builder: add real ThinLTO mode

We use ThinLTO for linking, but we use it in a way that doesn't give
most of its benefits: we merge all the bitcode files into a single LLVM
module and run some optimizations on it before linking. Therefore, this
works more like a traditional "full" LTO link rather than a true thin
link.

This commit adds a new experimental -lto=thin option to do a true
ThinLTO link. The main benefit is that linking will be a lot faster,
especially for large programs consisting of many packages.

At the moment, it only works for programs that don't do interface type
asserts and don't call interface methods. It also probably won't work on
WebAssembly and baremetal systems. But it's part of a larger goal
towards a truly incremental build system:
https://github.com/tinygo-org/tinygo/issues/2870
Once interface type asserts and method calls are converted to a
vtable-like implementation, most programs should just work on
linux/darwin/windows.

builder/build.go

@@ -176,6 +176,7 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 		DefaultStackSize:   config.StackSize(),
 		NeedsStackObjects:  config.NeedsStackObjects(),
 		Debug:              true,
+		LTO:                config.LTO() != "legacy",
 	}
 
 	// Load the target machine, which is the LLVM object that contains all
@@ -452,18 +453,24 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 				if err != nil {
 					return err
 				}
-				if runtime.GOOS == "windows" {
-					// Work around a problem on Windows.
-					// For some reason, WriteBitcodeToFile causes TinyGo to
-					// exit with the following message:
-					//   LLVM ERROR: IO failure on output stream: Bad file descriptor
-					buf := llvm.WriteBitcodeToMemoryBuffer(mod)
+				if compilerConfig.LTO {
+					buf := llvm.WriteThinLTOBitcodeToMemoryBuffer(mod)
 					defer buf.Dispose()
 					_, err = f.Write(buf.Bytes())
 				} else {
-					// Otherwise, write bitcode directly to the file (probably
-					// faster).
-					err = llvm.WriteBitcodeToFile(mod, f)
+					if runtime.GOOS == "windows" {
+						// Work around a problem on Windows.
+						// For some reason, WriteBitcodeToFile causes TinyGo to
+						// exit with the following message:
+						//   LLVM ERROR: IO failure on output stream: Bad file descriptor
+						buf := llvm.WriteBitcodeToMemoryBuffer(mod)
+						defer buf.Dispose()
+						_, err = f.Write(buf.Bytes())
+					} else {
+						// Otherwise, write bitcode directly to the file (probably
+						// faster).
+						err = llvm.WriteBitcodeToFile(mod, f)
+					}
 				}
 				if err != nil {
 					// WriteBitcodeToFile doesn't produce a useful error on its
@@ -609,8 +616,38 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 		},
 	}
 
+	// Add job to create the runtime.initAll function in a new module.
+	initAllJob := &compileJob{
+		description: "create runtime.initAll",
+		result:      filepath.Join(tmpdir, "runtime-initAll.bc"),
+		run: func(job *compileJob) (err error) {
+			// Create module with runtime.initAll.
+			ctx := llvm.NewContext()
+			defer ctx.Dispose()
+			initAllMod := createInitAll(ctx, config, compilerConfig, lprogram.Sorted())
+			defer initAllMod.Dispose()
+
+			// Write module to bitcode file.
+			llvmBuf := llvm.WriteThinLTOBitcodeToMemoryBuffer(initAllMod)
+			defer llvmBuf.Dispose()
+			return os.WriteFile(job.result, llvmBuf.Bytes(), 0666)
+		},
+	}
+
 	// Prepare link command.
-	linkerDependencies := []*compileJob{outputObjectFileJob}
+	var linkerDependencies []*compileJob
+	switch config.LTO() {
+	case "legacy":
+		// Link all Go bitcode files together into a single large module and
+		// then do a ThinLTO link with the resulting large module + extra C
+		// files (from CGo etc).
+		linkerDependencies = append(linkerDependencies, outputObjectFileJob)
+	case "thin":
+		// Do a real thin link, with each Go package in a separate translation
+		// unit. This is faster than merging them into one big LTO module.
+		linkerDependencies = append(linkerDependencies, packageJobs...)
+		linkerDependencies = append(linkerDependencies, initAllJob)
+	}
 	result.Executable = filepath.Join(tmpdir, "main")
 	if config.GOOS() == "windows" {
 		result.Executable += ".exe"

compileopts/config.go

@@ -165,6 +165,14 @@ func (c *Config) OptLevels() (optLevel, sizeLevel int, inlinerThreshold uint) {
 	}
 }
 
+// LTO returns one of the possible LTO configurations: legacy or thin.
+func (c *Config) LTO() string {
+	if c.Options.LTO != "" {
+		return c.Options.LTO
+	}
+	return "legacy"
+}
+
 // PanicStrategy returns the panic strategy selected for this target. Valid
 // values are "print" (print the panic value, then exit) or "trap" (issue a trap
 // instruction).

compileopts/options.go

@@ -14,6 +14,7 @@ var (
 	validPrintSizeOptions     = []string{"none", "short", "full"}
 	validPanicStrategyOptions = []string{"print", "trap"}
 	validOptOptions           = []string{"none", "0", "1", "2", "s", "z"}
+	validLTOOptions           = []string{"legacy", "thin"}
 )
 
 // Options contains extra options to give to the compiler. These options are
@@ -25,6 +26,7 @@ type Options struct {
 	GOARM           string // environment variable (only used with GOARCH=arm)
 	Target          string
 	Opt             string
+	LTO             string
 	GC              string
 	PanicStrategy   string
 	Scheduler       string
@@ -107,6 +109,12 @@ func (o *Options) Verify() error {
 		}
 	}
 
+	if o.LTO != "" {
+		if !isInArray(validLTOOptions, o.LTO) {
+			return fmt.Errorf("invalid -lto=%s: valid values are %s", o.LTO, strings.Join(validLTOOptions, ", "))
+		}
+	}
+
 	return nil
 }
 

compiler/alias.go

@@ -34,7 +34,9 @@ var stdlibAliases = map[string]string{
 // createAlias implements the function (in the builder) as a call to the alias
 // function.
 func (b *builder) createAlias(alias llvm.Value) {
-	b.llvmFn.SetVisibility(llvm.HiddenVisibility)
+	if !b.LTO {
+		b.llvmFn.SetVisibility(llvm.HiddenVisibility)
+	}
 	b.llvmFn.SetUnnamedAddr(true)
 
 	if b.Debug {

compiler/compiler.go

@@ -55,6 +55,7 @@ type Config struct {
 	DefaultStackSize   uint64
 	NeedsStackObjects  bool
 	Debug              bool // Whether to emit debug information in the LLVM module.
+	LTO                bool // non-legacy LTO (meaning: package bitcode is merged by the linker)
 }
 
 // compilerContext contains function-independent data that should still be
@@ -897,7 +898,9 @@ func (c *compilerContext) createPackage(irbuilder llvm.Builder, pkg *ssa.Package
 				c.createEmbedGlobal(member, global, files)
 			} else if !info.extern {
 				global.SetInitializer(llvm.ConstNull(global.GlobalValueType()))
-				global.SetVisibility(llvm.HiddenVisibility)
+				if !c.LTO {
+					global.SetVisibility(llvm.HiddenVisibility)
+				}
 				if info.section != "" {
 					global.SetSection(info.section)
 				}
@@ -944,7 +947,9 @@ func (c *compilerContext) createEmbedGlobal(member *ssa.Global, global llvm.Valu
 		}
 		strObj := c.getEmbedFileString(files[0])
 		global.SetInitializer(strObj)
-		global.SetVisibility(llvm.HiddenVisibility)
+		if !c.LTO {
+			global.SetVisibility(llvm.HiddenVisibility)
+		}
 
 	case *types.Slice:
 		if typ.Elem().Underlying().(*types.Basic).Kind() != types.Byte {
@@ -968,7 +973,9 @@ func (c *compilerContext) createEmbedGlobal(member *ssa.Global, global llvm.Valu
 		sliceLen := llvm.ConstInt(c.uintptrType, file.Size, false)
 		sliceObj := c.ctx.ConstStruct([]llvm.Value{slicePtr, sliceLen, sliceLen}, false)
 		global.SetInitializer(sliceObj)
-		global.SetVisibility(llvm.HiddenVisibility)
+		if !c.LTO {
+			global.SetVisibility(llvm.HiddenVisibility)
+		}
 
 	case *types.Struct:
 		// Assume this is an embed.FS struct:
@@ -1051,7 +1058,9 @@ func (c *compilerContext) createEmbedGlobal(member *ssa.Global, global llvm.Valu
 		globalInitializer := llvm.ConstNull(c.getLLVMType(member.Type().(*types.Pointer).Elem()))
 		globalInitializer = c.builder.CreateInsertValue(globalInitializer, sliceGlobal, 0, "")
 		global.SetInitializer(globalInitializer)
-		global.SetVisibility(llvm.HiddenVisibility)
+		if !c.LTO {
+			global.SetVisibility(llvm.HiddenVisibility)
+		}
 		global.SetAlignment(c.targetData.ABITypeAlignment(globalInitializer.Type()))
 	}
 }
@@ -1098,7 +1107,8 @@ func (b *builder) createFunctionStart(intrinsic bool) {
 		// assertion error in llvm-project/llvm/include/llvm/IR/GlobalValue.h:236
 		// is thrown.
 		if b.llvmFn.Linkage() != llvm.InternalLinkage &&
-			b.llvmFn.Linkage() != llvm.PrivateLinkage {
+			b.llvmFn.Linkage() != llvm.PrivateLinkage &&
+			!b.LTO {
 			b.llvmFn.SetVisibility(llvm.HiddenVisibility)
 		}
 		b.llvmFn.SetUnnamedAddr(true)

compiler/interface.go

@@ -490,17 +490,22 @@ func (b *builder) createTypeAssert(expr *ssa.TypeAssert) llvm.Value {
 		commaOk = b.CreateCall(fn.GlobalValueType(), fn, []llvm.Value{actualTypeNum}, "")
 
 	} else {
-		globalName := "reflect/types.typeid:" + getTypeCodeName(expr.AssertedType)
-		assertedTypeCodeGlobal := b.mod.NamedGlobal(globalName)
-		if assertedTypeCodeGlobal.IsNil() {
-			// Create a new typecode global.
-			assertedTypeCodeGlobal = llvm.AddGlobal(b.mod, b.ctx.Int8Type(), globalName)
-			assertedTypeCodeGlobal.SetGlobalConstant(true)
+		if b.LTO {
+			assertedTypeCodeGlobal := b.getTypeCode(expr.AssertedType)
+			commaOk = b.CreateICmp(llvm.IntEQ, actualTypeNum, assertedTypeCodeGlobal, "commaok")
+		} else {
+			globalName := "reflect/types.typeid:" + getTypeCodeName(expr.AssertedType)
+			assertedTypeCodeGlobal := b.mod.NamedGlobal(globalName)
+			if assertedTypeCodeGlobal.IsNil() {
+				// Create a new typecode global.
+				assertedTypeCodeGlobal = llvm.AddGlobal(b.mod, b.ctx.Int8Type(), globalName)
+				assertedTypeCodeGlobal.SetGlobalConstant(true)
+			}
+			// Type assert on concrete type.
+			// Call runtime.typeAssert, which will be lowered to a simple icmp
+			// or const false in the interface lowering pass.
+			commaOk = b.createRuntimeCall("typeAssert", []llvm.Value{actualTypeNum, assertedTypeCodeGlobal}, "typecode")
 		}
-		// Type assert on concrete type.
-		// Call runtime.typeAssert, which will be lowered to a simple icmp or
-		// const false in the interface lowering pass.
-		commaOk = b.createRuntimeCall("typeAssert", []llvm.Value{actualTypeNum, assertedTypeCodeGlobal}, "typecode")
 	}
 
 	// Add 2 new basic blocks (that should get optimized away): one for the

compiler/interrupt.go

@@ -45,7 +45,9 @@ func (b *builder) createInterruptGlobal(instr *ssa.CallCommon) (llvm.Value, erro
 	globalLLVMType := b.getLLVMType(globalType)
 	globalName := b.fn.Package().Pkg.Path() + "$interrupt" + strconv.FormatInt(id.Int64(), 10)
 	global := llvm.AddGlobal(b.mod, globalLLVMType, globalName)
-	global.SetVisibility(llvm.HiddenVisibility)
+	if !b.LTO {
+		global.SetVisibility(llvm.HiddenVisibility)
+	}
 	global.SetGlobalConstant(true)
 	global.SetUnnamedAddr(true)
 	initializer := llvm.ConstNull(globalLLVMType)

main.go

@@ -1357,6 +1357,7 @@ func main() {
 	command := os.Args[1]
 
 	opt := flag.String("opt", "z", "optimization level: 0, 1, 2, s, z")
+	lto := flag.String("lto", "legacy", "LTO mode: legacy or thin")
 	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 (none, tasks, asyncify)")
@@ -1459,6 +1460,7 @@ func main() {
 		Target:          *target,
 		StackSize:       stackSize,
 		Opt:             *opt,
+		LTO:             *lto,
 		GC:              *gc,
 		PanicStrategy:   *panicStrategy,
 		Scheduler:       *scheduler,

main_test.go

@@ -105,6 +105,13 @@ func TestBuild(t *testing.T) {
 			runTestWithConfig("print.go", t, opts, nil, nil)
 		})
 
+		t.Run("lto=thin", func(t *testing.T) {
+			t.Parallel()
+			opts := optionsFromTarget("", sema)
+			opts.LTO = "thin"
+			runTestWithConfig("init.go", t, opts, nil, nil)
+		})
+
 		t.Run("ldflags", func(t *testing.T) {
 			t.Parallel()
 			opts := optionsFromTarget("", sema)