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:
#2870
Once interface type asserts and method calls are converted to a
vtable-like implementation, most programs should just work on
linux/darwin/windows.

Author: aykevl

builder/build.go

@@ -175,6 +175,7 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 		DefaultStackSize:   config.StackSize(),
 		NeedsStackObjects:  config.NeedsStackObjects(),
 		Debug:              true,
+		LTO:                config.Options.LTO != "legacy",
 	}
 
 	// Load the target machine, which is the LLVM object that contains all
@@ -609,8 +610,38 @@ 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.Options.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/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 +109,10 @@ func (o *Options) Verify() error {
 		}
 	}
 
+	if !isInArray(validLTOOptions, o.LTO) {
+		return fmt.Errorf("invalid -lto=%s: valid values are %s", o.LTO, strings.Join(validLTOOptions, ", "))
+	}
+
 	return nil
 }
 

compiler/compiler.go

@@ -54,6 +54,7 @@ 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

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)