ci: changed to generate only files that exist in targets

Signed-off-by: deadprogram <ron@hybridgroup.com>

.circleci/config.yml

@@ -177,6 +177,11 @@ commands:
       - run:
           name: "Test TinyGo"
           command: make ASSERT=1 test
+          environment:
+            # Note: -p=2 limits parallelism to two jobs at a time, which is
+            # necessary to keep memory consumption down and avoid OOM (for a
+            # 2CPU/4GB executor).
+            GOFLAGS: -p=2
       - save_cache:
           key: go-cache-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_BUILD_NUM }}
           paths:

.gitignore

@@ -19,3 +19,10 @@ src/device/kendryte/*.s
 vendor
 llvm-build
 llvm-project
+
+# Ignore files generated by smoketest
+test.gba
+test.hex
+test.nro
+test.wasm
+wasm.wasm

Makefile

@@ -107,7 +107,10 @@ fmt-check:
 	@unformatted=$$(gofmt -l $(FMT_PATHS)); [ -z "$$unformatted" ] && exit 0; echo "Unformatted:"; for fn in $$unformatted; do echo "  $$fn"; done; exit 1
 
 
-gen-device: gen-device-avr gen-device-esp gen-device-nrf gen-device-sam gen-device-sifive gen-device-stm32 gen-device-kendryte gen-device-nxp
+gen-device: gen-device-avr gen-device-esp gen-device-nrf gen-device-sam gen-device-sifive gen-device-kendryte gen-device-nxp
+ifneq ($(STM32), 0)
+gen-device: gen-device-stm32
+endif
 
 gen-device-avr:
 	@if [ ! -e lib/avr/README.md ]; then echo "Submodules have not been downloaded. Please download them using:\n  git submodule update --init"; exit 1; fi
@@ -117,22 +120,22 @@ gen-device-avr:
 	@GO111MODULE=off $(GO) fmt ./src/device/avr
 
 build/gen-device-svd: ./tools/gen-device-svd/*.go
-	$(GO) build -o $@ ./tools/gen-device-svd/
+	CGO_CPPFLAGS="$(CGO_CPPFLAGS)" CGO_CXXFLAGS="$(CGO_CXXFLAGS)" CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GO) build -o $@ ./tools/gen-device-svd/
 
 gen-device-esp: build/gen-device-svd
-	./build/gen-device-svd -source=https://github.com/posborne/cmsis-svd/tree/master/data/Espressif-Community -interrupts=software lib/cmsis-svd/data/Espressif-Community/ src/device/esp/
+	./build/gen-device-svd -source=https://github.com/posborne/cmsis-svd/tree/master/data/Espressif-Community -only-used -interrupts=software lib/cmsis-svd/data/Espressif-Community/ src/device/esp/
 	GO111MODULE=off $(GO) fmt ./src/device/esp
 
 gen-device-nrf: build/gen-device-svd
-	./build/gen-device-svd -source=https://github.com/NordicSemiconductor/nrfx/tree/master/mdk lib/nrfx/mdk/ src/device/nrf/
+	./build/gen-device-svd -source=https://github.com/NordicSemiconductor/nrfx/tree/master/mdk -only-used lib/nrfx/mdk/ src/device/nrf/
 	GO111MODULE=off $(GO) fmt ./src/device/nrf
 
 gen-device-nxp: build/gen-device-svd
-	./build/gen-device-svd -source=https://github.com/posborne/cmsis-svd/tree/master/data/NXP lib/cmsis-svd/data/NXP/ src/device/nxp/
+	./build/gen-device-svd -source=https://github.com/posborne/cmsis-svd/tree/master/data/NXP -only-used lib/cmsis-svd/data/NXP/ src/device/nxp/
 	GO111MODULE=off $(GO) fmt ./src/device/nxp
 
 gen-device-sam: build/gen-device-svd
-	./build/gen-device-svd -source=https://github.com/posborne/cmsis-svd/tree/master/data/Atmel lib/cmsis-svd/data/Atmel/ src/device/sam/
+	./build/gen-device-svd -source=https://github.com/posborne/cmsis-svd/tree/master/data/Atmel -only-used lib/cmsis-svd/data/Atmel/ src/device/sam/
 	GO111MODULE=off $(GO) fmt ./src/device/sam
 
 gen-device-sifive: build/gen-device-svd
@@ -140,11 +143,11 @@ gen-device-sifive: build/gen-device-svd
 	GO111MODULE=off $(GO) fmt ./src/device/sifive
 
 gen-device-kendryte: build/gen-device-svd
-	./build/gen-device-svd -source=https://github.com/posborne/cmsis-svd/tree/master/data/Kendryte-Community -interrupts=software lib/cmsis-svd/data/Kendryte-Community/ src/device/kendryte/
+	./build/gen-device-svd -source=https://github.com/posborne/cmsis-svd/tree/master/data/Kendryte-Community -only-used -interrupts=software lib/cmsis-svd/data/Kendryte-Community/ src/device/kendryte/
 	GO111MODULE=off $(GO) fmt ./src/device/kendryte
 
 gen-device-stm32: build/gen-device-svd
-	./build/gen-device-svd -source=https://github.com/tinygo-org/stm32-svd lib/stm32-svd/svd src/device/stm32/
+	./build/gen-device-svd -source=https://github.com/tinygo-org/stm32-svd -only-used lib/stm32-svd/svd src/device/stm32/
 	GO111MODULE=off $(GO) fmt ./src/device/stm32
 
 
@@ -177,7 +180,7 @@ tinygo:
 	CGO_CPPFLAGS="$(CGO_CPPFLAGS)" CGO_CXXFLAGS="$(CGO_CXXFLAGS)" CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GO) build -buildmode exe -o build/tinygo$(EXE) -tags byollvm -ldflags="-X main.gitSha1=`git rev-parse --short HEAD`" .
 
 test: wasi-libc
-	CGO_CPPFLAGS="$(CGO_CPPFLAGS)" CGO_CXXFLAGS="$(CGO_CXXFLAGS)" CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GO) test -v -buildmode exe -tags byollvm ./cgo ./compileopts ./compiler ./interp ./transform .
+	CGO_CPPFLAGS="$(CGO_CPPFLAGS)" CGO_CXXFLAGS="$(CGO_CXXFLAGS)" CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GO) test -v -buildmode exe -tags byollvm ./builder ./cgo ./compileopts ./compiler ./interp ./transform .
 
 # Test known-working standard library packages.
 # TODO: do this in one command, parallelize, and only show failing tests (no
@@ -257,8 +260,6 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=pca10031            examples/blinky1
 	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=bluepill            examples/blinky1
-	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=reelboard           examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=reelboard           examples/blinky2
@@ -267,6 +268,10 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=pca10056            examples/blinky2
 	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=pca10059            examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=pca10059            examples/blinky2
+	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=itsybitsy-m0        examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=feather-m0          examples/blinky1
@@ -275,14 +280,6 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=circuitplay-express examples/blinky1
 	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=stm32f4disco        examples/blinky1
-	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=stm32f4disco        examples/blinky2
-	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=stm32f4disco-1      examples/blinky1
-	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=feather-stm32f405   examples/blinky1
-	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=circuitplay-bluefruit examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=circuitplay-express examples/i2s
@@ -307,12 +304,8 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=particle-xenon      examples/blinky1
 	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=nucleo-f103rb       examples/blinky1
-	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=pinetime-devkit0    examples/blinky1
 	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=lgt92    examples/blinky1
-	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=x9pro               examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=pca10056-s140v7     examples/blinky1
@@ -339,10 +332,6 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=teensy36            examples/blinky1
 	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=nucleo-f722ze       examples/blinky1
-	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=nucleo-l552ze       examples/blinky1
-	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=p1am-100            examples/blinky1
 	@$(MD5SUM) test.hex
 	# test pwm
@@ -352,8 +341,30 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=feather-m4          examples/pwm
 	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=pyportal            examples/pwm
+ifneq ($(STM32), 0)
+	$(TINYGO) build -size short -o test.hex -target=bluepill            examples/blinky1
 	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=feather-stm32f405   examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=lgt92               examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=nucleo-f103rb       examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=nucleo-f722ze       examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=nucleo-l031k6       examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=nucleo-l432kc       examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=nucleo-l552ze       examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=stm32f4disco        examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=stm32f4disco        examples/blinky2
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=stm32f4disco-1      examples/blinky1
+	@$(MD5SUM) test.hex
+endif
 ifneq ($(AVR), 0)
 	$(TINYGO) build -size short -o test.hex -target=atmega1284p         examples/serial
 	@$(MD5SUM) test.hex
@@ -391,6 +402,9 @@ endif
 	@$(MD5SUM) test.hex
 	$(TINYGO) build             -o test.nro -target=nintendoswitch      examples/serial
 	@$(MD5SUM) test.nro
+	$(TINYGO) build -size short -o test.hex -target=pca10040 -opt=0     ./testdata/stdlib.go
+	@$(MD5SUM) test.hex
+
 
 wasmtest:
 	$(GO) test ./tests/wasm

builder/build.go

@@ -4,16 +4,19 @@
 package builder
 
 import (
+	"crypto/sha512"
 	"debug/elf"
 	"encoding/binary"
+	"encoding/hex"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"go/types"
 	"io/ioutil"
 	"os"
 	"path/filepath"
+	"runtime"
 	"sort"
-	"strconv"
 	"strings"
 
 	"github.com/tinygo-org/tinygo/compileopts"
@@ -38,13 +41,41 @@ type BuildResult struct {
 	MainDir string
 }
 
+// packageAction is the struct that is serialized to JSON and hashed, to work as
+// a cache key of compiled packages. It should contain all the information that
+// goes into a compiled package to avoid using stale data.
+//
+// Right now it's still important to include a hash of every import, because a
+// dependency might have a public constant that this package uses and thus this
+// package will need to be recompiled if that constant changes. In the future,
+// the type data should be serialized to disk which can then be used as cache
+// key, avoiding the need for recompiling all dependencies when only the
+// implementation of an imported package changes.
+type packageAction struct {
+	ImportPath      string
+	CompilerVersion int // compiler.Version
+	InterpVersion   int // interp.Version
+	LLVMVersion     string
+	Config          *compiler.Config
+	CFlags          []string
+	FileHashes      map[string]string // hash of every file that's part of the package
+	Imports         map[string]string // map from imported package to action ID hash
+}
+
 // Build performs a single package to executable Go build. It takes in a package
 // name, an output path, and set of compile options and from that it manages the
 // whole compilation process.
 //
 // The error value may be of type *MultiError. Callers will likely want to check
 // for this case and print such errors individually.
-func Build(pkgName, outpath string, config *compileopts.Config, preAction func() error, action func(BuildResult) error) error {
+func Build(pkgName, outpath string, config *compileopts.Config, action func(BuildResult) error) error {
+	// Create a temporary directory for intermediary files.
+	dir, err := ioutil.TempDir("", "tinygo")
+	if err != nil {
+		return err
+	}
+	defer os.RemoveAll(dir)
+
 	compilerConfig := &compiler.Config{
 		Triple:          config.Triple(),
 		CPU:             config.CPU(),
@@ -87,31 +118,216 @@ func Build(pkgName, outpath string, config *compileopts.Config, preAction func()
 	// Makefile target.
 	var jobs []*compileJob
 
-	if preAction != nil {
+	// Create the *ssa.Program. This does not yet build the entire SSA of the
-		// Add job to preAction.
+	// program so it's pretty fast and doesn't need to be parallelized.
-		jobs = append(jobs, &compileJob{
+	program := lprogram.LoadSSA()
-			description: "preAction",
+
-			run:         preAction,
+	// Add jobs to compile each package.
-		})
+	// Packages that have a cache hit will not be compiled again.
+	var packageJobs []*compileJob
+	packageBitcodePaths := make(map[string]string)
+	packageActionIDs := make(map[string]string)
+	for _, pkg := range lprogram.Sorted() {
+		pkg := pkg // necessary to avoid a race condition
+
+		// Create a cache key: a hash from the action ID below that contains all
+		// the parameters for the build.
+		actionID := packageAction{
+			ImportPath:      pkg.ImportPath,
+			CompilerVersion: compiler.Version,
+			InterpVersion:   interp.Version,
+			LLVMVersion:     llvm.Version,
+			Config:          compilerConfig,
+			CFlags:          pkg.CFlags,
+			FileHashes:      make(map[string]string, len(pkg.FileHashes)),
+			Imports:         make(map[string]string, len(pkg.Pkg.Imports())),
+		}
+		for filePath, hash := range pkg.FileHashes {
+			actionID.FileHashes[filePath] = hex.EncodeToString(hash)
+		}
+		for _, imported := range pkg.Pkg.Imports() {
+			hash, ok := packageActionIDs[imported.Path()]
+			if !ok {
+				return fmt.Errorf("package %s imports %s but couldn't find dependency", pkg.ImportPath, imported.Path())
+			}
+			actionID.Imports[imported.Path()] = hash
+		}
+		buf, err := json.Marshal(actionID)
+		if err != nil {
+			panic(err) // shouldn't happen
+		}
+		hash := sha512.Sum512_224(buf)
+		packageActionIDs[pkg.ImportPath] = hex.EncodeToString(hash[:])
+
+		// Determine the path of the bitcode file (which is a serialized version
+		// of a LLVM module).
+		cacheDir := goenv.Get("GOCACHE")
+		if cacheDir == "off" {
+			// Use temporary build directory instead, effectively disabling the
+			// build cache.
+			cacheDir = dir
+		}
+		bitcodePath := filepath.Join(cacheDir, "pkg-"+hex.EncodeToString(hash[:])+".bc")
+		packageBitcodePaths[pkg.ImportPath] = bitcodePath
+
+		// Check whether this package has been compiled before, and if so don't
+		// compile it again.
+		if _, err := os.Stat(bitcodePath); err == nil {
+			// Already cached, don't recreate this package.
+			continue
+		}
+
+		// The package has not yet been compiled, so create a job to do so.
+		job := &compileJob{
+			description: "compile package " + pkg.ImportPath,
+			run: func(*compileJob) error {
+				// Compile AST to IR. The compiler.CompilePackage function will
+				// build the SSA as needed.
+				mod, errs := compiler.CompilePackage(pkg.ImportPath, pkg, program.Package(pkg.Pkg), machine, compilerConfig, config.DumpSSA())
+				if errs != nil {
+					return newMultiError(errs)
+				}
+				if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
+					return errors.New("verification error after compiling package " + pkg.ImportPath)
+				}
+
+				// Try to interpret package initializers at compile time.
+				// It may only be possible to do this partially, in which case
+				// it is completed after all IR files are linked.
+				pkgInit := mod.NamedFunction(pkg.Pkg.Path() + ".init")
+				if pkgInit.IsNil() {
+					panic("init not found for " + pkg.Pkg.Path())
+				}
+				err := interp.RunFunc(pkgInit, config.DumpSSA())
+				if err != nil {
+					return err
+				}
+				if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
+					return errors.New("verification error after interpreting " + pkgInit.Name())
+				}
+
+				// Serialize the LLVM module as a bitcode file.
+				// Write to a temporary path that is renamed to the destination
+				// file to avoid race conditions with other TinyGo invocatiosn
+				// that might also be compiling this package at the same time.
+				f, err := ioutil.TempFile(filepath.Dir(bitcodePath), filepath.Base(bitcodePath))
+				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)
+					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
+					// own, so create a somewhat useful error message here.
+					return fmt.Errorf("failed to write bitcode for package %s to file %s", pkg.ImportPath, bitcodePath)
+				}
+				err = f.Close()
+				if err != nil {
+					return err
+				}
+				return os.Rename(f.Name(), bitcodePath)
+			},
+		}
+		jobs = append(jobs, job)
+		packageJobs = append(packageJobs, job)
 	}
 
-	// Add job to compile and optimize all Go files at once.
+	// Add job that links and optimizes all packages together.
-	// TODO: parallelize this.
 	var mod llvm.Module
 	var stackSizeLoads []string
 	programJob := &compileJob{
-		description: "compile Go files",
+		description:  "link+optimize packages (LTO)",
-		run: func() (err error) {
+		dependencies: packageJobs,
-			mod, err = compileWholeProgram(pkgName, config, compilerConfig, lprogram, machine)
+		run: func(*compileJob) error {
-			if err != nil {
+			// Load and link all the bitcode files. This does not yet optimize
-				return
+			// anything, it only links the bitcode files together.
+			ctx := llvm.NewContext()
+			mod = ctx.NewModule("")
+			for _, pkg := range lprogram.Sorted() {
+				pkgMod, err := ctx.ParseBitcodeFile(packageBitcodePaths[pkg.ImportPath])
+				if err != nil {
+					return fmt.Errorf("failed to load bitcode file: %w", err)
+				}
+				err = llvm.LinkModules(mod, pkgMod)
+				if err != nil {
+					return fmt.Errorf("failed to link module: %w", err)
+				}
 			}
+
+			// Create runtime.initAll function that calls the runtime
+			// initializer of each package.
+			llvmInitFn := mod.NamedFunction("runtime.initAll")
+			llvmInitFn.SetLinkage(llvm.InternalLinkage)
+			llvmInitFn.SetUnnamedAddr(true)
+			llvmInitFn.Param(0).SetName("context")
+			llvmInitFn.Param(1).SetName("parentHandle")
+			block := mod.Context().AddBasicBlock(llvmInitFn, "entry")
+			irbuilder := mod.Context().NewBuilder()
+			defer irbuilder.Dispose()
+			irbuilder.SetInsertPointAtEnd(block)
+			i8ptrType := llvm.PointerType(mod.Context().Int8Type(), 0)
+			for _, pkg := range lprogram.Sorted() {
+				pkgInit := mod.NamedFunction(pkg.Pkg.Path() + ".init")
+				if pkgInit.IsNil() {
+					panic("init not found for " + pkg.Pkg.Path())
+				}
+				irbuilder.CreateCall(pkgInit, []llvm.Value{llvm.Undef(i8ptrType), llvm.Undef(i8ptrType)}, "")
+			}
+			irbuilder.CreateRetVoid()
+
+			// After linking, functions should (as far as possible) be set to
+			// private linkage or internal linkage. The compiler package marks
+			// non-exported functions by setting the visibility to hidden or
+			// (for thunks) to linkonce_odr linkage. Change the linkage here to
+			// internal to benefit much more from interprocedural optimizations.
+			for fn := mod.FirstFunction(); !fn.IsNil(); fn = llvm.NextFunction(fn) {
+				if fn.Visibility() == llvm.HiddenVisibility {
+					fn.SetVisibility(llvm.DefaultVisibility)
+					fn.SetLinkage(llvm.InternalLinkage)
+				} else if fn.Linkage() == llvm.LinkOnceODRLinkage {
+					fn.SetLinkage(llvm.InternalLinkage)
+				}
+			}
+
+			// Do the same for globals.
+			for global := mod.FirstGlobal(); !global.IsNil(); global = llvm.NextGlobal(global) {
+				if global.Visibility() == llvm.HiddenVisibility {
+					global.SetVisibility(llvm.DefaultVisibility)
+					global.SetLinkage(llvm.InternalLinkage)
+				} else if global.Linkage() == llvm.LinkOnceODRLinkage {
+					global.SetLinkage(llvm.InternalLinkage)
+				}
+			}
+
+			if config.Options.PrintIR {
+				fmt.Println("; Generated LLVM IR:")
+				fmt.Println(mod.String())
+			}
+
+			// Run all optimization passes, which are much more effective now
+			// that the optimizer can see the whole program at once.
+			err := optimizeProgram(mod, config)
+			if err != nil {
+				return err
+			}
+
 			// Make sure stack sizes are loaded from a separate section so they can be
 			// modified after linking.
 			if config.AutomaticStackSize() {
 				stackSizeLoads = transform.CreateStackSizeLoads(mod, config)
 			}
-			return
+			return nil
 		},
 	}
 	jobs = append(jobs, programJob)
@@ -148,19 +364,13 @@ func Build(pkgName, outpath string, config *compileopts.Config, preAction func()
 	// First add all jobs necessary to build this object file, then afterwards
 	// run all jobs in parallel as far as possible.
 
-	// Create a temporary directory for intermediary files.
-	dir, err := ioutil.TempDir("", "tinygo")
-	if err != nil {
-		return err
-	}
-	defer os.RemoveAll(dir)
-
 	// Add job to write the output object file.
 	objfile := filepath.Join(dir, "main.o")
 	outputObjectFileJob := &compileJob{
 		description:  "generate output file",
 		dependencies: []*compileJob{programJob},
-		run: func() error {
+		result:       objfile,
+		run: func(*compileJob) error {
 			llvmBuf, err := machine.EmitToMemoryBuffer(mod, llvm.ObjectFile)
 			if err != nil {
 				return err
@@ -174,40 +384,32 @@ func Build(pkgName, outpath string, config *compileopts.Config, preAction func()
 	linkerDependencies := []*compileJob{outputObjectFileJob}
 	executable := filepath.Join(dir, "main")
 	tmppath := executable // final file
-	ldflags := append(config.LDFlags(), "-o", executable, objfile)
+	ldflags := append(config.LDFlags(), "-o", executable)
 
 	// Add compiler-rt dependency if needed. Usually this is a simple load from
 	// a cache.
 	if config.Target.RTLib == "compiler-rt" {
-		path, job, err := CompilerRT.load(config.Triple(), config.CPU(), dir)
+		job, err := CompilerRT.load(config.Triple(), config.CPU(), dir)
 		if err != nil {
 			return err
 		}
-		if job != nil {
+		jobs = append(jobs, job.dependencies...)
-			// The library was not loaded from cache so needs to be compiled
+		jobs = append(jobs, job)
-			// (and then stored in the cache).
+		linkerDependencies = append(linkerDependencies, job)
-			jobs = append(jobs, job.dependencies...)
-			jobs = append(jobs, job)
-			linkerDependencies = append(linkerDependencies, job)
-		}
-		ldflags = append(ldflags, path)
 	}
 
 	// Add libc dependency if needed.
 	root := goenv.Get("TINYGOROOT")
 	switch config.Target.Libc {
 	case "picolibc":
-		path, job, err := Picolibc.load(config.Triple(), config.CPU(), dir)
+		job, err := Picolibc.load(config.Triple(), config.CPU(), dir)
 		if err != nil {
 			return err
 		}
-		if job != nil {
+		// The library needs to be compiled (cache miss).
-			// The library needs to be compiled (cache miss).
+		jobs = append(jobs, job.dependencies...)
-			jobs = append(jobs, job.dependencies...)
+		jobs = append(jobs, job)
-			jobs = append(jobs, job)
+		linkerDependencies = append(linkerDependencies, job)
-			linkerDependencies = append(linkerDependencies, job)
-		}
-		ldflags = append(ldflags, path)
 	case "wasi-libc":
 		path := filepath.Join(root, "lib/wasi-libc/sysroot/lib/wasm32-wasi/libc.a")
 		if _, err := os.Stat(path); os.IsNotExist(err) {
@@ -223,44 +425,37 @@ func Build(pkgName, outpath string, config *compileopts.Config, preAction func()
 	// Add jobs to compile extra files. These files are in C or assembly and
 	// contain things like the interrupt vector table and low level operations
 	// such as stack switching.
-	for i, path := range config.ExtraFiles() {
+	for _, path := range config.ExtraFiles() {
 		abspath := filepath.Join(root, path)
-		outpath := filepath.Join(dir, "extra-"+strconv.Itoa(i)+"-"+filepath.Base(path)+".o")
 		job := &compileJob{
 			description: "compile extra file " + path,
-			run: func() error {
+			run: func(job *compileJob) error {
-				err := runCCompiler(config.Target.Compiler, append(config.CFlags(), "-c", "-o", outpath, abspath)...)
+				result, err := compileAndCacheCFile(abspath, dir, config.CFlags(), config)
-				if err != nil {
+				job.result = result
-					return &commandError{"failed to build", path, err}
+				return err
-				}
-				return nil
 			},
 		}
 		jobs = append(jobs, job)
 		linkerDependencies = append(linkerDependencies, job)
-		ldflags = append(ldflags, outpath)
 	}
 
 	// Add jobs to compile C files in all packages. This is part of CGo.
 	// TODO: do this as part of building the package to be able to link the
 	// bitcode files together.
-	for i, pkg := range lprogram.Sorted() {
+	for _, pkg := range lprogram.Sorted() {
-		for j, filename := range pkg.CFiles {
+		pkg := pkg
-			file := filepath.Join(pkg.Dir, filename)
+		for _, filename := range pkg.CFiles {
-			outpath := filepath.Join(dir, "pkg"+strconv.Itoa(i)+"."+strconv.Itoa(j)+"-"+filepath.Base(file)+".o")
+			abspath := filepath.Join(pkg.Dir, filename)
 			job := &compileJob{
-				description: "compile CGo file " + file,
+				description: "compile CGo file " + abspath,
-				run: func() error {
+				run: func(job *compileJob) error {
-					err := runCCompiler(config.Target.Compiler, append(config.CFlags(), "-c", "-o", outpath, file)...)
+					result, err := compileAndCacheCFile(abspath, dir, pkg.CFlags, config)
-					if err != nil {
+					job.result = result
-						return &commandError{"failed to build", file, err}
+					return err
-					}
-					return nil
 				},
 			}
 			jobs = append(jobs, job)
 			linkerDependencies = append(linkerDependencies, job)
-			ldflags = append(ldflags, outpath)
 		}
 	}
 
@@ -275,7 +470,16 @@ func Build(pkgName, outpath string, config *compileopts.Config, preAction func()
 	jobs = append(jobs, &compileJob{
 		description:  "link",
 		dependencies: linkerDependencies,
-		run: func() error {
+		run: func(job *compileJob) error {
+			for _, dependency := range job.dependencies {
+				if dependency.result == "" {
+					return errors.New("dependency without result: " + dependency.description)
+				}
+				ldflags = append(ldflags, dependency.result)
+			}
+			if config.Options.PrintCommands {
+				fmt.Printf("%s %s\n", config.Target.Linker, strings.Join(ldflags, " "))
+			}
 			err = link(config.Target.Linker, ldflags...)
 			if err != nil {
 				return &commandError{"failed to link", executable, err}
@@ -374,29 +578,25 @@ func Build(pkgName, outpath string, config *compileopts.Config, preAction func()
 	})
 }
 
-// compileWholeProgram compiles the entire *loader.Program to a LLVM module and
+// optimizeProgram runs a series of optimizations and transformations that are
-// applies most necessary optimizations and transformations.
+// needed to convert a program to its final form. Some transformations are not
-func compileWholeProgram(pkgName string, config *compileopts.Config, compilerConfig *compiler.Config, lprogram *loader.Program, machine llvm.TargetMachine) (llvm.Module, error) {
+// optional and must be run as the compiler expects them to run.
-	// Compile AST to IR.
+func optimizeProgram(mod llvm.Module, config *compileopts.Config) error {
-	mod, errs := compiler.CompileProgram(lprogram, machine, compilerConfig, config.DumpSSA())
-	if errs != nil {
-		return mod, newMultiError(errs)
-	}
-
-	if config.Options.PrintIR {
-		fmt.Println("; Generated LLVM IR:")
-		fmt.Println(mod.String())
-	}
-	if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
-		return mod, errors.New("verification error after IR construction")
-	}
-
 	err := interp.Run(mod, config.DumpSSA())
 	if err != nil {
-		return mod, err
+		return err
 	}
-	if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
+	if config.VerifyIR() {
-		return mod, errors.New("verification error after interpreting runtime.initAll")
+		// Only verify if we really need it.
+		// The IR has already been verified before writing the bitcode to disk
+		// and the interp function above doesn't need to do a lot as most of the
+		// package initializers have already run. Additionally, verifying this
+		// linked IR is _expensive_ because dead code hasn't been removed yet,
+		// easily costing a few hundred milliseconds. Therefore, only do it when
+		// specifically requested.
+		if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
+			return errors.New("verification error after interpreting runtime.initAll")
+		}
 	}
 
 	if config.GOOS() != "darwin" {
@@ -411,23 +611,16 @@ func compileWholeProgram(pkgName string, config *compileopts.Config, compilerCon
 	if config.WasmAbi() == "js" {
 		err := transform.ExternalInt64AsPtr(mod)
 		if err != nil {
-			return mod, err
+			return err
 		}
 	}
 
 	// Optimization levels here are roughly the same as Clang, but probably not
 	// exactly.
-	errs = nil
+	var errs []error
 	switch config.Options.Opt {
-	/*
+	case "none", "0":
-		Currently, turning optimizations off causes compile failures.
+		errs = transform.Optimize(mod, config, 0, 0, 0) // -O0
-		We rely on the optimizer removing some dead symbols.
-		Avoid providing an option that does not work right now.
-		In the future once everything has been fixed we can re-enable this.
-
-		case "none", "0":
-			errs = transform.Optimize(mod, config, 0, 0, 0) // -O0
-	*/
 	case "1":
 		errs = transform.Optimize(mod, config, 1, 0, 0) // -O1
 	case "2":
@@ -437,13 +630,13 @@ func compileWholeProgram(pkgName string, config *compileopts.Config, compilerCon
 	case "z":
 		errs = transform.Optimize(mod, config, 2, 2, 5) // -Oz, default
 	default:
-		errs = []error{errors.New("unknown optimization level: -opt=" + config.Options.Opt)}
+		return errors.New("unknown optimization level: -opt=" + config.Options.Opt)
 	}
 	if len(errs) > 0 {
-		return mod, newMultiError(errs)
+		return newMultiError(errs)
 	}
 	if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
-		return mod, errors.New("verification failure after LLVM optimization passes")
+		return errors.New("verification failure after LLVM optimization passes")
 	}
 
 	// LLVM 11 by default tries to emit tail calls (even with the target feature
@@ -457,7 +650,7 @@ func compileWholeProgram(pkgName string, config *compileopts.Config, compilerCon
 		transform.DisableTailCalls(mod)
 	}
 
-	return mod, nil
+	return nil
 }
 
 // functionStackSizes keeps stack size information about a single function

builder/cc.go

@@ -0,0 +1,305 @@
+package builder
+
+// This file implements a wrapper around the C compiler (Clang) which uses a
+// build cache.
+
+import (
+	"crypto/sha512"
+	"encoding/hex"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"io/ioutil"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"unicode"
+
+	"github.com/tinygo-org/tinygo/compileopts"
+	"github.com/tinygo-org/tinygo/goenv"
+	"tinygo.org/x/go-llvm"
+)
+
+// compileAndCacheCFile compiles a C or assembly file using a build cache.
+// Compiling the same file again (if nothing changed, including included header
+// files) the output is loaded from the build cache instead.
+//
+// Its operation is a bit complex (more complex than Go package build caching)
+// because the list of file dependencies is only known after the file is
+// compiled. However, luckily compilers have a flag to write a list of file
+// dependencies in Makefile syntax which can be used for caching.
+//
+// Because of this complexity, every file has in fact two cached build outputs:
+// the file itself, and the list of dependencies. Its operation is as follows:
+//
+//   depfile = hash(path, compiler, cflags, ...)
+//   if depfile exists:
+//     outfile = hash of all files and depfile name
+//     if outfile exists:
+//       # cache hit
+//       return outfile
+//   # cache miss
+//   tmpfile = compile file
+//   read dependencies (side effect of compile)
+//   write depfile
+//   outfile = hash of all files and depfile name
+//   rename tmpfile to outfile
+//
+// There are a few edge cases that are not handled:
+// - If a file is added to an include path, that file may be included instead of
+//   some other file. This would be fixed by also including lookup failures in the
+//   dependencies file, but I'm not aware of a compiler which does that.
+// - The Makefile syntax that compilers output has issues, see readDepFile for
+//   details.
+// - A header file may be changed to add/remove an include. This invalidates the
+//   depfile but without invalidating its name. For this reason, the depfile is
+//   written on each new compilation (even when it seems unnecessary). However, it
+//   could in rare cases lead to a stale file fetched from the cache.
+func compileAndCacheCFile(abspath, tmpdir string, cflags []string, config *compileopts.Config) (string, error) {
+	// Hash input file.
+	fileHash, err := hashFile(abspath)
+	if err != nil {
+		return "", err
+	}
+
+	// Create cache key for the dependencies file.
+	buf, err := json.Marshal(struct {
+		Path        string
+		Hash        string
+		Compiler    string
+		Flags       []string
+		LLVMVersion string
+	}{
+		Path:        abspath,
+		Hash:        fileHash,
+		Compiler:    config.Target.Compiler,
+		Flags:       config.CFlags(),
+		LLVMVersion: llvm.Version,
+	})
+	if err != nil {
+		panic(err) // shouldn't happen
+	}
+	depfileNameHashBuf := sha512.Sum512_224(buf)
+	depfileNameHash := hex.EncodeToString(depfileNameHashBuf[:])
+
+	// Load dependencies file, if possible.
+	depfileName := "dep-" + depfileNameHash + ".json"
+	depfileCachePath := filepath.Join(goenv.Get("GOCACHE"), depfileName)
+	depfileBuf, err := ioutil.ReadFile(depfileCachePath)
+	var dependencies []string // sorted list of dependency paths
+	if err == nil {
+		// There is a dependency file, that's great!
+		// Parse it first.
+		err := json.Unmarshal(depfileBuf, &dependencies)
+		if err != nil {
+			return "", fmt.Errorf("could not parse dependencies JSON: %w", err)
+		}
+
+		// Obtain hashes of all the files listed as a dependency.
+		outpath, err := makeCFileCachePath(dependencies, depfileNameHash)
+		if err == nil {
+			if _, err := os.Stat(outpath); err == nil {
+				return outpath, nil
+			} else if !os.IsNotExist(err) {
+				return "", err
+			}
+		}
+	} else if !os.IsNotExist(err) {
+		// expected either nil or IsNotExist
+		return "", err
+	}
+
+	objTmpFile, err := ioutil.TempFile(goenv.Get("GOCACHE"), "tmp-*.o")
+	if err != nil {
+		return "", err
+	}
+	objTmpFile.Close()
+	depTmpFile, err := ioutil.TempFile(tmpdir, "dep-*.d")
+	if err != nil {
+		return "", err
+	}
+	depTmpFile.Close()
+	flags := append([]string{}, cflags...)                                   // copy cflags
+	flags = append(flags, "-MD", "-MV", "-MTdeps", "-MF", depTmpFile.Name()) // autogenerate dependencies
+	flags = append(flags, "-c", "-o", objTmpFile.Name(), abspath)
+	if config.Options.PrintCommands {
+		fmt.Printf("%s %s\n", config.Target.Compiler, strings.Join(flags, " "))
+	}
+	err = runCCompiler(config.Target.Compiler, flags...)
+	if err != nil {
+		return "", &commandError{"failed to build", abspath, err}
+	}
+
+	// Create sorted and uniqued slice of dependencies.
+	dependencyPaths, err := readDepFile(depTmpFile.Name())
+	if err != nil {
+		return "", err
+	}
+	dependencyPaths = append(dependencyPaths, abspath) // necessary for .s files
+	dependencySet := make(map[string]struct{}, len(dependencyPaths))
+	var dependencySlice []string
+	for _, path := range dependencyPaths {
+		if _, ok := dependencySet[path]; ok {
+			continue
+		}
+		dependencySet[path] = struct{}{}
+		dependencySlice = append(dependencySlice, path)
+	}
+	sort.Strings(dependencySlice)
+
+	// Write dependencies file.
+	f, err := ioutil.TempFile(filepath.Dir(depfileCachePath), depfileName)
+	buf, err = json.MarshalIndent(dependencySlice, "", "\t")
+	if err != nil {
+		panic(err) // shouldn't happen
+	}
+	_, err = f.Write(buf)
+	if err != nil {
+		return "", err
+	}
+	err = f.Close()
+	if err != nil {
+		return "", err
+	}
+	err = os.Rename(f.Name(), depfileCachePath)
+	if err != nil {
+		return "", err
+	}
+
+	// Move temporary object file to final location.
+	outpath, err := makeCFileCachePath(dependencySlice, depfileNameHash)
+	if err != nil {
+		return "", err
+	}
+	err = os.Rename(objTmpFile.Name(), outpath)
+	if err != nil {
+		return "", err
+	}
+
+	return outpath, nil
+}
+
+// Create a cache path (a path in GOCACHE) to store the output of a compiler
+// job. This path is based on the dep file name (which is a hash of metadata
+// including compiler flags) and the hash of all input files in the paths slice.
+func makeCFileCachePath(paths []string, depfileNameHash string) (string, error) {
+	// Hash all input files.
+	fileHashes := make(map[string]string, len(paths))
+	for _, path := range paths {
+		hash, err := hashFile(path)
+		if err != nil {
+			return "", err
+		}
+		fileHashes[path] = hash
+	}
+
+	// Calculate a cache key based on the above hashes.
+	buf, err := json.Marshal(struct {
+		DepfileHash string
+		FileHashes  map[string]string
+	}{
+		DepfileHash: depfileNameHash,
+		FileHashes:  fileHashes,
+	})
+	if err != nil {
+		panic(err) // shouldn't happen
+	}
+	outFileNameBuf := sha512.Sum512_224(buf)
+	cacheKey := hex.EncodeToString(outFileNameBuf[:])
+
+	outpath := filepath.Join(goenv.Get("GOCACHE"), "obj-"+cacheKey+".o")
+	return outpath, nil
+}
+
+// hashFile hashes the given file path and returns the hash as a hex string.
+func hashFile(path string) (string, error) {
+	f, err := os.Open(path)
+	if err != nil {
+		return "", fmt.Errorf("failed to hash file: %w", err)
+	}
+	defer f.Close()
+	fileHasher := sha512.New512_224()
+	_, err = io.Copy(fileHasher, f)
+	if err != nil {
+		return "", fmt.Errorf("failed to hash file: %w", err)
+	}
+	return hex.EncodeToString(fileHasher.Sum(nil)), nil
+}
+
+// readDepFile reads a dependency file in NMake (Visual Studio make) format. The
+// file is assumed to have a single target named deps.
+//
+// There are roughly three make syntax variants:
+// - BSD make, which doesn't support any escaping. This means that many special
+//   characters are not supported in file names.
+// - GNU make, which supports escaping using a backslash but when it fails to
+//   find a file it tries to fall back with the literal path name (to match BSD
+//   make).
+// - NMake (Visual Studio) and Jom, which simply quote the string if there are
+//   any weird characters.
+// Clang supports two variants: a format that's a compromise between BSD and GNU
+// make (and is buggy to match GCC which is equally buggy), and NMake/Jom, which
+// is at least somewhat sane. This last format isn't perfect either: it does not
+// correctly handle filenames with quote marks in them. Those are generally not
+// allowed on Windows, but of course can be used on POSIX like systems. Still,
+// it's the most sane of any of the formats so readDepFile will use that format.
+func readDepFile(filename string) ([]string, error) {
+	buf, err := ioutil.ReadFile(filename)
+	if err != nil {
+		return nil, err
+	}
+	if len(buf) == 0 {
+		return nil, nil
+	}
+	return parseDepFile(string(buf))
+}
+
+func parseDepFile(s string) ([]string, error) {
+	// This function makes no attempt at parsing anything other than Clang -MD
+	// -MV output.
+
+	// For Windows: replace CRLF with LF to make the logic below simpler.
+	s = strings.ReplaceAll(s, "\r\n", "\n")
+
+	// Collapse all lines ending in a backslash. These backslashes are really
+	// just a way to continue a line without making very long lines.
+	s = strings.ReplaceAll(s, "\\\n", " ")
+
+	// Only use the first line, which is expected to begin with "deps:".
+	line := strings.SplitN(s, "\n", 2)[0]
+	if !strings.HasPrefix(line, "deps:") {
+		return nil, errors.New("readDepFile: expected 'deps:' prefix")
+	}
+	line = strings.TrimSpace(line[len("deps:"):])
+
+	var deps []string
+	for line != "" {
+		if line[0] == '"' {
+			// File path is quoted. Path ends with double quote.
+			// This does not handle double quotes in path names, which is a
+			// problem on non-Windows systems.
+			line = line[1:]
+			end := strings.IndexByte(line, '"')
+			if end < 0 {
+				return nil, errors.New("readDepFile: path is incorrectly quoted")
+			}
+			dep := line[:end]
+			line = strings.TrimSpace(line[end+1:])
+			deps = append(deps, dep)
+		} else {
+			// File path is not quoted. Path ends in space or EOL.
+			end := strings.IndexFunc(line, unicode.IsSpace)
+			if end < 0 {
+				// last dependency
+				deps = append(deps, line)
+				break
+			}
+			dep := line[:end]
+			line = strings.TrimSpace(line[end:])
+			deps = append(deps, dep)
+		}
+	}
+	return deps, nil
+}

builder/cc_test.go

@@ -0,0 +1,33 @@
+package builder
+
+import (
+	"reflect"
+	"testing"
+)
+
+func TestSplitDepFile(t *testing.T) {
+	for i, tc := range []struct {
+		in  string
+		out []string
+	}{
+		{`deps: foo bar`, []string{"foo", "bar"}},
+		{`deps: foo "bar"`, []string{"foo", "bar"}},
+		{`deps: "foo" bar`, []string{"foo", "bar"}},
+		{`deps: "foo bar"`, []string{"foo bar"}},
+		{`deps: "foo bar" `, []string{"foo bar"}},
+		{"deps: foo\nbar", []string{"foo"}},
+		{"deps: foo \\\nbar", []string{"foo", "bar"}},
+		{"deps: foo\\bar \\\nbaz", []string{"foo\\bar", "baz"}},
+		{"deps: foo\\bar \\\r\n baz", []string{"foo\\bar", "baz"}}, // Windows uses CRLF line endings
+	} {
+		out, err := parseDepFile(tc.in)
+		if err != nil {
+			t.Errorf("test #%d failed: %v", i, err)
+			continue
+		}
+		if !reflect.DeepEqual(out, tc.out) {
+			t.Errorf("test #%d failed: expected %#v but got %#v", i, tc.out, out)
+			continue
+		}
+	}
+}

builder/jobs.go

@@ -27,12 +27,23 @@ const (
 type compileJob struct {
 	description  string // description, only used for logging
 	dependencies []*compileJob
-	run          func() error
+	result       string // result (path)
+	run          func(*compileJob) (err error)
 	state        jobState
 	err          error         // error if finished
 	duration     time.Duration // how long it took to run this job (only set after finishing)
 }
 
+// dummyCompileJob returns a new *compileJob that produces an output without
+// doing anything. This can be useful where a *compileJob producing an output is
+// expected but nothing needs to be done, for example for a load from a cache.
+func dummyCompileJob(result string) *compileJob {
+	return &compileJob{
+		description: "<dummy>",
+		result:      result,
+	}
+}
+
 // readyToRun returns whether this job is ready to run: it is itself not yet
 // started and all dependencies are finished.
 func (job *compileJob) readyToRun() bool {
@@ -150,9 +161,11 @@ func nextJob(jobs []*compileJob) *compileJob {
 func jobWorker(workerChan, doneChan chan *compileJob) {
 	for job := range workerChan {
 		start := time.Now()
-		err := job.run()
+		if job.run != nil {
-		if err != nil {
+			err := job.run(job)
-			job.err = err
+			if err != nil {
+				job.err = err
+			}
 		}
 		job.duration = time.Since(start)
 		doneChan <- job

builder/library.go

@@ -42,30 +42,28 @@ func (l *Library) sourcePaths(target string) []string {
 // The resulting file is stored in the provided tmpdir, which is expected to be
 // removed after the Load call.
 func (l *Library) Load(target, tmpdir string) (path string, err error) {
-	path, job, err := l.load(target, "", tmpdir)
+	job, err := l.load(target, "", tmpdir)
 	if err != nil {
 		return "", err
 	}
-	if job != nil {
+	jobs := append([]*compileJob{job}, job.dependencies...)
-		jobs := append([]*compileJob{job}, job.dependencies...)
+	err = runJobs(jobs)
-		err = runJobs(jobs)
+	return job.result, err
-	}
-	return path, err
 }
 
-// load returns a path to the library file for the given target, loading it from
+// load returns a compile job to build this library file for the given target
-// cache if possible. It will return a non-zero compiler job if the library
+// and CPU. It may return a dummy compileJob if the library build is already
-// wasn't cached, this job (and its dependencies) must be run before the library
+// cached. The path is stored as job.result but is only valid if the job and
-// path is valid.
+// job.dependencies have been run.
 // The provided tmpdir will be used to store intermediary files and possibly the
 // output archive file, it is expected to be removed after use.
-func (l *Library) load(target, cpu, tmpdir string) (path string, job *compileJob, err error) {
+func (l *Library) load(target, cpu, tmpdir string) (job *compileJob, err error) {
 	// Try to load a precompiled library.
 	precompiledPath := filepath.Join(goenv.Get("TINYGOROOT"), "pkg", target, l.name+".a")
 	if _, err := os.Stat(precompiledPath); err == nil {
 		// Found a precompiled library for this OS/architecture. Return the path
 		// directly.
-		return precompiledPath, nil, nil
+		return dummyCompileJob(precompiledPath), nil
 	}
 
 	var outfile string
@@ -78,7 +76,7 @@ func (l *Library) load(target, cpu, tmpdir string) (path string, job *compileJob
 	// Try to fetch this library from the cache.
 	if path, err := cacheLoad(outfile, l.sourcePaths(target)); path != "" || err != nil {
 		// Cache hit.
-		return path, nil, err
+		return dummyCompileJob(path), nil
 	}
 	// Cache miss, build it now.
 
@@ -86,7 +84,7 @@ func (l *Library) load(target, cpu, tmpdir string) (path string, job *compileJob
 	dir := filepath.Join(tmpdir, "build-lib-"+l.name)
 	err = os.Mkdir(dir, 0777)
 	if err != nil {
-		return "", nil, err
+		return nil, err
 	}
 
 	// Precalculate the flags to the compiler invocation.
@@ -113,7 +111,8 @@ func (l *Library) load(target, cpu, tmpdir string) (path string, job *compileJob
 	arpath := filepath.Join(dir, l.name+".a")
 	job = &compileJob{
 		description: "ar " + l.name + ".a",
-		run: func() error {
+		result:      arpath,
+		run: func(*compileJob) error {
 			// Create an archive of all object files.
 			err := makeArchive(arpath, objs)
 			if err != nil {
@@ -133,7 +132,7 @@ func (l *Library) load(target, cpu, tmpdir string) (path string, job *compileJob
 		objs = append(objs, objpath)
 		job.dependencies = append(job.dependencies, &compileJob{
 			description: "compile " + srcpath,
-			run: func() error {
+			run: func(*compileJob) error {
 				var compileArgs []string
 				compileArgs = append(compileArgs, args...)
 				compileArgs = append(compileArgs, "-o", objpath, srcpath)
@@ -146,5 +145,5 @@ func (l *Library) load(target, cpu, tmpdir string) (path string, job *compileJob
 		})
 	}
 
-	return arpath, job, nil
+	return job, nil
 }

cgo/cgo.go

@@ -41,7 +41,9 @@ type cgoPackage struct {
 	elaboratedTypes map[string]*elaboratedTypeInfo
 	enums           map[string]enumInfo
 	anonStructNum   int
-	ldflags         []string
+	cflags          []string // CFlags from #cgo lines
+	ldflags         []string // LDFlags from #cgo lines
+	visitedFiles    map[string][]byte
 }
 
 // constantInfo stores some information about a CGo constant found by libclang
@@ -156,9 +158,10 @@ typedef unsigned long long  _Cgo_ulonglong;
 // Process extracts `import "C"` statements from the AST, parses the comment
 // with libclang, and modifies the AST to use this information. It returns a
 // newly created *ast.File that should be added to the list of to-be-parsed
-// files. If there is one or more error, it returns these in the []error slice
+// files, the CFLAGS and LDFLAGS found in #cgo lines, and a map of file hashes
-// but still modifies the AST.
+// of the accessed C header files. If there is one or more error, it returns
-func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string) (*ast.File, []string, []error) {
+// these in the []error slice but still modifies the AST.
+func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string) (*ast.File, []string, []string, map[string][]byte, []error) {
 	p := &cgoPackage{
 		dir:             dir,
 		fset:            fset,
@@ -170,13 +173,9 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 		typedefs:        map[string]*typedefInfo{},
 		elaboratedTypes: map[string]*elaboratedTypeInfo{},
 		enums:           map[string]enumInfo{},
+		visitedFiles:    map[string][]byte{},
 	}
 
-	// Disable _FORTIFY_SOURCE as it causes problems on macOS.
-	// Note that it is only disabled for memcpy (etc) calls made from Go, which
-	// have better alternatives anyway.
-	cflags = append(cflags, "-D_FORTIFY_SOURCE=0")
-
 	// Add a new location for the following file.
 	generatedTokenPos := p.fset.AddFile(dir+"/!cgo.go", -1, 0)
 	generatedTokenPos.SetLines([]int{0})
@@ -185,7 +184,7 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 	// Find the absolute path for this package.
 	packagePath, err := filepath.Abs(fset.File(files[0].Pos()).Name())
 	if err != nil {
-		return nil, nil, []error{
+		return nil, nil, nil, nil, []error{
 			scanner.Error{
 				Pos: fset.Position(files[0].Pos()),
 				Msg: "cgo: cannot find absolute path: " + err.Error(), // TODO: wrap this error
@@ -360,7 +359,7 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 						continue
 					}
 					makePathsAbsolute(flags, packagePath)
-					cflags = append(cflags, flags...)
+					p.cflags = append(p.cflags, flags...)
 				case "LDFLAGS":
 					flags, err := shlex.Split(value)
 					if err != nil {
@@ -383,6 +382,13 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 		}
 	}
 
+	// Define CFlags that will be used while parsing the package.
+	// Disable _FORTIFY_SOURCE as it causes problems on macOS.
+	// Note that it is only disabled for memcpy (etc) calls made from Go, which
+	// have better alternatives anyway.
+	cflagsForCGo := append([]string{"-D_FORTIFY_SOURCE=0"}, cflags...)
+	cflagsForCGo = append(cflagsForCGo, p.cflags...)
+
 	// Process all CGo imports.
 	for _, genDecl := range statements {
 		cgoComment := genDecl.Doc.Text()
@@ -392,7 +398,7 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 			pos = genDecl.Doc.Pos()
 		}
 		position := fset.PositionFor(pos, true)
-		p.parseFragment(cgoComment+cgoTypes, cflags, position.Filename, position.Line)
+		p.parseFragment(cgoComment+cgoTypes, cflagsForCGo, position.Filename, position.Line)
 	}
 
 	// Declare functions found by libclang.
@@ -427,7 +433,7 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 	// Print the newly generated in-memory AST, for debugging.
 	//ast.Print(fset, p.generated)
 
-	return p.generated, p.ldflags, p.errors
+	return p.generated, p.cflags, p.ldflags, p.visitedFiles, p.errors
 }
 
 // makePathsAbsolute converts some common path compiler flags (-I, -L) from

cgo/cgo_test.go

@@ -65,7 +65,7 @@ func TestCGo(t *testing.T) {
 			}
 
 			// Process the AST with CGo.
-			cgoAST, _, cgoErrors := Process([]*ast.File{f}, "testdata", fset, cflags)
+			cgoAST, _, _, _, cgoErrors := Process([]*ast.File{f}, "testdata", fset, cflags)
 
 			// Check the AST for type errors.
 			var typecheckErrors []error

cgo/libclang.go

@@ -4,6 +4,7 @@ package cgo
 // modification. It does not touch the AST itself.
 
 import (
+	"crypto/sha512"
 	"fmt"
 	"go/ast"
 	"go/scanner"
@@ -56,6 +57,7 @@ unsigned tinygo_clang_Cursor_isBitField(GoCXCursor c);
 int tinygo_clang_globals_visitor(GoCXCursor c, GoCXCursor parent, CXClientData client_data);
 int tinygo_clang_struct_visitor(GoCXCursor c, GoCXCursor parent, CXClientData client_data);
 int tinygo_clang_enum_visitor(GoCXCursor c, GoCXCursor parent, CXClientData client_data);
+void tinygo_clang_inclusion_visitor(CXFile included_file, CXSourceLocation *inclusion_stack, unsigned include_len, CXClientData client_data);
 */
 import "C"
 
@@ -114,6 +116,7 @@ func (p *cgoPackage) parseFragment(fragment string, cflags []string, posFilename
 	}
 	defer C.clang_disposeTranslationUnit(unit)
 
+	// Report parser and type errors.
 	if numDiagnostics := int(C.clang_getNumDiagnostics(unit)); numDiagnostics != 0 {
 		addDiagnostic := func(diagnostic C.CXDiagnostic) {
 			spelling := getString(C.clang_getDiagnosticSpelling(diagnostic))
@@ -134,10 +137,36 @@ func (p *cgoPackage) parseFragment(fragment string, cflags []string, posFilename
 		}
 	}
 
+	// Extract information required by CGo.
 	ref := storedRefs.Put(p)
 	defer storedRefs.Remove(ref)
 	cursor := C.tinygo_clang_getTranslationUnitCursor(unit)
 	C.tinygo_clang_visitChildren(cursor, C.CXCursorVisitor(C.tinygo_clang_globals_visitor), C.CXClientData(ref))
+
+	// Determine files read during CGo processing, for caching.
+	inclusionCallback := func(includedFile C.CXFile) {
+		// Get full file path.
+		path := getString(C.clang_getFileName(includedFile))
+
+		// Get contents of file (that should be in-memory).
+		size := C.size_t(0)
+		rawData := C.clang_getFileContents(unit, includedFile, &size)
+		if rawData == nil {
+			// Sanity check. This should (hopefully) never trigger.
+			panic("libclang: file contents was not loaded")
+		}
+		data := (*[1 << 24]byte)(unsafe.Pointer(rawData))[:size]
+
+		// Hash the contents if it isn't hashed yet.
+		if _, ok := p.visitedFiles[path]; !ok {
+			// already stored
+			sum := sha512.Sum512_224(data)
+			p.visitedFiles[path] = sum[:]
+		}
+	}
+	inclusionCallbackRef := storedRefs.Put(inclusionCallback)
+	defer storedRefs.Remove(inclusionCallbackRef)
+	C.clang_getInclusions(unit, C.CXInclusionVisitor(C.tinygo_clang_inclusion_visitor), C.CXClientData(inclusionCallbackRef))
 }
 
 //export tinygo_clang_globals_visitor
@@ -772,3 +801,9 @@ func tinygo_clang_enum_visitor(c, parent C.GoCXCursor, client_data C.CXClientDat
 	}
 	return C.CXChildVisit_Continue
 }
+
+//export tinygo_clang_inclusion_visitor
+func tinygo_clang_inclusion_visitor(includedFile C.CXFile, inclusionStack *C.CXSourceLocation, includeLen C.unsigned, clientData C.CXClientData) {
+	callback := storedRefs.Get(unsafe.Pointer(clientData)).(func(C.CXFile))
+	callback(includedFile)
+}

compileopts/target.go

@@ -7,6 +7,7 @@ import (
 	"errors"
 	"io"
 	"os"
+	"os/exec"
 	"path/filepath"
 	"reflect"
 	"runtime"
@@ -42,7 +43,7 @@ type TargetSpec struct {
 	ExtraFiles       []string `json:"extra-files"`
 	Emulator         []string `json:"emulator" override:"copy"` // inherited Emulator must not be append
 	FlashCommand     string   `json:"flash-command"`
-	GDB              string   `json:"gdb"`
+	GDB              []string `json:"gdb"`
 	PortReset        string   `json:"flash-1200-bps-reset"`
 	FlashMethod      string   `json:"flash-method"`
 	FlashVolume      string   `json:"msd-volume-name"`
@@ -229,7 +230,13 @@ func LoadTarget(target string) (*TargetSpec, error) {
 	}
 }
 
+// WindowsBuildNotSupportedErr is being thrown, when goos is windows and no target has been specified.
+var WindowsBuildNotSupportedErr = errors.New("Building Windows binaries is currently not supported. Try specifying a different target")
+
 func defaultTarget(goos, goarch, triple string) (*TargetSpec, error) {
+	if goos == "windows" {
+		return nil, WindowsBuildNotSupportedErr
+	}
 	// No target spec available. Use the default one, useful on most systems
 	// with a regular OS.
 	spec := TargetSpec{
@@ -240,7 +247,7 @@ func defaultTarget(goos, goarch, triple string) (*TargetSpec, error) {
 		Compiler:  "clang",
 		Linker:    "cc",
 		CFlags:    []string{"--target=" + triple},
-		GDB:       "gdb",
+		GDB:       []string{"gdb"},
 		PortReset: "false",
 	}
 	if goos == "darwin" {
@@ -253,7 +260,7 @@ func defaultTarget(goos, goarch, triple string) (*TargetSpec, error) {
 	}
 	if goarch != runtime.GOARCH {
 		// Some educated guesses as to how to invoke helper programs.
-		spec.GDB = "gdb-multiarch"
+		spec.GDB = []string{"gdb-multiarch"}
 		if goarch == "arm" && goos == "linux" {
 			spec.CFlags = append(spec.CFlags, "--sysroot=/usr/arm-linux-gnueabihf")
 			spec.Linker = "arm-linux-gnueabihf-gcc"
@@ -271,3 +278,17 @@ func defaultTarget(goos, goarch, triple string) (*TargetSpec, error) {
 	}
 	return &spec, nil
 }
+
+// LookupGDB looks up a gdb executable.
+func (spec *TargetSpec) LookupGDB() (string, error) {
+	if len(spec.GDB) == 0 {
+		return "", errors.New("gdb not configured in the target specification")
+	}
+	for _, d := range spec.GDB {
+		_, err := exec.LookPath(d)
+		if err == nil {
+			return d, nil
+		}
+	}
+	return "", errors.New("no gdb found configured in the target specification (" + strings.Join(spec.GDB, ", ") + ")")
+}

compiler/asserts.go

@@ -159,6 +159,9 @@ func (b *builder) createNilCheck(inst ssa.Value, ptr llvm.Value, blockPrefix str
 	}
 
 	switch inst := inst.(type) {
+	case *ssa.Alloc:
+		// An alloc is never nil.
+		return
 	case *ssa.IndexAddr:
 		// This pointer is the result of an index operation into a slice or
 		// array. Such slices/arrays are already bounds checked so the pointer

compiler/calls.go

@@ -58,10 +58,10 @@ func (b *builder) createCall(fn llvm.Value, args []llvm.Value, name string) llvm
 
 // Expand an argument type to a list that can be used in a function call
 // parameter list.
-func expandFormalParamType(t llvm.Type, name string, goType types.Type) []paramInfo {
+func (c *compilerContext) expandFormalParamType(t llvm.Type, name string, goType types.Type) []paramInfo {
 	switch t.TypeKind() {
 	case llvm.StructTypeKind:
-		fieldInfos := flattenAggregateType(t, name, goType)
+		fieldInfos := c.flattenAggregateType(t, name, goType)
 		if len(fieldInfos) <= maxFieldsPerParam {
 			return fieldInfos
 		} else {
@@ -105,7 +105,7 @@ func (b *builder) expandFormalParamOffsets(t llvm.Type) []uint64 {
 func (b *builder) expandFormalParam(v llvm.Value) []llvm.Value {
 	switch v.Type().TypeKind() {
 	case llvm.StructTypeKind:
-		fieldInfos := flattenAggregateType(v.Type(), "", nil)
+		fieldInfos := b.flattenAggregateType(v.Type(), "", nil)
 		if len(fieldInfos) <= maxFieldsPerParam {
 			fields := b.flattenAggregate(v)
 			if len(fields) != len(fieldInfos) {
@@ -124,12 +124,15 @@ func (b *builder) expandFormalParam(v llvm.Value) []llvm.Value {
 
 // Try to flatten a struct type to a list of types. Returns a 1-element slice
 // with the passed in type if this is not possible.
-func flattenAggregateType(t llvm.Type, name string, goType types.Type) []paramInfo {
+func (c *compilerContext) flattenAggregateType(t llvm.Type, name string, goType types.Type) []paramInfo {
 	typeFlags := getTypeFlags(goType)
 	switch t.TypeKind() {
 	case llvm.StructTypeKind:
-		paramInfos := make([]paramInfo, 0, t.StructElementTypesCount())
+		var paramInfos []paramInfo
 		for i, subfield := range t.StructElementTypes() {
+			if c.targetData.TypeAllocSize(subfield) == 0 {
+				continue
+			}
 			suffix := strconv.Itoa(i)
 			if goType != nil {
 				// Try to come up with a good suffix for this struct field,
@@ -152,7 +155,7 @@ func flattenAggregateType(t llvm.Type, name string, goType types.Type) []paramIn
 					suffix = []string{"context", "funcptr"}[i]
 				}
 			}
-			subInfos := flattenAggregateType(subfield, name+"."+suffix, extractSubfield(goType, i))
+			subInfos := c.flattenAggregateType(subfield, name+"."+suffix, extractSubfield(goType, i))
 			for i := range subInfos {
 				subInfos[i].flags |= typeFlags
 			}
@@ -218,8 +221,11 @@ func extractSubfield(t types.Type, field int) types.Type {
 func (c *compilerContext) flattenAggregateTypeOffsets(t llvm.Type) []uint64 {
 	switch t.TypeKind() {
 	case llvm.StructTypeKind:
-		fields := make([]uint64, 0, t.StructElementTypesCount())
+		var fields []uint64
 		for fieldIndex, field := range t.StructElementTypes() {
+			if c.targetData.TypeAllocSize(field) == 0 {
+				continue
+			}
 			suboffsets := c.flattenAggregateTypeOffsets(field)
 			offset := c.targetData.ElementOffset(t, fieldIndex)
 			for i := range suboffsets {
@@ -238,8 +244,11 @@ func (c *compilerContext) flattenAggregateTypeOffsets(t llvm.Type) []uint64 {
 func (b *builder) flattenAggregate(v llvm.Value) []llvm.Value {
 	switch v.Type().TypeKind() {
 	case llvm.StructTypeKind:
-		fields := make([]llvm.Value, 0, v.Type().StructElementTypesCount())
+		var fields []llvm.Value
-		for i := range v.Type().StructElementTypes() {
+		for i, field := range v.Type().StructElementTypes() {
+			if b.targetData.TypeAllocSize(field) == 0 {
+				continue
+			}
 			subfield := b.CreateExtractValue(v, i, "")
 			subfields := b.flattenAggregate(subfield)
 			fields = append(fields, subfields...)
@@ -266,10 +275,13 @@ func (b *builder) collapseFormalParam(t llvm.Type, fields []llvm.Value) llvm.Val
 func (b *builder) collapseFormalParamInternal(t llvm.Type, fields []llvm.Value) (llvm.Value, []llvm.Value) {
 	switch t.TypeKind() {
 	case llvm.StructTypeKind:
-		flattened := flattenAggregateType(t, "", nil)
+		flattened := b.flattenAggregateType(t, "", nil)
 		if len(flattened) <= maxFieldsPerParam {
 			value := llvm.ConstNull(t)
 			for i, subtyp := range t.StructElementTypes() {
+				if b.targetData.TypeAllocSize(subtyp) == 0 {
+					continue
+				}
 				structField, remaining := b.collapseFormalParamInternal(subtyp, fields)
 				fields = remaining
 				value = b.CreateInsertValue(value, structField, i, "")

compiler/compiler.go

@@ -20,6 +20,11 @@ import (
 	"tinygo.org/x/go-llvm"
 )
 
+// Version of the compiler pacakge. Must be incremented each time the compiler
+// package changes in a way that affects the generated LLVM module.
+// This version is independent of the TinyGo version number.
+const Version = 6 // last change: fix issue 1304
+
 func init() {
 	llvm.InitializeAllTargets()
 	llvm.InitializeAllTargetMCs()
@@ -84,12 +89,13 @@ type compilerContext struct {
 // importantly with a newly created LLVM context and module.
 func newCompilerContext(moduleName string, machine llvm.TargetMachine, config *Config, dumpSSA bool) *compilerContext {
 	c := &compilerContext{
-		Config:     config,
+		Config:      config,
-		DumpSSA:    dumpSSA,
+		DumpSSA:     dumpSSA,
-		difiles:    make(map[string]llvm.Metadata),
+		difiles:     make(map[string]llvm.Metadata),
-		ditypes:    make(map[types.Type]llvm.Metadata),
+		ditypes:     make(map[types.Type]llvm.Metadata),
-		machine:    machine,
+		machine:     machine,
-		targetData: machine.CreateTargetData(),
+		targetData:  machine.CreateTargetData(),
+		astComments: map[string]*ast.CommentGroup{},
 	}
 
 	c.ctx = llvm.NewContext()
@@ -241,21 +247,14 @@ func Sizes(machine llvm.TargetMachine) types.Sizes {
 	}
 }
 
-// CompileProgram compiles the given package path or .go file path. Return an
+// CompilePackage compiles a single package to a LLVM module.
-// error when this fails (in any stage). If successful it returns the LLVM
+func CompilePackage(moduleName string, pkg *loader.Package, ssaPkg *ssa.Package, machine llvm.TargetMachine, config *Config, dumpSSA bool) (llvm.Module, []error) {
-// module. If not, one or more errors will be returned.
+	c := newCompilerContext(moduleName, machine, config, dumpSSA)
-func CompileProgram(lprogram *loader.Program, machine llvm.TargetMachine, config *Config, dumpSSA bool) (llvm.Module, []error) {
+	c.runtimePkg = ssaPkg.Prog.ImportedPackage("runtime").Pkg
-	c := newCompilerContext("", machine, config, dumpSSA)
+	c.program = ssaPkg.Prog
 
-	c.program = lprogram.LoadSSA()
+	// Convert AST to SSA.
-	c.program.Build()
+	ssaPkg.Build()
-	c.runtimePkg = c.program.ImportedPackage("runtime").Pkg
-
-	// Run a simple dead code elimination pass.
-	functions, err := c.simpleDCE(lprogram)
-	if err != nil {
-		return llvm.Module{}, []error{err}
-	}
 
 	// Initialize debug information.
 	if c.Debug {
@@ -268,112 +267,37 @@ func CompileProgram(lprogram *loader.Program, machine llvm.TargetMachine, config
 		})
 	}
 
-	c.loadASTComments(lprogram)
+	// Load comments such as //go:extern on globals.
+	c.loadASTComments(pkg)
 
 	// Predeclare the runtime.alloc function, which is used by the wordpack
 	// functionality.
 	c.getFunction(c.program.ImportedPackage("runtime").Members["alloc"].(*ssa.Function))
 
-	// Add definitions to declarations.
+	// Compile all functions, methods, and global variables in this package.
-	var initFuncs []llvm.Value
 	irbuilder := c.ctx.NewBuilder()
 	defer irbuilder.Dispose()
-	for _, f := range functions {
+	c.createPackage(irbuilder, ssaPkg)
-		if f.Synthetic == "package initializer" {
-			initFuncs = append(initFuncs, c.getFunction(f))
-		}
-		if f.Blocks == nil {
-			continue // external function
-		}
-
-		// Create the function definition.
-		b := newBuilder(c, irbuilder, f)
-		b.createFunction()
-	}
-
-	// After all packages are imported, add a synthetic initializer function
-	// that calls the initializer of each package.
-	initFn := c.program.ImportedPackage("runtime").Members["initAll"].(*ssa.Function)
-	llvmInitFn := c.getFunction(initFn)
-	llvmInitFn.SetLinkage(llvm.InternalLinkage)
-	llvmInitFn.SetUnnamedAddr(true)
-	if c.Debug {
-		difunc := c.attachDebugInfo(initFn)
-		pos := c.program.Fset.Position(initFn.Pos())
-		irbuilder.SetCurrentDebugLocation(uint(pos.Line), uint(pos.Column), difunc, llvm.Metadata{})
-	}
-	llvmInitFn.Param(0).SetName("context")
-	llvmInitFn.Param(1).SetName("parentHandle")
-	block := c.ctx.AddBasicBlock(llvmInitFn, "entry")
-	irbuilder.SetInsertPointAtEnd(block)
-	for _, fn := range initFuncs {
-		irbuilder.CreateCall(fn, []llvm.Value{llvm.Undef(c.i8ptrType), llvm.Undef(c.i8ptrType)}, "")
-	}
-	irbuilder.CreateRetVoid()
 
 	// see: https://reviews.llvm.org/D18355
 	if c.Debug {
 		c.mod.AddNamedMetadataOperand("llvm.module.flags",
 			c.ctx.MDNode([]llvm.Metadata{
 				llvm.ConstInt(c.ctx.Int32Type(), 1, false).ConstantAsMetadata(), // Error on mismatch
-				llvm.GlobalContext().MDString("Debug Info Version"),
+				c.ctx.MDString("Debug Info Version"),
 				llvm.ConstInt(c.ctx.Int32Type(), 3, false).ConstantAsMetadata(), // DWARF version
 			}),
 		)
 		c.mod.AddNamedMetadataOperand("llvm.module.flags",
 			c.ctx.MDNode([]llvm.Metadata{
 				llvm.ConstInt(c.ctx.Int32Type(), 1, false).ConstantAsMetadata(),
-				llvm.GlobalContext().MDString("Dwarf Version"),
+				c.ctx.MDString("Dwarf Version"),
 				llvm.ConstInt(c.ctx.Int32Type(), 4, false).ConstantAsMetadata(),
 			}),
 		)
 		c.dibuilder.Finalize()
 	}
 
-	return c.mod, c.diagnostics
-}
-
-// CompilePackage compiles a single package to a LLVM module.
-func CompilePackage(moduleName string, pkg *loader.Package, machine llvm.TargetMachine, config *Config, dumpSSA bool) (llvm.Module, []error) {
-	c := newCompilerContext(moduleName, machine, config, dumpSSA)
-
-	// Build SSA from AST.
-	ssaPkg := pkg.LoadSSA()
-	ssaPkg.Build()
-
-	// Sort by position, so that the order of the functions in the IR matches
-	// the order of functions in the source file. This is useful for testing,
-	// for example.
-	var members []string
-	for name := range ssaPkg.Members {
-		members = append(members, name)
-	}
-	sort.Slice(members, func(i, j int) bool {
-		iPos := ssaPkg.Members[members[i]].Pos()
-		jPos := ssaPkg.Members[members[j]].Pos()
-		if i == j {
-			// Cannot sort by pos, so do it by name.
-			return members[i] < members[j]
-		}
-		return iPos < jPos
-	})
-
-	// Define all functions.
-	irbuilder := c.ctx.NewBuilder()
-	defer irbuilder.Dispose()
-	for _, name := range members {
-		member := ssaPkg.Members[name]
-		switch member := member.(type) {
-		case *ssa.Function:
-			if member.Blocks == nil {
-				continue // external function
-			}
-			// Create the function definition.
-			b := newBuilder(c, irbuilder, member)
-			b.createFunction()
-		}
-	}
-
 	return c.mod, nil
 }
 
@@ -720,8 +644,9 @@ func (c *compilerContext) attachDebugInfo(f *ssa.Function) llvm.Metadata {
 // debug info is added to the function.
 func (c *compilerContext) attachDebugInfoRaw(f *ssa.Function, llvmFn llvm.Value, suffix, filename string, line int) llvm.Metadata {
 	// Debug info for this function.
-	diparams := make([]llvm.Metadata, 0, len(f.Params))
+	params := getParams(f.Signature)
-	for _, param := range f.Params {
+	diparams := make([]llvm.Metadata, 0, len(params))
+	for _, param := range params {
 		diparams = append(diparams, c.getDIType(param.Type()))
 	}
 	diFuncType := c.dibuilder.CreateSubroutineType(llvm.DISubroutineType{
@@ -759,6 +684,83 @@ func (c *compilerContext) getDIFile(filename string) llvm.Metadata {
 	return c.difiles[filename]
 }
 
+// createPackage builds the LLVM IR for all types, methods, and global variables
+// in the given package.
+func (c *compilerContext) createPackage(irbuilder llvm.Builder, pkg *ssa.Package) {
+	// Sort by position, so that the order of the functions in the IR matches
+	// the order of functions in the source file. This is useful for testing,
+	// for example.
+	var members []string
+	for name := range pkg.Members {
+		members = append(members, name)
+	}
+	sort.Slice(members, func(i, j int) bool {
+		iPos := pkg.Members[members[i]].Pos()
+		jPos := pkg.Members[members[j]].Pos()
+		if i == j {
+			// Cannot sort by pos, so do it by name.
+			return members[i] < members[j]
+		}
+		return iPos < jPos
+	})
+
+	// Define all functions.
+	for _, name := range members {
+		member := pkg.Members[name]
+		switch member := member.(type) {
+		case *ssa.Function:
+			if member.Blocks == nil {
+				continue // external function
+			}
+			// Create the function definition.
+			b := newBuilder(c, irbuilder, member)
+			b.createFunction()
+		case *ssa.Type:
+			if types.IsInterface(member.Type()) {
+				// Interfaces don't have concrete methods.
+				continue
+			}
+
+			// Named type. We should make sure all methods are created.
+			// This includes both functions with pointer receivers and those
+			// without.
+			methods := getAllMethods(pkg.Prog, member.Type())
+			methods = append(methods, getAllMethods(pkg.Prog, types.NewPointer(member.Type()))...)
+			for _, method := range methods {
+				// Parse this method.
+				fn := pkg.Prog.MethodValue(method)
+				if fn.Blocks == nil {
+					continue // external function
+				}
+				if member.Type().String() != member.String() {
+					// This is a member on a type alias. Do not build such a
+					// function.
+					continue
+				}
+				if fn.Synthetic != "" && fn.Synthetic != "package initializer" {
+					// This function is a kind of wrapper function (created by
+					// the ssa package, not appearing in the source code) that
+					// is created by the getFunction method as needed.
+					// Therefore, don't build it here to avoid "function
+					// redeclared" errors.
+					continue
+				}
+				// Create the function definition.
+				b := newBuilder(c, irbuilder, fn)
+				b.createFunction()
+			}
+		case *ssa.Global:
+			// Global variable.
+			info := c.getGlobalInfo(member)
+			if !info.extern {
+				global := c.getGlobal(member)
+				global.SetInitializer(llvm.ConstNull(global.Type().ElementType()))
+				global.SetVisibility(llvm.HiddenVisibility)
+			}
+		}
+	}
+}
+
 // createFunction builds the LLVM IR implementation for this function. The
 // function must not yet be defined, otherwise this function will create a
 // diagnostic.
@@ -767,7 +769,7 @@ func (b *builder) createFunction() {
 		fmt.Printf("\nfunc %s:\n", b.fn)
 	}
 	if !b.llvmFn.IsDeclaration() {
-		errValue := b.fn.Name() + " redeclared in this program"
+		errValue := b.llvmFn.Name() + " redeclared in this program"
 		fnPos := getPosition(b.llvmFn)
 		if fnPos.IsValid() {
 			errValue += "\n\tprevious declaration at " + fnPos.String()
@@ -776,9 +778,15 @@ func (b *builder) createFunction() {
 		return
 	}
 	if !b.info.exported {
-		b.llvmFn.SetLinkage(llvm.InternalLinkage)
+		b.llvmFn.SetVisibility(llvm.HiddenVisibility)
 		b.llvmFn.SetUnnamedAddr(true)
 	}
+	if b.info.exported && strings.HasPrefix(b.Triple, "wasm") {
+		// Set the exported name. This is necessary for WebAssembly because
+		// otherwise the function is not exported.
+		functionAttr := b.ctx.CreateStringAttribute("wasm-export-name", b.info.linkName)
+		b.llvmFn.AddFunctionAttr(functionAttr)
+	}
 
 	// Some functions have a pragma controlling the inlining level.
 	switch b.info.inline {
@@ -821,7 +829,7 @@ func (b *builder) createFunction() {
 	for _, param := range b.fn.Params {
 		llvmType := b.getLLVMType(param.Type())
 		fields := make([]llvm.Value, 0, 1)
-		for _, info := range expandFormalParamType(llvmType, param.Name(), param.Type()) {
+		for _, info := range b.expandFormalParamType(llvmType, param.Name(), param.Type()) {
 			param := b.llvmFn.Param(llvmParamIndex)
 			param.SetName(info.name)
 			fields = append(fields, param)
@@ -948,6 +956,12 @@ func (b *builder) createFunction() {
 			b.trackValue(phi.llvm)
 		}
 	}
+
+	// Create anonymous functions (closures etc.).
+	for _, sub := range b.fn.AnonFuncs {
+		b := newBuilder(b.compilerContext, b.Builder, sub)
+		b.createFunction()
+	}
 }
 
 // createInstruction builds the LLVM IR equivalent instructions for the

compiler/compiler_test.go

@@ -4,7 +4,6 @@ import (
 	"flag"
 	"go/types"
 	"io/ioutil"
-	"regexp"
 	"strconv"
 	"strings"
 	"testing"
@@ -20,6 +19,19 @@ var flagUpdate = flag.Bool("update", false, "update tests based on test output")
 // Basic tests for the compiler. Build some Go files and compare the output with
 // the expected LLVM IR for regression testing.
 func TestCompiler(t *testing.T) {
+	// Check LLVM version.
+	llvmMajor, err := strconv.Atoi(strings.SplitN(llvm.Version, ".", 2)[0])
+	if err != nil {
+		t.Fatal("could not parse LLVM version:", llvm.Version)
+	}
+	if llvmMajor < 11 {
+		// It is likely this version needs to be bumped in the future.
+		// The goal is to at least test the LLVM version that's used by default
+		// in TinyGo and (if possible without too many workarounds) also some
+		// previous versions.
+		t.Skip("compiler tests require LLVM 11 or above, got LLVM ", llvm.Version)
+	}
+
 	target, err := compileopts.LoadTarget("i686--linux")
 	if err != nil {
 		t.Fatal("failed to load target:", err)
@@ -47,7 +59,9 @@ func TestCompiler(t *testing.T) {
 		"basic.go",
 		"pointer.go",
 		"slice.go",
+		"string.go",
 		"float.go",
+		"interface.go",
 	}
 
 	for _, testCase := range tests {
@@ -65,8 +79,9 @@ func TestCompiler(t *testing.T) {
 			}
 
 			// Compile AST to IR.
+			program := lprogram.LoadSSA()
 			pkg := lprogram.MainPkg()
-			mod, errs := CompilePackage(testCase, pkg, machine, compilerConfig, false)
+			mod, errs := CompilePackage(testCase, pkg, program.Package(pkg.Pkg), machine, compilerConfig, false)
 			if errs != nil {
 				for _, err := range errs {
 					t.Log("error:", err)
@@ -107,8 +122,6 @@ func TestCompiler(t *testing.T) {
 	}
 }
 
-var alignRegexp = regexp.MustCompile(", align [0-9]+$")
-
 // fuzzyEqualIR returns true if the two LLVM IR strings passed in are roughly
 // equal. That means, only relevant lines are compared (excluding comments
 // etc.).
@@ -120,15 +133,6 @@ func fuzzyEqualIR(s1, s2 string) bool {
 	}
 	for i, line1 := range lines1 {
 		line2 := lines2[i]
-		match1 := alignRegexp.MatchString(line1)
-		match2 := alignRegexp.MatchString(line2)
-		if match1 != match2 {
-			// Only one of the lines has the align keyword. Remove it.
-			// This is a change to make the test work in both LLVM 10 and LLVM
-			// 11 (LLVM 11 appears to automatically add alignment everywhere).
-			line1 = alignRegexp.ReplaceAllString(line1, "")
-			line2 = alignRegexp.ReplaceAllString(line2, "")
-		}
 		if line1 != line2 {
 			return false
 		}
@@ -142,12 +146,6 @@ func fuzzyEqualIR(s1, s2 string) bool {
 // stripped out.
 func filterIrrelevantIRLines(lines []string) []string {
 	var out []string
-	llvmVersion, err := strconv.Atoi(strings.Split(llvm.Version, ".")[0])
-	if err != nil {
-		// Note: this should never happen and if it does, it will always happen
-		// for a particular build because llvm.Version is a constant.
-		panic(err)
-	}
 	for _, line := range lines {
 		line = strings.Split(line, ";")[0]    // strip out comments/info
 		line = strings.TrimRight(line, "\r ") // drop '\r' on Windows and remove trailing spaces from comments
@@ -157,15 +155,6 @@ func filterIrrelevantIRLines(lines []string) []string {
 		if strings.HasPrefix(line, "source_filename = ") {
 			continue
 		}
-		if llvmVersion < 10 && strings.HasPrefix(line, "attributes ") {
-			// Ignore attribute groups. These may change between LLVM versions.
-			// Right now test outputs are for LLVM 10.
-			continue
-		}
-		if llvmVersion < 10 && strings.HasPrefix(line, "target datalayout ") {
-			// Ignore the target layout. This may change between LLVM versions.
-			continue
-		}
 		out = append(out, line)
 	}
 	return out

compiler/defer.go

@@ -352,7 +352,7 @@ func (b *builder) createRunDefers() {
 
 			// Get the real defer struct type and cast to it.
 			valueTypes := []llvm.Type{b.uintptrType, llvm.PointerType(b.getLLVMRuntimeType("_defer"), 0)}
-			for _, param := range callback.Params {
+			for _, param := range getParams(callback.Signature) {
 				valueTypes = append(valueTypes, b.getLLVMType(param.Type()))
 			}
 			deferFrameType := b.ctx.StructType(valueTypes, false)
@@ -361,7 +361,7 @@ func (b *builder) createRunDefers() {
 			// Extract the params from the struct.
 			forwardParams := []llvm.Value{}
 			zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
-			for i := range callback.Params {
+			for i := range getParams(callback.Signature) {
 				gep := b.CreateInBoundsGEP(deferFramePtr, []llvm.Value{zero, llvm.ConstInt(b.ctx.Int32Type(), uint64(i+2), false)}, "gep")
 				forwardParam := b.CreateLoad(gep, "param")
 				forwardParams = append(forwardParams, forwardParam)

compiler/func.go

@@ -37,7 +37,7 @@ func (c *compilerContext) createFuncValue(builder llvm.Builder, funcPtr, context
 			funcValueWithSignatureGlobal = llvm.AddGlobal(c.mod, funcValueWithSignatureType, funcValueWithSignatureGlobalName)
 			funcValueWithSignatureGlobal.SetInitializer(funcValueWithSignature)
 			funcValueWithSignatureGlobal.SetGlobalConstant(true)
-			funcValueWithSignatureGlobal.SetLinkage(llvm.InternalLinkage)
+			funcValueWithSignatureGlobal.SetLinkage(llvm.LinkOnceODRLinkage)
 		}
 		funcValueScalar = llvm.ConstPtrToInt(funcValueWithSignatureGlobal, c.uintptrType)
 	default:
@@ -124,13 +124,13 @@ func (c *compilerContext) getRawFuncType(typ *types.Signature) llvm.Type {
 			// The receiver is not an interface, but a i8* type.
 			recv = c.i8ptrType
 		}
-		for _, info := range expandFormalParamType(recv, "", nil) {
+		for _, info := range c.expandFormalParamType(recv, "", nil) {
 			paramTypes = append(paramTypes, info.llvmType)
 		}
 	}
 	for i := 0; i < typ.Params().Len(); i++ {
 		subType := c.getLLVMType(typ.Params().At(i).Type())
-		for _, info := range expandFormalParamType(subType, "", nil) {
+		for _, info := range c.expandFormalParamType(subType, "", nil) {
 			paramTypes = append(paramTypes, info.llvmType)
 		}
 	}

compiler/goroutine.go

@@ -84,7 +84,7 @@ func (c *compilerContext) createGoroutineStartWrapper(fn llvm.Value, prefix stri
 		// Create the wrapper.
 		wrapperType := llvm.FunctionType(c.ctx.VoidType(), []llvm.Type{c.i8ptrType}, false)
 		wrapper = llvm.AddFunction(c.mod, name+"$gowrapper", wrapperType)
-		wrapper.SetLinkage(llvm.InternalLinkage)
+		wrapper.SetLinkage(llvm.LinkOnceODRLinkage)
 		wrapper.SetUnnamedAddr(true)
 		wrapper.AddAttributeAtIndex(-1, c.ctx.CreateStringAttribute("tinygo-gowrapper", name))
 		entry := c.ctx.AddBasicBlock(wrapper, "entry")
@@ -141,7 +141,7 @@ func (c *compilerContext) createGoroutineStartWrapper(fn llvm.Value, prefix stri
 		// Create the wrapper.
 		wrapperType := llvm.FunctionType(c.ctx.VoidType(), []llvm.Type{c.i8ptrType}, false)
 		wrapper = llvm.AddFunction(c.mod, prefix+".gowrapper", wrapperType)
-		wrapper.SetLinkage(llvm.InternalLinkage)
+		wrapper.SetLinkage(llvm.LinkOnceODRLinkage)
 		wrapper.SetUnnamedAddr(true)
 		wrapper.AddAttributeAtIndex(-1, c.ctx.CreateStringAttribute("tinygo-gowrapper", ""))
 		entry := c.ctx.AddBasicBlock(wrapper, "entry")

compiler/interface.go

@@ -24,16 +24,7 @@ import (
 func (b *builder) createMakeInterface(val llvm.Value, typ types.Type, pos token.Pos) llvm.Value {
 	itfValue := b.emitPointerPack([]llvm.Value{val})
 	itfTypeCodeGlobal := b.getTypeCode(typ)
-	itfMethodSetGlobal := b.getTypeMethodSet(typ)
+	itfTypeCode := b.CreatePtrToInt(itfTypeCodeGlobal, b.uintptrType, "")
-	itfConcreteTypeGlobal := b.mod.NamedGlobal("typeInInterface:" + itfTypeCodeGlobal.Name())
-	if itfConcreteTypeGlobal.IsNil() {
-		typeInInterface := b.getLLVMRuntimeType("typeInInterface")
-		itfConcreteTypeGlobal = llvm.AddGlobal(b.mod, typeInInterface, "typeInInterface:"+itfTypeCodeGlobal.Name())
-		itfConcreteTypeGlobal.SetInitializer(llvm.ConstNamedStruct(typeInInterface, []llvm.Value{itfTypeCodeGlobal, itfMethodSetGlobal}))
-		itfConcreteTypeGlobal.SetGlobalConstant(true)
-		itfConcreteTypeGlobal.SetLinkage(llvm.PrivateLinkage)
-	}
-	itfTypeCode := b.CreatePtrToInt(itfConcreteTypeGlobal, b.uintptrType, "")
 	itf := llvm.Undef(b.getLLVMRuntimeType("_interface"))
 	itf = b.CreateInsertValue(itf, itfTypeCode, 0, "")
 	itf = b.CreateInsertValue(itf, itfValue, 1, "")
@@ -54,6 +45,7 @@ func (c *compilerContext) getTypeCode(typ types.Type) llvm.Value {
 		// reflect lowering simpler.
 		var references llvm.Value
 		var length int64
+		var methodSet llvm.Value
 		switch typ := typ.(type) {
 		case *types.Named:
 			references = c.getTypeCode(typ.Underlying())
@@ -70,17 +62,28 @@ func (c *compilerContext) getTypeCode(typ types.Type) llvm.Value {
 			// Take a pointer to the typecodeID of the first field (if it exists).
 			structGlobal := c.makeStructTypeFields(typ)
 			references = llvm.ConstBitCast(structGlobal, global.Type())
+		case *types.Interface:
+			methodSetGlobal := c.getInterfaceMethodSet(typ)
+			references = llvm.ConstBitCast(methodSetGlobal, global.Type())
 		}
-		if !references.IsNil() {
+		if _, ok := typ.Underlying().(*types.Interface); !ok {
+			methodSet = c.getTypeMethodSet(typ)
+		}
+		if !references.IsNil() || length != 0 || !methodSet.IsNil() {
 			// Set the 'references' field of the runtime.typecodeID struct.
 			globalValue := llvm.ConstNull(global.Type().ElementType())
-			globalValue = llvm.ConstInsertValue(globalValue, references, []uint32{0})
+			if !references.IsNil() {
+				globalValue = llvm.ConstInsertValue(globalValue, references, []uint32{0})
+			}
 			if length != 0 {
 				lengthValue := llvm.ConstInt(c.uintptrType, uint64(length), false)
 				globalValue = llvm.ConstInsertValue(globalValue, lengthValue, []uint32{1})
 			}
+			if !methodSet.IsNil() {
+				globalValue = llvm.ConstInsertValue(globalValue, methodSet, []uint32{2})
+			}
 			global.SetInitializer(globalValue)
-			global.SetLinkage(llvm.PrivateLinkage)
+			global.SetLinkage(llvm.LinkOnceODRLinkage)
 		}
 		global.SetGlobalConstant(true)
 	}
@@ -103,8 +106,8 @@ func (c *compilerContext) makeStructTypeFields(typ *types.Struct) llvm.Value {
 		fieldName.SetLinkage(llvm.PrivateLinkage)
 		fieldName.SetUnnamedAddr(true)
 		fieldName = llvm.ConstGEP(fieldName, []llvm.Value{
-			llvm.ConstInt(llvm.Int32Type(), 0, false),
+			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
-			llvm.ConstInt(llvm.Int32Type(), 0, false),
+			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
 		})
 		fieldGlobalValue = llvm.ConstInsertValue(fieldGlobalValue, fieldName, []uint32{1})
 		if typ.Tag(i) != "" {
@@ -112,8 +115,8 @@ func (c *compilerContext) makeStructTypeFields(typ *types.Struct) llvm.Value {
 			fieldTag.SetLinkage(llvm.PrivateLinkage)
 			fieldTag.SetUnnamedAddr(true)
 			fieldTag = llvm.ConstGEP(fieldTag, []llvm.Value{
-				llvm.ConstInt(llvm.Int32Type(), 0, false),
+				llvm.ConstInt(c.ctx.Int32Type(), 0, false),
-				llvm.ConstInt(llvm.Int32Type(), 0, false),
+				llvm.ConstInt(c.ctx.Int32Type(), 0, false),
 			})
 			fieldGlobalValue = llvm.ConstInsertValue(fieldGlobalValue, fieldTag, []uint32{2})
 		}
@@ -186,7 +189,7 @@ func getTypeCodeName(t types.Type) string {
 	case *types.Interface:
 		methods := make([]string, t.NumMethods())
 		for i := 0; i < t.NumMethods(); i++ {
-			methods[i] = getTypeCodeName(t.Method(i).Type())
+			methods[i] = t.Method(i).Name() + ":" + getTypeCodeName(t.Method(i).Type())
 		}
 		return "interface:" + "{" + strings.Join(methods, ",") + "}"
 	case *types.Map:
@@ -264,7 +267,7 @@ func (c *compilerContext) getTypeMethodSet(typ types.Type) llvm.Value {
 	global = llvm.AddGlobal(c.mod, arrayType, typ.String()+"$methodset")
 	global.SetInitializer(value)
 	global.SetGlobalConstant(true)
-	global.SetLinkage(llvm.PrivateLinkage)
+	global.SetLinkage(llvm.LinkOnceODRLinkage)
 	return llvm.ConstGEP(global, []llvm.Value{zero, zero})
 }
 
@@ -295,7 +298,7 @@ func (c *compilerContext) getInterfaceMethodSet(typ types.Type) llvm.Value {
 	global = llvm.AddGlobal(c.mod, value.Type(), name+"$interface")
 	global.SetInitializer(value)
 	global.SetGlobalConstant(true)
-	global.SetLinkage(llvm.PrivateLinkage)
+	global.SetLinkage(llvm.LinkOnceODRLinkage)
 	return llvm.ConstGEP(global, []llvm.Value{zero, zero})
 }
 
@@ -338,10 +341,16 @@ func (b *builder) createTypeAssert(expr *ssa.TypeAssert) llvm.Value {
 		commaOk = b.createRuntimeCall("interfaceImplements", []llvm.Value{actualTypeNum, methodSet}, "")
 
 	} else {
+		globalName := "reflect/types.type:" + getTypeCodeName(expr.AssertedType) + "$id"
+		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.
-		assertedTypeCodeGlobal := b.getTypeCode(expr.AssertedType)
 		commaOk = b.createRuntimeCall("typeAssert", []llvm.Value{actualTypeNum, assertedTypeCodeGlobal}, "typecode")
 	}
 
@@ -445,9 +454,9 @@ func (c *compilerContext) getInterfaceInvokeWrapper(fn *ssa.Function, llvmFn llv
 	}
 
 	// Get the expanded receiver type.
-	receiverType := c.getLLVMType(fn.Params[0].Type())
+	receiverType := c.getLLVMType(fn.Signature.Recv().Type())
 	var expandedReceiverType []llvm.Type
-	for _, info := range expandFormalParamType(receiverType, "", nil) {
+	for _, info := range c.expandFormalParamType(receiverType, "", nil) {
 		expandedReceiverType = append(expandedReceiverType, info.llvmType)
 	}
 
@@ -467,7 +476,7 @@ func (c *compilerContext) getInterfaceInvokeWrapper(fn *ssa.Function, llvmFn llv
 	wrapper = llvm.AddFunction(c.mod, wrapperName, wrapFnType)
 	wrapper.LastParam().SetName("parentHandle")
 
-	wrapper.SetLinkage(llvm.InternalLinkage)
+	wrapper.SetLinkage(llvm.LinkOnceODRLinkage)
 	wrapper.SetUnnamedAddr(true)
 
 	// Create a new builder just to create this wrapper.

compiler/interrupt.go

@@ -46,7 +46,7 @@ func (b *builder) createInterruptGlobal(instr *ssa.CallCommon) (llvm.Value, erro
 		return llvm.Value{}, b.makeError(instr.Pos(), "interrupt redeclared in this program")
 	}
 	global := llvm.AddGlobal(b.mod, globalLLVMType, globalName)
-	global.SetLinkage(llvm.PrivateLinkage)
+	global.SetVisibility(llvm.HiddenVisibility)
 	global.SetGlobalConstant(true)
 	global.SetUnnamedAddr(true)
 	initializer := llvm.ConstNull(globalLLVMType)

compiler/passes.go

@@ -1,164 +0,0 @@
-package compiler
-
-// This file implements a simple reachability analysis, to reduce compile time.
-// This DCE pass used to be necessary for improving other passes but now it
-// isn't necessary anymore.
-
-import (
-	"errors"
-	"go/types"
-	"sort"
-
-	"github.com/tinygo-org/tinygo/loader"
-	"golang.org/x/tools/go/ssa"
-)
-
-type dceState struct {
-	*compilerContext
-	functions   []*dceFunction
-	functionMap map[*ssa.Function]*dceFunction
-}
-
-type dceFunction struct {
-	*ssa.Function
-	functionInfo
-	flag bool // used by dead code elimination
-}
-
-func (p *dceState) addFunction(ssaFn *ssa.Function) {
-	if _, ok := p.functionMap[ssaFn]; ok {
-		return
-	}
-	f := &dceFunction{Function: ssaFn}
-	f.functionInfo = p.getFunctionInfo(ssaFn)
-	p.functions = append(p.functions, f)
-	p.functionMap[ssaFn] = f
-
-	for _, anon := range ssaFn.AnonFuncs {
-		p.addFunction(anon)
-	}
-}
-
-// simpleDCE returns a list of alive functions in the program. Compiling only
-// these functions makes the compiler faster.
-//
-// This functionality will likely be replaced in the future with build caching.
-func (c *compilerContext) simpleDCE(lprogram *loader.Program) ([]*ssa.Function, error) {
-	mainPkg := c.program.Package(lprogram.MainPkg().Pkg)
-	if mainPkg == nil {
-		panic("could not find main package")
-	}
-	p := &dceState{
-		compilerContext: c,
-		functionMap:     make(map[*ssa.Function]*dceFunction),
-	}
-
-	for _, pkg := range lprogram.Sorted() {
-		pkg := c.program.Package(pkg.Pkg)
-		memberNames := make([]string, 0)
-		for name := range pkg.Members {
-			memberNames = append(memberNames, name)
-		}
-		sort.Strings(memberNames)
-
-		for _, name := range memberNames {
-			member := pkg.Members[name]
-			switch member := member.(type) {
-			case *ssa.Function:
-				p.addFunction(member)
-			case *ssa.Type:
-				methods := getAllMethods(pkg.Prog, member.Type())
-				if !types.IsInterface(member.Type()) {
-					// named type
-					for _, method := range methods {
-						p.addFunction(pkg.Prog.MethodValue(method))
-					}
-				}
-			case *ssa.Global:
-				// Ignore. Globals are not handled here.
-			case *ssa.NamedConst:
-				// Ignore: these are already resolved.
-			default:
-				panic("unknown member type: " + member.String())
-			}
-		}
-	}
-
-	// Initial set of live functions. Include main.main, *.init and runtime.*
-	// functions.
-	main, ok := mainPkg.Members["main"].(*ssa.Function)
-	if !ok {
-		if mainPkg.Members["main"] == nil {
-			return nil, errors.New("function main is undeclared in the main package")
-		} else {
-			return nil, errors.New("cannot declare main - must be func")
-		}
-	}
-	runtimePkg := c.program.ImportedPackage("runtime")
-	mathPkg := c.program.ImportedPackage("math")
-	taskPkg := c.program.ImportedPackage("internal/task")
-	p.functionMap[main].flag = true
-	worklist := []*ssa.Function{main}
-	for _, f := range p.functions {
-		if f.exported || f.Synthetic == "package initializer" || f.Pkg == runtimePkg || f.Pkg == taskPkg || (f.Pkg == mathPkg && f.Pkg != nil) {
-			if f.flag {
-				continue
-			}
-			f.flag = true
-			worklist = append(worklist, f.Function)
-		}
-	}
-
-	// Mark all called functions recursively.
-	for len(worklist) != 0 {
-		f := worklist[len(worklist)-1]
-		worklist = worklist[:len(worklist)-1]
-		for _, block := range f.Blocks {
-			for _, instr := range block.Instrs {
-				if instr, ok := instr.(*ssa.MakeInterface); ok {
-					for _, sel := range getAllMethods(c.program, instr.X.Type()) {
-						fn := c.program.MethodValue(sel)
-						callee := p.functionMap[fn]
-						if callee == nil {
-							// TODO: why is this necessary?
-							p.addFunction(fn)
-							callee = p.functionMap[fn]
-						}
-						if !callee.flag {
-							callee.flag = true
-							worklist = append(worklist, callee.Function)
-						}
-					}
-				}
-				for _, operand := range instr.Operands(nil) {
-					if operand == nil || *operand == nil {
-						continue
-					}
-					switch operand := (*operand).(type) {
-					case *ssa.Function:
-						f := p.functionMap[operand]
-						if f == nil {
-							// FIXME HACK: this function should have been
-							// discovered already. It is not for bound methods.
-							p.addFunction(operand)
-							f = p.functionMap[operand]
-						}
-						if !f.flag {
-							f.flag = true
-							worklist = append(worklist, operand)
-						}
-					}
-				}
-			}
-		}
-	}
-
-	// Return all live functions.
-	liveFunctions := []*ssa.Function{}
-	for _, f := range p.functions {
-		if f.flag {
-			liveFunctions = append(liveFunctions, f.Function)
-		}
-	}
-	return liveFunctions, nil
-}

compiler/sizes.go

@@ -150,6 +150,9 @@ func (s *stdSizes) Sizeof(T types.Type) int64 {
 		return s.PtrSize * 2
 	case *types.Pointer:
 		return s.PtrSize
+	case *types.Signature:
+		// Func values in TinyGo are two words in size.
+		return s.PtrSize * 2
 	default:
 		panic("unknown type: " + t.String())
 	}

compiler/symbol.go

@@ -72,9 +72,9 @@ func (c *compilerContext) getFunction(fn *ssa.Function) llvm.Value {
 	}
 
 	var paramInfos []paramInfo
-	for _, param := range fn.Params {
+	for _, param := range getParams(fn.Signature) {
 		paramType := c.getLLVMType(param.Type())
-		paramFragmentInfos := expandFormalParamType(paramType, param.Name(), param.Type())
+		paramFragmentInfos := c.expandFormalParamType(paramType, param.Name(), param.Type())
 		paramInfos = append(paramInfos, paramFragmentInfos...)
 	}
 
@@ -172,6 +172,21 @@ func (c *compilerContext) getFunction(fn *ssa.Function) llvm.Value {
 		}
 	}
 
+	// Synthetic functions are functions that do not appear in the source code,
+	// they are artificially constructed. Usually they are wrapper functions
+	// that are not referenced anywhere except in a SSA call instruction so
+	// should be created right away.
+	// The exception is the package initializer, which does appear in the
+	// *ssa.Package members and so shouldn't be created here.
+	if fn.Synthetic != "" && fn.Synthetic != "package initializer" {
+		irbuilder := c.ctx.NewBuilder()
+		b := newBuilder(c, irbuilder, fn)
+		b.createFunction()
+		irbuilder.Dispose()
+		llvmFn.SetLinkage(llvm.LinkOnceODRLinkage)
+		llvmFn.SetUnnamedAddr(true)
+	}
+
 	return llvmFn
 }
 
@@ -278,6 +293,20 @@ func (info *functionInfo) parsePragmas(f *ssa.Function) {
 	}
 }
 
+// getParams returns the function parameters, including the receiver at the
+// start. This is an alternative to the Params member of *ssa.Function, which is
+// not yet populated when the package has not yet been built.
+func getParams(sig *types.Signature) []*types.Var {
+	params := []*types.Var{}
+	if sig.Recv() != nil {
+		params = append(params, sig.Recv())
+	}
+	for i := 0; i < sig.Params().Len(); i++ {
+		params = append(params, sig.Params().At(i))
+	}
+	return params
+}
+
 // globalInfo contains some information about a specific global. By default,
 // linkName is equal to .RelString(nil) on a global and extern is false, but for
 // some symbols this is different (due to //go:extern for example).
@@ -289,25 +318,22 @@ type globalInfo struct {
 
 // loadASTComments loads comments on globals from the AST, for use later in the
 // program. In particular, they are required for //go:extern pragmas on globals.
-func (c *compilerContext) loadASTComments(lprogram *loader.Program) {
+func (c *compilerContext) loadASTComments(pkg *loader.Package) {
-	c.astComments = map[string]*ast.CommentGroup{}
+	for _, file := range pkg.Files {
-	for _, pkgInfo := range lprogram.Sorted() {
+		for _, decl := range file.Decls {
-		for _, file := range pkgInfo.Files {
+			switch decl := decl.(type) {
-			for _, decl := range file.Decls {
+			case *ast.GenDecl:
-				switch decl := decl.(type) {
+				switch decl.Tok {
-				case *ast.GenDecl:
+				case token.VAR:
-					switch decl.Tok {
+					if len(decl.Specs) != 1 {
-					case token.VAR:
+						continue
-						if len(decl.Specs) != 1 {
+					}
-							continue
+					for _, spec := range decl.Specs {
-						}
+						switch spec := spec.(type) {
-						for _, spec := range decl.Specs {
+						case *ast.ValueSpec: // decl.Tok == token.VAR
-							switch spec := spec.(type) {
+							for _, name := range spec.Names {
-							case *ast.ValueSpec: // decl.Tok == token.VAR
+								id := pkg.Pkg.Path() + "." + name.Name
-								for _, name := range spec.Names {
+								c.astComments[id] = decl.Doc
-									id := pkgInfo.Pkg.Path() + "." + name.Name
-									c.astComments[id] = decl.Doc
-								}
 							}
 						}
 					}
@@ -326,10 +352,6 @@ func (c *compilerContext) getGlobal(g *ssa.Global) llvm.Value {
 		typ := g.Type().(*types.Pointer).Elem()
 		llvmType := c.getLLVMType(typ)
 		llvmGlobal = llvm.AddGlobal(c.mod, llvmType, info.linkName)
-		if !info.extern {
-			llvmGlobal.SetInitializer(llvm.ConstNull(llvmType))
-			llvmGlobal.SetLinkage(llvm.InternalLinkage)
-		}
 
 		// Set alignment from the //go:align comment.
 		var alignInBits uint32
@@ -347,8 +369,6 @@ func (c *compilerContext) getGlobal(g *ssa.Global) llvm.Value {
 
 		if c.Debug && !info.extern {
 			// Add debug info.
-			// TODO: this should be done for every global in the program, not just
-			// the ones that are referenced from some code.
 			pos := c.program.Fset.Position(g.Pos())
 			diglobal := c.dibuilder.CreateGlobalVariableExpression(c.difiles[pos.Filename], llvm.DIGlobalVariableExpression{
 				Name:        g.RelString(nil),

compiler/testdata/basic.ll

@@ -3,70 +3,72 @@ source_filename = "basic.go"
 target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-f64:32:64-f80:32-n8:16:32-S128"
 target triple = "i686--linux"
 
-define internal void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   ret void
 }
 
-define internal i32 @main.addInt(i32 %x, i32 %y, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i32 @main.addInt(i32 %x, i32 %y, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = add i32 %x, %y
   ret i32 %0
 }
 
-define internal i1 @main.equalInt(i32 %x, i32 %y, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i1 @main.equalInt(i32 %x, i32 %y, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = icmp eq i32 %x, %y
   ret i1 %0
 }
 
-define internal i1 @main.floatEQ(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i1 @main.floatEQ(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = fcmp oeq float %x, %y
   ret i1 %0
 }
 
-define internal i1 @main.floatNE(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i1 @main.floatNE(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = fcmp une float %x, %y
   ret i1 %0
 }
 
-define internal i1 @main.floatLower(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i1 @main.floatLower(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = fcmp olt float %x, %y
   ret i1 %0
 }
 
-define internal i1 @main.floatLowerEqual(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i1 @main.floatLowerEqual(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = fcmp ole float %x, %y
   ret i1 %0
 }
 
-define internal i1 @main.floatGreater(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i1 @main.floatGreater(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = fcmp ogt float %x, %y
   ret i1 %0
 }
 
-define internal i1 @main.floatGreaterEqual(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i1 @main.floatGreaterEqual(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = fcmp oge float %x, %y
   ret i1 %0
 }
 
-define internal float @main.complexReal(float %x.r, float %x.i, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden float @main.complexReal(float %x.r, float %x.i, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   ret float %x.r
 }
 
-define internal float @main.complexImag(float %x.r, float %x.i, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden float @main.complexImag(float %x.r, float %x.i, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   ret float %x.i
 }
 
-define internal { float, float } @main.complexAdd(float %x.r, float %x.i, float %y.r, float %y.i, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden { float, float } @main.complexAdd(float %x.r, float %x.i, float %y.r, float %y.i, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = fadd float %x.r, %y.r
   %1 = fadd float %x.i, %y.i
@@ -75,7 +77,7 @@ entry:
   ret { float, float } %3
 }
 
-define internal { float, float } @main.complexSub(float %x.r, float %x.i, float %y.r, float %y.i, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden { float, float } @main.complexSub(float %x.r, float %x.i, float %y.r, float %y.i, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = fsub float %x.r, %y.r
   %1 = fsub float %x.i, %y.i
@@ -84,7 +86,7 @@ entry:
   ret { float, float } %3
 }
 
-define internal { float, float } @main.complexMul(float %x.r, float %x.i, float %y.r, float %y.i, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden { float, float } @main.complexMul(float %x.r, float %x.i, float %y.r, float %y.i, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = fmul float %x.r, %y.r
   %1 = fmul float %x.i, %y.i

compiler/testdata/float.ll

@@ -3,12 +3,14 @@ source_filename = "float.go"
 target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-f64:32:64-f80:32-n8:16:32-S128"
 target triple = "i686--linux"
 
-define internal void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   ret void
 }
 
-define internal i32 @main.f32tou32(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i32 @main.f32tou32(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %positive = fcmp oge float %v, 0.000000e+00
   %withinmax = fcmp ole float %v, 0x41EFFFFFC0000000
@@ -19,22 +21,22 @@ entry:
   ret i32 %0
 }
 
-define internal float @main.maxu32f(i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden float @main.maxu32f(i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   ret float 0x41F0000000000000
 }
 
-define internal i32 @main.maxu32tof32(i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i32 @main.maxu32tof32(i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   ret i32 -1
 }
 
-define internal { i32, i32, i32, i32 } @main.inftoi32(i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden { i32, i32, i32, i32 } @main.inftoi32(i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   ret { i32, i32, i32, i32 } { i32 -1, i32 0, i32 2147483647, i32 -2147483648 }
 }
 
-define internal i32 @main.u32tof32tou32(i32 %v, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i32 @main.u32tof32tou32(i32 %v, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = uitofp i32 %v to float
   %withinmax = fcmp ole float %0, 0x41EFFFFFC0000000
@@ -43,7 +45,7 @@ entry:
   ret i32 %1
 }
 
-define internal float @main.f32tou32tof32(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden float @main.f32tou32tof32(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %positive = fcmp oge float %v, 0.000000e+00
   %withinmax = fcmp ole float %v, 0x41EFFFFFC0000000
@@ -55,7 +57,7 @@ entry:
   ret float %1
 }
 
-define internal i8 @main.f32tou8(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i8 @main.f32tou8(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %positive = fcmp oge float %v, 0.000000e+00
   %withinmax = fcmp ole float %v, 2.550000e+02
@@ -66,7 +68,7 @@ entry:
   ret i8 %0
 }
 
-define internal i8 @main.f32toi8(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i8 @main.f32toi8(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %abovemin = fcmp oge float %v, -1.280000e+02
   %belowmax = fcmp ole float %v, 1.270000e+02

compiler/testdata/interface.go

@@ -0,0 +1,54 @@
+// This file tests interface types and interface builtins.
+
+package main
+
+// Test interface construction.
+
+func simpleType() interface{} {
+	return 0
+}
+
+func pointerType() interface{} {
+	// Pointers have an element type, in this case int.
+	var v *int
+	return v
+}
+
+func interfaceType() interface{} {
+	// Interfaces can exist in interfaces, but only indirectly (through
+	// pointers).
+	var v *error
+	return v
+}
+
+func anonymousInterfaceType() interface{} {
+	var v *interface {
+		String() string
+	}
+	return v
+}
+
+// Test interface builtins.
+
+func isInt(itf interface{}) bool {
+	_, ok := itf.(int)
+	return ok
+}
+
+func isError(itf interface{}) bool {
+	// Interface assert on (builtin) named interface type.
+	_, ok := itf.(error)
+	return ok
+}
+
+func isStringer(itf interface{}) bool {
+	// Interface assert on anonymous interface type.
+	_, ok := itf.(interface {
+		String() string
+	})
+	return ok
+}
+
+func callErrorMethod(itf error) string {
+	return itf.Error()
+}

compiler/testdata/interface.ll

@@ -0,0 +1,100 @@
+; ModuleID = 'interface.go'
+source_filename = "interface.go"
+target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-f64:32:64-f80:32-n8:16:32-S128"
+target triple = "i686--linux"
+
+%runtime.typecodeID = type { %runtime.typecodeID*, i32, %runtime.interfaceMethodInfo* }
+%runtime.interfaceMethodInfo = type { i8*, i32 }
+%runtime._interface = type { i32, i8* }
+%runtime._string = type { i8*, i32 }
+
+@"reflect/types.type:basic:int" = linkonce_odr constant %runtime.typecodeID zeroinitializer
+@"reflect/types.type:pointer:basic:int" = linkonce_odr constant %runtime.typecodeID { %runtime.typecodeID* @"reflect/types.type:basic:int", i32 0, %runtime.interfaceMethodInfo* null }
+@"reflect/types.type:pointer:named:error" = linkonce_odr constant %runtime.typecodeID { %runtime.typecodeID* @"reflect/types.type:named:error", i32 0, %runtime.interfaceMethodInfo* null }
+@"reflect/types.type:named:error" = linkonce_odr constant %runtime.typecodeID { %runtime.typecodeID* @"reflect/types.type:interface:{Error:func:{}{basic:string}}", i32 0, %runtime.interfaceMethodInfo* null }
+@"reflect/types.type:interface:{Error:func:{}{basic:string}}" = linkonce_odr constant %runtime.typecodeID { %runtime.typecodeID* bitcast ([1 x i8*]* @"reflect/types.interface:interface{Error() string}$interface" to %runtime.typecodeID*), i32 0, %runtime.interfaceMethodInfo* null }
+@"func Error() string" = external constant i8
+@"reflect/types.interface:interface{Error() string}$interface" = linkonce_odr constant [1 x i8*] [i8* @"func Error() string"]
+@"reflect/types.type:pointer:interface:{String:func:{}{basic:string}}" = linkonce_odr constant %runtime.typecodeID { %runtime.typecodeID* @"reflect/types.type:interface:{String:func:{}{basic:string}}", i32 0, %runtime.interfaceMethodInfo* null }
+@"reflect/types.type:interface:{String:func:{}{basic:string}}" = linkonce_odr constant %runtime.typecodeID { %runtime.typecodeID* bitcast ([1 x i8*]* @"reflect/types.interface:interface{String() string}$interface" to %runtime.typecodeID*), i32 0, %runtime.interfaceMethodInfo* null }
+@"func String() string" = external constant i8
+@"reflect/types.interface:interface{String() string}$interface" = linkonce_odr constant [1 x i8*] [i8* @"func String() string"]
+@"reflect/types.type:basic:int$id" = external constant i8
+@"error$interface" = linkonce_odr constant [1 x i8*] [i8* @"func Error() string"]
+
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret void
+}
+
+define hidden %runtime._interface @main.simpleType(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret %runtime._interface { i32 ptrtoint (%runtime.typecodeID* @"reflect/types.type:basic:int" to i32), i8* null }
+}
+
+define hidden %runtime._interface @main.pointerType(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret %runtime._interface { i32 ptrtoint (%runtime.typecodeID* @"reflect/types.type:pointer:basic:int" to i32), i8* null }
+}
+
+define hidden %runtime._interface @main.interfaceType(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret %runtime._interface { i32 ptrtoint (%runtime.typecodeID* @"reflect/types.type:pointer:named:error" to i32), i8* null }
+}
+
+define hidden %runtime._interface @main.anonymousInterfaceType(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret %runtime._interface { i32 ptrtoint (%runtime.typecodeID* @"reflect/types.type:pointer:interface:{String:func:{}{basic:string}}" to i32), i8* null }
+}
+
+define hidden i1 @main.isInt(i32 %itf.typecode, i8* %itf.value, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %typecode = call i1 @runtime.typeAssert(i32 %itf.typecode, i8* nonnull @"reflect/types.type:basic:int$id", i8* undef, i8* null)
+  br i1 %typecode, label %typeassert.ok, label %typeassert.next
+
+typeassert.ok:                                    ; preds = %entry
+  br label %typeassert.next
+
+typeassert.next:                                  ; preds = %typeassert.ok, %entry
+  ret i1 %typecode
+}
+
+declare i1 @runtime.typeAssert(i32, i8* dereferenceable_or_null(1), i8*, i8*)
+
+define hidden i1 @main.isError(i32 %itf.typecode, i8* %itf.value, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = call i1 @runtime.interfaceImplements(i32 %itf.typecode, i8** getelementptr inbounds ([1 x i8*], [1 x i8*]* @"error$interface", i32 0, i32 0), i8* undef, i8* null)
+  br i1 %0, label %typeassert.ok, label %typeassert.next
+
+typeassert.ok:                                    ; preds = %entry
+  br label %typeassert.next
+
+typeassert.next:                                  ; preds = %typeassert.ok, %entry
+  ret i1 %0
+}
+
+declare i1 @runtime.interfaceImplements(i32, i8** dereferenceable_or_null(4), i8*, i8*)
+
+define hidden i1 @main.isStringer(i32 %itf.typecode, i8* %itf.value, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = call i1 @runtime.interfaceImplements(i32 %itf.typecode, i8** getelementptr inbounds ([1 x i8*], [1 x i8*]* @"reflect/types.interface:interface{String() string}$interface", i32 0, i32 0), i8* undef, i8* null)
+  br i1 %0, label %typeassert.ok, label %typeassert.next
+
+typeassert.ok:                                    ; preds = %entry
+  br label %typeassert.next
+
+typeassert.next:                                  ; preds = %typeassert.ok, %entry
+  ret i1 %0
+}
+
+define hidden %runtime._string @main.callErrorMethod(i32 %itf.typecode, i8* %itf.value, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %invoke.func = call i32 @runtime.interfaceMethod(i32 %itf.typecode, i8** getelementptr inbounds ([1 x i8*], [1 x i8*]* @"error$interface", i32 0, i32 0), i8* nonnull @"func Error() string", i8* undef, i8* null)
+  %invoke.func.cast = inttoptr i32 %invoke.func to %runtime._string (i8*, i8*, i8*)*
+  %0 = call %runtime._string %invoke.func.cast(i8* %itf.value, i8* undef, i8* undef)
+  ret %runtime._string %0
+}
+
+declare i32 @runtime.interfaceMethod(i32, i8** dereferenceable_or_null(4), i8* dereferenceable_or_null(1), i8*, i8*)

compiler/testdata/pointer.ll

@@ -3,46 +3,48 @@ source_filename = "pointer.go"
 target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-f64:32:64-f80:32-n8:16:32-S128"
 target triple = "i686--linux"
 
-define internal void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   ret void
 }
 
-define internal [0 x i32] @main.pointerDerefZero([0 x i32]* %x, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden [0 x i32] @main.pointerDerefZero([0 x i32]* %x, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   ret [0 x i32] zeroinitializer
 }
 
-define internal i32* @main.pointerCastFromUnsafe(i8* %x, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i32* @main.pointerCastFromUnsafe(i8* %x, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = bitcast i8* %x to i32*
   ret i32* %0
 }
 
-define internal i8* @main.pointerCastToUnsafe(i32* dereferenceable_or_null(4) %x, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i8* @main.pointerCastToUnsafe(i32* dereferenceable_or_null(4) %x, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = bitcast i32* %x to i8*
   ret i8* %0
 }
 
-define internal i8* @main.pointerCastToUnsafeNoop(i8* dereferenceable_or_null(1) %x, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i8* @main.pointerCastToUnsafeNoop(i8* dereferenceable_or_null(1) %x, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   ret i8* %x
 }
 
-define internal i8* @main.pointerUnsafeGEPFixedOffset(i8* dereferenceable_or_null(1) %ptr, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i8* @main.pointerUnsafeGEPFixedOffset(i8* dereferenceable_or_null(1) %ptr, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = getelementptr inbounds i8, i8* %ptr, i32 10
   ret i8* %0
 }
 
-define internal i8* @main.pointerUnsafeGEPByteOffset(i8* dereferenceable_or_null(1) %ptr, i32 %offset, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i8* @main.pointerUnsafeGEPByteOffset(i8* dereferenceable_or_null(1) %ptr, i32 %offset, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = getelementptr inbounds i8, i8* %ptr, i32 %offset
   ret i8* %0
 }
 
-define internal i32* @main.pointerUnsafeGEPIntOffset(i32* dereferenceable_or_null(4) %ptr, i32 %offset, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i32* @main.pointerUnsafeGEPIntOffset(i32* dereferenceable_or_null(4) %ptr, i32 %offset, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   %0 = getelementptr i32, i32* %ptr, i32 %offset
   ret i32* %0

compiler/testdata/slice.go

@@ -7,3 +7,19 @@ func sliceLen(ints []int) int {
 func sliceCap(ints []int) int {
 	return cap(ints)
 }
+
+func sliceElement(ints []int, index int) int {
+	return ints[index]
+}
+
+func sliceAppendValues(ints []int) []int {
+	return append(ints, 1, 2, 3)
+}
+
+func sliceAppendSlice(ints, added []int) []int {
+	return append(ints, added...)
+}
+
+func sliceCopy(dst, src []int) int {
+	return copy(dst, src)
+}

compiler/testdata/slice.ll

@@ -3,17 +3,86 @@ source_filename = "slice.go"
 target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-f64:32:64-f80:32-n8:16:32-S128"
 target triple = "i686--linux"
 
-define internal void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   ret void
 }
 
-define internal i32 @main.sliceLen(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i32 @main.sliceLen(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   ret i32 %ints.len
 }
 
-define internal i32 @main.sliceCap(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i32 @main.sliceCap(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
   ret i32 %ints.cap
 }
+
+define hidden i32 @main.sliceElement(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i32 %index, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %.not = icmp ult i32 %index, %ints.len
+  br i1 %.not, label %lookup.next, label %lookup.throw
+
+lookup.throw:                                     ; preds = %entry
+  call void @runtime.lookupPanic(i8* undef, i8* null)
+  unreachable
+
+lookup.next:                                      ; preds = %entry
+  %0 = getelementptr inbounds i32, i32* %ints.data, i32 %index
+  %1 = load i32, i32* %0, align 4
+  ret i32 %1
+}
+
+declare void @runtime.lookupPanic(i8*, i8*)
+
+define hidden { i32*, i32, i32 } @main.sliceAppendValues(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %varargs = call i8* @runtime.alloc(i32 12, i8* undef, i8* null)
+  %0 = bitcast i8* %varargs to i32*
+  store i32 1, i32* %0, align 4
+  %1 = getelementptr inbounds i8, i8* %varargs, i32 4
+  %2 = bitcast i8* %1 to i32*
+  store i32 2, i32* %2, align 4
+  %3 = getelementptr inbounds i8, i8* %varargs, i32 8
+  %4 = bitcast i8* %3 to i32*
+  store i32 3, i32* %4, align 4
+  %append.srcPtr = bitcast i32* %ints.data to i8*
+  %append.new = call { i8*, i32, i32 } @runtime.sliceAppend(i8* %append.srcPtr, i8* nonnull %varargs, i32 %ints.len, i32 %ints.cap, i32 3, i32 4, i8* undef, i8* null)
+  %append.newPtr = extractvalue { i8*, i32, i32 } %append.new, 0
+  %append.newBuf = bitcast i8* %append.newPtr to i32*
+  %append.newLen = extractvalue { i8*, i32, i32 } %append.new, 1
+  %append.newCap = extractvalue { i8*, i32, i32 } %append.new, 2
+  %5 = insertvalue { i32*, i32, i32 } undef, i32* %append.newBuf, 0
+  %6 = insertvalue { i32*, i32, i32 } %5, i32 %append.newLen, 1
+  %7 = insertvalue { i32*, i32, i32 } %6, i32 %append.newCap, 2
+  ret { i32*, i32, i32 } %7
+}
+
+declare { i8*, i32, i32 } @runtime.sliceAppend(i8*, i8*, i32, i32, i32, i32, i8*, i8*)
+
+define hidden { i32*, i32, i32 } @main.sliceAppendSlice(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i32* %added.data, i32 %added.len, i32 %added.cap, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %append.srcPtr = bitcast i32* %ints.data to i8*
+  %append.srcPtr1 = bitcast i32* %added.data to i8*
+  %append.new = call { i8*, i32, i32 } @runtime.sliceAppend(i8* %append.srcPtr, i8* %append.srcPtr1, i32 %ints.len, i32 %ints.cap, i32 %added.len, i32 4, i8* undef, i8* null)
+  %append.newPtr = extractvalue { i8*, i32, i32 } %append.new, 0
+  %append.newBuf = bitcast i8* %append.newPtr to i32*
+  %append.newLen = extractvalue { i8*, i32, i32 } %append.new, 1
+  %append.newCap = extractvalue { i8*, i32, i32 } %append.new, 2
+  %0 = insertvalue { i32*, i32, i32 } undef, i32* %append.newBuf, 0
+  %1 = insertvalue { i32*, i32, i32 } %0, i32 %append.newLen, 1
+  %2 = insertvalue { i32*, i32, i32 } %1, i32 %append.newCap, 2
+  ret { i32*, i32, i32 } %2
+}
+
+define hidden i32 @main.sliceCopy(i32* %dst.data, i32 %dst.len, i32 %dst.cap, i32* %src.data, i32 %src.len, i32 %src.cap, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %copy.dstPtr = bitcast i32* %dst.data to i8*
+  %copy.srcPtr = bitcast i32* %src.data to i8*
+  %copy.n = call i32 @runtime.sliceCopy(i8* %copy.dstPtr, i8* %copy.srcPtr, i32 %dst.len, i32 %src.len, i32 4, i8* undef, i8* null)
+  ret i32 %copy.n
+}
+
+declare i32 @runtime.sliceCopy(i8*, i8*, i32, i32, i32, i8*, i8*)

compiler/testdata/string.go

@@ -0,0 +1,21 @@
+package main
+
+func stringLen(s string) int {
+	return len(s)
+}
+
+func stringIndex(s string, index int) byte {
+	return s[index]
+}
+
+func stringCompareEqual(s1, s2 string) bool {
+	return s1 == s2
+}
+
+func stringCompareUnequal(s1, s2 string) bool {
+	return s1 != s2
+}
+
+func stringCompareLarger(s1, s2 string) bool {
+	return s1 > s2
+}

compiler/testdata/string.ll

@@ -0,0 +1,57 @@
+; ModuleID = 'string.go'
+source_filename = "string.go"
+target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-f64:32:64-f80:32-n8:16:32-S128"
+target triple = "i686--linux"
+
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret void
+}
+
+define hidden i32 @main.stringLen(i8* %s.data, i32 %s.len, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret i32 %s.len
+}
+
+define hidden i8 @main.stringIndex(i8* %s.data, i32 %s.len, i32 %index, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %.not = icmp ult i32 %index, %s.len
+  br i1 %.not, label %lookup.next, label %lookup.throw
+
+lookup.throw:                                     ; preds = %entry
+  call void @runtime.lookupPanic(i8* undef, i8* null)
+  unreachable
+
+lookup.next:                                      ; preds = %entry
+  %0 = getelementptr inbounds i8, i8* %s.data, i32 %index
+  %1 = load i8, i8* %0, align 1
+  ret i8 %1
+}
+
+declare void @runtime.lookupPanic(i8*, i8*)
+
+define hidden i1 @main.stringCompareEqual(i8* %s1.data, i32 %s1.len, i8* %s2.data, i32 %s2.len, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = call i1 @runtime.stringEqual(i8* %s1.data, i32 %s1.len, i8* %s2.data, i32 %s2.len, i8* undef, i8* null)
+  ret i1 %0
+}
+
+declare i1 @runtime.stringEqual(i8*, i32, i8*, i32, i8*, i8*)
+
+define hidden i1 @main.stringCompareUnequal(i8* %s1.data, i32 %s1.len, i8* %s2.data, i32 %s2.len, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = call i1 @runtime.stringEqual(i8* %s1.data, i32 %s1.len, i8* %s2.data, i32 %s2.len, i8* undef, i8* null)
+  %1 = xor i1 %0, true
+  ret i1 %1
+}
+
+define hidden i1 @main.stringCompareLarger(i8* %s1.data, i32 %s1.len, i8* %s2.data, i32 %s2.len, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = call i1 @runtime.stringLess(i8* %s1.data, i32 %s1.len, i8* %s2.data, i32 %s2.len, i8* undef, i8* null)
+  %1 = xor i1 %0, true
+  ret i1 %1
+}
+
+declare i1 @runtime.stringLess(i8*, i32, i8*, i32, i8*, i8*)

go.mod

@@ -12,5 +12,5 @@ require (
 	go.bug.st/serial v1.1.2
 	golang.org/x/sys v0.0.0-20210113181707-4bcb84eeeb78
 	golang.org/x/tools v0.0.0-20200216192241-b320d3a0f5a2
-	tinygo.org/x/go-llvm v0.0.0-20210206225315-7fe719483a0f
+	tinygo.org/x/go-llvm v0.0.0-20210325115028-e7b85195e81c
 )

go.sum

@@ -57,5 +57,7 @@ golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898 h1:/atklqdjdhuosWIl6AIbO
 golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
 gopkg.in/yaml.v2 v2.2.2/go.mod h1:hI93XBmqTisBFMUTm0b8Fm+jr3Dg1NNxqwp+5A1VGuI=
-tinygo.org/x/go-llvm v0.0.0-20210206225315-7fe719483a0f h1:FP5Do5omlQ/dLQ3Hfy7oyJo69VS5Hn46rZw004r0lGU=
+tinygo.org/x/go-llvm v0.0.0-20210308112806-9ef958b6bed4 h1:CMUHxVTb+UuUePuMf8vkWjZ3gTp9BBK91KrgOCwoNHs=
-tinygo.org/x/go-llvm v0.0.0-20210206225315-7fe719483a0f/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
+tinygo.org/x/go-llvm v0.0.0-20210308112806-9ef958b6bed4/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
+tinygo.org/x/go-llvm v0.0.0-20210325115028-e7b85195e81c h1:vn9IPshzYmzZis10UEVrsIBRv9FpykADw6M3/tHHROg=
+tinygo.org/x/go-llvm v0.0.0-20210325115028-e7b85195e81c/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=

interp/errors.go

@@ -20,6 +20,11 @@ var (
 	errMapAlreadyCreated      = errors.New("interp: map already created")
 )
 
+// This is one of the errors that can be returned from toLLVMValue when the
+// passed type does not fit the data to serialize. It is recoverable by
+// serializing without a type (using rawValue.rawLLVMValue).
+var errInvalidPtrToIntSize = errors.New("interp: ptrtoint integer size does not equal pointer size")
+
 func isRecoverableError(err error) bool {
 	return err == errIntegerAsPointer || err == errUnsupportedInst || err == errUnsupportedRuntimeInst || err == errMapAlreadyCreated
 }

interp/interp.go

@@ -11,6 +11,12 @@ import (
 	"tinygo.org/x/go-llvm"
 )
 
+// Version of the interp package. It must be incremented whenever the interp
+// package is changed in a way that affects the output so that cached package
+// builds will be invalidated.
+// This version is independent of the TinyGo version number.
+const Version = 1
+
 // Enable extra checks, which should be disabled by default.
 // This may help track down bugs by adding a few more sanity checks.
 const checks = true
@@ -32,9 +38,7 @@ type runner struct {
 	callsExecuted uint64
 }
 
-// Run evaluates runtime.initAll function as much as possible at compile time.
+func newRunner(mod llvm.Module, debug bool) *runner {
-// Set debug to true if it should print output while running.
-func Run(mod llvm.Module, debug bool) error {
 	r := runner{
 		mod:           mod,
 		targetData:    llvm.NewTargetData(mod.DataLayout()),
@@ -47,6 +51,13 @@ func Run(mod llvm.Module, debug bool) error {
 	r.pointerSize = uint32(r.targetData.PointerSize())
 	r.i8ptrType = llvm.PointerType(mod.Context().Int8Type(), 0)
 	r.maxAlign = r.targetData.PrefTypeAlignment(r.i8ptrType) // assume pointers are maximally aligned (this is not always the case)
+	return &r
+}
+
+// Run evaluates runtime.initAll function as much as possible at compile time.
+// Set debug to true if it should print output while running.
+func Run(mod llvm.Module, debug bool) error {
+	r := newRunner(mod, debug)
 
 	initAll := mod.NamedFunction("runtime.initAll")
 	bb := initAll.EntryBasicBlock()
@@ -117,7 +128,7 @@ func Run(mod llvm.Module, debug bool) error {
 	r.pkgName = ""
 
 	// Update all global variables in the LLVM module.
-	mem := memoryView{r: &r}
+	mem := memoryView{r: r}
 	for _, obj := range r.objects {
 		if obj.llvmGlobal.IsNil() {
 			continue
@@ -125,7 +136,34 @@ func Run(mod llvm.Module, debug bool) error {
 		if obj.buffer == nil {
 			continue
 		}
-		initializer := obj.buffer.toLLVMValue(obj.llvmGlobal.Type().ElementType(), &mem)
+		initializer, err := obj.buffer.toLLVMValue(obj.llvmGlobal.Type().ElementType(), &mem)
+		if err == errInvalidPtrToIntSize {
+			// This can happen when a previous interp run did not have the
+			// correct LLVM type for a global and made something up. In that
+			// case, some fields could be written out as a series of (null)
+			// bytes even though they actually contain a pointer value.
+			// As a fallback, use asRawValue to get something of the correct
+			// memory layout.
+			initializer, err := obj.buffer.asRawValue(r).rawLLVMValue(&mem)
+			if err != nil {
+				return err
+			}
+			initializerType := initializer.Type()
+			newGlobal := llvm.AddGlobal(mod, initializerType, obj.llvmGlobal.Name()+".tmp")
+			newGlobal.SetInitializer(initializer)
+			newGlobal.SetLinkage(obj.llvmGlobal.Linkage())
+			newGlobal.SetAlignment(obj.llvmGlobal.Alignment())
+			// TODO: copy debug info, unnamed_addr, ...
+			bitcast := llvm.ConstBitCast(newGlobal, obj.llvmGlobal.Type())
+			obj.llvmGlobal.ReplaceAllUsesWith(bitcast)
+			name := obj.llvmGlobal.Name()
+			obj.llvmGlobal.EraseFromParentAsGlobal()
+			newGlobal.SetName(name)
+			continue
+		}
+		if err != nil {
+			return err
+		}
 		if checks && initializer.Type() != obj.llvmGlobal.Type().ElementType() {
 			panic("initializer type mismatch")
 		}
@@ -135,6 +173,87 @@ func Run(mod llvm.Module, debug bool) error {
 	return nil
 }
 
+// RunFunc evaluates a single package initializer at compile time.
+// Set debug to true if it should print output while running.
+func RunFunc(fn llvm.Value, debug bool) error {
+	// Create and initialize *runner object.
+	mod := fn.GlobalParent()
+	r := newRunner(mod, debug)
+	initName := fn.Name()
+	if !strings.HasSuffix(initName, ".init") {
+		return errorAt(fn, "interp: unexpected function name (expected *.init)")
+	}
+	r.pkgName = initName[:len(initName)-len(".init")]
+
+	// Create new function with the interp result.
+	newFn := llvm.AddFunction(mod, fn.Name()+".tmp", fn.Type().ElementType())
+	newFn.SetLinkage(fn.Linkage())
+	newFn.SetVisibility(fn.Visibility())
+	entry := mod.Context().AddBasicBlock(newFn, "entry")
+
+	// Create a builder, to insert instructions that could not be evaluated at
+	// compile time.
+	r.builder = mod.Context().NewBuilder()
+	defer r.builder.Dispose()
+	r.builder.SetInsertPointAtEnd(entry)
+
+	// Copy debug information.
+	subprogram := fn.Subprogram()
+	if !subprogram.IsNil() {
+		newFn.SetSubprogram(subprogram)
+		r.builder.SetCurrentDebugLocation(subprogram.SubprogramLine(), 0, subprogram, llvm.Metadata{})
+	}
+
+	// Run the initializer, filling the .init.tmp function.
+	if r.debug {
+		fmt.Fprintln(os.Stderr, "interp:", fn.Name())
+	}
+	_, pkgMem, callErr := r.run(r.getFunction(fn), nil, nil, "    ")
+	if callErr != nil {
+		if isRecoverableError(callErr.Err) {
+			// Could not finish, but could recover from it.
+			if r.debug {
+				fmt.Fprintln(os.Stderr, "not interpreting", r.pkgName, "because of error:", callErr.Error())
+			}
+			newFn.EraseFromParentAsFunction()
+			return nil
+		}
+		return callErr
+	}
+	for index, obj := range pkgMem.objects {
+		r.objects[index] = obj
+	}
+
+	// Update globals with values determined while running the initializer above.
+	mem := memoryView{r: r}
+	for _, obj := range r.objects {
+		if obj.llvmGlobal.IsNil() {
+			continue
+		}
+		if obj.buffer == nil {
+			continue
+		}
+		initializer, err := obj.buffer.toLLVMValue(obj.llvmGlobal.Type().ElementType(), &mem)
+		if err != nil {
+			return err
+		}
+		if checks && initializer.Type() != obj.llvmGlobal.Type().ElementType() {
+			panic("initializer type mismatch")
+		}
+		obj.llvmGlobal.SetInitializer(initializer)
+	}
+
+	// Finalize: remove the old init function and replace it with the new
+	// (.init.tmp) function.
+	r.builder.CreateRetVoid()
+	fnName := fn.Name()
+	fn.ReplaceAllUsesWith(newFn)
+	fn.EraseFromParentAsFunction()
+	newFn.SetName(fnName)
+
+	return nil
+}
+
 // getFunction returns the compiled version of the given LLVM function. It
 // compiles the function if necessary and caches the result.
 func (r *runner) getFunction(llvmFn llvm.Value) *function {

interp/interpreter.go

@@ -155,11 +155,8 @@ func (r *runner) run(fn *function, params []value, parentMem *memoryView, indent
 				// which case this call won't even get to this point but will
 				// already be emitted in initAll.
 				continue
-			case callFn.name == "(reflect.Type).Elem" || strings.HasPrefix(callFn.name, "runtime.print") || callFn.name == "runtime._panic" || callFn.name == "runtime.hashmapGet":
+			case strings.HasPrefix(callFn.name, "runtime.print") || callFn.name == "runtime._panic" || callFn.name == "runtime.hashmapGet":
 				// These functions should be run at runtime. Specifically:
-				//   * (reflect.Type).Elem is a special function. It should
-				//     eventually be interpreted, but fall back to a runtime call
-				//     for now.
 				//   * Print and panic functions are best emitted directly without
 				//     interpreting them, otherwise we get a ton of putchar (etc.)
 				//     calls.
@@ -280,26 +277,58 @@ func (r *runner) run(fn *function, params []value, parentMem *memoryView, indent
 				copy(dstBuf.buf[dst.offset():dst.offset()+nBytes], srcBuf.buf[src.offset():])
 				dstObj.buffer = dstBuf
 				mem.put(dst.index(), dstObj)
+			case callFn.name == "(reflect.rawType).elem":
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"call (reflect.rawType).elem:", operands[1:])
+				}
+				// Extract the type code global from the first parameter.
+				typecodeIDPtrToInt, err := operands[1].toLLVMValue(inst.llvmInst.Operand(0).Type(), &mem)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				typecodeID := typecodeIDPtrToInt.Operand(0)
+
+				// Get the type class.
+				// See also: getClassAndValueFromTypeCode in transform/reflect.go.
+				typecodeName := typecodeID.Name()
+				const prefix = "reflect/types.type:"
+				if !strings.HasPrefix(typecodeName, prefix) {
+					panic("unexpected typecode name: " + typecodeName)
+				}
+				id := typecodeName[len(prefix):]
+				class := id[:strings.IndexByte(id, ':')]
+				value := id[len(class)+1:]
+				if class == "named" {
+					// Get the underlying type.
+					class = value[:strings.IndexByte(value, ':')]
+					value = value[len(class)+1:]
+				}
+
+				// Elem() is only valid for certain type classes.
+				switch class {
+				case "chan", "pointer", "slice", "array":
+					elementType := llvm.ConstExtractValue(typecodeID.Initializer(), []uint32{0})
+					uintptrType := r.mod.Context().IntType(int(mem.r.pointerSize) * 8)
+					locals[inst.localIndex] = r.getValue(llvm.ConstPtrToInt(elementType, uintptrType))
+				default:
+					return nil, mem, r.errorAt(inst, fmt.Errorf("(reflect.Type).Elem() called on %s type", class))
+				}
 			case callFn.name == "runtime.typeAssert":
 				// This function must be implemented manually as it is normally
 				// implemented by the interface lowering pass.
 				if r.debug {
 					fmt.Fprintln(os.Stderr, indent+"typeassert:", operands[1:])
 				}
-				typeInInterfacePtr, err := operands[1].asPointer(r)
+				assertedType, err := operands[2].toLLVMValue(inst.llvmInst.Operand(1).Type(), &mem)
 				if err != nil {
 					return nil, mem, r.errorAt(inst, err)
 				}
-				actualType, err := mem.load(typeInInterfacePtr, r.pointerSize).asPointer(r)
+				actualTypePtrToInt, err := operands[1].toLLVMValue(inst.llvmInst.Operand(0).Type(), &mem)
 				if err != nil {
 					return nil, mem, r.errorAt(inst, err)
 				}
-				assertedType, err := operands[2].asPointer(r)
+				actualType := actualTypePtrToInt.Operand(0)
-				if err != nil {
+				if actualType.Name()+"$id" == assertedType.Name() {
-					return nil, mem, r.errorAt(inst, err)
-				}
-				result := assertedType.asRawValue(r).equal(actualType.asRawValue(r))
-				if result {
 					locals[inst.localIndex] = literalValue{uint8(1)}
 				} else {
 					locals[inst.localIndex] = literalValue{uint8(0)}
@@ -310,11 +339,11 @@ func (r *runner) run(fn *function, params []value, parentMem *memoryView, indent
 				}
 
 				// Load various values for the interface implements check below.
-				typeInInterfacePtr, err := operands[1].asPointer(r)
+				typecodePtr, err := operands[1].asPointer(r)
 				if err != nil {
 					return nil, mem, r.errorAt(inst, err)
 				}
-				methodSetPtr, err := mem.load(typeInInterfacePtr.addOffset(r.pointerSize), r.pointerSize).asPointer(r)
+				methodSetPtr, err := mem.load(typecodePtr.addOffset(r.pointerSize*2), r.pointerSize).asPointer(r)
 				if err != nil {
 					return nil, mem, r.errorAt(inst, err)
 				}
@@ -349,6 +378,43 @@ func (r *runner) run(fn *function, params []value, parentMem *memoryView, indent
 				}
 				// If assertOk is still 1, the assertion succeeded.
 				locals[inst.localIndex] = literalValue{assertOk}
+			case callFn.name == "runtime.interfaceMethod":
+				// This builtin returns the function (which may be a thunk) to
+				// invoke a method on an interface. It does not call the method.
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"interface method:", operands[1:])
+				}
+
+				// Load the first param, which is the type code (ptrtoint of the
+				// type code global).
+				typecodeIDPtrToInt, err := operands[1].toLLVMValue(inst.llvmInst.Operand(0).Type(), &mem)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				typecodeID := typecodeIDPtrToInt.Operand(0).Initializer()
+
+				// Load the method set, which is part of the typecodeID object.
+				methodSet := llvm.ConstExtractValue(typecodeID, []uint32{2}).Operand(0).Initializer()
+
+				// We don't need to load the interface method set.
+
+				// Load the signature of the to-be-called function.
+				signature := inst.llvmInst.Operand(2)
+
+				// Iterate through all methods, looking for the one method that
+				// should be returned.
+				numMethods := methodSet.Type().ArrayLength()
+				var method llvm.Value
+				for i := 0; i < numMethods; i++ {
+					methodSignature := llvm.ConstExtractValue(methodSet, []uint32{uint32(i), 0})
+					if methodSignature == signature {
+						method = llvm.ConstExtractValue(methodSet, []uint32{uint32(i), 1}).Operand(0)
+					}
+				}
+				if method.IsNil() {
+					return nil, mem, r.errorAt(inst, errors.New("could not find method: "+signature.Name()))
+				}
+				locals[inst.localIndex] = r.getValue(method)
 			case callFn.name == "runtime.hashmapMake":
 				// Create a new map.
 				hashmapPointerType := inst.llvmInst.Type()
@@ -482,6 +548,13 @@ func (r *runner) run(fn *function, params []value, parentMem *memoryView, indent
 				continue
 			}
 			result := mem.load(ptr, uint32(size))
+			if result == nil {
+				err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+				if err != nil {
+					return nil, mem, err
+				}
+				continue
+			}
 			if r.debug {
 				fmt.Fprintln(os.Stderr, indent+"load:", ptr, "->", result)
 			}
@@ -504,7 +577,14 @@ func (r *runner) run(fn *function, params []value, parentMem *memoryView, indent
 			if r.debug {
 				fmt.Fprintln(os.Stderr, indent+"store:", val, ptr)
 			}
-			mem.store(val, ptr)
+			ok := mem.store(val, ptr)
+			if !ok {
+				// Could not store the value, do it at runtime.
+				err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+				if err != nil {
+					return nil, mem, err
+				}
+			}
 		case llvm.Alloca:
 			// Alloca normally allocates some stack memory. In the interpreter,
 			// it allocates a global instead.
@@ -836,7 +916,11 @@ func (r *runner) runAtRuntime(fn *function, inst instruction, locals []value, me
 	for i := 0; i < numOperands; i++ {
 		operand := inst.llvmInst.Operand(i)
 		if !operand.IsAInstruction().IsNil() || !operand.IsAArgument().IsNil() {
-			operand = locals[fn.locals[operand]].toLLVMValue(operand.Type(), mem)
+			var err error
+			operand, err = locals[fn.locals[operand]].toLLVMValue(operand.Type(), mem)
+			if err != nil {
+				return r.errorAt(inst, err)
+			}
 		}
 		operands[i] = operand
 	}

interp/memory.go

@@ -259,12 +259,17 @@ func (mv *memoryView) put(index uint32, obj object) {
 	mv.objects[index] = obj
 }
 
-// Load the value behind the given pointer.
+// Load the value behind the given pointer. Returns nil if the pointer points to
+// an external global.
 func (mv *memoryView) load(p pointerValue, size uint32) value {
 	if checks && mv.hasExternalStore(p) {
 		panic("interp: load from object with external store")
 	}
 	obj := mv.get(p.index())
+	if obj.buffer == nil {
+		// External global, return nil.
+		return nil
+	}
 	if p.offset() == 0 && size == obj.size {
 		return obj.buffer.clone()
 	}
@@ -280,12 +285,17 @@ func (mv *memoryView) load(p pointerValue, size uint32) value {
 
 // Store to the value behind the given pointer. This overwrites the value in the
 // memory view, so that the changed value is discarded when the memory view is
-// reverted.
+// reverted. Returns true on success, false if the object to store to is
-func (mv *memoryView) store(v value, p pointerValue) {
+// external.
+func (mv *memoryView) store(v value, p pointerValue) bool {
 	if checks && mv.hasExternalLoadOrStore(p) {
 		panic("interp: store to object with external load/store")
 	}
 	obj := mv.get(p.index())
+	if obj.buffer == nil {
+		// External global, return false (for a failure).
+		return false
+	}
 	if checks && p.offset()+v.len(mv.r) > obj.size {
 		panic("interp: store out of bounds")
 	}
@@ -301,6 +311,7 @@ func (mv *memoryView) store(v value, p pointerValue) {
 		}
 	}
 	mv.put(p.index(), obj)
+	return true // success
 }
 
 // value is some sort of value, comparable to a LLVM constant. It can be
@@ -314,7 +325,7 @@ type value interface {
 	asRawValue(*runner) rawValue
 	Uint() uint64
 	Int() int64
-	toLLVMValue(llvm.Type, *memoryView) llvm.Value
+	toLLVMValue(llvm.Type, *memoryView) (llvm.Value, error)
 	String() string
 }
 
@@ -405,25 +416,25 @@ func (v literalValue) Int() int64 {
 	}
 }
 
-func (v literalValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
+func (v literalValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) (llvm.Value, error) {
 	switch llvmType.TypeKind() {
 	case llvm.IntegerTypeKind:
 		switch value := v.value.(type) {
 		case uint64:
-			return llvm.ConstInt(llvmType, value, false)
+			return llvm.ConstInt(llvmType, value, false), nil
 		case uint32:
-			return llvm.ConstInt(llvmType, uint64(value), false)
+			return llvm.ConstInt(llvmType, uint64(value), false), nil
 		case uint16:
-			return llvm.ConstInt(llvmType, uint64(value), false)
+			return llvm.ConstInt(llvmType, uint64(value), false), nil
 		case uint8:
-			return llvm.ConstInt(llvmType, uint64(value), false)
+			return llvm.ConstInt(llvmType, uint64(value), false), nil
 		default:
-			panic("inpterp: unknown literal type")
+			return llvm.Value{}, errors.New("interp: unknown literal type")
 		}
 	case llvm.DoubleTypeKind:
-		return llvm.ConstFloat(llvmType, math.Float64frombits(v.value.(uint64)))
+		return llvm.ConstFloat(llvmType, math.Float64frombits(v.value.(uint64))), nil
 	case llvm.FloatTypeKind:
-		return llvm.ConstFloat(llvmType, float64(math.Float32frombits(v.value.(uint32))))
+		return llvm.ConstFloat(llvmType, float64(math.Float32frombits(v.value.(uint32)))), nil
 	default:
 		return v.asRawValue(mem.r).toLLVMValue(llvmType, mem)
 	}
@@ -507,7 +518,7 @@ func (v pointerValue) llvmValue(mem *memoryView) llvm.Value {
 // toLLVMValue returns the LLVM value for this pointer, which may be a GEP or
 // bitcast. The llvm.Type parameter is optional, if omitted the pointer type may
 // be different than expected.
-func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
+func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) (llvm.Value, error) {
 	// Obtain the llvmValue, creating it if it doesn't exist yet.
 	llvmValue := v.llvmValue(mem)
 	if llvmValue.IsNil() {
@@ -518,7 +529,10 @@ func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Valu
 		if obj.llvmType.IsNil() {
 			// Create an initializer without knowing the global type.
 			// This is probably the result of a runtime.alloc call.
-			initializer := obj.buffer.asRawValue(mem.r).rawLLVMValue(mem)
+			initializer, err := obj.buffer.asRawValue(mem.r).rawLLVMValue(mem)
+			if err != nil {
+				return llvm.Value{}, err
+			}
 			globalType := initializer.Type()
 			llvmValue = llvm.AddGlobal(mem.r.mod, globalType, obj.globalName)
 			llvmValue.SetInitializer(initializer)
@@ -537,9 +551,12 @@ func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Valu
 			// Set the initializer for the global. Do this after creation to avoid
 			// infinite recursion between creating the global and creating the
 			// contents of the global (if the global contains itself).
-			initializer := obj.buffer.toLLVMValue(globalType, mem)
+			initializer, err := obj.buffer.toLLVMValue(globalType, mem)
+			if err != nil {
+				return llvm.Value{}, err
+			}
 			if checks && initializer.Type() != globalType {
-				panic("allocated value does not match allocated type")
+				return llvm.Value{}, errors.New("interp: allocated value does not match allocated type")
 			}
 			llvmValue.SetInitializer(initializer)
 		}
@@ -552,7 +569,7 @@ func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Valu
 	}
 
 	if llvmType.IsNil() {
-		return llvmValue
+		return llvmValue, nil
 	}
 
 	if llvmType.TypeKind() != llvm.PointerTypeKind {
@@ -564,7 +581,7 @@ func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Valu
 		// This can be worked around by simply converting to a raw value,
 		// rawValue knows how to create such structs.
 		if v.offset() != 0 {
-			panic("offset set without known pointer type")
+			return llvm.Value{}, errors.New("interp: offset set without known pointer type")
 		}
 		return v.asRawValue(mem.r).toLLVMValue(llvmType, mem)
 	}
@@ -575,14 +592,14 @@ func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Valu
 		if v.offset() != 0 {
 			// This should never happen, if offset is non-zero, the types
 			// shouldn't match.
-			panic("offset set while there is no way to convert the type")
+			return llvm.Value{}, errors.New("interp: offset set while there is no way to convert the type")
 		}
-		return llvmValue
+		return llvmValue, nil
 	}
 
 	if v.offset() == 0 {
 		// Offset is zero, so we can just bitcast to get a correct pointer.
-		return llvm.ConstBitCast(llvmValue, llvmType)
+		return llvm.ConstBitCast(llvmValue, llvmType), nil
 	}
 
 	// We need to make a constant GEP for pointer arithmetic.
@@ -606,11 +623,11 @@ func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Valu
 			offset -= int64(mem.r.targetData.ElementOffset(objectElementType, element))
 			objectElementType = objectElementType.StructElementTypes()[element]
 		default:
-			panic("pointer index with something other than a struct or array?")
+			return llvm.Value{}, errors.New("interp: pointer index with something other than a struct or array?")
 		}
 	}
 	if offset < 0 {
-		panic("offset has somehow gone negative, this should be impossible")
+		return llvm.Value{}, errors.New("interp: offset has somehow gone negative, this should be impossible")
 	}
 
 	// Finally do the gep, using the above computed indices.
@@ -618,9 +635,9 @@ func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Valu
 	// the bits of the pointer are now correct, just not the type).
 	gep := llvm.ConstInBoundsGEP(llvmValue, indices)
 	if gep.Type() != llvmType {
-		return llvm.ConstBitCast(gep, llvmType)
+		return llvm.ConstBitCast(gep, llvmType), nil
 	}
-	return gep
+	return gep, nil
 }
 
 // mapValue implements a Go map which is created at compile time and stored as a
@@ -713,7 +730,11 @@ func (b *mapBucket) create(ctx llvm.Context, nextBucket llvm.Value, mem *memoryV
 	// Create data for keys.
 	var keyValues []llvm.Value
 	for _, key := range b.keys {
-		keyValues = append(keyValues, key.rawLLVMValue(mem))
+		keyValue, err := key.rawLLVMValue(mem)
+		if err != nil {
+			panic(err)
+		}
+		keyValues = append(keyValues, keyValue)
 	}
 	if len(b.keys) < 8 {
 		keyValues = append(keyValues, llvm.ConstNull(llvm.ArrayType(int8Type, int(b.m.keySize)*(8-len(b.keys)))))
@@ -726,7 +747,11 @@ func (b *mapBucket) create(ctx llvm.Context, nextBucket llvm.Value, mem *memoryV
 	// Create data for values.
 	var valueValues []llvm.Value
 	for _, value := range b.values {
-		valueValues = append(valueValues, value.rawLLVMValue(mem))
+		v, err := value.rawLLVMValue(mem)
+		if err != nil {
+			panic(err)
+		}
+		valueValues = append(valueValues, v)
 	}
 	if len(b.values) < 8 {
 		valueValues = append(valueValues, llvm.ConstNull(llvm.ArrayType(int8Type, int(b.m.valueSize)*(8-len(b.values)))))
@@ -750,9 +775,9 @@ func (b *mapBucket) create(ctx llvm.Context, nextBucket llvm.Value, mem *memoryV
 	return bucket
 }
 
-func (v *mapValue) toLLVMValue(hashmapType llvm.Type, mem *memoryView) llvm.Value {
+func (v *mapValue) toLLVMValue(hashmapType llvm.Type, mem *memoryView) (llvm.Value, error) {
 	if !v.hashmap.IsNil() {
-		return v.hashmap
+		return v.hashmap, nil
 	}
 
 	// Create a slice of buckets with all the keys and values in the hashmap.
@@ -772,12 +797,12 @@ func (v *mapValue) toLLVMValue(hashmapType llvm.Type, mem *memoryView) llvm.Valu
 			copy(keyValue.buf[v.keySize/2:], literalValue{key.size}.asRawValue(v.r).buf)
 		case rawValue:
 			if key.hasPointer() {
-				panic("todo: map key with pointer")
+				return llvm.Value{}, errors.New("interp: todo: map key with pointer")
 			}
 			data = key.buf
 			keyValue = key
 		default:
-			panic("unknown map key type")
+			return llvm.Value{}, errors.New("interp: unknown map key type")
 		}
 		buf := make([]byte, len(data))
 		for i, p := range data {
@@ -821,7 +846,7 @@ func (v *mapValue) toLLVMValue(hashmapType llvm.Type, mem *memoryView) llvm.Valu
 	})
 
 	v.hashmap = hashmap
-	return v.hashmap
+	return v.hashmap, nil
 }
 
 // putString does a map assign operation, assuming that the map is of type
@@ -1018,7 +1043,7 @@ func (v rawValue) equal(rhs rawValue) bool {
 // goes. The resulting value does not have a specified type, but it will be the
 // same size and have the same bytes if it was created with a provided LLVM type
 // (through toLLVMValue).
-func (v rawValue) rawLLVMValue(mem *memoryView) llvm.Value {
+func (v rawValue) rawLLVMValue(mem *memoryView) (llvm.Value, error) {
 	var structFields []llvm.Value
 	ctx := mem.r.mod.Context()
 	int8Type := ctx.Int8Type()
@@ -1042,13 +1067,16 @@ func (v rawValue) rawLLVMValue(mem *memoryView) llvm.Value {
 	for i := uint32(0); i < uint32(len(v.buf)); {
 		if v.buf[i] > 255 {
 			addBytes()
-			field := pointerValue{v.buf[i]}.toLLVMValue(llvm.Type{}, mem)
+			field, err := pointerValue{v.buf[i]}.toLLVMValue(llvm.Type{}, mem)
+			if err != nil {
+				return llvm.Value{}, err
+			}
 			elementType := field.Type().ElementType()
 			if elementType.TypeKind() == llvm.StructTypeKind {
 				// There are some special pointer types that should be used as a
 				// ptrtoint, so that they can be used in certain optimizations.
 				name := elementType.StructName()
-				if name == "runtime.typeInInterface" || name == "runtime.funcValueWithSignature" {
+				if name == "runtime.typecodeID" || name == "runtime.funcValueWithSignature" {
 					uintptrType := ctx.IntType(int(mem.r.pointerSize) * 8)
 					field = llvm.ConstPtrToInt(field, uintptrType)
 				}
@@ -1065,12 +1093,12 @@ func (v rawValue) rawLLVMValue(mem *memoryView) llvm.Value {
 
 	// Return the created data.
 	if len(structFields) == 1 {
-		return structFields[0]
+		return structFields[0], nil
 	}
-	return ctx.ConstStruct(structFields, false)
+	return ctx.ConstStruct(structFields, false), nil
 }
 
-func (v rawValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
+func (v rawValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) (llvm.Value, error) {
 	isZero := true
 	for _, p := range v.buf {
 		if p != 0 {
@@ -1079,7 +1107,7 @@ func (v rawValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
 		}
 	}
 	if isZero {
-		return llvm.ConstNull(llvmType)
+		return llvm.ConstNull(llvmType), nil
 	}
 	switch llvmType.TypeKind() {
 	case llvm.IntegerTypeKind:
@@ -1088,7 +1116,17 @@ func (v rawValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
 			if err != nil {
 				panic(err)
 			}
-			return llvm.ConstPtrToInt(ptr.toLLVMValue(llvm.Type{}, mem), llvmType)
+			if checks && mem.r.targetData.TypeAllocSize(llvmType) != mem.r.targetData.TypeAllocSize(mem.r.i8ptrType) {
+				// Probably trying to serialize a pointer to a byte array,
+				// perhaps as a result of rawLLVMValue() in a previous interp
+				// run.
+				return llvm.Value{}, errInvalidPtrToIntSize
+			}
+			v, err := ptr.toLLVMValue(llvm.Type{}, mem)
+			if err != nil {
+				return llvm.Value{}, err
+			}
+			return llvm.ConstPtrToInt(v, llvmType), nil
 		}
 		var n uint64
 		switch llvmType.IntTypeWidth() {
@@ -1108,7 +1146,7 @@ func (v rawValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
 		default:
 			panic("unknown integer size")
 		}
-		return llvm.ConstInt(llvmType, n, false)
+		return llvm.ConstInt(llvmType, n, false), nil
 	case llvm.StructTypeKind:
 		fieldTypes := llvmType.StructElementTypes()
 		fields := make([]llvm.Value, len(fieldTypes))
@@ -1117,12 +1155,16 @@ func (v rawValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
 			field := rawValue{
 				buf: v.buf[offset:],
 			}
-			fields[i] = field.toLLVMValue(fieldType, mem)
+			var err error
+			fields[i], err = field.toLLVMValue(fieldType, mem)
+			if err != nil {
+				return llvm.Value{}, err
+			}
 		}
 		if llvmType.StructName() != "" {
-			return llvm.ConstNamedStruct(llvmType, fields)
+			return llvm.ConstNamedStruct(llvmType, fields), nil
 		}
-		return llvmType.Context().ConstStruct(fields, false)
+		return llvmType.Context().ConstStruct(fields, false), nil
 	case llvm.ArrayTypeKind:
 		numElements := llvmType.ArrayLength()
 		childType := llvmType.ElementType()
@@ -1133,27 +1175,34 @@ func (v rawValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
 			field := rawValue{
 				buf: v.buf[offset:],
 			}
-			fields[i] = field.toLLVMValue(childType, mem)
+			var err error
+			fields[i], err = field.toLLVMValue(childType, mem)
+			if err != nil {
+				return llvm.Value{}, err
+			}
 			if checks && fields[i].Type() != childType {
 				panic("child type doesn't match")
 			}
 		}
-		return llvm.ConstArray(childType, fields)
+		return llvm.ConstArray(childType, fields), nil
 	case llvm.PointerTypeKind:
 		if v.buf[0] > 255 {
 			// This is a regular pointer.
-			llvmValue := pointerValue{v.buf[0]}.toLLVMValue(llvm.Type{}, mem)
+			llvmValue, err := pointerValue{v.buf[0]}.toLLVMValue(llvm.Type{}, mem)
+			if err != nil {
+				return llvm.Value{}, err
+			}
 			if llvmValue.Type() != llvmType {
 				llvmValue = llvm.ConstBitCast(llvmValue, llvmType)
 			}
-			return llvmValue
+			return llvmValue, nil
 		}
 		// This is either a null pointer or a raw pointer for memory-mapped I/O
 		// (such as 0xe000ed00).
 		ptr := rawValue{v.buf[:mem.r.pointerSize]}.Uint()
 		if ptr == 0 {
 			// Null pointer.
-			return llvm.ConstNull(llvmType)
+			return llvm.ConstNull(llvmType), nil
 		}
 		var ptrValue llvm.Value // the underlying int
 		switch mem.r.pointerSize {
@@ -1164,19 +1213,19 @@ func (v rawValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
 		case 2:
 			ptrValue = llvm.ConstInt(llvmType.Context().Int16Type(), ptr, false)
 		default:
-			panic("unknown pointer size")
+			return llvm.Value{}, errors.New("interp: unknown pointer size")
 		}
-		return llvm.ConstIntToPtr(ptrValue, llvmType)
+		return llvm.ConstIntToPtr(ptrValue, llvmType), nil
 	case llvm.DoubleTypeKind:
 		b := rawValue{v.buf[:8]}.Uint()
 		f := math.Float64frombits(b)
-		return llvm.ConstFloat(llvmType, f)
+		return llvm.ConstFloat(llvmType, f), nil
 	case llvm.FloatTypeKind:
 		b := uint32(rawValue{v.buf[:4]}.Uint())
 		f := math.Float32frombits(b)
-		return llvm.ConstFloat(llvmType, float64(f))
+		return llvm.ConstFloat(llvmType, float64(f)), nil
 	default:
-		panic("todo: raw value to LLVM value: " + llvmType.String())
+		return llvm.Value{}, errors.New("interp: todo: raw value to LLVM value: " + llvmType.String())
 	}
 }
 
@@ -1365,8 +1414,8 @@ func (v localValue) Int() int64 {
 	panic("interp: localValue.Int")
 }
 
-func (v localValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
+func (v localValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) (llvm.Value, error) {
-	return v.value
+	return v.value, nil
 }
 
 func (r *runner) getValue(llvmValue llvm.Value) value {

interp/testdata/interface.ll

@@ -1,17 +1,16 @@
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64--linux"
 
-%runtime.typecodeID = type { %runtime.typecodeID*, i64 }
+%runtime.typecodeID = type { %runtime.typecodeID*, i64, %runtime.interfaceMethodInfo* }
 %runtime.interfaceMethodInfo = type { i8*, i64 }
-%runtime.typeInInterface = type { %runtime.typecodeID*, %runtime.interfaceMethodInfo* }
 
 @main.v1 = global i1 0
-@"reflect/types.type:named:main.foo" = private constant %runtime.typecodeID { %runtime.typecodeID* @"reflect/types.type:basic:int", i64 0 }
+@"reflect/types.type:named:main.foo" = private constant %runtime.typecodeID { %runtime.typecodeID* @"reflect/types.type:basic:int", i64 0, %runtime.interfaceMethodInfo* null }
+@"reflect/types.type:named:main.foo$id" = external constant i8
 @"reflect/types.type:basic:int" = external constant %runtime.typecodeID
-@"typeInInterface:reflect/types.type:named:main.foo" = private constant %runtime.typeInInterface { %runtime.typecodeID* @"reflect/types.type:named:main.foo", %runtime.interfaceMethodInfo* null }
 
 
-declare i1 @runtime.typeAssert(i64, %runtime.typecodeID*, i8*, i8*)
+declare i1 @runtime.typeAssert(i64, i8*, i8*, i8*)
 
 define void @runtime.initAll() unnamed_addr {
 entry:
@@ -22,7 +21,7 @@ entry:
 define internal void @main.init() unnamed_addr {
 entry:
   ; Test type asserts.
-  %typecode = call i1 @runtime.typeAssert(i64 ptrtoint (%runtime.typeInInterface* @"typeInInterface:reflect/types.type:named:main.foo" to i64), %runtime.typecodeID* @"reflect/types.type:named:main.foo", i8* undef, i8* null)
+  %typecode = call i1 @runtime.typeAssert(i64 ptrtoint (%runtime.typecodeID* @"reflect/types.type:named:main.foo" to i64), i8* @"reflect/types.type:named:main.foo$id", i8* undef, i8* null)
   store i1 %typecode, i1* @main.v1
   ret void
 }

loader/loader.go

@@ -2,6 +2,7 @@ package loader
 
 import (
 	"bytes"
+	"crypto/sha512"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -11,6 +12,7 @@ import (
 	"go/token"
 	"go/types"
 	"io"
+	"io/ioutil"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -66,10 +68,12 @@ type PackageJSON struct {
 type Package struct {
 	PackageJSON
 
-	program *Program
+	program    *Program
-	Files   []*ast.File
+	Files      []*ast.File
-	Pkg     *types.Package
+	FileHashes map[string][]byte
-	info    types.Info
+	CFlags     []string // CFlags used during CGo preprocessing (only set if CGo is used)
+	Pkg        *types.Package
+	info       types.Info
 }
 
 // Load loads the given package with all dependencies (including the runtime
@@ -118,7 +122,8 @@ func Load(config *compileopts.Config, inputPkgs []string, clangHeaders string, t
 	decoder := json.NewDecoder(buf)
 	for {
 		pkg := &Package{
-			program: p,
+			program:    p,
+			FileHashes: make(map[string][]byte),
 			info: types.Info{
 				Types:      make(map[ast.Expr]types.TypeAndValue),
 				Defs:       make(map[*ast.Ident]types.Object),
@@ -277,17 +282,15 @@ func (p *Program) Parse() error {
 }
 
 // parseFile is a wrapper around parser.ParseFile.
-func (p *Program) parseFile(path string, mode parser.Mode) (*ast.File, error) {
+func (p *Package) parseFile(path string, mode parser.Mode) (*ast.File, error) {
-	if p.fset == nil {
+	originalPath := p.program.getOriginalPath(path)
-		p.fset = token.NewFileSet()
+	data, err := ioutil.ReadFile(path)
-	}
-
-	rd, err := os.Open(path)
 	if err != nil {
 		return nil, err
 	}
-	defer rd.Close()
+	sum := sha512.Sum512_224(data)
-	return parser.ParseFile(p.fset, p.getOriginalPath(path), rd, mode)
+	p.FileHashes[originalPath] = sum[:]
+	return parser.ParseFile(p.program.fset, originalPath, data, mode)
 }
 
 // Parse parses and typechecks this package.
@@ -363,7 +366,7 @@ func (p *Package) parseFiles() ([]*ast.File, error) {
 		if !filepath.IsAbs(file) {
 			file = filepath.Join(p.Dir, file)
 		}
-		f, err := p.program.parseFile(file, parser.ParseComments)
+		f, err := p.parseFile(file, parser.ParseComments)
 		if err != nil {
 			fileErrs = append(fileErrs, err)
 			return
@@ -379,13 +382,17 @@ func (p *Package) parseFiles() ([]*ast.File, error) {
 
 	// Do CGo processing.
 	if len(p.CgoFiles) != 0 {
-		var cflags []string
+		var initialCFlags []string
-		cflags = append(cflags, p.program.config.CFlags()...)
+		initialCFlags = append(initialCFlags, p.program.config.CFlags()...)
-		cflags = append(cflags, "-I"+p.Dir)
+		initialCFlags = append(initialCFlags, "-I"+p.Dir)
 		if p.program.clangHeaders != "" {
-			cflags = append(cflags, "-Xclang", "-internal-isystem", "-Xclang", p.program.clangHeaders)
+			initialCFlags = append(initialCFlags, "-Xclang", "-internal-isystem", "-Xclang", p.program.clangHeaders)
+		}
+		generated, cflags, ldflags, accessedFiles, errs := cgo.Process(files, p.program.workingDir, p.program.fset, initialCFlags)
+		p.CFlags = append(initialCFlags, cflags...)
+		for path, hash := range accessedFiles {
+			p.FileHashes[path] = hash
 		}
-		generated, ldflags, errs := cgo.Process(files, p.program.workingDir, p.program.fset, cflags)
 		if errs != nil {
 			fileErrs = append(fileErrs, errs...)
 		}

main.go

@@ -93,7 +93,7 @@ func copyFile(src, dst string) error {
 // executeCommand is a simple wrapper to exec.Cmd
 func executeCommand(options *compileopts.Options, name string, arg ...string) *exec.Cmd {
 	if options.PrintCommands {
-		fmt.Printf("%s %s\n ", name, strings.Join(arg, " "))
+		fmt.Printf("%s %s\n", name, strings.Join(arg, " "))
 	}
 	return exec.Command(name, arg...)
 }
@@ -105,7 +105,7 @@ func Build(pkgName, outpath string, options *compileopts.Options) error {
 		return err
 	}
 
-	return builder.Build(pkgName, outpath, config, nil, func(result builder.BuildResult) error {
+	return builder.Build(pkgName, outpath, config, func(result builder.BuildResult) error {
 		if err := os.Rename(result.Binary, outpath); err != nil {
 			// Moving failed. Do a file copy.
 			inf, err := os.Open(result.Binary)
@@ -141,7 +141,7 @@ func Test(pkgName string, options *compileopts.Options, testCompileOnly bool, ou
 		return err
 	}
 
-	return builder.Build(pkgName, outpath, config, nil, func(result builder.BuildResult) error {
+	return builder.Build(pkgName, outpath, config, func(result builder.BuildResult) error {
 		if testCompileOnly || outpath != "" {
 			// Write test binary to the specified file name.
 			if outpath == "" {
@@ -237,7 +237,7 @@ func Flash(pkgName, port string, options *compileopts.Options) error {
 		return errors.New("unknown flash method: " + flashMethod)
 	}
 
-	return builder.Build(pkgName, fileExt, config, func() error {
+	return builder.Build(pkgName, fileExt, config, func(result builder.BuildResult) error {
 		// do we need port reset to put MCU into bootloader mode?
 		if config.Target.PortReset == "true" && flashMethod != "openocd" {
 			if port == "" {
@@ -250,14 +250,12 @@ func Flash(pkgName, port string, options *compileopts.Options) error {
 
 			err := touchSerialPortAt1200bps(port)
 			if err != nil {
-				return &commandError{"failed to reset port", "", err}
+				return &commandError{"failed to reset port", result.Binary, err}
 			}
 			// give the target MCU a chance to restart into bootloader
 			time.Sleep(3 * time.Second)
 		}
-		return nil
 
-	}, func(result builder.BuildResult) error {
 		// this flashing method copies the binary data to a Mass Storage Device (msd)
 		switch flashMethod {
 		case "", "command":
@@ -346,11 +344,12 @@ func FlashGDB(pkgName string, ocdOutput bool, options *compileopts.Options) erro
 	if err != nil {
 		return err
 	}
-	if config.Target.GDB == "" {
+	gdb, err := config.Target.LookupGDB()
-		return errors.New("gdb not configured in the target specification")
+	if err != nil {
+		return err
 	}
 
-	return builder.Build(pkgName, "", config, nil, func(result builder.BuildResult) error {
+	return builder.Build(pkgName, "", config, func(result builder.BuildResult) error {
 		// Find a good way to run GDB.
 		gdbInterface, openocdInterface := config.Programmer()
 		switch gdbInterface {
@@ -492,7 +491,7 @@ func FlashGDB(pkgName string, ocdOutput bool, options *compileopts.Options) erro
 		for _, cmd := range gdbCommands {
 			params = append(params, "-ex", cmd)
 		}
-		cmd := executeCommand(config.Options, config.Target.GDB, params...)
+		cmd := executeCommand(config.Options, gdb, params...)
 		cmd.Stdin = os.Stdin
 		cmd.Stdout = os.Stdout
 		cmd.Stderr = os.Stderr
@@ -514,7 +513,7 @@ func Run(pkgName string, options *compileopts.Options) error {
 		return err
 	}
 
-	return builder.Build(pkgName, ".elf", config, nil, func(result builder.BuildResult) error {
+	return builder.Build(pkgName, ".elf", config, func(result builder.BuildResult) error {
 		if len(config.Target.Emulator) == 0 {
 			// Run directly.
 			cmd := executeCommand(config.Options, result.Binary)

main_test.go

@@ -59,7 +59,10 @@ func TestCompiler(t *testing.T) {
 		t.Run("Host", func(t *testing.T) {
 			runPlatTests("", matches, t)
 			if runtime.GOOS == "darwin" {
-				runTest("testdata/libc/env.go", "", t, []string{"ENV1=VALUE1", "ENV2=VALUE2"}...)
+				runTest("testdata/libc/filesystem.go", "", t,
+					nil, nil)
+				runTest("testdata/libc/env.go", "", t,
+					[]string{"ENV1=VALUE1", "ENV2=VALUE2"}, nil)
 			}
 		})
 	}
@@ -107,7 +110,9 @@ func TestCompiler(t *testing.T) {
 
 		t.Run("WASI", func(t *testing.T) {
 			runPlatTests("wasi", matches, t)
-			runTest("testdata/libc/env.go", "wasi", t, []string{"ENV1=VALUE1", "ENV2=VALUE2"}...)
+			runTest("testdata/libc/env.go", "wasi", t,
+				[]string{"--env", "ENV1=VALUE1", "--env", "ENV2=VALUE2"}, nil)
+			runTest("testdata/libc/filesystem.go", "wasi", t, nil, []string{"--dir=."})
 		})
 	}
 }
@@ -119,7 +124,7 @@ func runPlatTests(target string, matches []string, t *testing.T) {
 		path := path // redefine to avoid race condition
 		t.Run(filepath.Base(path), func(t *testing.T) {
 			t.Parallel()
-			runTest(path, target, t)
+			runTest(path, target, t, nil, nil)
 		})
 	}
 }
@@ -136,7 +141,7 @@ func runBuild(src, out string, opts *compileopts.Options) error {
 	return Build(src, out, opts)
 }
 
-func runTest(path, target string, t *testing.T, environmentVars ...string) {
+func runTest(path, target string, t *testing.T, environmentVars []string, additionalArgs []string) {
 	// Get the expected output for this test.
 	txtpath := path[:len(path)-3] + ".txt"
 	if path[len(path)-1] == os.PathSeparator {
@@ -195,7 +200,7 @@ func runTest(path, target string, t *testing.T, environmentVars ...string) {
 			cmd = exec.Command(binary)
 		} else {
 			args := append(spec.Emulator[1:], binary)
-			cmd = exec.Command(spec.Emulator[0], args...)
+			cmd = exec.Command(spec.Emulator[0], append(args, additionalArgs...)...)
 		}
 
 		if len(spec.Emulator) != 0 && spec.Emulator[0] == "wasmtime" {

src/examples/pwm/arduino.go

@@ -0,0 +1,12 @@
+// +build arduino
+
+package main
+
+import "machine"
+
+var (
+	// Configuration on an Arduino Uno.
+	pwm  = machine.Timer2
+	pinA = machine.PB3 // pin 11 on the Uno
+	pinB = machine.PD3 // pin 3 on the Uno
+)

src/examples/pwm/feather-m4.go

@@ -0,0 +1,11 @@
+// +build feather_m4
+
+package main
+
+import "machine"
+
+var (
+	pwm  = machine.TCC0
+	pinA = machine.D12
+	pinB = machine.D13
+)

src/examples/pwm/itsybitsy-m0.go

@@ -0,0 +1,11 @@
+// +build itsybitsy_m0
+
+package main
+
+import "machine"
+
+var (
+	pwm  = machine.TCC0
+	pinA = machine.D3
+	pinB = machine.D4
+)

src/examples/pwm/itsybitsy-m4.go

@@ -0,0 +1,11 @@
+// +build itsybitsy_m4
+
+package main
+
+import "machine"
+
+var (
+	pwm  = machine.TCC0
+	pinA = machine.D12
+	pinB = machine.D13
+)

src/examples/pwm/pwm.go

@@ -1,64 +1,74 @@
 package main
 
+// This example demonstrates some features of the PWM support.
+
 import (
 	"machine"
 	"time"
 )
 
-// This example assumes that an RGB LED is connected to pins 3, 5 and 6 on an Arduino.
+const delayBetweenPeriods = time.Second * 5
-// Change the values below to use different pins.
-const (
-	redPin   = machine.D4
-	greenPin = machine.D5
-	bluePin  = machine.D6
-)
-
-// cycleColor is just a placeholder until math/rand or some equivalent is working.
-func cycleColor(color uint8) uint8 {
-	if color < 10 {
-		return color + 1
-	} else if color < 200 {
-		return color + 10
-	} else {
-		return 0
-	}
-}
 
 func main() {
-	machine.InitPWM()
+	// Delay a bit on startup to easily catch the first messages.
+	time.Sleep(time.Second * 2)
 
-	red := machine.PWM{redPin}
+	// Configure the PWM with the given period.
-	err := red.Configure()
+	err := pwm.Configure(machine.PWMConfig{
-	checkError(err, "failed to configure red pin")
+		Period: 16384e3, // 16.384ms
+	})
+	if err != nil {
+		println("failed to configure PWM")
+		return
+	}
 
-	green := machine.PWM{greenPin}
+	// The top value is the highest value that can be passed to PWMChannel.Set.
-	err = green.Configure()
+	// It is usually an even number.
-	checkError(err, "failed to configure green pin")
+	println("top:", pwm.Top())
 
-	blue := machine.PWM{bluePin}
+	// Configure the two channels we'll use as outputs.
-	err = blue.Configure()
+	channelA, err := pwm.Channel(pinA)
-	checkError(err, "failed to configure blue pin")
+	if err != nil {
+		println("failed to configure channel A")
+		return
+	}
+	channelB, err := pwm.Channel(pinB)
+	if err != nil {
+		println("failed to configure channel B")
+		return
+	}
 
-	var rc uint8
+	// Invert one of the channels to demonstrate output polarity.
-	var gc uint8 = 20
+	pwm.SetInverting(channelB, true)
-	var bc uint8 = 30
+
+	// Test out various frequencies below, including some edge cases.
+
+	println("running at 0% duty cycle")
+	pwm.Set(channelA, 0)
+	pwm.Set(channelB, 0)
+	time.Sleep(delayBetweenPeriods)
+
+	println("running at 1")
+	pwm.Set(channelA, 1)
+	pwm.Set(channelB, 1)
+	time.Sleep(delayBetweenPeriods)
+
+	println("running at 25% duty cycle")
+	pwm.Set(channelA, pwm.Top()/4)
+	pwm.Set(channelB, pwm.Top()/4)
+	time.Sleep(delayBetweenPeriods)
+
+	println("running at top-1")
+	pwm.Set(channelA, pwm.Top()-1)
+	pwm.Set(channelB, pwm.Top()-1)
+	time.Sleep(delayBetweenPeriods)
+
+	println("running at 100% duty cycle")
+	pwm.Set(channelA, pwm.Top())
+	pwm.Set(channelB, pwm.Top())
+	time.Sleep(delayBetweenPeriods)
 
 	for {
-		rc = cycleColor(rc)
+		time.Sleep(time.Second)
-		gc = cycleColor(gc)
-		bc = cycleColor(bc)
-
-		red.Set(uint16(rc) << 8)
-		green.Set(uint16(gc) << 8)
-		blue.Set(uint16(bc) << 8)
-
-		time.Sleep(time.Millisecond * 500)
-	}
-}
-
-func checkError(err error, msg string) {
-	if err != nil {
-		print(msg, ": ", err.Error())
-		println()
 	}
 }

src/internal/task/task_coroutine.go

@@ -11,7 +11,7 @@ import (
 type rawState uint8
 
 //export llvm.coro.resume
-func (s *rawState) resume()
+func coroResume(*rawState)
 
 type state struct{ *rawState }
 
@@ -20,7 +20,7 @@ func noopState() *rawState
 
 // Resume the task until it pauses or completes.
 func (t *Task) Resume() {
-	t.state.resume()
+	coroResume(t.state.rawState)
 }
 
 // setState is used by the compiler to set the state of the function at the beginning of a function call.
@@ -77,22 +77,9 @@ func Current() *Task
 // This is implemented inside the compiler.
 func Pause()
 
-type taskHolder interface {
-	setState(*rawState) *rawState
-	returnTo(*rawState)
-	returnCurrent()
-	setReturnPtr(unsafe.Pointer)
-	getReturnPtr() unsafe.Pointer
-}
-
-// If there are no direct references to the task methods, they will not be discovered by the compiler, and this will trigger a compiler error.
-// Instantiating this interface forces discovery of these methods.
-var _ = taskHolder((*Task)(nil))
-
 func fake() {
 	// Hack to ensure intrinsics are discovered.
 	Current()
-	go func() {}()
 	Pause()
 }
 

src/machine/board_arduino_nano33_baremetal.go

@@ -32,7 +32,7 @@ func init() {
 
 // I2C on the Arduino Nano 33.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM4_I2CM,
 		SERCOM: 4,
 	}

src/machine/board_bluepill.go

@@ -85,6 +85,6 @@ const (
 
 // I2C pins
 const (
-	SDA_PIN = PB7
+	I2C0_SDA_PIN = PB7
-	SCL_PIN = PB6
+	I2C0_SCL_PIN = PB6
 )

src/machine/board_circuitplay_express.go

@@ -25,8 +25,8 @@ const (
 
 // Analog Pins
 const (
-	A0  = PA02 // PWM available, also ADC/AIN[0]
+	A0  = PA02 // ADC/AIN[0]
-	A1  = PA05 // ADC/AIN[5]
+	A1  = PA05 // PWM available, also ADC/AIN[5]
 	A2  = PA06 // PWM available, also ADC/AIN[6]
 	A3  = PA07 // PWM available, also ADC/AIN[7]
 	A4  = PB03 // PORTB

src/machine/board_circuitplay_express_baremetal.go

@@ -23,12 +23,12 @@ func init() {
 // I2C on the Circuit Playground Express.
 var (
 	// external device
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM5_I2CM,
 		SERCOM: 5,
 	}
 	// internal device
-	I2C1 = I2C{
+	I2C1 = &I2C{
 		Bus:    sam.SERCOM1_I2CM,
 		SERCOM: 1,
 	}

src/machine/board_feather-m0.go

@@ -75,7 +75,7 @@ const (
 
 // I2C on the Feather M0.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM3_I2CM,
 		SERCOM: 3,
 	}

src/machine/board_feather-m4_baremetal.go

@@ -28,7 +28,7 @@ func init() {
 
 // I2C on the Feather M4.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM2_I2CM,
 		SERCOM: 2,
 	}

src/machine/board_feather-stm32f405.go

@@ -120,19 +120,22 @@ const (
 
 var (
 	UART1 = UART{
-		Buffer:          NewRingBuffer(),
+		Buffer:            NewRingBuffer(),
-		Bus:             stm32.USART3,
+		Bus:               stm32.USART3,
-		AltFuncSelector: AF7_USART1_2_3,
+		TxAltFuncSelector: AF7_USART1_2_3,
+		RxAltFuncSelector: AF7_USART1_2_3,
 	}
 	UART2 = UART{
-		Buffer:          NewRingBuffer(),
+		Buffer:            NewRingBuffer(),
-		Bus:             stm32.USART6,
+		Bus:               stm32.USART6,
-		AltFuncSelector: AF8_USART4_5_6,
+		TxAltFuncSelector: AF8_USART4_5_6,
+		RxAltFuncSelector: AF8_USART4_5_6,
 	}
 	UART3 = UART{
-		Buffer:          NewRingBuffer(),
+		Buffer:            NewRingBuffer(),
-		Bus:             stm32.USART1,
+		Bus:               stm32.USART1,
-		AltFuncSelector: AF7_USART1_2_3,
+		TxAltFuncSelector: AF7_USART1_2_3,
+		RxAltFuncSelector: AF7_USART1_2_3,
 	}
 	UART0 = UART1
 )
@@ -227,15 +230,15 @@ const (
 )
 
 var (
-	I2C1 = I2C{
+	I2C1 = &I2C{
 		Bus:             stm32.I2C1,
 		AltFuncSelector: AF4_I2C1_2_3,
 	}
-	I2C2 = I2C{
+	I2C2 = &I2C{
 		Bus:             stm32.I2C2,
 		AltFuncSelector: AF4_I2C1_2_3,
 	}
-	I2C3 = I2C{
+	I2C3 = &I2C{
 		Bus:             stm32.I2C1,
 		AltFuncSelector: AF4_I2C1_2_3,
 	}

src/machine/board_grandcentral-m4_baremetal.go

@@ -40,11 +40,11 @@ var (
 
 // I2C on the Grand Central M4
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM3_I2CM,
 		SERCOM: 3,
 	}
-	I2C1 = I2C{
+	I2C1 = &I2C{
 		Bus:    sam.SERCOM6_I2CM,
 		SERCOM: 6,
 	}

src/machine/board_hifive1b_baremetal.go

@@ -10,10 +10,3 @@ var (
 		Bus: sifive.QSPI1,
 	}
 )
-
-// I2C on the HiFive1 rev B.
-var (
-	I2C0 = I2C{
-		Bus: sifive.I2C0,
-	}
-)

src/machine/board_itsybitsy-m0.go

@@ -75,7 +75,7 @@ const (
 
 // I2C on the ItsyBitsy M0.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM3_I2CM,
 		SERCOM: 3,
 	}

src/machine/board_itsybitsy-m4_baremetal.go

@@ -28,7 +28,7 @@ func init() {
 
 // I2C on the ItsyBitsy M4.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM2_I2CM,
 		SERCOM: 2,
 	}

src/machine/board_lgt92.go

@@ -41,31 +41,44 @@ const (
 	// LORA RFM95 Radio
 	RFM95_DIO0_PIN = PC13
 
-	//TinyGo UART is MCU LPUSART1
+	// TinyGo UART is MCU LPUSART1
 	UART_RX_PIN = PA13
 	UART_TX_PIN = PA14
 
-	//TinyGo UART1 is MCU USART1
+	// TinyGo UART1 is MCU USART1
 	UART1_RX_PIN = PB6
 	UART1_TX_PIN = PB7
+
+	// MPU9250 Nine-Axis (Gyro + Accelerometer + Compass)
+	I2C0_SCL_PIN = PA9
+	I2C0_SDA_PIN = PA10
 )
 
 var (
 
 	// Console UART (LPUSART1)
 	UART0 = UART{
-		Buffer:          NewRingBuffer(),
+		Buffer:            NewRingBuffer(),
-		Bus:             stm32.LPUART1,
+		Bus:               stm32.LPUART1,
-		AltFuncSelector: 6,
+		TxAltFuncSelector: 6,
+		RxAltFuncSelector: 6,
 	}
 
 	// Gps UART
 	UART1 = UART{
-		Buffer:          NewRingBuffer(),
+		Buffer:            NewRingBuffer(),
-		Bus:             stm32.USART1,
+		Bus:               stm32.USART1,
-		AltFuncSelector: 0,
+		TxAltFuncSelector: 0,
+		RxAltFuncSelector: 0,
 	}
 
+	// MPU9250 Nine-Axis (Gyro + Accelerometer + Compass)
+	I2C1 = &I2C{
+		Bus:             stm32.I2C1,
+		AltFuncSelector: 6,
+	}
+	I2C0 = I2C1
+
 	// SPI
 	SPI0 = SPI{
 		Bus: stm32.SPI1,

src/machine/board_maixbit_baremetal.go

@@ -13,16 +13,3 @@ var (
 		Bus: kendryte.SPI1,
 	}
 )
-
-// I2C on the MAix Bit.
-var (
-	I2C0 = I2C{
-		Bus: kendryte.I2C0,
-	}
-	I2C1 = I2C{
-		Bus: kendryte.I2C1,
-	}
-	I2C2 = I2C{
-		Bus: kendryte.I2C2,
-	}
-)

src/machine/board_matrixportal-m4_baremetal.go

@@ -29,7 +29,7 @@ func init() {
 
 // I2C on the MatrixPortal M4
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM5_I2CM,
 		SERCOM: 5,
 	}

src/machine/board_metro-m4-airlift_baremetal.go

@@ -28,7 +28,7 @@ func init() {
 
 // I2C on the Metro M4.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM5_I2CM,
 		SERCOM: 5,
 	}

src/machine/board_nrf52840-mdk-usb-dongle.go

@@ -44,7 +44,7 @@ const (
 // USB CDC identifiers
 const (
 	usb_STRING_PRODUCT      = "Makerdiary nRF52840 MDK USB Dongle"
-	usb_STRING_MANUFACTURER = "Makerdiary"
+	usb_STRING_MANUFACTURER = "Nordic Semiconductor ASA"
 )
 
 var (

src/machine/board_nrf52840-mdk.go

@@ -39,7 +39,7 @@ const (
 // USB CDC identifiers
 const (
 	usb_STRING_PRODUCT      = "Makerdiary nRF52840 MDK"
-	usb_STRING_MANUFACTURER = "Makerdiary"
+	usb_STRING_MANUFACTURER = "Nordic Semiconductor ASA"
 )
 
 var (

src/machine/board_nucleof103rb.go

@@ -120,6 +120,6 @@ const (
 
 // I2C pins
 const (
-	SCL_PIN = PB6
+	I2C0_SCL_PIN = PB6
-	SDA_PIN = PB7
+	I2C0_SDA_PIN = PB7
 )

src/machine/board_nucleof722ze.go

@@ -33,9 +33,10 @@ var (
 	// debugger to be exposed as virtual COM port over USB on Nucleo boards.
 	// Both UART0 and UART1 refer to USART2.
 	UART0 = UART{
-		Buffer:          NewRingBuffer(),
+		Buffer:            NewRingBuffer(),
-		Bus:             stm32.USART3,
+		Bus:               stm32.USART3,
-		AltFuncSelector: UART_ALT_FN,
+		TxAltFuncSelector: UART_ALT_FN,
+		RxAltFuncSelector: UART_ALT_FN,
 	}
 	UART1 = &UART0
 )
@@ -53,6 +54,15 @@ const (
 
 // I2C pins
 const (
-	SCL_PIN = PB6
+	I2C0_SCL_PIN = PB8
-	SDA_PIN = PB7
+	I2C0_SDA_PIN = PB9
+)
+
+var (
+	// I2C1 is documented, alias to I2C0 as well
+	I2C1 = &I2C{
+		Bus:             stm32.I2C1,
+		AltFuncSelector: 4,
+	}
+	I2C0 = I2C1
 )

src/machine/board_nucleol031k6.go

@@ -0,0 +1,92 @@
+// +build nucleol031k6
+
+package machine
+
+import (
+	"device/stm32"
+	"runtime/interrupt"
+)
+
+const (
+	LED         = LED_BUILTIN
+	LED_BUILTIN = LED_GREEN
+	LED_GREEN   = PB3
+)
+
+const (
+	// Arduino Pins
+	A0 = PA0 // ADC_IN0
+	A1 = PA1 // ADC_IN1
+	A2 = PA3 // ADC_IN3
+	A3 = PA4 // ADC_IN4
+	A4 = PA5 // ADC_IN5 || I2C1_SDA
+	A5 = PA6 // ADC_IN6 || I2C1_SCL
+	A6 = PA7 // ADC_IN7
+	A7 = PA2 // ADC_IN2
+
+	D0  = PA10 // USART1_TX
+	D1  = PA9  // USART1_RX
+	D2  = PA12
+	D3  = PB0 // TIM2_CH3
+	D4  = PB7
+	D5  = PB6  // TIM16_CH1N
+	D6  = PB1  // TIM14_CH1
+	D9  = PA8  // TIM1_CH1
+	D10 = PA11 // SPI_CS || TIM1_CH4
+	D11 = PB5  // SPI1_MOSI || TIM3_CH2
+	D12 = PB4  // SPI1_MISO
+	D13 = PB3  // SPI1_SCK
+)
+
+const (
+	// UART pins
+	// PA2 and PA15 are connected to the ST-Link Virtual Com Port (VCP)
+	UART_TX_PIN = PA2
+	UART_RX_PIN = PA15
+
+	// SPI
+	SPI1_SCK_PIN = PB3
+	SPI1_SDI_PIN = PB5
+	SPI1_SDO_PIN = PB4
+	SPI0_SCK_PIN = SPI1_SCK_PIN
+	SPI0_SDI_PIN = SPI1_SDI_PIN
+	SPI0_SDO_PIN = SPI1_SDO_PIN
+
+	// I2C pins
+	// PB6 and PB7 are mapped to CN4 pin 7 and CN4 pin 8 respectively with the
+	// default solder bridge settings
+	I2C0_SCL_PIN  = PB7
+	I2C0_SDA_PIN  = PB6
+	I2C0_ALT_FUNC = 1
+)
+
+var (
+	// USART2 is the hardware serial port connected to the onboard ST-LINK
+	// debugger to be exposed as virtual COM port over USB on Nucleo boards.
+	// Both UART0 and UART1 refer to USART2.
+	UART0 = UART{
+		Buffer:            NewRingBuffer(),
+		Bus:               stm32.USART2,
+		TxAltFuncSelector: 4,
+		RxAltFuncSelector: 4,
+	}
+	UART1 = &UART0
+
+	// I2C1 is documented, alias to I2C0 as well
+	I2C1 = &I2C{
+		Bus:             stm32.I2C1,
+		AltFuncSelector: 1,
+	}
+	I2C0 = I2C1
+
+	// SPI
+	SPI0 = SPI{
+		Bus:             stm32.SPI1,
+		AltFuncSelector: 0,
+	}
+	SPI1 = &SPI0
+)
+
+func init() {
+	UART0.Interrupt = interrupt.New(stm32.IRQ_USART2, UART0.handleInterrupt)
+}

src/machine/board_nucleol432kc.go

@@ -0,0 +1,56 @@
+// +build nucleol432kc
+
+package machine
+
+import (
+	"device/stm32"
+	"runtime/interrupt"
+)
+
+const (
+	LED         = LED_BUILTIN
+	LED_BUILTIN = LED_GREEN
+	LED_GREEN   = PB3
+)
+
+// UART pins
+const (
+	// PA2 and PA15 are connected to the ST-Link Virtual Com Port (VCP)
+	UART_TX_PIN = PA2
+	UART_RX_PIN = PA15
+)
+
+// I2C pins
+const (
+	// With default solder bridge settings:
+	//    PB6 / Arduino D5 / CN3 Pin 8 is SCL
+	//    PB7 / Arduino D4 / CN3 Pin 7 is SDA
+	I2C0_SCL_PIN = PB6
+	I2C0_SDA_PIN = PB7
+)
+
+var (
+	// USART2 is the hardware serial port connected to the onboard ST-LINK
+	// debugger to be exposed as virtual COM port over USB on Nucleo boards.
+	// Both UART0 and UART1 refer to USART2.
+	UART0 = UART{
+		Buffer:            NewRingBuffer(),
+		Bus:               stm32.USART2,
+		TxAltFuncSelector: 7,
+		RxAltFuncSelector: 3,
+	}
+	UART1 = &UART0
+)
+
+var (
+	// I2C1 is documented, alias to I2C0 as well
+	I2C1 = &I2C{
+		Bus:             stm32.I2C1,
+		AltFuncSelector: 4,
+	}
+	I2C0 = I2C1
+)
+
+func init() {
+	UART0.Interrupt = interrupt.New(stm32.IRQ_USART2, UART0.handleInterrupt)
+}

src/machine/board_nucleol552ze.go

@@ -33,13 +33,28 @@ var (
 	// debugger to be exposed as virtual COM port over USB on Nucleo boards.
 	// Both UART0 and UART1 refer to LPUART1.
 	UART0 = UART{
-		Buffer:          NewRingBuffer(),
+		Buffer:            NewRingBuffer(),
-		Bus:             stm32.LPUART1,
+		Bus:               stm32.LPUART1,
-		AltFuncSelector: UART_ALT_FN,
+		TxAltFuncSelector: UART_ALT_FN,
+		RxAltFuncSelector: UART_ALT_FN,
 	}
 	UART1 = &UART0
 )
 
+const (
+	I2C0_SCL_PIN = PB8
+	I2C0_SDA_PIN = PB9
+)
+
+var (
+	// I2C1 is documented, alias to I2C0 as well
+	I2C1 = &I2C{
+		Bus:             stm32.I2C1,
+		AltFuncSelector: 4,
+	}
+	I2C0 = I2C1
+)
+
 func init() {
 	UART0.Interrupt = interrupt.New(stm32.IRQ_LPUART1, UART0.handleInterrupt)
 }

src/machine/board_p1am-100_baremetal.go

@@ -22,7 +22,7 @@ func init() {
 
 // I2C on the P1AM-100.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM0_I2CM,
 		SERCOM: 0,
 	}

src/machine/board_pca10059.go

@@ -0,0 +1,64 @@
+// +build pca10059
+
+package machine
+
+// The PCA10040 has a low-frequency (32kHz) crystal oscillator on board.
+const HasLowFrequencyCrystal = true
+
+// LEDs on the PCA10059 (nRF52840 dongle)
+const (
+	LED  Pin = LED1
+	LED1 Pin = 6
+	LED2 Pin = 8
+	LED3 Pin = (1 << 5) | 9
+	LED4 Pin = 12
+)
+
+// Buttons on the PCA10059 (nRF52840 dongle)
+const (
+	BUTTON  Pin = BUTTON1
+	BUTTON1 Pin = (1 << 5) | 6
+)
+
+// ADC pins
+const (
+	ADC1 Pin = 2
+	ADC2 Pin = 4
+	ADC3 Pin = 29
+	ADC4 Pin = 31
+)
+
+// UART pins
+const (
+	UART_TX_PIN Pin = NoPin
+	UART_RX_PIN Pin = NoPin
+)
+
+// UART0 is the USB device
+var (
+	UART0 = USB
+)
+
+// I2C pins (unused)
+const (
+	SDA_PIN = NoPin
+	SCL_PIN = NoPin
+)
+
+// SPI pins (unused)
+const (
+	SPI0_SCK_PIN = NoPin
+	SPI0_SDO_PIN = NoPin
+	SPI0_SDI_PIN = NoPin
+)
+
+// USB CDC identifiers
+const (
+	usb_STRING_PRODUCT      = "nRF52840 Dongle"
+	usb_STRING_MANUFACTURER = "Nordic Semiconductor ASA"
+)
+
+var (
+	usb_VID uint16 = 0x1915
+	usb_PID uint16 = 0xCAFE
+)

src/machine/board_pybadge_baremetal.go

@@ -28,7 +28,7 @@ func init() {
 
 // I2C on the ItsyBitsy M4.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM2_I2CM,
 		SERCOM: 2,
 	}

src/machine/board_pygamer.go

@@ -99,7 +99,7 @@ const (
 
 // I2C on the PyGamer.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM2_I2CM,
 		SERCOM: 2,
 	}

src/machine/board_pyportal_baremetal.go

@@ -21,7 +21,7 @@ func init() {
 
 // I2C on the PyPortal.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM5_I2CM,
 		SERCOM: 5,
 	}

src/machine/board_qtpy.go

@@ -93,7 +93,7 @@ const (
 
 // I2C on the QT Py M0.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM2_I2CM,
 		SERCOM: 2,
 	}

src/machine/board_stm32.go

@@ -1,5 +0,0 @@
-// +build bluepill nucleof103rb stm32f4
-
-package machine
-
-// Peripheral abstraction layer for the stm32.

src/machine/board_stm32f4disco.go

@@ -28,9 +28,10 @@ const (
 
 var (
 	UART0 = UART{
-		Buffer:          NewRingBuffer(),
+		Buffer:            NewRingBuffer(),
-		Bus:             stm32.USART2,
+		Bus:               stm32.USART2,
-		AltFuncSelector: AF7_USART1_2_3,
+		TxAltFuncSelector: AF7_USART1_2_3,
+		RxAltFuncSelector: AF7_USART1_2_3,
 	}
 	UART1 = &UART0
 )
@@ -73,7 +74,7 @@ const (
 )
 
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:             stm32.I2C1,
 		AltFuncSelector: AF4_I2C1_2_3,
 	}

src/machine/board_trinket.go

@@ -81,7 +81,7 @@ const (
 
 // I2C on the Trinket M0.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM2_I2CM,
 		SERCOM: 2,
 	}

src/machine/board_wioterminal_baremetal.go

@@ -29,12 +29,12 @@ func init() {
 
 // I2C on the Wio Terminal
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM4_I2CM,
 		SERCOM: 4,
 	}
 
-	I2C1 = I2C{
+	I2C1 = &I2C{
 		Bus:    sam.SERCOM4_I2CM,
 		SERCOM: 4,
 	}

src/machine/board_xiao.go

@@ -81,7 +81,7 @@ const (
 
 // I2C on the Xiao
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM2_I2CM,
 		SERCOM: 2,
 	}

src/machine/i2c.go

@@ -1,4 +1,4 @@
-// +build avr nrf sam stm32,!stm32f7x2,!stm32l5x2,!stm32l0 fe310 k210
+// +build atmega nrf sam stm32 fe310 k210
 
 package machine
 
@@ -29,7 +29,7 @@ var (
 // Many I2C-compatible devices are organized in terms of registers. This method
 // is a shortcut to easily write to such registers. Also, it only works for
 // devices with 7-bit addresses, which is the vast majority.
-func (i2c I2C) WriteRegister(address uint8, register uint8, data []byte) error {
+func (i2c *I2C) WriteRegister(address uint8, register uint8, data []byte) error {
 	buf := make([]uint8, len(data)+1)
 	buf[0] = register
 	copy(buf[1:], data)
@@ -42,6 +42,6 @@ func (i2c I2C) WriteRegister(address uint8, register uint8, data []byte) error {
 // Many I2C-compatible devices are organized in terms of registers. This method
 // is a shortcut to easily read such registers. Also, it only works for devices
 // with 7-bit addresses, which is the vast majority.
-func (i2c I2C) ReadRegister(address uint8, register uint8, data []byte) error {
+func (i2c *I2C) ReadRegister(address uint8, register uint8, data []byte) error {
 	return i2c.Tx(uint16(address), []byte{register}, data)
 }

src/machine/machine.go

@@ -37,10 +37,6 @@ func (p Pin) Low() {
 	p.Set(false)
 }
 
-type PWM struct {
-	Pin Pin
-}
-
 type ADC struct {
 	Pin Pin
 }

src/machine/machine_atmega.go

@@ -9,13 +9,20 @@ import (
 	"unsafe"
 )
 
+// I2C on AVR.
+type I2C struct {
+}
+
+// I2C0 is the only I2C interface on most AVRs.
+var I2C0 *I2C = nil
+
 // I2CConfig is used to store config info for I2C.
 type I2CConfig struct {
 	Frequency uint32
 }
 
 // Configure is intended to setup the I2C interface.
-func (i2c I2C) Configure(config I2CConfig) error {
+func (i2c *I2C) Configure(config I2CConfig) error {
 	// Default I2C bus speed is 100 kHz.
 	if config.Frequency == 0 {
 		config.Frequency = TWI_FREQ_100KHZ
@@ -42,7 +49,7 @@ func (i2c I2C) Configure(config I2CConfig) error {
 // Tx does a single I2C transaction at the specified address.
 // It clocks out the given address, writes the bytes in w, reads back len(r)
 // bytes and stores them in r, and generates a stop condition on the bus.
-func (i2c I2C) Tx(addr uint16, w, r []byte) error {
+func (i2c *I2C) Tx(addr uint16, w, r []byte) error {
 	if len(w) != 0 {
 		i2c.start(uint8(addr), true) // start transmission for writing
 		for _, b := range w {
@@ -63,7 +70,7 @@ func (i2c I2C) Tx(addr uint16, w, r []byte) error {
 }
 
 // start starts an I2C communication session.
-func (i2c I2C) start(address uint8, write bool) {
+func (i2c *I2C) start(address uint8, write bool) {
 	// Clear TWI interrupt flag, put start condition on SDA, and enable TWI.
 	avr.TWCR.Set((avr.TWCR_TWINT | avr.TWCR_TWSTA | avr.TWCR_TWEN))
 
@@ -80,7 +87,7 @@ func (i2c I2C) start(address uint8, write bool) {
 }
 
 // stop ends an I2C communication session.
-func (i2c I2C) stop() {
+func (i2c *I2C) stop() {
 	// Send stop condition.
 	avr.TWCR.Set(avr.TWCR_TWEN | avr.TWCR_TWINT | avr.TWCR_TWSTO)
 
@@ -90,7 +97,7 @@ func (i2c I2C) stop() {
 }
 
 // writeByte writes a single byte to the I2C bus.
-func (i2c I2C) writeByte(data byte) {
+func (i2c *I2C) writeByte(data byte) {
 	// Write data to register.
 	avr.TWDR.Set(data)
 
@@ -103,7 +110,7 @@ func (i2c I2C) writeByte(data byte) {
 }
 
 // readByte reads a single byte from the I2C bus.
-func (i2c I2C) readByte() byte {
+func (i2c *I2C) readByte() byte {
 	// Clear TWI interrupt flag and enable TWI.
 	avr.TWCR.Set(avr.TWCR_TWEN | avr.TWCR_TWINT | avr.TWCR_TWEA)
 

src/machine/machine_atmega328p.go

@@ -4,6 +4,7 @@ package machine
 
 import (
 	"device/avr"
+	"runtime/interrupt"
 	"runtime/volatile"
 )
 
@@ -21,71 +22,432 @@ func (p Pin) getPortMask() (*volatile.Register8, uint8) {
 	}
 }
 
-// InitPWM initializes the registers needed for PWM.
+// PWM is one PWM peripheral, which consists of a counter and two output
-func InitPWM() {
+// channels (that can be connected to two fixed pins). You can set the frequency
-	// use waveform generation
+// using SetPeriod, but only for all the channels in this PWM peripheral at
-	avr.TCCR0A.SetBits(avr.TCCR0A_WGM00)
+// once.
-
+type PWM struct {
-	// set timer 0 prescale factor to 64
+	num uint8
-	avr.TCCR0B.SetBits(avr.TCCR0B_CS01 | avr.TCCR0B_CS00)
-
-	// set timer 1 prescale factor to 64
-	avr.TCCR1B.SetBits(avr.TCCR1B_CS11)
-
-	// put timer 1 in 8-bit phase correct pwm mode
-	avr.TCCR1A.SetBits(avr.TCCR1A_WGM10)
-
-	// set timer 2 prescale factor to 64
-	avr.TCCR2B.SetBits(avr.TCCR2B_CS22)
-
-	// configure timer 2 for phase correct pwm (8-bit)
-	avr.TCCR2A.SetBits(avr.TCCR2A_WGM20)
 }
 
-// Configure configures a PWM pin for output.
+var (
-func (pwm PWM) Configure() error {
+	Timer0 = PWM{0} // 8 bit timer for PD5 and PD6
-	switch pwm.Pin / 8 {
+	Timer1 = PWM{1} // 16 bit timer for PB1 and PB2
-	case 0: // port B
+	Timer2 = PWM{2} // 8 bit timer for PB3 and PD3
-		avr.DDRB.SetBits(1 << uint8(pwm.Pin))
+)
-	case 2: // port D
+
-		avr.DDRD.SetBits(1 << uint8(pwm.Pin-16))
+// Configure enables and configures this PWM.
+//
+// For the two 8 bit timers, there is only a limited number of periods
+// available, namely the CPU frequency divided by 256 and again divided by 1, 8,
+// 64, 256, or 1024. For a MCU running at 16MHz, this would be a period of 16µs,
+// 128µs, 1024µs, 4096µs, or 16384µs.
+func (pwm PWM) Configure(config PWMConfig) error {
+	switch pwm.num {
+	case 0, 2: // 8-bit timers (Timer/counter 0 and Timer/counter 2)
+		// Calculate the timer prescaler.
+		// While we could configure a flexible top, that would sacrifice one of
+		// the PWM output compare registers and thus a PWM channel. I've chosen
+		// to instead limit this timer to a fixed number of frequencies.
+		var prescaler uint8
+		switch config.Period {
+		case 0, (uint64(1e9) * 256 * 1) / uint64(CPUFrequency()):
+			prescaler = 1
+		case (uint64(1e9) * 256 * 8) / uint64(CPUFrequency()):
+			prescaler = 2
+		case (uint64(1e9) * 256 * 64) / uint64(CPUFrequency()):
+			prescaler = 3
+		case (uint64(1e9) * 256 * 256) / uint64(CPUFrequency()):
+			prescaler = 4
+		case (uint64(1e9) * 256 * 1024) / uint64(CPUFrequency()):
+			prescaler = 5
+		default:
+			return ErrPWMPeriodTooLong
+		}
+
+		if pwm.num == 0 {
+			avr.TCCR0B.Set(prescaler)
+			// Set the PWM mode to fast PWM (mode = 3).
+			avr.TCCR0A.Set(avr.TCCR0A_WGM00 | avr.TCCR0A_WGM01)
+		} else {
+			avr.TCCR2B.Set(prescaler)
+			// Set the PWM mode to fast PWM (mode = 3).
+			avr.TCCR2A.Set(avr.TCCR2A_WGM20 | avr.TCCR2A_WGM21)
+		}
+	case 1: // Timer/counter 1
+		// The top value is the number of PWM ticks a PWM period takes. It is
+		// initially picked assuming an unlimited counter top and no PWM
+		// prescaler.
+		var top uint64
+		if config.Period == 0 {
+			// Use a top appropriate for LEDs. Picking a relatively low period
+			// here (0xff) for consistency with the other timers.
+			top = 0xff
+		} else {
+			// The formula below calculates the following formula, optimized:
+			//     top = period * (CPUFrequency() / 1e9)
+			// By dividing the CPU frequency first (an operation that is easily
+			// optimized away) the period has less chance of overflowing.
+			top = config.Period * (uint64(CPUFrequency()) / 1000000) / 1000
+		}
+
+		avr.TCCR1A.Set(avr.TCCR1A_WGM11)
+
+		// The ideal PWM period may be larger than would fit in the PWM counter,
+		// which is 16 bits (see maxTop). Therefore, try to make the PWM clock
+		// speed lower with a prescaler to make the top value fit the maximum
+		// top value.
+		const maxTop = 0x10000
+		switch {
+		case top <= maxTop:
+			avr.TCCR1B.Set(3<<3 | 1) // no prescaling
+		case top/8 <= maxTop:
+			avr.TCCR1B.Set(3<<3 | 2) // divide by 8
+			top /= 8
+		case top/64 <= maxTop:
+			avr.TCCR1B.Set(3<<3 | 3) // divide by 64
+			top /= 64
+		case top/256 <= maxTop:
+			avr.TCCR1B.Set(3<<3 | 4) // divide by 256
+			top /= 256
+		case top/1024 <= maxTop:
+			avr.TCCR1B.Set(3<<3 | 5) // divide by 1024
+			top /= 1024
+		default:
+			return ErrPWMPeriodTooLong
+		}
+
+		// A top of 0x10000 is at 100% duty cycle. Subtract one because the
+		// counter counts from 0, not 1 (avoiding an off-by-one).
+		top -= 1
+
+		avr.ICR1H.Set(uint8(top >> 8))
+		avr.ICR1L.Set(uint8(top))
 	}
 	return nil
 }
 
-// Set turns on the duty cycle for a PWM pin using the provided value. On the AVR this is normally a
+// SetPeriod updates the period of this PWM peripheral.
-// 8-bit value ranging from 0 to 255.
+// To set a particular frequency, use the following formula:
-func (pwm PWM) Set(value uint16) {
+//
-	value8 := uint8(value >> 8)
+//     period = 1e9 / frequency
-	switch pwm.Pin {
+//
-	case PD3:
+// If you use a period of 0, a period that works well for LEDs will be picked.
-		// connect pwm to pin on timer 2, channel B
+//
-		avr.TCCR2A.SetBits(avr.TCCR2A_COM2B1)
+// SetPeriod will not change the prescaler, but also won't change the current
-		avr.OCR2B.Set(value8) // set pwm duty
+// value in any of the channels. This means that you may need to update the
-	case PD5:
+// value for the particular channel.
-		// connect pwm to pin on timer 0, channel B
+//
-		avr.TCCR0A.SetBits(avr.TCCR0A_COM0B1)
+// Note that you cannot pick any arbitrary period after the PWM peripheral has
-		avr.OCR0B.Set(value8) // set pwm duty
+// been configured. If you want to switch between frequencies, pick the lowest
-	case PD6:
+// frequency (longest period) once when calling Configure and adjust the
-		// connect pwm to pin on timer 0, channel A
+// frequency here as needed.
-		avr.TCCR0A.SetBits(avr.TCCR0A_COM0A1)
+func (pwm PWM) SetPeriod(period uint64) error {
-		avr.OCR0A.Set(value8) // set pwm duty
+	if pwm.num != 1 {
-	case PB1:
+		return ErrPWMPeriodTooLong // TODO better error message
-		// connect pwm to pin on timer 1, channel A
+	}
-		avr.TCCR1A.SetBits(avr.TCCR1A_COM1A1)
+
-		// this is a 16-bit value, but we only currently allow the low order bits to be set
+	// The top value is the number of PWM ticks a PWM period takes. It is
-		avr.OCR1AL.Set(value8) // set pwm duty
+	// initially picked assuming an unlimited counter top and no PWM
-	case PB2:
+	// prescaler.
-		// connect pwm to pin on timer 1, channel B
+	var top uint64
-		avr.TCCR1A.SetBits(avr.TCCR1A_COM1B1)
+	if period == 0 {
-		// this is a 16-bit value, but we only currently allow the low order bits to be set
+		// Use a top appropriate for LEDs. Picking a relatively low period
-		avr.OCR1BL.Set(value8) // set pwm duty
+		// here (0xff) for consistency with the other timers.
-	case PB3:
+		top = 0xff
-		// connect pwm to pin on timer 2, channel A
+	} else {
-		avr.TCCR2A.SetBits(avr.TCCR2A_COM2A1)
+		// The formula below calculates the following formula, optimized:
-		avr.OCR2A.Set(value8) // set pwm duty
+		//     top = period * (CPUFrequency() / 1e9)
-	default:
+		// By dividing the CPU frequency first (an operation that is easily
-		panic("Invalid PWM pin")
+		// optimized away) the period has less chance of overflowing.
+		top = period * (uint64(CPUFrequency()) / 1000000) / 1000
+	}
+
+	prescaler := avr.TCCR1B.Get() & 0x7
+	switch prescaler {
+	case 1:
+		top /= 1
+	case 2:
+		top /= 8
+	case 3:
+		top /= 64
+	case 4:
+		top /= 256
+	case 5:
+		top /= 1024
+	}
+
+	// A top of 0x10000 is at 100% duty cycle. Subtract one because the counter
+	// counts from 0, not 1 (avoiding an off-by-one).
+	top -= 1
+
+	if top > 0xffff {
+		return ErrPWMPeriodTooLong
+	}
+
+	// Warning: this change is not atomic!
+	avr.ICR1H.Set(uint8(top >> 8))
+	avr.ICR1L.Set(uint8(top))
+
+	// ... and because of that, set the counter back to zero to avoid most of
+	// the effects of this non-atomicity.
+	avr.TCNT1H.Set(0)
+	avr.TCNT1L.Set(0)
+
+	return nil
+}
+
+// Top returns the current counter top, for use in duty cycle calculation. It
+// will only change with a call to Configure or SetPeriod, otherwise it is
+// constant.
+//
+// The value returned here is hardware dependent. In general, it's best to treat
+// it as an opaque value that can be divided by some number and passed to Set
+// (see Set documentation for more information).
+func (pwm PWM) Top() uint32 {
+	if pwm.num == 1 {
+		// Timer 1 has a configurable top value.
+		low := avr.ICR1L.Get()
+		high := avr.ICR1H.Get()
+		return uint32(high)<<8 | uint32(low) + 1
+	}
+	// Other timers go from 0 to 0xff (0x100 or 256 in total).
+	return 256
+}
+
+// Counter returns the current counter value of the timer in this PWM
+// peripheral. It may be useful for debugging.
+func (pwm PWM) Counter() uint32 {
+	switch pwm.num {
+	case 0:
+		return uint32(avr.TCNT0.Get())
+	case 1:
+		mask := interrupt.Disable()
+		low := avr.TCNT1L.Get()
+		high := avr.TCNT1H.Get()
+		interrupt.Restore(mask)
+		return uint32(high)<<8 | uint32(low)
+	case 2:
+		return uint32(avr.TCNT2.Get())
+	}
+	// Unknown PWM.
+	return 0
+}
+
+// Period returns the used PWM period in nanoseconds. It might deviate slightly
+// from the configured period due to rounding.
+func (pwm PWM) Period() uint64 {
+	var prescaler uint8
+	switch pwm.num {
+	case 0:
+		prescaler = avr.TCCR0B.Get() & 0x7
+	case 1:
+		prescaler = avr.TCCR1B.Get() & 0x7
+	case 2:
+		prescaler = avr.TCCR2B.Get() & 0x7
+	}
+	top := uint64(pwm.Top())
+	switch prescaler {
+	case 1: // prescaler 1
+		return 1 * top * 1000 / uint64(CPUFrequency()/1e6)
+	case 2: // prescaler 8
+		return 8 * top * 1000 / uint64(CPUFrequency()/1e6)
+	case 3: // prescaler 64
+		return 64 * top * 1000 / uint64(CPUFrequency()/1e6)
+	case 4: // prescaler 256
+		return 256 * top * 1000 / uint64(CPUFrequency()/1e6)
+	case 5: // prescaler 1024
+		return 1024 * top * 1000 / uint64(CPUFrequency()/1e6)
+	default: // unknown clock source
+		return 0
+	}
+}
+
+// Channel returns a PWM channel for the given pin.
+func (pwm PWM) Channel(pin Pin) (uint8, error) {
+	pin.Configure(PinConfig{Mode: PinOutput})
+	pin.Low()
+	switch pwm.num {
+	case 0:
+		switch pin {
+		case PD6: // channel A
+			avr.TCCR0A.SetBits(avr.TCCR0A_COM0A1)
+			return 0, nil
+		case PD5: // channel B
+			avr.TCCR0A.SetBits(avr.TCCR0A_COM0B1)
+			return 1, nil
+		}
+	case 1:
+		switch pin {
+		case PB1: // channel A
+			avr.TCCR1A.SetBits(avr.TCCR1A_COM1A1)
+			return 0, nil
+		case PB2: // channel B
+			avr.TCCR1A.SetBits(avr.TCCR1A_COM1B1)
+			return 1, nil
+		}
+	case 2:
+		switch pin {
+		case PB3: // channel A
+			avr.TCCR2A.SetBits(avr.TCCR2A_COM2A1)
+			return 0, nil
+		case PD3: // channel B
+			avr.TCCR2A.SetBits(avr.TCCR2A_COM2B1)
+			return 1, nil
+		}
+	}
+	return 0, ErrInvalidOutputPin
+}
+
+// SetInverting sets whether to invert the output of this channel.
+// Without inverting, a 25% duty cycle would mean the output is high for 25% of
+// the time and low for the rest. Inverting flips the output as if a NOT gate
+// was placed at the output, meaning that the output would be 25% low and 75%
+// high with a duty cycle of 25%.
+//
+// Note: the invert state may not be applied on the AVR until the next call to
+// ch.Set().
+func (pwm PWM) SetInverting(channel uint8, inverting bool) {
+	switch pwm.num {
+	case 0:
+		switch channel {
+		case 0: // channel A
+			if inverting {
+				avr.PORTB.SetBits(1 << 6) // PB6 high
+				avr.TCCR0A.SetBits(avr.TCCR0A_COM0A0)
+			} else {
+				avr.PORTB.ClearBits(1 << 6) // PB6 low
+				avr.TCCR0A.ClearBits(avr.TCCR0A_COM0A0)
+			}
+		case 1: // channel B
+			if inverting {
+				avr.PORTB.SetBits(1 << 5) // PB5 high
+				avr.TCCR0A.SetBits(avr.TCCR0A_COM0B0)
+			} else {
+				avr.PORTB.ClearBits(1 << 5) // PB5 low
+				avr.TCCR0A.ClearBits(avr.TCCR0A_COM0B0)
+			}
+		}
+	case 1:
+		// Note: the COM1A0/COM1B0 bit is not set with the configuration below.
+		// It will be set the following call to Set(), however.
+		switch channel {
+		case 0: // channel A, PB1
+			if inverting {
+				avr.PORTB.SetBits(1 << 1) // PB1 high
+			} else {
+				avr.PORTB.ClearBits(1 << 1) // PB1 low
+			}
+		case 1: // channel B, PB2
+			if inverting {
+				avr.PORTB.SetBits(1 << 2) // PB2 high
+			} else {
+				avr.PORTB.ClearBits(1 << 2) // PB2 low
+			}
+		}
+	case 2:
+		switch channel {
+		case 0: // channel A
+			if inverting {
+				avr.PORTB.SetBits(1 << 3) // PB3 high
+				avr.TCCR2A.SetBits(avr.TCCR2A_COM2A0)
+			} else {
+				avr.PORTB.ClearBits(1 << 3) // PB3 low
+				avr.TCCR2A.ClearBits(avr.TCCR2A_COM2A0)
+			}
+		case 1: // channel B
+			if inverting {
+				avr.PORTD.SetBits(1 << 3) // PD3 high
+				avr.TCCR2A.SetBits(avr.TCCR2A_COM2B0)
+			} else {
+				avr.PORTD.ClearBits(1 << 3) // PD3 low
+				avr.TCCR2A.ClearBits(avr.TCCR2A_COM2B0)
+			}
+		}
+	}
+}
+
+// Set updates the channel value. This is used to control the channel duty
+// cycle, in other words the fraction of time the channel output is high (or low
+// when inverted). For example, to set it to a 25% duty cycle, use:
+//
+//     pwm.Set(channel, pwm.Top() / 4)
+//
+// pwm.Set(channel, 0) will set the output to low and pwm.Set(channel,
+// pwm.Top()) will set the output to high, assuming the output isn't inverted.
+func (pwm PWM) Set(channel uint8, value uint32) {
+	switch pwm.num {
+	case 0:
+		value := uint16(value)
+		switch channel {
+		case 0: // channel A
+			if value == 0 {
+				avr.TCCR0A.ClearBits(avr.TCCR0A_COM0A1)
+			} else {
+				avr.OCR0A.Set(uint8(value - 1))
+				avr.TCCR0A.SetBits(avr.TCCR0A_COM0A1)
+			}
+		case 1: // channel B
+			if value == 0 {
+				avr.TCCR0A.ClearBits(avr.TCCR0A_COM0B1)
+			} else {
+				avr.OCR0B.Set(uint8(value) - 1)
+				avr.TCCR0A.SetBits(avr.TCCR0A_COM0B1)
+			}
+		}
+	case 1:
+		mask := interrupt.Disable()
+		switch channel {
+		case 0: // channel A, PB1
+			if value == 0 {
+				avr.TCCR1A.ClearBits(avr.TCCR1A_COM1A1 | avr.TCCR1A_COM1A0)
+			} else {
+				value := uint16(value) - 1 // yes, this is safe (it relies on underflow)
+				avr.OCR1AH.Set(uint8(value >> 8))
+				avr.OCR1AL.Set(uint8(value))
+				if avr.PORTB.HasBits(1 << 1) { // is PB1 high?
+					// Yes, set the inverting bit.
+					avr.TCCR1A.SetBits(avr.TCCR1A_COM1A1 | avr.TCCR1A_COM1A0)
+				} else {
+					// No, output is non-inverting.
+					avr.TCCR1A.SetBits(avr.TCCR1A_COM1A1)
+				}
+			}
+		case 1: // channel B, PB2
+			if value == 0 {
+				avr.TCCR1A.ClearBits(avr.TCCR1A_COM1B1 | avr.TCCR1A_COM1B0)
+			} else {
+				value := uint16(value) - 1 // yes, this is safe (it relies on underflow)
+				avr.OCR1BH.Set(uint8(value >> 8))
+				avr.OCR1BL.Set(uint8(value))
+				if avr.PORTB.HasBits(1 << 2) { // is PB2 high?
+					// Yes, set the inverting bit.
+					avr.TCCR1A.SetBits(avr.TCCR1A_COM1B1 | avr.TCCR1A_COM1B0)
+				} else {
+					// No, output is non-inverting.
+					avr.TCCR1A.SetBits(avr.TCCR1A_COM1B1)
+				}
+			}
+		}
+		interrupt.Restore(mask)
+	case 2:
+		value := uint16(value)
+		switch channel {
+		case 0: // channel A
+			if value == 0 {
+				avr.TCCR2A.ClearBits(avr.TCCR2A_COM2A1)
+			} else {
+				avr.OCR2A.Set(uint8(value - 1))
+				avr.TCCR2A.SetBits(avr.TCCR2A_COM2A1)
+			}
+		case 1: // channel B
+			if value == 0 {
+				avr.TCCR2A.ClearBits(avr.TCCR2A_COM2B1)
+			} else {
+				avr.OCR2B.Set(uint8(value - 1))
+				avr.TCCR2A.SetBits(avr.TCCR2A_COM2B1)
+			}
+		}
 	}
 }
 

src/machine/machine_atsamd21.go

@@ -8,8 +8,10 @@
 package machine
 
 import (
+	"device"
 	"device/arm"
 	"device/sam"
+	"errors"
 	"runtime/interrupt"
 	"runtime/volatile"
 	"unsafe"
@@ -29,8 +31,8 @@ const (
 	PinInput         PinMode = 9
 	PinInputPullup   PinMode = 10
 	PinOutput        PinMode = 11
-	PinPWM           PinMode = PinTimer
+	PinTCC           PinMode = PinTimer
-	PinPWMAlt        PinMode = PinTimerAlt
+	PinTCCAlt        PinMode = PinTimerAlt
 	PinInputPulldown PinMode = 12
 )
 
@@ -667,7 +669,7 @@ const (
 const i2cTimeout = 1000
 
 // Configure is intended to setup the I2C interface.
-func (i2c I2C) Configure(config I2CConfig) error {
+func (i2c *I2C) Configure(config I2CConfig) error {
 	// Default I2C bus speed is 100 kHz.
 	if config.Frequency == 0 {
 		config.Frequency = TWI_FREQ_100KHZ
@@ -723,7 +725,7 @@ func (i2c I2C) Configure(config I2CConfig) error {
 }
 
 // SetBaudRate sets the communication speed for the I2C.
-func (i2c I2C) SetBaudRate(br uint32) {
+func (i2c *I2C) SetBaudRate(br uint32) {
 	// Synchronous arithmetic baudrate, via Arduino SAMD implementation:
 	// SystemCoreClock / ( 2 * baudrate) - 5 - (((SystemCoreClock / 1000000) * WIRE_RISE_TIME_NANOSECONDS) / (2 * 1000));
 	baud := CPUFrequency()/(2*br) - 5 - (((CPUFrequency() / 1000000) * riseTimeNanoseconds) / (2 * 1000))
@@ -733,7 +735,7 @@ func (i2c I2C) SetBaudRate(br uint32) {
 // Tx does a single I2C transaction at the specified address.
 // It clocks out the given address, writes the bytes in w, reads back len(r)
 // bytes and stores them in r, and generates a stop condition on the bus.
-func (i2c I2C) Tx(addr uint16, w, r []byte) error {
+func (i2c *I2C) Tx(addr uint16, w, r []byte) error {
 	var err error
 	if len(w) != 0 {
 		// send start/address for write
@@ -810,7 +812,7 @@ func (i2c I2C) Tx(addr uint16, w, r []byte) error {
 }
 
 // WriteByte writes a single byte to the I2C bus.
-func (i2c I2C) WriteByte(data byte) error {
+func (i2c *I2C) WriteByte(data byte) error {
 	// Send data byte
 	i2c.Bus.DATA.Set(data)
 
@@ -835,7 +837,7 @@ func (i2c I2C) WriteByte(data byte) error {
 }
 
 // sendAddress sends the address and start signal
-func (i2c I2C) sendAddress(address uint16, write bool) error {
+func (i2c *I2C) sendAddress(address uint16, write bool) error {
 	data := (address << 1)
 	if !write {
 		data |= 1 // set read flag
@@ -855,7 +857,7 @@ func (i2c I2C) sendAddress(address uint16, write bool) error {
 	return nil
 }
 
-func (i2c I2C) signalStop() error {
+func (i2c *I2C) signalStop() error {
 	i2c.Bus.CTRLB.SetBits(wireCmdStop << sam.SERCOM_I2CM_CTRLB_CMD_Pos) // Stop command
 	timeout := i2cTimeout
 	for i2c.Bus.SYNCBUSY.HasBits(sam.SERCOM_I2CM_SYNCBUSY_SYSOP) {
@@ -867,7 +869,7 @@ func (i2c I2C) signalStop() error {
 	return nil
 }
 
-func (i2c I2C) signalRead() error {
+func (i2c *I2C) signalRead() error {
 	i2c.Bus.CTRLB.SetBits(wireCmdRead << sam.SERCOM_I2CM_CTRLB_CMD_Pos) // Read command
 	timeout := i2cTimeout
 	for i2c.Bus.SYNCBUSY.HasBits(sam.SERCOM_I2CM_SYNCBUSY_SYSOP) {
@@ -879,7 +881,7 @@ func (i2c I2C) signalRead() error {
 	return nil
 }
 
-func (i2c I2C) readByte() byte {
+func (i2c *I2C) readByte() byte {
 	for !i2c.Bus.INTFLAG.HasBits(sam.SERCOM_I2CM_INTFLAG_SB) {
 	}
 	return byte(i2c.Bus.DATA.Get())
@@ -1265,201 +1267,527 @@ func (spi SPI) Transfer(w byte) (byte, error) {
 	return byte(spi.Bus.DATA.Get()), nil
 }
 
-// PWM
+var (
-const period = 0xFFFF
+	ErrTxInvalidSliceSize = errors.New("SPI write and read slices must be same size")
+)
 
-// InitPWM initializes the PWM interface.
+// Tx handles read/write operation for SPI interface. Since SPI is a syncronous write/read
-func InitPWM() {
+// interface, there must always be the same number of bytes written as bytes read.
-	// turn on timer clocks used for PWM
+// The Tx method knows about this, and offers a few different ways of calling it.
-	sam.PM.APBCMASK.SetBits(sam.PM_APBCMASK_TCC0_ | sam.PM_APBCMASK_TCC1_ | sam.PM_APBCMASK_TCC2_)
+//
+// This form sends the bytes in tx buffer, putting the resulting bytes read into the rx buffer.
+// Note that the tx and rx buffers must be the same size:
+//
+// 		spi.Tx(tx, rx)
+//
+// This form sends the tx buffer, ignoring the result. Useful for sending "commands" that return zeros
+// until all the bytes in the command packet have been received:
+//
+// 		spi.Tx(tx, nil)
+//
+// This form sends zeros, putting the result into the rx buffer. Good for reading a "result packet":
+//
+// 		spi.Tx(nil, rx)
+//
+func (spi SPI) Tx(w, r []byte) error {
+	if spi.Bus.BAUD.Get() == 0x00 {
+		// When the SPI Freq is 24MHz, special processing is performed to improve the speed.
 
-	// Use GCLK0 for TCC0/TCC1
+		switch {
-	sam.GCLK.CLKCTRL.Set((sam.GCLK_CLKCTRL_ID_TCC0_TCC1 << sam.GCLK_CLKCTRL_ID_Pos) |
+		case w == nil:
-		(sam.GCLK_CLKCTRL_GEN_GCLK0 << sam.GCLK_CLKCTRL_GEN_Pos) |
+			// read only, so write zero and read a result.
-		sam.GCLK_CLKCTRL_CLKEN)
+			spi.rx(r)
-	for sam.GCLK.STATUS.HasBits(sam.GCLK_STATUS_SYNCBUSY) {
+		case r == nil:
-	}
+			// write only
+			spi.tx24mhz(w)
 
-	// Use GCLK0 for TCC2/TC3
+		default:
-	sam.GCLK.CLKCTRL.Set((sam.GCLK_CLKCTRL_ID_TCC2_TC3 << sam.GCLK_CLKCTRL_ID_Pos) |
+			// write/read
-		(sam.GCLK_CLKCTRL_GEN_GCLK0 << sam.GCLK_CLKCTRL_GEN_Pos) |
+			if len(w) != len(r) {
-		sam.GCLK_CLKCTRL_CLKEN)
+				return ErrTxInvalidSliceSize
-	for sam.GCLK.STATUS.HasBits(sam.GCLK_STATUS_SYNCBUSY) {
+			}
-	}
-}
 
-// Configure configures a PWM pin for output.
+			spi.txrx24mhz(w, r)
-func (pwm PWM) Configure() error {
+		}
-	// figure out which TCCX timer for this pin
-	timer := pwm.getTimer()
-	if timer == nil {
-		return ErrInvalidOutputPin
-	}
 
-	// disable timer
-	timer.CTRLA.ClearBits(sam.TCC_CTRLA_ENABLE)
-	// Wait for synchronization
-	for timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_ENABLE) {
-	}
-
-	// Use "Normal PWM" (single-slope PWM)
-	timer.WAVE.SetBits(sam.TCC_WAVE_WAVEGEN_NPWM)
-	// Wait for synchronization
-	for timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_WAVE) {
-	}
-
-	// Set the period (the number to count to (TOP) before resetting timer)
-	//TCC0->PER.reg = period;
-	timer.PER.Set(period)
-	// Wait for synchronization
-	for timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_PER) {
-	}
-
-	// Set pin as output
-	sam.PORT.DIRSET0.Set(1 << uint8(pwm.Pin))
-	// Set pin to low
-	sam.PORT.OUTCLR0.Set(1 << uint8(pwm.Pin))
-
-	// Enable the port multiplexer for pin
-	pwm.setPinCfg(sam.PORT_PINCFG0_PMUXEN)
-
-	// Connect TCCX timer to pin.
-	// we normally use the F channel aka ALT
-	pwmConfig := PinPWMAlt
-
-	// in the case of PA6 or PA7 we have to use E channel
-	if pwm.Pin == 6 || pwm.Pin == 7 {
-		pwmConfig = PinPWM
-	}
-
-	if pwm.Pin&1 > 0 {
-		// odd pin, so save the even pins
-		val := pwm.getPMux() & sam.PORT_PMUX0_PMUXE_Msk
-		pwm.setPMux(val | uint8(pwmConfig<<sam.PORT_PMUX0_PMUXO_Pos))
 	} else {
-		// even pin, so save the odd pins
+		switch {
-		val := pwm.getPMux() & sam.PORT_PMUX0_PMUXO_Msk
+		case w == nil:
-		pwm.setPMux(val | uint8(pwmConfig<<sam.PORT_PMUX0_PMUXE_Pos))
+			// read only, so write zero and read a result.
+			spi.rx(r)
+		case r == nil:
+			// write only
+			spi.tx(w)
+
+		default:
+			// write/read
+			if len(w) != len(r) {
+				return ErrTxInvalidSliceSize
+			}
+
+			spi.txrx(w, r)
+		}
 	}
 
 	return nil
 }
 
-// Set turns on the duty cycle for a PWM pin using the provided value.
+func (spi SPI) tx(tx []byte) {
-func (pwm PWM) Set(value uint16) {
+	for i := 0; i < len(tx); i++ {
-	// figure out which TCCX timer for this pin
+		for !spi.Bus.INTFLAG.HasBits(sam.SERCOM_SPI_INTFLAG_DRE) {
-	timer := pwm.getTimer()
+		}
-	if timer == nil {
+		spi.Bus.DATA.Set(uint32(tx[i]))
-		// The Configure call above cannot have succeeded, so simply ignore this
+	}
-		// error.
+	for !spi.Bus.INTFLAG.HasBits(sam.SERCOM_SPI_INTFLAG_TXC) {
-		return
 	}
 
-	// disable output
+	// read to clear RXC register
-	timer.CTRLA.ClearBits(sam.TCC_CTRLA_ENABLE)
+	for spi.Bus.INTFLAG.HasBits(sam.SERCOM_SPI_INTFLAG_RXC) {
+		spi.Bus.DATA.Get()
+	}
+}
 
-	// Wait for synchronization
+func (spi SPI) rx(rx []byte) {
-	for timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_ENABLE) {
+	spi.Bus.DATA.Set(0)
+	for !spi.Bus.INTFLAG.HasBits(sam.SERCOM_SPI_INTFLAG_DRE) {
 	}
 
+	for i := 1; i < len(rx); i++ {
+		spi.Bus.DATA.Set(0)
+		for !spi.Bus.INTFLAG.HasBits(sam.SERCOM_SPI_INTFLAG_RXC) {
+		}
+		rx[i-1] = byte(spi.Bus.DATA.Get())
+	}
+	for !spi.Bus.INTFLAG.HasBits(sam.SERCOM_SPI_INTFLAG_RXC) {
+	}
+	rx[len(rx)-1] = byte(spi.Bus.DATA.Get())
+}
+
+func (spi SPI) txrx(tx, rx []byte) {
+	spi.Bus.DATA.Set(uint32(tx[0]))
+	for !spi.Bus.INTFLAG.HasBits(sam.SERCOM_SPI_INTFLAG_DRE) {
+	}
+
+	for i := 1; i < len(rx); i++ {
+		spi.Bus.DATA.Set(uint32(tx[i]))
+		for !spi.Bus.INTFLAG.HasBits(sam.SERCOM_SPI_INTFLAG_RXC) {
+		}
+		rx[i-1] = byte(spi.Bus.DATA.Get())
+	}
+	for !spi.Bus.INTFLAG.HasBits(sam.SERCOM_SPI_INTFLAG_RXC) {
+	}
+	rx[len(rx)-1] = byte(spi.Bus.DATA.Get())
+}
+
+// tx24mhz is a special tx/rx function for CPU Clock 48 Mhz and SPI Freq 24 Mhz
+func (spi SPI) tx24mhz(tx []byte) {
+	spi.Bus.DATA.Set(uint32(tx[0]))
+	device.Asm("nop")
+	device.Asm("nop")
+	device.Asm("nop")
+	device.Asm("nop")
+	device.Asm("nop")
+	device.Asm("nop")
+
+	for i := 1; i < len(tx); i++ {
+		spi.Bus.DATA.Set(uint32(tx[i]))
+		device.Asm("nop")
+		device.Asm("nop")
+		spi.Bus.DATA.Get()
+	}
+	for !spi.Bus.INTFLAG.HasBits(sam.SERCOM_SPI_INTFLAG_RXC) {
+	}
+	spi.Bus.DATA.Get()
+}
+
+// txrx24mhz is a special tx/rx function for CPU Clock 48 Mhz and SPI Freq 24 Mhz
+func (spi SPI) txrx24mhz(tx, rx []byte) {
+	spi.Bus.DATA.Set(uint32(tx[0]))
+	device.Asm("nop")
+	device.Asm("nop")
+	device.Asm("nop")
+	device.Asm("nop")
+	device.Asm("nop")
+	device.Asm("nop")
+
+	for i := 1; i < len(rx); i++ {
+		spi.Bus.DATA.Set(uint32(tx[i]))
+		device.Asm("nop")
+		device.Asm("nop")
+		rx[i-1] = byte(spi.Bus.DATA.Get())
+	}
+	for !spi.Bus.INTFLAG.HasBits(sam.SERCOM_SPI_INTFLAG_RXC) {
+	}
+	rx[len(rx)-1] = byte(spi.Bus.DATA.Get())
+}
+
+// TCC is one timer/counter peripheral, which consists of a counter and multiple
+// output channels (that can be connected to actual pins). You can set the
+// frequency using SetPeriod, but only for all the channels in this TCC
+// peripheral at once.
+type TCC sam.TCC_Type
+
+// The SAM D21 has three TCC peripherals, which have PWM as one feature.
+var (
+	TCC0 = (*TCC)(sam.TCC0)
+	TCC1 = (*TCC)(sam.TCC1)
+	TCC2 = (*TCC)(sam.TCC2)
+)
+
+//go:inline
+func (tcc *TCC) timer() *sam.TCC_Type {
+	return (*sam.TCC_Type)(tcc)
+}
+
+// Configure enables and configures this TCC.
+func (tcc *TCC) Configure(config PWMConfig) error {
+	// Enable the clock source for this timer.
+	switch tcc.timer() {
+	case sam.TCC0:
+		sam.PM.APBCMASK.SetBits(sam.PM_APBCMASK_TCC0_)
+		// Use GCLK0 for TCC0/TCC1
+		sam.GCLK.CLKCTRL.Set((sam.GCLK_CLKCTRL_ID_TCC0_TCC1 << sam.GCLK_CLKCTRL_ID_Pos) |
+			(sam.GCLK_CLKCTRL_GEN_GCLK0 << sam.GCLK_CLKCTRL_GEN_Pos) |
+			sam.GCLK_CLKCTRL_CLKEN)
+		for sam.GCLK.STATUS.HasBits(sam.GCLK_STATUS_SYNCBUSY) {
+		}
+	case sam.TCC1:
+		sam.PM.APBCMASK.SetBits(sam.PM_APBCMASK_TCC1_)
+		// Use GCLK0 for TCC0/TCC1
+		sam.GCLK.CLKCTRL.Set((sam.GCLK_CLKCTRL_ID_TCC0_TCC1 << sam.GCLK_CLKCTRL_ID_Pos) |
+			(sam.GCLK_CLKCTRL_GEN_GCLK0 << sam.GCLK_CLKCTRL_GEN_Pos) |
+			sam.GCLK_CLKCTRL_CLKEN)
+		for sam.GCLK.STATUS.HasBits(sam.GCLK_STATUS_SYNCBUSY) {
+		}
+	case sam.TCC2:
+		sam.PM.APBCMASK.SetBits(sam.PM_APBCMASK_TCC2_)
+		// Use GCLK0 for TCC2/TC3
+		sam.GCLK.CLKCTRL.Set((sam.GCLK_CLKCTRL_ID_TCC2_TC3 << sam.GCLK_CLKCTRL_ID_Pos) |
+			(sam.GCLK_CLKCTRL_GEN_GCLK0 << sam.GCLK_CLKCTRL_GEN_Pos) |
+			sam.GCLK_CLKCTRL_CLKEN)
+		for sam.GCLK.STATUS.HasBits(sam.GCLK_STATUS_SYNCBUSY) {
+		}
+	}
+
+	// Disable timer (if it was enabled). This is necessary because
+	// tcc.setPeriod may want to change the prescaler bits in CTRLA, which is
+	// only allowed when the TCC is disabled.
+	tcc.timer().CTRLA.ClearBits(sam.TCC_CTRLA_ENABLE)
+
+	// Use "Normal PWM" (single-slope PWM)
+	tcc.timer().WAVE.Set(sam.TCC_WAVE_WAVEGEN_NPWM)
+
+	// Wait for synchronization of all changed registers.
+	for tcc.timer().SYNCBUSY.Get() != 0 {
+	}
+
+	// Set the period and prescaler.
+	err := tcc.setPeriod(config.Period, true)
+
+	// Enable the timer.
+	tcc.timer().CTRLA.SetBits(sam.TCC_CTRLA_ENABLE)
+
+	// Wait for synchronization of all changed registers.
+	for tcc.timer().SYNCBUSY.Get() != 0 {
+	}
+
+	// Return any error that might have occured in the tcc.setPeriod call.
+	return err
+}
+
+// SetPeriod updates the period of this TCC peripheral.
+// To set a particular frequency, use the following formula:
+//
+//     period = 1e9 / frequency
+//
+// If you use a period of 0, a period that works well for LEDs will be picked.
+//
+// SetPeriod will not change the prescaler, but also won't change the current
+// value in any of the channels. This means that you may need to update the
+// value for the particular channel.
+//
+// Note that you cannot pick any arbitrary period after the TCC peripheral has
+// been configured. If you want to switch between frequencies, pick the lowest
+// frequency (longest period) once when calling Configure and adjust the
+// frequency here as needed.
+func (tcc *TCC) SetPeriod(period uint64) error {
+	err := tcc.setPeriod(period, false)
+	if err == nil {
+		if tcc.Counter() >= tcc.Top() {
+			// When setting the timer to a shorter period, there is a chance
+			// that it passes the counter value and thus goes all the way to MAX
+			// before wrapping back to zero.
+			// To avoid this, reset the counter back to 0.
+			tcc.timer().COUNT.Set(0)
+		}
+	}
+	return err
+}
+
+// setPeriod sets the period of this TCC, possibly updating the prescaler as
+// well. The prescaler can only modified when the TCC is disabled, that is, in
+// the Configure function.
+func (tcc *TCC) setPeriod(period uint64, updatePrescaler bool) error {
+	var top uint64
+	if period == 0 {
+		// Make sure the TOP value is at 0xffff (enough for a 16-bit timer).
+		top = 0xffff
+	} else {
+		// The formula below calculates the following formula, optimized:
+		//     period * (48e6 / 1e9)
+		// This assumes that the chip is running at the (default) 48MHz speed.
+		top = period * 6 / 125
+	}
+
+	maxTop := uint64(0xffffff)
+	if tcc.timer() == sam.TCC2 {
+		// TCC2 is a 16-bit timer, not a 24-bit timer.
+		maxTop = 0xffff
+	}
+
+	if updatePrescaler {
+		// This function was called during Configure(), with the timer disabled.
+		// Note that updating the prescaler can only happen while the peripheral
+		// is disabled.
+		var prescaler uint32
+		switch {
+		case top <= maxTop:
+			prescaler = sam.TCC_CTRLA_PRESCALER_DIV1
+		case top/2 <= maxTop:
+			prescaler = sam.TCC_CTRLA_PRESCALER_DIV2
+			top = top / 2
+		case top/4 <= maxTop:
+			prescaler = sam.TCC_CTRLA_PRESCALER_DIV4
+			top = top / 4
+		case top/8 <= maxTop:
+			prescaler = sam.TCC_CTRLA_PRESCALER_DIV8
+			top = top / 8
+		case top/16 <= maxTop:
+			prescaler = sam.TCC_CTRLA_PRESCALER_DIV16
+			top = top / 16
+		case top/64 <= maxTop:
+			prescaler = sam.TCC_CTRLA_PRESCALER_DIV64
+			top = top / 64
+		case top/256 <= maxTop:
+			prescaler = sam.TCC_CTRLA_PRESCALER_DIV256
+			top = top / 256
+		case top/1024 <= maxTop:
+			prescaler = sam.TCC_CTRLA_PRESCALER_DIV1024
+			top = top / 1024
+		default:
+			return ErrPWMPeriodTooLong
+		}
+		tcc.timer().CTRLA.Set((tcc.timer().CTRLA.Get() &^ sam.TCC_CTRLA_PRESCALER_Msk) | (prescaler << sam.TCC_CTRLA_PRESCALER_Pos))
+	} else {
+		// Do not update the prescaler, but use the already-configured
+		// prescaler. This is the normal SetPeriod case, where the prescaler
+		// must not be changed.
+		prescaler := (tcc.timer().CTRLA.Get() & sam.TCC_CTRLA_PRESCALER_Msk) >> sam.TCC_CTRLA_PRESCALER_Pos
+		switch prescaler {
+		case sam.TCC_CTRLA_PRESCALER_DIV1:
+			top /= 1 // no-op
+		case sam.TCC_CTRLA_PRESCALER_DIV2:
+			top /= 2
+		case sam.TCC_CTRLA_PRESCALER_DIV4:
+			top /= 4
+		case sam.TCC_CTRLA_PRESCALER_DIV8:
+			top /= 8
+		case sam.TCC_CTRLA_PRESCALER_DIV16:
+			top /= 16
+		case sam.TCC_CTRLA_PRESCALER_DIV64:
+			top /= 64
+		case sam.TCC_CTRLA_PRESCALER_DIV256:
+			top /= 256
+		case sam.TCC_CTRLA_PRESCALER_DIV1024:
+			top /= 1024
+		default:
+			// unreachable
+		}
+		if top > maxTop {
+			return ErrPWMPeriodTooLong
+		}
+	}
+
+	// Set the period (the counter top).
+	tcc.timer().PER.Set(uint32(top) - 1)
+
+	// Wait for synchronization of CTRLA.PRESCALER and PER registers.
+	for tcc.timer().SYNCBUSY.Get() != 0 {
+	}
+
+	return nil
+}
+
+// Top returns the current counter top, for use in duty cycle calculation. It
+// will only change with a call to Configure or SetPeriod, otherwise it is
+// constant.
+//
+// The value returned here is hardware dependent. In general, it's best to treat
+// it as an opaque value that can be divided by some number and passed to Set
+// (see Set documentation for more information).
+func (tcc *TCC) Top() uint32 {
+	return tcc.timer().PER.Get() + 1
+}
+
+// Counter returns the current counter value of the timer in this TCC
+// peripheral. It may be useful for debugging.
+func (tcc *TCC) Counter() uint32 {
+	tcc.timer().CTRLBSET.Set(sam.TCC_CTRLBSET_CMD_READSYNC << sam.TCC_CTRLBSET_CMD_Pos)
+	for tcc.timer().SYNCBUSY.Get() != 0 {
+	}
+	return tcc.timer().COUNT.Get()
+}
+
+// Some constans to make pinTimerMapping below easier to read.
+const (
+	pinTCC0     = 1
+	pinTCC1     = 2
+	pinTCC2     = 3
+	pinTimerCh0 = 0 << 3
+	pinTimerCh2 = 1 << 3
+	pinTCC0Ch0  = pinTCC0 | pinTimerCh0
+	pinTCC0Ch2  = pinTCC0 | pinTimerCh2
+	pinTCC1Ch0  = pinTCC1 | pinTimerCh0
+	pinTCC1Ch2  = pinTCC1 | pinTimerCh2
+	pinTCC2Ch0  = pinTCC2 | pinTimerCh0
+)
+
+// Mapping from pin number to TCC peripheral and channel using a special
+// encoding. Note that only TCC0-TCC2 are included, not TC3 and up.
+// Every byte is split in two nibbles where the low nibble describes PinTCC and
+// the high nibble describes PinTCCAlt. Within a nibble, there is one bit that
+// indicates Ch0/Ch1 or Ch2/Ch3, and three other bits that contain the TCC
+// peripheral number plus one (to distinguish between TCC0Ch0 and 0).
+//
+// The encoding can be so compact because all pins are configured in pairs, so
+// if you know PA00 you can infer the configuration of PA01. And only channel 0
+// or 2 need to be included (taking up just one bit), because channel 0 and 2
+// are only ever used on odd pins and channel 1 and 3 on even pins, again using
+// the pin pair pattern to reduce the amount of information needed to be stored.
+//
+// Datasheet: https://cdn.sparkfun.com/datasheets/Dev/Arduino/Boards/Atmel-42181-SAM-D21_Datasheet.pdf
+var pinTimerMapping = [...]uint8{
+	// page 21
+	PA00 / 2: pinTCC2Ch0 | 0,
+	PA04 / 2: pinTCC0Ch0 | 0,
+	PA06 / 2: pinTCC1Ch0 | 0,
+	PA08 / 2: pinTCC0Ch0 | pinTCC1Ch2<<4,
+	PA10 / 2: pinTCC1Ch0 | pinTCC0Ch2<<4,
+	// page 22
+	PB10 / 2: 0 | pinTCC0Ch0<<4,
+	PB12 / 2: 0 | pinTCC0Ch2<<4,
+	PA12 / 2: pinTCC2Ch0 | pinTCC0Ch2<<4,
+	PA14 / 2: 0 | pinTCC0Ch0<<4,
+	PA16 / 2: pinTCC2Ch0 | pinTCC0Ch2<<4,
+	PA18 / 2: 0 | pinTCC0Ch2<<4,
+	PB16 / 2: 0 | pinTCC0Ch0<<4,
+	PA20 / 2: 0 | pinTCC0Ch2<<4,
+	PA22 / 2: 0 | pinTCC0Ch0<<4,
+	PA24 / 2: 0 | pinTCC1Ch2<<4,
+	// page 23
+	PA30 / 2: 0 | pinTCC1Ch0<<4,
+	PB30 / 2: pinTCC0Ch0 | pinTCC1Ch2<<4,
+}
+
+// findPinPadMapping returns the pin mode (PinTCC or PinTCCAlt) and the channel
+// number for a given timer and pin. A zero PinMode is returned if no mapping
+// could be found.
+func findPinTimerMapping(timer uint8, pin Pin) (PinMode, uint8) {
+	mapping := pinTimerMapping[pin/2]
+	// evenChannel below indicates the channel 0 or 2, for the even part of the
+	// pin pair. The next pin will also have the next channel (1 or 3).
+	if mapping&0x07 == timer+1 {
+		// PWM output is on peripheral function E.
+		evenChannel := ((mapping >> 3) & 1) * 2
+		return PinTCC, evenChannel + uint8(pin&1)
+	}
+	if (mapping&0x70)>>4 == timer+1 {
+		// PWM output is on peripheral function F.
+		evenChannel := ((mapping >> 7) & 1) * 2
+		return PinTCCAlt, evenChannel + uint8(pin&1)
+	}
+	return 0, 0
+}
+
+// Channel returns a PWM channel for the given pin. Note that one channel may be
+// shared between multiple pins, and so will have the same duty cycle. If this
+// is not desirable, look for a different TCC peripheral or consider using a
+// different pin.
+func (tcc *TCC) Channel(pin Pin) (uint8, error) {
+	var pinMode PinMode
+	var channel uint8
+	switch tcc.timer() {
+	case sam.TCC0:
+		pinMode, channel = findPinTimerMapping(0, pin)
+	case sam.TCC1:
+		pinMode, channel = findPinTimerMapping(1, pin)
+	case sam.TCC2:
+		pinMode, channel = findPinTimerMapping(2, pin)
+	}
+
+	if pinMode == 0 {
+		// No pin could be found.
+		return 0, ErrInvalidOutputPin
+	}
+
+	// Enable the port multiplexer for pin
+	pin.setPinCfg(sam.PORT_PINCFG0_PMUXEN)
+
+	if pin&1 > 0 {
+		// odd pin, so save the even pins
+		val := pin.getPMux() & sam.PORT_PMUX0_PMUXE_Msk
+		pin.setPMux(val | uint8(pinMode<<sam.PORT_PMUX0_PMUXO_Pos))
+	} else {
+		// even pin, so save the odd pins
+		val := pin.getPMux() & sam.PORT_PMUX0_PMUXO_Msk
+		pin.setPMux(val | uint8(pinMode<<sam.PORT_PMUX0_PMUXE_Pos))
+	}
+	return channel, nil
+}
+
+// SetInverting sets whether to invert the output of this channel.
+// Without inverting, a 25% duty cycle would mean the output is high for 25% of
+// the time and low for the rest. Inverting flips the output as if a NOT gate
+// was placed at the output, meaning that the output would be 25% low and 75%
+// high with a duty cycle of 25%.
+func (tcc *TCC) SetInverting(channel uint8, inverting bool) {
+	if inverting {
+		tcc.timer().WAVE.SetBits(1 << (sam.TCC_WAVE_POL0_Pos + channel))
+	} else {
+		tcc.timer().WAVE.ClearBits(1 << (sam.TCC_WAVE_POL0_Pos + channel))
+	}
+
+	// Wait for synchronization of the WAVE register.
+	for tcc.timer().SYNCBUSY.Get() != 0 {
+	}
+}
+
+// Set updates the channel value. This is used to control the channel duty
+// cycle, in other words the fraction of time the channel output is high (or low
+// when inverted). For example, to set it to a 25% duty cycle, use:
+//
+//     tcc.Set(channel, tcc.Top() / 4)
+//
+// tcc.Set(channel, 0) will set the output to low and tcc.Set(channel,
+// tcc.Top()) will set the output to high, assuming the output isn't inverted.
+func (tcc *TCC) Set(channel uint8, value uint32) {
 	// Set PWM signal to output duty cycle
-	pwm.setChannel(timer, uint32(value))
+	switch channel {
-
+	case 0:
-	// Wait for synchronization on all channels
+		tcc.timer().CC0.Set(value)
-	for timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_CC0 |
+	case 1:
-		sam.TCC_SYNCBUSY_CC1 |
+		tcc.timer().CC1.Set(value)
-		sam.TCC_SYNCBUSY_CC2 |
+	case 2:
-		sam.TCC_SYNCBUSY_CC3) {
+		tcc.timer().CC2.Set(value)
-	}
+	case 3:
-
+		tcc.timer().CC3.Set(value)
-	// enable
-	timer.CTRLA.SetBits(sam.TCC_CTRLA_ENABLE)
-	// Wait for synchronization
-	for timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_ENABLE) {
-	}
-}
-
-// getPMux returns the value for the correct PMUX register for this pin.
-func (pwm PWM) getPMux() uint8 {
-	return pwm.Pin.getPMux()
-}
-
-// setPMux sets the value for the correct PMUX register for this pin.
-func (pwm PWM) setPMux(val uint8) {
-	pwm.Pin.setPMux(val)
-}
-
-// getPinCfg returns the value for the correct PINCFG register for this pin.
-func (pwm PWM) getPinCfg() uint8 {
-	return pwm.Pin.getPinCfg()
-}
-
-// setPinCfg sets the value for the correct PINCFG register for this pin.
-func (pwm PWM) setPinCfg(val uint8) {
-	pwm.Pin.setPinCfg(val)
-}
-
-// getTimer returns the timer to be used for PWM on this pin
-func (pwm PWM) getTimer() *sam.TCC_Type {
-	switch pwm.Pin {
-	case 6:
-		return sam.TCC1
-	case 7:
-		return sam.TCC1
-	case 8:
-		return sam.TCC1
-	case 9:
-		return sam.TCC1
-	case 14:
-		return sam.TCC0
-	case 15:
-		return sam.TCC0
-	case 16:
-		return sam.TCC0
-	case 17:
-		return sam.TCC0
-	case 18:
-		return sam.TCC0
-	case 19:
-		return sam.TCC0
-	case 20:
-		return sam.TCC0
-	case 21:
-		return sam.TCC0
 	default:
-		return nil // not supported on this pin
+		// invalid PWM channel, ignore.
 	}
-}
 
-// setChannel sets the value for the correct channel for PWM on this pin
+	// Wait for synchronization on all channels (or anything in this peripheral,
-func (pwm PWM) setChannel(timer *sam.TCC_Type, val uint32) {
+	// really).
-	switch pwm.Pin {
+	for tcc.timer().SYNCBUSY.Get() != 0 {
-	case 6:
-		timer.CC0.Set(val)
-	case 7:
-		timer.CC1.Set(val)
-	case 8:
-		timer.CC0.Set(val)
-	case 9:
-		timer.CC1.Set(val)
-	case 14:
-		timer.CC0.Set(val)
-	case 15:
-		timer.CC1.Set(val)
-	case 16:
-		timer.CC2.Set(val)
-	case 17:
-		timer.CC3.Set(val)
-	case 18:
-		timer.CC2.Set(val)
-	case 19:
-		timer.CC3.Set(val)
-	case 20:
-		timer.CC2.Set(val)
-	case 21:
-		timer.CC3.Set(val)
-	default:
-		return // not supported on this pin
 	}
 }
 
@@ -1490,6 +1818,7 @@ func (usbcdc *USBCDC) Flush() error {
 
 			if usbcdc.waitTxc {
 				// waiting for the next flush(), because the transmission is not complete
+				usbcdc.waitTxcRetryCount++
 				return nil
 			}
 			usbcdc.waitTxc = true
@@ -1544,7 +1873,7 @@ func (usbcdc *USBCDC) WriteByte(c byte) error {
 				mask := interrupt.Disable()
 				UART0.waitTxc = false
 				UART0.waitTxcRetryCount = 0
-				usbcdc.TxIdx.Set(0)
+				UART0.TxIdx.Set(0)
 				usbLineInfo.lineState = 0
 				interrupt.Restore(mask)
 				break
@@ -1712,6 +2041,7 @@ func handleUSB(intr interrupt.Interrupt) {
 
 	// Start of frame
 	if (flags & sam.USB_DEVICE_INTFLAG_SOF) > 0 {
+		UART0.Flush()
 		// if you want to blink LED showing traffic, this would be the place...
 	}
 
@@ -1775,13 +2105,7 @@ func handleUSB(intr interrupt.Interrupt) {
 				}
 			}
 		}
-
-		if i == usb_CDC_ENDPOINT_IN && UART0.waitTxc {
-			UART0.waitTxcRetryCount++
-		}
 	}
-
-	UART0.Flush()
 }
 
 func initEndpoint(ep, config uint32) {

src/machine/machine_atsamd51.go

@@ -40,9 +40,9 @@ const (
 	PinInput         PinMode = 15
 	PinInputPullup   PinMode = 16
 	PinOutput        PinMode = 17
-	PinPWME          PinMode = PinTimer
+	PinTCCE          PinMode = PinTimer
-	PinPWMF          PinMode = PinTimerAlt
+	PinTCCF          PinMode = PinTimerAlt
-	PinPWMG          PinMode = PinTCCPDEC
+	PinTCCG          PinMode = PinTCCPDEC
 	PinInputPulldown PinMode = 18
 )
 
@@ -1104,7 +1104,7 @@ const (
 const i2cTimeout = 1000
 
 // Configure is intended to setup the I2C interface.
-func (i2c I2C) Configure(config I2CConfig) error {
+func (i2c *I2C) Configure(config I2CConfig) error {
 	// Default I2C bus speed is 100 kHz.
 	if config.Frequency == 0 {
 		config.Frequency = TWI_FREQ_100KHZ
@@ -1163,7 +1163,7 @@ func (i2c I2C) Configure(config I2CConfig) error {
 }
 
 // SetBaudRate sets the communication speed for the I2C.
-func (i2c I2C) SetBaudRate(br uint32) {
+func (i2c *I2C) SetBaudRate(br uint32) {
 	// Synchronous arithmetic baudrate, via Adafruit SAMD51 implementation:
 	// sercom->I2CM.BAUD.bit.BAUD = SERCOM_FREQ_REF / ( 2 * baudrate) - 1 ;
 	baud := SERCOM_FREQ_REF/(2*br) - 1
@@ -1173,7 +1173,7 @@ func (i2c I2C) SetBaudRate(br uint32) {
 // Tx does a single I2C transaction at the specified address.
 // It clocks out the given address, writes the bytes in w, reads back len(r)
 // bytes and stores them in r, and generates a stop condition on the bus.
-func (i2c I2C) Tx(addr uint16, w, r []byte) error {
+func (i2c *I2C) Tx(addr uint16, w, r []byte) error {
 	var err error
 	if len(w) != 0 {
 		// send start/address for write
@@ -1250,7 +1250,7 @@ func (i2c I2C) Tx(addr uint16, w, r []byte) error {
 }
 
 // WriteByte writes a single byte to the I2C bus.
-func (i2c I2C) WriteByte(data byte) error {
+func (i2c *I2C) WriteByte(data byte) error {
 	// Send data byte
 	i2c.Bus.DATA.Set(data)
 
@@ -1275,7 +1275,7 @@ func (i2c I2C) WriteByte(data byte) error {
 }
 
 // sendAddress sends the address and start signal
-func (i2c I2C) sendAddress(address uint16, write bool) error {
+func (i2c *I2C) sendAddress(address uint16, write bool) error {
 	data := (address << 1)
 	if !write {
 		data |= 1 // set read flag
@@ -1295,7 +1295,7 @@ func (i2c I2C) sendAddress(address uint16, write bool) error {
 	return nil
 }
 
-func (i2c I2C) signalStop() error {
+func (i2c *I2C) signalStop() error {
 	i2c.Bus.CTRLB.SetBits(wireCmdStop << sam.SERCOM_I2CM_CTRLB_CMD_Pos) // Stop command
 	timeout := i2cTimeout
 	for i2c.Bus.SYNCBUSY.HasBits(sam.SERCOM_I2CM_SYNCBUSY_SYSOP) {
@@ -1307,7 +1307,7 @@ func (i2c I2C) signalStop() error {
 	return nil
 }
 
-func (i2c I2C) signalRead() error {
+func (i2c *I2C) signalRead() error {
 	i2c.Bus.CTRLB.SetBits(wireCmdRead << sam.SERCOM_I2CM_CTRLB_CMD_Pos) // Read command
 	timeout := i2cTimeout
 	for i2c.Bus.SYNCBUSY.HasBits(sam.SERCOM_I2CM_SYNCBUSY_SYSOP) {
@@ -1319,7 +1319,7 @@ func (i2c I2C) signalRead() error {
 	return nil
 }
 
-func (i2c I2C) readByte() byte {
+func (i2c *I2C) readByte() byte {
 	for !i2c.Bus.INTFLAG.HasBits(sam.SERCOM_I2CM_INTFLAG_SB) {
 	}
 	return byte(i2c.Bus.DATA.Get())
@@ -1570,256 +1570,350 @@ const (
 	QSPI_DATA3 = PA11
 )
 
-// PWM
+// TCC is one timer peripheral, which consists of a counter and multiple output
-const period = 0xFFFF
+// channels (that can be connected to actual pins). You can set the frequency
+// using SetPeriod, but only for all the channels in this timer peripheral at
+// once.
+type TCC sam.TCC_Type
 
-// Configure configures a PWM pin for output.
+//go:inline
-func (pwm PWM) Configure() error {
+func (tcc *TCC) timer() *sam.TCC_Type {
-	// Set pin as output
+	return (*sam.TCC_Type)(tcc)
-	sam.PORT.GROUP[0].DIRSET.Set(1 << uint8(pwm.Pin))
+}
-	// Set pin to low
-	sam.PORT.GROUP[0].OUTCLR.Set(1 << uint8(pwm.Pin))
 
-	// Enable the port multiplexer for pin
+// Configure enables and configures this TCC.
-	pwm.setPinCfg(sam.PORT_GROUP_PINCFG_PMUXEN)
+func (tcc *TCC) Configure(config PWMConfig) error {
+	// Enable the TCC clock to be able to use the TCC.
+	tcc.configureClock()
 
-	// Connect timer/mux to pin.
+	// Disable timer (if it was enabled). This is necessary because
-	pwmConfig := pwm.getMux()
+	// tcc.setPeriod may want to change the prescaler bits in CTRLA, which is
-
+	// only allowed when the TCC is disabled.
-	if pwm.Pin&1 > 0 {
+	tcc.timer().CTRLA.ClearBits(sam.TCC_CTRLA_ENABLE)
-		// odd pin, so save the even pins
-		val := pwm.getPMux() & sam.PORT_GROUP_PMUX_PMUXE_Msk
-		pwm.setPMux(val | uint8(pwmConfig<<sam.PORT_GROUP_PMUX_PMUXO_Pos))
-	} else {
-		// even pin, so save the odd pins
-		val := pwm.getPMux() & sam.PORT_GROUP_PMUX_PMUXO_Msk
-		pwm.setPMux(val | uint8(pwmConfig<<sam.PORT_GROUP_PMUX_PMUXE_Pos))
-	}
-
-	// figure out which TCCX timer for this pin
-	timer := pwm.getTimer()
-	if timer == nil {
-		return ErrInvalidOutputPin
-	}
-
-	// disable timer
-	timer.CTRLA.ClearBits(sam.TCC_CTRLA_ENABLE)
-	// Wait for synchronization
-	for timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_ENABLE) {
-	}
-
-	// Set prescaler to 1/256
-	// TCCx->CTRLA.reg = TCC_CTRLA_PRESCALER_DIV256 | TCC_CTRLA_PRESCSYNC_GCLK;
-	timer.CTRLA.SetBits(sam.TCC_CTRLA_PRESCALER_DIV256 | sam.TCC_CTRLA_PRESCSYNC_GCLK)
 
 	// Use "Normal PWM" (single-slope PWM)
-	timer.WAVE.SetBits(sam.TCC_WAVE_WAVEGEN_NPWM)
+	tcc.timer().WAVE.Set(sam.TCC_WAVE_WAVEGEN_NPWM)
-	// Wait for synchronization
+
-	for timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_WAVE) {
+	// Wait for synchronization of all changed registers.
+	for tcc.timer().SYNCBUSY.Get() != 0 {
 	}
 
-	// while (TCCx->SYNCBUSY.bit.CC0 || TCCx->SYNCBUSY.bit.CC1);
+	// Set the period and prescaler.
-	for timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_CC0) ||
+	err := tcc.setPeriod(config.Period, true)
-		timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_CC1) {
+
+	// Enable the timer.
+	tcc.timer().CTRLA.SetBits(sam.TCC_CTRLA_ENABLE)
+
+	// Wait for synchronization of all changed registers.
+	for tcc.timer().SYNCBUSY.Get() != 0 {
 	}
 
-	// Set the initial value
+	// Return any error that might have occured in the tcc.setPeriod call.
-	// TCCx->CC[tcChannel].reg = (uint32_t) value;
+	return err
-	pwm.setChannel(timer, 0)
+}
 
-	for timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_CC0) ||
+// SetPeriod updates the period of this TCC peripheral.
-		timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_CC1) {
+// To set a particular frequency, use the following formula:
+//
+//     period = 1e9 / frequency
+//
+// If you use a period of 0, a period that works well for LEDs will be picked.
+//
+// SetPeriod will not change the prescaler, but also won't change the current
+// value in any of the channels. This means that you may need to update the
+// value for the particular channel.
+//
+// Note that you cannot pick any arbitrary period after the TCC peripheral has
+// been configured. If you want to switch between frequencies, pick the lowest
+// frequency (longest period) once when calling Configure and adjust the
+// frequency here as needed.
+func (tcc *TCC) SetPeriod(period uint64) error {
+	return tcc.setPeriod(period, false)
+}
+
+// setPeriod sets the period of this TCC, possibly updating the prescaler as
+// well. The prescaler can only modified when the TCC is disabled, that is, in
+// the Configure function.
+func (tcc *TCC) setPeriod(period uint64, updatePrescaler bool) error {
+	var top uint64
+	if period == 0 {
+		// Make sure the TOP value is at 0xffff (enough for a 16-bit timer).
+		top = 0xffff
+	} else {
+		// The formula below calculates the following formula, optimized:
+		//     period * (120e6 / 1e9)
+		// This assumes that the chip is running from generic clock generator 0
+		// at 120MHz.
+		top = period * 3 / 25
 	}
 
-	// Set the period (the number to count to (TOP) before resetting timer)
+	maxTop := uint64(0xffff)
-	//TCC0->PER.reg = period;
+	if tcc.timer() == sam.TCC0 || tcc.timer() == sam.TCC1 {
-	timer.PER.Set(period)
+		// Only TCC0 and TCC1 are 24-bit timers, the rest are 16-bit.
-	// Wait for synchronization
+		maxTop = 0xffffff
-	for timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_PER) {
 	}
 
-	// enable timer
+	if updatePrescaler {
-	timer.CTRLA.SetBits(sam.TCC_CTRLA_ENABLE)
+		// This function was called during Configure(), with the timer disabled.
-	// Wait for synchronization
+		// Note that updating the prescaler can only happen while the peripheral
-	for timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_ENABLE) {
+		// is disabled.
+		var prescaler uint32
+		switch {
+		case top <= maxTop:
+			prescaler = sam.TCC_CTRLA_PRESCALER_DIV1
+		case top/2 <= maxTop:
+			prescaler = sam.TCC_CTRLA_PRESCALER_DIV2
+			top = top / 2
+		case top/4 <= maxTop:
+			prescaler = sam.TCC_CTRLA_PRESCALER_DIV4
+			top = top / 4
+		case top/8 <= maxTop:
+			prescaler = sam.TCC_CTRLA_PRESCALER_DIV8
+			top = top / 8
+		case top/16 <= maxTop:
+			prescaler = sam.TCC_CTRLA_PRESCALER_DIV16
+			top = top / 16
+		case top/64 <= maxTop:
+			prescaler = sam.TCC_CTRLA_PRESCALER_DIV64
+			top = top / 64
+		case top/256 <= maxTop:
+			prescaler = sam.TCC_CTRLA_PRESCALER_DIV256
+			top = top / 256
+		case top/1024 <= maxTop:
+			prescaler = sam.TCC_CTRLA_PRESCALER_DIV1024
+			top = top / 1024
+		default:
+			return ErrPWMPeriodTooLong
+		}
+		tcc.timer().CTRLA.Set((tcc.timer().CTRLA.Get() &^ sam.TCC_CTRLA_PRESCALER_Msk) | (prescaler << sam.TCC_CTRLA_PRESCALER_Pos))
+	} else {
+		// Do not update the prescaler, but use the already-configured
+		// prescaler. This is the normal SetPeriod case, where the prescaler
+		// must not be changed.
+		prescaler := (tcc.timer().CTRLA.Get() & sam.TCC_CTRLA_PRESCALER_Msk) >> sam.TCC_CTRLA_PRESCALER_Pos
+		switch prescaler {
+		case sam.TCC_CTRLA_PRESCALER_DIV1:
+			top /= 1 // no-op
+		case sam.TCC_CTRLA_PRESCALER_DIV2:
+			top /= 2
+		case sam.TCC_CTRLA_PRESCALER_DIV4:
+			top /= 4
+		case sam.TCC_CTRLA_PRESCALER_DIV8:
+			top /= 8
+		case sam.TCC_CTRLA_PRESCALER_DIV16:
+			top /= 16
+		case sam.TCC_CTRLA_PRESCALER_DIV64:
+			top /= 64
+		case sam.TCC_CTRLA_PRESCALER_DIV256:
+			top /= 256
+		case sam.TCC_CTRLA_PRESCALER_DIV1024:
+			top /= 1024
+		default:
+			// unreachable
+		}
+		if top > maxTop {
+			return ErrPWMPeriodTooLong
+		}
+	}
+
+	// Set the period (the counter top).
+	tcc.timer().PER.Set(uint32(top) - 1)
+
+	// Wait for synchronization of CTRLA.PRESCALER and PER registers.
+	for tcc.timer().SYNCBUSY.Get() != 0 {
 	}
 
 	return nil
 }
 
-// Set turns on the duty cycle for a PWM pin using the provided value.
+// Top returns the current counter top, for use in duty cycle calculation. It
-func (pwm PWM) Set(value uint16) {
+// will only change with a call to Configure or SetPeriod, otherwise it is
-	// figure out which TCCX timer for this pin
+// constant.
-	timer := pwm.getTimer()
+//
-	if timer == nil {
+// The value returned here is hardware dependent. In general, it's best to treat
-		// The Configure call above cannot have succeeded, so simply ignore this
+// it as an opaque value that can be divided by some number and passed to
-		// error.
+// tcc.Set (see tcc.Set for more information).
-		return
+func (tcc *TCC) Top() uint32 {
+	return tcc.timer().PER.Get() + 1
+}
+
+// Counter returns the current counter value of the timer in this TCC
+// peripheral. It may be useful for debugging.
+func (tcc *TCC) Counter() uint32 {
+	tcc.timer().CTRLBSET.Set(sam.TCC_CTRLBSET_CMD_READSYNC << sam.TCC_CTRLBSET_CMD_Pos)
+	for tcc.timer().SYNCBUSY.Get() != 0 {
+	}
+	return tcc.timer().COUNT.Get()
+}
+
+// Constants that encode a TCC number and WO number together in a single byte.
+const (
+	pinTCC0   = 1 << 4 // keep the value 0 usable as "no value"
+	pinTCC1   = 2 << 4
+	pinTCC2   = 3 << 4
+	pinTCC3   = 4 << 4
+	pinTCC4   = 5 << 4
+	pinTCC0_0 = pinTCC0 | 0
+	pinTCC0_1 = pinTCC0 | 1
+	pinTCC0_2 = pinTCC0 | 2
+	pinTCC0_3 = pinTCC0 | 3
+	pinTCC0_4 = pinTCC0 | 4
+	pinTCC0_5 = pinTCC0 | 5
+	pinTCC0_6 = pinTCC0 | 6
+	pinTCC1_0 = pinTCC1 | 0
+	pinTCC1_2 = pinTCC1 | 2
+	pinTCC1_4 = pinTCC1 | 4
+	pinTCC1_6 = pinTCC1 | 6
+	pinTCC2_0 = pinTCC2 | 0
+	pinTCC2_2 = pinTCC2 | 2
+	pinTCC3_0 = pinTCC3 | 0
+	pinTCC4_0 = pinTCC4 | 0
+)
+
+// This is a copy of columns F and G (the TCC columns) of table 6-1 in the
+// datasheet:
+// http://ww1.microchip.com/downloads/en/DeviceDoc/60001507E.pdf
+// For example, "TCC0/WO[2]" is converted to pinTCC0_2.
+// Only the even pin numbers are stored here. The odd pin numbers are left out,
+// because their PWM output can be determined from the even number: just add one
+// to the wave output (WO) number.
+var pinTimerMapping = [...]struct{ F, G uint8 }{
+	// page 33
+	PC04 / 2: {pinTCC0_0, 0},
+	PA08 / 2: {pinTCC0_0, pinTCC1_4},
+	PA10 / 2: {pinTCC0_2, pinTCC1_6},
+	PB10 / 2: {pinTCC0_4, pinTCC1_0},
+	PB12 / 2: {pinTCC3_0, pinTCC0_0},
+	PB14 / 2: {pinTCC4_0, pinTCC0_2},
+	PD08 / 2: {pinTCC0_1, 0},
+	PD10 / 2: {pinTCC0_3, 0},
+	PD12 / 2: {pinTCC0_5, 0},
+	PC10 / 2: {pinTCC0_0, pinTCC1_4},
+	// page 34
+	PC12 / 2: {pinTCC0_2, pinTCC1_6},
+	PC14 / 2: {pinTCC0_4, pinTCC1_0},
+	PA12 / 2: {pinTCC0_6, pinTCC1_2},
+	PA14 / 2: {pinTCC2_0, pinTCC1_2},
+	PA16 / 2: {pinTCC1_0, pinTCC0_4},
+	PA18 / 2: {pinTCC1_2, pinTCC0_6},
+	PC16 / 2: {pinTCC0_0, 0},
+	PC18 / 2: {pinTCC0_2, 0},
+	PC20 / 2: {pinTCC0_4, 0},
+	PC22 / 2: {pinTCC0_6, 0},
+	PD20 / 2: {pinTCC1_0, 0},
+	PB16 / 2: {pinTCC3_0, pinTCC0_4},
+	PB18 / 2: {pinTCC1_0, 0},
+	// page 35
+	PB20 / 2: {pinTCC1_2, 0},
+	PA20 / 2: {pinTCC1_4, pinTCC0_0},
+	PA22 / 2: {pinTCC1_6, pinTCC0_2},
+	PA24 / 2: {pinTCC2_2, 0},
+	PB26 / 2: {pinTCC1_2, 0},
+	PB28 / 2: {pinTCC1_4, 0},
+	PA30 / 2: {pinTCC2_0, 0},
+	// page 36
+	PB30 / 2: {pinTCC4_0, pinTCC0_6},
+	PB02 / 2: {pinTCC2_2, 0},
+}
+
+// findPinPadMapping returns the pin mode (PinTCCF or PinTCCG) and the channel
+// number for a given timer and pin. A zero PinMode is returned if no mapping
+// could be found.
+func findPinTimerMapping(timer uint8, pin Pin) (PinMode, uint8) {
+	if int(pin/2) >= len(pinTimerMapping) {
+		return 0, 0 // invalid pin number
 	}
 
-	// Wait for synchronization
+	mapping := pinTimerMapping[pin/2]
-	for timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_CTRLB) {
+
-	}
+	// Check for column F in the datasheet.
-	for timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_CC0) ||
+	if mapping.F>>4-1 == timer {
-		timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_CC1) {
+		return PinTCCF, mapping.F&0x0f + uint8(pin)&1
 	}
 
-	// TCCx->CCBUF[tcChannel].reg = (uint32_t) value;
+	// Check for column G in the datasheet.
-	pwm.setChannelBuffer(timer, uint32(value))
+	if mapping.G>>4-1 == timer {
-
+		return PinTCCG, mapping.G&0x0f + uint8(pin)&1
-	for timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_CC0) ||
-		timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_CC1) {
 	}
 
-	// TCCx->CTRLBCLR.bit.LUPD = 1;
+	// Nothing found.
-	timer.CTRLBCLR.SetBits(sam.TCC_CTRLBCLR_LUPD)
+	return 0, 0
-	for timer.SYNCBUSY.HasBits(sam.TCC_SYNCBUSY_CTRLB) {
+}
+
+// Channel returns a PWM channel for the given pin. Note that one channel may be
+// shared between multiple pins, and so will have the same duty cycle. If this
+// is not desirable, look for a different TCC or consider using a different pin.
+func (tcc *TCC) Channel(pin Pin) (uint8, error) {
+	pinMode, woOutput := findPinTimerMapping(tcc.timerNum(), pin)
+
+	if pinMode == 0 {
+		// No pin could be found.
+		return 0, ErrInvalidOutputPin
+	}
+
+	// Convert from waveform output to channel, assuming WEXCTRL.OTMX equals 0.
+	// See table 49-4 "Output Matrix Channel Pin Routing Configuration" on page
+	// 1829 of the datasheet.
+	// The number of channels varies by TCC instance, hence the need to switch
+	// over them. For TCC2-4 the number of channels is equal to the number of
+	// waveform outputs, so the WO number maps directly to the channel number.
+	// For TCC0 and TCC1 this is not the case so they will need some special
+	// handling.
+	channel := woOutput
+	switch tcc.timer() {
+	case sam.TCC0:
+		channel = woOutput % 6
+	case sam.TCC1:
+		channel = woOutput % 4
+	}
+
+	// Enable the port multiplexer for pin
+	pin.setPinCfg(sam.PORT_GROUP_PINCFG_PMUXEN)
+
+	// Connect timer/mux to pin.
+	if pin&1 > 0 {
+		// odd pin, so save the even pins
+		val := pin.getPMux() & sam.PORT_GROUP_PMUX_PMUXE_Msk
+		pin.setPMux(val | uint8(pinMode<<sam.PORT_GROUP_PMUX_PMUXO_Pos))
+	} else {
+		// even pin, so save the odd pins
+		val := pin.getPMux() & sam.PORT_GROUP_PMUX_PMUXO_Msk
+		pin.setPMux(val | uint8(pinMode<<sam.PORT_GROUP_PMUX_PMUXE_Pos))
+	}
+
+	return channel, nil
+}
+
+// SetInverting sets whether to invert the output of this channel.
+// Without inverting, a 25% duty cycle would mean the output is high for 25% of
+// the time and low for the rest. Inverting flips the output as if a NOT gate
+// was placed at the output, meaning that the output would be 25% low and 75%
+// high with a duty cycle of 25%.
+func (tcc *TCC) SetInverting(channel uint8, inverting bool) {
+	if inverting {
+		tcc.timer().WAVE.SetBits(1 << (sam.TCC_WAVE_POL0_Pos + channel))
+	} else {
+		tcc.timer().WAVE.ClearBits(1 << (sam.TCC_WAVE_POL0_Pos + channel))
+	}
+
+	// Wait for synchronization of the WAVE register.
+	for tcc.timer().SYNCBUSY.Get() != 0 {
 	}
 }
 
-// getPMux returns the value for the correct PMUX register for this pin.
+// Set updates the channel value. This is used to control the channel duty
-func (pwm PWM) getPMux() uint8 {
+// cycle, in other words the fraction of time the channel output is high (or low
-	return pwm.Pin.getPMux()
+// when inverted). For example, to set it to a 25% duty cycle, use:
-}
+//
-
+//     tcc.Set(channel, tcc.Top() / 4)
-// setPMux sets the value for the correct PMUX register for this pin.
+//
-func (pwm PWM) setPMux(val uint8) {
+// tcc.Set(channel, 0) will set the output to low and tcc.Set(channel,
-	pwm.Pin.setPMux(val)
+// tcc.Top()) will set the output to high, assuming the output isn't inverted.
-}
+func (tcc *TCC) Set(channel uint8, value uint32) {
-
+	// Update CCBUF, which provides double buffering. The update is applied on
-// getPinCfg returns the value for the correct PINCFG register for this pin.
+	// the next cycle.
-func (pwm PWM) getPinCfg() uint8 {
+	tcc.timer().CCBUF[channel].Set(value)
-	return pwm.Pin.getPinCfg()
+	for tcc.timer().SYNCBUSY.Get() != 0 {
-}
-
-// setPinCfg sets the value for the correct PINCFG register for this pin.
-func (pwm PWM) setPinCfg(val uint8) {
-	pwm.Pin.setPinCfg(val)
-}
-
-// setChannel sets the value for the correct channel for PWM on this pin.
-func (pwm PWM) setChannel(timer *sam.TCC_Type, val uint32) {
-	switch pwm.Pin {
-	case PA14:
-		timer.CC[0].Set(val)
-	case PA15:
-		timer.CC[1].Set(val)
-	case PA16:
-		timer.CC[0].Set(val)
-	case PA17:
-		timer.CC[1].Set(val)
-	case PA18:
-		timer.CC[2].Set(val)
-	case PA19:
-		timer.CC[3].Set(val)
-	case PA20:
-		timer.CC[0].Set(val)
-	case PA21:
-		timer.CC[1].Set(val)
-	case PA22:
-		timer.CC[2].Set(val)
-	case PA23:
-		timer.CC[3].Set(val)
-	case PB12:
-		timer.CC[0].Set(val)
-	case PB13:
-		timer.CC[1].Set(val)
-	case PB14:
-		timer.CC[0].Set(val)
-	case PB15:
-		timer.CC[1].Set(val)
-	case PB16:
-		timer.CC[4].Set(val)
-	case PB17:
-		timer.CC[5].Set(val)
-	case PB31:
-		timer.CC[1].Set(val)
-	default:
-		return // not supported on this pin
-	}
-}
-
-// setChannelBuffer sets the value for the correct channel buffer for PWM on this pin
-func (pwm PWM) setChannelBuffer(timer *sam.TCC_Type, val uint32) {
-	switch pwm.Pin {
-	case PA14:
-		timer.CCBUF[0].Set(val)
-	case PA15:
-		timer.CCBUF[1].Set(val)
-	case PA16:
-		timer.CCBUF[0].Set(val)
-	case PA17:
-		timer.CCBUF[1].Set(val)
-	case PA18:
-		timer.CCBUF[2].Set(val)
-	case PA19:
-		timer.CCBUF[3].Set(val)
-	case PA20:
-		timer.CCBUF[0].Set(val)
-	case PA21:
-		timer.CCBUF[1].Set(val)
-	case PA22:
-		timer.CCBUF[2].Set(val)
-	case PA23:
-		timer.CCBUF[3].Set(val)
-	case PB12:
-		timer.CCBUF[0].Set(val)
-	case PB13:
-		timer.CCBUF[1].Set(val)
-	case PB14:
-		timer.CCBUF[0].Set(val)
-	case PB15:
-		timer.CCBUF[1].Set(val)
-	case PB16:
-		timer.CCBUF[4].Set(val)
-	case PB17:
-		timer.CCBUF[5].Set(val)
-	case PB31:
-		timer.CCBUF[1].Set(val)
-	default:
-		return // not supported on this pin
-	}
-}
-
-// getMux returns the pin mode mux to be used for PWM on this pin.
-func (pwm PWM) getMux() PinMode {
-	switch pwm.Pin {
-	case PA14:
-		return PinPWMF
-	case PA15:
-		return PinPWMF
-	case PA16:
-		return PinPWMF
-	case PA17:
-		return PinPWMF
-	case PA18:
-		return PinPWMF
-	case PA19:
-		return PinPWMF
-	case PA20:
-		return PinPWMG
-	case PA21:
-		return PinPWMG
-	case PA22:
-		return PinPWMG
-	case PA23:
-		return PinPWMG
-	case PB12:
-		return PinPWMF
-	case PB13:
-		return PinPWMF
-	case PB14:
-		return PinPWMF
-	case PB15:
-		return PinPWMF
-	case PB16:
-		return PinPWMG
-	case PB17:
-		return PinPWMG
-	case PB31:
-		return PinPWMF
-	default:
-		return 0 // not supported on this pin
 	}
 }
 
@@ -1850,6 +1944,7 @@ func (usbcdc *USBCDC) Flush() error {
 
 			if usbcdc.waitTxc {
 				// waiting for the next flush(), because the transmission is not complete
+				usbcdc.waitTxcRetryCount++
 				return nil
 			}
 			usbcdc.waitTxc = true
@@ -1904,7 +1999,7 @@ func (usbcdc *USBCDC) WriteByte(c byte) error {
 				mask := interrupt.Disable()
 				UART0.waitTxc = false
 				UART0.waitTxcRetryCount = 0
-				usbcdc.TxIdx.Set(0)
+				UART0.TxIdx.Set(0)
 				usbLineInfo.lineState = 0
 				interrupt.Restore(mask)
 				break
@@ -2074,6 +2169,7 @@ func handleUSBIRQ(interrupt.Interrupt) {
 
 	// Start of frame
 	if (flags & sam.USB_DEVICE_INTFLAG_SOF) > 0 {
+		UART0.Flush()
 		// if you want to blink LED showing traffic, this would be the place...
 	}
 
@@ -2137,13 +2233,7 @@ func handleUSBIRQ(interrupt.Interrupt) {
 				}
 			}
 		}
-
-		if i == usb_CDC_ENDPOINT_IN && UART0.waitTxc {
-			UART0.waitTxcRetryCount++
-		}
 	}
-
-	UART0.Flush()
 }
 
 func initEndpoint(ep, config uint32) {

src/machine/machine_atsamd51g19.go

@@ -11,43 +11,37 @@ import "device/sam"
 
 const HSRAM_SIZE = 0x00030000
 
-// InitPWM initializes the PWM interface.
+// This chip has three TCC peripherals, which have PWM as one feature.
-func InitPWM() {
+var (
-	// turn on timer clocks used for PWM
+	TCC0 = (*TCC)(sam.TCC0)
-	sam.MCLK.APBBMASK.SetBits(sam.MCLK_APBBMASK_TCC0_ | sam.MCLK_APBBMASK_TCC1_)
+	TCC1 = (*TCC)(sam.TCC1)
-	sam.MCLK.APBCMASK.SetBits(sam.MCLK_APBCMASK_TCC2_)
+	TCC2 = (*TCC)(sam.TCC2)
+)
 
-	//use clock generator 0
+func (tcc *TCC) configureClock() {
-	sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC0].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) |
+	// Turn on timer clocks used for TCC and use generic clock generator 0.
-		sam.GCLK_PCHCTRL_CHEN)
+	switch tcc.timer() {
-	sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC2].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) |
+	case sam.TCC0:
-		sam.GCLK_PCHCTRL_CHEN)
+		sam.MCLK.APBBMASK.SetBits(sam.MCLK_APBBMASK_TCC0_)
-}
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC0].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-
+	case sam.TCC1:
-// getTimer returns the timer to be used for PWM on this pin
+		sam.MCLK.APBBMASK.SetBits(sam.MCLK_APBBMASK_TCC1_)
-func (pwm PWM) getTimer() *sam.TCC_Type {
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC1].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-	switch pwm.Pin {
+	case sam.TCC2:
-	case PA14:
+		sam.MCLK.APBCMASK.SetBits(sam.MCLK_APBCMASK_TCC2_)
-		return sam.TCC2
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC2].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-	case PA15:
+	}
-		return sam.TCC2
+}
-	case PA16:
+
-		return sam.TCC1
+func (tcc *TCC) timerNum() uint8 {
-	case PA17:
+	switch tcc.timer() {
-		return sam.TCC1
+	case sam.TCC0:
-	case PA18:
+		return 0
-		return sam.TCC1
+	case sam.TCC1:
-	case PA19:
+		return 1
-		return sam.TCC1
+	case sam.TCC2:
-	case PA20:
+		return 2
-		return sam.TCC0
+	default:
-	case PA21:
+		return 0x0f // should not happen
-		return sam.TCC0
-	case PA22:
-		return sam.TCC0
-	case PA23:
-		return sam.TCC0
-	default:
-		return nil // not supported on this pin
 	}
 }

src/machine/machine_atsamd51j19.go

@@ -11,60 +11,49 @@ import "device/sam"
 
 const HSRAM_SIZE = 0x00030000
 
-// InitPWM initializes the PWM interface.
+// This chip has five TCC peripherals, which have PWM as one feature.
-func InitPWM() {
+var (
-	// turn on timer clocks used for PWM
+	TCC0 = (*TCC)(sam.TCC0)
-	sam.MCLK.APBBMASK.SetBits(sam.MCLK_APBBMASK_TCC0_ | sam.MCLK_APBBMASK_TCC1_)
+	TCC1 = (*TCC)(sam.TCC1)
-	sam.MCLK.APBCMASK.SetBits(sam.MCLK_APBCMASK_TCC2_ | sam.MCLK_APBCMASK_TCC3_)
+	TCC2 = (*TCC)(sam.TCC2)
-	sam.MCLK.APBDMASK.SetBits(sam.MCLK_APBDMASK_TCC4_)
+	TCC3 = (*TCC)(sam.TCC3)
+	TCC4 = (*TCC)(sam.TCC4)
+)
 
-	//use clock generator 0
+func (tcc *TCC) configureClock() {
-	sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC0].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) |
+	// Turn on timer clocks used for the TCC and use generic clock generator 0.
-		sam.GCLK_PCHCTRL_CHEN)
+	switch tcc.timer() {
-	sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC2].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) |
+	case sam.TCC0:
-		sam.GCLK_PCHCTRL_CHEN)
+		sam.MCLK.APBBMASK.SetBits(sam.MCLK_APBBMASK_TCC0_)
-	sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC4].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) |
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC0].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-		sam.GCLK_PCHCTRL_CHEN)
+	case sam.TCC1:
-}
+		sam.MCLK.APBBMASK.SetBits(sam.MCLK_APBBMASK_TCC1_)
-
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC1].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-// getTimer returns the timer to be used for PWM on this pin
+	case sam.TCC2:
-func (pwm PWM) getTimer() *sam.TCC_Type {
+		sam.MCLK.APBCMASK.SetBits(sam.MCLK_APBCMASK_TCC2_)
-	switch pwm.Pin {
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC2].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-	case PA14:
+	case sam.TCC3:
-		return sam.TCC2
+		sam.MCLK.APBCMASK.SetBits(sam.MCLK_APBCMASK_TCC3_)
-	case PA15:
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC3].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-		return sam.TCC2
+	case sam.TCC4:
-	case PA16:
+		sam.MCLK.APBDMASK.SetBits(sam.MCLK_APBDMASK_TCC4_)
-		return sam.TCC1
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC4].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-	case PA17:
+	}
-		return sam.TCC1
+}
-	case PA18:
+
-		return sam.TCC1
+func (tcc *TCC) timerNum() uint8 {
-	case PA19:
+	switch tcc.timer() {
-		return sam.TCC1
+	case sam.TCC0:
-	case PA20:
+		return 0
-		return sam.TCC0
+	case sam.TCC1:
-	case PA21:
+		return 1
-		return sam.TCC0
+	case sam.TCC2:
-	case PA22:
+		return 2
-		return sam.TCC0
+	case sam.TCC3:
-	case PA23:
+		return 3
-		return sam.TCC0
+	case sam.TCC4:
-	case PB12:
+		return 4
-		return sam.TCC3
+	default:
-	case PB13:
+		return 0x0f // should not happen
-		return sam.TCC3
-	case PB14:
-		return sam.TCC4
-	case PB15:
-		return sam.TCC4
-	case PB16:
-		return sam.TCC0
-	case PB17:
-		return sam.TCC0
-	case PB31:
-		return sam.TCC4
-	default:
-		return nil // not supported on this pin
 	}
 }

src/machine/machine_atsamd51j20.go

@@ -11,60 +11,49 @@ import "device/sam"
 
 const HSRAM_SIZE = 0x00040000
 
-// InitPWM initializes the PWM interface.
+// This chip has five TCC peripherals, which have PWM as one feature.
-func InitPWM() {
+var (
-	// turn on timer clocks used for PWM
+	TCC0 = (*TCC)(sam.TCC0)
-	sam.MCLK.APBBMASK.SetBits(sam.MCLK_APBBMASK_TCC0_ | sam.MCLK_APBBMASK_TCC1_)
+	TCC1 = (*TCC)(sam.TCC1)
-	sam.MCLK.APBCMASK.SetBits(sam.MCLK_APBCMASK_TCC2_ | sam.MCLK_APBCMASK_TCC3_)
+	TCC2 = (*TCC)(sam.TCC2)
-	sam.MCLK.APBDMASK.SetBits(sam.MCLK_APBDMASK_TCC4_)
+	TCC3 = (*TCC)(sam.TCC3)
+	TCC4 = (*TCC)(sam.TCC4)
+)
 
-	//use clock generator 0
+func (tcc *TCC) configureClock() {
-	sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC0].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) |
+	// Turn on timer clocks used for TCC and use generic clock generator 0.
-		sam.GCLK_PCHCTRL_CHEN)
+	switch tcc.timer() {
-	sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC2].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) |
+	case sam.TCC0:
-		sam.GCLK_PCHCTRL_CHEN)
+		sam.MCLK.APBBMASK.SetBits(sam.MCLK_APBBMASK_TCC0_)
-	sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC4].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) |
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC0].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-		sam.GCLK_PCHCTRL_CHEN)
+	case sam.TCC1:
-}
+		sam.MCLK.APBBMASK.SetBits(sam.MCLK_APBBMASK_TCC1_)
-
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC1].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-// getTimer returns the timer to be used for PWM on this pin
+	case sam.TCC2:
-func (pwm PWM) getTimer() *sam.TCC_Type {
+		sam.MCLK.APBCMASK.SetBits(sam.MCLK_APBCMASK_TCC2_)
-	switch pwm.Pin {
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC2].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-	case PA14:
+	case sam.TCC3:
-		return sam.TCC2
+		sam.MCLK.APBCMASK.SetBits(sam.MCLK_APBCMASK_TCC3_)
-	case PA15:
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC3].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-		return sam.TCC2
+	case sam.TCC4:
-	case PA16:
+		sam.MCLK.APBDMASK.SetBits(sam.MCLK_APBDMASK_TCC4_)
-		return sam.TCC1
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC4].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-	case PA17:
+	}
-		return sam.TCC1
+}
-	case PA18:
+
-		return sam.TCC1
+func (tcc *TCC) timerNum() uint8 {
-	case PA19:
+	switch tcc.timer() {
-		return sam.TCC1
+	case sam.TCC0:
-	case PA20:
+		return 0
-		return sam.TCC0
+	case sam.TCC1:
-	case PA21:
+		return 1
-		return sam.TCC0
+	case sam.TCC2:
-	case PA22:
+		return 2
-		return sam.TCC0
+	case sam.TCC3:
-	case PA23:
+		return 3
-		return sam.TCC0
+	case sam.TCC4:
-	case PB12:
+		return 4
-		return sam.TCC3
+	default:
-	case PB13:
+		return 0x0f // should not happen
-		return sam.TCC3
-	case PB14:
-		return sam.TCC4
-	case PB15:
-		return sam.TCC4
-	case PB16:
-		return sam.TCC0
-	case PB17:
-		return sam.TCC0
-	case PB31:
-		return sam.TCC4
-	default:
-		return nil // not supported on this pin
 	}
 }

src/machine/machine_atsamd51p19.go

@@ -11,60 +11,49 @@ import "device/sam"
 
 const HSRAM_SIZE = 0x00030000
 
-// InitPWM initializes the PWM interface.
+// This chip has five TCC peripherals, which have PWM as one feature.
-func InitPWM() {
+var (
-	// turn on timer clocks used for PWM
+	TCC0 = (*TCC)(sam.TCC0)
-	sam.MCLK.APBBMASK.SetBits(sam.MCLK_APBBMASK_TCC0_ | sam.MCLK_APBBMASK_TCC1_)
+	TCC1 = (*TCC)(sam.TCC1)
-	sam.MCLK.APBCMASK.SetBits(sam.MCLK_APBCMASK_TCC2_ | sam.MCLK_APBCMASK_TCC3_)
+	TCC2 = (*TCC)(sam.TCC2)
-	sam.MCLK.APBDMASK.SetBits(sam.MCLK_APBDMASK_TCC4_)
+	TCC3 = (*TCC)(sam.TCC3)
+	TCC4 = (*TCC)(sam.TCC4)
+)
 
-	//use clock generator 0
+func (tcc *TCC) configureClock() {
-	sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC0].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) |
+	// Turn on timer clocks used for TCC and use generic clock generator 0.
-		sam.GCLK_PCHCTRL_CHEN)
+	switch tcc.timer() {
-	sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC2].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) |
+	case sam.TCC0:
-		sam.GCLK_PCHCTRL_CHEN)
+		sam.MCLK.APBBMASK.SetBits(sam.MCLK_APBBMASK_TCC0_)
-	sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC4].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) |
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC0].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-		sam.GCLK_PCHCTRL_CHEN)
+	case sam.TCC1:
-}
+		sam.MCLK.APBBMASK.SetBits(sam.MCLK_APBBMASK_TCC1_)
-
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC1].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-// getTimer returns the timer to be used for PWM on this pin
+	case sam.TCC2:
-func (pwm PWM) getTimer() *sam.TCC_Type {
+		sam.MCLK.APBCMASK.SetBits(sam.MCLK_APBCMASK_TCC2_)
-	switch pwm.Pin {
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC2].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-	case PA14:
+	case sam.TCC3:
-		return sam.TCC2
+		sam.MCLK.APBCMASK.SetBits(sam.MCLK_APBCMASK_TCC3_)
-	case PA15:
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC3].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-		return sam.TCC2
+	case sam.TCC4:
-	case PA16:
+		sam.MCLK.APBDMASK.SetBits(sam.MCLK_APBDMASK_TCC4_)
-		return sam.TCC1
+		sam.GCLK.PCHCTRL[sam.PCHCTRL_GCLK_TCC4].Set((sam.GCLK_PCHCTRL_GEN_GCLK0 << sam.GCLK_PCHCTRL_GEN_Pos) | sam.GCLK_PCHCTRL_CHEN)
-	case PA17:
+	}
-		return sam.TCC1
+}
-	case PA18:
+
-		return sam.TCC1
+func (tcc *TCC) timerNum() uint8 {
-	case PA19:
+	switch tcc.timer() {
-		return sam.TCC1
+	case sam.TCC0:
-	case PA20:
+		return 0
-		return sam.TCC0
+	case sam.TCC1:
-	case PA21:
+		return 1
-		return sam.TCC0
+	case sam.TCC2:
-	case PA22:
+		return 2
-		return sam.TCC0
+	case sam.TCC3:
-	case PA23:
+		return 3
-		return sam.TCC0
+	case sam.TCC4:
-	case PB12:
+		return 4
-		return sam.TCC3
+	default:
-	case PB13:
+		return 0x0f // should not happen
-		return sam.TCC3
-	case PB14:
-		return sam.TCC4
-	case PB15:
-		return sam.TCC4
-	case PB16:
-		return sam.TCC0
-	case PB17:
-		return sam.TCC0
-	case PB31:
-		return sam.TCC4
-	default:
-		return nil // not supported on this pin
 	}
 }

src/machine/machine_attiny.go

@@ -1,9 +0,0 @@
-// +build avr,attiny
-
-package machine
-
-// Tx is a dummy implementation. I2C has not been implemented for ATtiny
-// devices.
-func (i2c I2C) Tx(addr uint16, w, r []byte) error {
-	return nil
-}

src/machine/machine_avr.go

@@ -141,10 +141,3 @@ func (a ADC) Get() uint16 {
 
 	return uint16(avr.ADCL.Get()) | uint16(avr.ADCH.Get())<<8
 }
-
-// I2C on AVR.
-type I2C struct {
-}
-
-// I2C0 is the only I2C interface on most AVRs.
-var I2C0 = I2C{}

src/machine/machine_fe310.go

@@ -5,6 +5,7 @@ package machine
 import (
 	"device/sifive"
 	"runtime/interrupt"
+	"unsafe"
 )
 
 func CPUFrequency() uint32 {
@@ -185,9 +186,13 @@ func (spi SPI) Transfer(w byte) (byte, error) {
 
 // I2C on the FE310-G002.
 type I2C struct {
-	Bus *sifive.I2C_Type
+	Bus sifive.I2C_Type
 }
 
+var (
+	I2C0 = (*I2C)(unsafe.Pointer(sifive.I2C0))
+)
+
 // I2CConfig is used to store config info for I2C.
 type I2CConfig struct {
 	Frequency uint32
@@ -196,7 +201,7 @@ type I2CConfig struct {
 }
 
 // Configure is intended to setup the I2C interface.
-func (i2c I2C) Configure(config I2CConfig) error {
+func (i2c *I2C) Configure(config I2CConfig) error {
 	var i2cClockFrequency uint32 = 32000000
 	if config.Frequency == 0 {
 		config.Frequency = TWI_FREQ_100KHZ
@@ -228,7 +233,7 @@ func (i2c I2C) Configure(config I2CConfig) error {
 // Tx does a single I2C transaction at the specified address.
 // It clocks out the given address, writes the bytes in w, reads back len(r)
 // bytes and stores them in r, and generates a stop condition on the bus.
-func (i2c I2C) Tx(addr uint16, w, r []byte) error {
+func (i2c *I2C) Tx(addr uint16, w, r []byte) error {
 	var err error
 	if len(w) != 0 {
 		// send start/address for write
@@ -276,7 +281,7 @@ func (i2c I2C) Tx(addr uint16, w, r []byte) error {
 }
 
 // Writes a single byte to the I2C bus.
-func (i2c I2C) writeByte(data byte) error {
+func (i2c *I2C) writeByte(data byte) error {
 	// Send data byte
 	i2c.Bus.TXR_RXR.Set(uint32(data))
 
@@ -295,7 +300,7 @@ func (i2c I2C) writeByte(data byte) error {
 }
 
 // Reads a single byte from the I2C bus.
-func (i2c I2C) readByte() byte {
+func (i2c *I2C) readByte() byte {
 	i2c.Bus.CR_SR.Set(sifive.I2C_CR_RD)
 
 	// wait until transmission complete
@@ -306,7 +311,7 @@ func (i2c I2C) readByte() byte {
 }
 
 // Sends the address and start signal.
-func (i2c I2C) sendAddress(address uint16, write bool) error {
+func (i2c *I2C) sendAddress(address uint16, write bool) error {
 	data := (address << 1)
 	if !write {
 		data |= 1 // set read flag in transmit register

src/machine/machine_generic.go

@@ -6,7 +6,7 @@ package machine
 
 var (
 	SPI0  = SPI{0}
-	I2C0  = I2C{0}
+	I2C0  = &I2C{0}
 	UART0 = UART{0}
 )
 
@@ -84,24 +84,6 @@ func (adc ADC) Get() uint16 {
 //export __tinygo_adc_read
 func adcRead(pin Pin) uint16
 
-// InitPWM enables support for PWM peripherals.
-func InitPWM() {
-	// Nothing to do here.
-}
-
-// Configure configures a PWM pin for output.
-func (pwm PWM) Configure() error {
-	return nil
-}
-
-// Set turns on the duty cycle for a PWM pin using the provided value.
-func (pwm PWM) Set(value uint16) {
-	pwmSet(pwm.Pin, value)
-}
-
-//export __tinygo_pwm_set
-func pwmSet(pin Pin, value uint16)
-
 // I2C is a generic implementation of the Inter-IC communication protocol.
 type I2C struct {
 	Bus uint8
@@ -115,13 +97,13 @@ type I2CConfig struct {
 }
 
 // Configure is intended to setup the I2C interface.
-func (i2c I2C) Configure(config I2CConfig) error {
+func (i2c *I2C) Configure(config I2CConfig) error {
 	i2cConfigure(i2c.Bus, config.SCL, config.SDA)
 	return nil
 }
 
 // Tx does a single I2C transaction at the specified address.
-func (i2c I2C) Tx(addr uint16, w, r []byte) error {
+func (i2c *I2C) Tx(addr uint16, w, r []byte) error {
 	i2cTransfer(i2c.Bus, &w[0], len(w), &r[0], len(r))
 	// TODO: do something with the returned error code.
 	return nil

src/machine/machine_k210.go

@@ -7,6 +7,7 @@ import (
 	"device/riscv"
 	"errors"
 	"runtime/interrupt"
+	"unsafe"
 )
 
 func CPUFrequency() uint32 {
@@ -493,9 +494,15 @@ func (spi SPI) Transfer(w byte) (byte, error) {
 
 // I2C on the K210.
 type I2C struct {
-	Bus *kendryte.I2C_Type
+	Bus kendryte.I2C_Type
 }
 
+var (
+	I2C0 = (*I2C)(unsafe.Pointer(kendryte.I2C0))
+	I2C1 = (*I2C)(unsafe.Pointer(kendryte.I2C1))
+	I2C2 = (*I2C)(unsafe.Pointer(kendryte.I2C2))
+)
+
 // I2CConfig is used to store config info for I2C.
 type I2CConfig struct {
 	Frequency uint32
@@ -504,7 +511,7 @@ type I2CConfig struct {
 }
 
 // Configure is intended to setup the I2C interface.
-func (i2c I2C) Configure(config I2CConfig) error {
+func (i2c *I2C) Configure(config I2CConfig) error {
 
 	if config.Frequency == 0 {
 		config.Frequency = TWI_FREQ_100KHZ
@@ -518,7 +525,7 @@ func (i2c I2C) Configure(config I2CConfig) error {
 	// Enable APB0 clock.
 	kendryte.SYSCTL.CLK_EN_CENT.SetBits(kendryte.SYSCTL_CLK_EN_CENT_APB0_CLK_EN)
 
-	switch i2c.Bus {
+	switch &i2c.Bus {
 	case kendryte.I2C0:
 		// Initialize I2C0 clock.
 		kendryte.SYSCTL.CLK_EN_PERI.SetBits(kendryte.SYSCTL_CLK_EN_PERI_I2C0_CLK_EN)
@@ -567,7 +574,7 @@ func (i2c I2C) Configure(config I2CConfig) error {
 // Tx does a single I2C transaction at the specified address.
 // It clocks out the given address, writes the bytes in w, reads back len(r)
 // bytes and stores them in r, and generates a stop condition on the bus.
-func (i2c I2C) Tx(addr uint16, w, r []byte) error {
+func (i2c *I2C) Tx(addr uint16, w, r []byte) error {
 	// Set peripheral address.
 	i2c.Bus.TAR.Set(uint32(addr))
 	// Enable controller.

src/machine/machine_nrf.go

@@ -6,6 +6,7 @@ import (
 	"device/nrf"
 	"errors"
 	"runtime/interrupt"
+	"unsafe"
 )
 
 var (
@@ -203,13 +204,13 @@ func (uart *UART) handleInterrupt(interrupt.Interrupt) {
 
 // I2C on the NRF.
 type I2C struct {
-	Bus *nrf.TWI_Type
+	Bus nrf.TWI_Type
 }
 
 // There are 2 I2C interfaces on the NRF.
 var (
-	I2C0 = I2C{Bus: nrf.TWI0}
+	I2C0 = (*I2C)(unsafe.Pointer(nrf.TWI0))
-	I2C1 = I2C{Bus: nrf.TWI1}
+	I2C1 = (*I2C)(unsafe.Pointer(nrf.TWI1))
 )
 
 // I2CConfig is used to store config info for I2C.
@@ -220,7 +221,7 @@ type I2CConfig struct {
 }
 
 // Configure is intended to setup the I2C interface.
-func (i2c I2C) Configure(config I2CConfig) error {
+func (i2c *I2C) Configure(config I2CConfig) error {
 	// Default I2C bus speed is 100 kHz.
 	if config.Frequency == 0 {
 		config.Frequency = TWI_FREQ_100KHZ
@@ -261,7 +262,7 @@ func (i2c I2C) Configure(config I2CConfig) error {
 // Tx does a single I2C transaction at the specified address.
 // It clocks out the given address, writes the bytes in w, reads back len(r)
 // bytes and stores them in r, and generates a stop condition on the bus.
-func (i2c I2C) Tx(addr uint16, w, r []byte) (err error) {
+func (i2c *I2C) Tx(addr uint16, w, r []byte) (err error) {
 	i2c.Bus.ADDRESS.Set(uint32(addr))
 
 	if len(w) != 0 {
@@ -299,7 +300,7 @@ cleanUp:
 // signalStop sends a stop signal when writing or tells the I2C peripheral that
 // it must generate a stop condition after the next character is retrieved when
 // reading.
-func (i2c I2C) signalStop() {
+func (i2c *I2C) signalStop() {
 	i2c.Bus.TASKS_STOP.Set(1)
 	for i2c.Bus.EVENTS_STOPPED.Get() == 0 {
 	}
@@ -307,7 +308,7 @@ func (i2c I2C) signalStop() {
 }
 
 // writeByte writes a single byte to the I2C bus.
-func (i2c I2C) writeByte(data byte) error {
+func (i2c *I2C) writeByte(data byte) error {
 	i2c.Bus.TXD.Set(uint32(data))
 	for i2c.Bus.EVENTS_TXDSENT.Get() == 0 {
 		if e := i2c.Bus.EVENTS_ERROR.Get(); e != 0 {
@@ -320,7 +321,7 @@ func (i2c I2C) writeByte(data byte) error {
 }
 
 // readByte reads a single byte from the I2C bus.
-func (i2c I2C) readByte() (byte, error) {
+func (i2c *I2C) readByte() (byte, error) {
 	for i2c.Bus.EVENTS_RXDREADY.Get() == 0 {
 		if e := i2c.Bus.EVENTS_ERROR.Get(); e != 0 {
 			i2c.Bus.EVENTS_ERROR.Set(0)

src/machine/machine_nrf51.go

@@ -24,7 +24,7 @@ func (uart UART) setPins(tx, rx Pin) {
 	nrf.UART0.PSELRXD.Set(uint32(rx))
 }
 
-func (i2c I2C) setPins(scl, sda Pin) {
+func (i2c *I2C) setPins(scl, sda Pin) {
 	i2c.Bus.PSELSCL.Set(uint32(scl))
 	i2c.Bus.PSELSDA.Set(uint32(sda))
 }

src/machine/machine_nrf52.go

@@ -56,14 +56,14 @@ func (uart UART) setPins(tx, rx Pin) {
 	nrf.UART0.PSELRXD.Set(uint32(rx))
 }
 
-func (i2c I2C) setPins(scl, sda Pin) {
+func (i2c *I2C) setPins(scl, sda Pin) {
 	i2c.Bus.PSELSCL.Set(uint32(scl))
 	i2c.Bus.PSELSDA.Set(uint32(sda))
 }
 
 // PWM
 var (
-	pwmChannelPins     = [3]uint32{0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF}
+	PWM0 = &PWM{PWM: nrf.PWM0}
-	pwms               = [3]*nrf.PWM_Type{nrf.PWM0, nrf.PWM1, nrf.PWM2}
+	PWM1 = &PWM{PWM: nrf.PWM1}
-	pwmChannelSequence [3]uint16
+	PWM2 = &PWM{PWM: nrf.PWM2}
 )