windows: add windows/386 support

.github/workflows/build-macos.yml

@@ -101,7 +101,7 @@ jobs:
       - name: make gen-device
         run: make -j3 gen-device
       - name: Test TinyGo
-        run: make test GOTESTFLAGS="-short"
+        run: make test GOTESTFLAGS="-only-current-os"
       - name: Build TinyGo release tarball
         run: make release -j3
       - name: Test stdlib packages

.github/workflows/windows.yml

@@ -114,7 +114,7 @@ jobs:
         run: make -j3 gen-device
       - name: Test TinyGo
         shell: bash
-        run: make test GOTESTFLAGS="-short"
+        run: make test GOTESTFLAGS="-only-current-os"
       - name: Build TinyGo release tarball
         shell: bash
         run: make build/release -j4

builder/builder_test.go

@@ -67,6 +67,7 @@ func TestClangAttributes(t *testing.T) {
 		{GOOS: "linux", GOARCH: "mipsle", GOMIPS: "softfloat"},
 		{GOOS: "darwin", GOARCH: "amd64"},
 		{GOOS: "darwin", GOARCH: "arm64"},
+		{GOOS: "windows", GOARCH: "386"},
 		{GOOS: "windows", GOARCH: "amd64"},
 		{GOOS: "windows", GOARCH: "arm64"},
 	} {

builder/mingw-w64.go

@@ -93,6 +93,9 @@ func makeMinGWExtraLibs(tmpdir, goarch string) []*compileJob {
 				defpath := inpath
 				var archDef, emulation string
 				switch goarch {
+				case "386":
+					archDef = "-DDEF_I386"
+					emulation = "i386pe"
 				case "amd64":
 					archDef = "-DDEF_X64"
 					emulation = "i386pep"

compileopts/target.go

@@ -430,14 +430,20 @@ func defaultTarget(options *Options) (*TargetSpec, error) {
 		spec.GC = "boehm"
 		spec.Linker = "ld.lld"
 		spec.Libc = "mingw-w64"
-		// Note: using a medium code model, low image base and no ASLR
-		// because Go doesn't really need those features. ASLR patches
-		// around issues for unsafe languages like C/C++ that are not
-		// normally present in Go (without explicitly opting in).
-		// For more discussion:
-		// https://groups.google.com/g/Golang-nuts/c/Jd9tlNc6jUE/m/Zo-7zIP_m3MJ?pli=1
 		switch options.GOARCH {
+		case "386":
+			spec.LDFlags = append(spec.LDFlags,
+				"-m", "i386pe",
+			)
+			// __udivdi3 is not present in ucrt it seems.
+			spec.RTLib = "compiler-rt"
 		case "amd64":
+			// Note: using a medium code model, low image base and no ASLR
+			// because Go doesn't really need those features. ASLR patches
+			// around issues for unsafe languages like C/C++ that are not
+			// normally present in Go (without explicitly opting in).
+			// For more discussion:
+			// https://groups.google.com/g/Golang-nuts/c/Jd9tlNc6jUE/m/Zo-7zIP_m3MJ?pli=1
 			spec.LDFlags = append(spec.LDFlags,
 				"-m", "i386pep",
 				"--image-base", "0x400000",

compiler/calls.go

@@ -76,7 +76,15 @@ func (b *builder) createCall(fnType llvm.Type, fn llvm.Value, args []llvm.Value,
 		fragments := b.expandFormalParam(arg)
 		expanded = append(expanded, fragments...)
 	}
-	return b.CreateCall(fnType, fn, expanded, name)
+	call := b.CreateCall(fnType, fn, expanded, name)
+	if !fn.IsAFunction().IsNil() {
+		if cc := fn.FunctionCallConv(); cc != llvm.CCallConv {
+			// Set a different calling convention if needed.
+			// This is needed for GetModuleHandleExA on Windows, for example.
+			call.SetInstructionCallConv(cc)
+		}
+	}
+	return call
 }
 
 // createInvoke is like createCall but continues execution at the landing pad if

compiler/llvm.go

@@ -452,6 +452,21 @@ func (b *builder) readStackPointer() llvm.Value {
 	return b.CreateCall(stacksave.GlobalValueType(), stacksave, nil, "")
 }
 
+// writeStackPointer emits a LLVM intrinsic call that updates the current stack
+// pointer.
+func (b *builder) writeStackPointer(sp llvm.Value) {
+	name := "llvm.stackrestore.p0"
+	if llvmutil.Version() < 18 {
+		name = "llvm.stackrestore" // backwards compatibility with LLVM 17 and below
+	}
+	stackrestore := b.mod.NamedFunction(name)
+	if stackrestore.IsNil() {
+		fnType := llvm.FunctionType(b.ctx.VoidType(), []llvm.Type{b.dataPtrType}, false)
+		stackrestore = llvm.AddFunction(b.mod, name, fnType)
+	}
+	b.CreateCall(stackrestore.GlobalValueType(), stackrestore, []llvm.Value{sp}, "")
+}
+
 // createZExtOrTrunc lets the input value fit in the output type bits, by zero
 // extending or truncating the integer.
 func (b *builder) createZExtOrTrunc(value llvm.Value, t llvm.Type) llvm.Value {

compiler/symbol.go

@@ -208,6 +208,12 @@ func (c *compilerContext) getFunction(fn *ssa.Function) (llvm.Type, llvm.Value)
 			// > circumstances, and should not be exposed to source languages.
 			llvmutil.AppendToGlobal(c.mod, "llvm.compiler.used", llvmFn)
 		}
+	case "GetModuleHandleExA", "GetProcAddress", "GetSystemInfo", "GetSystemTimeAsFileTime", "LoadLibraryExW", "QueryUnbiasedInterruptTime", "SetEnvironmentVariableA", "Sleep", "VirtualAlloc":
+		// On Windows we need to use a special calling convention for some
+		// external calls.
+		if c.GOOS == "windows" && c.GOARCH == "386" {
+			llvmFn.SetFunctionCallConv(llvm.X86StdcallCallConv)
+		}
 	}
 
 	// External/exported functions may not retain pointer values.

compiler/syscall.go

@@ -268,6 +268,8 @@ func (b *builder) createSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 		// The signature looks like this:
 		//   func Syscall(trap, nargs, a1, a2, a3 uintptr) (r1, r2 uintptr, err Errno)
 
+		isI386 := strings.HasPrefix(b.Triple, "i386-")
+
 		// Prepare input values.
 		var paramTypes []llvm.Type
 		var params []llvm.Value
@@ -285,11 +287,17 @@ func (b *builder) createSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 		if setLastError.IsNil() {
 			llvmType := llvm.FunctionType(b.ctx.VoidType(), []llvm.Type{b.ctx.Int32Type()}, false)
 			setLastError = llvm.AddFunction(b.mod, "SetLastError", llvmType)
+			if isI386 {
+				setLastError.SetFunctionCallConv(llvm.X86StdcallCallConv)
+			}
 		}
 		getLastError := b.mod.NamedFunction("GetLastError")
 		if getLastError.IsNil() {
 			llvmType := llvm.FunctionType(b.ctx.Int32Type(), nil, false)
 			getLastError = llvm.AddFunction(b.mod, "GetLastError", llvmType)
+			if isI386 {
+				getLastError.SetFunctionCallConv(llvm.X86StdcallCallConv)
+			}
 		}
 
 		// Now do the actual call. Pseudocode:
@@ -300,9 +308,24 @@ func (b *builder) createSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 		// Note that SetLastError/GetLastError could be replaced with direct
 		// access to the thread control block, which is probably smaller and
 		// faster. The Go runtime does this in assembly.
-		b.CreateCall(setLastError.GlobalValueType(), setLastError, []llvm.Value{llvm.ConstNull(b.ctx.Int32Type())}, "")
+		// On windows/386, we also need to save/restore the stack pointer. I'm
+		// not entirely sure why this is needed, but without it these calls
+		// change the stack pointer leading to a crash soon after.
+		setLastErrorCall := b.CreateCall(setLastError.GlobalValueType(), setLastError, []llvm.Value{llvm.ConstNull(b.ctx.Int32Type())}, "")
+		var sp llvm.Value
+		if isI386 {
+			setLastErrorCall.SetInstructionCallConv(llvm.X86StdcallCallConv)
+			sp = b.readStackPointer()
+		}
 		syscallResult := b.CreateCall(llvmType, fnPtr, params, "")
+		if isI386 {
+			syscallResult.SetInstructionCallConv(llvm.X86StdcallCallConv)
+			b.writeStackPointer(sp)
+		}
 		errResult := b.CreateCall(getLastError.GlobalValueType(), getLastError, nil, "err")
+		if isI386 {
+			errResult.SetInstructionCallConv(llvm.X86StdcallCallConv)
+		}
 		if b.uintptrType != b.ctx.Int32Type() {
 			errResult = b.CreateZExt(errResult, b.uintptrType, "err.uintptr")
 		}

main_test.go

@@ -35,6 +35,8 @@ const TESTDATA = "testdata"
 
 var testTarget = flag.String("target", "", "override test target")
 
+var testOnlyCurrentOS = flag.Bool("only-current-os", false, "")
+
 var supportedLinuxArches = map[string]string{
 	"AMD64Linux": "linux/amd64",
 	"X86Linux":   "linux/386",
@@ -158,20 +160,35 @@ func TestBuild(t *testing.T) {
 		return
 	}
 
-	t.Run("EmulatedCortexM3", func(t *testing.T) {
+	if !*testOnlyCurrentOS {
-		t.Parallel()
+		t.Run("EmulatedCortexM3", func(t *testing.T) {
-		runPlatTests(optionsFromTarget("cortex-m-qemu", sema), tests, t)
+			t.Parallel()
-	})
+			runPlatTests(optionsFromTarget("cortex-m-qemu", sema), tests, t)
+		})
 
-	t.Run("EmulatedRISCV", func(t *testing.T) {
+		t.Run("EmulatedRISCV", func(t *testing.T) {
-		t.Parallel()
+			t.Parallel()
-		runPlatTests(optionsFromTarget("riscv-qemu", sema), tests, t)
+			runPlatTests(optionsFromTarget("riscv-qemu", sema), tests, t)
-	})
+		})
 
-	t.Run("AVR", func(t *testing.T) {
+		t.Run("AVR", func(t *testing.T) {
-		t.Parallel()
+			t.Parallel()
-		runPlatTests(optionsFromTarget("simavr", sema), tests, t)
+			runPlatTests(optionsFromTarget("simavr", sema), tests, t)
-	})
+		})
+
+		t.Run("WebAssembly", func(t *testing.T) {
+			t.Parallel()
+			runPlatTests(optionsFromTarget("wasm", sema), tests, t)
+		})
+		t.Run("WASI", func(t *testing.T) {
+			t.Parallel()
+			runPlatTests(optionsFromTarget("wasip1", sema), tests, t)
+		})
+		t.Run("WASIp2", func(t *testing.T) {
+			t.Parallel()
+			runPlatTests(optionsFromTarget("wasip2", sema), tests, t)
+		})
+	}
 
 	if runtime.GOOS == "linux" {
 		for name, osArch := range supportedLinuxArches {
@@ -191,18 +208,13 @@ func TestBuild(t *testing.T) {
 			options := optionsFromOSARCH("linux/mipsle/softfloat", sema)
 			runTest("cgo/", options, t, nil, nil)
 		})
-		t.Run("WebAssembly", func(t *testing.T) {
+	} else if runtime.GOOS == "windows" {
-			t.Parallel()
+		if runtime.GOARCH != "386" {
-			runPlatTests(optionsFromTarget("wasm", sema), tests, t)
+			t.Run("Windows386", func(t *testing.T) {
-		})
+				t.Parallel()
-		t.Run("WASI", func(t *testing.T) {
+				runPlatTests(optionsFromOSARCH("windows/386", sema), tests, t)
-			t.Parallel()
+			})
-			runPlatTests(optionsFromTarget("wasip1", sema), tests, t)
+		}
-		})
-		t.Run("WASIp2", func(t *testing.T) {
-			t.Parallel()
-			runPlatTests(optionsFromTarget("wasip2", sema), tests, t)
-		})
 	}
 }
 

src/internal/task/task_stack_386.S

@@ -1,7 +1,12 @@
+#ifdef _WIN32
+.global  _tinygo_startTask
+_tinygo_startTask:
+#else // Linux etc
 .section .text.tinygo_startTask
 .global  tinygo_startTask
 .type    tinygo_startTask, %function
 tinygo_startTask:
+#endif
     .cfi_startproc
     // Small assembly stub for starting a goroutine. This is already run on the
     // new stack, with the callee-saved registers already loaded.
@@ -24,12 +29,21 @@ tinygo_startTask:
     addl $4, %esp
 
     // After return, exit this goroutine. This is a tail call.
+    #ifdef _WIN32
+    jmp _tinygo_pause
+    #else
     jmp tinygo_pause
+    #endif
     .cfi_endproc
 
+#ifdef _WIN32
+.global _tinygo_swapTask
+_tinygo_swapTask:
+#else
 .global tinygo_swapTask
 .type tinygo_swapTask, %function
 tinygo_swapTask:
+#endif
     // This function gets the following parameters:
     movl 4(%esp), %eax // newStack uintptr
     movl 8(%esp), %ecx // oldStack *uintptr

src/runtime/asm_386.S

@@ -1,7 +1,12 @@
+#ifdef _WIN32
+.global _tinygo_scanCurrentStack
+_tinygo_scanCurrentStack:
+#else
 .section .text.tinygo_scanCurrentStack
 .global tinygo_scanCurrentStack
 .type tinygo_scanCurrentStack, %function
 tinygo_scanCurrentStack:
+#endif
     // Sources:
     //   * https://stackoverflow.com/questions/18024672/what-registers-are-preserved-through-a-linux-x86-64-function-call
     //   * https://godbolt.org/z/q7e8dn
@@ -15,7 +20,11 @@ tinygo_scanCurrentStack:
     // Scan the stack.
     subl $8, %esp // adjust the stack before the call to maintain 16-byte alignment
     pushl %esp
+    #ifdef _WIN32
+    calll _tinygo_scanstack
+    #else
     calll tinygo_scanstack
+    #endif
 
     // Restore the stack pointer. Registers do not need to be restored as they
     // were only pushed to be discoverable by the GC.
@@ -23,9 +32,14 @@ tinygo_scanCurrentStack:
     retl
 
 
+#ifdef _WIN32
+.global _tinygo_longjmp
+_tinygo_longjmp:
+#else
 .section .text.tinygo_longjmp
 .global tinygo_longjmp
 tinygo_longjmp:
+#endif
     // Note: the code we jump to assumes eax is set to a non-zero value if we
     // jump from here.
     movl 4(%esp), %eax