Implement WebAssembly threads proposal with interpreter (tetratelabs#…

…1460)

Signed-off-by: Anuraag Agrawal <anuraaga@gmail.com>

.github/workflows/spectest.yaml

@@ -35,6 +35,7 @@ jobs:
         spec-version:
           - "v1"
           - "v2"
+          - "threads"
 
     steps:
       - uses: actions/checkout@v3
@@ -61,6 +62,7 @@ jobs:
         spec-version:
           - "v1"
           - "v2"
+          - "threads"
 
     steps:
       - uses: actions/checkout@v3

Makefile

@@ -125,11 +125,15 @@ spectest_v2_dir := $(spectest_base_dir)/v2
 spectest_v2_testdata_dir := $(spectest_v2_dir)/testdata
 # Latest draft state as of Dec 16, 2022.
 spec_version_v2 := 1782235239ddebaf2cb079b00fdaa2d2c4dedba3
+spectest_threads_dir := $(spectest_base_dir)/threads
+spectest_threads_testdata_dir := $(spectest_threads_dir)/testdata
+spec_version_threads := cc01bf0d17ba3fb1dc59fb7c5c725838aff18b50
 
 .PHONY: build.spectest
 build.spectest:
 	@$(MAKE) build.spectest.v1
 	@$(MAKE) build.spectest.v2
+	@$(MAKE) build.spectest.threads
 
 .PHONY: build.spectest.v1
 build.spectest.v1: # Note: wabt by default uses >1.0 features, so wast2json flags might drift as they include more. See WebAssembly/wabt#1878
@@ -169,9 +173,21 @@ build.spectest.v2: # Note: SIMD cases are placed in the "simd" subdirectory.
 		wast2json --debug-names $$f; \
 	done
 
+.PHONY: build.spectest.threads
+build.spectest.threads:
+	@mkdir -p $(spectest_threads_testdata_dir)
+	@cd $(spectest_threads_testdata_dir) \
+		&& curl -sSL 'https://api.github.com/repos/WebAssembly/threads/contents/test/core?ref=$(spec_version_threads)' | jq -r '.[]| .download_url' | grep -E "atomic.wast" | xargs -Iurl curl -sJL url -O
+# Fix broken CAS spectests
+# https://github.com/WebAssembly/threads/issues/195#issuecomment-1318429506
+	@cd $(spectest_threads_testdata_dir) && patch < ../atomic.wast.patch
+	@cd $(spectest_threads_testdata_dir) && for f in `find . -name '*.wast'`; do \
+		wast2json --enable-threads --debug-names $$f; \
+	done
+
 .PHONY: test
 test:
-	@go test $(go_test_options) $$(go list ./... | grep -vE '$(spectest_v1_dir)|$(spectest_v2_dir)')
+	@go test $(go_test_options) $$(go list ./... | grep -vE '$(spectest_v1_dir)|$(spectest_v2_dir)|$(spectest_threads_dir)')
 	@cd internal/version/testdata && go test $(go_test_options) ./...
 
 .PHONY: coverage
@@ -185,13 +201,17 @@ coverage: ## Generate test coverage
 spectest:
 	@$(MAKE) spectest.v1
 	@$(MAKE) spectest.v2
+	@$(MAKE) spectest.threads
 
 spectest.v1:
 	@go test $(go_test_options) $$(go list ./... | grep $(spectest_v1_dir))
 
 spectest.v2:
 	@go test $(go_test_options) $$(go list ./... | grep $(spectest_v2_dir))
 
+spectest.threads:
+	@go test $(go_test_options) $$(go list ./... | grep $(spectest_threads_dir))
+
 golangci_lint_path := $(shell go env GOPATH)/bin/golangci-lint
 
 $(golangci_lint_path):

api/features.go

@@ -143,6 +143,8 @@ const (
 	// Note: The instruction list is too long to enumerate in godoc.
 	// See https://github.com/WebAssembly/spec/blob/wg-2.0.draft1/proposals/simd/SIMD.md
 	CoreFeatureSIMD
+
+	// Update experimental/features.go when adding elements here.
 )
 
 // SetEnabled enables or disables the feature or group of features.
@@ -207,6 +209,9 @@ func featureName(f CoreFeatures) string {
 	case CoreFeatureSIMD:
 		// match https://github.com/WebAssembly/spec/blob/wg-2.0.draft1/proposals/simd/SIMD.md
 		return "simd"
+	case CoreFeatureSIMD << 1: // Defined in experimental/features.go
+		// match https://github.com/WebAssembly/threads/blob/main/proposals/threads/Overview.md
+		return "threads"
 	}
 	return ""
 }

experimental/features.go

@@ -0,0 +1,14 @@
+package experimental
+
+import "github.com/tetratelabs/wazero/api"
+
+// CoreFeaturesThreads enables threads instructions ("threads").
+//
+// # Notes
+//
+//   - This is not yet implemented by default, so you will need to use
+//     wazero.NewRuntimeConfigInterpreter
+//   - The instruction list is too long to enumerate in godoc.
+//     See https://github.com/WebAssembly/threads/blob/main/proposals/threads/Overview.md
+//   - Atomic operations are guest-only until api.Memory or otherwise expose them to host functions.
+const CoreFeaturesThreads = api.CoreFeatureSIMD << 1 // TODO: Implement the compiler engine

experimental/features_example_test.go

@@ -0,0 +1,94 @@
+package experimental_test
+
+import (
+	"context"
+	_ "embed"
+	"fmt"
+	"log"
+	"sync"
+
+	"github.com/tetratelabs/wazero"
+	"github.com/tetratelabs/wazero/api"
+	"github.com/tetratelabs/wazero/experimental"
+	"github.com/tetratelabs/wazero/imports/wasi_snapshot_preview1"
+)
+
+// pthreadWasm was generated by the following:
+//
+//	docker run -it --rm -v `pwd`/testdata:/workspace ghcr.io/webassembly/wasi-sdk:wasi-sdk-20 sh -c '$CC -o /workspace/pthread.wasm /workspace/pthread.c --target=wasm32-wasi-threads --sysroot=/wasi-sysroot -pthread -mexec-model=reactor -Wl,--export=run -Wl,--export=get'
+//
+// TODO: Use zig cc instead of wasi-sdk to compile when it supports wasm32-wasi-threads
+// https://github.com/ziglang/zig/issues/15484
+//
+//go:embed testdata/pthread.wasm
+var pthreadWasm []byte
+
+// This shows how to use a WebAssembly module compiled with the threads feature.
+func ExampleCoreFeaturesThreads() {
+	// Use a default context
+	ctx := context.Background()
+
+	// Threads support is currently only supported with interpreter, so the config
+	// must explicitly specify it.
+	cfg := wazero.NewRuntimeConfigInterpreter()
+
+	// Threads support must be enabled explicitly in addition to standard V2 features.
+	cfg = cfg.WithCoreFeatures(api.CoreFeaturesV2 | experimental.CoreFeaturesThreads)
+
+	r := wazero.NewRuntimeWithConfig(ctx, cfg)
+	defer r.Close(ctx)
+
+	// Because we are using wasi-sdk to compile the guest, we must initialize WASI.
+	wasi_snapshot_preview1.MustInstantiate(ctx, r)
+
+	mod, err := r.Instantiate(ctx, pthreadWasm)
+	if err != nil {
+		log.Panicln(err)
+	}
+
+	// Channel to synchronize start of goroutines before running.
+	startCh := make(chan struct{})
+	// Channel to synchronize end of goroutines.
+	endCh := make(chan struct{})
+
+	// Unfortunately, while memory accesses are thread-safe using atomic operations, compilers such
+	// as LLVM still have global state that is not handled thread-safe, preventing actually making
+	// concurrent invocations. We go ahead and add a global lock for now until this is resolved.
+	// TODO: Remove this lock once functions can actually be called concurrently.
+	var mu sync.Mutex
+
+	// We start up 8 goroutines and run for 6000 iterations each. The count should reach
+	// 48000, at the end, but it would not if threads weren't working!
+	for i := 0; i < 8; i++ {
+		go func() {
+			defer func() { endCh <- struct{}{} }()
+			<-startCh
+
+			mu.Lock()
+			defer mu.Unlock()
+
+			// ExportedFunction must be called within each goroutine to have independent call stacks.
+			// This incurs some overhead, a sync.Pool can be used to reduce this overhead if neeeded.
+			fn := mod.ExportedFunction("run")
+			for i := 0; i < 6000; i++ {
+				_, err := fn.Call(ctx)
+				if err != nil {
+					log.Panicln(err)
+				}
+			}
+		}()
+	}
+	for i := 0; i < 8; i++ {
+		startCh <- struct{}{}
+	}
+	for i := 0; i < 8; i++ {
+		<-endCh
+	}
+
+	res, err := mod.ExportedFunction("get").Call(ctx)
+	if err != nil {
+		log.Panicln(err)
+	}
+	fmt.Println(res[0])
+	// Output: 48000
+}

experimental/testdata/pthread.c

@@ -0,0 +1,18 @@
+#include <pthread.h>
+
+pthread_mutex_t mutex;
+int count = 0;
+
+void run() {
+  pthread_mutex_lock(&mutex);
+  count++;
+  pthread_mutex_unlock(&mutex);
+}
+
+int get() {
+  int res;
+  pthread_mutex_lock(&mutex);
+  res = count;
+  pthread_mutex_unlock(&mutex);
+  return res;
+}