil {
 					return &commandError{"failed to build", path, err}
 				}
@@ -228,7 +287,7 @@ func Compile(pkgName, outpath string, spec *TargetSpec, config *BuildConfig, act
 		}
 
 		// Link the object files together.
-		err = Link(sourceDir(), spec.Linker, ldflags...)
+		err = Link(spec.Linker, ldflags...)
 		if err != nil {
 			return &commandError{"failed to link", executable, err}
 		}
@@ -305,6 +364,33 @@ func Build(pkgName, outpath, target string, config *BuildConfig) error {
 	})
 }
 
+func Test(pkgName, target string, config *BuildConfig) error {
+	spec, err := LoadTarget(target)
+	if err != nil {
+		return err
+	}
+
+	spec.BuildTags = append(spec.BuildTags, "test")
+	config.testConfig.CompileTestBinary = true
+	return Compile(pkgName, ".elf", spec, 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 &commandError{"failed to run compiled binary", tmppath, err}
+		}
+		return nil
+	})
+}
+
 func Flash(pkgName, target, port string, config *BuildConfig) error {
 	spec, err := LoadTarget(target)
 	if err != nil {
@@ -463,6 +549,30 @@ func Run(pkgName, target string, config *BuildConfig) error {
 	})
 }
 
+// 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
+}
+
 func usage() {
 	fmt.Fprintln(os.Stderr, "TinyGo is a Go compiler for small places.")
 	fmt.Fprintln(os.Stderr, "version:", version)
@@ -470,6 +580,7 @@ func usage() {
 	fmt.Fprintln(os.Stderr, "\ncommands:")
 	fmt.Fprintln(os.Stderr, "  build: compile packages and dependencies")
 	fmt.Fprintln(os.Stderr, "  run:   compile and run immediately")
+	fmt.Fprintln(os.Stderr, "  test:  test packages")
 	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, "  clean: empty cache directory ("+cacheDir()+")")
@@ -480,19 +591,24 @@ func usage() {
 
 func handleCompilerError(err error) {
 	if err != nil {
-		if errUnsupported, ok := err.(*interp.Unsupported); ok {
+		switch err := err.(type) {
+		case *interp.Unsupported:
 			// hit an unknown/unsupported instruction
 			fmt.Fprintln(os.Stderr, "unsupported instruction during init evaluation:")
-			errUnsupported.Inst.Dump()
+			err.Inst.Dump()
 			fmt.Fprintln(os.Stderr)
-		} else if errCompiler, ok := err.(types.Error); ok {
-			fmt.Fprintln(os.Stderr, errCompiler)
-		} else if errLoader, ok := err.(loader.Errors); ok {
-			fmt.Fprintln(os.Stderr, "#", errLoader.Pkg.ImportPath)
-			for _, err := range errLoader.Errs {
+		case types.Error:
+			fmt.Fprintln(os.Stderr, err)
+		case loader.Errors:
+			fmt.Fprintln(os.Stderr, "#", err.Pkg.ImportPath)
+			for _, err := range err.Errs {
 				fmt.Fprintln(os.Stderr, err)
 			}
-		} else {
+		case *multiError:
+			for _, err := range err.Errs {
+				fmt.Fprintln(os.Stderr, err)
+			}
+		default:
 			fmt.Fprintln(os.Stderr, "error:", err)
 		}
 		os.Exit(1)
@@ -502,10 +618,14 @@ 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, dumb, marksweep)")
+	gc := flag.String("gc", "", "garbage collector to use (none, leaking, 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")
-	target := flag.String("target", "", "LLVM target")
+	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")
@@ -513,6 +633,7 @@ func main() {
 	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.")
@@ -523,13 +644,17 @@ func main() {
 
 	flag.CommandLine.Parse(os.Args[2:])
 	config := &BuildConfig{
-		opt:        *opt,
-		gc:         *gc,
-		printIR:    *printIR,
-		dumpSSA:    *dumpSSA,
-		debug:      !*nodebug,
-		printSizes: *printSize,
-		wasmAbi:    *wasmAbi,
+		opt:           *opt,
+		gc:            *gc,
+		panicStrategy: *panicStrategy,
+		scheduler:     *scheduler,
+		printIR:       *printIR,
+		dumpSSA:       *dumpSSA,
+		verifyIR:      *verifyIR,
+		debug:         !*nodebug,
+		printSizes:    *printSize,
+		tags:          *tags,
+		wasmAbi:       *wasmAbi,
 	}
 
 	if *cFlags != "" {
@@ -540,6 +665,19 @@ func main() {
 		config.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 config.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)
 
 	switch command {
@@ -549,8 +687,11 @@ func main() {
 			usage()
 			os.Exit(1)
 		}
-		if flag.NArg() != 1 {
-			fmt.Fprintln(os.Stderr, "No package specified.")
+		pkgName := "."
+		if flag.NArg() == 1 {
+			pkgName = 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)
 		}
@@ -558,7 +699,7 @@ func main() {
 		if target == "" && filepath.Ext(*outpath) == ".wasm" {
 			target = "wasm"
 		}
-		err := Build(flag.Arg(0), *outpath, target, config)
+		err := Build(pkgName, *outpath, target, config)
 		handleCompilerError(err)
 	case "build-builtins":
 		// Note: this command is only meant to be used while making a release!
@@ -600,6 +741,17 @@ func main() {
 		}
 		err := Run(flag.Arg(0), *target, config)
 		handleCompilerError(err)
+	case "test":
+		pkgName := "."
+		if flag.NArg() == 1 {
+			pkgName = 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, *target, config)
+		handleCompilerError(err)
 	case "clean":
 		// remove cache directory
 		dir := cacheDir()

main_test.go

@@ -13,17 +13,19 @@ import (
 	"runtime"
 	"sort"
 	"testing"
+
+	"github.com/tinygo-org/tinygo/loader"
 )
 
 const TESTDATA = "testdata"
 
 func TestCompiler(t *testing.T) {
-	matches, err := filepath.Glob(TESTDATA + "/*.go")
+	matches, err := filepath.Glob(filepath.Join(TESTDATA, "*.go"))
 	if err != nil {
 		t.Fatal("could not read test files:", err)
 	}
 
-	dirMatches, err := filepath.Glob(TESTDATA + "/*/main.go")
+	dirMatches, err := filepath.Glob(filepath.Join(TESTDATA, "*", "main.go"))
 	if err != nil {
 		t.Fatal("could not read test packages:", err)
 	}
@@ -64,7 +66,7 @@ func TestCompiler(t *testing.T) {
 	if runtime.GOOS == "linux" {
 		t.Log("running tests for linux/arm...")
 		for _, path := range matches {
-			if path == "testdata/cgo/" {
+			if path == filepath.Join("testdata", "cgo")+string(filepath.Separator) {
 				continue // TODO: improve CGo
 			}
 			t.Run(path, func(t *testing.T) {
@@ -74,7 +76,7 @@ func TestCompiler(t *testing.T) {
 
 		t.Log("running tests for linux/arm64...")
 		for _, path := range matches {
-			if path == "testdata/cgo/" {
+			if path == filepath.Join("testdata", "cgo")+string(filepath.Separator) {
 				continue // TODO: improve CGo
 			}
 			t.Run(path, func(t *testing.T) {
@@ -84,7 +86,7 @@ func TestCompiler(t *testing.T) {
 
 		t.Log("running tests for WebAssembly...")
 		for _, path := range matches {
-			if path == "testdata/gc.go" {
+			if path == filepath.Join("testdata", "gc.go") {
 				continue // known to fail
 			}
 			t.Run(path, func(t *testing.T) {
@@ -97,7 +99,7 @@ func TestCompiler(t *testing.T) {
 func runTest(path, tmpdir string, target string, t *testing.T) {
 	// Get the expected output for this test.
 	txtpath := path[:len(path)-3] + ".txt"
-	if path[len(path)-1] == '/' {
+	if path[len(path)-1] == os.PathSeparator {
 		txtpath = path + "out.txt"
 	}
 	f, err := os.Open(txtpath)
@@ -114,6 +116,7 @@ func runTest(path, tmpdir string, target string, t *testing.T) {
 		opt:        "z",
 		printIR:    false,
 		dumpSSA:    false,
+		verifyIR:   true,
 		debug:      false,
 		printSizes: "",
 		wasmAbi:    "js",
@@ -121,7 +124,13 @@ func runTest(path, tmpdir string, target string, t *testing.T) {
 	binary := filepath.Join(tmpdir, "test")
 	err = Build("./"+path, binary, target, config)
 	if err != nil {
-		t.Log("failed to build:", err)
+		if errLoader, ok := err.(loader.Errors); ok {
+			for _, err := range errLoader.Errs {
+				t.Log("failed to build:", err)
+			}
+		} else {
+			t.Log("failed to build:", err)
+		}
 		t.Fail()
 		return
 	}

src/device/arm/arm.go

@@ -30,6 +30,7 @@
 package arm
 
 import (
+	"runtime/volatile"
 	"unsafe"
 )
 
@@ -69,37 +70,58 @@ func SVCall3(num uintptr, a1, a2, a3 interface{}) uintptr
 // Run the following system call (SVCall) with 4 arguments.
 func SVCall4(num uintptr, a1, a2, a3, a4 interface{}) uintptr
 
-//go:volatile
-type RegValue uint32
-
 const (
 	SCS_BASE  = 0xE000E000
 	NVIC_BASE = SCS_BASE + 0x0100
+	SCB_BASE  = SCS_BASE + 0x0D00
 )
 
+const (
+	SCB_AIRCR_VECTKEY_Pos     = 16
+	SCB_AIRCR_SYSRESETREQ_Pos = 2
+	SCB_AIRCR_SYSRESETREQ_Msk = 1 << SCB_AIRCR_SYSRESETREQ_Pos
+)
+
+// System Control Block (SCB)
+//
+// SCB_Type provides the definitions for the System Control Block Registers.
+type SCB_Type struct {
+	CPUID volatile.Register32    // CPUID Base Register
+	ICSR  volatile.Register32    // Interrupt Control and State Register
+	VTOR  volatile.Register32    // Vector Table Offset Register
+	AIRCR volatile.Register32    // Application Interrupt and Reset Control Register
+	SCR   volatile.Register32    // System Control Register
+	CCR   volatile.Register32    // Configuration Control Register
+	_     volatile.Register32    // RESERVED1;
+	SHP   [2]volatile.Register32 // System Handlers Priority Registers. [0] is RESERVED
+	SHCSR volatile.Register32    // System Handler Control and State Register
+}
+
+var SCB = (*SCB_Type)(unsafe.Pointer(uintptr(SCB_BASE)))
+
 // Nested Vectored Interrupt Controller (NVIC).
 //
 // Source:
 // http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dui0553a/CIHIGCIF.html
 type NVIC_Type struct {
-	ISER [8]RegValue // Interrupt Set-enable Registers
+	ISER [8]volatile.Register32 // Interrupt Set-enable Registers
 	_    [24]uint32
-	ICER [8]RegValue // Interrupt Clear-enable Registers
+	ICER [8]volatile.Register32 // Interrupt Clear-enable Registers
 	_    [24]uint32
-	ISPR [8]RegValue // Interrupt Set-pending Registers
+	ISPR [8]volatile.Register32 // Interrupt Set-pending Registers
 	_    [24]uint32
-	ICPR [8]RegValue // Interrupt Clear-pending Registers
+	ICPR [8]volatile.Register32 // Interrupt Clear-pending Registers
 	_    [24]uint32
-	IABR [8]RegValue // Interrupt Active Bit Registers
+	IABR [8]volatile.Register32 // Interrupt Active Bit Registers
 	_    [56]uint32
-	IPR  [60]RegValue // Interrupt Priority Registers
+	IPR  [60]volatile.Register32 // Interrupt Priority Registers
 }
 
 var NVIC = (*NVIC_Type)(unsafe.Pointer(uintptr(NVIC_BASE)))
 
 // Enable the given interrupt number.
 func EnableIRQ(irq uint32) {
-	NVIC.ISER[irq>>5] = 1 << (irq & 0x1F)
+	NVIC.ISER[irq>>5].Set(1 << (irq & 0x1F))
 }
 
 // Set the priority of the given interrupt number.
@@ -116,7 +138,7 @@ func SetPriority(irq uint32, priority uint32) {
 	regpos := irq % 4
 	mask := uint32(0xff) << (regpos * 8) // bits to clear
 	priority = priority << (regpos * 8)  // bits to set
-	NVIC.IPR[regnum] = RegValue((uint32(NVIC.IPR[regnum]) &^ mask) | priority)
+	NVIC.IPR[regnum].Set((uint32(NVIC.IPR[regnum].Get()) &^ mask) | priority)
 }
 
 // DisableInterrupts disables all interrupts, and returns the old state.
@@ -136,3 +158,14 @@ func DisableInterrupts() uintptr {
 func EnableInterrupts(mask uintptr) {
 	Asm("cpsie if")
 }
+
+// SystemReset performs a hard system reset.
+func SystemReset() {
+	// SCB->AIRCR  = ((0x5FA << SCB_AIRCR_VECTKEY_Pos)      |
+	//              SCB_AIRCR_SYSRESETREQ_Msk);
+	SCB.AIRCR.Set((0x5FA << SCB_AIRCR_VECTKEY_Pos) | SCB_AIRCR_SYSRESETREQ_Msk)
+
+	for {
+		Asm("wfi")
+	}
+}

src/device/arm/cortexm.s

@@ -0,0 +1,21 @@
+.syntax unified
+
+.section .text.HardFault_Handler
+.global  HardFault_Handler
+.type    HardFault_Handler, %function
+HardFault_Handler:
+    // Put the old stack pointer in the first argument, for easy debugging. This
+    // is especially useful on Cortex-M0, which supports far fewer debug
+    // facilities.
+    mov r0, sp
+
+    // Load the default stack pointer from address 0 so that we can call normal
+    // functions again that expect a working stack. However, it will corrupt the
+    // old stack so the function below must not attempt to recover from this
+    // fault.
+    movs r3, #0
+    ldr r3, [r3]
+    mov sp, r3
+
+    // Continue handling this error in Go.
+    bl handleHardFault

src/device/riscv/riscv.go

@@ -0,0 +1,10 @@
+package riscv
+
+// Run the given assembly code. The code will be marked as having side effects,
+// as it doesn't produce output and thus would normally be eliminated by the
+// optimizer.
+func Asm(asm string)
+
+// ReadRegister returns the contents of the specified register. The register
+// must be a processor register, reachable with the "mov" instruction.
+func ReadRegister(name string) uintptr

src/device/riscv/start.S

@@ -0,0 +1,13 @@
+.section .init
+.global _start
+.type _start,@function
+
+_start:
+    // Workaround for missing support of the la pseudo-instruction in Clang 8:
+    // https://reviews.llvm.org/D55325
+    lui sp,      %hi(_stack_top)
+    addi sp, sp, %lo(_stack_top)
+    // see https://gnu-mcu-eclipse.github.io/arch/riscv/programmer/#the-gp-global-pointer-register
+    lui gp,      %hi(__global_pointer$)
+    addi gp, gp, %lo(__global_pointer$)
+    call main

src/examples/adc/adc.go

@@ -11,8 +11,8 @@ import (
 func main() {
 	machine.InitADC()
 
-	led := machine.GPIO{machine.LED}
-	led.Configure(machine.GPIOConfig{Mode: machine.GPIO_OUTPUT})
+	led := machine.LED
+	led.Configure(machine.PinConfig{Mode: machine.PinOutput})
 
 	sensor := machine.ADC{machine.ADC2}
 	sensor.Configure()

src/examples/blinky1/blinky1.go

@@ -8,8 +8,8 @@ import (
 )
 
 func main() {
-	led := machine.GPIO{machine.LED}
-	led.Configure(machine.GPIOConfig{Mode: machine.GPIO_OUTPUT})
+	led := machine.LED
+	led.Configure(machine.PinConfig{Mode: machine.PinOutput})
 	for {
 		led.Low()
 		time.Sleep(time.Millisecond * 500)

src/examples/blinky2/blinky2.go

@@ -16,8 +16,8 @@ func main() {
 }
 
 func led1() {
-	led := machine.GPIO{machine.LED}
-	led.Configure(machine.GPIOConfig{Mode: machine.GPIO_OUTPUT})
+	led := machine.LED1
+	led.Configure(machine.PinConfig{Mode: machine.PinOutput})
 	for {
 		println("+")
 		led.Low()
@@ -30,8 +30,8 @@ func led1() {
 }
 
 func led2() {
-	led := machine.GPIO{machine.LED2}
-	led.Configure(machine.GPIOConfig{Mode: machine.GPIO_OUTPUT})
+	led := machine.LED2
+	led.Configure(machine.PinConfig{Mode: machine.PinOutput})
 	for {
 		println("  +")
 		led.Low()

src/examples/button/button.go

@@ -7,14 +7,14 @@ import (
 
 // This example assumes that the button is connected to pin 8. Change the value
 // below to use a different pin.
-const buttonPin = 8
+const (
+	led    = machine.LED
+	button = machine.Pin(8)
+)
 
 func main() {
-	led := machine.GPIO{machine.LED}
-	led.Configure(machine.GPIOConfig{Mode: machine.GPIO_OUTPUT})
-
-	button := machine.GPIO{buttonPin}
-	button.Configure(machine.GPIOConfig{Mode: machine.GPIO_INPUT})
+	led.Configure(machine.PinConfig{Mode: machine.PinOutput})
+	button.Configure(machine.PinConfig{Mode: machine.PinInput})
 
 	for {
 		if button.Get() {

src/examples/button2/button2.go

@@ -8,29 +8,29 @@ import (
 // This example assumes that you are using the pca10040 board
 
 func main() {
-	led1 := machine.GPIO{machine.LED1}
-	led1.Configure(machine.GPIOConfig{Mode: machine.GPIO_OUTPUT})
+	led1 := machine.LED1
+	led1.Configure(machine.PinConfig{Mode: machine.PinOutput})
 
-	led2 := machine.GPIO{machine.LED2}
-	led2.Configure(machine.GPIOConfig{Mode: machine.GPIO_OUTPUT})
+	led2 := machine.LED2
+	led2.Configure(machine.PinConfig{Mode: machine.PinOutput})
 
-	led3 := machine.GPIO{machine.LED3}
-	led3.Configure(machine.GPIOConfig{Mode: machine.GPIO_OUTPUT})
+	led3 := machine.LED3
+	led3.Configure(machine.PinConfig{Mode: machine.PinOutput})
 
-	led4 := machine.GPIO{machine.LED4}
-	led4.Configure(machine.GPIOConfig{Mode: machine.GPIO_OUTPUT})
+	led4 := machine.LED4
+	led4.Configure(machine.PinConfig{Mode: machine.PinOutput})
 
-	button1 := machine.GPIO{machine.BUTTON1}
-	button1.Configure(machine.GPIOConfig{Mode: machine.GPIO_INPUT_PULLUP})
+	button1 := machine.BUTTON1
+	button1.Configure(machine.PinConfig{Mode: machine.PinInputPullup})
 
-	button2 := machine.GPIO{machine.BUTTON2}
-	button2.Configure(machine.GPIOConfig{Mode: machine.GPIO_INPUT_PULLUP})
+	button2 := machine.BUTTON2
+	button2.Configure(machine.PinConfig{Mode: machine.PinInputPullup})
 
-	button3 := machine.GPIO{machine.BUTTON3}
-	button3.Configure(machine.GPIOConfig{Mode: machine.GPIO_INPUT_PULLUP})
+	button3 := machine.BUTTON3
+	button3.Configure(machine.PinConfig{Mode: machine.PinInputPullup})
 
-	button4 := machine.GPIO{machine.BUTTON4}
-	button4.Configure(machine.GPIOConfig{Mode: machine.GPIO_INPUT_PULLUP})
+	button4 := machine.BUTTON4
+	button4.Configure(machine.PinConfig{Mode: machine.PinInputPullup})
 
 	for {
 		led1.Set(button1.Get())

src/examples/colorlamp/colorlamp.go

@@ -1,79 +0,0 @@
-// This program runs on an Arduino that has the following four devices connected:
-// - Button connected to D2
-// - Rotary analog dial connected to A0
-// - RGB LED connected to D3, D5, and D6 used as PWM pins
-// - BlinkM I2C RGB LED
-//
-// Pushing the button switches which color is selected.
-// Rotating the dial changes the value for the currently selected color.
-// Changing the color value updates the color displayed on both the
-// PWM-controlled RGB LED and the I2C-controlled BlinkM.
-package main
-
-import (
-	"machine"
-	"time"
-)
-
-const (
-	buttonPin = 2
-	redPin    = 3
-	greenPin  = 5
-	bluePin   = 6
-
-	red   = 0
-	green = 1
-	blue  = 2
-)
-
-func main() {
-	machine.InitADC()
-	machine.InitPWM()
-	machine.I2C0.Configure(machine.I2CConfig{})
-
-	// Init BlinkM
-	machine.I2C0.WriteTo(0x09, []byte("o"))
-
-	button := machine.GPIO{buttonPin}
-	button.Configure(machine.GPIOConfig{Mode: machine.GPIO_INPUT})
-
-	dial := machine.ADC{machine.ADC0}
-	dial.Configure()
-
-	redLED := machine.PWM{redPin}
-	redLED.Configure()
-
-	greenLED := machine.PWM{greenPin}
-	greenLED.Configure()
-
-	blueLED := machine.PWM{bluePin}
-	blueLED.Configure()
-
-	selectedColor := red
-	colors := []uint16{0, 0, 0}
-
-	for {
-		// If we pushed the button, switch active color.
-		if !button.Get() {
-			if selectedColor == blue {
-				selectedColor = red
-			} else {
-				selectedColor++
-			}
-		}
-
-		// Change the intensity for the currently selected color based on the dial setting.
-		colors[selectedColor] = (dial.Get())
-
-		// Update the RGB LED.
-		redLED.Set(colors[red])
-		greenLED.Set(colors[green])
-		blueLED.Set(colors[blue])
-
-		// Update the BlinkM.
-		machine.I2C0.WriteTo(0x09, []byte("n"))
-		machine.I2C0.WriteTo(0x09, []byte{byte(colors[red] >> 8), byte(colors[green] >> 8), byte(colors[blue] >> 8)})
-
-		time.Sleep(time.Millisecond * 100)
-	}
-}

src/examples/echo/echo.go

@@ -9,9 +9,9 @@ import (
 
 // change these to test a different UART or pins if available
 var (
-	uart       = machine.UART0
-	tx   uint8 = machine.UART_TX_PIN
-	rx   uint8 = machine.UART_RX_PIN
+	uart = machine.UART0
+	tx   = machine.UART_TX_PIN
+	rx   = machine.UART_RX_PIN
 )
 
 func main() {

src/examples/gba-display/gba-display.go

@@ -0,0 +1,21 @@
+package main
+
+// Draw a red square on the GameBoy Advance screen.
+
+import (
+	"image/color"
+	"machine"
+)
+
+var display = machine.Display
+
+func main() {
+	display.Configure()
+
+	for x := int16(30); x < 50; x++ {
+		for y := int16(80); y < 100; y++ {
+			display.SetPixel(x, y, color.RGBA{255, 0, 0, 255})
+		}
+	}
+	display.Display()
+}

src/examples/i2s/i2s.go

@@ -0,0 +1,25 @@
+// Example using the i2s hardware interface on the Adafruit Circuit Playground Express
+// to read data from the onboard MEMS microphone.
+//
+package main
+
+import (
+	"machine"
+)
+
+func main() {
+	machine.I2S0.Configure(machine.I2SConfig{
+		Mode:        machine.I2SModePDM,
+		ClockSource: machine.I2SClockSourceExternal,
+		Stereo:      true,
+	})
+
+	data := make([]uint32, 64)
+
+	for {
+		// get the next group of samples
+		machine.I2S0.Read(data)
+
+		println("data", data[0], data[1], data[2], data[4], "...")
+	}
+}

src/examples/mcp3008/mcp3008.go

@@ -8,19 +8,17 @@ import (
 	"time"
 )
 
-// CS_PIN is the pin used for Chip Select (CS). Change to whatever is in use on your board.
-const CS_PIN = 3
+// cs is the pin used for Chip Select (CS). Change to whatever is in use on your board.
+const cs = machine.Pin(3)
 
 var (
 	tx          []byte
 	rx          []byte
 	val, result uint16
-	cs          machine.GPIO
 )
 
 func main() {
-	cs = machine.GPIO{CS_PIN}
-	cs.Configure(machine.GPIOConfig{Mode: machine.GPIO_OUTPUT})
+	cs.Configure(machine.PinConfig{Mode: machine.PinOutput})
 
 	machine.SPI0.Configure(machine.SPIConfig{
 		Frequency: 4000000,

src/examples/microbit-blink/microbit-blink.go

@@ -0,0 +1,24 @@
+// blink program for the BBC micro:bit
+package main
+
+import (
+	"machine"
+	"time"
+)
+
+// The LED matrix in the micro:bit is a multiplexed display: https://en.wikipedia.org/wiki/Multiplexed_display
+// Driver for easier control: https://github.com/tinygo-org/drivers/tree/master/microbitmatrix
+func main() {
+	ledrow := machine.LED_ROW_1
+	ledrow.Configure(machine.PinConfig{Mode: machine.PinOutput})
+	ledcol := machine.LED_COL_1
+	ledcol.Configure(machine.PinConfig{Mode: machine.PinOutput})
+	ledcol.Low()
+	for {
+		ledrow.Low()
+		time.Sleep(time.Millisecond * 500)
+
+		ledrow.High()
+		time.Sleep(time.Millisecond * 500)
+	}
+}

src/examples/wasm/.gitignore

@@ -0,0 +1 @@
+html/*

src/examples/wasm/Makefile

@@ -0,0 +1,31 @@
+invoke: clean wasm_exec
+	tinygo build -o ./html/wasm.wasm -target wasm -no-debug ./invoke/wasm.go
+	cp ./invoke/wasm.js ./html/
+	cp ./invoke/index.html ./html/
+
+export: clean wasm_exec
+	tinygo build -o ./html/wasm.wasm -target wasm -no-debug ./export/wasm.go
+	cp ./export/wasm.js ./html/
+	cp ./export/index.html ./html/
+
+callback: clean wasm_exec
+	tinygo build -o ./html/wasm.wasm -target wasm ./callback/wasm.go
+	cp ./callback/wasm.js ./html/
+	cp ./callback/index.html ./html/
+
+slices: clean wasm_exec
+	tinygo build -o ./html/wasm.wasm -target wasm -no-debug ./slices/wasm.go
+	cp ./slices/wasm.js ./html/
+	cp ./slices/index.html ./html/
+
+
+main: clean wasm_exec
+	tinygo build -o ./html/wasm.wasm -target wasm -no-debug ./main/main.go
+	cp ./main/index.html ./html/
+
+wasm_exec:
+	cp ../../../targets/wasm_exec.js ./html/
+
+clean:
+	rm -rf ./html
+	mkdir ./html

src/examples/wasm/README.md

@@ -0,0 +1,133 @@
+# TinyGo WebAssembly examples
+
+The examples here show two different ways of using WebAssembly with TinyGo:
+
+1. Defining and exporting functions via the `//go:export <name>` directive. See
+[the export folder](./export) for an example of this.  Additionally, the Wasm
+module (which has a default value of `env`) can be specified using
+`//go:wasm-module <module>`.
+1. Defining and executing a `func main()`. This is similar to how the Go
+standard library implementation works. See [the main folder](./main) for an
+example of this.
+
+## Building
+
+Build using the `tinygo` compiler:
+
+```bash
+$ tinygo build -o ./wasm.wasm -target wasm ./main/main.go
+```
+
+This creates a `wasm.wasm` file, which we can load in JavaScript and execute in
+a browser.
+
+This examples folder contains two examples that can be built using `make`:
+
+```bash
+$ make export
+```
+
+```bash
+$ make main
+```
+
+## Running
+
+Start the local web server:
+
+```bash
+$ go run server.go
+Serving ./html on http://localhost:8080
+```
+
+Use your web browser to visit http://localhost:8080.
+
+* The wasm "export" example displays a simple math equation using HTML, with
+the result calculated dynamically using WebAssembly.  Changing any of the
+values on the left hand side triggers the exported wasm `update` function to
+recalculate the result.
+* The wasm "main" example uses `println` to write to your browser JavaScript
+console. You may need to open the browser development tools console to see it.
+
+## How it works
+
+Execution of the contents require a few JavaScript helper functions which are
+called from WebAssembly.
+
+We have defined these in [wasm_exec.js](../../../targets/wasm_exec.js).  It is
+based on `$GOROOT/misc/wasm/wasm_exec.js` from the standard library, but is
+slightly different. Ensure you are using the same version of `wasm_exec.js` as
+the version of `tinygo` you are using to compile.
+
+The general steps required to run the WebAssembly file in the browser includes
+loading it into JavaScript with `WebAssembly.instantiateStreaming`, or
+`WebAssembly.instantiate` in some browsers:
+
+```js
+const go = new Go(); // Defined in wasm_exec.js
+const WASM_URL = 'wasm.wasm';
+
+var wasm;
+
+if ('instantiateStreaming' in WebAssembly) {
+	WebAssembly.instantiateStreaming(fetch(WASM_URL), go.importObject).then(function (obj) {
+		wasm = obj.instance;
+		go.run(wasm);
+	})
+} else {
+	fetch(WASM_URL).then(resp =>
+		resp.arrayBuffer()
+	).then(bytes =>
+		WebAssembly.instantiate(bytes, go.importObject).then(function (obj) {
+			wasm = obj.instance;
+			go.run(wasm);
+		})
+	)
+}
+```
+
+If you have used explicit exports, you can call them by invoking them under the
+`wasm.exports` namespace. See the [`export`](./export/wasm.js) directory for an
+example of this.
+
+In addition to the JavaScript, it is important the wasm file is served with the
+[`Content-Type`](https://developer.mozilla.org/en-US/docs/Web/HTTP/Headers/Content-Type)
+header set to `application/wasm`.  Without it, most browsers won't run it.
+
+```go
+package main
+
+import (
+	"log"
+	"net/http"
+	"strings"
+)
+
+const dir = "./html"
+
+func main() {
+	fs := http.FileServer(http.Dir(dir))
+	log.Print("Serving " + dir + " on http://localhost:8080")
+	http.ListenAndServe(":8080", http.HandlerFunc(func(resp http.ResponseWriter, req *http.Request) {
+		resp.Header().Add("Cache-Control", "no-cache")
+		if strings.HasSuffix(req.URL.Path, ".wasm") {
+			resp.Header().Set("content-type", "application/wasm")
+		}
+		fs.ServeHTTP(resp, req)
+	}))}
+```
+
+This simple server serves anything inside the `./html` directory on port
+`8080`, setting any `*.wasm` files `Content-Type` header appropriately.
+
+For development purposes (**only!**), it also sets the `Cache-Control` header
+so your browser doesn't cache the files.  This is useful while developing, to
+ensure your browser displays the newest wasm when you recompile.
+
+In a production environment you **probably wouldn't** want to set the
+`Cache-Control` header like this.  Caching is generally beneficial for end
+users.
+
+Further information on the `Cache-Control` header can be found here:
+
+* https://developer.mozilla.org/en-US/docs/Web/HTTP/Headers/Cache-Control

src/examples/wasm/callback/index.html

@@ -0,0 +1,19 @@
+<!DOCTYPE html>
+
+<html>
+
+<head>
+	<meta charset="utf-8" />
+	<title>Go WebAssembly</title>
+	<meta name="viewport" content="width=device-width, initial-scale=1" />
+	<script src="wasm_exec.js" defer></script>
+	<script src="wasm.js" defer></script>
+</head>
+
+<body>
+	<h1>WebAssembly</h1>
+	<p>Add two numbers, using WebAssembly:</p>
+	<input type="number" id="a" value="0" /> + <input type="number" id="b" value="0" /> = <input type="number" id="result" readonly />
+</body>
+
+</html>

src/examples/wasm/callback/wasm.go

@@ -0,0 +1,27 @@
+package main
+
+import (
+	"strconv"
+	"syscall/js"
+)
+
+var a, b int
+
+func main() {
+	document := js.Global().Get("document")
+	document.Call("getElementById", "a").Set("oninput", updater(&a))
+	document.Call("getElementById", "b").Set("oninput", updater(&b))
+	update()
+}
+
+func updater(n *int) js.Func {
+	return js.FuncOf(func(this js.Value, args []js.Value) interface{} {
+		*n, _ = strconv.Atoi(this.Get("value").String())
+		update()
+		return nil
+	})
+}
+
+func update() {
+	js.Global().Get("document").Call("getElementById", "result").Set("value", a+b)
+}

src/examples/wasm/callback/wasm.js

@@ -0,0 +1,26 @@
+'use strict';
+
+const WASM_URL = 'wasm.wasm';
+
+var wasm;
+
+function init() {
+  const go = new Go();
+  if ('instantiateStreaming' in WebAssembly) {
+    WebAssembly.instantiateStreaming(fetch(WASM_URL), go.importObject).then(function (obj) {
+      wasm = obj.instance;
+      go.run(wasm);
+    })
+  } else {
+    fetch(WASM_URL).then(resp =>
+      resp.arrayBuffer()
+    ).then(bytes =>
+      WebAssembly.instantiate(bytes, go.importObject).then(function (obj) {
+        wasm = obj.instance;
+        go.run(wasm);
+      })
+    )
+  }
+}
+
+init();

src/examples/wasm/export/index.html

@@ -0,0 +1,20 @@
+<!DOCTYPE html>
+
+<html>
+
+<head>
+	<meta charset="utf-8" />
+	<title>Go WebAssembly</title>
+	<meta name="viewport" content="width=device-width, initial-scale=1" />
+	<script src="wasm_exec.js" defer></script>
+	<script src="wasm.js" defer></script>
+</head>
+
+<body>
+	<h1>WebAssembly</h1>
+	<p>Add two numbers, using WebAssembly:</p>
+	<input type="number" id="a" value="2" /> + <input type="number" id="b" value="2" /> = <input type="number"
+		id="result" readonly />
+</body>
+
+</html>

src/examples/wasm/export/wasm.go

@@ -16,10 +16,10 @@ func add(a, b int) int {
 //go:export update
 func update() {
 	document := js.Global().Get("document")
-	a_str := document.Call("getElementById", "a").Get("value").String()
-	b_str := document.Call("getElementById", "b").Get("value").String()
-	a, _ := strconv.Atoi(a_str)
-	b, _ := strconv.Atoi(b_str)
-	result := a + b
+	aStr := document.Call("getElementById", "a").Get("value").String()
+	bStr := document.Call("getElementById", "b").Get("value").String()
+	a, _ := strconv.Atoi(aStr)
+	b, _ := strconv.Atoi(bStr)
+	result := add(a, b)
 	document.Call("getElementById", "result").Set("value", result)
 }

src/examples/wasm/export/wasm.js

@@ -1,6 +1,6 @@
 'use strict';
 
-const WASM_URL = '../../../wasm.wasm';
+const WASM_URL = 'wasm.wasm';
 
 var wasm;
 
@@ -14,7 +14,7 @@ function init() {
 
   const go = new Go();
   if ('instantiateStreaming' in WebAssembly) {
-    WebAssembly.instantiateStreaming(fetch(WASM_URL), go.importObject).then(function(obj) {
+    WebAssembly.instantiateStreaming(fetch(WASM_URL), go.importObject).then(function (obj) {
       wasm = obj.instance;
       go.run(wasm);
       updateResult();
@@ -23,7 +23,7 @@ function init() {
     fetch(WASM_URL).then(resp =>
       resp.arrayBuffer()
     ).then(bytes =>
-      WebAssembly.instantiate(bytes, go.importObject).then(function(obj) {
+      WebAssembly.instantiate(bytes, go.importObject).then(function (obj) {
         wasm = obj.instance;
         go.run(wasm);
         updateResult();

src/examples/wasm/invoke/index.html

@@ -0,0 +1,19 @@
+<!DOCTYPE html>
+
+<html>
+
+<head>
+    <meta charset="utf-8"/>
+    <title>Go WebAssembly</title>
+    <meta name="viewport" content="width=device-width, initial-scale=1"/>
+    <script src="wasm_exec.js" defer></script>
+    <script src="wasm.js" defer></script>
+</head>
+
+<body>
+    <h1>WebAssembly</h1>
+    <p>Edit on either side to mimic values, using WebAssembly:</p>
+    <input type="text" id="a" value=""/>==<input type="text" id="b" value=""/>
+</body>
+
+</html>

src/examples/wasm/invoke/wasm.go

@@ -0,0 +1,15 @@
+package main
+
+import (
+	"syscall/js"
+)
+
+func runner(this js.Value, args []js.Value) interface{} {
+	return args[0].Invoke(args[1]).String()
+}
+
+func main() {
+	wait := make(chan struct{}, 0)
+	js.Global().Set("runner", js.FuncOf(runner))
+	<-wait
+}

src/examples/wasm/invoke/wasm.js

@@ -0,0 +1,43 @@
+'use strict';
+
+const WASM_URL = 'wasm.wasm';
+
+var wasm;
+
+function updateRight() {
+    const value = document.getElementById("a").value;
+    window.runner(function (value) {
+        document.getElementById("b").value = value;
+    }, value);
+}
+
+function updateLeft() {
+    const value = document.getElementById("b").value;
+    window.runner(function (value) {
+        document.getElementById("a").value = value;
+    }, value);
+}
+
+function init() {
+    document.querySelector('#a').oninput = updateRight;
+    document.querySelector('#b').oninput = updateLeft;
+
+    const go = new Go();
+    if ('instantiateStreaming' in WebAssembly) {
+        WebAssembly.instantiateStreaming(fetch(WASM_URL), go.importObject).then(function (obj) {
+            wasm = obj.instance;
+            go.run(wasm);
+        })
+    } else {
+        fetch(WASM_URL).then(resp =>
+            resp.arrayBuffer()
+        ).then(bytes =>
+            WebAssembly.instantiate(bytes, go.importObject).then(function (obj) {
+                wasm = obj.instance;
+                go.run(wasm);
+            })
+        )
+    }
+}
+
+init();

src/examples/wasm/main/README.md

@@ -0,0 +1,8 @@
+# WebAssembly main execution example
+
+A simple hello world that prints to the browser console.
+
+## License
+
+Note that `index.html` is copied almost verbatim from the Go 1.12 source at
+`$GOROOT/misc/wasm/wasm_exec.html`. Its license applies to this file.

src/examples/wasm/main/index.html

@@ -0,0 +1,49 @@
+<!doctype html>
+<!--
+Copyright 2018 The Go Authors. All rights reserved.
+Use of this source code is governed by a BSD-style
+license that can be found in the LICENSE file.
+-->
+<html>
+
+<head>
+	<meta charset="utf-8">
+	<title>Go wasm</title>
+</head>
+
+<body>
+	<!--
+	Add the following polyfill for Microsoft Edge 17/18 support:
+	<script src="https://cdn.jsdelivr.net/npm/text-encoding@0.7.0/lib/encoding.min.js"></script>
+	(see https://caniuse.com/#feat=textencoder)
+	-->
+	<script src="wasm_exec.js"></script>
+	<script>
+		if (!WebAssembly.instantiateStreaming) { // polyfill
+			WebAssembly.instantiateStreaming = async (resp, importObject) => {
+				const source = await (await resp).arrayBuffer();
+				return await WebAssembly.instantiate(source, importObject);
+			};
+		}
+
+		const go = new Go();
+		let mod, inst;
+		WebAssembly.instantiateStreaming(fetch("wasm.wasm"), go.importObject).then((result) => {
+			mod = result.module;
+			inst = result.instance;
+			document.getElementById("runButton").disabled = false;
+		}).catch((err) => {
+			console.error(err);
+		});
+
+		async function run() {
+			console.clear();
+			await go.run(inst);
+			inst = await WebAssembly.instantiate(mod, go.importObject); // reset instance
+		}
+	</script>
+
+	<button onClick="run();" id="runButton" disabled>Run</button>
+</body>
+
+</html>

src/examples/wasm/main/main.go

@@ -0,0 +1,5 @@
+package main
+
+func main() {
+	println("Hello world!")
+}

src/examples/wasm/server.go

@@ -0,0 +1,21 @@
+package main
+
+import (
+	"log"
+	"net/http"
+	"strings"
+)
+
+const dir = "./html"
+
+func main() {
+	fs := http.FileServer(http.Dir(dir))
+	log.Print("Serving " + dir + " on http://localhost:8080")
+	http.ListenAndServe(":8080", http.HandlerFunc(func(resp http.ResponseWriter, req *http.Request) {
+		resp.Header().Add("Cache-Control", "no-cache")
+		if strings.HasSuffix(req.URL.Path, ".wasm") {
+			resp.Header().Set("content-type", "application/wasm")
+		}
+		fs.ServeHTTP(resp, req)
+	}))
+}

src/examples/wasm/slices/index.html

@@ -0,0 +1,19 @@
+<!DOCTYPE html>
+
+<html>
+
+<head>
+    <meta charset="utf-8"/>
+    <title>Go WebAssembly</title>
+    <meta name="viewport" content="width=device-width, initial-scale=1"/>
+    <script src="wasm_exec.js" defer></script>
+    <script src="wasm.js" defer></script>
+</head>
+
+<body>
+    <h1>WebAssembly</h1>
+    <p>type values separated by comma, using WebAssembly:</p>
+    <input type="text" id="a" value=""/>==<div id="b"></div>
+</body>
+
+</html>

src/examples/wasm/slices/wasm.go

@@ -0,0 +1,23 @@
+package main
+
+import (
+	"strings"
+	"syscall/js"
+)
+
+func splitter(this js.Value, args []js.Value) interface{} {
+	values := strings.Split(args[0].String(), ",")
+
+	result := make([]interface{}, 0)
+	for _, each := range values {
+		result = append(result, each)
+	}
+
+	return js.ValueOf(result)
+}
+
+func main() {
+	wait := make(chan struct{}, 0)
+	js.Global().Set("splitter", js.FuncOf(splitter))
+	<-wait
+}

src/examples/wasm/slices/wasm.js

@@ -0,0 +1,33 @@
+'use strict';
+
+const WASM_URL = 'wasm.wasm';
+
+var wasm;
+
+function update() {
+    const value = document.getElementById("a").value;
+    document.getElementById("b").innerHTML = JSON.stringify(window.splitter(value));
+}
+
+function init() {
+    document.querySelector('#a').oninput = update;
+
+    const go = new Go();
+    if ('instantiateStreaming' in WebAssembly) {
+        WebAssembly.instantiateStreaming(fetch(WASM_URL), go.importObject).then(function (obj) {
+            wasm = obj.instance;
+            go.run(wasm);
+        })
+    } else {
+        fetch(WASM_URL).then(resp =>
+            resp.arrayBuffer()
+        ).then(bytes =>
+            WebAssembly.instantiate(bytes, go.importObject).then(function (obj) {
+                wasm = obj.instance;
+                go.run(wasm);
+            })
+        )
+    }
+}
+
+init();

src/examples/wasm/wasm.html

@@ -1,16 +0,0 @@
-<!DOCTYPE html>
-
-<html>
-	<head>
-		<meta charset="utf-8"/>
-		<title>Go WebAssembly</title>
-		<meta name="viewport" content="width=device-width, initial-scale=1"/>
-		<script src="../../../targets/wasm_exec.js" defer></script>
-		<script src="wasm.js" defer></script>
-	</head>
-	<body>
-		<h1>WebAssembly</h1>
-		<p>Add two numbers, using WebAssembly:</p>
-		<input type="number" id="a" value="2"/> + <input type="number" id="b" value="2"/> = <input type="number" id="result" readonly/>
-	</body>
-</html>

src/machine/board_arduino.go

@@ -1,24 +1,24 @@
-// +build avr,arduino
+// +build arduino
 
 package machine
 
 const CPU_FREQUENCY = 16000000
 
 // LED on the Arduino
-const LED = 13
+const LED Pin = 13
 
 // ADC on the Arduino
 const (
-	ADC0 = 0
-	ADC1 = 1
-	ADC2 = 2
-	ADC3 = 3
-	ADC4 = 4 // Used by TWI for SDA
-	ADC5 = 5 // Used by TWI for SCL
+	ADC0 Pin = 0
+	ADC1 Pin = 1
+	ADC2 Pin = 2
+	ADC3 Pin = 3
+	ADC4 Pin = 4 // Used by TWI for SDA
+	ADC5 Pin = 5 // Used by TWI for SCL
 )
 
 // UART pins
 const (
-	UART_TX_PIN = 1
-	UART_RX_PIN = 0
+	UART_TX_PIN Pin = 1
+	UART_RX_PIN Pin = 0
 )