diff --git a/builder/build.go b/builder/build.go
index e5977770..0361dd64 100644
--- a/builder/build.go
+++ b/builder/build.go
@@ -4,14 +4,18 @@
 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"
@@ -38,6 +42,26 @@ 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
+	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.
@@ -45,6 +69,13 @@ type BuildResult struct {
 // 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, 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(),
@@ -87,23 +118,200 @@ func Build(pkgName, outpath string, config *compileopts.Config, action func(Buil
 	// Makefile target.
 	var jobs []*compileJob
 
-	// Add job to compile and optimize all Go files at once.
-	// TODO: parallelize this.
+	// Create the *ssa.Program. This does not yet build the entire SSA of the
+	// program so it's pretty fast and doesn't need to be parallelized.
+	program := lprogram.LoadSSA()
+
+	// 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,
+			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() 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)
+				}
+
+				// 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 that links and optimizes all packages together.
 	var mod llvm.Module
 	var stackSizeLoads []string
 	programJob := &compileJob{
-		description: "compile Go files",
-		run: func() (err error) {
-			mod, err = compileWholeProgram(pkgName, config, compilerConfig, lprogram, machine)
-			if err != nil {
-				return
+		description:  "link+optimize packages (LTO)",
+		dependencies: packageJobs,
+		run: func() error {
+			// Load and link all the bitcode files. This does not yet optimize
+			// 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)
@@ -140,13 +348,6 @@ func Build(pkgName, outpath string, config *compileopts.Config, action func(Buil
 	// 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{
@@ -366,29 +567,16 @@ func Build(pkgName, outpath string, config *compileopts.Config, action func(Buil
 	})
 }
 
-// compileWholeProgram compiles the entire *loader.Program to a LLVM module and
-// applies most necessary optimizations and transformations.
-func compileWholeProgram(pkgName string, config *compileopts.Config, compilerConfig *compiler.Config, lprogram *loader.Program, machine llvm.TargetMachine) (llvm.Module, error) {
-	// Compile AST to IR.
-	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")
-	}
-
+// optimizeProgram runs a series of optimizations and transformations that are
+// needed to convert a program to its final form. Some transformations are not
+// optional and must be run as the compiler expects them to run.
+func optimizeProgram(mod llvm.Module, config *compileopts.Config) error {
 	err := interp.Run(mod, config.DumpSSA())
 	if err != nil {
-		return mod, err
+		return err
 	}
 	if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
-		return mod, errors.New("verification error after interpreting runtime.initAll")
+		return errors.New("verification error after interpreting runtime.initAll")
 	}
 
 	if config.GOOS() != "darwin" {
@@ -403,13 +591,13 @@ 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":
 		errs = transform.Optimize(mod, config, 0, 0, 0) // -O0
@@ -422,13 +610,13 @@ 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
@@ -442,7 +630,7 @@ 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
diff --git a/cgo/cgo.go b/cgo/cgo.go
index 73d26e1d..c496cfef 100644
--- a/cgo/cgo.go
+++ b/cgo/cgo.go
@@ -42,6 +42,7 @@ type cgoPackage struct {
 	enums           map[string]enumInfo
 	anonStructNum   int
 	ldflags         []string
+	visitedFiles    map[string][]byte
 }
 
 // constantInfo stores some information about a CGo constant found by libclang
@@ -156,9 +157,10 @@ 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
-// but still modifies the AST.
-func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string) (*ast.File, []string, []error) {
+// files, the LDFLAGS for this package, and a map of file hashes of the accessed
+// C header files. If there is one or more error, it returns 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, map[string][]byte, []error) {
 	p := &cgoPackage{
 		dir:             dir,
 		fset:            fset,
@@ -170,6 +172,7 @@ 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.
@@ -185,7 +188,7 @@ 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, []error{
 			scanner.Error{
 				Pos: fset.Position(files[0].Pos()),
 				Msg: "cgo: cannot find absolute path: " + err.Error(), // TODO: wrap this error
@@ -427,7 +430,7 @@ 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.ldflags, p.visitedFiles, p.errors
 }
 
 // makePathsAbsolute converts some common path compiler flags (-I, -L) from
diff --git a/cgo/cgo_test.go b/cgo/cgo_test.go
index e0fbc633..33a49b96 100644
--- a/cgo/cgo_test.go
+++ b/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
diff --git a/cgo/libclang.go b/cgo/libclang.go
index f1ba3baa..8e9af8a0 100644
--- a/cgo/libclang.go
+++ b/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)
+}
diff --git a/compiler/compiler.go b/compiler/compiler.go
index e25cc673..abd7be64 100644
--- a/compiler/compiler.go
+++ b/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 = 1
+
 func init() {
 	llvm.InitializeAllTargets()
 	llvm.InitializeAllTargetMCs()
@@ -242,15 +247,14 @@ func Sizes(machine llvm.TargetMachine) types.Sizes {
 	}
 }
 
-// CompileProgram compiles the given package path or .go file path. Return an
-// error when this fails (in any stage). If successful it returns the LLVM
-// module. If not, one or more errors will be returned.
-func CompileProgram(lprogram *loader.Program, machine llvm.TargetMachine, config *Config, dumpSSA bool) (llvm.Module, []error) {
-	c := newCompilerContext("", machine, config, dumpSSA)
+// CompilePackage compiles a single package to a LLVM module.
+func CompilePackage(moduleName string, pkg *loader.Package, ssaPkg *ssa.Package, machine llvm.TargetMachine, config *Config, dumpSSA bool) (llvm.Module, []error) {
+	c := newCompilerContext(moduleName, machine, config, dumpSSA)
+	c.runtimePkg = ssaPkg.Prog.ImportedPackage("runtime").Pkg
+	c.program = ssaPkg.Prog
 
-	c.program = lprogram.LoadSSA()
-	c.program.Build()
-	c.runtimePkg = c.program.ImportedPackage("runtime").Pkg
+	// Convert AST to SSA.
+	ssaPkg.Build()
 
 	// Initialize debug information.
 	if c.Debug {
@@ -263,43 +267,17 @@ func CompileProgram(lprogram *loader.Program, machine llvm.TargetMachine, config
 		})
 	}
 
-	for _, pkg := range lprogram.Sorted() {
-		c.loadASTComments(pkg)
-	}
+	// 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))
 
-	// Define all functions.
-	var initFuncs []llvm.Value
+	// Compile all functions, methods, and global variables in this package.
 	irbuilder := c.ctx.NewBuilder()
 	defer irbuilder.Dispose()
-	for _, pkg := range lprogram.Sorted() {
-		ssaPkg := c.program.Package(pkg.Pkg)
-		c.createPackage(irbuilder, ssaPkg)
-		initFuncs = append(initFuncs, c.getFunction(ssaPkg.Members["init"].(*ssa.Function)))
-	}
-
-	// 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()
+	c.createPackage(irbuilder, ssaPkg)
 
 	// see: https://reviews.llvm.org/D18355
 	if c.Debug {
@@ -320,22 +298,6 @@ func CompileProgram(lprogram *loader.Program, machine llvm.TargetMachine, config
 		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()
-
-	// Compile all functions, methods, and global variables in this package.
-	irbuilder := c.ctx.NewBuilder()
-	defer irbuilder.Dispose()
-	c.createPackage(irbuilder, ssaPkg)
-
 	return c.mod, nil
 }
 
@@ -682,8 +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))
-	for _, param := range f.Params {
+	params := getParams(f.Signature)
+	diparams := make([]llvm.Metadata, 0, len(params))
+	for _, param := range params {
 		diparams = append(diparams, c.getDIType(param.Type()))
 	}
 	diFuncType := c.dibuilder.CreateSubroutineType(llvm.DISubroutineType{
@@ -792,7 +755,7 @@ func (c *compilerContext) createPackage(irbuilder llvm.Builder, pkg *ssa.Package
 			if !info.extern {
 				global := c.getGlobal(member)
 				global.SetInitializer(llvm.ConstNull(global.Type().ElementType()))
-				global.SetLinkage(llvm.InternalLinkage)
+				global.SetVisibility(llvm.HiddenVisibility)
 			}
 		}
 	}
@@ -815,7 +778,7 @@ func (b *builder) createFunction() {
 		return
 	}
 	if !b.info.exported {
-		b.llvmFn.SetLinkage(llvm.InternalLinkage)
+		b.llvmFn.SetVisibility(llvm.HiddenVisibility)
 		b.llvmFn.SetUnnamedAddr(true)
 	}
 
diff --git a/compiler/compiler_test.go b/compiler/compiler_test.go
index 3919dd81..6802d56d 100644
--- a/compiler/compiler_test.go
+++ b/compiler/compiler_test.go
@@ -65,8 +65,9 @@ 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)
diff --git a/compiler/defer.go b/compiler/defer.go
index 5d356f52..f8ebb9f0 100644
--- a/compiler/defer.go
+++ b/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)
diff --git a/compiler/func.go b/compiler/func.go
index 37e8dac7..31162c1d 100644
--- a/compiler/func.go
+++ b/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:
diff --git a/compiler/goroutine.go b/compiler/goroutine.go
index 94965833..55090a9b 100644
--- a/compiler/goroutine.go
+++ b/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")
diff --git a/compiler/interface.go b/compiler/interface.go
index 9427c638..b21476b6 100644
--- a/compiler/interface.go
+++ b/compiler/interface.go
@@ -31,7 +31,7 @@ func (b *builder) createMakeInterface(val llvm.Value, typ types.Type, pos token.
 		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)
+		itfConcreteTypeGlobal.SetLinkage(llvm.LinkOnceODRLinkage)
 	}
 	itfTypeCode := b.CreatePtrToInt(itfConcreteTypeGlobal, b.uintptrType, "")
 	itf := llvm.Undef(b.getLLVMRuntimeType("_interface"))
@@ -80,7 +80,7 @@ func (c *compilerContext) getTypeCode(typ types.Type) llvm.Value {
 				globalValue = llvm.ConstInsertValue(globalValue, lengthValue, []uint32{1})
 			}
 			global.SetInitializer(globalValue)
-			global.SetLinkage(llvm.PrivateLinkage)
+			global.SetLinkage(llvm.LinkOnceODRLinkage)
 		}
 		global.SetGlobalConstant(true)
 	}
@@ -264,7 +264,7 @@ 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 +295,7 @@ 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})
 }
 
@@ -445,7 +445,7 @@ 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) {
 		expandedReceiverType = append(expandedReceiverType, info.llvmType)
@@ -467,7 +467,7 @@ 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.
diff --git a/compiler/interrupt.go b/compiler/interrupt.go
index 15278751..d763ce78 100644
--- a/compiler/interrupt.go
+++ b/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)
diff --git a/compiler/symbol.go b/compiler/symbol.go
index ea0b01d1..9857c881 100644
--- a/compiler/symbol.go
+++ b/compiler/symbol.go
@@ -72,7 +72,7 @@ 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())
 		paramInfos = append(paramInfos, paramFragmentInfos...)
@@ -183,7 +183,7 @@ func (c *compilerContext) getFunction(fn *ssa.Function) llvm.Value {
 		b := newBuilder(c, irbuilder, fn)
 		b.createFunction()
 		irbuilder.Dispose()
-		llvmFn.SetLinkage(llvm.InternalLinkage)
+		llvmFn.SetLinkage(llvm.LinkOnceODRLinkage)
 		llvmFn.SetUnnamedAddr(true)
 	}
 
@@ -293,6 +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).
diff --git a/compiler/testdata/basic.ll b/compiler/testdata/basic.ll
index dd5c9fa0..0c4f06c3 100644
--- a/compiler/testdata/basic.ll
+++ b/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
diff --git a/compiler/testdata/float.ll b/compiler/testdata/float.ll
index 817ee091..009b92c0 100644
--- a/compiler/testdata/float.ll
+++ b/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
diff --git a/compiler/testdata/pointer.ll b/compiler/testdata/pointer.ll
index a5f4c889..26a584a9 100644
--- a/compiler/testdata/pointer.ll
+++ b/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
diff --git a/compiler/testdata/slice.ll b/compiler/testdata/slice.ll
index 72301eef..1fb054c5 100644
--- a/compiler/testdata/slice.ll
+++ b/compiler/testdata/slice.ll
@@ -3,17 +3,19 @@ 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
 }
diff --git a/go.mod b/go.mod
index bc9c6fe3..45900b54 100644
--- a/go.mod
+++ b/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 v0.0.0-20210308112806-9ef958b6bed4
 )
diff --git a/go.sum b/go.sum
index b32b598c..96850fa8 100644
--- a/go.sum
+++ b/go.sum
@@ -57,5 +57,5 @@ 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 v0.0.0-20210206225315-7fe719483a0f/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
+tinygo.org/x/go-llvm v0.0.0-20210308112806-9ef958b6bed4 h1:CMUHxVTb+UuUePuMf8vkWjZ3gTp9BBK91KrgOCwoNHs=
+tinygo.org/x/go-llvm v0.0.0-20210308112806-9ef958b6bed4/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
diff --git a/loader/loader.go b/loader/loader.go
index 42347013..1eed4aae 100644
--- a/loader/loader.go
+++ b/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
-	Files   []*ast.File
-	Pkg     *types.Package
-	info    types.Info
+	program    *Program
+	Files      []*ast.File
+	FileHashes map[string][]byte
+	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) {
-	if p.fset == nil {
-		p.fset = token.NewFileSet()
-	}
-
-	rd, err := os.Open(path)
+func (p *Package) parseFile(path string, mode parser.Mode) (*ast.File, error) {
+	originalPath := p.program.getOriginalPath(path)
+	data, err := ioutil.ReadFile(path)
 	if err != nil {
 		return nil, err
 	}
-	defer rd.Close()
-	return parser.ParseFile(p.fset, p.getOriginalPath(path), rd, mode)
+	sum := sha512.Sum512_224(data)
+	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
@@ -385,7 +388,11 @@ func (p *Package) parseFiles() ([]*ast.File, error) {
 		if p.program.clangHeaders != "" {
 			cflags = append(cflags, "-Xclang", "-internal-isystem", "-Xclang", p.program.clangHeaders)
 		}
-		generated, ldflags, errs := cgo.Process(files, p.program.workingDir, p.program.fset, cflags)
+		p.CFlags = cflags
+		generated, ldflags, accessedFiles, errs := cgo.Process(files, p.program.workingDir, p.program.fset, cflags)
+		for path, hash := range accessedFiles {
+			p.FileHashes[path] = hash
+		}
 		if errs != nil {
 			fileErrs = append(fileErrs, errs...)
 		}
diff --git a/src/internal/task/task_coroutine.go b/src/internal/task/task_coroutine.go
index bfeedc30..33d910ed 100644
--- a/src/internal/task/task_coroutine.go
+++ b/src/internal/task/task_coroutine.go
@@ -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()
 }
 
diff --git a/transform/interface-lowering.go b/transform/interface-lowering.go
index 6aa3bb13..19546c58 100644
--- a/transform/interface-lowering.go
+++ b/transform/interface-lowering.go
@@ -173,11 +173,12 @@ func (p *lowerInterfacesPass) run() error {
 			// Only the name of the global is relevant, the object itself is
 			// discarded afterwards.
 			name := global.Name()
-			t := &typeInfo{
-				name:     name,
-				typecode: global,
+			if _, ok := p.types[name]; !ok {
+				p.types[name] = &typeInfo{
+					name:     name,
+					typecode: global,
+				}
 			}
-			p.types[name] = t
 		case typeInInterfacePtr:
 			// Count per type how often it is put in an interface. Also, collect
 			// all methods this type has (if it is named).
@@ -185,7 +186,15 @@ func (p *lowerInterfacesPass) run() error {
 			initializer := global.Initializer()
 			typecode := llvm.ConstExtractValue(initializer, []uint32{0})
 			methodSet := llvm.ConstExtractValue(initializer, []uint32{1})
-			t := p.types[typecode.Name()]
+			typecodeName := typecode.Name()
+			t := p.types[typecodeName]
+			if t == nil {
+				t = &typeInfo{
+					name:     typecodeName,
+					typecode: typecode,
+				}
+				p.types[typecodeName] = t
+			}
 			p.addTypeMethods(t, methodSet)
 
 			// Count the number of MakeInterface instructions, for sorting the