internal/task: use asyncify on webassembly

This change implements a new "scheduler" for WebAssembly using binaryen's asyncify transform. This is more reliable than the current "coroutines" transform, and works with non-Go code in the call stack. runtime (js/wasm): handle scheduler nesting If WASM calls into JS which calls back into WASM, it is possible for the scheduler to nest. The event from the callback must be handled immediately, so the task cannot simply be deferred to the outer scheduler. This creates a minimal scheduler loop which is used to handle such nesting.
2021-07-21 14:30:01 -04:00 · 2021-07-21 14:30:01 -04:00 · 641dcd7c16
--- a/.circleci/config.yml
+++ b/.circleci/config.yml
@ -25,7 +25,9 @@ commands:
                qemu-system-arm \
                qemu-user \
                gcc-avr \
-                avr-libc
+                avr-libc \
+                cmake \
+                ninja-build
  install-node:
    steps:
      - run:
@ -49,6 +51,13 @@ commands:
          command: |
            curl https://wasmtime.dev/install.sh -sSf | bash
            sudo ln -s ~/.wasmtime/bin/wasmtime /usr/local/bin/wasmtime
+  install-cmake:
+    steps:
+      - run:
+          name: "Install CMake"
+          command: |
+            wget https://github.com/Kitware/CMake/releases/download/v3.21.4/cmake-3.21.4-linux-x86_64.tar.gz
+            sudo tar --strip-components=1 -C /usr/local -xf cmake-3.21.4-linux-x86_64.tar.gz
  install-xtensa-toolchain:
    parameters:
      variant:
@ -76,6 +85,39 @@ commands:
            - llvm-project/clang/include
            - llvm-project/lld/include
            - llvm-project/llvm/include
+  hack-ninja-jobs:
+    steps:
+      - run:
+          name: "Hack Ninja to use less jobs"
+          command: |
+            echo -e '#!/bin/sh\n/usr/bin/ninja -j3 "$@"' > /go/bin/ninja
+            chmod +x /go/bin/ninja
+  build-binaryen-linux:
+    steps:
+      - restore_cache:
+          keys:
+            - binaryen-linux-v1
+      - run:
+          name: "Build Binaryen"
+          command: |
+            make binaryen
+      - save_cache:
+          key: binaryen-linux-v1
+          paths:
+            - build/wasm-opt
+  build-binaryen-linux-stretch:
+    steps:
+      - restore_cache:
+          keys:
+            - binaryen-linux-stretch-v1
+      - run:
+          name: "Build Binaryen"
+          command: |
+            CC=$PWD/llvm-build/bin/clang make binaryen
+      - save_cache:
+          key: binaryen-linux-stretch-v1
+          paths:
+            - build/wasm-opt
  build-wasi-libc:
    steps:
      - restore_cache:
@ -100,6 +142,8 @@ commands:
      - install-node
      - install-chrome
      - install-wasmtime
+      - hack-ninja-jobs
+      - build-binaryen-linux
      - restore_cache:
          keys:
            - go-cache-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_PREVIOUS_BUILD_NUM }}
@ -141,9 +185,13 @@ commands:
                qemu-system-arm \
                qemu-user \
                gcc-avr \
-                avr-libc
+                avr-libc \
+                ninja-build \
+                python3
      - install-node
      - install-wasmtime
+      - install-cmake
+      - hack-ninja-jobs
      - install-xtensa-toolchain:
          variant: "linux-amd64"
      - restore_cache:
@ -159,14 +207,9 @@ commands:
          command: |
            if [ ! -f llvm-build/lib/liblldELF.a ]
            then
-              # fetch LLVM source
+              # fetch LLVM source (may only have headers right now)
              rm -rf llvm-project
              make llvm-source
-              # install dependencies
-              sudo apt-get install cmake ninja-build
-              # hack ninja to use less jobs
-              echo -e '#!/bin/sh\n/usr/bin/ninja -j3 "$@"' > /go/bin/ninja
-              chmod +x /go/bin/ninja
              # build!
              make ASSERT=1 llvm-build
              find llvm-build -name CMakeFiles -prune -exec rm -r '{}' \;
@ -175,6 +218,7 @@ commands:
          key: llvm-build-11-linux-v4-assert
          paths:
            llvm-build
+      - build-binaryen-linux-stretch
      - run: make ASSERT=1
      - build-wasi-libc
      - run:
@ -206,9 +250,13 @@ commands:
                qemu-system-arm \
                qemu-user \
                gcc-avr \
-                avr-libc
+                avr-libc \
+                ninja-build \
+                python3
      - install-node
      - install-wasmtime
+      - install-cmake
+      - hack-ninja-jobs
      - install-xtensa-toolchain:
          variant: "linux-amd64"
      - restore_cache:
@ -224,14 +272,9 @@ commands:
          command: |
            if [ ! -f llvm-build/lib/liblldELF.a ]
            then
-              # fetch LLVM source
+              # fetch LLVM source (may only have headers right now)
              rm -rf llvm-project
              make llvm-source
-              # install dependencies
-              sudo apt-get install cmake ninja-build
-              # hack ninja to use less jobs
-              echo -e '#!/bin/sh\n/usr/bin/ninja -j3 "$@"' > /go/bin/ninja
-              chmod +x /go/bin/ninja
              # build!
              make llvm-build
              find llvm-build -name CMakeFiles -prune -exec rm -r '{}' \;
@ -240,6 +283,7 @@ commands:
          key: llvm-build-11-linux-v4-noassert
          paths:
            llvm-build
+      - build-binaryen-linux-stretch
      - build-wasi-libc
      - run:
          name: "Test TinyGo"
@ -283,7 +327,7 @@ commands:
            curl https://dl.google.com/go/go1.17.darwin-amd64.tar.gz -o go1.17.darwin-amd64.tar.gz
            sudo tar -C /usr/local -xzf go1.17.darwin-amd64.tar.gz
            ln -s /usr/local/go/bin/go /usr/local/bin/go
-            HOMEBREW_NO_AUTO_UPDATE=1 brew install qemu
+            HOMEBREW_NO_AUTO_UPDATE=1 brew install qemu cmake ninja
      - install-xtensa-toolchain:
          variant: "macos"
      - restore_cache:
@ -311,11 +355,9 @@ commands:
          command: |
            if [ ! -f llvm-build/lib/liblldELF.a ]
            then
-              # fetch LLVM source
+              # fetch LLVM source (may only have headers right now)
              rm -rf llvm-project
              make llvm-source
-              # install dependencies
-              HOMEBREW_NO_AUTO_UPDATE=1 brew install cmake ninja
              # build!
              make llvm-build
              find llvm-build -name CMakeFiles -prune -exec rm -r '{}' \;
@ -324,6 +366,20 @@ commands:
          key: llvm-build-11-macos-v5
          paths:
            llvm-build
+      - restore_cache:
+          keys:
+            - binaryen-macos-v1
+      - run:
+          name: "Build Binaryen"
+          command: |
+            if [ ! -f build/wasm-opt ]
+            then
+              make binaryen
+            fi
+      - save_cache:
+          key: binaryen-macos-v1
+          paths:
+            - build/wasm-opt
      - restore_cache:
          keys:
            - wasi-libc-sysroot-macos-v4
--- a/.github/workflows/windows.yml
+++ b/.github/workflows/windows.yml
@ -20,6 +20,10 @@ jobs:
        run: |
          choco install qemu --version=2020.06.12
          echo "C:\Program Files\QEMU" >> $GITHUB_PATH
+      - name: Install Ninja
+        shell: bash
+        run: |
+          choco install ninja
      - name: Checkout
        uses: actions/checkout@v2
        with:
@ -50,8 +54,6 @@ jobs:
          # fetch LLVM source
          rm -rf llvm-project
          make llvm-source
-          # install dependencies
-          choco install ninja
          # build!
          make llvm-build
          # Remove unnecessary object files (to reduce cache size).
@ -65,6 +67,15 @@ jobs:
      - name: Build wasi-libc
        if: steps.cache-wasi-libc.outputs.cache-hit != 'true'
        run: make wasi-libc
+      - name: Cache Binaryen
+        uses: actions/cache@v2
+        id: cache-binaryen
+        with:
+          key: binaryen-v1
+          path: build/binaryen
+      - name: Build Binaryen
+        if: steps.cache-binaryen.outputs.cache-hit != 'true'
+        run: make binaryen
      - name: Test TinyGo
        shell: bash
        run: make test
--- a/.gitignore
+++ b/.gitignore
@ -1,4 +1,3 @@
-build
 docs/_build
 src/device/avr/*.go
 src/device/avr/*.ld
--- a/.gitmodules
+++ b/.gitmodules
@ -26,3 +26,6 @@
 [submodule "lib/musl"]
 	path = lib/musl
 	url = git://git.musl-libc.org/musl
+[submodule "lib/binaryen"]
+	path = lib/binaryen
+	url = https://github.com/WebAssembly/binaryen.git
--- a/2
+++ b/2
@ -30,7 +30,7 @@ COPY --from=tinygo-base /tinygo/targets /tinygo/targets
 RUN cd /tinygo/ && \
    apt-get update && \
    apt-get install -y make clang-11 libllvm11 lld-11 && \
-    make wasi-libc
+    make wasi-libc binaryen

 # tinygo-avr stage installs the needed dependencies to compile TinyGo programs for AVR microcontrollers.
 FROM tinygo-base AS tinygo-avr
--- a/15
+++ b/15
@ -54,6 +54,8 @@ ifeq ($(OS),Windows_NT)
    CGO_LDFLAGS += -static -static-libgcc -static-libstdc++
    CGO_LDFLAGS_EXTRA += -lversion

+	BINARYEN_OPTION += -DCMAKE_EXE_LINKER_FLAGS='-static-libgcc -static-libstdc++'
+
    LIBCLANG_NAME = libclang

 else ifeq ($(shell uname -s),Darwin)
@ -163,12 +165,18 @@ llvm-source: $(LLVM_PROJECTDIR)/llvm
 # Configure LLVM.
 TINYGO_SOURCE_DIR=$(shell pwd)
 $(LLVM_BUILDDIR)/build.ninja: llvm-source
-	mkdir -p $(LLVM_BUILDDIR); cd $(LLVM_BUILDDIR); cmake -G Ninja $(TINYGO_SOURCE_DIR)/$(LLVM_PROJECTDIR)/llvm "-DLLVM_TARGETS_TO_BUILD=X86;ARM;AArch64;RISCV;WebAssembly" "-DLLVM_EXPERIMENTAL_TARGETS_TO_BUILD=AVR;Xtensa" -DCMAKE_BUILD_TYPE=Release -DLIBCLANG_BUILD_STATIC=ON -DLLVM_ENABLE_TERMINFO=OFF -DLLVM_ENABLE_ZLIB=OFF -DLLVM_ENABLE_LIBEDIT=OFF -DLLVM_ENABLE_Z3_SOLVER=OFF -DLLVM_ENABLE_OCAMLDOC=OFF -DLLVM_ENABLE_LIBXML2=OFF -DLLVM_ENABLE_PROJECTS="clang;lld" -DLLVM_TOOL_CLANG_TOOLS_EXTRA_BUILD=OFF -DCLANG_ENABLE_STATIC_ANALYZER=OFF -DCLANG_ENABLE_ARCMT=OFF $(LLVM_OPTION)
+	mkdir -p $(LLVM_BUILDDIR) && cd $(LLVM_BUILDDIR) && cmake -G Ninja $(TINYGO_SOURCE_DIR)/$(LLVM_PROJECTDIR)/llvm "-DLLVM_TARGETS_TO_BUILD=X86;ARM;AArch64;RISCV;WebAssembly" "-DLLVM_EXPERIMENTAL_TARGETS_TO_BUILD=AVR;Xtensa" -DCMAKE_BUILD_TYPE=Release -DLIBCLANG_BUILD_STATIC=ON -DLLVM_ENABLE_TERMINFO=OFF -DLLVM_ENABLE_ZLIB=OFF -DLLVM_ENABLE_LIBEDIT=OFF -DLLVM_ENABLE_Z3_SOLVER=OFF -DLLVM_ENABLE_OCAMLDOC=OFF -DLLVM_ENABLE_LIBXML2=OFF -DLLVM_ENABLE_PROJECTS="clang;lld" -DLLVM_TOOL_CLANG_TOOLS_EXTRA_BUILD=OFF -DCLANG_ENABLE_STATIC_ANALYZER=OFF -DCLANG_ENABLE_ARCMT=OFF $(LLVM_OPTION)

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

+# Build Binaryen
+.PHONY: binaryen
+binaryen: build/wasm-opt
+build/wasm-opt:
+	cd lib/binaryen && cmake -G Ninja . -DBUILD_STATIC_LIB=ON $(BINARYEN_OPTION) && ninja
+	cp lib/binaryen/bin/wasm-opt build/wasm-opt

 # Build wasi-libc sysroot
 .PHONY: wasi-libc
@ -476,7 +484,7 @@ endif
 wasmtest:
 	$(GO) test ./tests/wasm

-build/release: tinygo gen-device wasi-libc
+build/release: tinygo gen-device wasi-libc binaryen
 	@mkdir -p build/release/tinygo/bin
 	@mkdir -p build/release/tinygo/lib/clang/include
 	@mkdir -p build/release/tinygo/lib/CMSIS/CMSIS
@ -493,6 +501,7 @@ build/release: tinygo gen-device wasi-libc
 	@mkdir -p build/release/tinygo/pkg/armv7em-unknown-unknown-eabi
 	@echo copying source files
 	@cp -p  build/tinygo$(EXE)           build/release/tinygo/bin
+	@cp -p  build/wasm-opt$(EXE)         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
--- a/build/.gitignore
+++ b/build/.gitignore
@ -0,0 +1,2 @@
+*
+!.gitignore
--- a/builder/build.go
+++ b/builder/build.go
@ -16,9 +16,11 @@ import (
 	"io/ioutil"
 	"math/bits"
 	"os"
+	"os/exec"
 	"path/filepath"
 	"runtime"
 	"sort"
+	"strconv"
 	"strings"

 	"github.com/tinygo-org/tinygo/cgo"
@ -658,6 +660,37 @@ func Build(pkgName, outpath string, config *compileopts.Config, action func(Buil
 				}
 			}

+			// Run wasm-opt if necessary.
+			if config.Scheduler() == "asyncify" {
+				var optLevel, shrinkLevel int
+				switch config.Options.Opt {
+				case "none", "0":
+				case "1":
+					optLevel = 1
+				case "2":
+					optLevel = 2
+				case "s":
+					optLevel = 2
+					shrinkLevel = 1
+				case "z":
+					optLevel = 2
+					shrinkLevel = 2
+				default:
+					return fmt.Errorf("unknown opt level: %q", config.Options.Opt)
+				}
+				cmd := exec.Command(goenv.Get("WASMOPT"), "--asyncify", "-g",
+					"--optimize-level", strconv.Itoa(optLevel),
+					"--shrink-level", strconv.Itoa(shrinkLevel),
+					executable, "--output", executable)
+				cmd.Stdout = os.Stdout
+				cmd.Stderr = os.Stderr
+
+				err := cmd.Run()
+				if err != nil {
+					return fmt.Errorf("wasm-opt failed: %w", err)
+				}
+			}
+
 			// Print code size if requested.
 			if config.Options.PrintSizes == "short" || config.Options.PrintSizes == "full" {
 				packagePathMap := make(map[string]string, len(lprogram.Packages))
--- a/compileopts/config.go
+++ b/compileopts/config.go
@ -157,7 +157,7 @@ func (c *Config) OptLevels() (optLevel, sizeLevel int, inlinerThreshold uint) {
 // target.
 func (c *Config) FuncImplementation() string {
 	switch c.Scheduler() {
-	case "tasks":
+	case "tasks", "asyncify":
 		// 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
--- a/compileopts/options.go
+++ b/compileopts/options.go
@ -8,7 +8,7 @@ import (

 var (
 	validGCOptions            = []string{"none", "leaking", "extalloc", "conservative"}
-	validSchedulerOptions     = []string{"none", "tasks", "coroutines"}
+	validSchedulerOptions     = []string{"none", "tasks", "coroutines", "asyncify"}
 	validSerialOptions        = []string{"none", "uart", "usb"}
 	validPrintSizeOptions     = []string{"none", "short", "full"}
 	validPanicStrategyOptions = []string{"print", "trap"}
--- a/compileopts/options_test.go
+++ b/compileopts/options_test.go
@ -10,7 +10,7 @@ import (
 func TestVerifyOptions(t *testing.T) {

 	expectedGCError := errors.New(`invalid gc option 'incorrect': valid values are none, leaking, extalloc, conservative`)
-	expectedSchedulerError := errors.New(`invalid scheduler option 'incorrect': valid values are none, tasks, coroutines`)
+	expectedSchedulerError := errors.New(`invalid scheduler option 'incorrect': valid values are none, tasks, coroutines, asyncify`)
 	expectedPrintSizeError := errors.New(`invalid size option 'incorrect': valid values are none, short, full`)
 	expectedPanicStrategyError := errors.New(`invalid panic option 'incorrect': valid values are print, trap`)

--- a/compileopts/target.go
+++ b/compileopts/target.go
@ -208,6 +208,11 @@ func LoadTarget(options *Options) (*TargetSpec, error) {
 	if err != nil {
 		return nil, err
 	}
+
+	if spec.Scheduler == "asyncify" {
+		spec.ExtraFiles = append(spec.ExtraFiles, "src/internal/task/task_asyncify_wasm.S")
+	}
+
 	return spec, nil
 }

--- a/compiler/compiler_test.go
+++ b/compiler/compiler_test.go
@ -18,8 +18,9 @@ import (
 var flagUpdate = flag.Bool("update", false, "update tests based on test output")

 type testCase struct {
-	file   string
-	target string
+	file      string
+	target    string
+	scheduler string
 }

 // Basic tests for the compiler. Build some Go files and compare the output with
@ -41,20 +42,21 @@ func TestCompiler(t *testing.T) {
 	}

 	tests := []testCase{
-		{"basic.go", ""},
-		{"pointer.go", ""},
-		{"slice.go", ""},
-		{"string.go", ""},
-		{"float.go", ""},
-		{"interface.go", ""},
-		{"func.go", ""},
-		{"pragma.go", ""},
-		{"goroutine.go", "wasm"},
-		{"goroutine.go", "cortex-m-qemu"},
-		{"channel.go", ""},
-		{"intrinsics.go", "cortex-m-qemu"},
-		{"intrinsics.go", "wasm"},
-		{"gc.go", ""},
+		{"basic.go", "", ""},
+		{"pointer.go", "", ""},
+		{"slice.go", "", ""},
+		{"string.go", "", ""},
+		{"float.go", "", ""},
+		{"interface.go", "", ""},
+		{"func.go", "", "coroutines"},
+		{"pragma.go", "", ""},
+		{"goroutine.go", "wasm", "asyncify"},
+		{"goroutine.go", "wasm", "coroutines"},
+		{"goroutine.go", "cortex-m-qemu", "tasks"},
+		{"channel.go", "", ""},
+		{"intrinsics.go", "cortex-m-qemu", ""},
+		{"intrinsics.go", "wasm", ""},
+		{"gc.go", "", ""},
 	}

 	_, minor, err := goenv.GetGorootVersion(goenv.Get("GOROOT"))
@ -62,7 +64,7 @@ func TestCompiler(t *testing.T) {
 		t.Fatal("could not read Go version:", err)
 	}
 	if minor >= 17 {
-		tests = append(tests, testCase{"go1.17.go", ""})
+		tests = append(tests, testCase{"go1.17.go", "", ""})
 	}

 	for _, tc := range tests {
@ -70,7 +72,10 @@ func TestCompiler(t *testing.T) {
 		targetString := "wasm"
 		if tc.target != "" {
 			targetString = tc.target
-			name = tc.file + "-" + tc.target
+			name += "-" + tc.target
+		}
+		if tc.scheduler != "" {
+			name += "-" + tc.scheduler
 		}

 		t.Run(name, func(t *testing.T) {
@ -81,6 +86,9 @@ func TestCompiler(t *testing.T) {
 			if err != nil {
 				t.Fatal("failed to load target:", err)
 			}
+			if tc.scheduler != "" {
+				options.Scheduler = tc.scheduler
+			}
 			config := &compileopts.Config{
 				Options: options,
 				Target:  target,
@ -94,6 +102,7 @@ func TestCompiler(t *testing.T) {
 				Scheduler:          config.Scheduler(),
 				FuncImplementation: config.FuncImplementation(),
 				AutomaticStackSize: config.AutomaticStackSize(),
+				DefaultStackSize:   config.Target.DefaultStackSize,
 			}
 			machine, err := NewTargetMachine(compilerConfig)
 			if err != nil {
@ -142,6 +151,9 @@ func TestCompiler(t *testing.T) {
 			if tc.target != "" {
 				outFilePrefix += "-" + tc.target
 			}
+			if tc.scheduler != "" {
+				outFilePrefix += "-" + tc.scheduler
+			}
 			outPath := "./testdata/" + outFilePrefix + ".ll"

 			// Update test if needed. Do not check the result.
--- a/compiler/goroutine.go
+++ b/compiler/goroutine.go
@ -100,7 +100,7 @@ func (b *builder) createGo(instr *ssa.Go) {
 		switch b.Scheduler {
 		case "none", "coroutines":
 			// There are no additional parameters needed for the goroutine start operation.
-		case "tasks":
+		case "tasks", "asyncify":
 			// Add the function pointer as a parameter to start the goroutine.
 			params = append(params, funcPtr)
 		default:
@ -112,7 +112,7 @@ func (b *builder) createGo(instr *ssa.Go) {
 	paramBundle := b.emitPointerPack(params)
 	var callee, stackSize llvm.Value
 	switch b.Scheduler {
-	case "none", "tasks":
+	case "none", "tasks", "asyncify":
 		callee = b.createGoroutineStartWrapper(funcPtr, prefix, hasContext, instr.Pos())
 		if b.AutomaticStackSize {
 			// The stack size is not known until after linking. Call a dummy
@ -124,7 +124,7 @@ func (b *builder) createGo(instr *ssa.Go) {
 		} else {
 			// The stack size is fixed at compile time. By emitting it here as a
 			// constant, it can be optimized.
-			if b.Scheduler == "tasks" && b.DefaultStackSize == 0 {
+			if (b.Scheduler == "tasks" || b.Scheduler == "asyncify") && b.DefaultStackSize == 0 {
 				b.addError(instr.Pos(), "default stack size for goroutines is not set")
 			}
 			stackSize = llvm.ConstInt(b.uintptrType, b.DefaultStackSize, false)
@ -170,6 +170,11 @@ func (c *compilerContext) createGoroutineStartWrapper(fn llvm.Value, prefix stri
 	builder := c.ctx.NewBuilder()
 	defer builder.Dispose()

+	var deadlock llvm.Value
+	if c.Scheduler == "asyncify" {
+		deadlock = c.getFunction(c.program.ImportedPackage("runtime").Members["deadlock"].(*ssa.Function))
+	}
+
 	if !fn.IsAFunction().IsNil() {
 		// See whether this wrapper has already been created. If so, return it.
 		name := fn.Name()
@ -225,6 +230,12 @@ func (c *compilerContext) createGoroutineStartWrapper(fn llvm.Value, prefix stri
 		// Create the call.
 		builder.CreateCall(fn, params, "")

+		if c.Scheduler == "asyncify" {
+			builder.CreateCall(deadlock, []llvm.Value{
+				llvm.Undef(c.i8ptrType), llvm.Undef(c.i8ptrType),
+			}, "")
+		}
+
 	} else {
 		// For a function pointer like this:
 		//
@ -292,11 +303,22 @@ func (c *compilerContext) createGoroutineStartWrapper(fn llvm.Value, prefix stri

 		// Create the call.
 		builder.CreateCall(fnPtr, params, "")
+
+		if c.Scheduler == "asyncify" {
+			builder.CreateCall(deadlock, []llvm.Value{
+				llvm.Undef(c.i8ptrType), llvm.Undef(c.i8ptrType),
+			}, "")
+		}
 	}

-	// Finish the function. Every basic block must end in a terminator, and
-	// because goroutines never return a value we can simply return void.
-	builder.CreateRetVoid()
+	if c.Scheduler == "asyncify" {
+		// The goroutine was terminated via deadlock.
+		builder.CreateUnreachable()
+	} else {
+		// Finish the function. Every basic block must end in a terminator, and
+		// because goroutines never return a value we can simply return void.
+		builder.CreateRetVoid()
+	}

 	// Return a ptrtoint of the wrapper, not the function itself.
 	return builder.CreatePtrToInt(wrapper, c.uintptrType, "")
--- a/compiler/testdata/channel.ll
+++ b/compiler/testdata/channel.ll
@ -6,7 +6,8 @@ target triple = "wasm32-unknown-wasi"
 %runtime.channel = type { i32, i32, i8, %runtime.channelBlockedList*, i32, i32, i32, i8* }
 %runtime.channelBlockedList = type { %runtime.channelBlockedList*, %"internal/task.Task"*, %runtime.chanSelectState*, { %runtime.channelBlockedList*, i32, i32 } }
 %"internal/task.Task" = type { %"internal/task.Task"*, i8*, i64, %"internal/task.state" }
-%"internal/task.state" = type { i8* }
+%"internal/task.state" = type { i32, i8*, %"internal/task.stackState", i1 }
+%"internal/task.stackState" = type { i32, i32 }
 %runtime.chanSelectState = type { %runtime.channel*, i8* }

 declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*, i8*)
--- a/compiler/testdata/func-coroutines.ll
+++ b/compiler/testdata/func-coroutines.ll
--- a/compiler/testdata/gc.ll
+++ b/compiler/testdata/gc.ll
@ -5,7 +5,6 @@ target triple = "wasm32-unknown-wasi"

 %runtime.typecodeID = type { %runtime.typecodeID*, i32, %runtime.interfaceMethodInfo*, %runtime.typecodeID*, i32 }
 %runtime.interfaceMethodInfo = type { i8*, i32 }
-%runtime.funcValue = type { i8*, i32 }
 %runtime._interface = type { i32, i8* }

@main.scalar1 = hidden global i8* null, align 4
@ -72,11 +71,11 @@ entry:
 }

 ; Function Attrs: nounwind
-define hidden %runtime.funcValue* @main.newFuncValue(i8* %context, i8* %parentHandle) unnamed_addr #0 {
+define hidden { i8*, void ()* }* @main.newFuncValue(i8* %context, i8* %parentHandle) unnamed_addr #0 {
 entry:
  %new = call i8* @runtime.alloc(i32 8, i8* nonnull inttoptr (i32 197 to i8*), i8* undef, i8* null) #0
-  %0 = bitcast i8* %new to %runtime.funcValue*
-  ret %runtime.funcValue* %0
+  %0 = bitcast i8* %new to { i8*, void ()* }*
+  ret { i8*, void ()* }* %0
 }

 ; Function Attrs: nounwind
--- a/compiler/testdata/goroutine-cortex-m-qemu-tasks.ll
+++ b/compiler/testdata/goroutine-cortex-m-qemu-tasks.ll
--- a/compiler/testdata/goroutine-wasm-asyncify.ll
+++ b/compiler/testdata/goroutine-wasm-asyncify.ll
@ -0,0 +1,210 @@
+; ModuleID = 'goroutine.go'
+source_filename = "goroutine.go"
+target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
+target triple = "wasm32-unknown-wasi"
+
+%runtime.channel = type { i32, i32, i8, %runtime.channelBlockedList*, i32, i32, i32, i8* }
+%runtime.channelBlockedList = type { %runtime.channelBlockedList*, %"internal/task.Task"*, %runtime.chanSelectState*, { %runtime.channelBlockedList*, i32, i32 } }
+%"internal/task.Task" = type { %"internal/task.Task"*, i8*, i64, %"internal/task.state" }
+%"internal/task.state" = type { i32, i8*, %"internal/task.stackState", i1 }
+%"internal/task.stackState" = type { i32, i32 }
+%runtime.chanSelectState = type { %runtime.channel*, i8* }
+
+@"main$string" = internal unnamed_addr constant [4 x i8] c"test", align 1
+
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*, i8*)
+
+; Function Attrs: nounwind
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr #0 {
+entry:
+  ret void
+}
+
+; Function Attrs: nounwind
+define hidden void @main.regularFunctionGoroutine(i8* %context, i8* %parentHandle) unnamed_addr #0 {
+entry:
+  call void @"internal/task.start"(i32 ptrtoint (void (i8*)* @"main.regularFunction$gowrapper" to i32), i8* nonnull inttoptr (i32 5 to i8*), i32 8192, i8* undef, i8* null) #0
+  ret void
+}
+
+declare void @main.regularFunction(i32, i8*, i8*)
+
+declare void @runtime.deadlock(i8*, i8*)
+
+; Function Attrs: nounwind
+define linkonce_odr void @"main.regularFunction$gowrapper"(i8* %0) unnamed_addr #1 {
+entry:
+  %unpack.int = ptrtoint i8* %0 to i32
+  call void @main.regularFunction(i32 %unpack.int, i8* undef, i8* undef) #0
+  call void @runtime.deadlock(i8* undef, i8* undef) #0
+  unreachable
+}
+
+declare void @"internal/task.start"(i32, i8*, i32, i8*, i8*)
+
+; Function Attrs: nounwind
+define hidden void @main.inlineFunctionGoroutine(i8* %context, i8* %parentHandle) unnamed_addr #0 {
+entry:
+  call void @"internal/task.start"(i32 ptrtoint (void (i8*)* @"main.inlineFunctionGoroutine$1$gowrapper" to i32), i8* nonnull inttoptr (i32 5 to i8*), i32 8192, i8* undef, i8* null) #0
+  ret void
+}
+
+; Function Attrs: nounwind
+define hidden void @"main.inlineFunctionGoroutine$1"(i32 %x, i8* %context, i8* %parentHandle) unnamed_addr #0 {
+entry:
+  ret void
+}
+
+; Function Attrs: nounwind
+define linkonce_odr void @"main.inlineFunctionGoroutine$1$gowrapper"(i8* %0) unnamed_addr #2 {
+entry:
+  %unpack.int = ptrtoint i8* %0 to i32
+  call void @"main.inlineFunctionGoroutine$1"(i32 %unpack.int, i8* undef, i8* undef)
+  call void @runtime.deadlock(i8* undef, i8* undef) #0
+  unreachable
+}
+
+; Function Attrs: nounwind
+define hidden void @main.closureFunctionGoroutine(i8* %context, i8* %parentHandle) unnamed_addr #0 {
+entry:
+  %n = call i8* @runtime.alloc(i32 4, i8* nonnull inttoptr (i32 3 to i8*), i8* undef, i8* null) #0
+  %0 = bitcast i8* %n to i32*
+  store i32 3, i32* %0, align 4
+  %1 = call i8* @runtime.alloc(i32 8, i8* null, i8* undef, i8* null) #0
+  %2 = bitcast i8* %1 to i32*
+  store i32 5, i32* %2, align 4
+  %3 = getelementptr inbounds i8, i8* %1, i32 4
+  %4 = bitcast i8* %3 to i8**
+  store i8* %n, i8** %4, align 4
+  call void @"internal/task.start"(i32 ptrtoint (void (i8*)* @"main.closureFunctionGoroutine$1$gowrapper" to i32), i8* nonnull %1, i32 8192, i8* undef, i8* null) #0
+  %5 = load i32, i32* %0, align 4
+  call void @runtime.printint32(i32 %5, i8* undef, i8* null) #0
+  ret void
+}
+
+; Function Attrs: nounwind
+define hidden void @"main.closureFunctionGoroutine$1"(i32 %x, i8* %context, i8* %parentHandle) unnamed_addr #0 {
+entry:
+  %unpack.ptr = bitcast i8* %context to i32*
+  store i32 7, i32* %unpack.ptr, align 4
+  ret void
+}
+
+; Function Attrs: nounwind
+define linkonce_odr void @"main.closureFunctionGoroutine$1$gowrapper"(i8* %0) unnamed_addr #3 {
+entry:
+  %1 = bitcast i8* %0 to i32*
+  %2 = load i32, i32* %1, align 4
+  %3 = getelementptr inbounds i8, i8* %0, i32 4
+  %4 = bitcast i8* %3 to i8**
+  %5 = load i8*, i8** %4, align 4
+  call void @"main.closureFunctionGoroutine$1"(i32 %2, i8* %5, i8* undef)
+  call void @runtime.deadlock(i8* undef, i8* undef) #0
+  unreachable
+}
+
+declare void @runtime.printint32(i32, i8*, i8*)
+
+; Function Attrs: nounwind
+define hidden void @main.funcGoroutine(i8* %fn.context, void ()* %fn.funcptr, i8* %context, i8* %parentHandle) unnamed_addr #0 {
+entry:
+  %0 = call i8* @runtime.alloc(i32 12, i8* null, i8* undef, i8* null) #0
+  %1 = bitcast i8* %0 to i32*
+  store i32 5, i32* %1, align 4
+  %2 = getelementptr inbounds i8, i8* %0, i32 4
+  %3 = bitcast i8* %2 to i8**
+  store i8* %fn.context, i8** %3, align 4
+  %4 = getelementptr inbounds i8, i8* %0, i32 8
+  %5 = bitcast i8* %4 to void ()**
+  store void ()* %fn.funcptr, void ()** %5, align 4
+  call void @"internal/task.start"(i32 ptrtoint (void (i8*)* @main.funcGoroutine.gowrapper to i32), i8* nonnull %0, i32 8192, i8* undef, i8* null) #0
+  ret void
+}
+
+; Function Attrs: nounwind
+define linkonce_odr void @main.funcGoroutine.gowrapper(i8* %0) unnamed_addr #4 {
+entry:
+  %1 = bitcast i8* %0 to i32*
+  %2 = load i32, i32* %1, align 4
+  %3 = getelementptr inbounds i8, i8* %0, i32 4
+  %4 = bitcast i8* %3 to i8**
+  %5 = load i8*, i8** %4, align 4
+  %6 = getelementptr inbounds i8, i8* %0, i32 8
+  %7 = bitcast i8* %6 to void (i32, i8*, i8*)**
+  %8 = load void (i32, i8*, i8*)*, void (i32, i8*, i8*)** %7, align 4
+  call void %8(i32 %2, i8* %5, i8* undef) #0
+  call void @runtime.deadlock(i8* undef, i8* undef) #0
+  unreachable
+}
+
+; Function Attrs: nounwind
+define hidden void @main.recoverBuiltinGoroutine(i8* %context, i8* %parentHandle) unnamed_addr #0 {
+entry:
+  ret void
+}
+
+; Function Attrs: nounwind
+define hidden void @main.copyBuiltinGoroutine(i8* %dst.data, i32 %dst.len, i32 %dst.cap, i8* %src.data, i32 %src.len, i32 %src.cap, i8* %context, i8* %parentHandle) unnamed_addr #0 {
+entry:
+  %copy.n = call i32 @runtime.sliceCopy(i8* %dst.data, i8* %src.data, i32 %dst.len, i32 %src.len, i32 1, i8* undef, i8* null) #0
+  ret void
+}
+
+declare i32 @runtime.sliceCopy(i8* nocapture writeonly, i8* nocapture readonly, i32, i32, i32, i8*, i8*)
+
+; Function Attrs: nounwind
+define hidden void @main.closeBuiltinGoroutine(%runtime.channel* dereferenceable_or_null(32) %ch, i8* %context, i8* %parentHandle) unnamed_addr #0 {
+entry:
+  call void @runtime.chanClose(%runtime.channel* %ch, i8* undef, i8* null) #0
+  ret void
+}
+
+declare void @runtime.chanClose(%runtime.channel* dereferenceable_or_null(32), i8*, i8*)
+
+; Function Attrs: nounwind
+define hidden void @main.startInterfaceMethod(i32 %itf.typecode, i8* %itf.value, i8* %context, i8* %parentHandle) unnamed_addr #0 {
+entry:
+  %0 = call i8* @runtime.alloc(i32 16, i8* null, i8* undef, i8* null) #0
+  %1 = bitcast i8* %0 to i8**
+  store i8* %itf.value, i8** %1, align 4
+  %2 = getelementptr inbounds i8, i8* %0, i32 4
+  %.repack = bitcast i8* %2 to i8**
+  store i8* getelementptr inbounds ([4 x i8], [4 x i8]* @"main$string", i32 0, i32 0), i8** %.repack, align 4
+  %.repack1 = getelementptr inbounds i8, i8* %0, i32 8
+  %3 = bitcast i8* %.repack1 to i32*
+  store i32 4, i32* %3, align 4
+  %4 = getelementptr inbounds i8, i8* %0, i32 12
+  %5 = bitcast i8* %4 to i32*
+  store i32 %itf.typecode, i32* %5, align 4
+  call void @"internal/task.start"(i32 ptrtoint (void (i8*)* @"interface:{Print:func:{basic:string}{}}.Print$invoke$gowrapper" to i32), i8* nonnull %0, i32 8192, i8* undef, i8* null) #0
+  ret void
+}
+
+declare void @"interface:{Print:func:{basic:string}{}}.Print$invoke"(i8*, i8*, i32, i32, i8*, i8*) #5
+
+; Function Attrs: nounwind
+define linkonce_odr void @"interface:{Print:func:{basic:string}{}}.Print$invoke$gowrapper"(i8* %0) unnamed_addr #6 {
+entry:
+  %1 = bitcast i8* %0 to i8**
+  %2 = load i8*, i8** %1, align 4
+  %3 = getelementptr inbounds i8, i8* %0, i32 4
+  %4 = bitcast i8* %3 to i8**
+  %5 = load i8*, i8** %4, align 4
+  %6 = getelementptr inbounds i8, i8* %0, i32 8
+  %7 = bitcast i8* %6 to i32*
+  %8 = load i32, i32* %7, align 4
+  %9 = getelementptr inbounds i8, i8* %0, i32 12
+  %10 = bitcast i8* %9 to i32*
+  %11 = load i32, i32* %10, align 4
+  call void @"interface:{Print:func:{basic:string}{}}.Print$invoke"(i8* %2, i8* %5, i32 %8, i32 %11, i8* undef, i8* undef) #0
+  call void @runtime.deadlock(i8* undef, i8* undef) #0
+  unreachable
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind "tinygo-gowrapper"="main.regularFunction" }
+attributes #2 = { nounwind "tinygo-gowrapper"="main.inlineFunctionGoroutine$1" }
+attributes #3 = { nounwind "tinygo-gowrapper"="main.closureFunctionGoroutine$1" }
+attributes #4 = { nounwind "tinygo-gowrapper" }
+attributes #5 = { "tinygo-invoke"="reflect/methods.Print(string)" "tinygo-methods"="reflect/methods.Print(string)" }
+attributes #6 = { nounwind "tinygo-gowrapper"="interface:{Print:func:{basic:string}{}}.Print$invoke" }
--- a/compiler/testdata/goroutine-wasm-coroutines.ll
+++ b/compiler/testdata/goroutine-wasm-coroutines.ll
--- a/goenv/goenv.go
+++ b/goenv/goenv.go
@ -3,12 +3,15 @@
 package goenv

 import (
+	"bytes"
+	"errors"
 	"fmt"
 	"os"
 	"os/exec"
 	"os/user"
 	"path/filepath"
 	"runtime"
+	"strings"
 )

 // Keys is a slice of all available environment variable keys.
@ -67,11 +70,98 @@ func Get(name string) string {
 		return "1"
 	case "TINYGOROOT":
 		return sourceDir()
+	case "WASMOPT":
+		if path := os.Getenv("WASMOPT"); path != "" {
+			err := wasmOptCheckVersion(path)
+			if err != nil {
+				fmt.Fprintf(os.Stderr, "cannot use %q as wasm-opt (from WASMOPT environment variable): %s", path, err.Error())
+				os.Exit(1)
+			}
+
+			return path
+		}
+
+		return findWasmOpt()
 	default:
 		return ""
 	}
 }

+// Find wasm-opt, or exit with an error.
+func findWasmOpt() string {
+	tinygoroot := sourceDir()
+	searchPaths := []string{
+		tinygoroot + "/bin/wasm-opt",
+		tinygoroot + "/build/wasm-opt",
+	}
+
+	var paths []string
+	for _, path := range searchPaths {
+		if runtime.GOOS == "windows" {
+			path += ".exe"
+		}
+
+		_, err := os.Stat(path)
+		if err != nil && os.IsNotExist(err) {
+			continue
+		}
+
+		paths = append(paths, path)
+	}
+
+	if path, err := exec.LookPath("wasm-opt"); err == nil {
+		paths = append(paths, path)
+	}
+
+	if len(paths) == 0 {
+		fmt.Fprintln(os.Stderr, "error: could not find wasm-opt, set the WASMOPT environment variable to override")
+		os.Exit(1)
+	}
+
+	errs := make([]error, len(paths))
+	for i, path := range paths {
+		err := wasmOptCheckVersion(path)
+		if err == nil {
+			return path
+		}
+
+		errs[i] = err
+	}
+	fmt.Fprintln(os.Stderr, "no usable wasm-opt found, update or run \"make binaryen\"")
+	for i, path := range paths {
+		fmt.Fprintf(os.Stderr, "\t%s: %s\n", path, errs[i].Error())
+	}
+	os.Exit(1)
+	panic("unreachable")
+}
+
+// wasmOptCheckVersion checks if a copy of wasm-opt is usable.
+func wasmOptCheckVersion(path string) error {
+	cmd := exec.Command(path, "--version")
+	var buf bytes.Buffer
+	cmd.Stdout = &buf
+	cmd.Stderr = os.Stderr
+	err := cmd.Run()
+	if err != nil {
+		return err
+	}
+
+	str := buf.String()
+	if strings.Contains(str, "(") {
+		// The git tag may be placed in parentheses after the main version string.
+		str = strings.Split(str, "(")[0]
+	}
+
+	str = strings.TrimSpace(str)
+	var ver uint
+	_, err = fmt.Sscanf(str, "wasm-opt version %d", &ver)
+	if err != nil || ver < 102 {
+		return errors.New("incompatible wasm-opt (need 102 or newer)")
+	}
+
+	return nil
+}
+
 // Return the TINYGOROOT, or exit with an error.
 func sourceDir() string {
 	// Use $TINYGOROOT as root, if available.
--- a/lib/binaryen
+++ b/lib/binaryen
@ -0,0 +1 @@
+Subproject commit 96f7acf09aae1ec6e8bc573dfa8f309c4f892a40
--- a/src/internal/task/task_asyncify.go
+++ b/src/internal/task/task_asyncify.go
@ -0,0 +1,127 @@
+//go:build scheduler.asyncify
+// +build scheduler.asyncify
+
+package task
+
+import (
+	"unsafe"
+)
+
+// Stack canary, to detect a stack overflow. The number is a random number
+// generated by random.org. The bit fiddling dance is necessary because
+// otherwise Go wouldn't allow the cast to a smaller integer size.
+const stackCanary = uintptr(uint64(0x670c1333b83bf575) & uint64(^uintptr(0)))
+
+//go:linkname runtimePanic runtime.runtimePanic
+func runtimePanic(str string)
+
+// state is a structure which holds a reference to the state of the task.
+// When the task is suspended, the stack pointers are saved here.
+type state struct {
+	// entry is the entry function of the task.
+	// This is needed every time the function is invoked so that asyncify knows what to rewind.
+	entry uintptr
+
+	// args are a pointer to a struct holding the arguments of the function.
+	args unsafe.Pointer
+
+	// stackState is the state of the stack while unwound.
+	stackState
+
+	launched bool
+}
+
+// stackState is the saved state of a stack while unwound.
+// The stack is arranged with asyncify at the bottom, C stack at the top, and a gap of available stack space between the two.
+type stackState struct {
+	// asyncify is the stack pointer of the asyncify stack.
+	// This starts from the bottom and grows upwards.
+	asyncifysp uintptr
+
+	// asyncify is stack pointer of the C stack.
+	// This starts from the top and grows downwards.
+	csp uintptr
+}
+
+// start creates and starts a new goroutine with the given function and arguments.
+// The new goroutine is immediately started.
+func start(fn uintptr, args unsafe.Pointer, stackSize uintptr) {
+	t := &Task{}
+	t.state.initialize(fn, args, stackSize)
+	runqueuePushBack(t)
+}
+
+//export tinygo_launch
+func (*state) launch()
+
+//go:linkname align runtime.align
+func align(p uintptr) uintptr
+
+// initialize the state and prepare to call the specified function with the specified argument bundle.
+func (s *state) initialize(fn uintptr, args unsafe.Pointer, stackSize uintptr) {
+	// Save the entry call.
+	s.entry = fn
+	s.args = args
+
+	// Create a stack.
+	stack := make([]uintptr, stackSize/unsafe.Sizeof(uintptr(0)))
+
+	// Calculate stack base addresses.
+	s.asyncifysp = uintptr(unsafe.Pointer(&stack[0]))
+	s.csp = uintptr(unsafe.Pointer(&stack[0])) + uintptr(len(stack))*unsafe.Sizeof(uintptr(0))
+	stack[0] = stackCanary
+}
+
+//go:linkname runqueuePushBack runtime.runqueuePushBack
+func runqueuePushBack(*Task)
+
+// currentTask is the current running task, or nil if currently in the scheduler.
+var currentTask *Task
+
+// Current returns the current active task.
+func Current() *Task {
+	return currentTask
+}
+
+// Pause suspends the current task and returns to the scheduler.
+// This function may only be called when running on a goroutine stack, not when running on the system stack.
+func Pause() {
+	// This is mildly unsafe but this is also the only place we can do this.
+	if *(*uintptr)(unsafe.Pointer(currentTask.state.asyncifysp)) != stackCanary {
+		runtimePanic("stack overflow")
+	}
+
+	currentTask.state.unwind()
+
+	*(*uintptr)(unsafe.Pointer(currentTask.state.asyncifysp)) = stackCanary
+}
+
+//export tinygo_unwind
+func (*stackState) unwind()
+
+// Resume the task until it pauses or completes.
+// This may only be called from the scheduler.
+func (t *Task) Resume() {
+	// The current task must be saved and restored because this can nest on WASM with JS.
+	prevTask := currentTask
+	currentTask = t
+	if !t.state.launched {
+		t.state.launch()
+		t.state.launched = true
+	} else {
+		t.state.rewind()
+	}
+	currentTask = prevTask
+	if t.state.asyncifysp > t.state.csp {
+		runtimePanic("stack overflow")
+	}
+}
+
+//export tinygo_rewind
+func (*state) rewind()
+
+// OnSystemStack returns whether the caller is running on the system stack.
+func OnSystemStack() bool {
+	// If there is not an active goroutine, then this must be running on the system stack.
+	return Current() == nil
+}
--- a/src/internal/task/task_asyncify_wasm.S
+++ b/src/internal/task/task_asyncify_wasm.S
@ -0,0 +1,99 @@
+.globaltype __stack_pointer, i32
+
+.global  tinygo_unwind
+.type    tinygo_unwind,@function
+tinygo_unwind: // func (state *stackState) unwind()
+    .functype tinygo_unwind (i32) -> ()
+    // Check if we are rewinding.
+    i32.const 0
+    i32.load8_u tinygo_rewinding
+    if // if tinygo_rewinding {
+    // Stop rewinding.
+    call stop_rewind
+    i32.const 0
+    i32.const 0
+    i32.store8 tinygo_rewinding // tinygo_rewinding = false;
+    else
+    // Save the C stack pointer (destination structure pointer is in local 0).
+    local.get 0
+    global.get __stack_pointer
+    i32.store 4 // state.csp = getCurrentStackPointer()
+    // Ask asyncify to unwind.
+    // When resuming, asyncify will return this function with tinygo_rewinding set to true.
+    local.get 0
+    call start_unwind // asyncify.start_unwind(state)
+    end_if
+    return
+    end_function
+
+.global tinygo_launch
+.type tinygo_launch,@function
+tinygo_launch: // func (state *state) launch()
+    .functype tinygo_launch (i32) -> ()
+    // Switch to the goroutine's C stack.
+    global.get __stack_pointer // prev := getCurrentStackPointer()
+    local.get 0
+    i32.load 12
+    global.set __stack_pointer // setStackPointer(state.csp)
+    // Get the argument pack and entry pointer.
+    local.get 0
+    i32.load 4 // args := state.args
+    local.get 0
+    i32.load 0 // fn := state.entry
+    // Launch the entry function.
+    call_indirect (i32) -> () // fn(args)
+    // Stop unwinding.
+    call stop_unwind
+    // Restore the C stack.
+    global.set __stack_pointer // setStackPointer(prev)
+    return
+    end_function
+
+.global  tinygo_rewind
+.type    tinygo_rewind,@function
+tinygo_rewind: // func (state *state) rewind()
+    .functype tinygo_rewind (i32) -> ()
+    // Switch to the goroutine's C stack.
+    global.get __stack_pointer // prev := getCurrentStackPointer()
+    local.get 0
+    i32.load 12
+    global.set __stack_pointer // setStackPointer(state.csp)
+    // Get the argument pack and entry pointer.
+    local.get 0
+    i32.load 4 // args := state.args
+    local.get 0
+    i32.load 0 // fn := state.entry
+    // Prepare to rewind.
+    i32.const 0
+    i32.const 1
+    i32.store8 tinygo_rewinding // tinygo_rewinding = true;
+    local.get 0
+    i32.const 8
+    i32.add
+    call start_rewind // asyncify.start_rewind(&state.stackState)
+    // Launch the entry function.
+    // This will actually rewind the call stack.
+    call_indirect (i32) -> () // fn(args)
+    // Stop unwinding.
+    call stop_unwind
+    // Restore the C stack.
+    global.set __stack_pointer // setStackPointer(prev)
+    return
+    end_function
+
+.functype start_unwind (i32) -> ()
+.import_module start_unwind, asyncify
+.functype stop_unwind () -> ()
+.import_module stop_unwind, asyncify
+.functype start_rewind (i32) -> ()
+.import_module start_rewind, asyncify
+.functype stop_rewind () -> ()
+.import_module stop_rewind, asyncify
+
+        .hidden tinygo_rewinding                # @tinygo_rewinding
+        .type   tinygo_rewinding,@object
+        .section        .bss.tinygo_rewinding,"",@
+        .globl  tinygo_rewinding
+tinygo_rewinding:
+        .int8   0                               # 0x0
+        .size   tinygo_rewinding, 1
--- a/src/runtime/gc_conservative.go
+++ b/src/runtime/gc_conservative.go
@ -1,3 +1,4 @@
+//go:build gc.conservative
 // +build gc.conservative

 package runtime
--- a/src/runtime/runtime_wasm_js.go
+++ b/src/runtime/runtime_wasm_js.go
@ -1,3 +1,4 @@
+//go:build wasm && !wasi
 // +build wasm,!wasi

 package runtime
@ -6,13 +7,19 @@ import "unsafe"

 type timeUnit float64 // time in milliseconds, just like Date.now() in JavaScript

+// wasmNested is used to detect scheduler nesting (WASM calls into JS calls back into WASM).
+// When this happens, we need to use a reduced version of the scheduler.
+var wasmNested bool
+
 //export _start
 func _start() {
 	// These need to be initialized early so that the heap can be initialized.
 	heapStart = uintptr(unsafe.Pointer(&heapStartSymbol))
 	heapEnd = uintptr(wasm_memory_size(0) * wasmPageSize)

+	wasmNested = true
 	run()
+	wasmNested = false
 }

 var handleEvent func()
@ -27,12 +34,27 @@ func resume() {
 	go func() {
 		handleEvent()
 	}()
+
+	if wasmNested {
+		minSched()
+		return
+	}
+
+	wasmNested = true
 	scheduler()
+	wasmNested = false
 }

 //export go_scheduler
 func go_scheduler() {
+	if wasmNested {
+		minSched()
+		return
+	}
+
+	wasmNested = true
 	scheduler()
+	wasmNested = false
 }

 func ticksToNanoseconds(ticks timeUnit) int64 {
--- a/src/runtime/scheduler.go
+++ b/src/runtime/scheduler.go
@ -172,6 +172,24 @@ func scheduler() {
 	}
 }

+// This horrible hack exists to make WASM work properly.
+// When a WASM program calls into JS which calls back into WASM, the event with which we called back in needs to be handled before returning.
+// Thus there are two copies of the scheduler running at once.
+// This is a reduced version of the scheduler which does not deal with the timer queue (that is a problem for the outer scheduler).
+func minSched() {
+	scheduleLog("start nested scheduler")
+	for !schedulerDone {
+		t := runqueue.Pop()
+		if t == nil {
+			break
+		}
+
+		scheduleLogTask("  run:", t)
+		t.Resume()
+	}
+	scheduleLog("stop nested scheduler")
+}
+
 func Gosched() {
 	runqueue.Push(task.Current())
 	task.Pause()
--- a/targets/wasi.json
+++ b/targets/wasi.json
@ -6,6 +6,8 @@
 	"goarch":        "arm",
 	"linker":        "wasm-ld",
 	"libc":          "wasi-libc",
+	"scheduler":     "asyncify",
+	"default-stack-size": 8192,
 	"ldflags": [
 		"--allow-undefined",
 		"--stack-first",
--- a/targets/wasm.json
+++ b/targets/wasm.json
@ -6,6 +6,8 @@
 	"goarch":        "wasm",
 	"linker":        "wasm-ld",
 	"libc":          "wasi-libc",
+	"scheduler":     "asyncify",
+	"default-stack-size": 8192,
 	"ldflags": [
 		"--allow-undefined",
 		"--stack-first",
--- a/tests/wasm/chan_test.go
+++ b/tests/wasm/chan_test.go
@ -24,13 +24,12 @@ func TestChan(t *testing.T) {
 	err = chromedp.Run(ctx,
 		chromedp.Navigate(server.URL+"/run?file=chan.wasm"),
 		waitLog(`1
-2
 4
+2
 3
 true`),
 	)
 	if err != nil {
 		t.Fatal(err)
 	}
-
 }
--- a/transform/optimizer.go
+++ b/transform/optimizer.go
@ -125,7 +125,7 @@ func Optimize(mod llvm.Module, config *compileopts.Config, optLevel, sizeLevel i
 		if err != nil {
 			return []error{err}
 		}
-	case "tasks":
+	case "tasks", "asyncify":
 		// No transformations necessary.
 	case "none":
 		// Check for any goroutine starts.
@ -219,7 +219,7 @@ func getFunctionsUsedInTransforms(config *compileopts.Config) []string {
 	case "none":
 	case "coroutines":
 		fnused = append(append([]string{}, fnused...), coroFunctionsUsedInTransforms...)
-	case "tasks":
+	case "tasks", "asyncify":
 		fnused = append(append([]string{}, fnused...), taskFunctionsUsedInTransforms...)
 	default:
 		panic(fmt.Errorf("invalid scheduler %q", config.Scheduler()))
				`@ -0,0 +1 @@`
				`Subproject commit 96f7acf09aae1ec6e8bc573dfa8f309c4f892a40`