interp: support GEP on fixed (MMIO) addresses

GetElementPtr would not work on values that weren't pointers. Because
fixed addresses (often used in memory-mapped I/O) are integers rather
than pointers in interp, it would return an error.

This resulted in the teensy40 target not compiling correctly since the
interp package rewrite. This commit should fix that.

.circleci/config.yml

@@ -68,12 +68,12 @@ commands:
     steps:
       - restore_cache:
           keys:
-            - llvm-source-10-v1
+            - llvm-source-11-v1
       - run:
           name: "Fetch LLVM source"
           command: make llvm-source
       - save_cache:
-          key: llvm-source-10-v1
+          key: llvm-source-11-v1
           paths:
             - llvm-project
   build-wasi-libc:
@@ -153,17 +153,14 @@ commands:
       - llvm-source-linux
       - restore_cache:
           keys:
-            - llvm-build-10-linux-v1-assert
+            - llvm-build-11-linux-v1-assert
       - run:
           name: "Build LLVM"
           command: |
             if [ ! -f llvm-build/lib/liblldELF.a ]
             then
               # install dependencies
-              sudo apt-get install cmake clang ninja-build
-              # make build faster
-              export CC=clang
-              export CXX=clang++
+              sudo apt-get install cmake ninja-build
               # hack ninja to use less jobs
               echo -e '#!/bin/sh\n/usr/bin/ninja -j3 "$@"' > /go/bin/ninja
               chmod +x /go/bin/ninja
@@ -171,7 +168,7 @@ commands:
               make ASSERT=1 llvm-build
             fi
       - save_cache:
-          key: llvm-build-10-linux-v1-assert
+          key: llvm-build-11-linux-v1-assert
           paths:
             llvm-build
       - run: make ASSERT=1
@@ -214,17 +211,14 @@ commands:
       - llvm-source-linux
       - restore_cache:
           keys:
-            - llvm-build-10-linux-v1
+            - llvm-build-11-linux-v1-noassert
       - run:
           name: "Build LLVM"
           command: |
             if [ ! -f llvm-build/lib/liblldELF.a ]
             then
               # install dependencies
-              sudo apt-get install cmake clang ninja-build
-              # make build faster
-              export CC=clang
-              export CXX=clang++
+              sudo apt-get install cmake ninja-build
               # hack ninja to use less jobs
               echo -e '#!/bin/sh\n/usr/bin/ninja -j3 "$@"' > /go/bin/ninja
               chmod +x /go/bin/ninja
@@ -232,7 +226,7 @@ commands:
               make llvm-build
             fi
       - save_cache:
-          key: llvm-build-10-linux-v1
+          key: llvm-build-11-linux-v1-noassert
           paths:
             llvm-build
       - build-wasi-libc
@@ -274,8 +268,8 @@ commands:
       - run:
           name: "Install dependencies"
           command: |
-            curl https://dl.google.com/go/go1.14.darwin-amd64.tar.gz -o go1.14.darwin-amd64.tar.gz
-            sudo tar -C /usr/local -xzf go1.14.darwin-amd64.tar.gz
+            curl https://dl.google.com/go/go1.15.5.darwin-amd64.tar.gz -o go1.15.5.darwin-amd64.tar.gz
+            sudo tar -C /usr/local -xzf go1.15.5.darwin-amd64.tar.gz
             ln -s /usr/local/go/bin/go /usr/local/bin/go
             HOMEBREW_NO_AUTO_UPDATE=1 brew install qemu
       - install-xtensa-toolchain:
@@ -286,17 +280,17 @@ commands:
             - go-cache-macos-v2-{{ checksum "go.mod" }}
       - restore_cache:
           keys:
-            - llvm-source-10-macos-v1
+            - llvm-source-11-macos-v1
       - run:
           name: "Fetch LLVM source"
           command: make llvm-source
       - save_cache:
-          key: llvm-source-10-macos-v1
+          key: llvm-source-11-macos-v1
           paths:
             - llvm-project
       - restore_cache:
           keys:
-            - llvm-build-10-macos-v1
+            - llvm-build-11-macos-v1
       - run:
           name: "Build LLVM"
           command: |
@@ -308,7 +302,7 @@ commands:
               make llvm-build
             fi
       - save_cache:
-          key: llvm-build-10-macos-v1
+          key: llvm-build-11-macos-v1
           paths:
             llvm-build
       - restore_cache:
@@ -344,60 +338,6 @@ commands:
           paths:
             - ~/.cache/go-build
             - /go/pkg/mod
-  arch-release:
-    steps:
-      - run:
-          name: Install dependencies
-          command: pacman -Sy --noconfirm openssh git pacman-contrib binutils
-      - run:
-          name: Create TinyGo user
-          command: useradd -m tinygo
-      - run:
-          name: Start SSH Agent
-          user: tinygo
-          command: eval $(ssh-agent -s)
-      - run:
-          name: Add ARCH_RELEASE_SSH_PRIVATE_KEY identity
-          user: tinygo
-          command: echo "${ARCH_RELEASE_SSH_PRIVATE_KEY}" | tr -d '\r' | ssh-add -
-      - run:
-          name: Create SSH directory
-          user: tinygo
-          command: mkdir -p ~/.ssh && chmod 700 ~/.ssh
-      - run:
-          name: Add aur.archlinux.org to known hosts
-          user: tinygo
-          command: ssh-keyscan aur.archlinux.org >> ~/.ssh/known_hosts
-      - run:
-          name: Clone tinygo-bin repo
-          user: tinygo
-          command: git clone ssh://aur@aur.archlinux.org/tinygo-bin.git ~/tinygo-bin
-      - run:
-          name: Update package version
-          user: tinygo
-          command: sed -i -E "s/(pkgver=)(.*)$/\1${CIRCLE_TAG}/" ~/tinygo-bin/PKGBUILD
-      - run:
-          name: Update file checksums
-          user: tinygo
-          command: cd ~/tinygo-bin && updpkgsums
-      - run:
-          name: Update .SRCINFO
-          user: tinygo
-          command: cd ~/tinygo-bin && makepkg --printsrcinfo > .SRCINFO
-      # Commit the update
-      - run:
-          name: Set git commit config
-          user: tinygo
-          command: |
-            git config --global user.email "tinygo-bot@tinygo.org" &&
-            git config --global user.name "TinyGo Release Bot"
-      - run:
-          name: Commit and push changes
-          user: tinygo
-          command: |
-            cd ~/tinygo-bin &&
-            git commit -a -m "Update tinygo-bin to v${CIRCLE_TAG}" &&
-            git push origin master
 
 jobs:
   test-llvm9-go111:
@@ -445,12 +385,6 @@ jobs:
       xcode: "10.1.0"
     steps:
       - build-macos
-  arch-release:
-    docker:
-      - image: archlinux:latest
-    steps:
-      - arch-release
-
 
 
 
@@ -465,12 +399,3 @@ workflows:
       - build-linux
       - build-macos
       - assert-test-linux
-  release:
-    jobs:
-      - arch-release:
-          filters:
-            branches:
-              ignore: /.*/
-            tags:
-              # Runs on every semver release
-              only: /v[0-9]+(\.[0-9]+)*(-.*)*/

Makefile

@@ -48,23 +48,13 @@ ifeq ($(OS),Windows_NT)
     # LLVM compiled using MinGW on Windows appears to have problems with threads.
     # Without this flag, linking results in errors like these:
     #     libLLVMSupport.a(Threading.cpp.obj):Threading.cpp:(.text+0x55): undefined reference to `std::thread::hardware_concurrency()'
-    LLVM_OPTION += -DLLVM_ENABLE_THREADS=OFF
+    LLVM_OPTION += -DLLVM_ENABLE_THREADS=OFF -DLLVM_ENABLE_PIC=OFF
 
+    CGO_CPPFLAGS += -DCINDEX_NO_EXPORTS
     CGO_LDFLAGS += -static -static-libgcc -static-libstdc++
     CGO_LDFLAGS_EXTRA += -lversion
 
-    # Build libclang manually because the CMake-based build system on Windows
-    # doesn't allow building libclang as a static library.
-    LIBCLANG_PATH = $(abspath build/libclang-custom.a)
-    LIBCLANG_FILES = $(abspath $(wildcard $(LLVM_BUILDDIR)/tools/clang/tools/libclang/CMakeFiles/libclang.dir/*.cpp.obj))
-
-    # Add the libclang dependency to the tinygo binary target.
-tinygo: $(LIBCLANG_PATH)
-test: $(LIBCLANG_PATH)
-    # Build libclang.
-$(LIBCLANG_PATH): $(LIBCLANG_FILES)
-	@mkdir -p build
-	ar rcs $(LIBCLANG_PATH) $^
+    LIBCLANG_PATH = $(abspath $(LLVM_BUILDDIR))/lib/liblibclang.a
 
 else ifeq ($(shell uname -s),Darwin)
     MD5SUM = md5
@@ -87,7 +77,7 @@ LLD_LIBS = $(START_GROUP) -llldCOFF -llldCommon -llldCore -llldDriver -llldELF -
 
 # For static linking.
 ifneq ("$(wildcard $(LLVM_BUILDDIR)/bin/llvm-config*)","")
-    CGO_CPPFLAGS=$(shell $(LLVM_BUILDDIR)/bin/llvm-config --cppflags) -I$(abspath $(LLVM_BUILDDIR))/tools/clang/include -I$(abspath $(CLANG_SRC))/include -I$(abspath $(LLD_SRC))/include
+    CGO_CPPFLAGS+=$(shell $(LLVM_BUILDDIR)/bin/llvm-config --cppflags) -I$(abspath $(LLVM_BUILDDIR))/tools/clang/include -I$(abspath $(CLANG_SRC))/include -I$(abspath $(LLD_SRC))/include
     CGO_CXXFLAGS=-std=c++14
     CGO_LDFLAGS+=$(LIBCLANG_PATH) -L$(abspath $(LLVM_BUILDDIR)/lib) $(CLANG_LIBS) $(LLD_LIBS) $(shell $(LLVM_BUILDDIR)/bin/llvm-config --ldflags --libs --system-libs $(LLVM_COMPONENTS)) -lstdc++ $(CGO_LDFLAGS_EXTRA)
 endif
@@ -146,7 +136,7 @@ gen-device-stm32: build/gen-device-svd
 
 # Get LLVM sources.
 $(LLVM_PROJECTDIR)/README.md:
-	git clone -b xtensa_release_10.0.1 --depth=1 https://github.com/tinygo-org/llvm-project $(LLVM_PROJECTDIR)
+	git clone -b xtensa_release_11.0.0 --depth=1 https://github.com/tinygo-org/llvm-project $(LLVM_PROJECTDIR)
 llvm-source: $(LLVM_PROJECTDIR)/README.md
 
 # Configure LLVM.
@@ -187,9 +177,11 @@ tinygo-test:
 	$(TINYGO) test encoding/ascii85
 	$(TINYGO) test encoding/base32
 	$(TINYGO) test encoding/hex
+	$(TINYGO) test hash/adler32
 	$(TINYGO) test hash/fnv
 	$(TINYGO) test hash/crc64
 	$(TINYGO) test math
+	$(TINYGO) test math/cmplx
 	$(TINYGO) test text/scanner
 	$(TINYGO) test unicode/utf8
 
@@ -329,6 +321,8 @@ 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
 	# test pwm
 	$(TINYGO) build -size short -o test.hex -target=itsybitsy-m0        examples/pwm
 	@$(MD5SUM) test.hex

README.md

@@ -148,6 +148,6 @@ The original reasoning was: if [Python](https://micropython.org/) can run on mic
 
 This project is licensed under the BSD 3-clause license, just like the [Go project](https://golang.org/LICENSE) itself.
 
-Some code has been copied from the LLVM project and is therefore licensed under [a variant of the Apache 2.0 license](http://releases.llvm.org/10.0.0/LICENSE.TXT). This has been clearly indicated in the header of these files.
+Some code has been copied from the LLVM project and is therefore licensed under [a variant of the Apache 2.0 license](http://releases.llvm.org/11.0.0/LICENSE.TXT). This has been clearly indicated in the header of these files.
 
 Some code has been copied and/or ported from Paul Stoffregen's Teensy libraries and is therefore licensed under PJRC's license. This has been clearly indicated in the header of these files.

azure-pipelines.yml

@@ -18,7 +18,7 @@ jobs:
     - task: Cache@2
       displayName: Cache LLVM source
       inputs:
-        key: llvm-source-10-windows-v1
+        key: llvm-source-11-windows-v1
         path: llvm-project
     - task: Bash@3
       displayName: Download LLVM source
@@ -32,7 +32,7 @@ jobs:
     - task: CacheBeta@0
       displayName: Cache LLVM build
       inputs:
-        key: llvm-build-10-windows-v1
+        key: llvm-build-11-windows-v3
         path: llvm-build
     - task: Bash@3
       displayName: Build LLVM
@@ -41,7 +41,11 @@ jobs:
         script: |
           if [ ! -f llvm-build/lib/liblldELF.a ]
           then
+            # install dependencies
             choco install ninja
+            # hack ninja to use fewer jobs
+            echo -e 'C:\\ProgramData\\Chocolatey\\bin\\ninja -j4 %*' > /usr/bin/ninja.bat
+            # build!
             make llvm-build
           fi
     - task: Bash@3

builder/cc1as.cpp

@@ -101,7 +101,7 @@ bool AssemblerInvocation::CreateFromArgs(AssemblerInvocation &Opts,
 
   // Target Options
   Opts.Triple = llvm::Triple::normalize(Args.getLastArgValue(OPT_triple));
-  Opts.CPU = Args.getLastArgValue(OPT_target_cpu);
+  Opts.CPU = std::string(Args.getLastArgValue(OPT_target_cpu));
   Opts.Features = Args.getAllArgValues(OPT_target_feature);
 
   // Use the default target triple if unspecified.
@@ -132,13 +132,19 @@ bool AssemblerInvocation::CreateFromArgs(AssemblerInvocation &Opts,
 
   Opts.RelaxELFRelocations = Args.hasArg(OPT_mrelax_relocations);
   Opts.DwarfVersion = getLastArgIntValue(Args, OPT_dwarf_version_EQ, 2, Diags);
-  Opts.DwarfDebugFlags = Args.getLastArgValue(OPT_dwarf_debug_flags);
-  Opts.DwarfDebugProducer = Args.getLastArgValue(OPT_dwarf_debug_producer);
-  Opts.DebugCompilationDir = Args.getLastArgValue(OPT_fdebug_compilation_dir);
-  Opts.MainFileName = Args.getLastArgValue(OPT_main_file_name);
+  Opts.DwarfDebugFlags =
+      std::string(Args.getLastArgValue(OPT_dwarf_debug_flags));
+  Opts.DwarfDebugProducer =
+      std::string(Args.getLastArgValue(OPT_dwarf_debug_producer));
+  Opts.DebugCompilationDir =
+      std::string(Args.getLastArgValue(OPT_fdebug_compilation_dir));
+  Opts.MainFileName = std::string(Args.getLastArgValue(OPT_main_file_name));
 
-  for (const auto &Arg : Args.getAllArgValues(OPT_fdebug_prefix_map_EQ))
-    Opts.DebugPrefixMap.insert(StringRef(Arg).split('='));
+  for (const auto &Arg : Args.getAllArgValues(OPT_fdebug_prefix_map_EQ)) {
+    auto Split = StringRef(Arg).split('=');
+    Opts.DebugPrefixMap.insert(
+        {std::string(Split.first), std::string(Split.second)});
+  }
 
   // Frontend Options
   if (Args.hasArg(OPT_INPUT)) {
@@ -154,8 +160,9 @@ bool AssemblerInvocation::CreateFromArgs(AssemblerInvocation &Opts,
     }
   }
   Opts.LLVMArgs = Args.getAllArgValues(OPT_mllvm);
-  Opts.OutputPath = Args.getLastArgValue(OPT_o);
-  Opts.SplitDwarfOutput = Args.getLastArgValue(OPT_split_dwarf_output);
+  Opts.OutputPath = std::string(Args.getLastArgValue(OPT_o));
+  Opts.SplitDwarfOutput =
+      std::string(Args.getLastArgValue(OPT_split_dwarf_output));
   if (Arg *A = Args.getLastArg(OPT_filetype)) {
     StringRef Name = A->getValue();
     unsigned OutputType = StringSwitch<unsigned>(Name)
@@ -183,8 +190,9 @@ bool AssemblerInvocation::CreateFromArgs(AssemblerInvocation &Opts,
   Opts.NoExecStack = Args.hasArg(OPT_mno_exec_stack);
   Opts.FatalWarnings = Args.hasArg(OPT_massembler_fatal_warnings);
   Opts.NoWarn = Args.hasArg(OPT_massembler_no_warn);
-  Opts.RelocationModel = Args.getLastArgValue(OPT_mrelocation_model, "pic");
-  Opts.TargetABI = Args.getLastArgValue(OPT_target_abi);
+  Opts.RelocationModel =
+      std::string(Args.getLastArgValue(OPT_mrelocation_model, "pic"));
+  Opts.TargetABI = std::string(Args.getLastArgValue(OPT_target_abi));
   Opts.IncrementalLinkerCompatible =
       Args.hasArg(OPT_mincremental_linker_compatible);
   Opts.SymbolDefs = Args.getAllArgValues(OPT_defsym);
@@ -314,12 +322,7 @@ bool ExecuteAssembler(AssemblerInvocation &Opts, DiagnosticsEngine &Diags) {
                             SrcMgr.getMemoryBuffer(BufferIndex)->getBuffer());
 
   // Build up the feature string from the target feature list.
-  std::string FS;
-  if (!Opts.Features.empty()) {
-    FS = Opts.Features[0];
-    for (unsigned i = 1, e = Opts.Features.size(); i != e; ++i)
-      FS += "," + Opts.Features[i];
-  }
+  std::string FS = llvm::join(Opts.Features, ",");
 
   std::unique_ptr<MCStreamer> Str;
 
@@ -383,7 +386,7 @@ bool ExecuteAssembler(AssemblerInvocation &Opts, DiagnosticsEngine &Diags) {
     MCSection *AsmLabel = Ctx.getMachOSection(
         "__LLVM", "__asm", MachO::S_REGULAR, 4, SectionKind::getReadOnly());
     Str.get()->SwitchSection(AsmLabel);
-    Str.get()->EmitZeros(1);
+    Str.get()->emitZeros(1);
   }
 
   // Assembly to object compilation should leverage assembly info.

builder/clang.cpp

@@ -11,6 +11,7 @@
 #include <clang/FrontendTool/Utils.h>
 #include <llvm/ADT/IntrusiveRefCntPtr.h>
 #include <llvm/Option/Option.h>
+#include <llvm/Support/Host.h>
 
 using namespace llvm;
 using namespace clang;

compiler/interface.go

@@ -465,9 +465,7 @@ func (c *compilerContext) getInterfaceInvokeWrapper(f *ir.Function) llvm.Value {
 	paramTypes := append([]llvm.Type{c.i8ptrType}, fnType.ParamTypes()[len(expandedReceiverType):]...)
 	wrapFnType := llvm.FunctionType(fnType.ReturnType(), paramTypes, false)
 	wrapper = llvm.AddFunction(c.mod, wrapperName, wrapFnType)
-	if f.LLVMFn.LastParam().Name() == "parentHandle" {
-		wrapper.LastParam().SetName("parentHandle")
-	}
+	wrapper.LastParam().SetName("parentHandle")
 
 	wrapper.SetLinkage(llvm.InternalLinkage)
 	wrapper.SetUnnamedAddr(true)

compiler/llvmutil/wordpack.go

@@ -26,7 +26,6 @@ func EmitPointerPack(builder llvm.Builder, mod llvm.Module, config *compileopts.
 	packedType := ctx.StructType(valueTypes, false)
 
 	// Allocate memory for the packed data.
-	var packedAlloc, packedHeapAlloc llvm.Value
 	size := targetData.TypeAllocSize(packedType)
 	if size == 0 {
 		return llvm.ConstPointerNull(i8ptrType)
@@ -39,9 +38,39 @@ func EmitPointerPack(builder llvm.Builder, mod llvm.Module, config *compileopts.
 			// Try to keep this cast in SSA form.
 			return builder.CreateIntToPtr(values[0], i8ptrType, "pack.int")
 		}
+
 		// Because packedType is a struct and we have to cast it to a *i8, store
-		// it in an alloca first for bitcasting (store+bitcast+load).
-		packedAlloc, _, _ = CreateTemporaryAlloca(builder, mod, packedType, "")
+		// it in a *i8 alloca first and load the *i8 value from there. This is
+		// effectively a bitcast.
+		packedAlloc, _, _ := CreateTemporaryAlloca(builder, mod, i8ptrType, "")
+
+		if size < targetData.TypeAllocSize(i8ptrType) {
+			// The alloca is bigger than the value that will be stored in it.
+			// To avoid having some bits undefined, zero the alloca first.
+			// Hopefully this will get optimized away.
+			builder.CreateStore(llvm.ConstNull(i8ptrType), packedAlloc)
+		}
+
+		// Store all values in the alloca.
+		packedAllocCast := builder.CreateBitCast(packedAlloc, llvm.PointerType(packedType, 0), "")
+		for i, value := range values {
+			indices := []llvm.Value{
+				llvm.ConstInt(ctx.Int32Type(), 0, false),
+				llvm.ConstInt(ctx.Int32Type(), uint64(i), false),
+			}
+			gep := builder.CreateInBoundsGEP(packedAllocCast, indices, "")
+			builder.CreateStore(value, gep)
+		}
+
+		// Load value (the *i8) from the alloca.
+		result := builder.CreateLoad(packedAlloc, "")
+
+		// End the lifetime of the alloca, to help the optimizer.
+		packedPtr := builder.CreateBitCast(packedAlloc, i8ptrType, "")
+		packedSize := llvm.ConstInt(ctx.Int64Type(), targetData.TypeAllocSize(packedAlloc.Type()), false)
+		EmitLifetimeEnd(builder, mod, packedPtr, packedSize)
+
+		return result
 	} else {
 		// Check if the values are all constants.
 		constant := true
@@ -67,7 +96,7 @@ func EmitPointerPack(builder llvm.Builder, mod llvm.Module, config *compileopts.
 		// Packed data is bigger than a pointer, so allocate it on the heap.
 		sizeValue := llvm.ConstInt(uintptrType, size, false)
 		alloc := mod.NamedFunction("runtime.alloc")
-		packedHeapAlloc = builder.CreateCall(alloc, []llvm.Value{
+		packedHeapAlloc := builder.CreateCall(alloc, []llvm.Value{
 			sizeValue,
 			llvm.Undef(i8ptrType),            // unused context parameter
 			llvm.ConstPointerNull(i8ptrType), // coroutine handle
@@ -80,28 +109,19 @@ func EmitPointerPack(builder llvm.Builder, mod llvm.Module, config *compileopts.
 				llvm.ConstPointerNull(i8ptrType), // coroutine handle
 			}, "")
 		}
-		packedAlloc = builder.CreateBitCast(packedHeapAlloc, llvm.PointerType(packedType, 0), "")
-	}
-	// Store all values in the alloca or heap pointer.
-	for i, value := range values {
-		indices := []llvm.Value{
-			llvm.ConstInt(ctx.Int32Type(), 0, false),
-			llvm.ConstInt(ctx.Int32Type(), uint64(i), false),
-		}
-		gep := builder.CreateInBoundsGEP(packedAlloc, indices, "")
-		builder.CreateStore(value, gep)
-	}
+		packedAlloc := builder.CreateBitCast(packedHeapAlloc, llvm.PointerType(packedType, 0), "")
 
-	if packedHeapAlloc.IsNil() {
-		// Load value (as *i8) from the alloca.
-		packedAlloc = builder.CreateBitCast(packedAlloc, llvm.PointerType(i8ptrType, 0), "")
-		result := builder.CreateLoad(packedAlloc, "")
-		packedPtr := builder.CreateBitCast(packedAlloc, i8ptrType, "")
-		packedSize := llvm.ConstInt(ctx.Int64Type(), targetData.TypeAllocSize(packedAlloc.Type()), false)
-		EmitLifetimeEnd(builder, mod, packedPtr, packedSize)
-		return result
-	} else {
-		// Get the original heap allocation pointer, which already is an *i8.
+		// Store all values in the heap pointer.
+		for i, value := range values {
+			indices := []llvm.Value{
+				llvm.ConstInt(ctx.Int32Type(), 0, false),
+				llvm.ConstInt(ctx.Int32Type(), uint64(i), false),
+			}
+			gep := builder.CreateInBoundsGEP(packedAlloc, indices, "")
+			builder.CreateStore(value, gep)
+		}
+
+		// Return the original heap allocation pointer, which already is an *i8.
 		return packedHeapAlloc
 	}
 }

goenv/version.go

@@ -12,7 +12,7 @@ import (
 
 // Version of TinyGo.
 // Update this value before release of new version of software.
-const Version = "0.16.0"
+const Version = "0.17.0-dev"
 
 // GetGorootVersion returns the major and minor version for a given GOROOT path.
 // If the goroot cannot be determined, (0, 0) is returned.

interp/README.md

@@ -6,50 +6,81 @@ possible and only run unknown expressions (e.g. external calls) at runtime. This
 is in practice a partial evaluator of the `runtime.initAll` function, which
 calls each package initializer.
 
-It works by directly interpreting LLVM IR:
+This package is a rewrite of a previous partial evaluator that worked
+directly on LLVM IR and used the module and LLVM constants as intermediate
+values. This newer version instead uses a mostly Go intermediate form. It
+compiles functions and extracts relevant data first (compiler.go), then
+executes those functions (interpreter.go) in a memory space that can be
+rolled back per function (memory.go). This means that it is not necessary to
+scan functions to see whether they can be run at compile time, which was very
+error prone. Instead it just tries to execute everything and if it hits
+something it cannot interpret (such as a store to memory-mapped I/O) it rolls
+back the execution of that function and runs the function at runtime instead.
+All in all, this design provides several benefits:
 
-  * Almost all operations work directly on constants, and are implemented using
-    the llvm.Const* set of functions that are evaluated directly.
-  * External function calls and some other operations (inline assembly, volatile
-    load, volatile store) are seen as having limited side effects. Limited in
-    the sense that it is known at compile time which globals it affects, which
-    then are marked 'dirty' (meaning, further operations on it must be done at
-    runtime). These operations are emitted directly in the `runtime.initAll`
-    function. Return values are also considered 'dirty'.
-  * Such 'dirty' objects and local values must be executed at runtime instead of
-    at compile time. This dirtyness propagates further through the IR, for
-    example storing a dirty local value to a global also makes the global dirty,
-    meaning that the global may not be read or written at compile time as it's
-    contents at that point during interpretation is unknown.
-  * There are some heuristics in place to avoid doing too much with dirty
-    values. For example, a branch based on a dirty local marks the whole
-    function itself as having side effect (as if it is an external function).
-    However, all globals it touches are still taken into account and when a call
-    is inserted in `runtime.initAll`, all globals it references are also marked
-    dirty.
-  * Heap allocation (`runtime.alloc`) is emulated by creating new objects. The
-    value in the allocation is the initializer of the global, the zero value is
-    the zero initializer.
-  * Stack allocation (`alloca`) is often emulated using a fake alloca object,
-    until the address of the alloca is taken in which case it is also created as
-    a real `alloca` in `runtime.initAll` and marked dirty. This may be necessary
-    when calling an external function with the given alloca as paramter.
+  * Much better error handling. By being able to revert to runtime execution
+    without the need for scanning functions, this version is able to
+    automatically work around many bugs in the previous implementation.
+  * More correct memory model. This is not inherent to the new design, but the
+    new design also made the memory model easier to reason about.
+  * Faster execution of initialization code. While it is not much faster for
+    normal interpretation (maybe 25% or so) due to the compilation overhead,
+    it should be a whole lot faster for loops as it doesn't have to call into
+    LLVM (via CGo) for every operation.
+
+As mentioned, this partial evaulator comes in three parts: a compiler, an
+interpreter, and a memory manager.
+
+## Compiler
+
+The main task of the compiler is that it extracts all necessary data from
+every instruction in a function so that when this instruction is interpreted,
+no additional CGo calls are necessary. This is not currently done for all
+instructions (`runtime.alloc` is a notable exception), but at least it does
+so for the vast majority of instructions.
+
+## Interpreter
+
+The interpreter runs an instruction just like it would if it were executed
+'for real'. The vast majority of instructions can be executed at compile
+time. As indicated above, some instructions need to be executed at runtime
+instead.
+
+## Memory
+
+Memory is represented as objects (the `object` type) that contains data that
+will eventually be stored in a global and values (the `value` interface) that
+can be worked with while running the interpreter. Values therefore are only
+used locally and are always passed by value (just like most LLVM constants)
+while objects represent the backing storage (like LLVM globals). Some values
+are pointer values, and point to an object.
+
+Importantly, this partial evaluator can roll back the execution of a
+function. This is implemented by creating a new memory view per function
+activation, which makes sure that any change to a global (such as a store
+instruction) is stored in the memory view. It creates a copy of the object
+and stores that in the memory view to be modified. Once the function has
+executed successfully, all these modified objects are then copied into the
+parent function, up to the root function invocation which (on successful
+execution) writes the values back into the LLVM module. This way, function
+invocations can be rolled back without leaving a trace.
+
+Pointer values point to memory objects, but not to a particular memory
+object. Every memory object is given an index, and pointers use that index to
+look up the current active object for the pointer to load from or to copy
+when storing to it.
+
+Rolling back a function should roll back everyting, including the few
+instructions emitted at runtime. This is done by treating instructions much
+like memory objects and removing the created instructions when necessary.
 
 ## Why is this necessary?
 
 A partial evaluator is hard to get right, so why go through all the trouble of
 writing one?
 
-The main reason is that the previous attempt wasn't complete and wasn't sound.
-It simply tried to evaluate Go SSA directly, which was good but more difficult
-than necessary. An IR based interpreter needs to understand fewer instructions
-as the LLVM IR simply has less (complex) instructions than Go SSA. Also, LLVM
-provides some useful tools like easily getting all uses of a function or global,
-which Go SSA does not provide.
-
-But why is it necessary at all? The answer is that globals with initializers are
-much easier to optimize by LLVM than initialization code. Also, there are a few
-other benefits:
+The answer is that globals with initializers are much easier to optimize by
+LLVM than initialization code. Also, there are a few other benefits:
 
   * Dead globals are trivial to optimize away.
   * Constant globals are easier to detect. Remember that Go does not have global
@@ -60,5 +91,29 @@ other benefits:
   * Constants are much more efficent on microcontrollers, as they can be
     allocated in flash instead of RAM.
 
+The Go SSA package does not create constant initializers for globals.
+Instead, it emits initialization functions, so if you write the following:
+
+```go
+var foo = []byte{1, 2, 3, 4}
+```
+
+It would generate something like this:
+
+```go
+var foo []byte
+
+func init() {
+    foo = make([]byte, 4)
+    foo[0] = 1
+    foo[1] = 2
+    foo[2] = 3
+    foo[3] = 4
+}
+```
+
+This is of course hugely wasteful, it's much better to create `foo` as a
+global array instead of initializing it at runtime.
+
 For more details, see [this section of the
 documentation](https://tinygo.org/compiler-internals/differences-from-go/).

interp/compiler.go

@@ -0,0 +1,410 @@
+package interp
+
+// This file compiles the LLVM IR to a form that's easy to efficiently
+// interpret.
+
+import (
+	"strings"
+
+	"tinygo.org/x/go-llvm"
+)
+
+// A function is a compiled LLVM function, which means that interpreting it
+// avoids most CGo calls necessary. This is done in a separate step so the
+// result can be cached.
+// Functions are in SSA form, just like the LLVM version if it. The first block
+// (blocks[0]) is the entry block.
+type function struct {
+	llvmFn llvm.Value
+	name   string       // precalculated llvmFn.Name()
+	params []llvm.Value // precalculated llvmFn.Params()
+	blocks []*basicBlock
+	locals map[llvm.Value]int
+}
+
+// basicBlock represents a LLVM basic block and contains a slice of
+// instructions. The last instruction must be a terminator instruction.
+type basicBlock struct {
+	instructions []instruction
+}
+
+// instruction is a precompiled LLVM IR instruction. The operands can be either
+// an already known value (such as literalValue or pointerValue) but can also be
+// the special localValue, which means that the value is a function parameter or
+// is produced by another instruction in the function. In that case, the
+// interpreter will replace the operand with that local value.
+type instruction struct {
+	opcode     llvm.Opcode
+	localIndex int
+	operands   []value
+	llvmInst   llvm.Value
+	name       string
+}
+
+// String returns a nice human-readable version of this instruction.
+func (inst *instruction) String() string {
+	operands := make([]string, len(inst.operands))
+	for i, op := range inst.operands {
+		operands[i] = op.String()
+	}
+
+	name := instructionNameMap[inst.opcode]
+	if name == "" {
+		name = "<unknown op>"
+	}
+	return name + " " + strings.Join(operands, " ")
+}
+
+// compileFunction compiles a given LLVM function to an easier to interpret
+// version of the function. As far as possible, all operands are preprocessed so
+// that the interpreter doesn't have to call into LLVM.
+func (r *runner) compileFunction(llvmFn llvm.Value) *function {
+	fn := &function{
+		llvmFn: llvmFn,
+		name:   llvmFn.Name(),
+		params: llvmFn.Params(),
+		locals: make(map[llvm.Value]int),
+	}
+	if llvmFn.IsDeclaration() {
+		// Nothing to do.
+		return fn
+	}
+
+	for i, param := range fn.params {
+		fn.locals[param] = i
+	}
+
+	// Make a map of all the blocks, to quickly find the block number for a
+	// given branch instruction.
+	blockIndices := make(map[llvm.Value]int)
+	for llvmBB := llvmFn.FirstBasicBlock(); !llvmBB.IsNil(); llvmBB = llvm.NextBasicBlock(llvmBB) {
+		index := len(blockIndices)
+		blockIndices[llvmBB.AsValue()] = index
+	}
+
+	// Compile every block.
+	for llvmBB := llvmFn.FirstBasicBlock(); !llvmBB.IsNil(); llvmBB = llvm.NextBasicBlock(llvmBB) {
+		bb := &basicBlock{}
+		fn.blocks = append(fn.blocks, bb)
+
+		// Compile every instruction in the block.
+		for llvmInst := llvmBB.FirstInstruction(); !llvmInst.IsNil(); llvmInst = llvm.NextInstruction(llvmInst) {
+			// Create instruction skeleton.
+			opcode := llvmInst.InstructionOpcode()
+			inst := instruction{
+				opcode:     opcode,
+				localIndex: len(fn.locals),
+				llvmInst:   llvmInst,
+			}
+			fn.locals[llvmInst] = len(fn.locals)
+
+			// Add operands specific for this instruction.
+			switch opcode {
+			case llvm.Ret:
+				// Return instruction, which can either be a `ret void` (no
+				// return value) or return a value.
+				numOperands := llvmInst.OperandsCount()
+				if numOperands != 0 {
+					inst.operands = []value{
+						r.getValue(llvmInst.Operand(0)),
+					}
+				}
+			case llvm.Br:
+				// Branch instruction. Can be either a conditional branch (with
+				// 3 operands) or unconditional branch (with just one basic
+				// block operand).
+				numOperands := llvmInst.OperandsCount()
+				switch numOperands {
+				case 3:
+					// Conditional jump to one of two blocks. Comparable to an
+					// if/else in procedural languages.
+					thenBB := llvmInst.Operand(2)
+					elseBB := llvmInst.Operand(1)
+					inst.operands = []value{
+						r.getValue(llvmInst.Operand(0)),
+						literalValue{uint32(blockIndices[thenBB])},
+						literalValue{uint32(blockIndices[elseBB])},
+					}
+				case 1:
+					// Unconditional jump to a target basic block. Comparable to
+					// a jump in C and Go.
+					jumpBB := llvmInst.Operand(0)
+					inst.operands = []value{
+						literalValue{uint32(blockIndices[jumpBB])},
+					}
+				default:
+					panic("unknown number of operands")
+				}
+			case llvm.PHI:
+				inst.name = llvmInst.Name()
+				incomingCount := inst.llvmInst.IncomingCount()
+				for i := 0; i < incomingCount; i++ {
+					incomingBB := inst.llvmInst.IncomingBlock(i)
+					incomingValue := inst.llvmInst.IncomingValue(i)
+					inst.operands = append(inst.operands,
+						literalValue{uint32(blockIndices[incomingBB.AsValue()])},
+						r.getValue(incomingValue),
+					)
+				}
+			case llvm.Select:
+				// Select is a special instruction that is much like a ternary
+				// operator. It produces operand 1 or 2 based on the boolean
+				// that is operand 0.
+				inst.name = llvmInst.Name()
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+					r.getValue(llvmInst.Operand(1)),
+					r.getValue(llvmInst.Operand(2)),
+				}
+			case llvm.Call:
+				// Call is a regular function call but could also be a runtime
+				// intrinsic. Some runtime intrinsics are treated specially by
+				// the interpreter, such as runtime.alloc. We don't
+				// differentiate between them here because these calls may also
+				// need to be run at runtime, in which case they should all be
+				// created in the same way.
+				llvmCalledValue := llvmInst.CalledValue()
+				if !llvmCalledValue.IsAFunction().IsNil() {
+					name := llvmCalledValue.Name()
+					if name == "llvm.dbg.value" || strings.HasPrefix(name, "llvm.lifetime.") {
+						// These intrinsics should not be interpreted, they are not
+						// relevant to the execution of this function.
+						continue
+					}
+				}
+				inst.name = llvmInst.Name()
+				numOperands := llvmInst.OperandsCount()
+				inst.operands = append(inst.operands, r.getValue(llvmCalledValue))
+				for i := 0; i < numOperands-1; i++ {
+					inst.operands = append(inst.operands, r.getValue(llvmInst.Operand(i)))
+				}
+			case llvm.Load:
+				// Load instruction. The interpreter will load from the
+				// appropriate memory view.
+				// Also provide the memory size to be loaded, which is necessary
+				// with a lack of type information.
+				inst.name = llvmInst.Name()
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+					literalValue{r.targetData.TypeAllocSize(llvmInst.Type())},
+				}
+			case llvm.Store:
+				// Store instruction. The interpreter will create a new object
+				// in the memory view of the function invocation and store to
+				// that, to make it possible to roll back this store.
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+					r.getValue(llvmInst.Operand(1)),
+				}
+			case llvm.Alloca:
+				// Alloca allocates stack space for local variables.
+				numElements := r.getValue(inst.llvmInst.Operand(0)).(literalValue).value.(uint32)
+				elementSize := r.targetData.TypeAllocSize(inst.llvmInst.Type().ElementType())
+				inst.operands = []value{
+					literalValue{elementSize * uint64(numElements)},
+				}
+			case llvm.GetElementPtr:
+				// GetElementPtr does pointer arithmetic.
+				inst.name = llvmInst.Name()
+				ptr := llvmInst.Operand(0)
+				n := llvmInst.OperandsCount()
+				elementType := ptr.Type().ElementType()
+				// gep: [source ptr, dest value size, pairs of indices...]
+				inst.operands = []value{
+					r.getValue(ptr),
+					literalValue{r.targetData.TypeAllocSize(llvmInst.Type().ElementType())},
+					r.getValue(llvmInst.Operand(1)),
+					literalValue{r.targetData.TypeAllocSize(elementType)},
+				}
+				for i := 2; i < n; i++ {
+					operand := r.getValue(llvmInst.Operand(i))
+					if elementType.TypeKind() == llvm.StructTypeKind {
+						index := operand.(literalValue).value.(uint32)
+						elementOffset := r.targetData.ElementOffset(elementType, int(index))
+						// Encode operands in a special way. The elementOffset
+						// is just the offset in bytes. The elementSize is a
+						// negative number (when cast to a int64) by flipping
+						// all the bits. This allows the interpreter to detect
+						// this is a struct field and that it should not
+						// multiply it with the elementOffset to get the offset.
+						// It is important for the interpreter to know the
+						// struct field index for when the GEP must be done at
+						// runtime.
+						inst.operands = append(inst.operands, literalValue{elementOffset}, literalValue{^uint64(index)})
+						elementType = elementType.StructElementTypes()[index]
+					} else {
+						elementType = elementType.ElementType()
+						elementSize := r.targetData.TypeAllocSize(elementType)
+						elementSizeOperand := literalValue{elementSize}
+						// Add operand * elementSizeOperand bytes to the pointer.
+						inst.operands = append(inst.operands, operand, elementSizeOperand)
+					}
+				}
+			case llvm.BitCast, llvm.IntToPtr, llvm.PtrToInt:
+				// Bitcasts are ususally used to cast a pointer from one type to
+				// another leaving the pointer itself intact.
+				inst.name = llvmInst.Name()
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+				}
+			case llvm.ExtractValue:
+				inst.name = llvmInst.Name()
+				agg := llvmInst.Operand(0)
+				var offset uint64
+				indexingType := agg.Type()
+				for _, index := range inst.llvmInst.Indices() {
+					switch indexingType.TypeKind() {
+					case llvm.StructTypeKind:
+						offset += r.targetData.ElementOffset(indexingType, int(index))
+						indexingType = indexingType.StructElementTypes()[index]
+					default: // ArrayTypeKind
+						indexingType = indexingType.ElementType()
+						elementSize := r.targetData.TypeAllocSize(indexingType)
+						offset += elementSize * uint64(index)
+					}
+				}
+				size := r.targetData.TypeAllocSize(inst.llvmInst.Type())
+				// extractvalue [agg, byteOffset, byteSize]
+				inst.operands = []value{
+					r.getValue(agg),
+					literalValue{offset},
+					literalValue{size},
+				}
+			case llvm.InsertValue:
+				inst.name = llvmInst.Name()
+				agg := llvmInst.Operand(0)
+				var offset uint64
+				indexingType := agg.Type()
+				for _, index := range inst.llvmInst.Indices() {
+					switch indexingType.TypeKind() {
+					case llvm.StructTypeKind:
+						offset += r.targetData.ElementOffset(indexingType, int(index))
+						indexingType = indexingType.StructElementTypes()[index]
+					default: // ArrayTypeKind
+						indexingType = indexingType.ElementType()
+						elementSize := r.targetData.TypeAllocSize(indexingType)
+						offset += elementSize * uint64(index)
+					}
+				}
+				// insertvalue [agg, elt, byteOffset]
+				inst.operands = []value{
+					r.getValue(agg),
+					r.getValue(llvmInst.Operand(1)),
+					literalValue{offset},
+				}
+			case llvm.ICmp:
+				inst.name = llvmInst.Name()
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+					r.getValue(llvmInst.Operand(1)),
+					literalValue{uint8(llvmInst.IntPredicate())},
+				}
+			case llvm.FCmp:
+				inst.name = llvmInst.Name()
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+					r.getValue(llvmInst.Operand(1)),
+					literalValue{uint8(llvmInst.FloatPredicate())},
+				}
+			case llvm.Add, llvm.Sub, llvm.Mul, llvm.UDiv, llvm.SDiv, llvm.URem, llvm.SRem, llvm.Shl, llvm.LShr, llvm.AShr, llvm.And, llvm.Or, llvm.Xor:
+				// Integer binary operations.
+				inst.name = llvmInst.Name()
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+					r.getValue(llvmInst.Operand(1)),
+				}
+			case llvm.SExt, llvm.ZExt, llvm.Trunc:
+				// Extend or shrink an integer size.
+				// No sign extension going on so easy to do.
+				// zext: [value, bitwidth]
+				// trunc: [value, bitwidth]
+				inst.name = llvmInst.Name()
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+					literalValue{uint64(llvmInst.Type().IntTypeWidth())},
+				}
+			case llvm.SIToFP, llvm.UIToFP:
+				// Convert an integer to a floating point instruction.
+				// opcode: [value, bitwidth]
+				inst.name = llvmInst.Name()
+				inst.operands = []value{
+					r.getValue(llvmInst.Operand(0)),
+					literalValue{uint64(r.targetData.TypeAllocSize(llvmInst.Type()) * 8)},
+				}
+			default:
+				// Unknown instruction, which is already set in inst.opcode so
+				// is detectable.
+				// This error is handled when actually trying to interpret this
+				// instruction (to not trigger on code that won't be executed).
+			}
+			bb.instructions = append(bb.instructions, inst)
+		}
+	}
+	return fn
+}
+
+// instructionNameMap maps from instruction opcodes to instruction names. This
+// can be useful for debug logging.
+var instructionNameMap = [...]string{
+	llvm.Ret:         "ret",
+	llvm.Br:          "br",
+	llvm.Switch:      "switch",
+	llvm.IndirectBr:  "indirectbr",
+	llvm.Invoke:      "invoke",
+	llvm.Unreachable: "unreachable",
+
+	// Standard Binary Operators
+	llvm.Add:  "add",
+	llvm.FAdd: "fadd",
+	llvm.Sub:  "sub",
+	llvm.FSub: "fsub",
+	llvm.Mul:  "mul",
+	llvm.FMul: "fmul",
+	llvm.UDiv: "udiv",
+	llvm.SDiv: "sdiv",
+	llvm.FDiv: "fdiv",
+	llvm.URem: "urem",
+	llvm.SRem: "srem",
+	llvm.FRem: "frem",
+
+	// Logical Operators
+	llvm.Shl:  "shl",
+	llvm.LShr: "lshr",
+	llvm.AShr: "ashr",
+	llvm.And:  "and",
+	llvm.Or:   "or",
+	llvm.Xor:  "xor",
+
+	// Memory Operators
+	llvm.Alloca:        "alloca",
+	llvm.Load:          "load",
+	llvm.Store:         "store",
+	llvm.GetElementPtr: "getelementptr",
+
+	// Cast Operators
+	llvm.Trunc:    "trunc",
+	llvm.ZExt:     "zext",
+	llvm.SExt:     "sext",
+	llvm.FPToUI:   "fptoui",
+	llvm.FPToSI:   "fptosi",
+	llvm.UIToFP:   "uitofp",
+	llvm.SIToFP:   "sitofp",
+	llvm.FPTrunc:  "fptrunc",
+	llvm.FPExt:    "fpext",
+	llvm.PtrToInt: "ptrtoint",
+	llvm.IntToPtr: "inttoptr",
+	llvm.BitCast:  "bitcast",
+
+	// Other Operators
+	llvm.ICmp:           "icmp",
+	llvm.FCmp:           "fcmp",
+	llvm.PHI:            "phi",
+	llvm.Call:           "call",
+	llvm.Select:         "select",
+	llvm.VAArg:          "vaarg",
+	llvm.ExtractElement: "extractelement",
+	llvm.InsertElement:  "insertelement",
+	llvm.ShuffleVector:  "shufflevector",
+	llvm.ExtractValue:   "extractvalue",
+	llvm.InsertValue:    "insertvalue",
+}

interp/errors.go

@@ -11,15 +11,17 @@ import (
 	"tinygo.org/x/go-llvm"
 )
 
-// errUnreachable is returned when an unreachable instruction is executed. This
-// error should not be visible outside of the interp package.
-var errUnreachable = &Error{Err: errors.New("interp: unreachable executed")}
+// These errors are expected during normal execution and can be recovered from
+// by running the affected function at runtime instead of compile time.
+var (
+	errIntegerAsPointer       = errors.New("interp: trying to use an integer as a pointer (memory-mapped I/O?)")
+	errUnsupportedInst        = errors.New("interp: unsupported instruction")
+	errUnsupportedRuntimeInst = errors.New("interp: unsupported instruction (to be emitted at runtime)")
+	errMapAlreadyCreated      = errors.New("interp: map already created")
+)
 
-// unsupportedInstructionError returns a new "unsupported instruction" error for
-// the given instruction. It includes source location information, when
-// available.
-func (e *evalPackage) unsupportedInstructionError(inst llvm.Value) *Error {
-	return e.errorAt(inst, errors.New("interp: unsupported instruction"))
+func isRecoverableError(err error) bool {
+	return err == errIntegerAsPointer || err == errUnsupportedInst || err == errUnsupportedRuntimeInst || err == errMapAlreadyCreated
 }
 
 // ErrorLine is one line in a traceback. The position may be missing.
@@ -46,13 +48,13 @@ func (e *Error) Error() string {
 // errorAt returns an error value for the currently interpreted package at the
 // location of the instruction. The location information may not be complete as
 // it depends on debug information in the IR.
-func (e *evalPackage) errorAt(inst llvm.Value, err error) *Error {
-	pos := getPosition(inst)
+func (r *runner) errorAt(inst instruction, err error) *Error {
+	pos := getPosition(inst.llvmInst)
 	return &Error{
-		ImportPath: e.packagePath,
+		ImportPath: r.pkgName,
 		Pos:        pos,
 		Err:        err,
-		Traceback:  []ErrorLine{{pos, inst}},
+		Traceback:  []ErrorLine{{pos, inst.llvmInst}},
 	}
 }
 

interp/frame.go

@@ -1,708 +0,0 @@
-package interp
-
-// This file implements the core interpretation routines, interpreting single
-// functions.
-
-import (
-	"errors"
-	"strings"
-
-	"tinygo.org/x/go-llvm"
-)
-
-type frame struct {
-	*evalPackage
-	fn     llvm.Value
-	locals map[llvm.Value]Value
-}
-
-// evalBasicBlock evaluates a single basic block, returning the return value (if
-// ending with a ret instruction), a list of outgoing basic blocks (if not
-// ending with a ret instruction), or an error on failure.
-// Most of it works at compile time. Some calls get translated into calls to be
-// executed at runtime: calls to functions with side effects, external calls,
-// and operations on the result of such instructions.
-func (fr *frame) evalBasicBlock(bb, incoming llvm.BasicBlock, indent string) (retval Value, outgoing []llvm.Value, err *Error) {
-	for inst := bb.FirstInstruction(); !inst.IsNil(); inst = llvm.NextInstruction(inst) {
-		if fr.Debug {
-			print(indent)
-			inst.Dump()
-			println()
-		}
-		switch {
-		case !inst.IsABinaryOperator().IsNil():
-			lhs := fr.getLocal(inst.Operand(0)).(*LocalValue).Underlying
-			rhs := fr.getLocal(inst.Operand(1)).(*LocalValue).Underlying
-
-			switch inst.InstructionOpcode() {
-			// Standard binary operators
-			case llvm.Add:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateAdd(lhs, rhs, "")}
-			case llvm.FAdd:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFAdd(lhs, rhs, "")}
-			case llvm.Sub:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateSub(lhs, rhs, "")}
-			case llvm.FSub:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFSub(lhs, rhs, "")}
-			case llvm.Mul:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateMul(lhs, rhs, "")}
-			case llvm.FMul:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFMul(lhs, rhs, "")}
-			case llvm.UDiv:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateUDiv(lhs, rhs, "")}
-			case llvm.SDiv:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateSDiv(lhs, rhs, "")}
-			case llvm.FDiv:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFDiv(lhs, rhs, "")}
-			case llvm.URem:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateURem(lhs, rhs, "")}
-			case llvm.SRem:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateSRem(lhs, rhs, "")}
-			case llvm.FRem:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFRem(lhs, rhs, "")}
-
-			// Logical operators
-			case llvm.Shl:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateShl(lhs, rhs, "")}
-			case llvm.LShr:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateLShr(lhs, rhs, "")}
-			case llvm.AShr:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateAShr(lhs, rhs, "")}
-			case llvm.And:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateAnd(lhs, rhs, "")}
-			case llvm.Or:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateOr(lhs, rhs, "")}
-			case llvm.Xor:
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateXor(lhs, rhs, "")}
-
-			default:
-				return nil, nil, fr.unsupportedInstructionError(inst)
-			}
-
-		// Memory operators
-		case !inst.IsAAllocaInst().IsNil():
-			allocType := inst.Type().ElementType()
-			alloca := llvm.AddGlobal(fr.Mod, allocType, fr.packagePath+"$alloca")
-			alloca.SetInitializer(llvm.ConstNull(allocType))
-			alloca.SetLinkage(llvm.InternalLinkage)
-			fr.locals[inst] = &LocalValue{
-				Underlying: alloca,
-				Eval:       fr.Eval,
-			}
-		case !inst.IsALoadInst().IsNil():
-			operand := fr.getLocal(inst.Operand(0)).(*LocalValue)
-			var value llvm.Value
-			if !operand.IsConstant() || inst.IsVolatile() || (!operand.Underlying.IsAConstantExpr().IsNil() && operand.Underlying.Opcode() == llvm.BitCast) {
-				value = fr.builder.CreateLoad(operand.Value(), inst.Name())
-			} else {
-				var err error
-				value, err = operand.Load()
-				if err != nil {
-					return nil, nil, fr.errorAt(inst, err)
-				}
-			}
-			if value.Type() != inst.Type() {
-				return nil, nil, fr.errorAt(inst, errors.New("interp: load: type does not match"))
-			}
-			fr.locals[inst] = fr.getValue(value)
-		case !inst.IsAStoreInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			ptr := fr.getLocal(inst.Operand(1))
-			if inst.IsVolatile() {
-				fr.builder.CreateStore(value.Value(), ptr.Value())
-			} else {
-				err := ptr.Store(value.Value())
-				if err != nil {
-					return nil, nil, fr.errorAt(inst, err)
-				}
-			}
-		case !inst.IsAGetElementPtrInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			llvmIndices := make([]llvm.Value, inst.OperandsCount()-1)
-			for i := range llvmIndices {
-				llvmIndices[i] = inst.Operand(i + 1)
-			}
-			indices := make([]uint32, len(llvmIndices))
-			for i, llvmIndex := range llvmIndices {
-				operand := fr.getLocal(llvmIndex)
-				if !operand.IsConstant() {
-					// Not a constant operation.
-					// This should be detected by the scanner, but isn't at the
-					// moment.
-					return nil, nil, fr.errorAt(inst, errors.New("todo: non-const gep"))
-				}
-				indices[i] = uint32(operand.Value().ZExtValue())
-			}
-			result, err := value.GetElementPtr(indices)
-			if err != nil {
-				return nil, nil, fr.errorAt(inst, err)
-			}
-			if result.Type() != inst.Type() {
-				return nil, nil, fr.errorAt(inst, errors.New("interp: gep: type does not match"))
-			}
-			fr.locals[inst] = result
-
-		// Cast operators
-		case !inst.IsATruncInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateTrunc(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsAZExtInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateZExt(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsASExtInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateSExt(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsAFPToUIInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFPToUI(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsAFPToSIInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFPToSI(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsAUIToFPInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateUIToFP(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsASIToFPInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateSIToFP(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsAFPTruncInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFPTrunc(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsAFPExtInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFPExt(value.(*LocalValue).Value(), inst.Type(), "")}
-		case !inst.IsAPtrToIntInst().IsNil():
-			value := fr.getLocal(inst.Operand(0))
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreatePtrToInt(value.Value(), inst.Type(), "")}
-		case !inst.IsABitCastInst().IsNil() && inst.Type().TypeKind() == llvm.PointerTypeKind:
-			operand := inst.Operand(0)
-			if !operand.IsACallInst().IsNil() {
-				fn := operand.CalledValue()
-				if !fn.IsAFunction().IsNil() && fn.Name() == "runtime.alloc" {
-					continue // special case: bitcast of alloc
-				}
-			}
-			if _, ok := fr.getLocal(operand).(*MapValue); ok {
-				// Special case for runtime.trackPointer calls.
-				// Note: this might not be entirely sound in some rare cases
-				// where the map is stored in a dirty global.
-				uses := getUses(inst)
-				if len(uses) == 1 {
-					use := uses[0]
-					if !use.IsACallInst().IsNil() && !use.CalledValue().IsAFunction().IsNil() && use.CalledValue().Name() == "runtime.trackPointer" {
-						continue
-					}
-				}
-				// It is not possible in Go to bitcast a map value to a pointer.
-				return nil, nil, fr.errorAt(inst, errors.New("unimplemented: bitcast of map"))
-			}
-			value := fr.getLocal(operand).(*LocalValue)
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateBitCast(value.Value(), inst.Type(), "")}
-
-		// Other operators
-		case !inst.IsAICmpInst().IsNil():
-			lhs := fr.getLocal(inst.Operand(0)).(*LocalValue).Underlying
-			rhs := fr.getLocal(inst.Operand(1)).(*LocalValue).Underlying
-			predicate := inst.IntPredicate()
-			if predicate == llvm.IntEQ {
-				var lhsZero, rhsZero bool
-				var ok1, ok2 bool
-				if lhs.Type().TypeKind() == llvm.PointerTypeKind {
-					// Unfortunately, the const propagation in the IR builder
-					// doesn't handle pointer compares of inttoptr values. So we
-					// implement it manually here.
-					lhsZero, ok1 = isPointerNil(lhs)
-					rhsZero, ok2 = isPointerNil(rhs)
-				}
-				if lhs.Type().TypeKind() == llvm.IntegerTypeKind {
-					lhsZero, ok1 = isZero(lhs)
-					rhsZero, ok2 = isZero(rhs)
-				}
-				if ok1 && ok2 {
-					if lhsZero && rhsZero {
-						// Both are zero, so this icmp is always evaluated to true.
-						fr.locals[inst] = &LocalValue{fr.Eval, llvm.ConstInt(fr.Mod.Context().Int1Type(), 1, false)}
-						continue
-					}
-					if lhsZero != rhsZero {
-						// Only one of them is zero, so this comparison must return false.
-						fr.locals[inst] = &LocalValue{fr.Eval, llvm.ConstInt(fr.Mod.Context().Int1Type(), 0, false)}
-						continue
-					}
-				}
-			}
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateICmp(predicate, lhs, rhs, "")}
-		case !inst.IsAFCmpInst().IsNil():
-			lhs := fr.getLocal(inst.Operand(0)).(*LocalValue).Underlying
-			rhs := fr.getLocal(inst.Operand(1)).(*LocalValue).Underlying
-			predicate := inst.FloatPredicate()
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateFCmp(predicate, lhs, rhs, "")}
-		case !inst.IsAPHINode().IsNil():
-			for i := 0; i < inst.IncomingCount(); i++ {
-				if inst.IncomingBlock(i) == incoming {
-					fr.locals[inst] = fr.getLocal(inst.IncomingValue(i))
-				}
-			}
-		case !inst.IsACallInst().IsNil():
-			callee := inst.CalledValue()
-			switch {
-			case callee.Name() == "runtime.alloc":
-				// heap allocation
-				users := getUses(inst)
-				var resultInst = inst
-				if len(users) == 1 && !users[0].IsABitCastInst().IsNil() {
-					// happens when allocating something other than i8*
-					resultInst = users[0]
-				}
-				size := fr.getLocal(inst.Operand(0)).(*LocalValue).Underlying.ZExtValue()
-				allocType := resultInst.Type().ElementType()
-				typeSize := fr.TargetData.TypeAllocSize(allocType)
-				elementCount := 1
-				if size != typeSize {
-					// allocate an array
-					if size%typeSize != 0 {
-						return nil, nil, fr.unsupportedInstructionError(inst)
-					}
-					elementCount = int(size / typeSize)
-					allocType = llvm.ArrayType(allocType, elementCount)
-				}
-				alloc := llvm.AddGlobal(fr.Mod, allocType, fr.packagePath+"$alloc")
-				alloc.SetInitializer(llvm.ConstNull(allocType))
-				alloc.SetLinkage(llvm.InternalLinkage)
-				result := &LocalValue{
-					Underlying: alloc,
-					Eval:       fr.Eval,
-				}
-				if elementCount == 1 {
-					fr.locals[resultInst] = result
-				} else {
-					result, err := result.GetElementPtr([]uint32{0, 0})
-					if err != nil {
-						return nil, nil, fr.errorAt(inst, err)
-					}
-					fr.locals[resultInst] = result
-				}
-			case callee.Name() == "runtime.hashmapMake":
-				// create a map
-				keySize := inst.Operand(0).ZExtValue()
-				valueSize := inst.Operand(1).ZExtValue()
-				fr.locals[inst] = &MapValue{
-					Eval:      fr.Eval,
-					PkgName:   fr.packagePath,
-					KeySize:   int(keySize),
-					ValueSize: int(valueSize),
-				}
-			case callee.Name() == "runtime.hashmapStringSet":
-				// set a string key in the map
-				keyBuf := fr.getLocal(inst.Operand(1)).(*LocalValue)
-				keyLen := fr.getLocal(inst.Operand(2)).(*LocalValue)
-				valPtr := fr.getLocal(inst.Operand(3)).(*LocalValue)
-				m, ok := fr.getLocal(inst.Operand(0)).(*MapValue)
-				if !ok || !keyBuf.IsConstant() || !keyLen.IsConstant() || !valPtr.IsConstant() {
-					// The mapassign operation could not be done at compile
-					// time. Do it at runtime instead.
-					m := fr.getLocal(inst.Operand(0)).Value()
-					fr.markDirty(m)
-					llvmParams := []llvm.Value{
-						m,                                    // *runtime.hashmap
-						fr.getLocal(inst.Operand(1)).Value(), // key.ptr
-						fr.getLocal(inst.Operand(2)).Value(), // key.len
-						fr.getLocal(inst.Operand(3)).Value(), // value (unsafe.Pointer)
-						fr.getLocal(inst.Operand(4)).Value(), // context
-						fr.getLocal(inst.Operand(5)).Value(), // parentHandle
-					}
-					fr.builder.CreateCall(callee, llvmParams, "")
-					continue
-				}
-				// "key" is a Go string value, which in the TinyGo calling convention is split up
-				// into separate pointer and length parameters.
-				err := m.PutString(keyBuf, keyLen, valPtr)
-				if err != nil {
-					return nil, nil, fr.errorAt(inst, err)
-				}
-			case callee.Name() == "runtime.hashmapBinarySet":
-				// set a binary (int etc.) key in the map
-				keyBuf := fr.getLocal(inst.Operand(1)).(*LocalValue)
-				valPtr := fr.getLocal(inst.Operand(2)).(*LocalValue)
-				m, ok := fr.getLocal(inst.Operand(0)).(*MapValue)
-				if !ok || !keyBuf.IsConstant() || !valPtr.IsConstant() {
-					// The mapassign operation could not be done at compile
-					// time. Do it at runtime instead.
-					m := fr.getLocal(inst.Operand(0)).Value()
-					fr.markDirty(m)
-					llvmParams := []llvm.Value{
-						m,                                    // *runtime.hashmap
-						fr.getLocal(inst.Operand(1)).Value(), // key
-						fr.getLocal(inst.Operand(2)).Value(), // value
-						fr.getLocal(inst.Operand(3)).Value(), // context
-						fr.getLocal(inst.Operand(4)).Value(), // parentHandle
-					}
-					fr.builder.CreateCall(callee, llvmParams, "")
-					continue
-				}
-				err := m.PutBinary(keyBuf, valPtr)
-				if err != nil {
-					return nil, nil, fr.errorAt(inst, err)
-				}
-			case callee.Name() == "runtime.stringConcat":
-				// adding two strings together
-				buf1Ptr := fr.getLocal(inst.Operand(0))
-				buf1Len := fr.getLocal(inst.Operand(1))
-				buf2Ptr := fr.getLocal(inst.Operand(2))
-				buf2Len := fr.getLocal(inst.Operand(3))
-				buf1, err := getStringBytes(buf1Ptr, buf1Len.Value())
-				if err != nil {
-					return nil, nil, fr.errorAt(inst, err)
-				}
-				buf2, err := getStringBytes(buf2Ptr, buf2Len.Value())
-				if err != nil {
-					return nil, nil, fr.errorAt(inst, err)
-				}
-				result := []byte(string(buf1) + string(buf2))
-				vals := make([]llvm.Value, len(result))
-				for i := range vals {
-					vals[i] = llvm.ConstInt(fr.Mod.Context().Int8Type(), uint64(result[i]), false)
-				}
-				globalType := llvm.ArrayType(fr.Mod.Context().Int8Type(), len(result))
-				globalValue := llvm.ConstArray(fr.Mod.Context().Int8Type(), vals)
-				global := llvm.AddGlobal(fr.Mod, globalType, fr.packagePath+"$stringconcat")
-				global.SetInitializer(globalValue)
-				global.SetLinkage(llvm.InternalLinkage)
-				global.SetGlobalConstant(true)
-				global.SetUnnamedAddr(true)
-				stringType := fr.Mod.GetTypeByName("runtime._string")
-				retPtr := llvm.ConstGEP(global, getLLVMIndices(fr.Mod.Context().Int32Type(), []uint32{0, 0}))
-				retLen := llvm.ConstInt(stringType.StructElementTypes()[1], uint64(len(result)), false)
-				ret := llvm.ConstNull(stringType)
-				ret = llvm.ConstInsertValue(ret, retPtr, []uint32{0})
-				ret = llvm.ConstInsertValue(ret, retLen, []uint32{1})
-				fr.locals[inst] = &LocalValue{fr.Eval, ret}
-			case callee.Name() == "runtime.sliceCopy":
-				elementSize := fr.getLocal(inst.Operand(4)).(*LocalValue).Value().ZExtValue()
-				dstArray := fr.getLocal(inst.Operand(0)).(*LocalValue).stripPointerCasts()
-				srcArray := fr.getLocal(inst.Operand(1)).(*LocalValue).stripPointerCasts()
-				dstLen := fr.getLocal(inst.Operand(2)).(*LocalValue)
-				srcLen := fr.getLocal(inst.Operand(3)).(*LocalValue)
-				if elementSize != 1 && dstArray.Type().ElementType().TypeKind() == llvm.ArrayTypeKind && srcArray.Type().ElementType().TypeKind() == llvm.ArrayTypeKind {
-					// Slice data pointers are created by adding a global array
-					// and getting the address of the first element using a GEP.
-					// However, before the compiler can pass it to
-					// runtime.sliceCopy, it has to perform a bitcast to a *i8,
-					// to make it a unsafe.Pointer. Now, when the IR builder
-					// sees a bitcast of a GEP with zero indices, it will make
-					// a bitcast of the original array instead of the GEP,
-					// which breaks our assumptions.
-					// Re-add this GEP, in the hope that it it is then of the correct type...
-					dstArrayValue, err := dstArray.GetElementPtr([]uint32{0, 0})
-					if err != nil {
-						return nil, nil, fr.errorAt(inst, err)
-					}
-					dstArray = dstArrayValue.(*LocalValue)
-					srcArrayValue, err := srcArray.GetElementPtr([]uint32{0, 0})
-					if err != nil {
-						return nil, nil, fr.errorAt(inst, err)
-					}
-					srcArray = srcArrayValue.(*LocalValue)
-				}
-				if fr.Eval.TargetData.TypeAllocSize(dstArray.Type().ElementType()) != elementSize {
-					return nil, nil, fr.errorAt(inst, errors.New("interp: slice dst element size does not match pointer type"))
-				}
-				if fr.Eval.TargetData.TypeAllocSize(srcArray.Type().ElementType()) != elementSize {
-					return nil, nil, fr.errorAt(inst, errors.New("interp: slice src element size does not match pointer type"))
-				}
-				if dstArray.Type() != srcArray.Type() {
-					return nil, nil, fr.errorAt(inst, errors.New("interp: slice element types don't match"))
-				}
-				length := dstLen.Value().SExtValue()
-				if srcLength := srcLen.Value().SExtValue(); srcLength < length {
-					length = srcLength
-				}
-				if length < 0 {
-					return nil, nil, fr.errorAt(inst, errors.New("interp: trying to copy a slice with negative length?"))
-				}
-				for i := int64(0); i < length; i++ {
-					// *dst = *src
-					val, err := srcArray.Load()
-					if err != nil {
-						return nil, nil, fr.errorAt(inst, err)
-					}
-					err = dstArray.Store(val)
-					if err != nil {
-						return nil, nil, fr.errorAt(inst, err)
-					}
-					// dst++
-					dstArrayValue, err := dstArray.GetElementPtr([]uint32{1})
-					if err != nil {
-						return nil, nil, fr.errorAt(inst, err)
-					}
-					dstArray = dstArrayValue.(*LocalValue)
-					// src++
-					srcArrayValue, err := srcArray.GetElementPtr([]uint32{1})
-					if err != nil {
-						return nil, nil, fr.errorAt(inst, err)
-					}
-					srcArray = srcArrayValue.(*LocalValue)
-				}
-			case callee.Name() == "runtime.stringToBytes":
-				// convert a string to a []byte
-				bufPtr := fr.getLocal(inst.Operand(0))
-				bufLen := fr.getLocal(inst.Operand(1))
-				result, err := getStringBytes(bufPtr, bufLen.Value())
-				if err != nil {
-					return nil, nil, fr.errorAt(inst, err)
-				}
-				vals := make([]llvm.Value, len(result))
-				for i := range vals {
-					vals[i] = llvm.ConstInt(fr.Mod.Context().Int8Type(), uint64(result[i]), false)
-				}
-				globalType := llvm.ArrayType(fr.Mod.Context().Int8Type(), len(result))
-				globalValue := llvm.ConstArray(fr.Mod.Context().Int8Type(), vals)
-				global := llvm.AddGlobal(fr.Mod, globalType, fr.packagePath+"$bytes")
-				global.SetInitializer(globalValue)
-				global.SetLinkage(llvm.InternalLinkage)
-				global.SetGlobalConstant(true)
-				global.SetUnnamedAddr(true)
-				sliceType := inst.Type()
-				retPtr := llvm.ConstGEP(global, getLLVMIndices(fr.Mod.Context().Int32Type(), []uint32{0, 0}))
-				retLen := llvm.ConstInt(sliceType.StructElementTypes()[1], uint64(len(result)), false)
-				ret := llvm.ConstNull(sliceType)
-				ret = llvm.ConstInsertValue(ret, retPtr, []uint32{0}) // ptr
-				ret = llvm.ConstInsertValue(ret, retLen, []uint32{1}) // len
-				ret = llvm.ConstInsertValue(ret, retLen, []uint32{2}) // cap
-				fr.locals[inst] = &LocalValue{fr.Eval, ret}
-			case callee.Name() == "runtime.typeAssert":
-				actualTypeInt := fr.getLocal(inst.Operand(0)).(*LocalValue).Underlying
-				assertedType := fr.getLocal(inst.Operand(1)).(*LocalValue).Underlying
-				if actualTypeInt.IsAConstantExpr().IsNil() || actualTypeInt.Opcode() != llvm.PtrToInt {
-					return nil, nil, fr.errorAt(inst, errors.New("interp: expected typecode in runtime.typeAssert to be a ptrtoint"))
-				}
-				actualType := actualTypeInt.Operand(0)
-				if actualType.IsAConstant().IsNil() || assertedType.IsAConstant().IsNil() {
-					return nil, nil, fr.errorAt(inst, errors.New("interp: unimplemented: type assert with non-constant interface value"))
-				}
-				assertOk := uint64(0)
-				if llvm.ConstExtractValue(actualType.Initializer(), []uint32{0}) == assertedType {
-					assertOk = 1
-				}
-				fr.locals[inst] = &LocalValue{fr.Eval, llvm.ConstInt(fr.Mod.Context().Int1Type(), assertOk, false)}
-			case callee.Name() == "runtime.interfaceImplements":
-				typecode := fr.getLocal(inst.Operand(0)).(*LocalValue).Underlying
-				interfaceMethodSet := fr.getLocal(inst.Operand(1)).(*LocalValue).Underlying
-				if typecode.IsAConstantExpr().IsNil() || typecode.Opcode() != llvm.PtrToInt {
-					return nil, nil, fr.errorAt(inst, errors.New("interp: expected typecode to be a ptrtoint"))
-				}
-				typecode = typecode.Operand(0)
-				if interfaceMethodSet.IsAConstantExpr().IsNil() || interfaceMethodSet.Opcode() != llvm.GetElementPtr {
-					return nil, nil, fr.errorAt(inst, errors.New("interp: expected method set in runtime.interfaceImplements to be a constant gep"))
-				}
-				interfaceMethodSet = interfaceMethodSet.Operand(0).Initializer()
-				methodSet := llvm.ConstExtractValue(typecode.Initializer(), []uint32{1})
-				if methodSet.IsAConstantExpr().IsNil() || methodSet.Opcode() != llvm.GetElementPtr {
-					return nil, nil, fr.errorAt(inst, errors.New("interp: expected method set to be a constant gep"))
-				}
-				methodSet = methodSet.Operand(0).Initializer()
-
-				// Make a set of all the methods on the concrete type, for
-				// easier checking in the next step.
-				definedMethods := map[string]struct{}{}
-				for i := 0; i < methodSet.Type().ArrayLength(); i++ {
-					methodInfo := llvm.ConstExtractValue(methodSet, []uint32{uint32(i)})
-					name := llvm.ConstExtractValue(methodInfo, []uint32{0}).Name()
-					definedMethods[name] = struct{}{}
-				}
-				// Check whether all interface methods are also in the list
-				// of defined methods calculated above.
-				implements := uint64(1) // i1 true
-				for i := 0; i < interfaceMethodSet.Type().ArrayLength(); i++ {
-					name := llvm.ConstExtractValue(interfaceMethodSet, []uint32{uint32(i)}).Name()
-					if _, ok := definedMethods[name]; !ok {
-						// There is a method on the interface that is not
-						// implemented by the type.
-						implements = 0 // i1 false
-						break
-					}
-				}
-				fr.locals[inst] = &LocalValue{fr.Eval, llvm.ConstInt(fr.Mod.Context().Int1Type(), implements, false)}
-			case callee.Name() == "runtime.nanotime":
-				fr.locals[inst] = &LocalValue{fr.Eval, llvm.ConstInt(fr.Mod.Context().Int64Type(), 0, false)}
-			case callee.Name() == "llvm.dbg.value":
-				// do nothing
-			case strings.HasPrefix(callee.Name(), "llvm.lifetime."):
-				// do nothing
-			case callee.Name() == "runtime.trackPointer":
-				// do nothing
-			case strings.HasPrefix(callee.Name(), "runtime.print") || callee.Name() == "runtime._panic":
-				// This are all print instructions, which necessarily have side
-				// effects but no results.
-				// TODO: print an error when executing runtime._panic (with the
-				// exact error message it would print at runtime).
-				var params []llvm.Value
-				for i := 0; i < inst.OperandsCount()-1; i++ {
-					operand := fr.getLocal(inst.Operand(i)).Value()
-					fr.markDirty(operand)
-					params = append(params, operand)
-				}
-				// TODO: accurate debug info, including call chain
-				fr.builder.CreateCall(callee, params, inst.Name())
-			case !callee.IsAFunction().IsNil() && callee.IsDeclaration():
-				// external functions
-				var params []llvm.Value
-				for i := 0; i < inst.OperandsCount()-1; i++ {
-					operand := fr.getLocal(inst.Operand(i)).Value()
-					fr.markDirty(operand)
-					params = append(params, operand)
-				}
-				// TODO: accurate debug info, including call chain
-				result := fr.builder.CreateCall(callee, params, inst.Name())
-				if inst.Type().TypeKind() != llvm.VoidTypeKind {
-					fr.markDirty(result)
-					fr.locals[inst] = &LocalValue{fr.Eval, result}
-				}
-			case !callee.IsAFunction().IsNil():
-				// regular function
-				var params []Value
-				dirtyParams := false
-				for i := 0; i < inst.OperandsCount()-1; i++ {
-					local := fr.getLocal(inst.Operand(i))
-					if !local.IsConstant() {
-						dirtyParams = true
-					}
-					params = append(params, local)
-				}
-				var ret Value
-				scanResult, err := fr.hasSideEffects(callee)
-				if err != nil {
-					return nil, nil, err
-				}
-				if scanResult.severity == sideEffectLimited || dirtyParams && scanResult.severity != sideEffectAll {
-					// Side effect is bounded. This means the operation invokes
-					// side effects (like calling an external function) but it
-					// is known at compile time which side effects it invokes.
-					// This means the function can be called at runtime and the
-					// affected globals can be marked dirty at compile time.
-					llvmParams := make([]llvm.Value, len(params))
-					for i, param := range params {
-						llvmParams[i] = param.Value()
-					}
-					result := fr.builder.CreateCall(callee, llvmParams, inst.Name())
-					ret = &LocalValue{fr.Eval, result}
-					// mark all mentioned globals as dirty
-					for global := range scanResult.mentionsGlobals {
-						fr.markDirty(global)
-					}
-				} else {
-					// Side effect is one of:
-					//   * None: no side effects, can be fully interpreted at
-					//     compile time.
-					//   * Unbounded: cannot call at runtime so we'll try to
-					//     interpret anyway and hope for the best.
-					ret, err = fr.function(callee, params, indent+"    ")
-					if err != nil {
-						// Record this function call in the backtrace.
-						err.Traceback = append(err.Traceback, ErrorLine{
-							Pos:  getPosition(inst),
-							Inst: inst,
-						})
-						return nil, nil, err
-					}
-				}
-				if inst.Type().TypeKind() != llvm.VoidTypeKind {
-					fr.locals[inst] = ret
-				}
-			default:
-				// function pointers, etc.
-				return nil, nil, fr.unsupportedInstructionError(inst)
-			}
-		case !inst.IsAExtractValueInst().IsNil():
-			agg := fr.getLocal(inst.Operand(0)).(*LocalValue) // must be constant
-			indices := inst.Indices()
-			if agg.Underlying.IsConstant() {
-				newValue := llvm.ConstExtractValue(agg.Underlying, indices)
-				fr.locals[inst] = fr.getValue(newValue)
-			} else {
-				if len(indices) != 1 {
-					return nil, nil, fr.errorAt(inst, errors.New("interp: cannot handle extractvalue with not exactly 1 index"))
-				}
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateExtractValue(agg.Underlying, int(indices[0]), inst.Name())}
-			}
-		case !inst.IsAInsertValueInst().IsNil():
-			agg := fr.getLocal(inst.Operand(0)).(*LocalValue) // must be constant
-			val := fr.getLocal(inst.Operand(1))
-			indices := inst.Indices()
-			if agg.IsConstant() && val.IsConstant() {
-				newValue := llvm.ConstInsertValue(agg.Underlying, val.Value(), indices)
-				fr.locals[inst] = &LocalValue{fr.Eval, newValue}
-			} else {
-				if len(indices) != 1 {
-					return nil, nil, fr.errorAt(inst, errors.New("interp: cannot handle insertvalue with not exactly 1 index"))
-				}
-				fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateInsertValue(agg.Underlying, val.Value(), int(indices[0]), inst.Name())}
-			}
-		case !inst.IsASelectInst().IsNil():
-			// var result T
-			// if cond {
-			//   result = x
-			// } else {
-			//   result = y
-			// }
-			// return result
-			cond := fr.getLocal(inst.Operand(0)).(*LocalValue).Underlying
-			x := fr.getLocal(inst.Operand(1)).(*LocalValue).Underlying
-			y := fr.getLocal(inst.Operand(2)).(*LocalValue).Underlying
-			fr.locals[inst] = &LocalValue{fr.Eval, fr.builder.CreateSelect(cond, x, y, "")}
-
-		case !inst.IsAReturnInst().IsNil() && inst.OperandsCount() == 0:
-			return nil, nil, nil // ret void
-		case !inst.IsAReturnInst().IsNil() && inst.OperandsCount() == 1:
-			return fr.getLocal(inst.Operand(0)), nil, nil
-		case !inst.IsABranchInst().IsNil() && inst.OperandsCount() == 3:
-			// conditional branch (if/then/else)
-			cond := fr.getLocal(inst.Operand(0)).Value()
-			if cond.Type() != fr.Mod.Context().Int1Type() {
-				return nil, nil, fr.errorAt(inst, errors.New("expected an i1 in a branch instruction"))
-			}
-			thenBB := inst.Operand(1)
-			elseBB := inst.Operand(2)
-			if !cond.IsAInstruction().IsNil() {
-				return nil, nil, fr.errorAt(inst, errors.New("interp: branch on a non-constant"))
-			}
-			if !cond.IsAConstantExpr().IsNil() {
-				// This may happen when the instruction builder could not
-				// const-fold some instructions.
-				return nil, nil, fr.errorAt(inst, errors.New("interp: branch on a non-const-propagated constant expression"))
-			}
-			switch cond {
-			case llvm.ConstInt(fr.Mod.Context().Int1Type(), 0, false): // false
-				return nil, []llvm.Value{thenBB}, nil // then
-			case llvm.ConstInt(fr.Mod.Context().Int1Type(), 1, false): // true
-				return nil, []llvm.Value{elseBB}, nil // else
-			default:
-				return nil, nil, fr.errorAt(inst, errors.New("branch was not true or false"))
-			}
-		case !inst.IsABranchInst().IsNil() && inst.OperandsCount() == 1:
-			// unconditional branch (goto)
-			return nil, []llvm.Value{inst.Operand(0)}, nil
-		case !inst.IsAUnreachableInst().IsNil():
-			// Unreachable was reached (e.g. after a call to panic()).
-			// Report this as an error, as it is not supposed to happen.
-			// This is a sentinel error value.
-			return nil, nil, errUnreachable
-
-		default:
-			return nil, nil, fr.unsupportedInstructionError(inst)
-		}
-	}
-
-	panic("interp: reached end of basic block without terminator")
-}
-
-// Get the Value for an operand, which is a constant value of some sort.
-func (fr *frame) getLocal(v llvm.Value) Value {
-	if ret, ok := fr.locals[v]; ok {
-		return ret
-	} else if value := fr.getValue(v); value != nil {
-		return value
-	} else {
-		// This should not happen under normal circumstances.
-		panic("cannot find value")
-	}
-}

interp/interp.go

@@ -1,59 +1,77 @@
-// Package interp interprets Go package initializers as much as possible. This
-// avoid running them at runtime, improving code size and making other
-// optimizations possible.
+// Package interp is a partial evaluator of code run at package init time. See
+// the README in this package for details.
 package interp
 
-// This file provides the overarching Eval object with associated (utility)
-// methods.
-
 import (
+	"fmt"
+	"os"
 	"strings"
+	"time"
 
 	"tinygo.org/x/go-llvm"
 )
 
-type Eval struct {
-	Mod             llvm.Module
-	TargetData      llvm.TargetData
-	Debug           bool
-	builder         llvm.Builder
-	dirtyGlobals    map[llvm.Value]struct{}
-	sideEffectFuncs map[llvm.Value]*sideEffectResult // cache of side effect scan results
+// 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
+
+// runner contains all state related to one interp run.
+type runner struct {
+	mod           llvm.Module
+	targetData    llvm.TargetData
+	builder       llvm.Builder
+	pointerSize   uint32                   // cached pointer size from the TargetData
+	i8ptrType     llvm.Type                // often used type so created in advance
+	maxAlign      int                      // maximum alignment of an object, alignment of runtime.alloc() result
+	debug         bool                     // log debug messages
+	pkgName       string                   // package name of the currently executing package
+	functionCache map[llvm.Value]*function // cache of compiled functions
+	objects       []object                 // slice of objects in memory
+	globals       map[llvm.Value]int       // map from global to index in objects slice
+	start         time.Time
+	callsExecuted uint64
 }
 
-// evalPackage encapsulates the Eval type for just a single package. The Eval
-// type keeps state across the whole program, the evalPackage type keeps extra
-// state for the currently interpreted package.
-type evalPackage struct {
-	*Eval
-	packagePath string
-}
-
-// Run evaluates the function with the given name and then eliminates all
-// callers.
+// 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 {
-	if debug {
-		println("\ncompile-time evaluation:")
+	r := runner{
+		mod:           mod,
+		targetData:    llvm.NewTargetData(mod.DataLayout()),
+		debug:         debug,
+		functionCache: make(map[llvm.Value]*function),
+		objects:       []object{{}},
+		globals:       make(map[llvm.Value]int),
+		start:         time.Now(),
 	}
+	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)
 
-	name := "runtime.initAll"
-	e := &Eval{
-		Mod:          mod,
-		TargetData:   llvm.NewTargetData(mod.DataLayout()),
-		Debug:        debug,
-		dirtyGlobals: map[llvm.Value]struct{}{},
-	}
-	e.builder = mod.Context().NewBuilder()
-
-	initAll := mod.NamedFunction(name)
+	initAll := mod.NamedFunction("runtime.initAll")
 	bb := initAll.EntryBasicBlock()
+
+	// Create a builder, to insert instructions that could not be evaluated at
+	// compile time.
+	r.builder = mod.Context().NewBuilder()
+	defer r.builder.Dispose()
+
 	// Create a dummy alloca in the entry block that we can set the insert point
 	// to. This is necessary because otherwise we might be removing the
 	// instruction (init call) that we are removing after successful
 	// interpretation.
-	e.builder.SetInsertPointBefore(bb.FirstInstruction())
-	dummy := e.builder.CreateAlloca(e.Mod.Context().Int8Type(), "dummy")
-	e.builder.SetInsertPointBefore(dummy)
+	r.builder.SetInsertPointBefore(bb.FirstInstruction())
+	dummy := r.builder.CreateAlloca(r.mod.Context().Int8Type(), "dummy")
+	r.builder.SetInsertPointBefore(dummy)
+	defer dummy.EraseFromParentAsInstruction()
+
+	// Get a list if init calls. A runtime.initAll might look something like this:
+	// func initAll() {
+	//     unsafe.init()
+	//     machine.init()
+	//     runtime.init()
+	// }
+	// This function gets a list of these call instructions.
 	var initCalls []llvm.Value
 	for inst := bb.FirstInstruction(); !inst.IsNil(); inst = llvm.NextInstruction(inst) {
 		if inst == dummy {
@@ -63,99 +81,67 @@ func Run(mod llvm.Module, debug bool) error {
 			break // ret void
 		}
 		if inst.IsACallInst().IsNil() || inst.CalledValue().IsAFunction().IsNil() {
-			return errorAt(inst, "interp: expected all instructions in "+name+" to be direct calls")
+			return errorAt(inst, "interp: expected all instructions in "+initAll.Name()+" to be direct calls")
 		}
 		initCalls = append(initCalls, inst)
 	}
 
-	// Do this in a separate step to avoid corrupting the iterator above.
-	undefPtr := llvm.Undef(llvm.PointerType(mod.Context().Int8Type(), 0))
+	// Run initializers for each package. Once the package initializer is
+	// finished, the call to the package initializer can be removed.
 	for _, call := range initCalls {
 		initName := call.CalledValue().Name()
 		if !strings.HasSuffix(initName, ".init") {
-			return errorAt(call, "interp: expected all instructions in "+name+" to be *.init() calls")
+			return errorAt(call, "interp: expected all instructions in "+initAll.Name()+" to be *.init() calls")
 		}
-		pkgName := initName[:len(initName)-5]
+		r.pkgName = initName[:len(initName)-len(".init")]
 		fn := call.CalledValue()
+		if r.debug {
+			fmt.Fprintln(os.Stderr, "call:", fn.Name())
+		}
+		_, mem, callErr := r.run(r.getFunction(fn), nil, nil, "    ")
+		if callErr != nil {
+			if isRecoverableError(callErr.Err) {
+				if r.debug {
+					fmt.Fprintln(os.Stderr, "not interpreting", r.pkgName, "because of error:", callErr.Error())
+				}
+				mem.revert()
+				continue
+			}
+			return callErr
+		}
 		call.EraseFromParentAsInstruction()
-		evalPkg := evalPackage{
-			Eval:        e,
-			packagePath: pkgName,
+		for index, obj := range mem.objects {
+			r.objects[index] = obj
 		}
-		_, err := evalPkg.function(fn, []Value{&LocalValue{e, undefPtr}, &LocalValue{e, undefPtr}}, "")
-		if err == errUnreachable {
-			break
+	}
+	r.pkgName = ""
+
+	// Update all global variables in the LLVM module.
+	mem := memoryView{r: &r}
+	for _, obj := range r.objects {
+		if obj.llvmGlobal.IsNil() {
+			continue
 		}
-		if err != nil {
-			return err
+		if obj.buffer == nil {
+			continue
 		}
+		initializer := obj.buffer.toLLVMValue(obj.llvmGlobal.Type().ElementType(), &mem)
+		if checks && initializer.Type() != obj.llvmGlobal.Type().ElementType() {
+			panic("initializer type mismatch")
+		}
+		obj.llvmGlobal.SetInitializer(initializer)
 	}
 
 	return nil
 }
 
-// function interprets the given function. The params are the function params
-// and the indent is the string indentation to use when dumping all interpreted
-// instructions.
-func (e *evalPackage) function(fn llvm.Value, params []Value, indent string) (Value, *Error) {
-	fr := frame{
-		evalPackage: e,
-		fn:          fn,
-		locals:      make(map[llvm.Value]Value),
-	}
-	for i, param := range fn.Params() {
-		fr.locals[param] = params[i]
-	}
-
-	bb := fn.EntryBasicBlock()
-	var lastBB llvm.BasicBlock
-	for {
-		retval, outgoing, err := fr.evalBasicBlock(bb, lastBB, indent)
-		if outgoing == nil {
-			// returned something (a value or void, or an error)
-			return retval, err
-		}
-		if len(outgoing) > 1 {
-			panic("unimplemented: multiple outgoing blocks")
-		}
-		next := outgoing[0]
-		if next.IsABasicBlock().IsNil() {
-			panic("did not switch to a basic block")
-		}
-		lastBB = bb
-		bb = next.AsBasicBlock()
-	}
-}
-
-// getValue determines what kind of LLVM value it gets and returns the
-// appropriate Value type.
-func (e *Eval) getValue(v llvm.Value) Value {
-	return &LocalValue{e, v}
-}
-
-// markDirty marks the passed-in LLVM value dirty, recursively. For example,
-// when it encounters a constant GEP on a global, it marks the global dirty.
-func (e *Eval) markDirty(v llvm.Value) {
-	if !v.IsAGlobalVariable().IsNil() {
-		if v.IsGlobalConstant() {
-			return
-		}
-		if _, ok := e.dirtyGlobals[v]; !ok {
-			e.dirtyGlobals[v] = struct{}{}
-			e.sideEffectFuncs = nil // re-calculate all side effects
-		}
-	} else if v.IsConstant() {
-		if v.OperandsCount() >= 2 && !v.Operand(0).IsAGlobalVariable().IsNil() {
-			// looks like a constant getelementptr of a global.
-			// TODO: find a way to make sure it really is: v.Opcode() returns 0.
-			e.markDirty(v.Operand(0))
-			return
-		}
-		return // nothing to mark
-	} else if !v.IsAGetElementPtrInst().IsNil() {
-		panic("interp: todo: GEP")
-	} else {
-		// Not constant and not a global or GEP so doesn't have to be marked
-		// non-constant.
+// 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 {
+	if fn, ok := r.functionCache[llvmFn]; ok {
+		return fn
 	}
+	fn := r.compileFunction(llvmFn)
+	r.functionCache[llvmFn] = fn
+	return fn
 }

interp/interp_test.go

@@ -42,9 +42,29 @@ func runTest(t *testing.T, pathPrefix string) {
 	// Perform the transform.
 	err = Run(mod, false)
 	if err != nil {
+		if err, match := err.(*Error); match {
+			println(err.Error())
+			if !err.Inst.IsNil() {
+				err.Inst.Dump()
+				println()
+			}
+			if len(err.Traceback) > 0 {
+				println("\ntraceback:")
+				for _, line := range err.Traceback {
+					println(line.Pos.String() + ":")
+					line.Inst.Dump()
+					println()
+				}
+			}
+		}
 		t.Fatal(err)
 	}
 
+	// To be sure, verify that the module is still valid.
+	if llvm.VerifyModule(mod, llvm.PrintMessageAction) != nil {
+		t.FailNow()
+	}
+
 	// Run some cleanup passes to get easy-to-read outputs.
 	pm := llvm.NewPassManager()
 	defer pm.Dispose()

interp/interpreter.go

@@ -0,0 +1,932 @@
+package interp
+
+import (
+	"errors"
+	"fmt"
+	"math"
+	"os"
+	"strings"
+	"time"
+
+	"tinygo.org/x/go-llvm"
+)
+
+func (r *runner) run(fn *function, params []value, parentMem *memoryView, indent string) (value, memoryView, *Error) {
+	mem := memoryView{r: r, parent: parentMem}
+	locals := make([]value, len(fn.locals))
+	r.callsExecuted++
+
+	if time.Since(r.start) > time.Minute {
+		// Running for more than a minute. This should never happen.
+		return nil, mem, r.errorAt(fn.blocks[0].instructions[0], fmt.Errorf("interp: running for more than a minute, timing out (executed calls: %d)", r.callsExecuted))
+	}
+
+	// Parameters are considered a kind of local values.
+	for i, param := range params {
+		locals[i] = param
+	}
+
+	// Start with the first basic block and the first instruction.
+	// Branch instructions may modify both bb and instIndex when branching.
+	bb := fn.blocks[0]
+	currentBB := 0
+	lastBB := -1 // last basic block is undefined, only defined after a branch
+	var operands []value
+	for instIndex := 0; instIndex < len(bb.instructions); instIndex++ {
+		inst := bb.instructions[instIndex]
+		operands = operands[:0]
+		isRuntimeInst := false
+		if inst.opcode != llvm.PHI {
+			for _, v := range inst.operands {
+				if v, ok := v.(localValue); ok {
+					if localVal := locals[fn.locals[v.value]]; localVal == nil {
+						return nil, mem, r.errorAt(inst, errors.New("interp: local not defined"))
+					} else {
+						operands = append(operands, localVal)
+						if _, ok := localVal.(localValue); ok {
+							isRuntimeInst = true
+						}
+						continue
+					}
+				}
+				operands = append(operands, v)
+			}
+		}
+		if isRuntimeInst {
+			err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+			if err != nil {
+				return nil, mem, err
+			}
+			continue
+		}
+		switch inst.opcode {
+		case llvm.Ret:
+			if len(operands) != 0 {
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"ret", operands[0])
+				}
+				// Return instruction has a value to return.
+				return operands[0], mem, nil
+			}
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"ret")
+			}
+			// Return instruction doesn't return anything, it's just 'ret void'.
+			return nil, mem, nil
+		case llvm.Br:
+			switch len(operands) {
+			case 1:
+				// Unconditional branch: [nextBB]
+				lastBB = currentBB
+				currentBB = int(operands[0].(literalValue).value.(uint32))
+				bb = fn.blocks[currentBB]
+				instIndex = -1 // start at 0 the next cycle
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"br", operands, "->", currentBB)
+				}
+			case 3:
+				// Conditional branch: [cond, thenBB, elseBB]
+				lastBB = currentBB
+				switch operands[0].Uint() {
+				case 1: // true -> thenBB
+					currentBB = int(operands[1].(literalValue).value.(uint32))
+				case 0: // false -> elseBB
+					currentBB = int(operands[2].(literalValue).value.(uint32))
+				default:
+					panic("bool should be 0 or 1")
+				}
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"br", operands, "->", currentBB)
+				}
+				bb = fn.blocks[currentBB]
+				instIndex = -1 // start at 0 the next cycle
+			default:
+				panic("unknown operands length")
+			}
+			break // continue with next block
+		case llvm.PHI:
+			var result value
+			for i := 0; i < len(inst.operands); i += 2 {
+				if int(inst.operands[i].(literalValue).value.(uint32)) == lastBB {
+					incoming := inst.operands[i+1]
+					if local, ok := incoming.(localValue); ok {
+						result = locals[fn.locals[local.value]]
+					} else {
+						result = incoming
+					}
+					break
+				}
+			}
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"phi", inst.operands, "->", result)
+			}
+			if result == nil {
+				panic("could not find PHI input")
+			}
+			locals[inst.localIndex] = result
+		case llvm.Select:
+			// Select is much like a ternary operator: it picks a result from
+			// the second and third operand based on the boolean first operand.
+			var result value
+			switch operands[0].Uint() {
+			case 1:
+				result = operands[1]
+			case 0:
+				result = operands[2]
+			default:
+				panic("boolean must be 0 or 1")
+			}
+			locals[inst.localIndex] = result
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"select", operands, "->", result)
+			}
+		case llvm.Call:
+			// A call instruction can either be a regular call or a runtime intrinsic.
+			fnPtr, err := operands[0].asPointer(r)
+			if err != nil {
+				return nil, mem, r.errorAt(inst, err)
+			}
+			callFn := r.getFunction(fnPtr.llvmValue(&mem))
+			switch {
+			case callFn.name == "runtime.trackPointer":
+				// Allocas and such are created as globals, so don't need a
+				// runtime.trackPointer.
+				// Unless the object is allocated at runtime for example, in
+				// 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":
+				// 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.
+				//   * runtime.hashmapGet tries to access the map value directly.
+				//     This is not possible as the map value is treated as a special
+				//     kind of object in this package.
+				err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+				if err != nil {
+					return nil, mem, err
+				}
+			case callFn.name == "runtime.nanotime" && r.pkgName == "time":
+				// The time package contains a call to runtime.nanotime.
+				// This appears to be to work around a limitation in Windows
+				// Server 2008:
+				//   > Monotonic times are reported as offsets from startNano.
+				//   > We initialize startNano to runtimeNano() - 1 so that on systems where
+				//   > monotonic time resolution is fairly low (e.g. Windows 2008
+				//   > which appears to have a default resolution of 15ms),
+				//   > we avoid ever reporting a monotonic time of 0.
+				//   > (Callers may want to use 0 as "time not set".)
+				// Simply let runtime.nanotime return 0 in this case, which
+				// should be fine and avoids a call to runtime.nanotime. It
+				// means that monotonic time in the time package is counted from
+				// time.Time{}.Sub(1), which should be fine.
+				locals[inst.localIndex] = literalValue{uint64(0)}
+			case callFn.name == "runtime.alloc":
+				// Allocate heap memory. At compile time, this is instead done
+				// by creating a global variable.
+
+				// Get the requested memory size to be allocated.
+				size := operands[1].Uint()
+
+				// Create the object.
+				alloc := object{
+					globalName: r.pkgName + "$alloc",
+					buffer:     newRawValue(uint32(size)),
+					size:       uint32(size),
+				}
+				index := len(r.objects)
+				r.objects = append(r.objects, alloc)
+
+				// And create a pointer to this object, for working with it (so
+				// that stores to it copy it, etc).
+				ptr := newPointerValue(r, index, 0)
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"runtime.alloc:", size, "->", ptr)
+				}
+				locals[inst.localIndex] = ptr
+			case callFn.name == "runtime.sliceCopy":
+				// sliceCopy implements the built-in copy function for slices.
+				// It is implemented here so that it can be used even if the
+				// runtime implementation is not available. Doing it this way
+				// may also be faster.
+				// Code:
+				// func sliceCopy(dst, src unsafe.Pointer, dstLen, srcLen uintptr, elemSize uintptr) int {
+				//     n := srcLen
+				//     if n > dstLen {
+				//         n = dstLen
+				//     }
+				//     memmove(dst, src, n*elemSize)
+				//     return int(n)
+				// }
+				dstLen := operands[3].Uint()
+				srcLen := operands[4].Uint()
+				elemSize := operands[5].Uint()
+				n := srcLen
+				if n > dstLen {
+					n = dstLen
+				}
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"copy:", operands[1], operands[2], n)
+				}
+				if n != 0 {
+					// Only try to copy bytes when there are any bytes to copy.
+					// This is not just an optimization. If one of the slices
+					// (or both) are nil, the asPointer method call will fail
+					// even though copying a nil slice is allowed.
+					dst, err := operands[1].asPointer(r)
+					if err != nil {
+						return nil, mem, r.errorAt(inst, err)
+					}
+					src, err := operands[2].asPointer(r)
+					if err != nil {
+						return nil, mem, r.errorAt(inst, err)
+					}
+					nBytes := uint32(n * elemSize)
+					dstObj := mem.getWritable(dst.index())
+					dstBuf := dstObj.buffer.asRawValue(r)
+					srcBuf := mem.get(src.index()).buffer.asRawValue(r)
+					copy(dstBuf.buf[dst.offset():dst.offset()+nBytes], srcBuf.buf[src.offset():])
+					dstObj.buffer = dstBuf
+					mem.put(dst.index(), dstObj)
+				}
+				switch inst.llvmInst.Type().IntTypeWidth() {
+				case 16:
+					locals[inst.localIndex] = literalValue{uint16(n)}
+				case 32:
+					locals[inst.localIndex] = literalValue{uint32(n)}
+				case 64:
+					locals[inst.localIndex] = literalValue{uint64(n)}
+				default:
+					panic("unknown integer type width")
+				}
+			case strings.HasPrefix(callFn.name, "llvm.memcpy.p0i8.p0i8.") || strings.HasPrefix(callFn.name, "llvm.memmove.p0i8.p0i8."):
+				// Copy a block of memory from one pointer to another.
+				dst, err := operands[1].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				src, err := operands[2].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				nBytes := uint32(operands[3].Uint())
+				dstObj := mem.getWritable(dst.index())
+				dstBuf := dstObj.buffer.asRawValue(r)
+				srcBuf := mem.get(src.index()).buffer.asRawValue(r)
+				copy(dstBuf.buf[dst.offset():dst.offset()+nBytes], srcBuf.buf[src.offset():])
+				dstObj.buffer = dstBuf
+				mem.put(dst.index(), dstObj)
+			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)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				actualType, err := mem.load(typeInInterfacePtr, r.pointerSize).asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				assertedType, err := operands[2].asPointer(r)
+				if err != nil {
+					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)}
+				}
+			case callFn.name == "runtime.interfaceImplements":
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"interface assert:", operands[1:])
+				}
+
+				// Load various values for the interface implements check below.
+				typeInInterfacePtr, 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)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				methodSet := mem.get(methodSetPtr.index()).llvmGlobal.Initializer()
+				interfaceMethodSetPtr, err := operands[2].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				interfaceMethodSet := mem.get(interfaceMethodSetPtr.index()).llvmGlobal.Initializer()
+
+				// Make a set of all the methods on the concrete type, for
+				// easier checking in the next step.
+				concreteTypeMethods := map[string]struct{}{}
+				for i := 0; i < methodSet.Type().ArrayLength(); i++ {
+					methodInfo := llvm.ConstExtractValue(methodSet, []uint32{uint32(i)})
+					name := llvm.ConstExtractValue(methodInfo, []uint32{0}).Name()
+					concreteTypeMethods[name] = struct{}{}
+				}
+
+				// Check whether all interface methods are also in the list
+				// of defined methods calculated above. This is the interface
+				// assert itself.
+				assertOk := uint8(1) // i1 true
+				for i := 0; i < interfaceMethodSet.Type().ArrayLength(); i++ {
+					name := llvm.ConstExtractValue(interfaceMethodSet, []uint32{uint32(i)}).Name()
+					if _, ok := concreteTypeMethods[name]; !ok {
+						// There is a method on the interface that is not
+						// implemented by the type. The assertion will fail.
+						assertOk = 0 // i1 false
+						break
+					}
+				}
+				// If assertOk is still 1, the assertion succeeded.
+				locals[inst.localIndex] = literalValue{assertOk}
+			case callFn.name == "runtime.hashmapMake":
+				// Create a new map.
+				hashmapPointerType := inst.llvmInst.Type()
+				keySize := uint32(operands[1].Uint())
+				valueSize := uint32(operands[2].Uint())
+				m := newMapValue(r, hashmapPointerType, keySize, valueSize)
+				alloc := object{
+					llvmType:   hashmapPointerType,
+					globalName: r.pkgName + "$map",
+					buffer:     m,
+					size:       m.len(r),
+				}
+				index := len(r.objects)
+				r.objects = append(r.objects, alloc)
+
+				// Create a pointer to this map. Maps are reference types, so
+				// are implemented as pointers.
+				ptr := newPointerValue(r, index, 0)
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"runtime.hashmapMake:", keySize, valueSize, "->", ptr)
+				}
+				locals[inst.localIndex] = ptr
+			case callFn.name == "runtime.hashmapBinarySet":
+				// Do a mapassign operation with a binary key (that is, without
+				// a string key).
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"runtime.hashmapBinarySet:", operands[1:])
+				}
+				mapPtr, err := operands[1].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				m := mem.getWritable(mapPtr.index()).buffer.(*mapValue)
+				keyPtr, err := operands[2].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				valuePtr, err := operands[3].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				err = m.putBinary(&mem, keyPtr, valuePtr)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+			case callFn.name == "runtime.hashmapStringSet":
+				// Do a mapassign operation with a string key.
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"runtime.hashmapBinarySet:", operands[1:])
+				}
+				mapPtr, err := operands[1].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				m := mem.getWritable(mapPtr.index()).buffer.(*mapValue)
+				stringPtr, err := operands[2].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				stringLen := operands[3].Uint()
+				valuePtr, err := operands[4].asPointer(r)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				err = m.putString(&mem, stringPtr, stringLen, valuePtr)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+			default:
+				if len(callFn.blocks) == 0 {
+					// Call to a function declaration without a definition
+					// available.
+					err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+					if err != nil {
+						return nil, mem, err
+					}
+					continue
+				}
+				// Call a function with a definition available. Run it as usual,
+				// possibly trying to recover from it if it failed to execute.
+				if r.debug {
+					argStrings := make([]string, len(operands)-1)
+					for i := range argStrings {
+						argStrings[i] = operands[i+1].String()
+					}
+					fmt.Fprintln(os.Stderr, indent+"call:", callFn.name+"("+strings.Join(argStrings, ", ")+")")
+				}
+				retval, callMem, callErr := r.run(callFn, operands[1:], &mem, indent+"    ")
+				if callErr != nil {
+					if isRecoverableError(callErr.Err) {
+						// This error can be recovered by doing the call at
+						// runtime instead of at compile time. But we need to
+						// revert any changes made by the call first.
+						if r.debug {
+							fmt.Fprintln(os.Stderr, indent+"!! revert because of error:", callErr.Err)
+						}
+						callMem.revert()
+						err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+						if err != nil {
+							return nil, mem, err
+						}
+						continue
+					}
+					// Add to the traceback, so that error handling code can see
+					// how this function got called.
+					callErr.Traceback = append(callErr.Traceback, ErrorLine{
+						Pos:  getPosition(inst.llvmInst),
+						Inst: inst.llvmInst,
+					})
+					return nil, mem, callErr
+				}
+				locals[inst.localIndex] = retval
+				mem.extend(callMem)
+			}
+		case llvm.Load:
+			// Load instruction, loading some data from the topmost memory view.
+			ptr, err := operands[0].asPointer(r)
+			if err != nil {
+				return nil, mem, r.errorAt(inst, err)
+			}
+			size := operands[1].(literalValue).value.(uint64)
+			if mem.hasExternalStore(ptr) {
+				// If there could be an external store (for example, because a
+				// pointer to the object was passed to a function that could not
+				// be interpreted at compile time) then the load must be done at
+				// runtime.
+				err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+				if err != nil {
+					return nil, mem, err
+				}
+				continue
+			}
+			result := mem.load(ptr, uint32(size))
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"load:", ptr, "->", result)
+			}
+			locals[inst.localIndex] = result
+		case llvm.Store:
+			// Store instruction. Create a new object in the memory view and
+			// store to that, to make it possible to roll back this store.
+			ptr, err := operands[1].asPointer(r)
+			if err != nil {
+				return nil, mem, r.errorAt(inst, err)
+			}
+			if mem.hasExternalLoadOrStore(ptr) {
+				err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+				if err != nil {
+					return nil, mem, err
+				}
+				continue
+			}
+			val := operands[0]
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"store:", val, ptr)
+			}
+			mem.store(val, ptr)
+		case llvm.Alloca:
+			// Alloca normally allocates some stack memory. In the interpreter,
+			// it allocates a global instead.
+			// This can likely be optimized, as all it really needs is an alloca
+			// in the initAll function and creating a global is wasteful for
+			// this purpose.
+
+			// Create the new object.
+			size := operands[0].(literalValue).value.(uint64)
+			alloca := object{
+				llvmType:   inst.llvmInst.Type(),
+				globalName: r.pkgName + "$alloca",
+				buffer:     newRawValue(uint32(size)),
+				size:       uint32(size),
+			}
+			index := len(r.objects)
+			r.objects = append(r.objects, alloca)
+
+			// Create a pointer to this object (an alloca produces a pointer).
+			ptr := newPointerValue(r, index, 0)
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"alloca:", operands, "->", ptr)
+			}
+			locals[inst.localIndex] = ptr
+		case llvm.GetElementPtr:
+			// GetElementPtr does pointer arithmetic, changing the offset of the
+			// pointer into the underlying object.
+			var offset uint64
+			var gepOperands []uint64
+			for i := 2; i < len(operands); i += 2 {
+				index := operands[i].Uint()
+				elementSize := operands[i+1].Uint()
+				if int64(elementSize) < 0 {
+					// This is a struct field.
+					// The field number is encoded by flipping all the bits.
+					gepOperands = append(gepOperands, ^elementSize)
+					offset += index
+				} else {
+					// This is a normal GEP, probably an array index.
+					gepOperands = append(gepOperands, index)
+					offset += elementSize * index
+				}
+			}
+			ptr, err := operands[0].asPointer(r)
+			if err != nil {
+				if err != errIntegerAsPointer {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				// GEP on fixed pointer value (for example, memory-mapped I/O).
+				ptrValue := operands[0].Uint() + offset
+				switch operands[0].len(r) {
+				case 8:
+					locals[inst.localIndex] = literalValue{uint64(ptrValue)}
+				case 4:
+					locals[inst.localIndex] = literalValue{uint32(ptrValue)}
+				case 2:
+					locals[inst.localIndex] = literalValue{uint16(ptrValue)}
+				default:
+					panic("pointer operand is not of a known pointer size")
+				}
+				continue
+			}
+			ptr = ptr.addOffset(uint32(offset))
+			locals[inst.localIndex] = ptr
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"gep:", operands, "->", ptr)
+			}
+		case llvm.BitCast, llvm.IntToPtr, llvm.PtrToInt:
+			// Various bitcast-like instructions that all keep the same bits
+			// while changing the LLVM type.
+			// Because interp doesn't preserve the type, these operations are
+			// identity operations.
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+instructionNameMap[inst.opcode]+":", operands[0])
+			}
+			locals[inst.localIndex] = operands[0]
+		case llvm.ExtractValue:
+			agg := operands[0].asRawValue(r)
+			offset := operands[1].(literalValue).value.(uint64)
+			size := operands[2].(literalValue).value.(uint64)
+			elt := rawValue{
+				buf: agg.buf[offset : offset+size],
+			}
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"extractvalue:", operands, "->", elt)
+			}
+			locals[inst.localIndex] = elt
+		case llvm.InsertValue:
+			agg := operands[0].asRawValue(r)
+			elt := operands[1].asRawValue(r)
+			offset := int(operands[2].(literalValue).value.(uint64))
+			newagg := newRawValue(uint32(len(agg.buf)))
+			copy(newagg.buf, agg.buf)
+			copy(newagg.buf[offset:], elt.buf)
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"insertvalue:", operands, "->", newagg)
+			}
+			locals[inst.localIndex] = newagg
+		case llvm.ICmp:
+			predicate := llvm.IntPredicate(operands[2].(literalValue).value.(uint8))
+			var result bool
+			lhs := operands[0]
+			rhs := operands[1]
+			switch predicate {
+			case llvm.IntEQ, llvm.IntNE:
+				lhsPointer, lhsErr := lhs.asPointer(r)
+				rhsPointer, rhsErr := rhs.asPointer(r)
+				if (lhsErr == nil) != (rhsErr == nil) {
+					// Fast path: only one is a pointer, so they can't be equal.
+					result = false
+				} else if lhsErr == nil {
+					// Both must be nil, so both are pointers.
+					// Compare them directly.
+					result = lhsPointer.equal(rhsPointer)
+				} else {
+					// Fall back to generic comparison.
+					result = lhs.asRawValue(r).equal(rhs.asRawValue(r))
+				}
+				if predicate == llvm.IntNE {
+					result = !result
+				}
+			case llvm.IntUGT:
+				result = lhs.Uint() > rhs.Uint()
+			case llvm.IntUGE:
+				result = lhs.Uint() >= rhs.Uint()
+			case llvm.IntULT:
+				result = lhs.Uint() < rhs.Uint()
+			case llvm.IntULE:
+				result = lhs.Uint() <= rhs.Uint()
+			case llvm.IntSGT:
+				result = lhs.Int() > rhs.Int()
+			case llvm.IntSGE:
+				result = lhs.Int() >= rhs.Int()
+			case llvm.IntSLT:
+				result = lhs.Int() < rhs.Int()
+			case llvm.IntSLE:
+				result = lhs.Int() <= rhs.Int()
+			default:
+				return nil, mem, r.errorAt(inst, errors.New("interp: unsupported icmp"))
+			}
+			if result {
+				locals[inst.localIndex] = literalValue{uint8(1)}
+			} else {
+				locals[inst.localIndex] = literalValue{uint8(0)}
+			}
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"icmp:", operands[0], intPredicateString(predicate), operands[1], "->", result)
+			}
+		case llvm.FCmp:
+			predicate := llvm.FloatPredicate(operands[2].(literalValue).value.(uint8))
+			var result bool
+			var lhs, rhs float64
+			switch operands[0].len(r) {
+			case 8:
+				lhs = math.Float64frombits(operands[0].Uint())
+				rhs = math.Float64frombits(operands[1].Uint())
+			case 4:
+				lhs = float64(math.Float32frombits(uint32(operands[0].Uint())))
+				rhs = float64(math.Float32frombits(uint32(operands[1].Uint())))
+			default:
+				panic("unknown float type")
+			}
+			switch predicate {
+			case llvm.FloatOEQ:
+				result = lhs == rhs
+			case llvm.FloatUNE:
+				result = lhs != rhs
+			case llvm.FloatOGT:
+				result = lhs > rhs
+			case llvm.FloatOGE:
+				result = lhs >= rhs
+			case llvm.FloatOLT:
+				result = lhs < rhs
+			case llvm.FloatOLE:
+				result = lhs <= rhs
+			default:
+				return nil, mem, r.errorAt(inst, errors.New("interp: unsupported fcmp"))
+			}
+			if result {
+				locals[inst.localIndex] = literalValue{uint8(1)}
+			} else {
+				locals[inst.localIndex] = literalValue{uint8(0)}
+			}
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+"fcmp:", operands[0], predicate, operands[1], "->", result)
+			}
+		case llvm.Add, llvm.Sub, llvm.Mul, llvm.UDiv, llvm.SDiv, llvm.URem, llvm.SRem, llvm.Shl, llvm.LShr, llvm.AShr, llvm.And, llvm.Or, llvm.Xor:
+			// Integer binary operations.
+			lhs := operands[0]
+			rhs := operands[1]
+			lhsPtr, err := lhs.asPointer(r)
+			if err == nil {
+				// The lhs is a pointer. This sometimes happens for particular
+				// pointer tricks.
+				switch inst.opcode {
+				case llvm.Add:
+					// This likely means this is part of a
+					// unsafe.Pointer(uintptr(ptr) + offset) pattern.
+					lhsPtr = lhsPtr.addOffset(uint32(rhs.Uint()))
+					locals[inst.localIndex] = lhsPtr
+					continue
+				case llvm.Xor:
+					if rhs.Uint() == 0 {
+						// Special workaround for strings.noescape, see
+						// src/strings/builder.go in the Go source tree. This is
+						// the identity operator, so we can return the input.
+						locals[inst.localIndex] = lhs
+						continue
+					}
+				default:
+					// Catch-all for weird operations that should just be done
+					// at runtime.
+					err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+					if err != nil {
+						return nil, mem, err
+					}
+					continue
+				}
+			}
+			var result uint64
+			switch inst.opcode {
+			case llvm.Add:
+				result = lhs.Uint() + rhs.Uint()
+			case llvm.Sub:
+				result = lhs.Uint() - rhs.Uint()
+			case llvm.Mul:
+				result = lhs.Uint() * rhs.Uint()
+			case llvm.UDiv:
+				result = lhs.Uint() / rhs.Uint()
+			case llvm.SDiv:
+				result = uint64(lhs.Int() / rhs.Int())
+			case llvm.URem:
+				result = lhs.Uint() % rhs.Uint()
+			case llvm.SRem:
+				result = uint64(lhs.Int() % rhs.Int())
+			case llvm.Shl:
+				result = lhs.Uint() << rhs.Uint()
+			case llvm.LShr:
+				result = lhs.Uint() >> rhs.Uint()
+			case llvm.AShr:
+				result = uint64(lhs.Int() >> rhs.Uint())
+			case llvm.And:
+				result = lhs.Uint() & rhs.Uint()
+			case llvm.Or:
+				result = lhs.Uint() | rhs.Uint()
+			case llvm.Xor:
+				result = lhs.Uint() ^ rhs.Uint()
+			default:
+				panic("unreachable")
+			}
+			switch lhs.len(r) {
+			case 8:
+				locals[inst.localIndex] = literalValue{result}
+			case 4:
+				locals[inst.localIndex] = literalValue{uint32(result)}
+			case 2:
+				locals[inst.localIndex] = literalValue{uint16(result)}
+			case 1:
+				locals[inst.localIndex] = literalValue{uint8(result)}
+			default:
+				panic("unknown integer size")
+			}
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+instructionNameMap[inst.opcode]+":", lhs, rhs, "->", result)
+			}
+		case llvm.SExt, llvm.ZExt, llvm.Trunc:
+			// Change the size of an integer to a larger or smaller bit width.
+			// We make use of the fact that the Uint() function already
+			// zero-extends the value and that Int() already sign-extends the
+			// value, so we only need to truncate it to the appropriate bit
+			// width. This means we can implement sext, zext and trunc in the
+			// same way, by first {zero,sign}extending all the way up to uint64
+			// and then truncating it as necessary.
+			var value uint64
+			if inst.opcode == llvm.SExt {
+				value = uint64(operands[0].Int())
+			} else {
+				value = operands[0].Uint()
+			}
+			bitwidth := operands[1].Uint()
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+instructionNameMap[inst.opcode]+":", value, bitwidth)
+			}
+			switch bitwidth {
+			case 64:
+				locals[inst.localIndex] = literalValue{value}
+			case 32:
+				locals[inst.localIndex] = literalValue{uint32(value)}
+			case 16:
+				locals[inst.localIndex] = literalValue{uint16(value)}
+			case 8:
+				locals[inst.localIndex] = literalValue{uint8(value)}
+			default:
+				panic("unknown integer size in sext/zext/trunc")
+			}
+		case llvm.SIToFP, llvm.UIToFP:
+			var value float64
+			switch inst.opcode {
+			case llvm.SIToFP:
+				value = float64(operands[0].Int())
+			case llvm.UIToFP:
+				value = float64(operands[0].Uint())
+			}
+			bitwidth := operands[1].Uint()
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+instructionNameMap[inst.opcode]+":", value, bitwidth)
+			}
+			switch bitwidth {
+			case 64:
+				locals[inst.localIndex] = literalValue{math.Float64bits(value)}
+			case 32:
+				locals[inst.localIndex] = literalValue{math.Float32bits(float32(value))}
+			default:
+				panic("unknown integer size in sitofp/uitofp")
+			}
+		default:
+			if r.debug {
+				fmt.Fprintln(os.Stderr, indent+inst.String())
+			}
+			return nil, mem, r.errorAt(inst, errUnsupportedInst)
+		}
+	}
+	return nil, mem, r.errorAt(bb.instructions[len(bb.instructions)-1], errors.New("interp: reached end of basic block without terminator"))
+}
+
+func (r *runner) runAtRuntime(fn *function, inst instruction, locals []value, mem *memoryView, indent string) *Error {
+	numOperands := inst.llvmInst.OperandsCount()
+	operands := make([]llvm.Value, numOperands)
+	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)
+		}
+		operands[i] = operand
+	}
+	if r.debug {
+		fmt.Fprintln(os.Stderr, indent+inst.String())
+	}
+	var result llvm.Value
+	switch inst.opcode {
+	case llvm.Call:
+		llvmFn := operands[len(operands)-1]
+		args := operands[:len(operands)-1]
+		for _, arg := range args {
+			if arg.Type().TypeKind() == llvm.PointerTypeKind {
+				mem.markExternalStore(arg)
+			}
+		}
+		result = r.builder.CreateCall(llvmFn, args, inst.name)
+	case llvm.Load:
+		mem.markExternalLoad(operands[0])
+		result = r.builder.CreateLoad(operands[0], inst.name)
+		if inst.llvmInst.IsVolatile() {
+			result.SetVolatile(true)
+		}
+	case llvm.Store:
+		mem.markExternalStore(operands[1])
+		result = r.builder.CreateStore(operands[0], operands[1])
+		if inst.llvmInst.IsVolatile() {
+			result.SetVolatile(true)
+		}
+	case llvm.BitCast:
+		result = r.builder.CreateBitCast(operands[0], inst.llvmInst.Type(), inst.name)
+	case llvm.ExtractValue:
+		indices := inst.llvmInst.Indices()
+		if len(indices) != 1 {
+			panic("expected exactly one index")
+		}
+		result = r.builder.CreateExtractValue(operands[0], int(indices[0]), inst.name)
+	case llvm.InsertValue:
+		indices := inst.llvmInst.Indices()
+		if len(indices) != 1 {
+			panic("expected exactly one index")
+		}
+		result = r.builder.CreateInsertValue(operands[0], operands[1], int(indices[0]), inst.name)
+	case llvm.Add:
+		result = r.builder.CreateAdd(operands[0], operands[1], inst.name)
+	case llvm.Sub:
+		result = r.builder.CreateSub(operands[0], operands[1], inst.name)
+	case llvm.Mul:
+		result = r.builder.CreateMul(operands[0], operands[1], inst.name)
+	case llvm.UDiv:
+		result = r.builder.CreateUDiv(operands[0], operands[1], inst.name)
+	case llvm.SDiv:
+		result = r.builder.CreateSDiv(operands[0], operands[1], inst.name)
+	case llvm.URem:
+		result = r.builder.CreateURem(operands[0], operands[1], inst.name)
+	case llvm.SRem:
+		result = r.builder.CreateSRem(operands[0], operands[1], inst.name)
+	case llvm.ZExt:
+		result = r.builder.CreateZExt(operands[0], inst.llvmInst.Type(), inst.name)
+	default:
+		return r.errorAt(inst, errUnsupportedRuntimeInst)
+	}
+	locals[inst.localIndex] = localValue{result}
+	mem.instructions = append(mem.instructions, result)
+	return nil
+}
+
+func intPredicateString(predicate llvm.IntPredicate) string {
+	switch predicate {
+	case llvm.IntEQ:
+		return "eq"
+	case llvm.IntNE:
+		return "ne"
+	case llvm.IntUGT:
+		return "ugt"
+	case llvm.IntUGE:
+		return "uge"
+	case llvm.IntULT:
+		return "ult"
+	case llvm.IntULE:
+		return "ule"
+	case llvm.IntSGT:
+		return "sgt"
+	case llvm.IntSGE:
+		return "sge"
+	case llvm.IntSLT:
+		return "slt"
+	case llvm.IntSLE:
+		return "sle"
+	default:
+		return "cmp?"
+	}
+}

interp/scan.go

@@ -1,259 +0,0 @@
-package interp
-
-import (
-	"errors"
-	"strings"
-
-	"tinygo.org/x/go-llvm"
-)
-
-type sideEffectSeverity int
-
-func (severity sideEffectSeverity) String() string {
-	switch severity {
-	case sideEffectInProgress:
-		return "in progress"
-	case sideEffectNone:
-		return "none"
-	case sideEffectLimited:
-		return "limited"
-	case sideEffectAll:
-		return "all"
-	default:
-		return "unknown"
-	}
-}
-
-const (
-	sideEffectInProgress sideEffectSeverity = iota // computing side effects is in progress (for recursive functions)
-	sideEffectNone                                 // no side effects at all (pure)
-	sideEffectLimited                              // has side effects, but the effects are known
-	sideEffectAll                                  // has unknown side effects
-)
-
-// sideEffectResult contains the scan results after scanning a function for side
-// effects (recursively).
-type sideEffectResult struct {
-	severity        sideEffectSeverity
-	mentionsGlobals map[llvm.Value]struct{}
-}
-
-// hasSideEffects scans this function and all descendants, recursively. It
-// returns whether this function has side effects and if it does, which globals
-// it mentions anywhere in this function or any called functions.
-func (e *evalPackage) hasSideEffects(fn llvm.Value) (*sideEffectResult, *Error) {
-	name := fn.Name()
-	switch {
-	case name == "runtime.alloc":
-		// Cannot be scanned but can be interpreted.
-		return &sideEffectResult{severity: sideEffectNone}, nil
-	case name == "runtime.nanotime":
-		// Fixed value at compile time.
-		return &sideEffectResult{severity: sideEffectNone}, nil
-	case name == "runtime._panic":
-		return &sideEffectResult{severity: sideEffectLimited}, nil
-	case name == "runtime.typeAssert":
-		return &sideEffectResult{severity: sideEffectNone}, nil
-	case name == "runtime.interfaceImplements":
-		return &sideEffectResult{severity: sideEffectNone}, nil
-	case name == "runtime.sliceCopy":
-		return &sideEffectResult{severity: sideEffectNone}, nil
-	case name == "runtime.trackPointer":
-		return &sideEffectResult{severity: sideEffectNone}, nil
-	case name == "llvm.dbg.value":
-		return &sideEffectResult{severity: sideEffectNone}, nil
-	case name == "(*sync/atomic.Value).Load" || name == "(*sync/atomic.Value).Store":
-		// These functions do some unsafe pointer loading/storing but are
-		// otherwise safe.
-		return &sideEffectResult{severity: sideEffectLimited}, nil
-	case strings.HasPrefix(name, "llvm.lifetime."):
-		return &sideEffectResult{severity: sideEffectNone}, nil
-	}
-	if fn.IsDeclaration() {
-		return &sideEffectResult{severity: sideEffectLimited}, nil
-	}
-	if e.sideEffectFuncs == nil {
-		e.sideEffectFuncs = make(map[llvm.Value]*sideEffectResult)
-	}
-	if se, ok := e.sideEffectFuncs[fn]; ok {
-		return se, nil
-	}
-	result := &sideEffectResult{
-		severity:        sideEffectInProgress,
-		mentionsGlobals: map[llvm.Value]struct{}{},
-	}
-	e.sideEffectFuncs[fn] = result
-	dirtyLocals := map[llvm.Value]struct{}{}
-	for bb := fn.EntryBasicBlock(); !bb.IsNil(); bb = llvm.NextBasicBlock(bb) {
-		for inst := bb.FirstInstruction(); !inst.IsNil(); inst = llvm.NextInstruction(inst) {
-			if inst.IsAInstruction().IsNil() {
-				// Should not happen in valid IR.
-				panic("not an instruction")
-			}
-
-			// Check for any globals mentioned anywhere in the function. Assume
-			// any mentioned globals may be read from or written to when
-			// executed, thus must be marked dirty with a call.
-			for i := 0; i < inst.OperandsCount(); i++ {
-				operand := inst.Operand(i)
-				if !operand.IsAGlobalVariable().IsNil() {
-					result.mentionsGlobals[operand] = struct{}{}
-				}
-			}
-
-			switch inst.InstructionOpcode() {
-			case llvm.IndirectBr, llvm.Invoke:
-				// Not emitted by the compiler.
-				return nil, e.errorAt(inst, errors.New("unknown instructions"))
-			case llvm.Call:
-				child := inst.CalledValue()
-				if !child.IsAInlineAsm().IsNil() {
-					// Inline assembly. This most likely has side effects.
-					// Assume they're only limited side effects, similar to
-					// external function calls.
-					result.updateSeverity(sideEffectLimited)
-					continue
-				}
-				if child.IsAFunction().IsNil() {
-					// Indirect call?
-					// In any case, we can't know anything here about what it
-					// affects exactly so mark this function as invoking all
-					// possible side effects.
-					result.updateSeverity(sideEffectAll)
-					continue
-				}
-				if child.IsDeclaration() {
-					// External function call. Assume only limited side effects
-					// (no affected globals, etc.).
-					switch child.Name() {
-					case "runtime.alloc":
-						continue
-					case "runtime.typeAssert":
-						continue // implemented in interp
-					case "runtime.interfaceImplements":
-						continue // implemented in interp
-					}
-					if e.hasLocalSideEffects(dirtyLocals, inst) {
-						result.updateSeverity(sideEffectLimited)
-					}
-					continue
-				}
-				childSideEffects, err := e.hasSideEffects(child)
-				if err != nil {
-					return nil, err
-				}
-				switch childSideEffects.severity {
-				case sideEffectInProgress, sideEffectNone:
-					// no side effects or recursive function - continue scanning
-				case sideEffectLimited:
-					// The return value may be problematic.
-					if e.hasLocalSideEffects(dirtyLocals, inst) {
-						result.updateSeverity(sideEffectLimited)
-					}
-				case sideEffectAll:
-					result.updateSeverity(sideEffectAll)
-				default:
-					panic("unreachable")
-				}
-			case llvm.Load:
-				if inst.IsVolatile() {
-					result.updateSeverity(sideEffectLimited)
-				}
-				if _, ok := e.dirtyGlobals[inst.Operand(0)]; ok {
-					if e.hasLocalSideEffects(dirtyLocals, inst) {
-						result.updateSeverity(sideEffectLimited)
-					}
-				}
-			case llvm.Store:
-				if inst.IsVolatile() {
-					result.updateSeverity(sideEffectLimited)
-				}
-			case llvm.IntToPtr:
-				// Pointer casts are not yet supported.
-				result.updateSeverity(sideEffectLimited)
-			default:
-				// Ignore most instructions.
-				// Check this list for completeness:
-				// https://godoc.org/github.com/llvm-mirror/llvm/bindings/go/llvm#Opcode
-			}
-		}
-	}
-
-	if result.severity == sideEffectInProgress {
-		// No side effect was reported for this function.
-		result.severity = sideEffectNone
-	}
-	return result, nil
-}
-
-// hasLocalSideEffects checks whether the given instruction flows into a branch
-// or return instruction, in which case the whole function must be marked as
-// having side effects and be called at runtime.
-func (e *Eval) hasLocalSideEffects(dirtyLocals map[llvm.Value]struct{}, inst llvm.Value) bool {
-	if _, ok := dirtyLocals[inst]; ok {
-		// It is already known that this local is dirty.
-		return true
-	}
-
-	for use := inst.FirstUse(); !use.IsNil(); use = use.NextUse() {
-		user := use.User()
-		if user.IsAInstruction().IsNil() {
-			// Should not happen in valid IR.
-			panic("user not an instruction")
-		}
-		switch user.InstructionOpcode() {
-		case llvm.Br, llvm.Switch:
-			// A branch on a dirty value makes this function dirty: it cannot be
-			// interpreted at compile time so has to be run at runtime. It is
-			// marked as having side effects for this reason.
-			return true
-		case llvm.Ret:
-			// This function returns a dirty value so it is itself marked as
-			// dirty to make sure it is called at runtime.
-			return true
-		case llvm.Store:
-			ptr := user.Operand(1)
-			if !ptr.IsAGlobalVariable().IsNil() {
-				// Store to a global variable.
-				// Already handled in (*Eval).hasSideEffects.
-				continue
-			}
-			// This store might affect all kinds of values. While it is
-			// certainly possible to traverse through all of them, the easiest
-			// option right now is to just assume the worst and say that this
-			// function has side effects.
-			// TODO: traverse through all stores and mark all relevant allocas /
-			// globals dirty.
-			return true
-		default:
-			// All instructions that take 0 or more operands (1 or more if it
-			// was a use) and produce a result.
-			// For a list:
-			// https://godoc.org/github.com/llvm-mirror/llvm/bindings/go/llvm#Opcode
-			dirtyLocals[user] = struct{}{}
-			if e.hasLocalSideEffects(dirtyLocals, user) {
-				return true
-			}
-		}
-	}
-
-	// No side effects found.
-	return false
-}
-
-// updateSeverity sets r.severity to the max of r.severity and severity,
-// conservatively assuming the worst severity.
-func (r *sideEffectResult) updateSeverity(severity sideEffectSeverity) {
-	if severity > r.severity {
-		r.severity = severity
-	}
-}
-
-// updateSeverity updates the severity with the severity of the child severity,
-// like in a function call. This means it also copies the mentioned globals.
-func (r *sideEffectResult) update(child *sideEffectResult) {
-	r.updateSeverity(child.severity)
-	for global := range child.mentionsGlobals {
-		r.mentionsGlobals[global] = struct{}{}
-	}
-}

interp/scan_test.go

@@ -1,95 +0,0 @@
-package interp
-
-import (
-	"os"
-	"sort"
-	"testing"
-
-	"tinygo.org/x/go-llvm"
-)
-
-var scanTestTable = []struct {
-	name            string
-	severity        sideEffectSeverity
-	mentionsGlobals []string
-}{
-	{"returnsConst", sideEffectNone, nil},
-	{"returnsArg", sideEffectNone, nil},
-	{"externalCallOnly", sideEffectNone, nil},
-	{"externalCallAndReturn", sideEffectLimited, nil},
-	{"externalCallBranch", sideEffectLimited, nil},
-	{"readCleanGlobal", sideEffectNone, []string{"cleanGlobalInt"}},
-	{"readDirtyGlobal", sideEffectLimited, []string{"dirtyGlobalInt"}},
-	{"callFunctionPointer", sideEffectAll, []string{"functionPointer"}},
-	{"getDirtyPointer", sideEffectLimited, nil},
-	{"storeToPointer", sideEffectLimited, nil},
-	{"callTypeAssert", sideEffectNone, nil},
-	{"callInterfaceImplements", sideEffectNone, nil},
-}
-
-func TestScan(t *testing.T) {
-	t.Parallel()
-
-	// Read the input IR.
-	path := "testdata/scan.ll"
-	ctx := llvm.NewContext()
-	buf, err := llvm.NewMemoryBufferFromFile(path)
-	os.Stat(path) // make sure this file is tracked by `go test` caching
-	if err != nil {
-		t.Fatalf("could not read file %s: %v", path, err)
-	}
-	mod, err := ctx.ParseIR(buf)
-	if err != nil {
-		t.Fatalf("could not load module:\n%v", err)
-	}
-
-	// Check all to-be-tested functions.
-	for _, tc := range scanTestTable {
-		// Create an eval object, for testing.
-		e := &Eval{
-			Mod:          mod,
-			TargetData:   llvm.NewTargetData(mod.DataLayout()),
-			dirtyGlobals: map[llvm.Value]struct{}{},
-		}
-
-		// Mark some globals dirty, for testing.
-		e.markDirty(mod.NamedGlobal("dirtyGlobalInt"))
-
-		// Scan for side effects.
-		fn := mod.NamedFunction(tc.name)
-		if fn.IsNil() {
-			t.Errorf("scan test: could not find tested function %s in the IR", tc.name)
-			continue
-		}
-		evalPkg := &evalPackage{e, "testdata"}
-		result, err := evalPkg.hasSideEffects(fn)
-		if err != nil {
-			t.Errorf("scan test: failed to scan %s for side effects: %v", fn.Name(), err)
-		}
-
-		// Check whether the result is what we expect.
-		if result.severity != tc.severity {
-			t.Errorf("scan test: function %s should have severity %s but it has %s", tc.name, tc.severity, result.severity)
-		}
-
-		// Check whether the mentioned globals match with what we'd expect.
-		mentionsGlobalNames := make([]string, 0, len(result.mentionsGlobals))
-		for global := range result.mentionsGlobals {
-			mentionsGlobalNames = append(mentionsGlobalNames, global.Name())
-		}
-		sort.Strings(mentionsGlobalNames)
-		globalsMismatch := false
-		if len(result.mentionsGlobals) != len(tc.mentionsGlobals) {
-			globalsMismatch = true
-		} else {
-			for i, globalName := range mentionsGlobalNames {
-				if tc.mentionsGlobals[i] != globalName {
-					globalsMismatch = true
-				}
-			}
-		}
-		if globalsMismatch {
-			t.Errorf("scan test: expected %s to mention globals %v, but it mentions globals %v", tc.name, tc.mentionsGlobals, mentionsGlobalNames)
-		}
-	}
-}

interp/testdata/basic.ll

@@ -4,6 +4,10 @@ target triple = "x86_64--linux"
 @main.v1 = internal global i64 0
 @main.nonConst1 = global [4 x i64] zeroinitializer
 @main.nonConst2 = global i64 0
+@main.someArray = global [8 x {i16, i32}] zeroinitializer
+@main.exportedValue = global [1 x i16*] [i16* @main.exposedValue1]
+@main.exposedValue1 = global i16 0
+@main.exposedValue2 = global i16 0
 
 declare void @runtime.printint64(i64) unnamed_addr
 
@@ -47,6 +51,20 @@ entry:
   %value2 = load i64, i64* %gep2
   store i64 %value2, i64* @main.nonConst2
 
+  ; Test that the following GEP works:
+  ;     var someArray
+  ;     modifyExternal(&someArray[3].field1)
+  %gep3 = getelementptr [8 x {i16, i32}], [8 x {i16, i32}]* @main.someArray, i32 0, i32 3, i32 1
+  call void @modifyExternal(i32* %gep3)
+
+  ; Test that marking a value as external also marks all referenced values.
+  call void @modifyExternal(i32* bitcast ([1 x i16*]* @main.exportedValue to i32*))
+  store i16 5, i16* @main.exposedValue1
+
+  ; Test that this even propagates through functions.
+  call void @modifyExternal(i32* bitcast (void ()* @willModifyGlobal to i32*))
+  store i16 7, i16* @main.exposedValue2
+
   ret void
 }
 
@@ -58,3 +76,14 @@ entry:
 }
 
 declare i64 @someValue()
+
+declare void @modifyExternal(i32*)
+
+; This function will modify an external value. By passing this function as a
+; function pointer to an external function, @main.exposedValue2 should be
+; marked as external.
+define void @willModifyGlobal() {
+entry:
+  store i16 8, i16* @main.exposedValue2
+  ret void
+}

interp/testdata/basic.out.ll

@@ -3,6 +3,10 @@ target triple = "x86_64--linux"
 
 @main.nonConst1 = local_unnamed_addr global [4 x i64] zeroinitializer
 @main.nonConst2 = local_unnamed_addr global i64 0
+@main.someArray = global [8 x { i16, i32 }] zeroinitializer
+@main.exportedValue = global [1 x i16*] [i16* @main.exposedValue1]
+@main.exposedValue1 = global i16 0
+@main.exposedValue2 = local_unnamed_addr global i16 0
 
 declare void @runtime.printint64(i64) unnamed_addr
 
@@ -16,6 +20,11 @@ entry:
   store i64 %value1, i64* getelementptr inbounds ([4 x i64], [4 x i64]* @main.nonConst1, i32 0, i32 0)
   %value2 = load i64, i64* getelementptr inbounds ([4 x i64], [4 x i64]* @main.nonConst1, i32 0, i32 0)
   store i64 %value2, i64* @main.nonConst2
+  call void @modifyExternal(i32* getelementptr inbounds ([8 x { i16, i32 }], [8 x { i16, i32 }]* @main.someArray, i32 0, i32 3, i32 1))
+  call void @modifyExternal(i32* bitcast ([1 x i16*]* @main.exportedValue to i32*))
+  store i16 5, i16* @main.exposedValue1
+  call void @modifyExternal(i32* bitcast (void ()* @willModifyGlobal to i32*))
+  store i16 7, i16* @main.exposedValue2
   ret void
 }
 
@@ -27,3 +36,11 @@ entry:
 }
 
 declare i64 @someValue() local_unnamed_addr
+
+declare void @modifyExternal(i32*) local_unnamed_addr
+
+define void @willModifyGlobal() {
+entry:
+  store i16 8, i16* @main.exposedValue2
+  ret void
+}

interp/testdata/map.ll

@@ -48,8 +48,7 @@ entry:
 define internal void @main.testNonConstantBinarySet() {
   %hashmap.key = alloca i8
   %hashmap.value = alloca i8
-  ; Create hashmap from global. This breaks the normal hashmapBinarySet
-  ; optimization, to test the fallback.
+  ; Create hashmap from global.
   %map.new = call %runtime.hashmap* @runtime.hashmapMake(i8 1, i8 1, i32 1, i8* undef, i8* null)
   store %runtime.hashmap* %map.new, %runtime.hashmap** @main.binaryMap
   %map = load %runtime.hashmap*, %runtime.hashmap** @main.binaryMap
@@ -64,8 +63,7 @@ define internal void @main.testNonConstantBinarySet() {
 ; operations (with string keys).
 define internal void @main.testNonConstantStringSet() {
   %hashmap.value = alloca i8
-  ; Create hashmap from global. This breaks the normal hashmapStringSet
-  ; optimization, to test the fallback.
+  ; Create hashmap from global.
   %map.new = call %runtime.hashmap* @runtime.hashmapMake(i8 8, i8 1, i32 1, i8* undef, i8* null)
   store %runtime.hashmap* %map.new, %runtime.hashmap** @main.stringMap
   %map = load %runtime.hashmap*, %runtime.hashmap** @main.stringMap

interp/testdata/map.out.ll

@@ -2,27 +2,19 @@ target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64"
 target triple = "armv6m-none-eabi"
 
 %runtime.hashmap = type { %runtime.hashmap*, i8*, i32, i8, i8, i8 }
-%runtime._string = type { i8*, i32 }
 
 @main.m = local_unnamed_addr global %runtime.hashmap* @"main$map"
-@main.binaryMap = local_unnamed_addr global %runtime.hashmap* @"main$map.4"
-@main.stringMap = local_unnamed_addr global %runtime.hashmap* @"main$map.6"
+@main.binaryMap = local_unnamed_addr global %runtime.hashmap* @"main$map.1"
+@main.stringMap = local_unnamed_addr global %runtime.hashmap* @"main$map.3"
 @main.init.string = internal unnamed_addr constant [7 x i8] c"CONNECT"
-@"main$mapbucket" = internal unnamed_addr global { [8 x i8], i8*, [8 x i8], [8 x %runtime._string] } { [8 x i8] c"\04\00\00\00\00\00\00\00", i8* null, [8 x i8] c"\01\00\00\00\00\00\00\00", [8 x %runtime._string] [%runtime._string { i8* getelementptr inbounds ([7 x i8], [7 x i8]* @main.init.string, i32 0, i32 0), i32 7 }, %runtime._string zeroinitializer, %runtime._string zeroinitializer, %runtime._string zeroinitializer, %runtime._string zeroinitializer, %runtime._string zeroinitializer, %runtime._string zeroinitializer, %runtime._string zeroinitializer] }
-@"main$map" = internal unnamed_addr global %runtime.hashmap { %runtime.hashmap* null, i8* getelementptr inbounds ({ [8 x i8], i8*, [8 x i8], [8 x %runtime._string] }, { [8 x i8], i8*, [8 x i8], [8 x %runtime._string] }* @"main$mapbucket", i32 0, i32 0, i32 0), i32 1, i8 1, i8 8, i8 0 }
-@"main$alloca.2" = internal global i8 1
-@"main$alloca.3" = internal global i8 2
-@"main$map.4" = internal unnamed_addr global %runtime.hashmap { %runtime.hashmap* null, i8* null, i32 0, i8 1, i8 1, i8 0 }
-@"main$alloca.5" = internal global i8 2
-@"main$map.6" = internal unnamed_addr global %runtime.hashmap { %runtime.hashmap* null, i8* null, i32 0, i8 8, i8 1, i8 0 }
-
-declare void @runtime.hashmapBinarySet(%runtime.hashmap*, i8*, i8*, i8*, i8*) local_unnamed_addr
-
-declare void @runtime.hashmapStringSet(%runtime.hashmap*, i8*, i32, i8*, i8*, i8*) local_unnamed_addr
+@"main$map" = internal global %runtime.hashmap { %runtime.hashmap* null, i8* getelementptr inbounds ({ [8 x i8], i8*, { i8, [7 x i8] }, { { [7 x i8]*, [4 x i8] }, [56 x i8] } }, { [8 x i8], i8*, { i8, [7 x i8] }, { { [7 x i8]*, [4 x i8] }, [56 x i8] } }* @"main$mapbucket", i32 0, i32 0, i32 0), i32 1, i8 1, i8 8, i8 0 }
+@"main$mapbucket" = internal unnamed_addr global { [8 x i8], i8*, { i8, [7 x i8] }, { { [7 x i8]*, [4 x i8] }, [56 x i8] } } { [8 x i8] c"\04\00\00\00\00\00\00\00", i8* null, { i8, [7 x i8] } { i8 1, [7 x i8] zeroinitializer }, { { [7 x i8]*, [4 x i8] }, [56 x i8] } { { [7 x i8]*, [4 x i8] } { [7 x i8]* @main.init.string, [4 x i8] c"\07\00\00\00" }, [56 x i8] zeroinitializer } }
+@"main$map.1" = internal global %runtime.hashmap { %runtime.hashmap* null, i8* getelementptr inbounds ({ [8 x i8], i8*, { i8, [7 x i8] }, { i8, [7 x i8] } }, { [8 x i8], i8*, { i8, [7 x i8] }, { i8, [7 x i8] } }* @"main$mapbucket.2", i32 0, i32 0, i32 0), i32 1, i8 1, i8 1, i8 0 }
+@"main$mapbucket.2" = internal unnamed_addr global { [8 x i8], i8*, { i8, [7 x i8] }, { i8, [7 x i8] } } { [8 x i8] c"\04\00\00\00\00\00\00\00", i8* null, { i8, [7 x i8] } { i8 1, [7 x i8] zeroinitializer }, { i8, [7 x i8] } { i8 2, [7 x i8] zeroinitializer } }
+@"main$map.3" = internal global %runtime.hashmap { %runtime.hashmap* null, i8* getelementptr inbounds ({ [8 x i8], i8*, { { [7 x i8]*, [4 x i8] }, [56 x i8] }, { i8, [7 x i8] } }, { [8 x i8], i8*, { { [7 x i8]*, [4 x i8] }, [56 x i8] }, { i8, [7 x i8] } }* @"main$mapbucket.4", i32 0, i32 0, i32 0), i32 1, i8 8, i8 1, i8 0 }
+@"main$mapbucket.4" = internal unnamed_addr global { [8 x i8], i8*, { { [7 x i8]*, [4 x i8] }, [56 x i8] }, { i8, [7 x i8] } } { [8 x i8] c"x\00\00\00\00\00\00\00", i8* null, { { [7 x i8]*, [4 x i8] }, [56 x i8] } { { [7 x i8]*, [4 x i8] } { [7 x i8]* @main.init.string, [4 x i8] c"\07\00\00\00" }, [56 x i8] zeroinitializer }, { i8, [7 x i8] } { i8 2, [7 x i8] zeroinitializer } }
 
 define void @runtime.initAll() unnamed_addr {
 entry:
-  call void @runtime.hashmapBinarySet(%runtime.hashmap* @"main$map.4", i8* @"main$alloca.2", i8* @"main$alloca.3", i8* undef, i8* null)
-  call void @runtime.hashmapStringSet(%runtime.hashmap* @"main$map.6", i8* getelementptr inbounds ([7 x i8], [7 x i8]* @main.init.string, i32 0, i32 0), i32 7, i8* @"main$alloca.5", i8* undef, i8* null)
   ret void
 }

interp/testdata/scan.ll

@@ -1,78 +0,0 @@
-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 }
-
-declare i1 @runtime.typeAssert(i64, %runtime.typecodeID*, i8*, i8*)
-declare i1 @runtime.interfaceImplements(i64, i8**)
-
-define i64 @returnsConst() {
-  ret i64 0
-}
-
-define i64 @returnsArg(i64 %arg) {
-  ret i64 %arg
-}
-
-declare i64 @externalCall()
-
-define i64 @externalCallOnly() {
-  %result = call i64 @externalCall()
-  ret i64 0
-}
-
-define i64 @externalCallAndReturn() {
-  %result = call i64 @externalCall()
-  ret i64 %result
-}
-
-define i64 @externalCallBranch() {
-  %result = call i64 @externalCall()
-  %zero = icmp eq i64 %result, 0
-  br i1 %zero, label %if.then, label %if.done
-
-if.then:
-  ret i64 2
-
-if.done:
-  ret i64 4
-}
-
-@cleanGlobalInt = global i64 5
-define i64 @readCleanGlobal() {
-  %global = load i64, i64* @cleanGlobalInt
-  ret i64 %global
-}
-
-@dirtyGlobalInt = global i64 5
-define i64 @readDirtyGlobal() {
-  %global = load i64, i64* @dirtyGlobalInt
-  ret i64 %global
-}
-
-declare i64* @getDirtyPointer()
-
-define void @storeToPointer() {
-  %ptr = call i64* @getDirtyPointer()
-  store i64 3, i64* %ptr
-  ret void
-}
-
-@functionPointer = global i64()* null
-define i64 @callFunctionPointer() {
-  %fp = load i64()*, i64()** @functionPointer
-  %result = call i64 %fp()
-  ret i64 %result
-}
-
-define i1 @callTypeAssert() {
-  ; Note: parameters are not realistic.
-  %ok = call i1 @runtime.typeAssert(i64 0, %runtime.typecodeID* null, i8* undef, i8* null)
-  ret i1 %ok
-}
-
-define i1 @callInterfaceImplements() {
-  ; Note: parameters are not realistic.
-  %ok = call i1 @runtime.interfaceImplements(i64 0, i8** null)
-  ret i1 %ok
-}

interp/testdata/slice-copy.out.ll

@@ -1,6 +1,8 @@
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64--linux"
 
+@"main$alloc.1" = internal unnamed_addr constant [6 x i8] c"\05\00{\00\00\04"
+
 declare void @runtime.printuint8(i8) local_unnamed_addr
 
 declare void @runtime.printint16(i16) local_unnamed_addr
@@ -15,6 +17,7 @@ entry:
   call void @runtime.printuint8(i8 3)
   call void @runtime.printuint8(i8 3)
   call void @runtime.printint16(i16 5)
-  call void @runtime.printint16(i16 5)
+  %int16SliceDst.val = load i16, i16* bitcast ([6 x i8]* @"main$alloc.1" to i16*)
+  call void @runtime.printint16(i16 %int16SliceDst.val)
   ret void
 }

interp/utils.go

@@ -1,126 +0,0 @@
-package interp
-
-import (
-	"errors"
-
-	"tinygo.org/x/go-llvm"
-)
-
-// Return a list of values (actually, instructions) where this value is used as
-// an operand.
-func getUses(value llvm.Value) []llvm.Value {
-	var uses []llvm.Value
-	use := value.FirstUse()
-	for !use.IsNil() {
-		uses = append(uses, use.User())
-		use = use.NextUse()
-	}
-	return uses
-}
-
-// getStringBytes loads the byte slice of a Go string represented as a
-// {ptr, len} pair.
-func getStringBytes(strPtr Value, strLen llvm.Value) ([]byte, error) {
-	if !strLen.IsConstant() {
-		return nil, errors.New("getStringBytes with a non-constant length")
-	}
-	buf := make([]byte, strLen.ZExtValue())
-	for i := range buf {
-		gep, err := strPtr.GetElementPtr([]uint32{uint32(i)})
-		if err != nil {
-			return nil, err
-		}
-		c, err := gep.Load()
-		if err != nil {
-			return nil, err
-		}
-		buf[i] = byte(c.ZExtValue())
-	}
-	return buf, nil
-}
-
-// getLLVMIndices converts an []uint32 into an []llvm.Value, for use in
-// llvm.ConstGEP.
-func getLLVMIndices(int32Type llvm.Type, indices []uint32) []llvm.Value {
-	llvmIndices := make([]llvm.Value, len(indices))
-	for i, index := range indices {
-		llvmIndices[i] = llvm.ConstInt(int32Type, uint64(index), false)
-	}
-	return llvmIndices
-}
-
-// Return true if this type is a scalar value (integer or floating point), false
-// otherwise.
-func isScalar(t llvm.Type) bool {
-	switch t.TypeKind() {
-	case llvm.IntegerTypeKind, llvm.FloatTypeKind, llvm.DoubleTypeKind:
-		return true
-	default:
-		return false
-	}
-}
-
-// isPointerNil returns whether this is a nil pointer or not. The ok value
-// indicates whether the result is certain: if it is false the result boolean is
-// not valid.
-func isPointerNil(v llvm.Value) (result bool, ok bool) {
-	if !v.IsAConstantExpr().IsNil() {
-		switch v.Opcode() {
-		case llvm.IntToPtr:
-			// Whether a constant inttoptr is nil is easy to
-			// determine.
-			result, ok = isZero(v.Operand(0))
-			if ok {
-				return
-			}
-		case llvm.BitCast, llvm.GetElementPtr:
-			// These const instructions are just a kind of wrappers for the
-			// underlying pointer.
-			return isPointerNil(v.Operand(0))
-		}
-	}
-	if !v.IsAConstantPointerNull().IsNil() {
-		// A constant pointer null is always null, of course.
-		return true, true
-	}
-	if !v.IsAGlobalValue().IsNil() {
-		// A global value is never null.
-		return false, true
-	}
-	return false, false // not valid
-}
-
-// isZero returns whether the value in v is the integer zero, and whether that
-// can be known right now.
-func isZero(v llvm.Value) (result bool, ok bool) {
-	if !v.IsAConstantExpr().IsNil() {
-		switch v.Opcode() {
-		case llvm.PtrToInt:
-			return isPointerNil(v.Operand(0))
-		}
-	}
-	if !v.IsAConstantInt().IsNil() {
-		val := v.ZExtValue()
-		return val == 0, true
-	}
-	return false, false // not valid
-}
-
-// unwrap returns the underlying value, with GEPs removed. This can be useful to
-// get the underlying global of a GEP pointer.
-func unwrap(value llvm.Value) llvm.Value {
-	for {
-		if !value.IsAConstantExpr().IsNil() {
-			switch value.Opcode() {
-			case llvm.GetElementPtr:
-				value = value.Operand(0)
-				continue
-			}
-		} else if !value.IsAGetElementPtrInst().IsNil() {
-			value = value.Operand(0)
-			continue
-		}
-		break
-	}
-	return value
-}

interp/values.go

@@ -1,443 +0,0 @@
-package interp
-
-// This file provides a litte bit of abstraction around LLVM values.
-
-import (
-	"errors"
-	"strconv"
-
-	"tinygo.org/x/go-llvm"
-)
-
-// A Value is a LLVM value with some extra methods attached for easier
-// interpretation.
-type Value interface {
-	Value() llvm.Value                     // returns a LLVM value
-	Type() llvm.Type                       // equal to Value().Type()
-	IsConstant() bool                      // returns true if this value is a constant value
-	Load() (llvm.Value, error)             // dereference a pointer
-	Store(llvm.Value) error                // store to a pointer
-	GetElementPtr([]uint32) (Value, error) // returns an interior pointer
-	String() string                        // string representation, for debugging
-}
-
-// A type that simply wraps a LLVM constant value.
-type LocalValue struct {
-	Eval       *Eval
-	Underlying llvm.Value
-}
-
-// Value implements Value by returning the constant value itself.
-func (v *LocalValue) Value() llvm.Value {
-	return v.Underlying
-}
-
-func (v *LocalValue) Type() llvm.Type {
-	return v.Underlying.Type()
-}
-
-func (v *LocalValue) IsConstant() bool {
-	if _, ok := v.Eval.dirtyGlobals[unwrap(v.Underlying)]; ok {
-		return false
-	}
-	return v.Underlying.IsConstant()
-}
-
-// Load loads a constant value if this is a constant pointer.
-func (v *LocalValue) Load() (llvm.Value, error) {
-	if !v.Underlying.IsAGlobalVariable().IsNil() {
-		return v.Underlying.Initializer(), nil
-	}
-	switch v.Underlying.Opcode() {
-	case llvm.GetElementPtr:
-		indices := v.getConstGEPIndices()
-		if indices[0] != 0 {
-			return llvm.Value{}, errors.New("invalid GEP")
-		}
-		global := v.Eval.getValue(v.Underlying.Operand(0))
-		agg, err := global.Load()
-		if err != nil {
-			return llvm.Value{}, err
-		}
-		return llvm.ConstExtractValue(agg, indices[1:]), nil
-	case llvm.BitCast:
-		return llvm.Value{}, errors.New("interp: load from a bitcast")
-	default:
-		return llvm.Value{}, errors.New("interp: load from a constant")
-	}
-}
-
-// Store stores to the underlying value if the value type is a pointer type,
-// otherwise it returns an error.
-func (v *LocalValue) Store(value llvm.Value) error {
-	if !v.Underlying.IsAGlobalVariable().IsNil() {
-		if !value.IsConstant() {
-			v.MarkDirty()
-			v.Eval.builder.CreateStore(value, v.Underlying)
-		} else {
-			v.Underlying.SetInitializer(value)
-		}
-		return nil
-	}
-	if !value.IsConstant() {
-		v.MarkDirty()
-		v.Eval.builder.CreateStore(value, v.Underlying)
-		return nil
-	}
-	switch v.Underlying.Opcode() {
-	case llvm.GetElementPtr:
-		indices := v.getConstGEPIndices()
-		if indices[0] != 0 {
-			return errors.New("invalid GEP")
-		}
-		global := &LocalValue{v.Eval, v.Underlying.Operand(0)}
-		agg, err := global.Load()
-		if err != nil {
-			return err
-		}
-		agg = llvm.ConstInsertValue(agg, value, indices[1:])
-		return global.Store(agg)
-	default:
-		return errors.New("interp: store on a constant")
-	}
-}
-
-// GetElementPtr returns a GEP when the underlying value is of pointer type.
-func (v *LocalValue) GetElementPtr(indices []uint32) (Value, error) {
-	if !v.Underlying.IsAGlobalVariable().IsNil() {
-		int32Type := v.Underlying.Type().Context().Int32Type()
-		gep := llvm.ConstGEP(v.Underlying, getLLVMIndices(int32Type, indices))
-		return &LocalValue{v.Eval, gep}, nil
-	}
-	if !v.Underlying.IsAConstantExpr().IsNil() {
-		switch v.Underlying.Opcode() {
-		case llvm.GetElementPtr, llvm.IntToPtr, llvm.BitCast:
-			int32Type := v.Underlying.Type().Context().Int32Type()
-			llvmIndices := getLLVMIndices(int32Type, indices)
-			return &LocalValue{v.Eval, llvm.ConstGEP(v.Underlying, llvmIndices)}, nil
-		}
-	}
-	return nil, errors.New("interp: unknown GEP")
-}
-
-// stripPointerCasts removes all const bitcasts from pointer values, if there
-// are any.
-func (v *LocalValue) stripPointerCasts() *LocalValue {
-	value := v.Underlying
-	for {
-		if !value.IsAConstantExpr().IsNil() {
-			switch value.Opcode() {
-			case llvm.BitCast:
-				value = value.Operand(0)
-				continue
-			}
-		}
-		return &LocalValue{
-			Eval:       v.Eval,
-			Underlying: value,
-		}
-	}
-}
-
-func (v *LocalValue) String() string {
-	isConstant := "false"
-	if v.IsConstant() {
-		isConstant = "true"
-	}
-	return "&LocalValue{Type: " + v.Type().String() + ", IsConstant: " + isConstant + "}"
-}
-
-// getConstGEPIndices returns indices of this constant GEP, if this is a GEP
-// instruction. If it is not, the behavior is undefined.
-func (v *LocalValue) getConstGEPIndices() []uint32 {
-	indices := make([]uint32, v.Underlying.OperandsCount()-1)
-	for i := range indices {
-		operand := v.Underlying.Operand(i + 1)
-		indices[i] = uint32(operand.ZExtValue())
-	}
-	return indices
-}
-
-// MarkDirty marks this global as dirty, meaning that every load from and store
-// to this global (from now on) must be performed at runtime.
-func (v *LocalValue) MarkDirty() {
-	underlying := unwrap(v.Underlying)
-	if underlying.IsAGlobalVariable().IsNil() {
-		panic("trying to mark a non-global as dirty")
-	}
-	if !v.IsConstant() {
-		return // already dirty
-	}
-	v.Eval.dirtyGlobals[underlying] = struct{}{}
-}
-
-// MapValue implements a Go map which is created at compile time and stored as a
-// global variable.
-type MapValue struct {
-	Eval       *Eval
-	PkgName    string
-	Underlying llvm.Value
-	Keys       []Value
-	Values     []Value
-	KeySize    int
-	ValueSize  int
-	KeyType    llvm.Type
-	ValueType  llvm.Type
-}
-
-func (v *MapValue) newBucket() llvm.Value {
-	ctx := v.Eval.Mod.Context()
-	i8ptrType := llvm.PointerType(ctx.Int8Type(), 0)
-	bucketType := ctx.StructType([]llvm.Type{
-		llvm.ArrayType(ctx.Int8Type(), 8), // tophash
-		i8ptrType,                         // next bucket
-		llvm.ArrayType(v.KeyType, 8),      // key type
-		llvm.ArrayType(v.ValueType, 8),    // value type
-	}, false)
-	bucketValue := llvm.ConstNull(bucketType)
-	bucket := llvm.AddGlobal(v.Eval.Mod, bucketType, v.PkgName+"$mapbucket")
-	bucket.SetInitializer(bucketValue)
-	bucket.SetLinkage(llvm.InternalLinkage)
-	bucket.SetUnnamedAddr(true)
-	return bucket
-}
-
-// Value returns a global variable which is a pointer to the actual hashmap.
-func (v *MapValue) Value() llvm.Value {
-	if !v.Underlying.IsNil() {
-		return v.Underlying
-	}
-
-	ctx := v.Eval.Mod.Context()
-	i8ptrType := llvm.PointerType(ctx.Int8Type(), 0)
-
-	var firstBucketGlobal llvm.Value
-	if len(v.Keys) == 0 {
-		// there are no buckets
-		firstBucketGlobal = llvm.ConstPointerNull(i8ptrType)
-	} else {
-		// create initial bucket
-		firstBucketGlobal = v.newBucket()
-	}
-
-	// Insert each key/value pair in the hashmap.
-	bucketGlobal := firstBucketGlobal
-	for i, key := range v.Keys {
-		var keyBuf []byte
-		llvmKey := key.Value()
-		llvmValue := v.Values[i].Value()
-		if key.Type().TypeKind() == llvm.StructTypeKind && key.Type().StructName() == "runtime._string" {
-			keyPtr := llvm.ConstExtractValue(llvmKey, []uint32{0})
-			keyLen := llvm.ConstExtractValue(llvmKey, []uint32{1})
-			keyPtrVal := v.Eval.getValue(keyPtr)
-			var err error
-			keyBuf, err = getStringBytes(keyPtrVal, keyLen)
-			if err != nil {
-				panic(err) // TODO
-			}
-		} else if key.Type().TypeKind() == llvm.IntegerTypeKind {
-			keyBuf = make([]byte, v.Eval.TargetData.TypeAllocSize(key.Type()))
-			n := key.Value().ZExtValue()
-			for i := range keyBuf {
-				keyBuf[i] = byte(n)
-				n >>= 8
-			}
-		} else if key.Type().TypeKind() == llvm.ArrayTypeKind &&
-			key.Type().ElementType().TypeKind() == llvm.IntegerTypeKind &&
-			key.Type().ElementType().IntTypeWidth() == 8 {
-			keyBuf = make([]byte, v.Eval.TargetData.TypeAllocSize(key.Type()))
-			for i := range keyBuf {
-				keyBuf[i] = byte(llvm.ConstExtractValue(llvmKey, []uint32{uint32(i)}).ZExtValue())
-			}
-		} else {
-			panic("interp: map key type not implemented: " + key.Type().String())
-		}
-		hash := v.hash(keyBuf)
-
-		if i%8 == 0 && i != 0 {
-			// Bucket is full, create a new one.
-			newBucketGlobal := v.newBucket()
-			zero := llvm.ConstInt(ctx.Int32Type(), 0, false)
-			newBucketPtr := llvm.ConstInBoundsGEP(newBucketGlobal, []llvm.Value{zero})
-			newBucketPtrCast := llvm.ConstBitCast(newBucketPtr, i8ptrType)
-			// insert pointer into old bucket
-			bucket := bucketGlobal.Initializer()
-			bucket = llvm.ConstInsertValue(bucket, newBucketPtrCast, []uint32{1})
-			bucketGlobal.SetInitializer(bucket)
-			// switch to next bucket
-			bucketGlobal = newBucketGlobal
-		}
-
-		tophashValue := llvm.ConstInt(ctx.Int8Type(), uint64(v.topHash(hash)), false)
-		bucket := bucketGlobal.Initializer()
-		bucket = llvm.ConstInsertValue(bucket, tophashValue, []uint32{0, uint32(i % 8)})
-		bucket = llvm.ConstInsertValue(bucket, llvmKey, []uint32{2, uint32(i % 8)})
-		bucket = llvm.ConstInsertValue(bucket, llvmValue, []uint32{3, uint32(i % 8)})
-		bucketGlobal.SetInitializer(bucket)
-	}
-
-	// Create the hashmap itself.
-	zero := llvm.ConstInt(ctx.Int32Type(), 0, false)
-	bucketPtr := llvm.ConstInBoundsGEP(firstBucketGlobal, []llvm.Value{zero})
-	hashmapType := v.Type()
-	hashmap := llvm.ConstNamedStruct(hashmapType, []llvm.Value{
-		llvm.ConstPointerNull(llvm.PointerType(hashmapType, 0)),                        // next
-		llvm.ConstBitCast(bucketPtr, i8ptrType),                                        // buckets
-		llvm.ConstInt(hashmapType.StructElementTypes()[2], uint64(len(v.Keys)), false), // count
-		llvm.ConstInt(ctx.Int8Type(), uint64(v.KeySize), false),                        // keySize
-		llvm.ConstInt(ctx.Int8Type(), uint64(v.ValueSize), false),                      // valueSize
-		llvm.ConstInt(ctx.Int8Type(), 0, false),                                        // bucketBits
-	})
-
-	// Create a pointer to this hashmap.
-	hashmapPtr := llvm.AddGlobal(v.Eval.Mod, hashmap.Type(), v.PkgName+"$map")
-	hashmapPtr.SetInitializer(hashmap)
-	hashmapPtr.SetLinkage(llvm.InternalLinkage)
-	hashmapPtr.SetUnnamedAddr(true)
-	v.Underlying = llvm.ConstInBoundsGEP(hashmapPtr, []llvm.Value{zero})
-	return v.Underlying
-}
-
-// Type returns type runtime.hashmap, which is the actual hashmap type.
-func (v *MapValue) Type() llvm.Type {
-	return v.Eval.Mod.GetTypeByName("runtime.hashmap")
-}
-
-func (v *MapValue) IsConstant() bool {
-	return true // TODO: dirty maps
-}
-
-// Load panics: maps are of reference type so cannot be dereferenced.
-func (v *MapValue) Load() (llvm.Value, error) {
-	panic("interp: load from a map")
-}
-
-// Store returns an error: maps are of reference type so cannot be stored to.
-func (v *MapValue) Store(value llvm.Value) error {
-	// This must be a bug, but it might be helpful to indicate the location
-	// anyway.
-	return errors.New("interp: store on a map")
-}
-
-// GetElementPtr panics: maps are of reference type so their (interior)
-// addresses cannot be calculated.
-func (v *MapValue) GetElementPtr(indices []uint32) (Value, error) {
-	return nil, errors.New("interp: GEP on a map")
-}
-
-// PutString does a map assign operation, assuming that the map is of type
-// map[string]T.
-func (v *MapValue) PutString(keyBuf, keyLen, valPtr *LocalValue) error {
-	if !v.Underlying.IsNil() {
-		return errors.New("map already created")
-	}
-
-	if valPtr.Underlying.Opcode() == llvm.BitCast {
-		valPtr = &LocalValue{v.Eval, valPtr.Underlying.Operand(0)}
-	}
-	value, err := valPtr.Load()
-	if err != nil {
-		return err
-	}
-	if v.ValueType.IsNil() {
-		v.ValueType = value.Type()
-		if int(v.Eval.TargetData.TypeAllocSize(v.ValueType)) != v.ValueSize {
-			return errors.New("interp: map store value type has the wrong size")
-		}
-	} else {
-		if value.Type() != v.ValueType {
-			return errors.New("interp: map store value type is inconsistent")
-		}
-	}
-
-	keyType := v.Eval.Mod.GetTypeByName("runtime._string")
-	v.KeyType = keyType
-	key := llvm.ConstNull(keyType)
-	key = llvm.ConstInsertValue(key, keyBuf.Value(), []uint32{0})
-	key = llvm.ConstInsertValue(key, keyLen.Value(), []uint32{1})
-
-	// TODO: avoid duplicate keys
-	v.Keys = append(v.Keys, &LocalValue{v.Eval, key})
-	v.Values = append(v.Values, &LocalValue{v.Eval, value})
-
-	return nil
-}
-
-// PutBinary does a map assign operation.
-func (v *MapValue) PutBinary(keyPtr, valPtr *LocalValue) error {
-	if !v.Underlying.IsNil() {
-		return errors.New("map already created")
-	}
-
-	if valPtr.Underlying.Opcode() == llvm.BitCast {
-		valPtr = &LocalValue{v.Eval, valPtr.Underlying.Operand(0)}
-	}
-	value, err := valPtr.Load()
-	if err != nil {
-		return err
-	}
-	if v.ValueType.IsNil() {
-		v.ValueType = value.Type()
-		if int(v.Eval.TargetData.TypeAllocSize(v.ValueType)) != v.ValueSize {
-			return errors.New("interp: map store value type has the wrong size")
-		}
-	} else {
-		if value.Type() != v.ValueType {
-			return errors.New("interp: map store value type is inconsistent")
-		}
-	}
-
-	if !keyPtr.Underlying.IsAConstantExpr().IsNil() {
-		if keyPtr.Underlying.Opcode() == llvm.BitCast {
-			keyPtr = &LocalValue{v.Eval, keyPtr.Underlying.Operand(0)}
-		} else if keyPtr.Underlying.Opcode() == llvm.GetElementPtr {
-			keyPtr = &LocalValue{v.Eval, keyPtr.Underlying.Operand(0)}
-		}
-	}
-	key, err := keyPtr.Load()
-	if err != nil {
-		return err
-	}
-	if v.KeyType.IsNil() {
-		v.KeyType = key.Type()
-		if int(v.Eval.TargetData.TypeAllocSize(v.KeyType)) != v.KeySize {
-			return errors.New("interp: map store key type has the wrong size")
-		}
-	} else {
-		if key.Type() != v.KeyType {
-			return errors.New("interp: map store key type is inconsistent")
-		}
-	}
-
-	// TODO: avoid duplicate keys
-	v.Keys = append(v.Keys, &LocalValue{v.Eval, key})
-	v.Values = append(v.Values, &LocalValue{v.Eval, value})
-
-	return nil
-}
-
-// Get FNV-1a hash of this string.
-//
-// https://en.wikipedia.org/wiki/Fowler%E2%80%93Noll%E2%80%93Vo_hash_function#FNV-1a_hash
-func (v *MapValue) hash(data []byte) uint32 {
-	var result uint32 = 2166136261 // FNV offset basis
-	for _, c := range data {
-		result ^= uint32(c)
-		result *= 16777619 // FNV prime
-	}
-	return result
-}
-
-// Get the topmost 8 bits of the hash, without using a special value (like 0).
-func (v *MapValue) topHash(hash uint32) uint8 {
-	tophash := uint8(hash >> 24)
-	if tophash < 1 {
-		// 0 means empty slot, so make it bigger.
-		tophash += 1
-	}
-	return tophash
-}
-
-func (v *MapValue) String() string {
-	return "&MapValue{KeySize: " + strconv.Itoa(v.KeySize) + ", ValueSize: " + strconv.Itoa(v.ValueSize) + "}"
-}

main.go

@@ -348,12 +348,14 @@ func FlashGDB(pkgName string, ocdOutput bool, options *compileopts.Options) erro
 		switch gdbInterface {
 		case "msd", "command", "":
 			if len(config.Target.Emulator) != 0 {
-				// Assume QEMU as an emulator.
 				if config.Target.Emulator[0] == "mgba" {
 					gdbInterface = "mgba"
+				} else if config.Target.Emulator[0] == "simavr" {
+					gdbInterface = "simavr"
 				} else if strings.HasPrefix(config.Target.Emulator[0], "qemu-system-") {
 					gdbInterface = "qemu"
 				} else {
+					// Assume QEMU as an emulator.
 					gdbInterface = "qemu-user"
 				}
 			} else if openocdInterface != "" && config.Target.OpenOCDTarget != "" {
@@ -429,6 +431,14 @@ func FlashGDB(pkgName string, ocdOutput bool, options *compileopts.Options) erro
 			daemon = exec.Command(config.Target.Emulator[0], args...)
 			daemon.Stdout = os.Stdout
 			daemon.Stderr = os.Stderr
+		case "simavr":
+			gdbCommands = append(gdbCommands, "target remote :1234")
+
+			// Run in an emulator.
+			args := append(config.Target.Emulator[1:], "-g", result.Binary)
+			daemon = exec.Command(config.Target.Emulator[0], args...)
+			daemon.Stdout = os.Stdout
+			daemon.Stderr = os.Stderr
 		case "msd":
 			return errors.New("gdb is not supported for drag-and-drop programmable devices")
 		default:

src/device/esp/esp32.S

@@ -24,7 +24,7 @@ call_start_cpu0:
     wsr.ps a2
     rsync
 
-    // Set WINDOWBASE to 1 << WINDOWSTART.
+    // Set WINDOWSTART to 1 << WINDOWBASE.
     rsr.windowbase  a2
     ssl  a2
     movi a2, 1
@@ -42,8 +42,13 @@ call_start_cpu0:
     wsr.ps a2
     rsync
 
+    // Enable the FPU (coprocessor 0 so the lowest bit).
+    movi a2, 1
+    wsr.cpenable a2
+    rsync
+
     // Jump to the runtime start function written in Go.
-    j main
+    call4 main
 
 .section .text.tinygo_scanCurrentStack
 .global tinygo_scanCurrentStack

src/internal/task/task_stack.go

@@ -45,6 +45,11 @@ func Pause() {
 	currentTask.state.pause()
 }
 
+//export tinygo_pause
+func pause() {
+	Pause()
+}
+
 // Resume the task until it pauses or completes.
 // This may only be called from the scheduler.
 func (t *Task) Resume() {
@@ -58,10 +63,32 @@ func (s *state) initialize(fn uintptr, args unsafe.Pointer, stackSize uintptr) {
 	// Create a stack.
 	stack := make([]uintptr, stackSize/unsafe.Sizeof(uintptr(0)))
 
+	// Set up the stack canary, a random number that should be checked when
+	// switching from the task back to the scheduler. The stack canary pointer
+	// points to the first word of the stack. If it has changed between now and
+	// the next stack switch, there was a stack overflow.
+	s.canaryPtr = &stack[0]
+	*s.canaryPtr = stackCanary
+
+	// Get a pointer to the top of the stack, where the initial register values
+	// are stored. They will be popped off the stack on the first stack switch
+	// to the goroutine, and will start running tinygo_startTask (this setup
+	// happens in archInit).
+	r := (*calleeSavedRegs)(unsafe.Pointer(&stack[uintptr(len(stack))-(unsafe.Sizeof(calleeSavedRegs{})/unsafe.Sizeof(uintptr(0)))]))
+
 	// Invoke architecture-specific initialization.
-	s.archInit(stack, fn, args)
+	s.archInit(r, fn, args)
 }
 
+//export tinygo_swapTask
+func swapTask(oldStack uintptr, newStack *uintptr)
+
+// startTask is a small wrapper function that sets up the first (and only)
+// argument to the new goroutine and makes sure it is exited when the goroutine
+// finishes.
+//go:extern tinygo_startTask
+var startTask [0]uint8
+
 //go:linkname runqueuePushBack runtime.runqueuePushBack
 func runqueuePushBack(*Task)
 

src/internal/task/task_stack_avr.S

@@ -0,0 +1,99 @@
+.section .bss.tinygo_systemStack
+.global tinygo_systemStack
+.type   tinygo_systemStack, %object
+tinygo_systemStack:
+    .short 0
+
+.section .text.tinygo_startTask
+.global  tinygo_startTask
+.type    tinygo_startTask, %function
+tinygo_startTask:
+    // Small assembly stub for starting a goroutine. This is already run on the
+    // new stack, with the callee-saved registers already loaded.
+    // Most importantly, r2r3 contain the pc of the to-be-started function and
+    // r4r5 contain the only argument it is given. Multiple arguments are packed
+    // into one by storing them in a new allocation.
+
+    // Set the first argument of the goroutine start wrapper, which contains all
+    // the arguments.
+    movw  r24, r4
+
+    // Branch to the "goroutine start" function. Note that the Z register is
+    // call-clobbered, so does not need to be restored after use.
+    movw  Z, r2
+    icall
+
+    // After return, exit this goroutine. This is a tail call.
+#if __AVR_ARCH__ == 2 || __AVR_ARCH__ == 25
+    // Small memory devices (≤8kB flash) that do not have the long call
+    // instruction availble will need to use rcall instead.
+    // Note that they will probably not be able to run more than the main
+    // goroutine anyway, but this file is compiled for all AVRs so it needs to
+    // compile at least.
+    rcall  tinygo_pause
+#else
+    // Other devices can (and must) use the regular call instruction.
+    call tinygo_pause
+#endif
+
+.global tinygo_swapTask
+.type tinygo_swapTask, %function
+tinygo_swapTask:
+    // This function gets the following parameters:
+    // r24:r25 = newStack uintptr
+    // r22:r23 = oldStack *uintptr
+
+    // Save all call-saved registers:
+    // https://gcc.gnu.org/wiki/avr-gcc#Call-Saved_Registers
+    push r29 // Y
+    push r28 // Y
+    push r17
+    push r16
+    push r15
+    push r14
+    push r13
+    push r12
+    push r11
+    push r10
+    push r9
+    push r8
+    push r7
+    push r6
+    push r5
+    push r4
+    push r3
+    push r2
+
+    // Save the current stack pointer in oldStack.
+    in  r2, 0x3d; SPL
+    in  r3, 0x3e; SPH
+    movw Y, r22
+    std Y+0, r2
+    std Y+1, r3
+
+    // Switch to the new stack pointer.
+    out  0x3d, r24; SPL
+    out  0x3e, r25; SPH
+
+    // Load saved register from the new stack.
+    pop r2
+    pop r3
+    pop r4
+    pop r5
+    pop r6
+    pop r7
+    pop r8
+    pop r9
+    pop r10
+    pop r11
+    pop r12
+    pop r13
+    pop r14
+    pop r15
+    pop r16
+    pop r17
+    pop r28 // Y
+    pop r29 // Y
+
+    // Return into the new task, as if tinygo_swapTask was a regular call.
+    ret

src/internal/task/task_stack_avr.go

@@ -4,9 +4,12 @@ package task
 
 import "unsafe"
 
+//go:extern tinygo_systemStack
+var systemStack uintptr
+
 // calleeSavedRegs is the list of registers that must be saved and restored when
-// switching between tasks. Also see scheduler_avr.S that relies on the
-// exact layout of this struct.
+// switching between tasks. Also see task_stack_avr.S that relies on the exact
+// layout of this struct.
 //
 // https://gcc.gnu.org/wiki/avr-gcc#Call-Saved_Registers
 type calleeSavedRegs struct {
@@ -23,34 +26,15 @@ type calleeSavedRegs struct {
 	pc uintptr
 }
 
-// registers gets a pointer to the registers stored at the top of the stack.
-func (s *state) registers() *calleeSavedRegs {
-	return (*calleeSavedRegs)(unsafe.Pointer(s.sp + 1))
-}
-
-// startTask is a small wrapper function that sets up the first (and only)
-// argument to the new goroutine and makes sure it is exited when the goroutine
-// finishes.
-//go:extern tinygo_startTask
-var startTask [0]uint8
-
 // archInit runs architecture-specific setup for the goroutine startup.
 // Note: adding //go:noinline to work around an AVR backend bug.
 //go:noinline
-func (s *state) archInit(stack []uintptr, fn uintptr, args unsafe.Pointer) {
-	// Set up the stack canary, a random number that should be checked when
-	// switching from the task back to the scheduler. The stack canary pointer
-	// points to the first word of the stack. If it has changed between now and
-	// the next stack switch, there was a stack overflow.
-	s.canaryPtr = &stack[0]
-	*s.canaryPtr = stackCanary
-
+func (s *state) archInit(r *calleeSavedRegs, fn uintptr, args unsafe.Pointer) {
 	// Store the initial sp for the startTask function (implemented in assembly).
-	s.sp = uintptr(unsafe.Pointer(&stack[uintptr(len(stack))-(unsafe.Sizeof(calleeSavedRegs{})/unsafe.Sizeof(uintptr(0)))])) - 1
+	s.sp = uintptr(unsafe.Pointer(r)) - 1
 
 	// Initialize the registers.
 	// These will be popped off of the stack on the first resume of the goroutine.
-	r := s.registers()
 
 	// Start the function at tinygo_startTask.
 	startTask := uintptr(unsafe.Pointer(&startTask))
@@ -67,20 +51,17 @@ func (s *state) archInit(stack []uintptr, fn uintptr, args unsafe.Pointer) {
 }
 
 func (s *state) resume() {
-	switchToTask(s.sp)
+	swapTask(s.sp, &systemStack)
 }
 
-//export tinygo_switchToTask
-func switchToTask(uintptr)
-
-//export tinygo_switchToScheduler
-func switchToScheduler(*uintptr)
-
 func (s *state) pause() {
-	switchToScheduler(&s.sp)
+	newStack := systemStack
+	systemStack = 0
+	swapTask(newStack, &s.sp)
 }
 
-//export tinygo_pause
-func pause() {
-	Pause()
+// SystemStack returns the system stack pointer when called from a task stack.
+// When called from the system stack, it returns 0.
+func SystemStack() uintptr {
+	return systemStack
 }

src/internal/task/task_stack_cortexm.S

@@ -30,17 +30,6 @@ tinygo_startTask:
     .cfi_endproc
 .size tinygo_startTask, .-tinygo_startTask
 
-.section .text.tinygo_getSystemStackPointer
-.global  tinygo_getSystemStackPointer
-.type    tinygo_getSystemStackPointer, %function
-tinygo_getSystemStackPointer:
-    .cfi_startproc
-    // The system stack pointer is always stored in the MSP register.
-    mrs r0, MSP
-    bx lr
-    .cfi_endproc
-.size tinygo_getSystemStackPointer, .-tinygo_getSystemStackPointer
-
 .section .text.tinygo_switchToScheduler
 .global  tinygo_switchToScheduler
 .type    tinygo_switchToScheduler, %function

src/internal/task/task_stack_cortexm.go

@@ -2,10 +2,13 @@
 
 package task
 
-import "unsafe"
+import (
+	"device/arm"
+	"unsafe"
+)
 
 // calleeSavedRegs is the list of registers that must be saved and restored when
-// switching between tasks. Also see scheduler_cortexm.S that relies on the
+// switching between tasks. Also see task_stack_cortexm.S that relies on the
 // exact layout of this struct.
 type calleeSavedRegs struct {
 	r4  uintptr
@@ -20,34 +23,15 @@ type calleeSavedRegs struct {
 	pc uintptr
 }
 
-// registers gets a pointer to the registers stored at the top of the stack.
-func (s *state) registers() *calleeSavedRegs {
-	return (*calleeSavedRegs)(unsafe.Pointer(s.sp))
-}
-
-// startTask is a small wrapper function that sets up the first (and only)
-// argument to the new goroutine and makes sure it is exited when the goroutine
-// finishes.
-//go:extern tinygo_startTask
-var startTask [0]uint8
-
 // archInit runs architecture-specific setup for the goroutine startup.
-func (s *state) archInit(stack []uintptr, fn uintptr, args unsafe.Pointer) {
-	// Set up the stack canary, a random number that should be checked when
-	// switching from the task back to the scheduler. The stack canary pointer
-	// points to the first word of the stack. If it has changed between now and
-	// the next stack switch, there was a stack overflow.
-	s.canaryPtr = &stack[0]
-	*s.canaryPtr = stackCanary
-
+func (s *state) archInit(r *calleeSavedRegs, fn uintptr, args unsafe.Pointer) {
 	// Store the initial sp for the startTask function (implemented in assembly).
-	s.sp = uintptr(unsafe.Pointer(&stack[uintptr(len(stack))-(unsafe.Sizeof(calleeSavedRegs{})/unsafe.Sizeof(uintptr(0)))]))
+	s.sp = uintptr(unsafe.Pointer(r))
 
 	// Initialize the registers.
 	// These will be popped off of the stack on the first resume of the goroutine.
-	r := s.registers()
 
-	// Start the function at tinygo_startTask (defined in src/runtime/scheduler_cortexm.S).
+	// Start the function at tinygo_startTask (defined in src/internal/task/task_stack_cortexm.S).
 	// This assembly code calls a function (passed in r4) with a single argument (passed in r5).
 	// After the function returns, it calls Pause().
 	r.pc = uintptr(unsafe.Pointer(&startTask))
@@ -75,7 +59,8 @@ func (s *state) pause() {
 	switchToScheduler(&s.sp)
 }
 
-//export tinygo_pause
-func pause() {
-	Pause()
+// SystemStack returns the system stack pointer. On Cortex-M, it is always
+// available.
+func SystemStack() uintptr {
+	return arm.AsmFull("mrs {}, MSP", nil)
 }

src/internal/task/task_stack_esp32.S

@@ -0,0 +1,86 @@
+.section .text.tinygo_startTask,"ax",@progbits
+.global  tinygo_startTask
+.type    tinygo_startTask, %function
+tinygo_startTask:
+    // Small assembly stub for starting a goroutine. This already runs on the
+    // new stack, control reaches this function after returning from the initial
+    // tinygo_swapTask below (the retw.n instruction).
+    //
+    // The stack was set up in such a way that it looks as if this function was
+    // paused using tinygo_swapTask by setting up the parent register window and
+    // return pointer as a call4 instruction - except such a call never took
+    // place. Instead, the stack pointer is switched to the new stack after all
+    // live-but-invisible registers have been flushed to the stack. This means
+    // that all registers as present in tinygo_swapTask are moved four up (a2 in
+    // tinygo_swapTask is a6 in this function). We don't use any of those
+    // registers however. Instead, the retw.n instruction will load them through
+    // an underflow exception from the stack which means we get a0-a3 as defined
+    // in task_stack_esp32.go.
+
+    // Branch to the "goroutine start" function. The first (and only) parameter
+    // is stored in a2, but has to be moved to a6 to make it appear as a2 in the
+    // goroutine start function (due to changing the register window by four
+    // with callx4).
+    mov.n a6, a2
+    callx4 a3
+
+    // After return, exit this goroutine. This call never returns.
+    call4  tinygo_pause
+
+.section .text.tinygo_swapTask,"ax",@progbits
+.global tinygo_swapTask
+.type tinygo_swapTask, %function
+tinygo_swapTask:
+    // This function gets the following parameters:
+    // a2 = newStack uintptr
+    // a3 = oldStack *uintptr
+
+    // Reserve 32 bytes on the stack. It really needs to be 32 bytes, with 16
+    // extra at the bottom to adhere to the ABI.
+    entry sp, 32
+
+    // Disable interrupts while flushing registers. This is necessary because
+    // interrupts might want to use the stack pointer (at a2) which will be some
+    // arbitrary register while registers are flushed.
+    rsil a4, 3 // XCHAL_EXCM_LEVEL
+
+    // Flush all unsaved registers to the stack.
+    // This trick has been borrowed from the Zephyr project:
+    // https://github.com/zephyrproject-rtos/zephyr/blob/d79b003758/arch/xtensa/include/xtensa-asm2-s.h#L17
+    and a12, a12, a12
+    rotw 3
+    and a12, a12, a12
+    rotw 3
+    and a12, a12, a12
+    rotw 3
+    and a12, a12, a12
+    rotw 3
+    and a12, a12, a12
+    rotw 4
+
+    // Restore interrupts.
+    wsr.ps a4
+
+    // At this point, the following is true:
+    //     WindowStart == 1 << WindowBase
+    // Therefore, we don't need to do this manually.
+    // It also means that the stack pointer can now be safely modified.
+
+    // Save a0, which stores the return address and the parent register window
+    // in the upper two bits.
+    s32i.n a0, sp, 0
+
+    // Save the current stack pointer in oldStack.
+    s32i.n  sp, a3, 0
+
+    // Switch to the new stack pointer (newStack).
+    mov.n   sp, a2
+
+    // Load a0, which is the previous return addres from before the previous
+    // switch or the constructed return address to tinygo_startTask. This
+    // register also stores the parent register window.
+    l32i.n a0, sp, 0
+
+    // Return into the new stack. This instruction will trigger a window
+    // underflow, reloading the saved registers from the stack.
+    retw.n

src/internal/task/task_stack_esp32.go

@@ -0,0 +1,76 @@
+// +build scheduler.tasks,esp32
+
+package task
+
+// The windowed ABI (used on the ESP32) is as follows:
+//   a0:    return address (link register)
+//   a1:    stack pointer (must be 16-byte aligned)
+//   a2-a7: incoming arguments
+//   a7:    stack frame pointer (optional, normally unused in TinyGo)
+// Sources:
+//   http://cholla.mmto.org/esp8266/xtensa.html
+//   https://0x04.net/~mwk/doc/xtensa.pdf
+
+import (
+	"unsafe"
+)
+
+var systemStack uintptr
+
+// calleeSavedRegs is the list of registers that must be saved and restored when
+// switching between tasks. Also see task_stack_esp8266.S that relies on the
+// exact layout of this struct.
+type calleeSavedRegs struct {
+	// Registers in the register window of tinygo_startTask.
+	a0 uintptr
+	a1 uintptr
+	a2 uintptr
+	a3 uintptr
+
+	// Locals that can be used by tinygo_swapTask.
+	// The first field is the a0 loaded in tinygo_swapTask, the rest is unused.
+	locals [4]uintptr
+}
+
+// archInit runs architecture-specific setup for the goroutine startup.
+func (s *state) archInit(r *calleeSavedRegs, fn uintptr, args unsafe.Pointer) {
+	// Store the stack pointer for the tinygo_swapTask function (implemented in
+	// assembly). It needs to point to the locals field instead of a0 so that
+	// the retw.n at the end of tinygo_swapTask will return into
+	// tinygo_startTask with a0-a3 loaded (using the register window mechanism).
+	s.sp = uintptr(unsafe.Pointer(&r.locals[0]))
+
+	// Start the goroutine at tinygo_startTask (defined in
+	// src/internal/task/task_stack_esp32.S). The topmost two bits are not part
+	// of the address but instead store the register window of the caller.
+	// In this case there is no caller, instead we set up the return address as
+	// if tinygo_startTask called tinygo_swapTask with a call4 instruction.
+	r.locals[0] = uintptr(unsafe.Pointer(&startTask))&^(3<<30) | (1 << 30)
+
+	// Set up the stack pointer inside tinygo_startTask.
+	// Unlike most calling conventions, the windowed ABI actually saves the
+	// stack pointer on the stack to make register windowing work.
+	r.a1 = uintptr(unsafe.Pointer(r)) + 32
+
+	// Store the function pointer and the (only) parameter on the stack in a
+	// location that will be reloaded into registers when doing the
+	// pseudo-return to tinygo_startTask using the register window mechanism.
+	r.a3 = fn
+	r.a2 = uintptr(args)
+}
+
+func (s *state) resume() {
+	swapTask(s.sp, &systemStack)
+}
+
+func (s *state) pause() {
+	newStack := systemStack
+	systemStack = 0
+	swapTask(newStack, &s.sp)
+}
+
+// SystemStack returns the system stack pointer when called from a task stack.
+// When called from the system stack, it returns 0.
+func SystemStack() uintptr {
+	return systemStack
+}

src/internal/task/task_stack_esp8266.S

@@ -0,0 +1,54 @@
+.section .text.tinygo_startTask,"ax",@progbits
+.global  tinygo_startTask
+.type    tinygo_startTask, %function
+tinygo_startTask:
+    // Small assembly stub for starting a goroutine. This is already run on the
+    // new stack, with the callee-saved registers already loaded.
+    // Most importantly, r4 contains the pc of the to-be-started function and r5
+    // contains the only argument it is given. Multiple arguments are packed
+    // into one by storing them in a new allocation.
+
+    // Set the first argument of the goroutine start wrapper, which contains all
+    // the arguments.
+    mov.n  a2, a13
+
+    // Branch to the "goroutine start" function.
+    callx0 a12
+
+    // After return, exit this goroutine. This is a tail call.
+    call0  tinygo_pause
+.size tinygo_startTask, .-tinygo_startTask
+
+.global tinygo_swapTask
+.type tinygo_swapTask, %function
+tinygo_swapTask:
+    // This function gets the following parameters:
+    //   a2 = newStack uintptr
+    //   a3 = oldStack *uintptr
+    // Note:
+    //   a0 is the return address
+    //   a1 is the stack pointer (sp)
+
+    // Save all callee-saved registers:
+    addi    sp,  sp, -20
+    s32i.n  a12, sp, 0
+    s32i.n  a13, sp, 4
+    s32i.n  a14, sp, 8
+    s32i.n  a15, sp, 12
+    s32i.n  a0,  sp, 16
+
+    // Save the current stack pointer in oldStack.
+    s32i.n  sp, a3, 0
+
+    // Switch to the new stack pointer.
+    mov.n   sp, a2
+
+    // Load state from new task and branch to the previous position in the
+    // program.
+    l32i.n  a12, sp, 0
+    l32i.n  a13, sp, 4
+    l32i.n  a14, sp, 8
+    l32i.n  a15, sp, 12
+    l32i.n  a0,  sp, 16
+    addi    sp,  sp, 20
+    ret.n

src/internal/task/task_stack_esp8266.go

@@ -0,0 +1,70 @@
+// +build scheduler.tasks,esp8266
+
+package task
+
+// Stack switch implementation for the ESP8266, which does not use the windowed
+// ABI of Xtensa. Registers are assigned as follows:
+//   a0:      return address (link register)
+//   a1:      stack pointer (must be 16-byte aligned)
+//   a2-a7:   incoming arguments
+//   a8:      static chain (unused)
+//   a12-a15: callee-saved
+//   a15:     stack frame pointer (optional, unused)
+// Sources:
+// http://cholla.mmto.org/esp8266/xtensa.html
+// https://0x04.net/~mwk/doc/xtensa.pdf
+
+import "unsafe"
+
+var systemStack uintptr
+
+// calleeSavedRegs is the list of registers that must be saved and restored when
+// switching between tasks. Also see task_stack_esp8266.S that relies on the
+// exact layout of this struct.
+type calleeSavedRegs struct {
+	a12 uintptr
+	a13 uintptr
+	a14 uintptr
+	a15 uintptr
+
+	pc uintptr // also link register or r0
+}
+
+// archInit runs architecture-specific setup for the goroutine startup.
+func (s *state) archInit(r *calleeSavedRegs, fn uintptr, args unsafe.Pointer) {
+	// Store the initial sp for the startTask function (implemented in assembly).
+	s.sp = uintptr(unsafe.Pointer(r))
+
+	// Initialize the registers.
+	// These will be popped off of the stack on the first resume of the goroutine.
+
+	// Start the function at tinygo_startTask (defined in
+	// src/internal/task/task_stack_esp8266.S).
+	// This assembly code calls a function (passed in a12) with a single argument
+	// (passed in a13). After the function returns, it calls Pause().
+	r.pc = uintptr(unsafe.Pointer(&startTask))
+
+	// Pass the function to call in a12.
+	// This function is a compiler-generated wrapper which loads arguments out of a struct pointer.
+	// See createGoroutineStartWrapper (defined in compiler/goroutine.go) for more information.
+	r.a12 = fn
+
+	// Pass the pointer to the arguments struct in a13.
+	r.a13 = uintptr(args)
+}
+
+func (s *state) resume() {
+	swapTask(s.sp, &systemStack)
+}
+
+func (s *state) pause() {
+	newStack := systemStack
+	systemStack = 0
+	swapTask(newStack, &s.sp)
+}
+
+// SystemStack returns the system stack pointer when called from a task stack.
+// When called from the system stack, it returns 0.
+func SystemStack() uintptr {
+	return systemStack
+}

src/machine/board_arduino_mkr1000.go

@@ -0,0 +1,91 @@
+// +build arduino_mkr1000
+
+// This contains the pin mappings for the Arduino MKR1000 board.
+//
+// For more information, see: https://store.arduino.cc/usa/arduino-mkr1000-with-headers-mounted
+//
+package machine
+
+// used to reset into bootloader
+const RESET_MAGIC_VALUE = 0x07738135
+
+// GPIO Pins
+const (
+	RX0 Pin = PB23 // UART2 RX
+	TX1 Pin = PB22 // UART2 TX
+
+	D0 Pin = PA22 // PWM available
+	D1 Pin = PA23 // PWM available
+	D2 Pin = PA10 // PWM available
+	D3 Pin = PA11 // PWM available
+	D4 Pin = PB10 // PWM available
+	D5 Pin = PB11 // PWM available
+
+	D6  Pin = PA20 // PWM available
+	D7  Pin = PA21 // PWM available
+	D8  Pin = PA16 // PWM available
+	D9  Pin = PA17
+	D10 Pin = PA19 // PWM available
+	D11 Pin = PA08 // SDA
+	D12 Pin = PA09 // PWM available, SCL
+	D13 Pin = PB23 // RX
+	D14 Pin = PB22 // TX
+)
+
+// Analog pins
+const (
+	A0 Pin = PA02 // ADC0/AIN[0]
+	A1 Pin = PB02 // AIN[10]
+	A2 Pin = PB03 // AIN[11]
+	A3 Pin = PA04 // AIN[04]
+	A4 Pin = PA05 // AIN[05]
+	A5 Pin = PA06 // AIN[06]
+	A6 Pin = PA07 // AIN[07]
+)
+
+const (
+	LED = D6
+)
+
+// UART0 aka USBCDC pins
+const (
+	USBCDC_DM_PIN Pin = PA24
+	USBCDC_DP_PIN Pin = PA25
+)
+
+// UART1 pins
+const (
+	UART_TX_PIN Pin = PB22
+	UART_RX_PIN Pin = PB23
+)
+
+// I2C pins
+const (
+	SDA_PIN Pin = D11 // SDA
+	SCL_PIN Pin = D12 // SCL
+)
+
+// SPI pins
+const (
+	SPI0_SCK_PIN Pin = D9  // SCK: S1
+	SPI0_SDO_PIN Pin = D8  // SDO: S1
+	SPI0_SDI_PIN Pin = D10 // SDI: S1
+)
+
+// I2S pins
+const (
+	I2S_SCK_PIN Pin = PA10
+	I2S_SD_PIN  Pin = PA07
+	I2S_WS_PIN      = NoPin // TODO: figure out what this is on Arduino Nano 33.
+)
+
+// USB CDC identifiers
+const (
+	usb_STRING_PRODUCT      = "Arduino MKR1000"
+	usb_STRING_MANUFACTURER = "Arduino"
+)
+
+var (
+	usb_VID uint16 = 0x2341
+	usb_PID uint16 = 0x804e
+)

src/machine/board_matrixportal-m4.go

@@ -0,0 +1,178 @@
+// +build matrixportal_m4
+
+package machine
+
+// used to reset into bootloader
+const RESET_MAGIC_VALUE = 0xF01669EF
+
+// Digital pins
+const (
+	//    Pin   // Function          SERCOM  PWM  Interrupt
+	//    ----  // ----------------  ------  ---  ---------
+	D0  = PA01  // UART RX            1[1]   PWM  EXTI1
+	D1  = PA00  // UART TX            1[0]   PWM  EXTI0
+	D2  = PB22  // Button "Up"                    EXTI6
+	D3  = PB23  // Button "Down"                  EXTI7
+	D4  = PA23  // NeoPixel                       EXTI7
+	D5  = PB31  // I2C SDA            5[1]        EXTI15
+	D6  = PB30  // I2C SCL            5[0]        EXTI14
+	D7  = PB00  // HUB75 R1                       EXTI0
+	D8  = PB01  // HUB75 G1                       EXTI1
+	D9  = PB02  // HUB75 B1                       EXTI2
+	D10 = PB03  // HUB75 R2                       EXTI3
+	D11 = PB04  // HUB75 G2                       EXTI4
+	D12 = PB05  // HUB75 B2                       EXTI5
+	D13 = PA14  // LED                       PWM  EXTI14
+	D14 = PB06  // HUB75 CLK                      EXTI6
+	D15 = PB14  // HUB75 LAT                      EXTI14
+	D16 = PB12  // HUB75 OE                       EXTI12
+	D17 = PB07  // HUB75 ADDR A                   EXTI7
+	D18 = PB08  // HUB75 ADDR B                   EXTI8
+	D19 = PB09  // HUB75 ADDR C                   EXTI9
+	D20 = PB15  // HUB75 ADDR D                   EXTI15
+	D21 = PB13  // HUB75 ADDR E                   EXTI13
+	D22 = PA02  // ADC (A0)                       EXTI2
+	D23 = PA05  // ADC (A1)                       EXTI5
+	D24 = PA04  // ADC (A2)                  PWM  EXTI4
+	D25 = PA06  // ADC (A3)                  PWM  EXTI6
+	D26 = PA07  // ADC (A4)                       EXTI7
+	D27 = PA12  // ESP32 UART RX      4[1]   PWM  EXTI12
+	D28 = PA13  // ESP32 UART TX      4[0]   PWM  EXTI13
+	D29 = PA20  // ESP32 GPIO0               PWM  EXTI4
+	D30 = PA21  // ESP32 Reset               PWM  EXTI5
+	D31 = PA22  // ESP32 Busy                PWM  EXTI6
+	D32 = PA18  // ESP32 RTS                 PWM  EXTI2
+	D33 = PB17  // ESP32 SPI CS              PWM  EXTI1
+	D34 = PA16  // ESP32 SPI SCK      3[1]   PWM  EXTI0
+	D35 = PA17  // ESP32 SPI SDI      3[0]   PWM  EXTI1
+	D36 = PA19  // ESP32 SPI SDO      1[3]   PWM  EXTI3
+	D37 = NoPin // USB Host enable
+	D38 = PA24  // USB DM
+	D39 = PA27  // USB DP
+	D40 = PA03  // DAC/VREFP
+	D41 = PB10  // Flash QSPI SCK
+	D42 = PB11  // Flash QSPI CS
+	D43 = PA08  // Flash QSPI I00
+	D44 = PA09  // Flash QSPI IO1
+	D45 = PA10  // Flash QSPI IO2
+	D46 = PA11  // Flash QSPI IO3
+	D47 = PA27  // LIS3DH IRQ                     EXTI11
+	D48 = PA05  // SPI SCK            0[1]        EXTI5
+	D49 = PA04  // SPI SDO            0[0]   PWM  EXTI4
+	D50 = PA07  // SPI SDI            0[3]        EXTI7
+)
+
+// Analog pins
+const (
+	A0 = PA02 // ADC Channel 0
+	A1 = PA05 // ADC Channel 5
+	A2 = PA04 // ADC Channel 4
+	A3 = PA06 // ADC Channel 6
+	A4 = PA07 // ADC Channel 7
+)
+
+// LED pins
+const (
+	LED      = D13
+	NEOPIXEL = D4
+)
+
+// Button pins
+const (
+	BUTTON_UP   = D2
+	BUTTON_DOWN = D3
+)
+
+// UART pins
+const (
+	UART1_RX_PIN = D0 // SERCOM1[1]
+	UART1_TX_PIN = D1 // SERCOM1[0]
+
+	UART2_RX_PIN = D27 // SERCOM4[1] (ESP32 RX)
+	UART2_TX_PIN = D28 // SERCOM4[0] (ESP32 TX)
+
+	UART_RX_PIN = UART1_RX_PIN
+	UART_TX_PIN = UART1_TX_PIN
+)
+
+// SPI pins
+const (
+	SPI0_SCK_PIN = D34 // SERCOM3[1] (ESP32 SCK)
+	SPI0_SDO_PIN = D36 // SERCOM1[3] (ESP32 SDO)
+	SPI0_SDI_PIN = D35 // SERCOM3[0] (ESP32 SDI)
+
+	SPI1_SCK_PIN = D48 // SERCOM0[1]
+	SPI1_SDO_PIN = D49 // SERCOM0[0]
+	SPI1_SDI_PIN = D50 // SERCOM0[3]
+
+	SPI_SCK_PIN = SPI0_SCK_PIN
+	SPI_SDO_PIN = SPI0_SDO_PIN
+	SPI_SDI_PIN = SPI0_SDI_PIN
+)
+
+// I2C pins
+const (
+	I2C0_SDA_PIN = D5 // SERCOM5[1]
+	I2C0_SCL_PIN = D6 // SERCOM5[0]
+
+	I2C_SDA_PIN = I2C0_SDA_PIN
+	I2C_SCL_PIN = I2C0_SCL_PIN
+
+	SDA_PIN = I2C_SDA_PIN // awkward naming required by machine_atsamd51.go
+	SCL_PIN = I2C_SCL_PIN //
+)
+
+// ESP32 pins
+const (
+	NINA_ACK    = D31
+	NINA_GPIO0  = D29
+	NINA_RESETN = D30
+
+	NINA_RX  = UART2_RX_PIN
+	NINA_TX  = UART2_TX_PIN
+	NINA_RTS = D32
+
+	NINA_CS  = D33
+	NINA_SDO = SPI0_SDO_PIN
+	NINA_SDI = SPI0_SDI_PIN
+	NINA_SCK = SPI0_SCK_PIN
+)
+
+// HUB75 pins
+const (
+	HUB75_R1 = D7
+	HUB75_G1 = D8
+	HUB75_B1 = D9
+	HUB75_R2 = D10
+	HUB75_G2 = D11
+	HUB75_B2 = D12
+
+	HUB75_CLK    = D14
+	HUB75_LAT    = D15
+	HUB75_OE     = D16
+	HUB75_ADDR_A = D17
+	HUB75_ADDR_B = D18
+	HUB75_ADDR_C = D19
+	HUB75_ADDR_D = D20
+	HUB75_ADDR_E = D21
+)
+
+// USB CDC pins (UART0)
+const (
+	USBCDC_DM_PIN = D38
+	USBCDC_DP_PIN = D39
+
+	UART0_RX_PIN = USBCDC_DM_PIN
+	UART0_TX_PIN = USBCDC_DP_PIN
+)
+
+// USB CDC identifiers
+const (
+	usb_STRING_PRODUCT      = "Matrix Portal M4"
+	usb_STRING_MANUFACTURER = "Adafruit Industries"
+)
+
+var (
+	usb_VID uint16 = 0x239A
+	usb_PID uint16 = 0x80C9
+)

src/machine/board_matrixportal-m4_baremetal.go

@@ -0,0 +1,50 @@
+// +build sam,atsamd51,matrixportal_m4
+
+package machine
+
+import (
+	"device/sam"
+	"runtime/interrupt"
+)
+
+// UART on the MatrixPortal M4
+var (
+	UART1 = UART{
+		Buffer: NewRingBuffer(),
+		Bus:    sam.SERCOM1_USART_INT,
+		SERCOM: 1,
+	}
+
+	UART2 = UART{
+		Buffer: NewRingBuffer(),
+		Bus:    sam.SERCOM4_USART_INT,
+		SERCOM: 4,
+	}
+)
+
+func init() {
+	UART1.Interrupt = interrupt.New(sam.IRQ_SERCOM1_1, UART1.handleInterrupt)
+	UART2.Interrupt = interrupt.New(sam.IRQ_SERCOM4_1, UART2.handleInterrupt)
+}
+
+// I2C on the MatrixPortal M4
+var (
+	I2C0 = I2C{
+		Bus:    sam.SERCOM5_I2CM,
+		SERCOM: 5,
+	}
+)
+
+// SPI on the MatrixPortal M4
+var (
+	SPI0 = SPI{
+		Bus:    sam.SERCOM3_SPIM,
+		SERCOM: 3, // BUG: SDO on SERCOM1!
+	}
+	NINA_SPI = SPI0
+
+	SPI1 = SPI{
+		Bus:    sam.SERCOM0_SPIM,
+		SERCOM: 0,
+	}
+)

src/machine/board_nucleof722ze.go

@@ -0,0 +1,58 @@
+// +build nucleof722ze
+
+package machine
+
+import (
+	"device/stm32"
+	"runtime/interrupt"
+)
+
+const (
+	LED         = LED_BUILTIN
+	LED_BUILTIN = LED_GREEN
+	LED_GREEN   = PB0
+	LED_BLUE    = PB7
+	LED_RED     = PB14
+)
+
+const (
+	BUTTON      = BUTTON_USER
+	BUTTON_USER = PC13
+)
+
+// UART pins
+const (
+	// PD8 and PD9 are connected to the ST-Link Virtual Com Port (VCP)
+	UART_TX_PIN = PD8
+	UART_RX_PIN = PD9
+	UART_ALT_FN = 7 // GPIO_AF7_UART3
+)
+
+var (
+	// USART3 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.USART3,
+		AltFuncSelector: UART_ALT_FN,
+	}
+	UART1 = &UART0
+)
+
+func init() {
+	UART0.Interrupt = interrupt.New(stm32.IRQ_USART3, UART0.handleInterrupt)
+}
+
+// SPI pins
+const (
+	SPI0_SCK_PIN = PA5
+	SPI0_SDI_PIN = PA6
+	SPI0_SDO_PIN = PA7
+)
+
+// I2C pins
+const (
+	SCL_PIN = PB6
+	SDA_PIN = PB7
+)

src/machine/i2c.go

@@ -1,4 +1,4 @@
-// +build avr nrf sam stm32,!stm32f407 fe310 k210
+// +build avr nrf sam stm32,!stm32f407,!stm32f7x2 fe310 k210
 
 package machine
 

src/machine/machine_atsamd21.go

@@ -136,6 +136,11 @@ var pinPadMapping = [32]byte{
 // found" (indicated by returning ok=false). The pad number is returned to
 // calculate the DOPO/DIPO bitfields of the various serial peripherals.
 func findPinPadMapping(sercom uint8, pin Pin) (pinMode PinMode, pad uint32, ok bool) {
+	if int(pin)/2 >= len(pinPadMapping) {
+		// This is probably NoPin, for which no mapping is available.
+		return
+	}
+
 	nibbles := pinPadMapping[pin/2]
 	upper := nibbles >> 4
 	lower := nibbles & 0xf

src/machine/machine_atsamd51.go

@@ -322,6 +322,11 @@ var pinPadMapping = [64]uint16{
 // (indicated by returning ok=false). The pad number is returned to calculate
 // the DOPO/DIPO bitfields of the various serial peripherals.
 func findPinPadMapping(sercom uint8, pin Pin) (pinMode PinMode, pad uint32, ok bool) {
+	if int(pin)/2 >= len(pinPadMapping) {
+		// This is probably NoPin, for which no mapping is available.
+		return
+	}
+
 	bytes := pinPadMapping[pin/2]
 	upper := byte(bytes >> 8)
 	lower := byte(bytes & 0xff)

src/machine/machine_nrf.go

@@ -328,163 +328,3 @@ func (i2c I2C) readByte() (byte, error) {
 	i2c.Bus.EVENTS_RXDREADY.Set(0)
 	return byte(i2c.Bus.RXD.Get()), nil
 }
-
-// SPI on the NRF.
-type SPI struct {
-	Bus *nrf.SPI_Type
-}
-
-// There are 2 SPI interfaces on the NRF5x.
-var (
-	SPI0 = SPI{Bus: nrf.SPI0}
-	SPI1 = SPI{Bus: nrf.SPI1}
-)
-
-// SPIConfig is used to store config info for SPI.
-type SPIConfig struct {
-	Frequency uint32
-	SCK       Pin
-	SDO       Pin
-	SDI       Pin
-	LSBFirst  bool
-	Mode      uint8
-}
-
-// Configure is intended to setup the SPI interface.
-func (spi SPI) Configure(config SPIConfig) {
-	// Disable bus to configure it
-	spi.Bus.ENABLE.Set(nrf.SPI_ENABLE_ENABLE_Disabled)
-
-	// set frequency
-	var freq uint32
-
-	if config.Frequency == 0 {
-		config.Frequency = 4000000 // 4MHz
-	}
-
-	switch {
-	case config.Frequency >= 8000000:
-		freq = nrf.SPI_FREQUENCY_FREQUENCY_M8
-	case config.Frequency >= 4000000:
-		freq = nrf.SPI_FREQUENCY_FREQUENCY_M4
-	case config.Frequency >= 2000000:
-		freq = nrf.SPI_FREQUENCY_FREQUENCY_M2
-	case config.Frequency >= 1000000:
-		freq = nrf.SPI_FREQUENCY_FREQUENCY_M1
-	case config.Frequency >= 500000:
-		freq = nrf.SPI_FREQUENCY_FREQUENCY_K500
-	case config.Frequency >= 250000:
-		freq = nrf.SPI_FREQUENCY_FREQUENCY_K250
-	default: // below 250kHz, default to the lowest speed available
-		freq = nrf.SPI_FREQUENCY_FREQUENCY_K125
-	}
-	spi.Bus.FREQUENCY.Set(freq)
-
-	var conf uint32
-
-	// set bit transfer order
-	if config.LSBFirst {
-		conf = (nrf.SPI_CONFIG_ORDER_LsbFirst << nrf.SPI_CONFIG_ORDER_Pos)
-	}
-
-	// set mode
-	switch config.Mode {
-	case 0:
-		conf &^= (nrf.SPI_CONFIG_CPOL_ActiveHigh << nrf.SPI_CONFIG_CPOL_Pos)
-		conf &^= (nrf.SPI_CONFIG_CPHA_Leading << nrf.SPI_CONFIG_CPHA_Pos)
-	case 1:
-		conf &^= (nrf.SPI_CONFIG_CPOL_ActiveHigh << nrf.SPI_CONFIG_CPOL_Pos)
-		conf |= (nrf.SPI_CONFIG_CPHA_Trailing << nrf.SPI_CONFIG_CPHA_Pos)
-	case 2:
-		conf |= (nrf.SPI_CONFIG_CPOL_ActiveLow << nrf.SPI_CONFIG_CPOL_Pos)
-		conf &^= (nrf.SPI_CONFIG_CPHA_Leading << nrf.SPI_CONFIG_CPHA_Pos)
-	case 3:
-		conf |= (nrf.SPI_CONFIG_CPOL_ActiveLow << nrf.SPI_CONFIG_CPOL_Pos)
-		conf |= (nrf.SPI_CONFIG_CPHA_Trailing << nrf.SPI_CONFIG_CPHA_Pos)
-	default: // to mode
-		conf &^= (nrf.SPI_CONFIG_CPOL_ActiveHigh << nrf.SPI_CONFIG_CPOL_Pos)
-		conf &^= (nrf.SPI_CONFIG_CPHA_Leading << nrf.SPI_CONFIG_CPHA_Pos)
-	}
-	spi.Bus.CONFIG.Set(conf)
-
-	// set pins
-	spi.setPins(config.SCK, config.SDO, config.SDI)
-
-	// Re-enable bus now that it is configured.
-	spi.Bus.ENABLE.Set(nrf.SPI_ENABLE_ENABLE_Enabled)
-}
-
-// Transfer writes/reads a single byte using the SPI interface.
-func (spi SPI) Transfer(w byte) (byte, error) {
-	spi.Bus.TXD.Set(uint32(w))
-	for spi.Bus.EVENTS_READY.Get() == 0 {
-	}
-	r := spi.Bus.RXD.Get()
-	spi.Bus.EVENTS_READY.Set(0)
-
-	// TODO: handle SPI errors
-	return byte(r), nil
-}
-
-// Tx handles read/write operation for SPI interface. Since SPI is a syncronous write/read
-// interface, there must always be the same number of bytes written as bytes read.
-// The Tx method knows about this, and offers a few different ways of calling it.
-//
-// 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 {
-	var err error
-
-	switch {
-	case len(w) == 0:
-		// read only, so write zero and read a result.
-		for i := range r {
-			r[i], err = spi.Transfer(0)
-			if err != nil {
-				return err
-			}
-		}
-	case len(r) == 0:
-		// write only
-		spi.Bus.TXD.Set(uint32(w[0]))
-		w = w[1:]
-		for _, b := range w {
-			spi.Bus.TXD.Set(uint32(b))
-			for spi.Bus.EVENTS_READY.Get() == 0 {
-			}
-			spi.Bus.EVENTS_READY.Set(0)
-			_ = spi.Bus.RXD.Get()
-		}
-		for spi.Bus.EVENTS_READY.Get() == 0 {
-		}
-		spi.Bus.EVENTS_READY.Set(0)
-		_ = spi.Bus.RXD.Get()
-
-	default:
-		// write/read
-		if len(w) != len(r) {
-			return ErrTxInvalidSliceSize
-		}
-
-		for i, b := range w {
-			r[i], err = spi.Transfer(b)
-			if err != nil {
-				return err
-			}
-		}
-	}
-
-	return nil
-}

src/machine/machine_nrf51.go

@@ -29,18 +29,169 @@ func (i2c I2C) setPins(scl, sda Pin) {
 	i2c.Bus.PSELSDA.Set(uint32(sda))
 }
 
-// SPI
-func (spi SPI) setPins(sck, sdo, sdi Pin) {
-	if sck == 0 {
-		sck = SPI0_SCK_PIN
-	}
-	if sdo == 0 {
-		sdo = SPI0_SDO_PIN
-	}
-	if sdi == 0 {
-		sdi = SPI0_SDI_PIN
-	}
-	spi.Bus.PSELSCK.Set(uint32(sck))
-	spi.Bus.PSELMOSI.Set(uint32(sdo))
-	spi.Bus.PSELMISO.Set(uint32(sdi))
+// SPI on the NRF.
+type SPI struct {
+	Bus *nrf.SPI_Type
+}
+
+// There are 2 SPI interfaces on the NRF51.
+var (
+	SPI0 = SPI{Bus: nrf.SPI0}
+	SPI1 = SPI{Bus: nrf.SPI1}
+)
+
+// SPIConfig is used to store config info for SPI.
+type SPIConfig struct {
+	Frequency uint32
+	SCK       Pin
+	SDO       Pin
+	SDI       Pin
+	LSBFirst  bool
+	Mode      uint8
+}
+
+// Configure is intended to setup the SPI interface.
+func (spi SPI) Configure(config SPIConfig) {
+	// Disable bus to configure it
+	spi.Bus.ENABLE.Set(nrf.SPI_ENABLE_ENABLE_Disabled)
+
+	// set frequency
+	var freq uint32
+
+	if config.Frequency == 0 {
+		config.Frequency = 4000000 // 4MHz
+	}
+
+	switch {
+	case config.Frequency >= 8000000:
+		freq = nrf.SPI_FREQUENCY_FREQUENCY_M8
+	case config.Frequency >= 4000000:
+		freq = nrf.SPI_FREQUENCY_FREQUENCY_M4
+	case config.Frequency >= 2000000:
+		freq = nrf.SPI_FREQUENCY_FREQUENCY_M2
+	case config.Frequency >= 1000000:
+		freq = nrf.SPI_FREQUENCY_FREQUENCY_M1
+	case config.Frequency >= 500000:
+		freq = nrf.SPI_FREQUENCY_FREQUENCY_K500
+	case config.Frequency >= 250000:
+		freq = nrf.SPI_FREQUENCY_FREQUENCY_K250
+	default: // below 250kHz, default to the lowest speed available
+		freq = nrf.SPI_FREQUENCY_FREQUENCY_K125
+	}
+	spi.Bus.FREQUENCY.Set(freq)
+
+	var conf uint32
+
+	// set bit transfer order
+	if config.LSBFirst {
+		conf = (nrf.SPI_CONFIG_ORDER_LsbFirst << nrf.SPI_CONFIG_ORDER_Pos)
+	}
+
+	// set mode
+	switch config.Mode {
+	case 0:
+		conf &^= (nrf.SPI_CONFIG_CPOL_ActiveHigh << nrf.SPI_CONFIG_CPOL_Pos)
+		conf &^= (nrf.SPI_CONFIG_CPHA_Leading << nrf.SPI_CONFIG_CPHA_Pos)
+	case 1:
+		conf &^= (nrf.SPI_CONFIG_CPOL_ActiveHigh << nrf.SPI_CONFIG_CPOL_Pos)
+		conf |= (nrf.SPI_CONFIG_CPHA_Trailing << nrf.SPI_CONFIG_CPHA_Pos)
+	case 2:
+		conf |= (nrf.SPI_CONFIG_CPOL_ActiveLow << nrf.SPI_CONFIG_CPOL_Pos)
+		conf &^= (nrf.SPI_CONFIG_CPHA_Leading << nrf.SPI_CONFIG_CPHA_Pos)
+	case 3:
+		conf |= (nrf.SPI_CONFIG_CPOL_ActiveLow << nrf.SPI_CONFIG_CPOL_Pos)
+		conf |= (nrf.SPI_CONFIG_CPHA_Trailing << nrf.SPI_CONFIG_CPHA_Pos)
+	default: // to mode
+		conf &^= (nrf.SPI_CONFIG_CPOL_ActiveHigh << nrf.SPI_CONFIG_CPOL_Pos)
+		conf &^= (nrf.SPI_CONFIG_CPHA_Leading << nrf.SPI_CONFIG_CPHA_Pos)
+	}
+	spi.Bus.CONFIG.Set(conf)
+
+	// set pins
+	if config.SCK == 0 && config.SDO == 0 && config.SDI == 0 {
+		config.SCK = SPI0_SCK_PIN
+		config.SDO = SPI0_SDO_PIN
+		config.SDI = SPI0_SDI_PIN
+	}
+	spi.Bus.PSELSCK.Set(uint32(config.SCK))
+	spi.Bus.PSELMOSI.Set(uint32(config.SDO))
+	spi.Bus.PSELMISO.Set(uint32(config.SDI))
+
+	// Re-enable bus now that it is configured.
+	spi.Bus.ENABLE.Set(nrf.SPI_ENABLE_ENABLE_Enabled)
+}
+
+// Transfer writes/reads a single byte using the SPI interface.
+func (spi SPI) Transfer(w byte) (byte, error) {
+	spi.Bus.TXD.Set(uint32(w))
+	for spi.Bus.EVENTS_READY.Get() == 0 {
+	}
+	r := spi.Bus.RXD.Get()
+	spi.Bus.EVENTS_READY.Set(0)
+
+	// TODO: handle SPI errors
+	return byte(r), nil
+}
+
+// Tx handles read/write operation for SPI interface. Since SPI is a syncronous write/read
+// interface, there must always be the same number of bytes written as bytes read.
+// The Tx method knows about this, and offers a few different ways of calling it.
+//
+// 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 {
+	var err error
+
+	switch {
+	case len(w) == 0:
+		// read only, so write zero and read a result.
+		for i := range r {
+			r[i], err = spi.Transfer(0)
+			if err != nil {
+				return err
+			}
+		}
+	case len(r) == 0:
+		// write only
+		spi.Bus.TXD.Set(uint32(w[0]))
+		w = w[1:]
+		for _, b := range w {
+			spi.Bus.TXD.Set(uint32(b))
+			for spi.Bus.EVENTS_READY.Get() == 0 {
+			}
+			spi.Bus.EVENTS_READY.Set(0)
+			_ = spi.Bus.RXD.Get()
+		}
+		for spi.Bus.EVENTS_READY.Get() == 0 {
+		}
+		spi.Bus.EVENTS_READY.Set(0)
+		_ = spi.Bus.RXD.Get()
+
+	default:
+		// write/read
+		if len(w) != len(r) {
+			return ErrTxInvalidSliceSize
+		}
+
+		for i, b := range w {
+			r[i], err = spi.Transfer(b)
+			if err != nil {
+				return err
+			}
+		}
+	}
+
+	return nil
 }

src/machine/machine_nrf52.go

@@ -4,17 +4,48 @@ package machine
 
 import (
 	"device/nrf"
-	"unsafe"
+)
+
+// Hardware pins
+const (
+	P0_00 Pin = 0
+	P0_01 Pin = 1
+	P0_02 Pin = 2
+	P0_03 Pin = 3
+	P0_04 Pin = 4
+	P0_05 Pin = 5
+	P0_06 Pin = 6
+	P0_07 Pin = 7
+	P0_08 Pin = 8
+	P0_09 Pin = 9
+	P0_10 Pin = 10
+	P0_11 Pin = 11
+	P0_12 Pin = 12
+	P0_13 Pin = 13
+	P0_14 Pin = 14
+	P0_15 Pin = 15
+	P0_16 Pin = 16
+	P0_17 Pin = 17
+	P0_18 Pin = 18
+	P0_19 Pin = 19
+	P0_20 Pin = 20
+	P0_21 Pin = 21
+	P0_22 Pin = 22
+	P0_23 Pin = 23
+	P0_24 Pin = 24
+	P0_25 Pin = 25
+	P0_26 Pin = 26
+	P0_27 Pin = 27
+	P0_28 Pin = 28
+	P0_29 Pin = 29
+	P0_30 Pin = 30
+	P0_31 Pin = 31
 )
 
 var (
 	UART0 = NRF_UART0
 )
 
-func CPUFrequency() uint32 {
-	return 64000000
-}
-
 // Get peripheral and pin number for this GPIO pin.
 func (p Pin) getPortPin() (*nrf.GPIO_Type, uint32) {
 	return nrf.P0, uint32(p)
@@ -30,165 +61,9 @@ func (i2c I2C) setPins(scl, sda Pin) {
 	i2c.Bus.PSELSDA.Set(uint32(sda))
 }
 
-// SPI
-func (spi SPI) setPins(sck, sdo, sdi Pin) {
-	if sck == 0 {
-		sck = SPI0_SCK_PIN
-	}
-	if sdo == 0 {
-		sdo = SPI0_SDO_PIN
-	}
-	if sdi == 0 {
-		sdi = SPI0_SDI_PIN
-	}
-	spi.Bus.PSEL.SCK.Set(uint32(sck))
-	spi.Bus.PSEL.MOSI.Set(uint32(sdo))
-	spi.Bus.PSEL.MISO.Set(uint32(sdi))
-}
-
-// InitADC initializes the registers needed for ADC.
-func InitADC() {
-	return // no specific setup on nrf52 machine.
-}
-
-// Configure configures an ADC pin to be able to read analog data.
-func (a ADC) Configure() {
-	return // no pin specific setup on nrf52 machine.
-}
-
-// Get returns the current value of a ADC pin in the range 0..0xffff.
-func (a ADC) Get() uint16 {
-	var pwmPin uint32
-	var value int16
-
-	switch a.Pin {
-	case 2:
-		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput0
-
-	case 3:
-		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput1
-
-	case 4:
-		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput2
-
-	case 5:
-		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput3
-
-	case 28:
-		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput4
-
-	case 29:
-		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput5
-
-	case 30:
-		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput6
-
-	case 31:
-		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput7
-
-	default:
-		return 0
-	}
-
-	nrf.SAADC.RESOLUTION.Set(nrf.SAADC_RESOLUTION_VAL_12bit)
-
-	// Enable ADC.
-	nrf.SAADC.ENABLE.Set(nrf.SAADC_ENABLE_ENABLE_Enabled << nrf.SAADC_ENABLE_ENABLE_Pos)
-	for i := 0; i < 8; i++ {
-		nrf.SAADC.CH[i].PSELN.Set(nrf.SAADC_CH_PSELP_PSELP_NC)
-		nrf.SAADC.CH[i].PSELP.Set(nrf.SAADC_CH_PSELP_PSELP_NC)
-	}
-
-	// Configure ADC.
-	nrf.SAADC.CH[0].CONFIG.Set(((nrf.SAADC_CH_CONFIG_RESP_Bypass << nrf.SAADC_CH_CONFIG_RESP_Pos) & nrf.SAADC_CH_CONFIG_RESP_Msk) |
-		((nrf.SAADC_CH_CONFIG_RESP_Bypass << nrf.SAADC_CH_CONFIG_RESN_Pos) & nrf.SAADC_CH_CONFIG_RESN_Msk) |
-		((nrf.SAADC_CH_CONFIG_GAIN_Gain1_5 << nrf.SAADC_CH_CONFIG_GAIN_Pos) & nrf.SAADC_CH_CONFIG_GAIN_Msk) |
-		((nrf.SAADC_CH_CONFIG_REFSEL_Internal << nrf.SAADC_CH_CONFIG_REFSEL_Pos) & nrf.SAADC_CH_CONFIG_REFSEL_Msk) |
-		((nrf.SAADC_CH_CONFIG_TACQ_3us << nrf.SAADC_CH_CONFIG_TACQ_Pos) & nrf.SAADC_CH_CONFIG_TACQ_Msk) |
-		((nrf.SAADC_CH_CONFIG_MODE_SE << nrf.SAADC_CH_CONFIG_MODE_Pos) & nrf.SAADC_CH_CONFIG_MODE_Msk))
-
-	// Set pin to read.
-	nrf.SAADC.CH[0].PSELN.Set(pwmPin)
-	nrf.SAADC.CH[0].PSELP.Set(pwmPin)
-
-	// Destination for sample result.
-	nrf.SAADC.RESULT.PTR.Set(uint32(uintptr(unsafe.Pointer(&value))))
-	nrf.SAADC.RESULT.MAXCNT.Set(1) // One sample
-
-	// Start tasks.
-	nrf.SAADC.TASKS_START.Set(1)
-	for nrf.SAADC.EVENTS_STARTED.Get() == 0 {
-	}
-	nrf.SAADC.EVENTS_STARTED.Set(0x00)
-
-	// Start the sample task.
-	nrf.SAADC.TASKS_SAMPLE.Set(1)
-
-	// Wait until the sample task is done.
-	for nrf.SAADC.EVENTS_END.Get() == 0 {
-	}
-	nrf.SAADC.EVENTS_END.Set(0x00)
-
-	// Stop the ADC
-	nrf.SAADC.TASKS_STOP.Set(1)
-	for nrf.SAADC.EVENTS_STOPPED.Get() == 0 {
-	}
-	nrf.SAADC.EVENTS_STOPPED.Set(0)
-
-	// Disable the ADC.
-	nrf.SAADC.ENABLE.Set(nrf.SAADC_ENABLE_ENABLE_Disabled << nrf.SAADC_ENABLE_ENABLE_Pos)
-
-	if value < 0 {
-		value = 0
-	}
-
-	// Return 16-bit result from 12-bit value.
-	return uint16(value << 4)
-}
-
 // PWM
 var (
 	pwmChannelPins     = [3]uint32{0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF}
 	pwms               = [3]*nrf.PWM_Type{nrf.PWM0, nrf.PWM1, nrf.PWM2}
 	pwmChannelSequence [3]uint16
 )
-
-// InitPWM initializes the registers needed for PWM.
-func InitPWM() {
-	return
-}
-
-// 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) {
-	for i := 0; i < 3; i++ {
-		if pwmChannelPins[i] == 0xFFFFFFFF || pwmChannelPins[i] == uint32(pwm.Pin) {
-			pwmChannelPins[i] = uint32(pwm.Pin)
-			pwmChannelSequence[i] = (value >> 2) | 0x8000 // set bit 15 to invert polarity
-
-			p := pwms[i]
-
-			p.PSEL.OUT[0].Set(uint32(pwm.Pin))
-			p.PSEL.OUT[1].Set(uint32(pwm.Pin))
-			p.PSEL.OUT[2].Set(uint32(pwm.Pin))
-			p.PSEL.OUT[3].Set(uint32(pwm.Pin))
-			p.ENABLE.Set(nrf.PWM_ENABLE_ENABLE_Enabled << nrf.PWM_ENABLE_ENABLE_Pos)
-			p.PRESCALER.Set(nrf.PWM_PRESCALER_PRESCALER_DIV_2)
-			p.MODE.Set(nrf.PWM_MODE_UPDOWN_Up)
-			p.COUNTERTOP.Set(16384) // frequency
-			p.LOOP.Set(0)
-			p.DECODER.Set((nrf.PWM_DECODER_LOAD_Common << nrf.PWM_DECODER_LOAD_Pos) | (nrf.PWM_DECODER_MODE_RefreshCount << nrf.PWM_DECODER_MODE_Pos))
-			p.SEQ[0].PTR.Set(uint32(uintptr(unsafe.Pointer(&pwmChannelSequence[i]))))
-			p.SEQ[0].CNT.Set(1)
-			p.SEQ[0].REFRESH.Set(1)
-			p.SEQ[0].ENDDELAY.Set(0)
-			p.TASKS_SEQSTART[0].Set(1)
-
-			break
-		}
-	}
-}

src/machine/machine_nrf52840.go

@@ -4,13 +4,8 @@ package machine
 
 import (
 	"device/nrf"
-	"unsafe"
 )
 
-func CPUFrequency() uint32 {
-	return 64000000
-}
-
 // Hardware pins
 const (
 	P0_00 Pin = 0
@@ -82,164 +77,9 @@ func (i2c I2C) setPins(scl, sda Pin) {
 	i2c.Bus.PSEL.SDA.Set(uint32(sda))
 }
 
-// SPI
-func (spi SPI) setPins(sck, sdo, sdi Pin) {
-	if sck == 0 {
-		sck = SPI0_SCK_PIN
-	}
-	if sdo == 0 {
-		sdo = SPI0_SDO_PIN
-	}
-	if sdi == 0 {
-		sdi = SPI0_SDI_PIN
-	}
-	spi.Bus.PSEL.SCK.Set(uint32(sck))
-	spi.Bus.PSEL.MOSI.Set(uint32(sdo))
-	spi.Bus.PSEL.MISO.Set(uint32(sdi))
-}
-
-// InitADC initializes the registers needed for ADC.
-func InitADC() {
-	return // no specific setup on nrf52840 machine.
-}
-
-// Configure configures an ADC pin to be able to read analog data.
-func (a ADC) Configure() error {
-	return nil // no pin specific setup on nrf52840 machine.
-}
-
-// Get returns the current value of a ADC pin in the range 0..0xffff.
-func (a ADC) Get() uint16 {
-	var pwmPin uint32
-	var value int16
-
-	switch a.Pin {
-	case 2:
-		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput0
-
-	case 3:
-		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput1
-
-	case 4:
-		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput2
-
-	case 5:
-		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput3
-
-	case 28:
-		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput4
-
-	case 29:
-		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput5
-
-	case 30:
-		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput6
-
-	case 31:
-		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput7
-
-	default:
-		return 0
-	}
-
-	nrf.SAADC.RESOLUTION.Set(nrf.SAADC_RESOLUTION_VAL_12bit)
-
-	// Enable ADC.
-	nrf.SAADC.ENABLE.Set(nrf.SAADC_ENABLE_ENABLE_Enabled << nrf.SAADC_ENABLE_ENABLE_Pos)
-	for i := 0; i < 8; i++ {
-		nrf.SAADC.CH[i].PSELN.Set(nrf.SAADC_CH_PSELP_PSELP_NC)
-		nrf.SAADC.CH[i].PSELP.Set(nrf.SAADC_CH_PSELP_PSELP_NC)
-	}
-
-	// Configure ADC.
-	nrf.SAADC.CH[0].CONFIG.Set(((nrf.SAADC_CH_CONFIG_RESP_Bypass << nrf.SAADC_CH_CONFIG_RESP_Pos) & nrf.SAADC_CH_CONFIG_RESP_Msk) |
-		((nrf.SAADC_CH_CONFIG_RESP_Bypass << nrf.SAADC_CH_CONFIG_RESN_Pos) & nrf.SAADC_CH_CONFIG_RESN_Msk) |
-		((nrf.SAADC_CH_CONFIG_GAIN_Gain1_5 << nrf.SAADC_CH_CONFIG_GAIN_Pos) & nrf.SAADC_CH_CONFIG_GAIN_Msk) |
-		((nrf.SAADC_CH_CONFIG_REFSEL_Internal << nrf.SAADC_CH_CONFIG_REFSEL_Pos) & nrf.SAADC_CH_CONFIG_REFSEL_Msk) |
-		((nrf.SAADC_CH_CONFIG_TACQ_3us << nrf.SAADC_CH_CONFIG_TACQ_Pos) & nrf.SAADC_CH_CONFIG_TACQ_Msk) |
-		((nrf.SAADC_CH_CONFIG_MODE_SE << nrf.SAADC_CH_CONFIG_MODE_Pos) & nrf.SAADC_CH_CONFIG_MODE_Msk))
-
-	// Set pin to read.
-	nrf.SAADC.CH[0].PSELN.Set(pwmPin)
-	nrf.SAADC.CH[0].PSELP.Set(pwmPin)
-
-	// Destination for sample result.
-	nrf.SAADC.RESULT.PTR.Set(uint32(uintptr(unsafe.Pointer(&value))))
-	nrf.SAADC.RESULT.MAXCNT.Set(1) // One sample
-
-	// Start tasks.
-	nrf.SAADC.TASKS_START.Set(1)
-	for nrf.SAADC.EVENTS_STARTED.Get() == 0 {
-	}
-	nrf.SAADC.EVENTS_STARTED.Set(0x00)
-
-	// Start the sample task.
-	nrf.SAADC.TASKS_SAMPLE.Set(1)
-
-	// Wait until the sample task is done.
-	for nrf.SAADC.EVENTS_END.Get() == 0 {
-	}
-	nrf.SAADC.EVENTS_END.Set(0x00)
-
-	// Stop the ADC
-	nrf.SAADC.TASKS_STOP.Set(1)
-	for nrf.SAADC.EVENTS_STOPPED.Get() == 0 {
-	}
-	nrf.SAADC.EVENTS_STOPPED.Set(0)
-
-	// Disable the ADC.
-	nrf.SAADC.ENABLE.Set(nrf.SAADC_ENABLE_ENABLE_Disabled << nrf.SAADC_ENABLE_ENABLE_Pos)
-
-	if value < 0 {
-		value = 0
-	}
-
-	// Return 16-bit result from 12-bit value.
-	return uint16(value << 4)
-}
-
 // PWM
 var (
 	pwmChannelPins     = [4]uint32{0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF}
 	pwms               = [4]*nrf.PWM_Type{nrf.PWM0, nrf.PWM1, nrf.PWM2, nrf.PWM3}
 	pwmChannelSequence [4]uint16
 )
-
-// InitPWM initializes the registers needed for PWM.
-func InitPWM() {
-	return
-}
-
-// Configure configures a PWM pin for output.
-func (pwm PWM) Configure() {
-}
-
-// Set turns on the duty cycle for a PWM pin using the provided value.
-func (pwm PWM) Set(value uint16) {
-	for i := 0; i < 4; i++ {
-		if pwmChannelPins[i] == 0xFFFFFFFF || pwmChannelPins[i] == uint32(pwm.Pin) {
-			pwmChannelPins[i] = uint32(pwm.Pin)
-			pwmChannelSequence[i] = (value >> 2) | 0x8000 // set bit 15 to invert polarity
-
-			p := pwms[i]
-
-			p.PSEL.OUT[0].Set(uint32(pwm.Pin))
-			p.PSEL.OUT[1].Set(uint32(pwm.Pin))
-			p.PSEL.OUT[2].Set(uint32(pwm.Pin))
-			p.PSEL.OUT[3].Set(uint32(pwm.Pin))
-			p.ENABLE.Set(nrf.PWM_ENABLE_ENABLE_Enabled << nrf.PWM_ENABLE_ENABLE_Pos)
-			p.PRESCALER.Set(nrf.PWM_PRESCALER_PRESCALER_DIV_2)
-			p.MODE.Set(nrf.PWM_MODE_UPDOWN_Up)
-			p.COUNTERTOP.Set(16384) // frequency
-			p.LOOP.Set(0)
-			p.DECODER.Set((nrf.PWM_DECODER_LOAD_Common << nrf.PWM_DECODER_LOAD_Pos) | (nrf.PWM_DECODER_MODE_RefreshCount << nrf.PWM_DECODER_MODE_Pos))
-			p.SEQ[0].PTR.Set(uint32(uintptr(unsafe.Pointer(&pwmChannelSequence[i]))))
-			p.SEQ[0].CNT.Set(1)
-			p.SEQ[0].REFRESH.Set(1)
-			p.SEQ[0].ENDDELAY.Set(0)
-			p.TASKS_SEQSTART[0].Set(1)
-
-			break
-		}
-	}
-}

src/machine/machine_nrf528xx.go

@@ -0,0 +1,296 @@
+// +build nrf52 nrf52840
+
+package machine
+
+import (
+	"device/nrf"
+	"unsafe"
+)
+
+func CPUFrequency() uint32 {
+	return 64000000
+}
+
+// InitADC initializes the registers needed for ADC.
+func InitADC() {
+	return // no specific setup on nrf52 machine.
+}
+
+// Configure configures an ADC pin to be able to read analog data.
+func (a ADC) Configure() {
+	return // no pin specific setup on nrf52 machine.
+}
+
+// Get returns the current value of a ADC pin in the range 0..0xffff.
+func (a ADC) Get() uint16 {
+	var pwmPin uint32
+	var value int16
+
+	switch a.Pin {
+	case 2:
+		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput0
+
+	case 3:
+		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput1
+
+	case 4:
+		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput2
+
+	case 5:
+		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput3
+
+	case 28:
+		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput4
+
+	case 29:
+		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput5
+
+	case 30:
+		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput6
+
+	case 31:
+		pwmPin = nrf.SAADC_CH_PSELP_PSELP_AnalogInput7
+
+	default:
+		return 0
+	}
+
+	nrf.SAADC.RESOLUTION.Set(nrf.SAADC_RESOLUTION_VAL_12bit)
+
+	// Enable ADC.
+	nrf.SAADC.ENABLE.Set(nrf.SAADC_ENABLE_ENABLE_Enabled << nrf.SAADC_ENABLE_ENABLE_Pos)
+	for i := 0; i < 8; i++ {
+		nrf.SAADC.CH[i].PSELN.Set(nrf.SAADC_CH_PSELP_PSELP_NC)
+		nrf.SAADC.CH[i].PSELP.Set(nrf.SAADC_CH_PSELP_PSELP_NC)
+	}
+
+	// Configure ADC.
+	nrf.SAADC.CH[0].CONFIG.Set(((nrf.SAADC_CH_CONFIG_RESP_Bypass << nrf.SAADC_CH_CONFIG_RESP_Pos) & nrf.SAADC_CH_CONFIG_RESP_Msk) |
+		((nrf.SAADC_CH_CONFIG_RESP_Bypass << nrf.SAADC_CH_CONFIG_RESN_Pos) & nrf.SAADC_CH_CONFIG_RESN_Msk) |
+		((nrf.SAADC_CH_CONFIG_GAIN_Gain1_5 << nrf.SAADC_CH_CONFIG_GAIN_Pos) & nrf.SAADC_CH_CONFIG_GAIN_Msk) |
+		((nrf.SAADC_CH_CONFIG_REFSEL_Internal << nrf.SAADC_CH_CONFIG_REFSEL_Pos) & nrf.SAADC_CH_CONFIG_REFSEL_Msk) |
+		((nrf.SAADC_CH_CONFIG_TACQ_3us << nrf.SAADC_CH_CONFIG_TACQ_Pos) & nrf.SAADC_CH_CONFIG_TACQ_Msk) |
+		((nrf.SAADC_CH_CONFIG_MODE_SE << nrf.SAADC_CH_CONFIG_MODE_Pos) & nrf.SAADC_CH_CONFIG_MODE_Msk))
+
+	// Set pin to read.
+	nrf.SAADC.CH[0].PSELN.Set(pwmPin)
+	nrf.SAADC.CH[0].PSELP.Set(pwmPin)
+
+	// Destination for sample result.
+	nrf.SAADC.RESULT.PTR.Set(uint32(uintptr(unsafe.Pointer(&value))))
+	nrf.SAADC.RESULT.MAXCNT.Set(1) // One sample
+
+	// Start tasks.
+	nrf.SAADC.TASKS_START.Set(1)
+	for nrf.SAADC.EVENTS_STARTED.Get() == 0 {
+	}
+	nrf.SAADC.EVENTS_STARTED.Set(0x00)
+
+	// Start the sample task.
+	nrf.SAADC.TASKS_SAMPLE.Set(1)
+
+	// Wait until the sample task is done.
+	for nrf.SAADC.EVENTS_END.Get() == 0 {
+	}
+	nrf.SAADC.EVENTS_END.Set(0x00)
+
+	// Stop the ADC
+	nrf.SAADC.TASKS_STOP.Set(1)
+	for nrf.SAADC.EVENTS_STOPPED.Get() == 0 {
+	}
+	nrf.SAADC.EVENTS_STOPPED.Set(0)
+
+	// Disable the ADC.
+	nrf.SAADC.ENABLE.Set(nrf.SAADC_ENABLE_ENABLE_Disabled << nrf.SAADC_ENABLE_ENABLE_Pos)
+
+	if value < 0 {
+		value = 0
+	}
+
+	// Return 16-bit result from 12-bit value.
+	return uint16(value << 4)
+}
+
+// SPI on the NRF.
+type SPI struct {
+	Bus *nrf.SPIM_Type
+}
+
+// There are 3 SPI interfaces on the NRF528xx.
+var (
+	SPI0 = SPI{Bus: nrf.SPIM0}
+	SPI1 = SPI{Bus: nrf.SPIM1}
+	SPI2 = SPI{Bus: nrf.SPIM2}
+)
+
+// SPIConfig is used to store config info for SPI.
+type SPIConfig struct {
+	Frequency uint32
+	SCK       Pin
+	SDO       Pin
+	SDI       Pin
+	LSBFirst  bool
+	Mode      uint8
+}
+
+// Configure is intended to setup the SPI interface.
+func (spi SPI) Configure(config SPIConfig) {
+	// Disable bus to configure it
+	spi.Bus.ENABLE.Set(nrf.SPIM_ENABLE_ENABLE_Disabled)
+
+	// Pick a default frequency.
+	if config.Frequency == 0 {
+		config.Frequency = 4000000 // 4MHz
+	}
+
+	// set frequency
+	var freq uint32
+	switch {
+	case config.Frequency >= 8000000:
+		freq = nrf.SPIM_FREQUENCY_FREQUENCY_M8
+	case config.Frequency >= 4000000:
+		freq = nrf.SPIM_FREQUENCY_FREQUENCY_M4
+	case config.Frequency >= 2000000:
+		freq = nrf.SPIM_FREQUENCY_FREQUENCY_M2
+	case config.Frequency >= 1000000:
+		freq = nrf.SPIM_FREQUENCY_FREQUENCY_M1
+	case config.Frequency >= 500000:
+		freq = nrf.SPIM_FREQUENCY_FREQUENCY_K500
+	case config.Frequency >= 250000:
+		freq = nrf.SPIM_FREQUENCY_FREQUENCY_K250
+	default: // below 250kHz, default to the lowest speed available
+		freq = nrf.SPIM_FREQUENCY_FREQUENCY_K125
+	}
+	spi.Bus.FREQUENCY.Set(freq)
+
+	var conf uint32
+
+	// set bit transfer order
+	if config.LSBFirst {
+		conf = (nrf.SPIM_CONFIG_ORDER_LsbFirst << nrf.SPIM_CONFIG_ORDER_Pos)
+	}
+
+	// set mode
+	switch config.Mode {
+	case 0:
+		conf &^= (nrf.SPIM_CONFIG_CPOL_ActiveHigh << nrf.SPIM_CONFIG_CPOL_Pos)
+		conf &^= (nrf.SPIM_CONFIG_CPHA_Leading << nrf.SPIM_CONFIG_CPHA_Pos)
+	case 1:
+		conf &^= (nrf.SPIM_CONFIG_CPOL_ActiveHigh << nrf.SPIM_CONFIG_CPOL_Pos)
+		conf |= (nrf.SPIM_CONFIG_CPHA_Trailing << nrf.SPIM_CONFIG_CPHA_Pos)
+	case 2:
+		conf |= (nrf.SPIM_CONFIG_CPOL_ActiveLow << nrf.SPIM_CONFIG_CPOL_Pos)
+		conf &^= (nrf.SPIM_CONFIG_CPHA_Leading << nrf.SPIM_CONFIG_CPHA_Pos)
+	case 3:
+		conf |= (nrf.SPIM_CONFIG_CPOL_ActiveLow << nrf.SPIM_CONFIG_CPOL_Pos)
+		conf |= (nrf.SPIM_CONFIG_CPHA_Trailing << nrf.SPIM_CONFIG_CPHA_Pos)
+	default: // to mode
+		conf &^= (nrf.SPIM_CONFIG_CPOL_ActiveHigh << nrf.SPIM_CONFIG_CPOL_Pos)
+		conf &^= (nrf.SPIM_CONFIG_CPHA_Leading << nrf.SPIM_CONFIG_CPHA_Pos)
+	}
+	spi.Bus.CONFIG.Set(conf)
+
+	// set pins
+	if config.SCK == 0 && config.SDO == 0 && config.SDI == 0 {
+		config.SCK = SPI0_SCK_PIN
+		config.SDO = SPI0_SDO_PIN
+		config.SDI = SPI0_SDI_PIN
+	}
+	spi.Bus.PSEL.SCK.Set(uint32(config.SCK))
+	spi.Bus.PSEL.MOSI.Set(uint32(config.SDO))
+	spi.Bus.PSEL.MISO.Set(uint32(config.SDI))
+
+	// Re-enable bus now that it is configured.
+	spi.Bus.ENABLE.Set(nrf.SPIM_ENABLE_ENABLE_Enabled)
+}
+
+// Transfer writes/reads a single byte using the SPI interface.
+func (spi SPI) Transfer(w byte) (byte, error) {
+	var wbuf, rbuf [1]byte
+	wbuf[0] = w
+	err := spi.Tx(wbuf[:], rbuf[:])
+	return rbuf[0], err
+}
+
+// Tx handles read/write operation for SPI interface. Since SPI is a syncronous
+// write/read interface, there must always be the same number of bytes written
+// as bytes read. Therefore, if the number of bytes don't match it will be
+// padded until they fit: if len(w) > len(r) the extra bytes received will be
+// dropped and if len(w) < len(r) extra 0 bytes will be sent.
+func (spi SPI) Tx(w, r []byte) error {
+	// Unfortunately the hardware (on the nrf52832) only supports up to 255
+	// bytes in the buffers, so if either w or r is longer than that the
+	// transfer needs to be broken up in pieces.
+	// The nrf52840 supports far larger buffers however, which isn't yet
+	// supported.
+	for len(r) != 0 || len(w) != 0 {
+		// Prepare the SPI transfer: set the DMA pointers and lengths.
+		if len(r) != 0 {
+			spi.Bus.RXD.PTR.Set(uint32(uintptr(unsafe.Pointer(&r[0]))))
+			n := uint32(len(r))
+			if n > 255 {
+				n = 255
+			}
+			spi.Bus.RXD.MAXCNT.Set(n)
+			r = r[n:]
+		}
+		if len(w) != 0 {
+			spi.Bus.TXD.PTR.Set(uint32(uintptr(unsafe.Pointer(&w[0]))))
+			n := uint32(len(w))
+			if n > 255 {
+				n = 255
+			}
+			spi.Bus.TXD.MAXCNT.Set(n)
+			w = w[n:]
+		}
+
+		// Do the transfer.
+		// Note: this can be improved by not waiting until the transfer is
+		// finished if the transfer is send-only (a common case).
+		spi.Bus.TASKS_START.Set(1)
+		for spi.Bus.EVENTS_END.Get() == 0 {
+		}
+		spi.Bus.EVENTS_END.Set(0)
+	}
+
+	return nil
+}
+
+// InitPWM initializes the registers needed for PWM.
+func InitPWM() {
+	return
+}
+
+// Configure configures a PWM pin for output.
+func (pwm PWM) Configure() {
+}
+
+// Set turns on the duty cycle for a PWM pin using the provided value.
+func (pwm PWM) Set(value uint16) {
+	for i := 0; i < len(pwmChannelPins); i++ {
+		if pwmChannelPins[i] == 0xFFFFFFFF || pwmChannelPins[i] == uint32(pwm.Pin) {
+			pwmChannelPins[i] = uint32(pwm.Pin)
+			pwmChannelSequence[i] = (value >> 2) | 0x8000 // set bit 15 to invert polarity
+
+			p := pwms[i]
+
+			p.PSEL.OUT[0].Set(uint32(pwm.Pin))
+			p.PSEL.OUT[1].Set(uint32(pwm.Pin))
+			p.PSEL.OUT[2].Set(uint32(pwm.Pin))
+			p.PSEL.OUT[3].Set(uint32(pwm.Pin))
+			p.ENABLE.Set(nrf.PWM_ENABLE_ENABLE_Enabled << nrf.PWM_ENABLE_ENABLE_Pos)
+			p.PRESCALER.Set(nrf.PWM_PRESCALER_PRESCALER_DIV_2)
+			p.MODE.Set(nrf.PWM_MODE_UPDOWN_Up)
+			p.COUNTERTOP.Set(16384) // frequency
+			p.LOOP.Set(0)
+			p.DECODER.Set((nrf.PWM_DECODER_LOAD_Common << nrf.PWM_DECODER_LOAD_Pos) | (nrf.PWM_DECODER_MODE_RefreshCount << nrf.PWM_DECODER_MODE_Pos))
+			p.SEQ[0].PTR.Set(uint32(uintptr(unsafe.Pointer(&pwmChannelSequence[i]))))
+			p.SEQ[0].CNT.Set(1)
+			p.SEQ[0].REFRESH.Set(1)
+			p.SEQ[0].ENDDELAY.Set(0)
+			p.TASKS_SEQSTART[0].Set(1)
+
+			break
+		}
+	}
+}

src/machine/machine_stm32_f7_uart.go

@@ -0,0 +1,62 @@
+// +build stm32f7
+
+package machine
+
+// Peripheral abstraction layer for UARTs on the stm32 family.
+
+import (
+	"device/stm32"
+	"runtime/interrupt"
+	"unsafe"
+)
+
+// Configure the UART.
+func (uart UART) Configure(config UARTConfig) {
+	// Default baud rate to 115200.
+	if config.BaudRate == 0 {
+		config.BaudRate = 115200
+	}
+
+	// Set the GPIO pins to defaults if they're not set
+	if config.TX == 0 && config.RX == 0 {
+		config.TX = UART_TX_PIN
+		config.RX = UART_RX_PIN
+	}
+
+	// Enable USART clock
+	enableAltFuncClock(unsafe.Pointer(uart.Bus))
+
+	uart.configurePins(config)
+
+	// Set baud rate
+	uart.SetBaudRate(config.BaudRate)
+
+	// Enable USART port, tx, rx and rx interrupts
+	uart.Bus.CR1.Set(stm32.USART_CR1_TE | stm32.USART_CR1_RE | stm32.USART_CR1_RXNEIE | stm32.USART_CR1_UE)
+
+	// Enable RX IRQ
+	uart.Interrupt.SetPriority(0xc0)
+	uart.Interrupt.Enable()
+}
+
+// handleInterrupt should be called from the appropriate interrupt handler for
+// this UART instance.
+func (uart *UART) handleInterrupt(interrupt.Interrupt) {
+	uart.Receive(byte((uart.Bus.RDR.Get() & 0xFF)))
+}
+
+// SetBaudRate sets the communication speed for the UART. Defer to chip-specific
+// routines for calculation
+func (uart UART) SetBaudRate(br uint32) {
+	divider := uart.getBaudRateDivisor(br)
+	uart.Bus.BRR.Set(divider)
+}
+
+// WriteByte writes a byte of data to the UART.
+func (uart UART) WriteByte(c byte) error {
+	uart.Bus.TDR.Set(uint32(c))
+
+	for !uart.Bus.ISR.HasBits(stm32.USART_ISR_TXE) {
+	}
+	return nil
+}

src/machine/machine_stm32_i2c.go

@@ -1,4 +1,4 @@
-// +build stm32,!stm32f103xx,!stm32f407
+// +build stm32,!stm32f103xx,!stm32f407,!stm32f7x2
 
 package machine
 

src/machine/machine_stm32_spi.go

@@ -1,4 +1,4 @@
-// +build stm32
+// +build stm32,!stm32f7x2
 
 package machine
 

src/machine/machine_stm32_uart.go

@@ -1,4 +1,4 @@
-// +build stm32
+// +build stm32,!stm32f7
 
 package machine
 

src/machine/machine_stm32f7.go

@@ -0,0 +1,259 @@
+// +build stm32f7
+
+package machine
+
+// Peripheral abstraction layer for the stm32f4
+
+import (
+	"device/stm32"
+	"unsafe"
+)
+
+const (
+	PA0  = portA + 0
+	PA1  = portA + 1
+	PA2  = portA + 2
+	PA3  = portA + 3
+	PA4  = portA + 4
+	PA5  = portA + 5
+	PA6  = portA + 6
+	PA7  = portA + 7
+	PA8  = portA + 8
+	PA9  = portA + 9
+	PA10 = portA + 10
+	PA11 = portA + 11
+	PA12 = portA + 12
+	PA13 = portA + 13
+	PA14 = portA + 14
+	PA15 = portA + 15
+
+	PB0  = portB + 0
+	PB1  = portB + 1
+	PB2  = portB + 2
+	PB3  = portB + 3
+	PB4  = portB + 4
+	PB5  = portB + 5
+	PB6  = portB + 6
+	PB7  = portB + 7
+	PB8  = portB + 8
+	PB9  = portB + 9
+	PB10 = portB + 10
+	PB11 = portB + 11
+	PB12 = portB + 12
+	PB13 = portB + 13
+	PB14 = portB + 14
+	PB15 = portB + 15
+
+	PC0  = portC + 0
+	PC1  = portC + 1
+	PC2  = portC + 2
+	PC3  = portC + 3
+	PC4  = portC + 4
+	PC5  = portC + 5
+	PC6  = portC + 6
+	PC7  = portC + 7
+	PC8  = portC + 8
+	PC9  = portC + 9
+	PC10 = portC + 10
+	PC11 = portC + 11
+	PC12 = portC + 12
+	PC13 = portC + 13
+	PC14 = portC + 14
+	PC15 = portC + 15
+
+	PD0  = portD + 0
+	PD1  = portD + 1
+	PD2  = portD + 2
+	PD3  = portD + 3
+	PD4  = portD + 4
+	PD5  = portD + 5
+	PD6  = portD + 6
+	PD7  = portD + 7
+	PD8  = portD + 8
+	PD9  = portD + 9
+	PD10 = portD + 10
+	PD11 = portD + 11
+	PD12 = portD + 12
+	PD13 = portD + 13
+	PD14 = portD + 14
+	PD15 = portD + 15
+
+	PE0  = portE + 0
+	PE1  = portE + 1
+	PE2  = portE + 2
+	PE3  = portE + 3
+	PE4  = portE + 4
+	PE5  = portE + 5
+	PE6  = portE + 6
+	PE7  = portE + 7
+	PE8  = portE + 8
+	PE9  = portE + 9
+	PE10 = portE + 10
+	PE11 = portE + 11
+	PE12 = portE + 12
+	PE13 = portE + 13
+	PE14 = portE + 14
+	PE15 = portE + 15
+
+	PF0  = portF + 0
+	PF1  = portF + 1
+	PF2  = portF + 2
+	PF3  = portF + 3
+	PF4  = portF + 4
+	PF5  = portF + 5
+	PF6  = portF + 6
+	PF7  = portF + 7
+	PF8  = portF + 8
+	PF9  = portF + 9
+	PF10 = portF + 10
+	PF11 = portF + 11
+	PF12 = portF + 12
+	PF13 = portF + 13
+	PF14 = portF + 14
+	PF15 = portF + 15
+
+	PG0  = portG + 0
+	PG1  = portG + 1
+	PG2  = portG + 2
+	PG3  = portG + 3
+	PG4  = portG + 4
+	PG5  = portG + 5
+	PG6  = portG + 6
+	PG7  = portG + 7
+	PG8  = portG + 8
+	PG9  = portG + 9
+	PG10 = portG + 10
+	PG11 = portG + 11
+	PG12 = portG + 12
+	PG13 = portG + 13
+	PG14 = portG + 14
+	PG15 = portG + 15
+
+	PH0 = portH + 0
+	PH1 = portH + 1
+)
+
+func (p Pin) getPort() *stm32.GPIO_Type {
+	switch p / 16 {
+	case 0:
+		return stm32.GPIOA
+	case 1:
+		return stm32.GPIOB
+	case 2:
+		return stm32.GPIOC
+	case 3:
+		return stm32.GPIOD
+	case 4:
+		return stm32.GPIOE
+	case 5:
+		return stm32.GPIOF
+	case 6:
+		return stm32.GPIOG
+	case 7:
+		return stm32.GPIOH
+	case 8:
+		return stm32.GPIOI
+	default:
+		panic("machine: unknown port")
+	}
+}
+
+// enableClock enables the clock for this desired GPIO port.
+func (p Pin) enableClock() {
+	switch p / 16 {
+	case 0:
+		stm32.RCC.AHB1ENR.SetBits(stm32.RCC_AHB1ENR_GPIOAEN)
+	case 1:
+		stm32.RCC.AHB1ENR.SetBits(stm32.RCC_AHB1ENR_GPIOBEN)
+	case 2:
+		stm32.RCC.AHB1ENR.SetBits(stm32.RCC_AHB1ENR_GPIOCEN)
+	case 3:
+		stm32.RCC.AHB1ENR.SetBits(stm32.RCC_AHB1ENR_GPIODEN)
+	case 4:
+		stm32.RCC.AHB1ENR.SetBits(stm32.RCC_AHB1ENR_GPIOEEN)
+	case 5:
+		stm32.RCC.AHB1ENR.SetBits(stm32.RCC_AHB1ENR_GPIOFEN)
+	case 6:
+		stm32.RCC.AHB1ENR.SetBits(stm32.RCC_AHB1ENR_GPIOGEN)
+	case 7:
+		stm32.RCC.AHB1ENR.SetBits(stm32.RCC_AHB1ENR_GPIOHEN)
+	case 8:
+		stm32.RCC.AHB1ENR.SetBits(stm32.RCC_AHB1ENR_GPIOIEN)
+	default:
+		panic("machine: unknown port")
+	}
+}
+
+// Enable peripheral clock
+func enableAltFuncClock(bus unsafe.Pointer) {
+	switch bus {
+	case unsafe.Pointer(stm32.DAC): // DAC interface clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_DACEN)
+	case unsafe.Pointer(stm32.PWR): // Power interface clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_PWREN)
+	case unsafe.Pointer(stm32.CAN1): // CAN 1 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_CAN1EN)
+	case unsafe.Pointer(stm32.I2C3): // I2C3 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_I2C3EN)
+	case unsafe.Pointer(stm32.I2C2): // I2C2 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_I2C2EN)
+	case unsafe.Pointer(stm32.I2C1): // I2C1 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_I2C1EN)
+	case unsafe.Pointer(stm32.UART5): // UART5 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_UART5EN)
+	case unsafe.Pointer(stm32.UART4): // UART4 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_UART4EN)
+	case unsafe.Pointer(stm32.USART3): // USART3 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_USART3EN)
+	case unsafe.Pointer(stm32.USART2): // USART2 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_USART2EN)
+	case unsafe.Pointer(stm32.SPI3): // SPI3 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_SPI3EN)
+	case unsafe.Pointer(stm32.SPI2): // SPI2 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_SPI2EN)
+	case unsafe.Pointer(stm32.WWDG): // Window watchdog clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_WWDGEN)
+	case unsafe.Pointer(stm32.TIM14): // TIM14 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_TIM14EN)
+	case unsafe.Pointer(stm32.TIM13): // TIM13 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_TIM13EN)
+	case unsafe.Pointer(stm32.TIM12): // TIM12 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_TIM12EN)
+	case unsafe.Pointer(stm32.TIM7): // TIM7 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_TIM7EN)
+	case unsafe.Pointer(stm32.TIM6): // TIM6 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_TIM6EN)
+	case unsafe.Pointer(stm32.TIM5): // TIM5 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_TIM5EN)
+	case unsafe.Pointer(stm32.TIM4): // TIM4 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_TIM4EN)
+	case unsafe.Pointer(stm32.TIM3): // TIM3 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_TIM3EN)
+	case unsafe.Pointer(stm32.TIM2): // TIM2 clock enable
+		stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_TIM2EN)
+	case unsafe.Pointer(stm32.TIM11): // TIM11 clock enable
+		stm32.RCC.APB2ENR.SetBits(stm32.RCC_APB2ENR_TIM11EN)
+	case unsafe.Pointer(stm32.TIM10): // TIM10 clock enable
+		stm32.RCC.APB2ENR.SetBits(stm32.RCC_APB2ENR_TIM10EN)
+	case unsafe.Pointer(stm32.TIM9): // TIM9 clock enable
+		stm32.RCC.APB2ENR.SetBits(stm32.RCC_APB2ENR_TIM9EN)
+	case unsafe.Pointer(stm32.SYSCFG): // System configuration controller clock enable
+		stm32.RCC.APB2ENR.SetBits(stm32.RCC_APB2ENR_SYSCFGEN)
+	case unsafe.Pointer(stm32.SPI1): // SPI1 clock enable
+		stm32.RCC.APB2ENR.SetBits(stm32.RCC_APB2ENR_SPI1EN)
+	case unsafe.Pointer(stm32.ADC3): // ADC3 clock enable
+		stm32.RCC.APB2ENR.SetBits(stm32.RCC_APB2ENR_ADC3EN)
+	case unsafe.Pointer(stm32.ADC2): // ADC2 clock enable
+		stm32.RCC.APB2ENR.SetBits(stm32.RCC_APB2ENR_ADC2EN)
+	case unsafe.Pointer(stm32.ADC1): // ADC1 clock enable
+		stm32.RCC.APB2ENR.SetBits(stm32.RCC_APB2ENR_ADC1EN)
+	case unsafe.Pointer(stm32.USART6): // USART6 clock enable
+		stm32.RCC.APB2ENR.SetBits(stm32.RCC_APB2ENR_USART6EN)
+	case unsafe.Pointer(stm32.USART1): // USART1 clock enable
+		stm32.RCC.APB2ENR.SetBits(stm32.RCC_APB2ENR_USART1EN)
+	case unsafe.Pointer(stm32.TIM8): // TIM8 clock enable
+		stm32.RCC.APB2ENR.SetBits(stm32.RCC_APB2ENR_TIM8EN)
+	case unsafe.Pointer(stm32.TIM1): // TIM1 clock enable
+		stm32.RCC.APB2ENR.SetBits(stm32.RCC_APB2ENR_TIM1EN)
+	}
+}

src/machine/machine_stm32f7x2.go

@@ -0,0 +1,44 @@
+// +build stm32f7x2
+
+package machine
+
+// Peripheral abstraction layer for the stm32f407
+
+import (
+	"device/stm32"
+	"runtime/interrupt"
+)
+
+func CPUFrequency() uint32 {
+	return 216000000
+}
+
+//---------- UART related types and code
+
+// UART representation
+type UART struct {
+	Buffer          *RingBuffer
+	Bus             *stm32.USART_Type
+	Interrupt       interrupt.Interrupt
+	AltFuncSelector stm32.AltFunc
+}
+
+// Configure the UART.
+func (uart UART) configurePins(config UARTConfig) {
+	// enable the alternate functions on the TX and RX pins
+	config.TX.ConfigureAltFunc(PinConfig{Mode: PinModeUARTTX}, uart.AltFuncSelector)
+	config.RX.ConfigureAltFunc(PinConfig{Mode: PinModeUARTRX}, uart.AltFuncSelector)
+}
+
+// UART baudrate calc based on the bus and clockspeed
+// NOTE: keep this in sync with the runtime/runtime_stm32f7x2.go clock init code
+func (uart UART) getBaudRateDivisor(baudRate uint32) uint32 {
+	var clock uint32
+	switch uart.Bus {
+	case stm32.USART1, stm32.USART6:
+		clock = CPUFrequency() / 2 // APB2 Frequency
+	case stm32.USART2, stm32.USART3, stm32.UART4, stm32.UART5:
+		clock = CPUFrequency() / 8 // APB1 Frequency
+	}
+	return clock / baudRate
+}

src/machine/spi.go

@@ -1,4 +1,4 @@
-// +build !baremetal sam stm32 fe310 k210
+// +build !baremetal sam stm32,!stm32f7x2 fe310 k210
 
 package machine
 

src/runtime/arch_xtensa.go

@@ -2,6 +2,8 @@
 
 package runtime
 
+import "device"
+
 const GOARCH = "arm" // xtensa pretends to be arm
 
 // The bitness of the CPU (e.g. 8, 32, 64).
@@ -12,4 +14,7 @@ func align(ptr uintptr) uintptr {
 	return (ptr + 3) &^ 3
 }
 
-func getCurrentStackPointer() uintptr
+func getCurrentStackPointer() uintptr {
+	// The stack pointer (sp) is a1.
+	return device.AsmFull("mov {}, sp", nil)
+}

src/runtime/interface.go

@@ -123,7 +123,7 @@ type structField struct {
 // than a function call. Also, by keeping the method set around it is easier to
 // implement interfaceImplements in the interp package.
 type typeInInterface struct {
-	typecode  *typecodeID
+	typecode  *typecodeID          // element type, underlying type, or reference to struct fields
 	methodSet *interfaceMethodInfo // nil or a GEP of an array
 }
 

src/runtime/runtime_stm32f7x2.go

@@ -0,0 +1,208 @@
+// +build stm32,stm32f7x2
+
+package runtime
+
+import (
+	"device/arm"
+	"device/stm32"
+	"machine"
+	"runtime/interrupt"
+	"runtime/volatile"
+)
+
+func init() {
+	initCLK()
+	initTIM3()
+	machine.UART0.Configure(machine.UARTConfig{})
+	initTIM7()
+}
+
+func putchar(c byte) {
+	machine.UART0.WriteByte(c)
+}
+
+const (
+	HSE_STARTUP_TIMEOUT = 0x0500
+	PLL_M               = 4
+	PLL_N               = 216
+	PLL_P               = 2
+	PLL_Q               = 2
+)
+
+/*
+   clock settings
+   +-------------+--------+
+   | HSE         | 8mhz   |
+   | SYSCLK      | 216mhz |
+   | HCLK        | 216mhz |
+   | APB1(PCLK1) | 27mhz  |
+   | APB2(PCLK2) | 108mhz |
+   +-------------+--------+
+*/
+func initCLK() {
+	// PWR_CLK_ENABLE
+	stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_PWREN)
+	_ = stm32.RCC.APB1ENR.Get()
+
+	// PWR_VOLTAGESCALING_CONFIG
+	stm32.PWR.CR1.ReplaceBits(0x3<<stm32.PWR_CR1_VOS_Pos, stm32.PWR_CR1_VOS_Msk, 0)
+	_ = stm32.PWR.CR1.Get()
+
+	// Initialize the High-Speed External Oscillator
+	initOsc()
+
+	// Set flash wait states (min 7 latency units) based on clock
+	if (stm32.FLASH.ACR.Get() & stm32.FLASH_ACR_LATENCY_Msk) < 7 {
+		stm32.FLASH.ACR.ReplaceBits(7, stm32.FLASH_ACR_LATENCY_Msk, 0)
+	}
+
+	// HCLK (0x1C00 = DIV_16, 0x0 = RCC_SYSCLK_DIV1) - ensure timers remain
+	// within spec as the SYSCLK source changes.
+	stm32.RCC.CFGR.ReplaceBits(0x00001C00, stm32.RCC_CFGR_PPRE1_Msk, 0)
+	stm32.RCC.CFGR.ReplaceBits(0x00001C00<<3, stm32.RCC_CFGR_PPRE2_Msk, 0)
+	stm32.RCC.CFGR.ReplaceBits(0, stm32.RCC_CFGR_HPRE_Msk, 0)
+
+	// Set SYSCLK source and wait
+	// (2 = PLLCLK, 3 = RCC_CFGR_SW mask, 3 << 3 = RCC_CFGR_SWS mask)
+	stm32.RCC.CFGR.ReplaceBits(2, 3, 0)
+	for stm32.RCC.CFGR.Get()&(3<<2) != (2 << 2) {
+	}
+
+	// Set flash wait states (max 7 latency units) based on clock
+	if (stm32.FLASH.ACR.Get() & stm32.FLASH_ACR_LATENCY_Msk) > 7 {
+		stm32.FLASH.ACR.ReplaceBits(7, stm32.FLASH_ACR_LATENCY_Msk, 0)
+	}
+
+	// Set APB1 and APB2 clocks (0x1800 = DIV8, 0x1000 = DIV2)
+	stm32.RCC.CFGR.ReplaceBits(0x1800, stm32.RCC_CFGR_PPRE1_Msk, 0)
+	stm32.RCC.CFGR.ReplaceBits(0x1000<<3, stm32.RCC_CFGR_PPRE2_Msk, 0)
+}
+
+func initOsc() {
+	// Enable HSE, wait until ready
+	stm32.RCC.CR.SetBits(stm32.RCC_CR_HSEON)
+	for !stm32.RCC.CR.HasBits(stm32.RCC_CR_HSERDY) {
+	}
+
+	// Disable the PLL, wait until disabled
+	stm32.RCC.CR.ClearBits(stm32.RCC_CR_PLLON)
+	for stm32.RCC.CR.HasBits(stm32.RCC_CR_PLLRDY) {
+	}
+
+	// Configure the PLL
+	stm32.RCC.PLLCFGR.Set(0x20000000 |
+		(1 << stm32.RCC_PLLCFGR_PLLSRC_Pos) | // 1 = HSE
+		PLL_M |
+		(PLL_N << stm32.RCC_PLLCFGR_PLLN0_Pos) |
+		(((PLL_P >> 1) - 1) << stm32.RCC_PLLCFGR_PLLP0_Pos) |
+		(PLL_Q << stm32.RCC_PLLCFGR_PLLQ0_Pos))
+
+	// Enable the PLL, wait until ready
+	stm32.RCC.CR.SetBits(stm32.RCC_CR_PLLON)
+	for !stm32.RCC.CR.HasBits(stm32.RCC_CR_PLLRDY) {
+	}
+}
+
+var (
+	// tick in milliseconds
+	tickCount timeUnit
+)
+
+var timerWakeup volatile.Register8
+
+func ticksToNanoseconds(ticks timeUnit) int64 {
+	return int64(ticks) * 1000
+}
+
+func nanosecondsToTicks(ns int64) timeUnit {
+	return timeUnit(ns / 1000)
+}
+
+// Enable the TIM3 clock.(sleep count)
+func initTIM3() {
+	stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_TIM3EN)
+
+	intr := interrupt.New(stm32.IRQ_TIM3, handleTIM3)
+	intr.SetPriority(0xc3)
+	intr.Enable()
+}
+
+// Enable the TIM7 clock.(tick count)
+func initTIM7() {
+	stm32.RCC.APB1ENR.SetBits(stm32.RCC_APB1ENR_TIM7EN)
+
+	// CK_INT = APB1 x2 = 54mhz
+	stm32.TIM7.PSC.Set(54000000/10000 - 1) // 54mhz to 10khz(0.1ms)
+	stm32.TIM7.ARR.Set(10 - 1)             // interrupt per 1ms
+
+	// Enable the hardware interrupt.
+	stm32.TIM7.DIER.SetBits(stm32.TIM_DIER_UIE)
+
+	// Enable the timer.
+	stm32.TIM7.CR1.SetBits(stm32.TIM_CR1_CEN)
+
+	intr := interrupt.New(stm32.IRQ_TIM7, handleTIM7)
+	intr.SetPriority(0xc1)
+	intr.Enable()
+}
+
+const asyncScheduler = false
+
+// sleepTicks should sleep for specific number of microseconds.
+func sleepTicks(d timeUnit) {
+	timerSleep(uint32(d))
+}
+
+// number of ticks (microseconds) since start.
+func ticks() timeUnit {
+	// milliseconds to microseconds
+	return tickCount * 1000
+}
+
+// ticks are in microseconds
+func timerSleep(ticks uint32) {
+	timerWakeup.Set(0)
+
+	// CK_INT = APB1 x2 = 54mhz
+	// prescale counter down from 54mhz to 10khz aka 0.1 ms frequency.
+	stm32.TIM3.PSC.Set(54000000/10000 - 1)
+
+	// set duty aka duration
+	arr := (ticks / 100) - 1 // convert from microseconds to 0.1 ms
+	if arr == 0 {
+		arr = 1 // avoid blocking
+	}
+	stm32.TIM3.ARR.Set(arr)
+
+	// Enable the hardware interrupt.
+	stm32.TIM3.DIER.SetBits(stm32.TIM_DIER_UIE)
+
+	// Enable the timer.
+	stm32.TIM3.CR1.SetBits(stm32.TIM_CR1_CEN)
+
+	// wait till timer wakes up
+	for timerWakeup.Get() == 0 {
+		arm.Asm("wfi")
+	}
+}
+
+func handleTIM3(interrupt.Interrupt) {
+	if stm32.TIM3.SR.HasBits(stm32.TIM_SR_UIF) {
+		// Disable the timer.
+		stm32.TIM3.CR1.ClearBits(stm32.TIM_CR1_CEN)
+
+		// clear the update flag
+		stm32.TIM3.SR.ClearBits(stm32.TIM_SR_UIF)
+
+		// timer was triggered
+		timerWakeup.Set(1)
+	}
+}
+
+func handleTIM7(interrupt.Interrupt) {
+	if stm32.TIM7.SR.HasBits(stm32.TIM_SR_UIF) {
+		// clear the update flag
+		stm32.TIM7.SR.ClearBits(stm32.TIM_SR_UIF)
+		tickCount++
+	}
+}

src/runtime/scheduler_avr.S

@@ -1,189 +0,0 @@
-.section .bss.tinygo_systemStack
-.global tinygo_systemStack
-.type   tinygo_systemStack, %object
-tinygo_systemStack:
-    .short 0
-
-.section .text.tinygo_startTask
-.global  tinygo_startTask
-.type    tinygo_startTask, %function
-tinygo_startTask:
-    // Small assembly stub for starting a goroutine. This is already run on the
-    // new stack, with the callee-saved registers already loaded.
-    // Most importantly, r2r3 contain the pc of the to-be-started function and
-    // r4r5 contain the only argument it is given. Multiple arguments are packed
-    // into one by storing them in a new allocation.
-
-    // Set the first argument of the goroutine start wrapper, which contains all
-    // the arguments.
-    movw  r24, r4
-
-    // Branch to the "goroutine start" function. Note that the Z register is
-    // call-clobbered, so does not need to be restored after use.
-    movw  Z, r2
-    icall
-
-    // After return, exit this goroutine. This is a tail call.
-#if __AVR_ARCH__ == 2 || __AVR_ARCH__ == 25
-    // Small memory devices (≤8kB flash) that do not have the long call
-    // instruction availble will need to use rcall instead.
-    // Note that they will probably not be able to run more than the main
-    // goroutine anyway, but this file is compiled for all AVRs so it needs to
-    // compile at least.
-    rcall  tinygo_pause
-#else
-    // Other devices can (and must) use the regular call instruction.
-    call tinygo_pause
-#endif
-
-// Get the system stack pointer, independent of whether we're currently on the
-// system stack or a task stack.
-.global tinygo_getSystemStackPointer
-.type tinygo_getSystemStackPointer, %function
-tinygo_getSystemStackPointer:
-    // Load system stack pointer.
-    lds r24, tinygo_systemStack
-    lds r25, tinygo_systemStack+1
-
-    // Compare against 0.
-    cp  r24, r1
-    cpc r25, r1
-
-    // Branch (and then return) if tinygo_systemStack has a non-zero value.
-    brne 1f
-
-    // tinygo_systemStack is zero, so return the current stack pointer.
-    in  r24, 0x3d; SPL
-    in  r25, 0x3e; SPH
-
-1:
-    ret
-
-.global tinygo_switchToTask
-.type tinygo_switchToTask, %function
-tinygo_switchToTask:
-    // The sp parameter is the only parameter, so it will take up r24:r25.
-    // r24:r25 = sp uintptr
-
-    // Save all call-saved registers:
-    // https://gcc.gnu.org/wiki/avr-gcc#Call-Saved_Registers
-    push r29 // Y
-    push r28 // Y
-    push r17
-    push r16
-    push r15
-    push r14
-    push r13
-    push r12
-    push r11
-    push r10
-    push r9
-    push r8
-    push r7
-    push r6
-    push r5
-    push r4
-    push r3
-    push r2
-
-    // Save the system stack pointer in a global.
-    in  r2, 0x3d; SPL
-    in  r3, 0x3e; SPH
-    sts tinygo_systemStack+0, r2
-    sts tinygo_systemStack+1, r3
-
-    // Switch to the task stack pointer.
-    out  0x3d, r24; SPL
-    out  0x3e, r25; SPH
-
-    // Load saved register from the task stack.
-    pop r2
-    pop r3
-    pop r4
-    pop r5
-    pop r6
-    pop r7
-    pop r8
-    pop r9
-    pop r10
-    pop r11
-    pop r12
-    pop r13
-    pop r14
-    pop r15
-    pop r16
-    pop r17
-    pop r28 // Y
-    pop r29 // Y
-
-    // Return into the new task, as if tinygo_switchToScheduler was a regular
-    // call.
-    ret
-
-.global tinygo_switchToScheduler
-.type tinygo_switchToScheduler, %function
-tinygo_switchToScheduler:
-    // The sp parameter is the only parameter, so it will take up r24:r25.
-    // r24:r25 = sp *uintptr
-
-    // Save all call-saved registers on the task stack:
-    // https://gcc.gnu.org/wiki/avr-gcc#Call-Saved_Registers
-    push r29 // Y
-    push r28 // Y
-    push r17
-    push r16
-    push r15
-    push r14
-    push r13
-    push r12
-    push r11
-    push r10
-    push r9
-    push r8
-    push r7
-    push r6
-    push r5
-    push r4
-    push r3
-    push r2
-
-    // Save the task stack.
-    in  r2, 0x3d; SPL
-    in  r3, 0x3e; SPH
-    movw Y, r24
-    std Y+0, r2
-    std Y+1, r3
-
-    // Switch to the system stack.
-    lds r2, tinygo_systemStack
-    lds r3, tinygo_systemStack+1
-    out  0x3d, r2; SPL
-    out  0x3e, r3; SPH
-
-    // Clear tinygo_systemStack to make sure tinygo_getSystemStackPointer knows
-    // which pointer to return.
-    sts tinygo_systemStack+0, r1
-    sts tinygo_systemStack+1, r1
-
-    // Load saved register from the system stack.
-    pop r2
-    pop r3
-    pop r4
-    pop r5
-    pop r6
-    pop r7
-    pop r8
-    pop r9
-    pop r10
-    pop r11
-    pop r12
-    pop r13
-    pop r14
-    pop r15
-    pop r16
-    pop r17
-    pop r28 // Y
-    pop r29 // Y
-
-    // Return into the scheduler, as if tinygo_switchToTask was a regular call.
-    ret

src/runtime/scheduler_tasks.go

@@ -2,7 +2,16 @@
 
 package runtime
 
+import "internal/task"
+
 // getSystemStackPointer returns the current stack pointer of the system stack.
 // This is not necessarily the same as the current stack pointer.
-//export tinygo_getSystemStackPointer
-func getSystemStackPointer() uintptr
+func getSystemStackPointer() uintptr {
+	// TODO: this always returns the correct stack on Cortex-M, so don't bother
+	// comparing against 0.
+	sp := task.SystemStack()
+	if sp == 0 {
+		sp = getCurrentStackPointer()
+	}
+	return sp
+}

targets/arduino-mkr1000.json

@@ -0,0 +1,6 @@
+{
+    "inherits": ["atsamd21g18a"],
+    "build-tags": ["arduino_mkr1000"],
+    "flash-command": "bossac -i -e -w -v -R -U --port={port} --offset=0x2000 {bin}",
+    "flash-1200-bps-reset": "true"
+}

targets/avr.json

@@ -13,7 +13,8 @@
 		"-Wl,--gc-sections"
 	],
 	"extra-files": [
-		"src/runtime/gc_avr.S",
-		"src/runtime/scheduler_avr.S"
-	]
+		"src/internal/task/task_stack_avr.S",
+		"src/runtime/gc_avr.S"
+	],
+	"gdb": "avr-gdb"
 }

targets/cortex-m.json

@@ -25,8 +25,8 @@
 	],
 	"extra-files": [
 		"src/device/arm/cortexm.s",
-		"src/runtime/gc_arm.S",
-		"src/runtime/scheduler_cortexm.S"
+		"src/internal/task/task_stack_cortexm.S",
+		"src/runtime/gc_arm.S"
 	],
 	"gdb": "gdb-multiarch"
 }

targets/cortex-m7.json

@@ -0,0 +1,9 @@
+{
+	"inherits": ["cortex-m"],
+	"llvm-target": "armv7em-none-eabi",
+	"cflags": [
+		"--target=armv7em-none-eabi",
+		"-Qunused-arguments"
+	],
+	"gdb": "arm-none-eabi-gdb"
+}

targets/esp32.json

@@ -2,13 +2,17 @@
 	"inherits": ["xtensa"],
 	"cpu": "esp32",
 	"build-tags": ["esp32", "esp"],
+	"scheduler": "tasks",
 	"linker": "xtensa-esp32-elf-ld",
+	"default-stack-size": 2048,
 	"cflags": [
 		"-mcpu=esp32"
 	],
+	"rtlib": "compiler-rt",
 	"linkerscript": "targets/esp32.ld",
 	"extra-files": [
-		"src/device/esp/esp32.S"
+		"src/device/esp/esp32.S",
+		"src/internal/task/task_stack_esp32.S"
 	],
 	"binary-format": "esp32",
 	"flash-command": "esptool.py --chip=esp32 --port {port} write_flash 0x1000 {bin} -ff 80m -fm dout"

targets/esp8266.json

@@ -2,13 +2,16 @@
 	"inherits": ["xtensa"],
 	"cpu": "esp8266",
 	"build-tags": ["esp8266", "esp"],
+	"scheduler": "tasks",
 	"linker": "xtensa-esp32-elf-ld",
+	"default-stack-size": 2048,
 	"cflags": [
 		"-mcpu=esp8266"
 	],
 	"linkerscript": "targets/esp8266.ld",
 	"extra-files": [
-		"src/device/esp/esp8266.S"
+		"src/device/esp/esp8266.S",
+		"src/internal/task/task_stack_esp8266.S"
 	],
 	"binary-format": "esp8266",
 	"flash-command": "esptool.py --chip=esp8266 --port {port} write_flash 0x00000 {bin} -fm qio"

targets/matrixportal-m4.json

@@ -0,0 +1,11 @@
+{
+    "inherits": ["atsamd51j19a"],
+    "build-tags": ["atsamd51j19a", "matrixportal_m4"],
+    "flash-1200-bps-reset": "true",
+    "flash-method": "msd",
+    "msd-volume-name": "MATRIXBOOT",
+    "msd-firmware-name": "firmware.uf2",
+    "openocd-transport": "swd",
+    "openocd-interface": "jlink",
+    "openocd-target": "atsame5x"
+}

targets/nucleo-f722ze.json

@@ -0,0 +1,11 @@
+{
+  "inherits": ["cortex-m7"],
+  "build-tags": ["nucleof722ze", "stm32f7x2", "stm32f7", "stm32"],
+  "linkerscript": "targets/stm32f7x2zetx.ld",
+  "extra-files": [
+    "src/device/stm32/stm32f7x2.s"
+  ],
+  "flash-method": "openocd",
+  "openocd-interface": "stlink-v2-1",
+  "openocd-target": "stm32f7x"
+}

targets/stm32f7x2zetx.ld

@@ -0,0 +1,10 @@
+
+MEMORY
+{
+    FLASH_TEXT (rw) : ORIGIN = 0x08000000, LENGTH = 512K
+    RAM (xrw)       : ORIGIN = 0x20000000, LENGTH = 256K
+}
+
+_stack_size = 4K;
+
+INCLUDE "targets/arm.ld"