wasm: add Boehm GC support

aykevl · deadprogram · commit c08b2dfe068b · 2025-06-13T14:42:23.000+02:00
This adds support for `-gc=boehm` on `-target=wasip1` and `-target=wasm`
(in a browser or NodeJS). Notably it does *not* add Boehm GC support for
`-target=wasip2`, since that target doesn't have a real libc.
diff --git a/builder/bdwgc.go b/builder/bdwgc.go
@@ -14,7 +14,7 @@ var BoehmGC = Library{
 	name: "bdwgc",
 	cflags: func(target, headerPath string) []string {
 		libdir := filepath.Join(goenv.Get("TINYGOROOT"), "lib/bdwgc")
-		return []string{
+		flags := []string{
 			// use a modern environment
 			"-DUSE_MMAP",              // mmap is available
 			"-DUSE_MUNMAP",            // return memory to the OS using munmap
@@ -30,6 +30,7 @@ var BoehmGC = Library{
 			// Use a minimal environment.
 			"-DNO_MSGBOX_ON_ERROR", // don't call MessageBoxA on Windows
 			"-DDONT_USE_ATEXIT",
+			"-DNO_GETENV",
 
 			// Special flag to work around the lack of __data_start in ld.lld.
 			// TODO: try to fix this in LLVM/lld directly so we don't have to
@@ -49,12 +50,13 @@ var BoehmGC = Library{
 
 			"-I" + libdir + "/include",
 		}
+		return flags
 	},
 	needsLibc: true,
 	sourceDir: func() string {
 		return filepath.Join(goenv.Get("TINYGOROOT"), "lib/bdwgc")
 	},
-	librarySources: func(target string) ([]string, error) {
+	librarySources: func(target string, _ bool) ([]string, error) {
 		sources := []string{
 			"allchblk.c",
 			"alloc.c",
diff --git a/builder/builtins.go b/builder/builtins.go
@@ -226,7 +226,7 @@ var libCompilerRT = Library{
 		// Development build.
 		return filepath.Join(goenv.Get("TINYGOROOT"), "lib/compiler-rt-builtins")
 	},
-	librarySources: func(target string) ([]string, error) {
+	librarySources: func(target string, _ bool) ([]string, error) {
 		builtins := append([]string{}, genericBuiltins...) // copy genericBuiltins
 		switch compileopts.CanonicalArchName(target) {
 		case "arm":
diff --git a/builder/library.go b/builder/library.go
@@ -38,7 +38,7 @@ type Library struct {
 	sourceDir func() string
 
 	// The source files, relative to sourceDir.
-	librarySources func(target string) ([]string, error)
+	librarySources func(target string, libcNeedsMalloc bool) ([]string, error)
 
 	// The source code for the crt1.o file, relative to sourceDir.
 	crt1Source string
@@ -226,7 +226,7 @@ func (l *Library) load(config *compileopts.Config, tmpdir string) (job *compileJ
 
 	// Create jobs to compile all sources. These jobs are depended upon by the
 	// archive job above, so must be run first.
-	paths, err := l.librarySources(target)
+	paths, err := l.librarySources(target, config.LibcNeedsMalloc())
 	if err != nil {
 		return nil, nil, err
 	}
diff --git a/builder/mingw-w64.go b/builder/mingw-w64.go
@@ -48,7 +48,7 @@ var libMinGW = Library{
 		}
 		return flags
 	},
-	librarySources: func(target string) ([]string, error) {
+	librarySources: func(target string, _ bool) ([]string, error) {
 		// These files are needed so that printf and the like are supported.
 		var sources []string
 		if strings.Split(target, "-")[0] == "i386" {
diff --git a/builder/musl.go b/builder/musl.go
@@ -121,7 +121,7 @@ var libMusl = Library{
 		return cflags
 	},
 	sourceDir: func() string { return filepath.Join(goenv.Get("TINYGOROOT"), "lib/musl/src") },
-	librarySources: func(target string) ([]string, error) {
+	librarySources: func(target string, _ bool) ([]string, error) {
 		arch := compileopts.MuslArchitecture(target)
 		globs := []string{
 			"conf/*.c",
diff --git a/builder/picolibc.go b/builder/picolibc.go
@@ -43,7 +43,7 @@ var libPicolibc = Library{
 		}
 	},
 	sourceDir: func() string { return filepath.Join(goenv.Get("TINYGOROOT"), "lib/picolibc/newlib") },
-	librarySources: func(target string) ([]string, error) {
+	librarySources: func(target string, _ bool) ([]string, error) {
 		sources := append([]string(nil), picolibcSources...)
 		if !strings.HasPrefix(target, "avr") {
 			// Small chips without long jumps can't compile many files (printf,
diff --git a/builder/wasilibc.go b/builder/wasilibc.go
@@ -120,7 +120,7 @@ var libWasiLibc = Library{
 		return nil
 	},
 	sourceDir: func() string { return filepath.Join(goenv.Get("TINYGOROOT"), "lib/wasi-libc") },
-	librarySources: func(target string) ([]string, error) {
+	librarySources: func(target string, libcNeedsMalloc bool) ([]string, error) {
 		type filePattern struct {
 			glob    string
 			exclude []string
@@ -169,6 +169,11 @@ var libWasiLibc = Library{
 			{glob: "libc-bottom-half/sources/*.c"},
 		}
 
+		// We're using the Boehm GC, so we need a heap implementation in the libc.
+		if libcNeedsMalloc {
+			globs = append(globs, filePattern{glob: "dlmalloc/src/dlmalloc.c"})
+		}
+
 		// See: LIBC_TOP_HALF_MUSL_SOURCES in the Makefile
 		sources := []string{
 			"libc-top-half/musl/src/misc/a64l.c",
diff --git a/builder/wasmbuiltins.go b/builder/wasmbuiltins.go
@@ -41,7 +41,7 @@ var libWasmBuiltins = Library{
 		}
 	},
 	sourceDir: func() string { return filepath.Join(goenv.Get("TINYGOROOT"), "lib/wasi-libc") },
-	librarySources: func(target string) ([]string, error) {
+	librarySources: func(target string, _ bool) ([]string, error) {
 		return []string{
 			// memory builtins needed for llvm.memcpy.*, llvm.memmove.*, and
 			// llvm.memset.* LLVM intrinsics.
diff --git a/compileopts/config.go b/compileopts/config.go
@@ -22,7 +22,8 @@ import (
 // builder.Library struct but that's hard to do since we want to know the
 // library path in advance in several places).
 var libVersions = map[string]int{
-	"musl": 3,
+	"musl":  3,
+	"bdwgc": 2,
 }
 
 // Config keeps all configuration affecting the build in a single struct.
@@ -138,7 +139,7 @@ func (c *Config) GC() string {
 // that can be traced by the garbage collector.
 func (c *Config) NeedsStackObjects() bool {
 	switch c.GC() {
-	case "conservative", "custom", "precise":
+	case "conservative", "custom", "precise", "boehm":
 		for _, tag := range c.BuildTags() {
 			if tag == "tinygo.wasm" {
 				return true
@@ -263,6 +264,15 @@ func MuslArchitecture(triple string) string {
 	return CanonicalArchName(triple)
 }
 
+// Returns true if the libc needs to include malloc, for the libcs where this
+// matters.
+func (c *Config) LibcNeedsMalloc() bool {
+	if c.GC() == "boehm" && c.Target.Libc == "wasi-libc" {
+		return true
+	}
+	return false
+}
+
 // LibraryPath returns the path to the library build directory. The path will be
 // a library path in the cache directory (which might not yet be built).
 func (c *Config) LibraryPath(name string) string {
@@ -286,9 +296,14 @@ func (c *Config) LibraryPath(name string) string {
 		archname += "-v" + strconv.Itoa(v)
 	}
 
+	options := ""
+	if c.LibcNeedsMalloc() {
+		options += "+malloc"
+	}
+
 	// No precompiled library found. Determine the path name that will be used
 	// in the build cache.
-	return filepath.Join(goenv.Get("GOCACHE"), name+"-"+archname)
+	return filepath.Join(goenv.Get("GOCACHE"), name+options+"-"+archname)
 }
 
 // DefaultBinaryExtension returns the default extension for binaries, such as
diff --git a/compileopts/target.go b/compileopts/target.go
@@ -470,6 +470,12 @@ func defaultTarget(options *Options) (*TargetSpec, error) {
 		return nil, fmt.Errorf("unknown GOOS=%s", options.GOOS)
 	}
 
+	if spec.GC == "boehm" {
+		// Add this file only when needed. This fixes a build failure on
+		// Windows.
+		spec.ExtraFiles = append(spec.ExtraFiles, "src/runtime/gc_boehm.c")
+	}
+
 	// Target triples (which actually have four components, but are called
 	// triples for historical reasons) have the form:
 	//   arch-vendor-os-environment
diff --git a/go.mod b/go.mod
@@ -3,7 +3,7 @@ module github.com/tinygo-org/tinygo
 go 1.22.0
 
 require (
-	github.com/aykevl/go-wasm v0.0.2-0.20240825160117-b76c3f9f0982
+	github.com/aykevl/go-wasm v0.0.2-0.20250317121156-42b86c494139
 	github.com/blakesmith/ar v0.0.0-20150311145944-8bd4349a67f2
 	github.com/chromedp/cdproto v0.0.0-20220113222801-0725d94bb6ee
 	github.com/chromedp/chromedp v0.7.6
diff --git a/go.sum b/go.sum
@@ -1,5 +1,5 @@
-github.com/aykevl/go-wasm v0.0.2-0.20240825160117-b76c3f9f0982 h1:cD7QfvrJdYmBw2tFP/VyKPT8ZESlcrwSwo7SvH9Y4dc=
-github.com/aykevl/go-wasm v0.0.2-0.20240825160117-b76c3f9f0982/go.mod h1:7sXyiaA0WtSogCu67R2252fQpVmJMh9JWJ9ddtGkpWw=
+github.com/aykevl/go-wasm v0.0.2-0.20250317121156-42b86c494139 h1:2O/WuAt8J5id3khcAtVB90czG80m+v0sfkLE07GrCVg=
+github.com/aykevl/go-wasm v0.0.2-0.20250317121156-42b86c494139/go.mod h1:7sXyiaA0WtSogCu67R2252fQpVmJMh9JWJ9ddtGkpWw=
 github.com/blakesmith/ar v0.0.0-20150311145944-8bd4349a67f2 h1:oMCHnXa6CCCafdPDbMh/lWRhRByN0VFLvv+g+ayx1SI=
 github.com/blakesmith/ar v0.0.0-20150311145944-8bd4349a67f2/go.mod h1:PkYb9DJNAwrSvRx5DYA+gUcOIgTGVMNkfSCbZM8cWpI=
 github.com/chromedp/cdproto v0.0.0-20211126220118-81fa0469ad77/go.mod h1:At5TxYYdxkbQL0TSefRjhLE3Q0lgvqKKMSFUglJ7i1U=
diff --git a/lib/bdwgc b/lib/bdwgc
@@ -1 +1 @@
-Subproject commit d1ff06cc503a74dca0150d5e988f2c93158b0cdf
+Subproject commit 1166f11f7dee08d7ad369296b24cf8c9582f8789
diff --git a/main_test.go b/main_test.go
@@ -178,11 +178,27 @@ func TestBuild(t *testing.T) {
 
 		t.Run("WebAssembly", func(t *testing.T) {
 			t.Parallel()
+
 			runPlatTests(optionsFromTarget("wasm", sema), tests, t)
+			// Test with -gc=boehm.
+			t.Run("gc.go-boehm", func(t *testing.T) {
+				t.Parallel()
+				optionsBoehm := optionsFromTarget("wasm", sema)
+				optionsBoehm.GC = "boehm"
+				runTest("gc.go", optionsBoehm, t, nil, nil)
+			})
 		})
-		t.Run("WASI", func(t *testing.T) {
+		t.Run("WASIp1", func(t *testing.T) {
 			t.Parallel()
 			runPlatTests(optionsFromTarget("wasip1", sema), tests, t)
+
+			// Test with -gc=boehm.
+			t.Run("gc.go-boehm", func(t *testing.T) {
+				t.Parallel()
+				optionsBoehm := optionsFromTarget("wasip1", sema)
+				optionsBoehm.GC = "boehm"
+				runTest("gc.go", optionsBoehm, t, nil, nil)
+			})
 		})
 		t.Run("WASIp2", func(t *testing.T) {
 			t.Parallel()
diff --git a/src/runtime/arch_tinygowasm_malloc.go b/src/runtime/arch_tinygowasm_malloc.go
@@ -1,4 +1,4 @@
-//go:build tinygo.wasm && !(custommalloc || wasm_unknown)
+//go:build tinygo.wasm && !(custommalloc || wasm_unknown || gc.boehm)
 
 package runtime
 
diff --git a/src/runtime/gc_boehm.c b/src/runtime/gc_boehm.c
@@ -0,0 +1,37 @@
+//go:build none
+
+// This file is included in the build on systems that support the Boehm GC,
+// despite the //go:build line above.
+
+#include <stdint.h>
+
+typedef void (* GC_push_other_roots_proc)(void);
+void GC_set_push_other_roots(GC_push_other_roots_proc);
+
+typedef void(* GC_warn_proc)(const char *msg, uintptr_t arg);
+void GC_set_warn_proc(GC_warn_proc p);
+
+void tinygo_runtime_bdwgc_callback(void);
+
+static void callback(void) {
+    tinygo_runtime_bdwgc_callback();
+}
+
+static void warn_proc(const char *msg, uintptr_t arg) {
+}
+
+void tinygo_runtime_bdwgc_init(void) {
+    GC_set_push_other_roots(callback);
+#if defined(__wasm__)
+    // There are a lot of warnings on WebAssembly in the form:
+    //
+    //     GC Warning: Repeated allocation of very large block (appr. size 68 KiB):
+    //         May lead to memory leak and poor performance
+    //
+    // The usual advice is to use something like GC_malloc_ignore_off_page but
+    // unfortunately for most allocations that's not allowed: Go allocations can
+    // legitimately hold pointers further than one page in the allocation. So
+    // instead we just disable the warning.
+    GC_set_warn_proc(warn_proc);
+#endif
+}
diff --git a/src/runtime/gc_boehm.go b/src/runtime/gc_boehm.go
@@ -20,7 +20,6 @@ package runtime
 
 import (
 	"internal/gclayout"
-	"internal/reflectlite"
 	"internal/task"
 	"unsafe"
 )
@@ -39,22 +38,30 @@ var needsResumeWorld bool
 func initHeap() {
 	libgc_init()
 
-	libgc_set_push_other_roots(gcCallbackPtr)
+	// Call GC_set_push_other_roots(gcCallback) in C because of function
+	// signature differences that do matter in WebAssembly.
+	gcInit()
 }
 
-var gcCallbackPtr = reflectlite.ValueOf(gcCallback).UnsafePointer()
+//export tinygo_runtime_bdwgc_init
+func gcInit()
 
+//export tinygo_runtime_bdwgc_callback
 func gcCallback() {
 	// Mark globals and all stacks, and stop the world if we're using threading.
 	gcMarkReachable()
 
-	if needsResumeWorld {
-		// Should never happen, check for it anyway.
-		runtimePanic("gc: world already stopped")
-	}
+	// If we use a scheduler with parallelism (the threads scheduler for
+	// example), we need to call gcResumeWorld() after scanning has finished.
+	if hasParallelism {
+		if needsResumeWorld {
+			// Should never happen, check for it anyway.
+			runtimePanic("gc: world already stopped")
+		}
 
-	// Note that we need to resume the world after finishing the GC call.
-	needsResumeWorld = true
+		// Note that we need to resume the world after finishing the GC call.
+		needsResumeWorld = true
+	}
 }
 
 func markRoots(start, end uintptr) {
diff --git a/src/runtime/gc_stack_portable.go b/src/runtime/gc_stack_portable.go
@@ -1,4 +1,4 @@
-//go:build (gc.conservative || gc.custom || gc.precise) && tinygo.wasm
+//go:build (gc.conservative || gc.custom || gc.precise || gc.boehm) && tinygo.wasm
 
 package runtime
 
diff --git a/targets/wasip1.json b/targets/wasip1.json
@@ -23,7 +23,8 @@
 		"--no-demangle"
 	],
 	"extra-files": [
-		"src/runtime/asm_tinygowasm.S"
+		"src/runtime/asm_tinygowasm.S",
+		"src/runtime/gc_boehm.c"
 	],
 	"emulator":      "wasmtime run --dir={tmpDir}::/tmp {}"
 }
diff --git a/targets/wasm.json b/targets/wasm.json
@@ -24,7 +24,8 @@
 		"--no-demangle"
 	],
 	"extra-files": [
-		"src/runtime/asm_tinygowasm.S"
+		"src/runtime/asm_tinygowasm.S",
+		"src/runtime/gc_boehm.c"
 	],
 	"emulator":      "node {root}/targets/wasm_exec.js {}"
 }
diff --git a/tests/runtime_wasi/malloc_test.go b/tests/runtime_wasi/malloc_test.go
@@ -124,32 +124,3 @@ func TestMallocFree(t *testing.T) {
 		})
 	}
 }
-
-func TestMallocEmpty(t *testing.T) {
-	ptr := libc_malloc(0)
-	if ptr != nil {
-		t.Errorf("expected nil pointer, got %p", ptr)
-	}
-}
-
-func TestCallocEmpty(t *testing.T) {
-	ptr := libc_calloc(0, 1)
-	if ptr != nil {
-		t.Errorf("expected nil pointer, got %p", ptr)
-	}
-	ptr = libc_calloc(1, 0)
-	if ptr != nil {
-		t.Errorf("expected nil pointer, got %p", ptr)
-	}
-}
-
-func TestReallocEmpty(t *testing.T) {
-	ptr := libc_malloc(1)
-	if ptr == nil {
-		t.Error("expected pointer but was nil")
-	}
-	ptr = libc_realloc(ptr, 0)
-	if ptr != nil {
-		t.Errorf("expected nil pointer, got %p", ptr)
-	}
-}

Original file line number	Diff line number	Diff line change
`@@ -48,7 +48,7 @@ var libMinGW = Library{`
`48`	`48`	`}`
`49`	`49`	`return flags`
`50`	`50`	`},`
`51`		`- librarySources: func(target string) ([]string, error) {`
	`51`	`+ librarySources: func(target string, _ bool) ([]string, error) {`
`52`	`52`	`// These files are needed so that printf and the like are supported.`
`53`	`53`	`var sources []string`
`54`	`54`	`if strings.Split(target, "-")[0] == "i386" {`
Original file line number	Diff line number	Diff line change
`@@ -43,7 +43,7 @@ var libPicolibc = Library{`
`43`	`43`	`}`
`44`	`44`	`},`
`45`	`45`	`sourceDir: func() string { return filepath.Join(goenv.Get("TINYGOROOT"), "lib/picolibc/newlib") },`
`46`		`- librarySources: func(target string) ([]string, error) {`
	`46`	`+ librarySources: func(target string, _ bool) ([]string, error) {`
`47`	`47`	`sources := append([]string(nil), picolibcSources...)`
`48`	`48`	`if !strings.HasPrefix(target, "avr") {`
`49`	`49`	`// Small chips without long jumps can't compile many files (printf,`
Original file line number	Diff line number	Diff line change
`@@ -41,7 +41,7 @@ var libWasmBuiltins = Library{`
`41`	`41`	`}`
`42`	`42`	`},`
`43`	`43`	`sourceDir: func() string { return filepath.Join(goenv.Get("TINYGOROOT"), "lib/wasi-libc") },`
`44`		`- librarySources: func(target string) ([]string, error) {`
	`44`	`+ librarySources: func(target string, _ bool) ([]string, error) {`
`45`	`45`	`return []string{`
`46`	`46`	`// memory builtins needed for llvm.memcpy., llvm.memmove., and`
`47`	`47`	`// llvm.memset.* LLVM intrinsics.`