Update README.md

main: version 0.18.0
wasm: scan globals conservatively
2026-07-27 07:08:42 +00:00 · 2021-05-12 12:22:08 +02:00 · 2021-05-12 07:23:49 +02:00 · 2021-05-11 18:15:36 +02:00 · 2021-05-11 00:17:03 +02:00 · 2021-05-10 13:27:59 +02:00
345 changed files with 12698 additions and 4829 deletions
@@ -68,14 +68,17 @@ commands:
    steps:
      - restore_cache:
          keys:
-            - llvm-source-11-v1
+            - llvm-source-11-v2
      - run:
          name: "Fetch LLVM source"
          command: make llvm-source
      - save_cache:
-          key: llvm-source-11-v1
+          key: llvm-source-11-v2
          paths:
-            - llvm-project
+            - llvm-project/clang/lib/Headers
+            - llvm-project/clang/include
+            - llvm-project/lld/include
+            - llvm-project/llvm/include
  build-wasi-libc:
    steps:
      - restore_cache:
@@ -154,12 +157,15 @@ commands:
      - llvm-source-linux
      - restore_cache:
          keys:
-            - llvm-build-11-linux-v2-assert
+            - llvm-build-11-linux-v3-assert
      - run:
          name: "Build LLVM"
          command: |
            if [ ! -f llvm-build/lib/liblldELF.a ]
            then
+              # fetch LLVM source
+              rm -rf llvm-project
+              make llvm-source
              # install dependencies
              sudo apt-get install cmake ninja-build
              # hack ninja to use less jobs
@@ -167,9 +173,10 @@ commands:
              chmod +x /go/bin/ninja
              # build!
              make ASSERT=1 llvm-build
+              find llvm-build -name CMakeFiles -prune -exec rm -r '{}' \;
            fi
      - save_cache:
-          key: llvm-build-11-linux-v2-assert
+          key: llvm-build-11-linux-v3-assert
          paths:
            llvm-build
      - run: make ASSERT=1
@@ -177,6 +184,11 @@ commands:
      - run:
          name: "Test TinyGo"
          command: make ASSERT=1 test
+          environment:
+            # Note: -p=2 limits parallelism to two jobs at a time, which is
+            # necessary to keep memory consumption down and avoid OOM (for a
+            # 2CPU/4GB executor).
+            GOFLAGS: -p=2
      - save_cache:
          key: go-cache-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_BUILD_NUM }}
          paths:
@@ -213,12 +225,15 @@ commands:
      - llvm-source-linux
      - restore_cache:
          keys:
-            - llvm-build-11-linux-v2-noassert
+            - llvm-build-11-linux-v3-noassert
      - run:
          name: "Build LLVM"
          command: |
            if [ ! -f llvm-build/lib/liblldELF.a ]
            then
+              # fetch LLVM source
+              rm -rf llvm-project
+              make llvm-source
              # install dependencies
              sudo apt-get install cmake ninja-build
              # hack ninja to use less jobs
@@ -226,9 +241,10 @@ commands:
              chmod +x /go/bin/ninja
              # build!
              make llvm-build
+              find llvm-build -name CMakeFiles -prune -exec rm -r '{}' \;
            fi
      - save_cache:
-          key: llvm-build-11-linux-v2-noassert
+          key: llvm-build-11-linux-v3-noassert
          paths:
            llvm-build
      - build-wasi-libc
@@ -278,43 +294,50 @@ commands:
          variant: "macos"
      - restore_cache:
          keys:
-            - go-cache-macos-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_PREVIOUS_BUILD_NUM }}
-            - go-cache-macos-v2-{{ checksum "go.mod" }}
+            - go-cache-macos-v3-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_PREVIOUS_BUILD_NUM }}
+            - go-cache-macos-v3-{{ checksum "go.mod" }}
      - restore_cache:
          keys:
-            - llvm-source-11-macos-v1
+            - llvm-source-11-macos-v3
      - run:
          name: "Fetch LLVM source"
          command: make llvm-source
      - save_cache:
-          key: llvm-source-11-macos-v1
+          key: llvm-source-11-macos-v3
          paths:
-            - llvm-project
+            - llvm-project/clang/lib/Headers
+            - llvm-project/clang/include
+            - llvm-project/lld/include
+            - llvm-project/llvm/include
      - restore_cache:
          keys:
-            - llvm-build-11-macos-v2
+            - llvm-build-11-macos-v4
      - run:
          name: "Build LLVM"
          command: |
            if [ ! -f llvm-build/lib/liblldELF.a ]
            then
+              # fetch LLVM source
+              rm -rf llvm-project
+              make llvm-source
              # install dependencies
              HOMEBREW_NO_AUTO_UPDATE=1 brew install cmake ninja
              # build!
              make llvm-build
+              find llvm-build -name CMakeFiles -prune -exec rm -r '{}' \;
            fi
      - save_cache:
-          key: llvm-build-11-macos-v2
+          key: llvm-build-11-macos-v4
          paths:
            llvm-build
      - restore_cache:
          keys:
-            - wasi-libc-sysroot-macos-v3
+            - wasi-libc-sysroot-macos-v4
      - run:
          name: "Build wasi-libc"
          command: make wasi-libc
      - save_cache:
-          key: wasi-libc-sysroot-macos-v3
+          key: wasi-libc-sysroot-macos-v4
          paths:
            - lib/wasi-libc/sysroot
      - run:
@@ -336,7 +359,7 @@ commands:
            tinygo version
      - run: make smoketest AVR=0
      - save_cache:
-          key: go-cache-macos-v2-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_BUILD_NUM }}
+          key: go-cache-macos-v3-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_BUILD_NUM }}
          paths:
            - ~/.cache/go-build
            - /go/pkg/mod
@@ -378,7 +401,7 @@ jobs:
      - build-linux
  build-macos:
    macos:
-      xcode: "10.1.0"
+      xcode: "11.1.0" # macOS 10.14
    steps:
      - build-macos

@@ -19,3 +19,10 @@ src/device/kendryte/*.s
 vendor
 llvm-build
 llvm-project
+
+# Ignore files generated by smoketest
+test.gba
+test.hex
+test.nro
+test.wasm
+wasm.wasm
@@ -1,3 +1,125 @@
+0.18.0
+---
+
+* **command line**
+  - drop support for Go 1.11 and 1.12
+  - throw an error when no target is specified on Windows
+  - improve error messages in `getDefaultPort()`, support for multiple ports
+  - remove `-cflags` and `-ldflags` flags
+  - implement `-ldflags="-X ..."`
+  - add `-print-allocs` flag that lets you print all heap allocations
+  - openocd commands in tinygo command line
+  - add `-llvm-features` parameter
+  - match `go test` output
+  - discover USB ports only, this will ignore f.ex. bluetooth
+  - use physicmal path instead of cached GOROOT in function getGoroot
+  - add goroot for snap installs
+* **compiler**
+  - `builder`: add support for `-opt=0`
+  - `builder`, `compiler`: compile and cache packages in parallel
+  - `builder`: run interp per package
+  - `builder`: cache C and assembly file outputs
+  - `builder`: add support for `-x` flag to print commands
+  - `builder`: add optsize attribute while building the package
+  - `builder`: run function passes per package
+  - `builder`: hard code Clang compiler
+  - `compiler`: do not use `llvm.GlobalContext()`
+  - `compiler`: remove SimpleDCE pass
+  - `compiler`: do not emit nil checks for `*ssa.Alloc` instructions
+  - `compiler`: merge `runtime.typecodeID` and runtime.typeInInterface
+  - `compiler`: do not check for impossible type asserts
+  - `compiler`: fix use of global context: `llvm.Int32Type()`
+  - `compiler`: add interface IR test
+  - `compiler`: fix lack of method name in interface matching
+  - `compiler`: fix "fragment covers entire variable" bug
+  - `compiler`: optimize string literals and globals
+  - `compiler`: decouple func lowering from interface type codes
+  - `compiler`: add function attributes to some runtime calls
+  - `compiler`: improve position information in error messages
+  - `cgo`: add support for CFLAGS in .c files
+  - `interp`: support GEP on fixed (MMIO) addresses
+  - `interp`: handle `(reflect.Type).Elem()`
+  - `interp`: add support for runtime.interfaceMethod
+  - `interp`: make toLLVMValue return an error instead of panicking
+  - `interp`: add support for switch statement
+  - `interp`: fix phi instruction
+  - `interp`: remove map support
+  - `interp`: support extractvalue/insertvalue with multiple operands
+  - `transform`: optimize string comparisons against ""
+  - `transform`: optimize `reflect.Type` `Implements()` method
+  - `transform`: fix bug in interface lowering when signatures are renamed
+  - `transform`: don't rely on struct name of `runtime.typecodeID`
+  - `transform`: use IPSCCP pass instead of the constant propagation pass
+  - `transform`: fix func lowering assertion failure
+  - `transform`: do not lower zero-sized alloc to alloca
+  - `transform`: split interface and reflect lowering
+* **standard library**
+  - `runtime`: add dummy debug package
+  - `machine`: fix data shift/mask in newUSBSetup
+  - `machine`: make `machine.I2C0` and similar objects pointers
+  - `machine`: unify usbcdc code
+  - `machine`: refactor PWM support
+  - `machine`: avoid heap allocations in USB code
+  - `reflect`: let `reflect.Type` be of interface type
+  - `reflect`: implement a number of stub functions
+  - `reflect`: check for access in the `Interface` method call
+  - `reflect`: fix `AssignableTo` and `Implements` methods
+  - `reflect`: implement `Value.CanAddr`
+  - `reflect`: implement `Sizeof` and `Alignof` for func values
+  - `reflect`: implement `New` function
+  - `runtime`: implement command line arguments in hosted environments
+  - `runtime`: implement environment variables for Linux
+  - `runtime`: improve timers on nrf, and samd chips
+* **targets**
+  - all: use -Qunused-arguments only for assembly files
+  - `atmega1280`: add PWM support
+  - `attiny`: remove dummy UART
+  - `atsamd21`: improve SPI
+  - `atsamd51`: fix PWM support in atsamd51p20
+  - `atsamd5x`: improve SPI
+  - `atsamd51`, `atsame5x`: unify samd51 and same5x
+  - `atsamd51`, `atsamd21`: fix `ADC.Get()` value at 8bit and 10bit
+  - `atsame5x`: add support for CAN
+  - `avr`: remove I2C stubs from attiny support
+  - `cortexm`: check for `arm-none-eabi-gdb` and `gdb-multiarch` commands
+  - `cortexm`: add `__isr_vector` symbol
+  - `cortexm`: disable FPU on Cortex-M4
+  - `cortexm`: clean up Cortex-M target files
+  - `fe310`: fix SPI read
+  - `gameboy-advance`: Fix RGBA color interpretation
+  - `nrf52833`: add PWM support
+  - `stm32l0`: use unified UART logic
+  - `stm32`: move f103 (bluepill) to common i2c code
+  - `stm32`: separate altfunc selection for UART Tx/Rx
+  - `stm32`: i2c implementation for F7, L5 and L4 MCUs
+  - `stm32`: make SPI CLK fast to fix data issue
+  - `stm32`: support SPI on L4 series
+  - `unix`: avoid possible heap allocation with `-opt=0`
+  - `unix`: use conservative GC by default
+  - `unix`: use the tasks scheduler instead of coroutines
+  - `wasi`: upgrade WASI version to wasi_snapshot_preview1
+  - `wasi`: darwin: support basic file io based on libc
+  - `wasm`: only export explicitly exported functions
+  - `wasm`: use WASI ABI for exit function
+  - `wasm`: scan globals conservatively
+* **boards**
+  - `arduino-mega1280`: add support for the Arduino Mega 1280
+  - `arduino-nano-new`: Add Arduino Nano w/ New Bootloader target
+  - `atsame54-xpro`: add initial support this board
+  - `feather-m4-can`: add initial support for this board
+  - `grandcentral-m4`: add board support for Adafruit Grand Central M4 (SAMD51)
+  - `lgt92`: update to new UART structure
+  - `microbit`: remove LED constant
+  - `microbit-v2`: add support for S113 SoftDevice
+  - `nucleol432`: add support for this board
+  - `nucleo-l031k6`: add this board
+  - `pca10059`: initial support for this board
+  - `qtpy`: fix msd-volume-name
+  - `qtpy`: fix i2c setting
+  - `teensy40`: move txBuffer allocation to UART declaration
+  - `teensy40`: add UART0 as alias for UART1
+
+
 0.17.0

 ---
@@ -107,7 +107,10 @@ fmt-check:
 	@unformatted=$$(gofmt -l $(FMT_PATHS)); [ -z "$$unformatted" ] && exit 0; echo "Unformatted:"; for fn in $$unformatted; do echo "  $$fn"; done; exit 1


-gen-device: gen-device-avr gen-device-esp gen-device-nrf gen-device-sam gen-device-sifive gen-device-stm32 gen-device-kendryte gen-device-nxp
+gen-device: gen-device-avr gen-device-esp gen-device-nrf gen-device-sam gen-device-sifive gen-device-kendryte gen-device-nxp
+ifneq ($(STM32), 0)
+gen-device: gen-device-stm32
+endif

 gen-device-avr:
 	@if [ ! -e lib/avr/README.md ]; then echo "Submodules have not been downloaded. Please download them using:\n  git submodule update --init"; exit 1; fi
@@ -149,9 +152,9 @@ gen-device-stm32: build/gen-device-svd


 # Get LLVM sources.
-$(LLVM_PROJECTDIR)/README.md:
+$(LLVM_PROJECTDIR)/llvm:
 	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
+llvm-source: $(LLVM_PROJECTDIR)/llvm

 # Configure LLVM.
 TINYGO_SOURCE_DIR=$(shell pwd)
@@ -177,27 +180,30 @@ tinygo:
 	CGO_CPPFLAGS="$(CGO_CPPFLAGS)" CGO_CXXFLAGS="$(CGO_CXXFLAGS)" CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GO) build -buildmode exe -o build/tinygo$(EXE) -tags byollvm -ldflags="-X main.gitSha1=`git rev-parse --short HEAD`" .

 test: wasi-libc
-	CGO_CPPFLAGS="$(CGO_CPPFLAGS)" CGO_CXXFLAGS="$(CGO_CXXFLAGS)" CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GO) test -v -buildmode exe -tags byollvm ./cgo ./compileopts ./compiler ./interp ./transform .
+	CGO_CPPFLAGS="$(CGO_CPPFLAGS)" CGO_CXXFLAGS="$(CGO_CXXFLAGS)" CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GO) test -v -buildmode exe -tags byollvm ./builder ./cgo ./compileopts ./compiler ./interp ./transform .
+
+TEST_PACKAGES = \
+	container/heap \
+	container/list \
+	container/ring \
+	crypto/des \
+	encoding \
+	encoding/ascii85 \
+	encoding/base32 \
+	encoding/hex \
+	hash/adler32 \
+	hash/fnv \
+	hash/crc64 \
+	math \
+	math/cmplx \
+	text/scanner \
+	unicode/utf8 \

 # Test known-working standard library packages.
-# TODO: do this in one command, parallelize, and only show failing tests (no
-# implied -v flag).
+# TODO: parallelize, and only show failing tests (no implied -v flag).
 .PHONY: tinygo-test
 tinygo-test:
-	$(TINYGO) test container/heap
-	$(TINYGO) test container/list
-	$(TINYGO) test container/ring
-	$(TINYGO) test crypto/des
-	$(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
+	$(TINYGO) test $(TEST_PACKAGES)

 .PHONY: smoketest
 smoketest:
@@ -223,8 +229,6 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=microbit            examples/microbit-blink
 	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=microbit-v2         examples/microbit-blink
-	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=pca10040            examples/pininterrupt
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=pca10040            examples/serial
@@ -253,12 +257,14 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=microbit-s110v8     examples/echo
 	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=microbit-v2         examples/microbit-blink
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=microbit-v2-s113v7  examples/microbit-blink
+	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=nrf52840-mdk        examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=pca10031            examples/blinky1
 	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=bluepill            examples/blinky1
-	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=reelboard           examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=reelboard           examples/blinky2
@@ -267,6 +273,10 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=pca10056            examples/blinky2
 	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=pca10059            examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=pca10059            examples/blinky2
+	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=itsybitsy-m0        examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=feather-m0          examples/blinky1
@@ -275,14 +285,6 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=circuitplay-express examples/blinky1
 	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=stm32f4disco        examples/blinky1
-	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=stm32f4disco        examples/blinky2
-	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=stm32f4disco-1      examples/blinky1
-	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=feather-stm32f405   examples/blinky1
-	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=circuitplay-bluefruit examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=circuitplay-express examples/i2s
@@ -291,6 +293,8 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.gba -target=gameboy-advance     examples/gba-display
 	@$(MD5SUM) test.gba
+	$(TINYGO) build -size short -o test.hex -target=grandcentral-m4     examples/blinky1
+	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=itsybitsy-m4        examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=feather-m4          examples/blinky1
@@ -307,12 +311,8 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=particle-xenon      examples/blinky1
 	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=nucleo-f103rb       examples/blinky1
-	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=pinetime-devkit0    examples/blinky1
 	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=lgt92    examples/blinky1
-	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=x9pro               examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=pca10056-s140v7     examples/blinky1
@@ -339,12 +339,16 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=teensy36            examples/blinky1
 	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=nucleo-f722ze       examples/blinky1
-	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=nucleo-l552ze       examples/blinky1
-	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=p1am-100            examples/blinky1
 	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=atsame54-xpro       examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=atsame54-xpro       examples/can
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=feather-m4-can      examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=feather-m4-can      examples/caninterrupt
+	@$(MD5SUM) test.hex
 	# test pwm
 	$(TINYGO) build -size short -o test.hex -target=itsybitsy-m0        examples/pwm
 	@$(MD5SUM) test.hex
@@ -352,8 +356,30 @@ smoketest:
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=feather-m4          examples/pwm
 	@$(MD5SUM) test.hex
-	$(TINYGO) build -size short -o test.hex -target=pyportal            examples/pwm
+ifneq ($(STM32), 0)
+	$(TINYGO) build -size short -o test.hex -target=bluepill            examples/blinky1
 	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=feather-stm32f405   examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=lgt92               examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=nucleo-f103rb       examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=nucleo-f722ze       examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=nucleo-l031k6       examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=nucleo-l432kc       examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=nucleo-l552ze       examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=stm32f4disco        examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=stm32f4disco        examples/blinky2
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=stm32f4disco-1      examples/blinky1
+	@$(MD5SUM) test.hex
+endif
 ifneq ($(AVR), 0)
 	$(TINYGO) build -size short -o test.hex -target=atmega1284p         examples/serial
 	@$(MD5SUM) test.hex
@@ -363,6 +389,10 @@ ifneq ($(AVR), 0)
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=arduino -scheduler=tasks  examples/blinky1
 	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=arduino-mega1280    examples/blinky1
+	@$(MD5SUM) test.hex
+	$(TINYGO) build -size short -o test.hex -target=arduino-mega1280    examples/pwm
+	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=arduino-nano        examples/blinky1
 	@$(MD5SUM) test.hex
 	$(TINYGO) build -size short -o test.hex -target=digispark           examples/blinky1
@@ -391,6 +421,9 @@ endif
 	@$(MD5SUM) test.hex
 	$(TINYGO) build             -o test.nro -target=nintendoswitch      examples/serial
 	@$(MD5SUM) test.nro
+	$(TINYGO) build -size short -o test.hex -target=pca10040 -opt=0     ./testdata/stdlib.go
+	@$(MD5SUM) test.hex
+

 wasmtest:
 	$(GO) test ./tests/wasm
@@ -43,15 +43,17 @@ See the [getting started instructions](https://tinygo.org/getting-started/) for

 You can compile TinyGo programs for microcontrollers, WebAssembly and Linux.

-The following 55 microcontroller boards are currently supported:
+The following 62 microcontroller boards are currently supported:

 * [Adafruit Circuit Playground Bluefruit](https://www.adafruit.com/product/4333)
 * [Adafruit Circuit Playground Express](https://www.adafruit.com/product/3333)
 * [Adafruit CLUE](https://www.adafruit.com/product/4500)
 * [Adafruit Feather M0](https://www.adafruit.com/product/2772)
 * [Adafruit Feather M4](https://www.adafruit.com/product/3857)
+* [Adafruit Feather M4 CAN](https://www.adafruit.com/product/4759)
 * [Adafruit Feather nRF52840 Express](https://www.adafruit.com/product/4062)
 * [Adafruit Feather STM32F405 Express](https://www.adafruit.com/product/4382)
+* [Adafruit Grand Central M4](https://www.adafruit.com/product/4064)
 * [Adafruit ItsyBitsy M0](https://www.adafruit.com/product/3727)
 * [Adafruit ItsyBitsy M4](https://www.adafruit.com/product/3800)
 * [Adafruit ItsyBitsy nRF52840](https://www.adafruit.com/product/4481)
@@ -62,6 +64,7 @@ The following 55 microcontroller boards are currently supported:
 * [Adafruit PyPortal](https://www.adafruit.com/product/4116)
 * [Adafruit QT Py](https://www.adafruit.com/product/4600)
 * [Adafruit Trinket M0](https://www.adafruit.com/product/3500)
+* [Arduino Mega 1280](https://www.arduino.cc/en/Main/arduinoBoardMega/)
 * [Arduino Mega 2560](https://store.arduino.cc/arduino-mega-2560-rev3)
 * [Arduino MKR1000](https://store.arduino.cc/arduino-mkr1000-wifi)
 * [Arduino Nano](https://store.arduino.cc/arduino-nano)
@@ -77,11 +80,13 @@ The following 55 microcontroller boards are currently supported:
 * [Game Boy Advance](https://en.wikipedia.org/wiki/Game_Boy_Advance)
 * [Makerdiary nRF52840-MDK](https://wiki.makerdiary.com/nrf52840-mdk/)
 * [Makerdiary nRF52840-MDK USB Dongle](https://wiki.makerdiary.com/nrf52840-mdk-usb-dongle/)
+* [Microchip SAM E54 Xplained Pro](https://www.microchip.com/developmenttools/productdetails/atsame54-xpro)
 * [nice!nano](https://docs.nicekeyboards.com/#/nice!nano/)
 * [Nintendo Switch](https://www.nintendo.com/switch/)
 * [Nordic Semiconductor PCA10031](https://www.nordicsemi.com/eng/Products/nRF51-Dongle)
 * [Nordic Semiconductor PCA10040](https://www.nordicsemi.com/eng/Products/Bluetooth-low-energy/nRF52-DK)
 * [Nordic Semiconductor PCA10056](https://www.nordicsemi.com/Software-and-Tools/Development-Kits/nRF52840-DK)
+* [Nordic Semiconductor pca10059](https://www.nordicsemi.com/Software-and-tools/Development-Kits/nRF52840-Dongle)
 * [Particle Argon](https://docs.particle.io/datasheets/wi-fi/argon-datasheet/)
 * [Particle Boron](https://docs.particle.io/datasheets/cellular/boron-datasheet/)
 * [Particle Xenon](https://docs.particle.io/datasheets/discontinued/xenon-datasheet/)
@@ -96,6 +101,8 @@ The following 55 microcontroller boards are currently supported:
 * [SiFIve HiFive1](https://www.sifive.com/boards/hifive1)
 * [ST Micro "Nucleo" F103RB](https://www.st.com/en/evaluation-tools/nucleo-f103rb.html)
 * [ST Micro "Nucleo" F722ZE](https://www.st.com/en/evaluation-tools/nucleo-f722ze.html)
+* [ST Micro "Nucleo" L031K6](https://www.st.com/ja/evaluation-tools/nucleo-l031k6.html)
+* [ST Micro "Nucleo" L432KC](https://www.st.com/ja/evaluation-tools/nucleo-l432kc.html)
 * [ST Micro "Nucleo" L552ZE](https://www.st.com/en/evaluation-tools/nucleo-l552ze-q.html)
 * [ST Micro STM32F103XX "Bluepill"](https://stm32-base.org/boards/STM32F103C8T6-Blue-Pill)
 * [ST Micro STM32F407 "Discovery"](https://www.st.com/en/evaluation-tools/stm32f4discovery.html)
@@ -4,16 +4,19 @@
 package builder

 import (
+	"crypto/sha512"
 	"debug/elf"
 	"encoding/binary"
+	"encoding/hex"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"go/types"
 	"io/ioutil"
 	"os"
 	"path/filepath"
+	"runtime"
 	"sort"
-	"strconv"
 	"strings"

 	"github.com/tinygo-org/tinygo/compileopts"
@@ -36,6 +39,35 @@ type BuildResult struct {
 	// The directory of the main package. This is useful for testing as the test
 	// binary must be run in the directory of the tested package.
 	MainDir string
+
+	// ImportPath is the import path of the main package. This is useful for
+	// correctly printing test results: the import path isn't always the same as
+	// the path listed on the command line.
+	ImportPath string
+}
+
+// packageAction is the struct that is serialized to JSON and hashed, to work as
+// a cache key of compiled packages. It should contain all the information that
+// goes into a compiled package to avoid using stale data.
+//
+// Right now it's still important to include a hash of every import, because a
+// dependency might have a public constant that this package uses and thus this
+// package will need to be recompiled if that constant changes. In the future,
+// the type data should be serialized to disk which can then be used as cache
+// key, avoiding the need for recompiling all dependencies when only the
+// implementation of an imported package changes.
+type packageAction struct {
+	ImportPath       string
+	CompilerVersion  int // compiler.Version
+	InterpVersion    int // interp.Version
+	LLVMVersion      string
+	Config           *compiler.Config
+	CFlags           []string
+	FileHashes       map[string]string // hash of every file that's part of the package
+	Imports          map[string]string // map from imported package to action ID hash
+	OptLevel         int               // LLVM optimization level (0-3)
+	SizeLevel        int               // LLVM optimization for size level (0-2)
+	UndefinedGlobals []string          // globals that are left as external globals (no initializer)
 }

 // Build performs a single package to executable Go build. It takes in a package
@@ -45,6 +77,13 @@ type BuildResult struct {
 // The error value may be of type *MultiError. Callers will likely want to check
 // for this case and print such errors individually.
 func Build(pkgName, outpath string, config *compileopts.Config, action func(BuildResult) error) error {
+	// Create a temporary directory for intermediary files.
+	dir, err := ioutil.TempDir("", "tinygo")
+	if err != nil {
+		return err
+	}
+	defer os.RemoveAll(dir)
+
 	compilerConfig := &compiler.Config{
 		Triple:          config.Triple(),
 		CPU:             config.CPU(),
@@ -60,6 +99,7 @@ func Build(pkgName, outpath string, config *compileopts.Config, action func(Buil
 		DefaultStackSize:   config.Target.DefaultStackSize,
 		NeedsStackObjects:  config.NeedsStackObjects(),
 		Debug:              config.Debug(),
+		LLVMFeatures:       config.LLVMFeatures(),
 	}

 	// Load the target machine, which is the LLVM object that contains all
@@ -87,23 +127,277 @@ func Build(pkgName, outpath string, config *compileopts.Config, action func(Buil
 	// Makefile target.
 	var jobs []*compileJob

-	// Add job to compile and optimize all Go files at once.
-	// TODO: parallelize this.
+	// Create the *ssa.Program. This does not yet build the entire SSA of the
+	// program so it's pretty fast and doesn't need to be parallelized.
+	program := lprogram.LoadSSA()
+
+	// Add jobs to compile each package.
+	// Packages that have a cache hit will not be compiled again.
+	var packageJobs []*compileJob
+	packageBitcodePaths := make(map[string]string)
+	packageActionIDs := make(map[string]string)
+	optLevel, sizeLevel, _ := config.OptLevels()
+	for _, pkg := range lprogram.Sorted() {
+		pkg := pkg // necessary to avoid a race condition
+
+		var undefinedGlobals []string
+		for name := range config.Options.GlobalValues[pkg.Pkg.Path()] {
+			undefinedGlobals = append(undefinedGlobals, name)
+		}
+		sort.Strings(undefinedGlobals)
+
+		// Create a cache key: a hash from the action ID below that contains all
+		// the parameters for the build.
+		actionID := packageAction{
+			ImportPath:       pkg.ImportPath,
+			CompilerVersion:  compiler.Version,
+			InterpVersion:    interp.Version,
+			LLVMVersion:      llvm.Version,
+			Config:           compilerConfig,
+			CFlags:           pkg.CFlags,
+			FileHashes:       make(map[string]string, len(pkg.FileHashes)),
+			Imports:          make(map[string]string, len(pkg.Pkg.Imports())),
+			OptLevel:         optLevel,
+			SizeLevel:        sizeLevel,
+			UndefinedGlobals: undefinedGlobals,
+		}
+		for filePath, hash := range pkg.FileHashes {
+			actionID.FileHashes[filePath] = hex.EncodeToString(hash)
+		}
+		for _, imported := range pkg.Pkg.Imports() {
+			hash, ok := packageActionIDs[imported.Path()]
+			if !ok {
+				return fmt.Errorf("package %s imports %s but couldn't find dependency", pkg.ImportPath, imported.Path())
+			}
+			actionID.Imports[imported.Path()] = hash
+		}
+		buf, err := json.Marshal(actionID)
+		if err != nil {
+			panic(err) // shouldn't happen
+		}
+		hash := sha512.Sum512_224(buf)
+		packageActionIDs[pkg.ImportPath] = hex.EncodeToString(hash[:])
+
+		// Determine the path of the bitcode file (which is a serialized version
+		// of a LLVM module).
+		cacheDir := goenv.Get("GOCACHE")
+		if cacheDir == "off" {
+			// Use temporary build directory instead, effectively disabling the
+			// build cache.
+			cacheDir = dir
+		}
+		bitcodePath := filepath.Join(cacheDir, "pkg-"+hex.EncodeToString(hash[:])+".bc")
+		packageBitcodePaths[pkg.ImportPath] = bitcodePath
+
+		// Check whether this package has been compiled before, and if so don't
+		// compile it again.
+		if _, err := os.Stat(bitcodePath); err == nil {
+			// Already cached, don't recreate this package.
+			continue
+		}
+
+		// The package has not yet been compiled, so create a job to do so.
+		job := &compileJob{
+			description: "compile package " + pkg.ImportPath,
+			run: func(*compileJob) error {
+				// Compile AST to IR. The compiler.CompilePackage function will
+				// build the SSA as needed.
+				mod, errs := compiler.CompilePackage(pkg.ImportPath, pkg, program.Package(pkg.Pkg), machine, compilerConfig, config.DumpSSA())
+				if errs != nil {
+					return newMultiError(errs)
+				}
+				if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
+					return errors.New("verification error after compiling package " + pkg.ImportPath)
+				}
+
+				// Erase all globals that are part of the undefinedGlobals list.
+				// This list comes from the -ldflags="-X pkg.foo=val" option.
+				// Instead of setting the value directly in the AST (which would
+				// mean the value, which may be a secret, is stored in the build
+				// cache), the global itself is left external (undefined) and is
+				// only set at the end of the compilation.
+				for _, name := range undefinedGlobals {
+					globalName := pkg.Pkg.Path() + "." + name
+					global := mod.NamedGlobal(globalName)
+					if global.IsNil() {
+						return errors.New("global not found: " + globalName)
+					}
+					name := global.Name()
+					newGlobal := llvm.AddGlobal(mod, global.Type().ElementType(), name+".tmp")
+					global.ReplaceAllUsesWith(newGlobal)
+					global.EraseFromParentAsGlobal()
+					newGlobal.SetName(name)
+				}
+
+				// Try to interpret package initializers at compile time.
+				// It may only be possible to do this partially, in which case
+				// it is completed after all IR files are linked.
+				pkgInit := mod.NamedFunction(pkg.Pkg.Path() + ".init")
+				if pkgInit.IsNil() {
+					panic("init not found for " + pkg.Pkg.Path())
+				}
+				err := interp.RunFunc(pkgInit, config.DumpSSA())
+				if err != nil {
+					return err
+				}
+				if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
+					return errors.New("verification error after interpreting " + pkgInit.Name())
+				}
+
+				if sizeLevel >= 2 {
+					// Set the "optsize" attribute to make slightly smaller
+					// binaries at the cost of some performance.
+					kind := llvm.AttributeKindID("optsize")
+					attr := mod.Context().CreateEnumAttribute(kind, 0)
+					for fn := mod.FirstFunction(); !fn.IsNil(); fn = llvm.NextFunction(fn) {
+						fn.AddFunctionAttr(attr)
+					}
+				}
+
+				// Run function passes for each function in the module.
+				// These passes are intended to be run on each function right
+				// after they're created to reduce IR size (and maybe also for
+				// cache locality to improve performance), but for now they're
+				// run here for each function in turn. Maybe this can be
+				// improved in the future.
+				builder := llvm.NewPassManagerBuilder()
+				defer builder.Dispose()
+				builder.SetOptLevel(optLevel)
+				builder.SetSizeLevel(sizeLevel)
+				funcPasses := llvm.NewFunctionPassManagerForModule(mod)
+				defer funcPasses.Dispose()
+				builder.PopulateFunc(funcPasses)
+				funcPasses.InitializeFunc()
+				for fn := mod.FirstFunction(); !fn.IsNil(); fn = llvm.NextFunction(fn) {
+					if fn.IsDeclaration() {
+						continue
+					}
+					funcPasses.RunFunc(fn)
+				}
+				funcPasses.FinalizeFunc()
+
+				// Serialize the LLVM module as a bitcode file.
+				// Write to a temporary path that is renamed to the destination
+				// file to avoid race conditions with other TinyGo invocatiosn
+				// that might also be compiling this package at the same time.
+				f, err := ioutil.TempFile(filepath.Dir(bitcodePath), filepath.Base(bitcodePath))
+				if err != nil {
+					return err
+				}
+				if runtime.GOOS == "windows" {
+					// Work around a problem on Windows.
+					// For some reason, WriteBitcodeToFile causes TinyGo to
+					// exit with the following message:
+					//   LLVM ERROR: IO failure on output stream: Bad file descriptor
+					buf := llvm.WriteBitcodeToMemoryBuffer(mod)
+					defer buf.Dispose()
+					_, err = f.Write(buf.Bytes())
+				} else {
+					// Otherwise, write bitcode directly to the file (probably
+					// faster).
+					err = llvm.WriteBitcodeToFile(mod, f)
+				}
+				if err != nil {
+					// WriteBitcodeToFile doesn't produce a useful error on its
+					// own, so create a somewhat useful error message here.
+					return fmt.Errorf("failed to write bitcode for package %s to file %s", pkg.ImportPath, bitcodePath)
+				}
+				err = f.Close()
+				if err != nil {
+					return err
+				}
+				return os.Rename(f.Name(), bitcodePath)
+			},
+		}
+		jobs = append(jobs, job)
+		packageJobs = append(packageJobs, job)
+	}
+
+	// Add job that links and optimizes all packages together.
 	var mod llvm.Module
 	var stackSizeLoads []string
 	programJob := &compileJob{
-		description: "compile Go files",
-		run: func() (err error) {
-			mod, err = compileWholeProgram(pkgName, config, compilerConfig, lprogram, machine)
-			if err != nil {
-				return
+		description:  "link+optimize packages (LTO)",
+		dependencies: packageJobs,
+		run: func(*compileJob) error {
+			// Load and link all the bitcode files. This does not yet optimize
+			// anything, it only links the bitcode files together.
+			ctx := llvm.NewContext()
+			mod = ctx.NewModule("")
+			for _, pkg := range lprogram.Sorted() {
+				pkgMod, err := ctx.ParseBitcodeFile(packageBitcodePaths[pkg.ImportPath])
+				if err != nil {
+					return fmt.Errorf("failed to load bitcode file: %w", err)
+				}
+				err = llvm.LinkModules(mod, pkgMod)
+				if err != nil {
+					return fmt.Errorf("failed to link module: %w", err)
+				}
 			}
+
+			// Create runtime.initAll function that calls the runtime
+			// initializer of each package.
+			llvmInitFn := mod.NamedFunction("runtime.initAll")
+			llvmInitFn.SetLinkage(llvm.InternalLinkage)
+			llvmInitFn.SetUnnamedAddr(true)
+			llvmInitFn.Param(0).SetName("context")
+			llvmInitFn.Param(1).SetName("parentHandle")
+			block := mod.Context().AddBasicBlock(llvmInitFn, "entry")
+			irbuilder := mod.Context().NewBuilder()
+			defer irbuilder.Dispose()
+			irbuilder.SetInsertPointAtEnd(block)
+			i8ptrType := llvm.PointerType(mod.Context().Int8Type(), 0)
+			for _, pkg := range lprogram.Sorted() {
+				pkgInit := mod.NamedFunction(pkg.Pkg.Path() + ".init")
+				if pkgInit.IsNil() {
+					panic("init not found for " + pkg.Pkg.Path())
+				}
+				irbuilder.CreateCall(pkgInit, []llvm.Value{llvm.Undef(i8ptrType), llvm.Undef(i8ptrType)}, "")
+			}
+			irbuilder.CreateRetVoid()
+
+			// After linking, functions should (as far as possible) be set to
+			// private linkage or internal linkage. The compiler package marks
+			// non-exported functions by setting the visibility to hidden or
+			// (for thunks) to linkonce_odr linkage. Change the linkage here to
+			// internal to benefit much more from interprocedural optimizations.
+			for fn := mod.FirstFunction(); !fn.IsNil(); fn = llvm.NextFunction(fn) {
+				if fn.Visibility() == llvm.HiddenVisibility {
+					fn.SetVisibility(llvm.DefaultVisibility)
+					fn.SetLinkage(llvm.InternalLinkage)
+				} else if fn.Linkage() == llvm.LinkOnceODRLinkage {
+					fn.SetLinkage(llvm.InternalLinkage)
+				}
+			}
+
+			// Do the same for globals.
+			for global := mod.FirstGlobal(); !global.IsNil(); global = llvm.NextGlobal(global) {
+				if global.Visibility() == llvm.HiddenVisibility {
+					global.SetVisibility(llvm.DefaultVisibility)
+					global.SetLinkage(llvm.InternalLinkage)
+				} else if global.Linkage() == llvm.LinkOnceODRLinkage {
+					global.SetLinkage(llvm.InternalLinkage)
+				}
+			}
+
+			if config.Options.PrintIR {
+				fmt.Println("; Generated LLVM IR:")
+				fmt.Println(mod.String())
+			}
+
+			// Run all optimization passes, which are much more effective now
+			// that the optimizer can see the whole program at once.
+			err := optimizeProgram(mod, config)
+			if err != nil {
+				return err
+			}
+
 			// Make sure stack sizes are loaded from a separate section so they can be
 			// modified after linking.
 			if config.AutomaticStackSize() {
 				stackSizeLoads = transform.CreateStackSizeLoads(mod, config)
 			}
-			return
+			return nil
 		},
 	}
 	jobs = append(jobs, programJob)
@@ -140,19 +434,13 @@ func Build(pkgName, outpath string, config *compileopts.Config, action func(Buil
 	// First add all jobs necessary to build this object file, then afterwards
 	// run all jobs in parallel as far as possible.

-	// Create a temporary directory for intermediary files.
-	dir, err := ioutil.TempDir("", "tinygo")
-	if err != nil {
-		return err
-	}
-	defer os.RemoveAll(dir)
-
 	// Add job to write the output object file.
 	objfile := filepath.Join(dir, "main.o")
 	outputObjectFileJob := &compileJob{
 		description:  "generate output file",
 		dependencies: []*compileJob{programJob},
-		run: func() error {
+		result:       objfile,
+		run: func(*compileJob) error {
 			llvmBuf, err := machine.EmitToMemoryBuffer(mod, llvm.ObjectFile)
 			if err != nil {
 				return err
@@ -166,40 +454,32 @@ func Build(pkgName, outpath string, config *compileopts.Config, action func(Buil
 	linkerDependencies := []*compileJob{outputObjectFileJob}
 	executable := filepath.Join(dir, "main")
 	tmppath := executable // final file
-	ldflags := append(config.LDFlags(), "-o", executable, objfile)
+	ldflags := append(config.LDFlags(), "-o", executable)

 	// Add compiler-rt dependency if needed. Usually this is a simple load from
 	// a cache.
 	if config.Target.RTLib == "compiler-rt" {
-		path, job, err := CompilerRT.load(config.Triple(), config.CPU(), dir)
+		job, err := CompilerRT.load(config.Triple(), config.CPU(), dir)
 		if err != nil {
 			return err
 		}
-		if job != nil {
-			// The library was not loaded from cache so needs to be compiled
-			// (and then stored in the cache).
-			jobs = append(jobs, job.dependencies...)
-			jobs = append(jobs, job)
-			linkerDependencies = append(linkerDependencies, job)
-		}
-		ldflags = append(ldflags, path)
+		jobs = append(jobs, job.dependencies...)
+		jobs = append(jobs, job)
+		linkerDependencies = append(linkerDependencies, job)
 	}

 	// Add libc dependency if needed.
 	root := goenv.Get("TINYGOROOT")
 	switch config.Target.Libc {
 	case "picolibc":
-		path, job, err := Picolibc.load(config.Triple(), config.CPU(), dir)
+		job, err := Picolibc.load(config.Triple(), config.CPU(), dir)
 		if err != nil {
 			return err
 		}
-		if job != nil {
-			// The library needs to be compiled (cache miss).
-			jobs = append(jobs, job.dependencies...)
-			jobs = append(jobs, job)
-			linkerDependencies = append(linkerDependencies, job)
-		}
-		ldflags = append(ldflags, path)
+		// The library needs to be compiled (cache miss).
+		jobs = append(jobs, job.dependencies...)
+		jobs = append(jobs, job)
+		linkerDependencies = append(linkerDependencies, job)
 	case "wasi-libc":
 		path := filepath.Join(root, "lib/wasi-libc/sysroot/lib/wasm32-wasi/libc.a")
 		if _, err := os.Stat(path); os.IsNotExist(err) {
@@ -215,44 +495,37 @@ func Build(pkgName, outpath string, config *compileopts.Config, action func(Buil
 	// Add jobs to compile extra files. These files are in C or assembly and
 	// contain things like the interrupt vector table and low level operations
 	// such as stack switching.
-	for i, path := range config.ExtraFiles() {
+	for _, path := range config.ExtraFiles() {
 		abspath := filepath.Join(root, path)
-		outpath := filepath.Join(dir, "extra-"+strconv.Itoa(i)+"-"+filepath.Base(path)+".o")
 		job := &compileJob{
 			description: "compile extra file " + path,
-			run: func() error {
-				err := runCCompiler(config.Target.Compiler, append(config.CFlags(), "-c", "-o", outpath, abspath)...)
-				if err != nil {
-					return &commandError{"failed to build", path, err}
-				}
-				return nil
+			run: func(job *compileJob) error {
+				result, err := compileAndCacheCFile(abspath, dir, config.CFlags(), config.Options.PrintCommands)
+				job.result = result
+				return err
 			},
 		}
 		jobs = append(jobs, job)
 		linkerDependencies = append(linkerDependencies, job)
-		ldflags = append(ldflags, outpath)
 	}

 	// Add jobs to compile C files in all packages. This is part of CGo.
 	// TODO: do this as part of building the package to be able to link the
 	// bitcode files together.
-	for i, pkg := range lprogram.Sorted() {
-		for j, filename := range pkg.CFiles {
-			file := filepath.Join(pkg.Dir, filename)
-			outpath := filepath.Join(dir, "pkg"+strconv.Itoa(i)+"."+strconv.Itoa(j)+"-"+filepath.Base(file)+".o")
+	for _, pkg := range lprogram.Sorted() {
+		pkg := pkg
+		for _, filename := range pkg.CFiles {
+			abspath := filepath.Join(pkg.Dir, filename)
 			job := &compileJob{
-				description: "compile CGo file " + file,
-				run: func() error {
-					err := runCCompiler(config.Target.Compiler, append(config.CFlags(), "-c", "-o", outpath, file)...)
-					if err != nil {
-						return &commandError{"failed to build", file, err}
-					}
-					return nil
+				description: "compile CGo file " + abspath,
+				run: func(job *compileJob) error {
+					result, err := compileAndCacheCFile(abspath, dir, pkg.CFlags, config.Options.PrintCommands)
+					job.result = result
+					return err
 				},
 			}
 			jobs = append(jobs, job)
 			linkerDependencies = append(linkerDependencies, job)
-			ldflags = append(ldflags, outpath)
 		}
 	}

@@ -267,7 +540,16 @@ func Build(pkgName, outpath string, config *compileopts.Config, action func(Buil
 	jobs = append(jobs, &compileJob{
 		description:  "link",
 		dependencies: linkerDependencies,
-		run: func() error {
+		run: func(job *compileJob) error {
+			for _, dependency := range job.dependencies {
+				if dependency.result == "" {
+					return errors.New("dependency without result: " + dependency.description)
+				}
+				ldflags = append(ldflags, dependency.result)
+			}
+			if config.Options.PrintCommands {
+				fmt.Printf("%s %s\n", config.Target.Linker, strings.Join(ldflags, " "))
+			}
 			err = link(config.Target.Linker, ldflags...)
 			if err != nil {
 				return &commandError{"failed to link", executable, err}
@@ -361,40 +643,43 @@ func Build(pkgName, outpath string, config *compileopts.Config, action func(Buil
 		return fmt.Errorf("unknown output binary format: %s", outputBinaryFormat)
 	}
 	return action(BuildResult{
-		Binary:  tmppath,
-		MainDir: lprogram.MainPkg().Dir,
+		Binary:     tmppath,
+		MainDir:    lprogram.MainPkg().Dir,
+		ImportPath: lprogram.MainPkg().ImportPath,
 	})
 }

-// compileWholeProgram compiles the entire *loader.Program to a LLVM module and
-// applies most necessary optimizations and transformations.
-func compileWholeProgram(pkgName string, config *compileopts.Config, compilerConfig *compiler.Config, lprogram *loader.Program, machine llvm.TargetMachine) (llvm.Module, error) {
-	// Compile AST to IR.
-	mod, errs := compiler.CompileProgram(lprogram, machine, compilerConfig, config.DumpSSA())
-	if errs != nil {
-		return mod, newMultiError(errs)
-	}
-
-	if config.Options.PrintIR {
-		fmt.Println("; Generated LLVM IR:")
-		fmt.Println(mod.String())
-	}
-	if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
-		return mod, errors.New("verification error after IR construction")
-	}
-
+// optimizeProgram runs a series of optimizations and transformations that are
+// needed to convert a program to its final form. Some transformations are not
+// optional and must be run as the compiler expects them to run.
+func optimizeProgram(mod llvm.Module, config *compileopts.Config) error {
 	err := interp.Run(mod, config.DumpSSA())
 	if err != nil {
-		return mod, err
+		return err
 	}
-	if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
-		return mod, errors.New("verification error after interpreting runtime.initAll")
+	if config.VerifyIR() {
+		// Only verify if we really need it.
+		// The IR has already been verified before writing the bitcode to disk
+		// and the interp function above doesn't need to do a lot as most of the
+		// package initializers have already run. Additionally, verifying this
+		// linked IR is _expensive_ because dead code hasn't been removed yet,
+		// easily costing a few hundred milliseconds. Therefore, only do it when
+		// specifically requested.
+		if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
+			return errors.New("verification error after interpreting runtime.initAll")
+		}
 	}

 	if config.GOOS() != "darwin" {
 		transform.ApplyFunctionSections(mod) // -ffunction-sections
 	}

+	// Insert values from -ldflags="-X ..." into the IR.
+	err = setGlobalValues(mod, config.Options.GlobalValues)
+	if err != nil {
+		return err
+	}
+
 	// Browsers cannot handle external functions that have type i64 because it
 	// cannot be represented exactly in JavaScript (JS only has doubles). To
 	// keep functions interoperable, pass int64 types as pointers to
@@ -403,39 +688,19 @@ func compileWholeProgram(pkgName string, config *compileopts.Config, compilerCon
 	if config.WasmAbi() == "js" {
 		err := transform.ExternalInt64AsPtr(mod)
 		if err != nil {
-			return mod, err
+			return err
 		}
 	}

 	// Optimization levels here are roughly the same as Clang, but probably not
 	// exactly.
-	errs = nil
-	switch config.Options.Opt {
-	/*
-		Currently, turning optimizations off causes compile failures.
-		We rely on the optimizer removing some dead symbols.
-		Avoid providing an option that does not work right now.
-		In the future once everything has been fixed we can re-enable this.
-
-		case "none", "0":
-			errs = transform.Optimize(mod, config, 0, 0, 0) // -O0
-	*/
-	case "1":
-		errs = transform.Optimize(mod, config, 1, 0, 0) // -O1
-	case "2":
-		errs = transform.Optimize(mod, config, 2, 0, 225) // -O2
-	case "s":
-		errs = transform.Optimize(mod, config, 2, 1, 225) // -Os
-	case "z":
-		errs = transform.Optimize(mod, config, 2, 2, 5) // -Oz, default
-	default:
-		errs = []error{errors.New("unknown optimization level: -opt=" + config.Options.Opt)}
-	}
+	optLevel, sizeLevel, inlinerThreshold := config.OptLevels()
+	errs := transform.Optimize(mod, config, optLevel, sizeLevel, inlinerThreshold)
 	if len(errs) > 0 {
-		return mod, newMultiError(errs)
+		return newMultiError(errs)
 	}
 	if err := llvm.VerifyModule(mod, llvm.PrintMessageAction); err != nil {
-		return mod, errors.New("verification failure after LLVM optimization passes")
+		return errors.New("verification failure after LLVM optimization passes")
 	}

 	// LLVM 11 by default tries to emit tail calls (even with the target feature
@@ -449,7 +714,72 @@ func compileWholeProgram(pkgName string, config *compileopts.Config, compilerCon
 		transform.DisableTailCalls(mod)
 	}

-	return mod, nil
+	return nil
+}
+
+// setGlobalValues sets the global values from the -ldflags="-X ..." compiler
+// option in the given module. An error may be returned if the global is not of
+// the expected type.
+func setGlobalValues(mod llvm.Module, globals map[string]map[string]string) error {
+	var pkgPaths []string
+	for pkgPath := range globals {
+		pkgPaths = append(pkgPaths, pkgPath)
+	}
+	sort.Strings(pkgPaths)
+	for _, pkgPath := range pkgPaths {
+		pkg := globals[pkgPath]
+		var names []string
+		for name := range pkg {
+			names = append(names, name)
+		}
+		sort.Strings(names)
+		for _, name := range names {
+			value := pkg[name]
+			globalName := pkgPath + "." + name
+			global := mod.NamedGlobal(globalName)
+			if global.IsNil() || !global.Initializer().IsNil() {
+				// The global either does not exist (optimized away?) or has
+				// some value, in which case it has already been initialized at
+				// package init time.
+				continue
+			}
+
+			// A strin is a {ptr, len} pair. We need these types to build the
+			// initializer.
+			initializerType := global.Type().ElementType()
+			if initializerType.TypeKind() != llvm.StructTypeKind || initializerType.StructName() == "" {
+				return fmt.Errorf("%s: not a string", globalName)
+			}
+			elementTypes := initializerType.StructElementTypes()
+			if len(elementTypes) != 2 {
+				return fmt.Errorf("%s: not a string", globalName)
+			}
+
+			// Create a buffer for the string contents.
+			bufInitializer := mod.Context().ConstString(value, false)
+			buf := llvm.AddGlobal(mod, bufInitializer.Type(), ".string")
+			buf.SetInitializer(bufInitializer)
+			buf.SetAlignment(1)
+			buf.SetUnnamedAddr(true)
+			buf.SetLinkage(llvm.PrivateLinkage)
+
+			// Create the string value, which is a {ptr, len} pair.
+			zero := llvm.ConstInt(mod.Context().Int32Type(), 0, false)
+			ptr := llvm.ConstGEP(buf, []llvm.Value{zero, zero})
+			if ptr.Type() != elementTypes[0] {
+				return fmt.Errorf("%s: not a string", globalName)
+			}
+			length := llvm.ConstInt(elementTypes[1], uint64(len(value)), false)
+			initializer := llvm.ConstNamedStruct(initializerType, []llvm.Value{
+				ptr,
+				length,
+			})
+
+			// Set the initializer. No initializer should be set at this point.
+			global.SetInitializer(initializer)
+		}
+	}
+	return nil
 }

 // functionStackSizes keeps stack size information about a single function
@@ -0,0 +1,309 @@
+package builder
+
+// This file implements a wrapper around the C compiler (Clang) which uses a
+// build cache.
+
+import (
+	"crypto/sha512"
+	"encoding/hex"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"io/ioutil"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"unicode"
+
+	"github.com/tinygo-org/tinygo/goenv"
+	"tinygo.org/x/go-llvm"
+)
+
+// compileAndCacheCFile compiles a C or assembly file using a build cache.
+// Compiling the same file again (if nothing changed, including included header
+// files) the output is loaded from the build cache instead.
+//
+// Its operation is a bit complex (more complex than Go package build caching)
+// because the list of file dependencies is only known after the file is
+// compiled. However, luckily compilers have a flag to write a list of file
+// dependencies in Makefile syntax which can be used for caching.
+//
+// Because of this complexity, every file has in fact two cached build outputs:
+// the file itself, and the list of dependencies. Its operation is as follows:
+//
+//   depfile = hash(path, compiler, cflags, ...)
+//   if depfile exists:
+//     outfile = hash of all files and depfile name
+//     if outfile exists:
+//       # cache hit
+//       return outfile
+//   # cache miss
+//   tmpfile = compile file
+//   read dependencies (side effect of compile)
+//   write depfile
+//   outfile = hash of all files and depfile name
+//   rename tmpfile to outfile
+//
+// There are a few edge cases that are not handled:
+// - If a file is added to an include path, that file may be included instead of
+//   some other file. This would be fixed by also including lookup failures in the
+//   dependencies file, but I'm not aware of a compiler which does that.
+// - The Makefile syntax that compilers output has issues, see readDepFile for
+//   details.
+// - A header file may be changed to add/remove an include. This invalidates the
+//   depfile but without invalidating its name. For this reason, the depfile is
+//   written on each new compilation (even when it seems unnecessary). However, it
+//   could in rare cases lead to a stale file fetched from the cache.
+func compileAndCacheCFile(abspath, tmpdir string, cflags []string, printCommands bool) (string, error) {
+	// Hash input file.
+	fileHash, err := hashFile(abspath)
+	if err != nil {
+		return "", err
+	}
+
+	// Create cache key for the dependencies file.
+	buf, err := json.Marshal(struct {
+		Path        string
+		Hash        string
+		Flags       []string
+		LLVMVersion string
+	}{
+		Path:        abspath,
+		Hash:        fileHash,
+		Flags:       cflags,
+		LLVMVersion: llvm.Version,
+	})
+	if err != nil {
+		panic(err) // shouldn't happen
+	}
+	depfileNameHashBuf := sha512.Sum512_224(buf)
+	depfileNameHash := hex.EncodeToString(depfileNameHashBuf[:])
+
+	// Load dependencies file, if possible.
+	depfileName := "dep-" + depfileNameHash + ".json"
+	depfileCachePath := filepath.Join(goenv.Get("GOCACHE"), depfileName)
+	depfileBuf, err := ioutil.ReadFile(depfileCachePath)
+	var dependencies []string // sorted list of dependency paths
+	if err == nil {
+		// There is a dependency file, that's great!
+		// Parse it first.
+		err := json.Unmarshal(depfileBuf, &dependencies)
+		if err != nil {
+			return "", fmt.Errorf("could not parse dependencies JSON: %w", err)
+		}
+
+		// Obtain hashes of all the files listed as a dependency.
+		outpath, err := makeCFileCachePath(dependencies, depfileNameHash)
+		if err == nil {
+			if _, err := os.Stat(outpath); err == nil {
+				return outpath, nil
+			} else if !os.IsNotExist(err) {
+				return "", err
+			}
+		}
+	} else if !os.IsNotExist(err) {
+		// expected either nil or IsNotExist
+		return "", err
+	}
+
+	objTmpFile, err := ioutil.TempFile(goenv.Get("GOCACHE"), "tmp-*.o")
+	if err != nil {
+		return "", err
+	}
+	objTmpFile.Close()
+	depTmpFile, err := ioutil.TempFile(tmpdir, "dep-*.d")
+	if err != nil {
+		return "", err
+	}
+	depTmpFile.Close()
+	flags := append([]string{}, cflags...)                                   // copy cflags
+	flags = append(flags, "-MD", "-MV", "-MTdeps", "-MF", depTmpFile.Name()) // autogenerate dependencies
+	flags = append(flags, "-c", "-o", objTmpFile.Name(), abspath)
+	if strings.ToLower(filepath.Ext(abspath)) == ".s" {
+		// If this is an assembly file (.s or .S, lowercase or uppercase), then
+		// we'll need to add -Qunused-arguments because many parameters are
+		// relevant to C, not assembly. And with -Werror, having meaningless
+		// flags (for the assembler) is a compiler error.
+		flags = append(flags, "-Qunused-arguments")
+	}
+	if printCommands {
+		fmt.Printf("clang %s\n", strings.Join(flags, " "))
+	}
+	err = runCCompiler(flags...)
+	if err != nil {
+		return "", &commandError{"failed to build", abspath, err}
+	}
+
+	// Create sorted and uniqued slice of dependencies.
+	dependencyPaths, err := readDepFile(depTmpFile.Name())
+	if err != nil {
+		return "", err
+	}
+	dependencyPaths = append(dependencyPaths, abspath) // necessary for .s files
+	dependencySet := make(map[string]struct{}, len(dependencyPaths))
+	var dependencySlice []string
+	for _, path := range dependencyPaths {
+		if _, ok := dependencySet[path]; ok {
+			continue
+		}
+		dependencySet[path] = struct{}{}
+		dependencySlice = append(dependencySlice, path)
+	}
+	sort.Strings(dependencySlice)
+
+	// Write dependencies file.
+	f, err := ioutil.TempFile(filepath.Dir(depfileCachePath), depfileName)
+	buf, err = json.MarshalIndent(dependencySlice, "", "\t")
+	if err != nil {
+		panic(err) // shouldn't happen
+	}
+	_, err = f.Write(buf)
+	if err != nil {
+		return "", err
+	}
+	err = f.Close()
+	if err != nil {
+		return "", err
+	}
+	err = os.Rename(f.Name(), depfileCachePath)
+	if err != nil {
+		return "", err
+	}
+
+	// Move temporary object file to final location.
+	outpath, err := makeCFileCachePath(dependencySlice, depfileNameHash)
+	if err != nil {
+		return "", err
+	}
+	err = os.Rename(objTmpFile.Name(), outpath)
+	if err != nil {
+		return "", err
+	}
+
+	return outpath, nil
+}
+
+// Create a cache path (a path in GOCACHE) to store the output of a compiler
+// job. This path is based on the dep file name (which is a hash of metadata
+// including compiler flags) and the hash of all input files in the paths slice.
+func makeCFileCachePath(paths []string, depfileNameHash string) (string, error) {
+	// Hash all input files.
+	fileHashes := make(map[string]string, len(paths))
+	for _, path := range paths {
+		hash, err := hashFile(path)
+		if err != nil {
+			return "", err
+		}
+		fileHashes[path] = hash
+	}
+
+	// Calculate a cache key based on the above hashes.
+	buf, err := json.Marshal(struct {
+		DepfileHash string
+		FileHashes  map[string]string
+	}{
+		DepfileHash: depfileNameHash,
+		FileHashes:  fileHashes,
+	})
+	if err != nil {
+		panic(err) // shouldn't happen
+	}
+	outFileNameBuf := sha512.Sum512_224(buf)
+	cacheKey := hex.EncodeToString(outFileNameBuf[:])
+
+	outpath := filepath.Join(goenv.Get("GOCACHE"), "obj-"+cacheKey+".o")
+	return outpath, nil
+}
+
+// hashFile hashes the given file path and returns the hash as a hex string.
+func hashFile(path string) (string, error) {
+	f, err := os.Open(path)
+	if err != nil {
+		return "", fmt.Errorf("failed to hash file: %w", err)
+	}
+	defer f.Close()
+	fileHasher := sha512.New512_224()
+	_, err = io.Copy(fileHasher, f)
+	if err != nil {
+		return "", fmt.Errorf("failed to hash file: %w", err)
+	}
+	return hex.EncodeToString(fileHasher.Sum(nil)), nil
+}
+
+// readDepFile reads a dependency file in NMake (Visual Studio make) format. The
+// file is assumed to have a single target named deps.
+//
+// There are roughly three make syntax variants:
+// - BSD make, which doesn't support any escaping. This means that many special
+//   characters are not supported in file names.
+// - GNU make, which supports escaping using a backslash but when it fails to
+//   find a file it tries to fall back with the literal path name (to match BSD
+//   make).
+// - NMake (Visual Studio) and Jom, which simply quote the string if there are
+//   any weird characters.
+// Clang supports two variants: a format that's a compromise between BSD and GNU
+// make (and is buggy to match GCC which is equally buggy), and NMake/Jom, which
+// is at least somewhat sane. This last format isn't perfect either: it does not
+// correctly handle filenames with quote marks in them. Those are generally not
+// allowed on Windows, but of course can be used on POSIX like systems. Still,
+// it's the most sane of any of the formats so readDepFile will use that format.
+func readDepFile(filename string) ([]string, error) {
+	buf, err := ioutil.ReadFile(filename)
+	if err != nil {
+		return nil, err
+	}
+	if len(buf) == 0 {
+		return nil, nil
+	}
+	return parseDepFile(string(buf))
+}
+
+func parseDepFile(s string) ([]string, error) {
+	// This function makes no attempt at parsing anything other than Clang -MD
+	// -MV output.
+
+	// For Windows: replace CRLF with LF to make the logic below simpler.
+	s = strings.ReplaceAll(s, "\r\n", "\n")
+
+	// Collapse all lines ending in a backslash. These backslashes are really
+	// just a way to continue a line without making very long lines.
+	s = strings.ReplaceAll(s, "\\\n", " ")
+
+	// Only use the first line, which is expected to begin with "deps:".
+	line := strings.SplitN(s, "\n", 2)[0]
+	if !strings.HasPrefix(line, "deps:") {
+		return nil, errors.New("readDepFile: expected 'deps:' prefix")
+	}
+	line = strings.TrimSpace(line[len("deps:"):])
+
+	var deps []string
+	for line != "" {
+		if line[0] == '"' {
+			// File path is quoted. Path ends with double quote.
+			// This does not handle double quotes in path names, which is a
+			// problem on non-Windows systems.
+			line = line[1:]
+			end := strings.IndexByte(line, '"')
+			if end < 0 {
+				return nil, errors.New("readDepFile: path is incorrectly quoted")
+			}
+			dep := line[:end]
+			line = strings.TrimSpace(line[end+1:])
+			deps = append(deps, dep)
+		} else {
+			// File path is not quoted. Path ends in space or EOL.
+			end := strings.IndexFunc(line, unicode.IsSpace)
+			if end < 0 {
+				// last dependency
+				deps = append(deps, line)
+				break
+			}
+			dep := line[:end]
+			line = strings.TrimSpace(line[end:])
+			deps = append(deps, dep)
+		}
+	}
+	return deps, nil
+}
@@ -0,0 +1,33 @@
+package builder
+
+import (
+	"reflect"
+	"testing"
+)
+
+func TestSplitDepFile(t *testing.T) {
+	for i, tc := range []struct {
+		in  string
+		out []string
+	}{
+		{`deps: foo bar`, []string{"foo", "bar"}},
+		{`deps: foo "bar"`, []string{"foo", "bar"}},
+		{`deps: "foo" bar`, []string{"foo", "bar"}},
+		{`deps: "foo bar"`, []string{"foo bar"}},
+		{`deps: "foo bar" `, []string{"foo bar"}},
+		{"deps: foo\nbar", []string{"foo"}},
+		{"deps: foo \\\nbar", []string{"foo", "bar"}},
+		{"deps: foo\\bar \\\nbaz", []string{"foo\\bar", "baz"}},
+		{"deps: foo\\bar \\\r\n baz", []string{"foo\\bar", "baz"}}, // Windows uses CRLF line endings
+	} {
+		out, err := parseDepFile(tc.in)
+		if err != nil {
+			t.Errorf("test #%d failed: %v", i, err)
+			continue
+		}
+		if !reflect.DeepEqual(out, tc.out) {
+			t.Errorf("test #%d failed: expected %#v but got %#v", i, tc.out, out)
+			continue
+		}
+	}
+}
@@ -17,10 +17,18 @@ func NewConfig(options *compileopts.Options) (*compileopts.Config, error) {
 	if err != nil {
 		return nil, err
 	}
+
+	if options.OpenOCDCommands != nil {
+		// Override the OpenOCDCommands from the target spec if specified on
+		// the command-line
+		spec.OpenOCDCommands = options.OpenOCDCommands
+	}
+
 	goroot := goenv.Get("GOROOT")
 	if goroot == "" {
 		return nil, errors.New("cannot locate $GOROOT, please set it manually")
 	}
+
 	major, minor, err := goenv.GetGorootVersion(goroot)
 	if err != nil {
 		return nil, fmt.Errorf("could not read version from GOROOT (%v): %v", goroot, err)
@@ -28,7 +36,9 @@ func NewConfig(options *compileopts.Options) (*compileopts.Config, error) {
 	if major != 1 || minor < 13 || minor > 16 {
 		return nil, fmt.Errorf("requires go version 1.13 through 1.16, got go%d.%d", major, minor)
 	}
+
 	clangHeaderPath := getClangHeaderPath(goenv.Get("TINYGOROOT"))
+
 	return &compileopts.Config{
 		Options:        options,
 		Target:         spec,
@@ -27,12 +27,23 @@ const (
 type compileJob struct {
 	description  string // description, only used for logging
 	dependencies []*compileJob
-	run          func() error
+	result       string // result (path)
+	run          func(*compileJob) (err error)
 	state        jobState
 	err          error         // error if finished
 	duration     time.Duration // how long it took to run this job (only set after finishing)
 }

+// dummyCompileJob returns a new *compileJob that produces an output without
+// doing anything. This can be useful where a *compileJob producing an output is
+// expected but nothing needs to be done, for example for a load from a cache.
+func dummyCompileJob(result string) *compileJob {
+	return &compileJob{
+		description: "<dummy>",
+		result:      result,
+	}
+}
+
 // readyToRun returns whether this job is ready to run: it is itself not yet
 // started and all dependencies are finished.
 func (job *compileJob) readyToRun() bool {
@@ -150,9 +161,11 @@ func nextJob(jobs []*compileJob) *compileJob {
 func jobWorker(workerChan, doneChan chan *compileJob) {
 	for job := range workerChan {
 		start := time.Now()
-		err := job.run()
-		if err != nil {
-			job.err = err
+		if job.run != nil {
+			err := job.run(job)
+			if err != nil {
+				job.err = err
+			}
 		}
 		job.duration = time.Since(start)
 		doneChan <- job
@@ -42,30 +42,28 @@ func (l *Library) sourcePaths(target string) []string {
 // The resulting file is stored in the provided tmpdir, which is expected to be
 // removed after the Load call.
 func (l *Library) Load(target, tmpdir string) (path string, err error) {
-	path, job, err := l.load(target, "", tmpdir)
+	job, err := l.load(target, "", tmpdir)
 	if err != nil {
 		return "", err
 	}
-	if job != nil {
-		jobs := append([]*compileJob{job}, job.dependencies...)
-		err = runJobs(jobs)
-	}
-	return path, err
+	jobs := append([]*compileJob{job}, job.dependencies...)
+	err = runJobs(jobs)
+	return job.result, err
 }

-// load returns a path to the library file for the given target, loading it from
-// cache if possible. It will return a non-zero compiler job if the library
-// wasn't cached, this job (and its dependencies) must be run before the library
-// path is valid.
+// load returns a compile job to build this library file for the given target
+// and CPU. It may return a dummy compileJob if the library build is already
+// cached. The path is stored as job.result but is only valid if the job and
+// job.dependencies have been run.
 // The provided tmpdir will be used to store intermediary files and possibly the
 // output archive file, it is expected to be removed after use.
-func (l *Library) load(target, cpu, tmpdir string) (path string, job *compileJob, err error) {
+func (l *Library) load(target, cpu, tmpdir string) (job *compileJob, err error) {
 	// Try to load a precompiled library.
 	precompiledPath := filepath.Join(goenv.Get("TINYGOROOT"), "pkg", target, l.name+".a")
 	if _, err := os.Stat(precompiledPath); err == nil {
 		// Found a precompiled library for this OS/architecture. Return the path
 		// directly.
-		return precompiledPath, nil, nil
+		return dummyCompileJob(precompiledPath), nil
 	}

 	var outfile string
@@ -78,7 +76,7 @@ func (l *Library) load(target, cpu, tmpdir string) (path string, job *compileJob
 	// Try to fetch this library from the cache.
 	if path, err := cacheLoad(outfile, l.sourcePaths(target)); path != "" || err != nil {
 		// Cache hit.
-		return path, nil, err
+		return dummyCompileJob(path), nil
 	}
 	// Cache miss, build it now.

@@ -86,7 +84,7 @@ func (l *Library) load(target, cpu, tmpdir string) (path string, job *compileJob
 	dir := filepath.Join(tmpdir, "build-lib-"+l.name)
 	err = os.Mkdir(dir, 0777)
 	if err != nil {
-		return "", nil, err
+		return nil, err
 	}

 	// Precalculate the flags to the compiler invocation.
@@ -113,7 +111,8 @@ func (l *Library) load(target, cpu, tmpdir string) (path string, job *compileJob
 	arpath := filepath.Join(dir, l.name+".a")
 	job = &compileJob{
 		description: "ar " + l.name + ".a",
-		run: func() error {
+		result:      arpath,
+		run: func(*compileJob) error {
 			// Create an archive of all object files.
 			err := makeArchive(arpath, objs)
 			if err != nil {
@@ -133,11 +132,11 @@ func (l *Library) load(target, cpu, tmpdir string) (path string, job *compileJob
 		objs = append(objs, objpath)
 		job.dependencies = append(job.dependencies, &compileJob{
 			description: "compile " + srcpath,
-			run: func() error {
+			run: func(*compileJob) error {
 				var compileArgs []string
 				compileArgs = append(compileArgs, args...)
 				compileArgs = append(compileArgs, "-o", objpath, srcpath)
-				err := runCCompiler("clang", compileArgs...)
+				err := runCCompiler(compileArgs...)
 				if err != nil {
 					return &commandError{"failed to build", srcpath, err}
 				}
@@ -146,5 +145,5 @@ func (l *Library) load(target, cpu, tmpdir string) (path string, job *compileJob
 		})
 	}

-	return arpath, job, nil
+	return job, nil
 }
@@ -9,8 +9,8 @@ import (
 )

 // runCCompiler invokes a C compiler with the given arguments.
-func runCCompiler(command string, flags ...string) error {
-	if hasBuiltinTools && command == "clang" {
+func runCCompiler(flags ...string) error {
+	if hasBuiltinTools {
 		// Compile this with the internal Clang compiler.
 		headerPath := getClangHeaderPath(goenv.Get("TINYGOROOT"))
 		if headerPath == "" {
@@ -23,17 +23,8 @@ func runCCompiler(command string, flags ...string) error {
 		return cmd.Run()
 	}

-	// Running some other compiler. Maybe it has been defined in the
-	// commands map (unlikely).
-	if cmdNames, ok := commands[command]; ok {
-		return execCommand(cmdNames, flags...)
-	}
-
-	// Alternatively, run the compiler directly.
-	cmd := exec.Command(command, flags...)
-	cmd.Stdout = os.Stdout
-	cmd.Stderr = os.Stderr
-	return cmd.Run()
+	// Compile this with an external invocation of the Clang compiler.
+	return execCommand(commands["clang"], flags...)
 }

 // link invokes a linker with the given name and flags.
@@ -41,7 +41,9 @@ type cgoPackage struct {
 	elaboratedTypes map[string]*elaboratedTypeInfo
 	enums           map[string]enumInfo
 	anonStructNum   int
-	ldflags         []string
+	cflags          []string // CFlags from #cgo lines
+	ldflags         []string // LDFlags from #cgo lines
+	visitedFiles    map[string][]byte
 }

 // constantInfo stores some information about a CGo constant found by libclang
@@ -156,9 +158,10 @@ typedef unsigned long long  _Cgo_ulonglong;
 // Process extracts `import "C"` statements from the AST, parses the comment
 // with libclang, and modifies the AST to use this information. It returns a
 // newly created *ast.File that should be added to the list of to-be-parsed
-// files. If there is one or more error, it returns these in the []error slice
-// but still modifies the AST.
-func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string) (*ast.File, []string, []error) {
+// files, the CFLAGS and LDFLAGS found in #cgo lines, and a map of file hashes
+// of the accessed C header files. If there is one or more error, it returns
+// these in the []error slice but still modifies the AST.
+func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string) (*ast.File, []string, []string, map[string][]byte, []error) {
 	p := &cgoPackage{
 		dir:             dir,
 		fset:            fset,
@@ -170,13 +173,9 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 		typedefs:        map[string]*typedefInfo{},
 		elaboratedTypes: map[string]*elaboratedTypeInfo{},
 		enums:           map[string]enumInfo{},
+		visitedFiles:    map[string][]byte{},
 	}

-	// Disable _FORTIFY_SOURCE as it causes problems on macOS.
-	// Note that it is only disabled for memcpy (etc) calls made from Go, which
-	// have better alternatives anyway.
-	cflags = append(cflags, "-D_FORTIFY_SOURCE=0")
-
 	// Add a new location for the following file.
 	generatedTokenPos := p.fset.AddFile(dir+"/!cgo.go", -1, 0)
 	generatedTokenPos.SetLines([]int{0})
@@ -185,7 +184,7 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 	// Find the absolute path for this package.
 	packagePath, err := filepath.Abs(fset.File(files[0].Pos()).Name())
 	if err != nil {
-		return nil, nil, []error{
+		return nil, nil, nil, nil, []error{
 			scanner.Error{
 				Pos: fset.Position(files[0].Pos()),
 				Msg: "cgo: cannot find absolute path: " + err.Error(), // TODO: wrap this error
@@ -360,7 +359,7 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 						continue
 					}
 					makePathsAbsolute(flags, packagePath)
-					cflags = append(cflags, flags...)
+					p.cflags = append(p.cflags, flags...)
 				case "LDFLAGS":
 					flags, err := shlex.Split(value)
 					if err != nil {
@@ -383,6 +382,13 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 		}
 	}

+	// Define CFlags that will be used while parsing the package.
+	// Disable _FORTIFY_SOURCE as it causes problems on macOS.
+	// Note that it is only disabled for memcpy (etc) calls made from Go, which
+	// have better alternatives anyway.
+	cflagsForCGo := append([]string{"-D_FORTIFY_SOURCE=0"}, cflags...)
+	cflagsForCGo = append(cflagsForCGo, p.cflags...)
+
 	// Process all CGo imports.
 	for _, genDecl := range statements {
 		cgoComment := genDecl.Doc.Text()
@@ -392,7 +398,7 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 			pos = genDecl.Doc.Pos()
 		}
 		position := fset.PositionFor(pos, true)
-		p.parseFragment(cgoComment+cgoTypes, cflags, position.Filename, position.Line)
+		p.parseFragment(cgoComment+cgoTypes, cflagsForCGo, position.Filename, position.Line)
 	}

 	// Declare functions found by libclang.
@@ -427,7 +433,7 @@ func Process(files []*ast.File, dir string, fset *token.FileSet, cflags []string
 	// Print the newly generated in-memory AST, for debugging.
 	//ast.Print(fset, p.generated)

-	return p.generated, p.ldflags, p.errors
+	return p.generated, p.cflags, p.ldflags, p.visitedFiles, p.errors
 }

 // makePathsAbsolute converts some common path compiler flags (-I, -L) from
@@ -65,7 +65,7 @@ func TestCGo(t *testing.T) {
 			}

 			// Process the AST with CGo.
-			cgoAST, _, cgoErrors := Process([]*ast.File{f}, "testdata", fset, cflags)
+			cgoAST, _, _, _, cgoErrors := Process([]*ast.File{f}, "testdata", fset, cflags)

 			// Check the AST for type errors.
 			var typecheckErrors []error
@@ -4,6 +4,7 @@ package cgo
 // modification. It does not touch the AST itself.

 import (
+	"crypto/sha512"
 	"fmt"
 	"go/ast"
 	"go/scanner"
@@ -56,6 +57,7 @@ unsigned tinygo_clang_Cursor_isBitField(GoCXCursor c);
 int tinygo_clang_globals_visitor(GoCXCursor c, GoCXCursor parent, CXClientData client_data);
 int tinygo_clang_struct_visitor(GoCXCursor c, GoCXCursor parent, CXClientData client_data);
 int tinygo_clang_enum_visitor(GoCXCursor c, GoCXCursor parent, CXClientData client_data);
+void tinygo_clang_inclusion_visitor(CXFile included_file, CXSourceLocation *inclusion_stack, unsigned include_len, CXClientData client_data);
 */
 import "C"

@@ -114,6 +116,7 @@ func (p *cgoPackage) parseFragment(fragment string, cflags []string, posFilename
 	}
 	defer C.clang_disposeTranslationUnit(unit)

+	// Report parser and type errors.
 	if numDiagnostics := int(C.clang_getNumDiagnostics(unit)); numDiagnostics != 0 {
 		addDiagnostic := func(diagnostic C.CXDiagnostic) {
 			spelling := getString(C.clang_getDiagnosticSpelling(diagnostic))
@@ -134,10 +137,36 @@ func (p *cgoPackage) parseFragment(fragment string, cflags []string, posFilename
 		}
 	}

+	// Extract information required by CGo.
 	ref := storedRefs.Put(p)
 	defer storedRefs.Remove(ref)
 	cursor := C.tinygo_clang_getTranslationUnitCursor(unit)
 	C.tinygo_clang_visitChildren(cursor, C.CXCursorVisitor(C.tinygo_clang_globals_visitor), C.CXClientData(ref))
+
+	// Determine files read during CGo processing, for caching.
+	inclusionCallback := func(includedFile C.CXFile) {
+		// Get full file path.
+		path := getString(C.clang_getFileName(includedFile))
+
+		// Get contents of file (that should be in-memory).
+		size := C.size_t(0)
+		rawData := C.clang_getFileContents(unit, includedFile, &size)
+		if rawData == nil {
+			// Sanity check. This should (hopefully) never trigger.
+			panic("libclang: file contents was not loaded")
+		}
+		data := (*[1 << 24]byte)(unsafe.Pointer(rawData))[:size]
+
+		// Hash the contents if it isn't hashed yet.
+		if _, ok := p.visitedFiles[path]; !ok {
+			// already stored
+			sum := sha512.Sum512_224(data)
+			p.visitedFiles[path] = sum[:]
+		}
+	}
+	inclusionCallbackRef := storedRefs.Put(inclusionCallback)
+	defer storedRefs.Remove(inclusionCallbackRef)
+	C.clang_getInclusions(unit, C.CXInclusionVisitor(C.tinygo_clang_inclusion_visitor), C.CXClientData(inclusionCallbackRef))
 }

 //export tinygo_clang_globals_visitor
@@ -772,3 +801,9 @@ func tinygo_clang_enum_visitor(c, parent C.GoCXCursor, client_data C.CXClientDat
 	}
 	return C.CXChildVisit_Continue
 }
+
+//export tinygo_clang_inclusion_visitor
+func tinygo_clang_inclusion_visitor(includedFile C.CXFile, inclusionStack *C.CXSourceLocation, includeLen C.unsigned, clientData C.CXClientData) {
+	callback := storedRefs.Get(unsafe.Pointer(clientData)).(func(C.CXFile))
+	callback(includedFile)
+}
@@ -80,12 +80,7 @@ func (c *Config) GC() string {
 	if c.Target.GC != "" {
 		return c.Target.GC
 	}
-	for _, tag := range c.Target.BuildTags {
-		if tag == "baremetal" || tag == "wasm" {
-			return "conservative"
-		}
-	}
-	return "extalloc"
+	return "conservative"
 }

 // NeedsStackObjects returns true if the compiler should insert stack objects
@@ -118,6 +113,27 @@ func (c *Config) Scheduler() string {
 	return "coroutines"
 }

+// OptLevels returns the optimization level (0-2), size level (0-2), and inliner
+// threshold as used in the LLVM optimization pipeline.
+func (c *Config) OptLevels() (optLevel, sizeLevel int, inlinerThreshold uint) {
+	switch c.Options.Opt {
+	case "none", "0":
+		return 0, 0, 0 // -O0
+	case "1":
+		return 1, 0, 0 // -O1
+	case "2":
+		return 2, 0, 225 // -O2
+	case "s":
+		return 2, 1, 225 // -Os
+	case "z":
+		return 2, 2, 5 // -Oz, default
+	default:
+		// This is not shown to the user: valid choices are already checked as
+		// part of Options.Verify(). It is here as a sanity check.
+		panic("unknown optimization level: -opt=" + c.Options.Opt)
+	}
+}
+
 // FuncImplementation picks an appropriate func value implementation for the
 // target.
 func (c *Config) FuncImplementation() string {
@@ -163,7 +179,7 @@ func (c *Config) AutomaticStackSize() bool {
 // CFlags returns the flags to pass to the C compiler. This is necessary for CGo
 // preprocessing.
 func (c *Config) CFlags() []string {
-	cflags := append([]string{}, c.Options.CFlags...)
+	var cflags []string
 	for _, flag := range c.Target.CFlags {
 		cflags = append(cflags, strings.ReplaceAll(flag, "{root}", goenv.Get("TINYGOROOT")))
 	}
@@ -184,7 +200,7 @@ func (c *Config) CFlags() []string {
 func (c *Config) LDFlags() []string {
 	root := goenv.Get("TINYGOROOT")
 	// Merge and adjust LDFlags.
-	ldflags := append([]string{}, c.Options.LDFlags...)
+	var ldflags []string
 	for _, flag := range c.Target.LDFlags {
 		ldflags = append(ldflags, strings.ReplaceAll(flag, "{root}", root))
 	}
@@ -321,6 +337,10 @@ func (c *Config) WasmAbi() string {
 	return c.Target.WasmAbi
 }

+func (c *Config) LLVMFeatures() string {
+	return c.Options.LLVMFeatures
+}
+
 type TestConfig struct {
 	CompileTestBinary bool
 	// TODO: Filter the test functions to run, include verbose flag, etc
@@ -2,6 +2,7 @@ package compileopts

 import (
 	"fmt"
+	"regexp"
 	"strings"
 )

@@ -10,29 +11,32 @@ var (
 	validSchedulerOptions     = []string{"none", "tasks", "coroutines"}
 	validPrintSizeOptions     = []string{"none", "short", "full"}
 	validPanicStrategyOptions = []string{"print", "trap"}
+	validOptOptions           = []string{"none", "0", "1", "2", "s", "z"}
 )

 // Options contains extra options to give to the compiler. These options are
 // usually passed from the command line.
 type Options struct {
-	Target        string
-	Opt           string
-	GC            string
-	PanicStrategy string
-	Scheduler     string
-	PrintIR       bool
-	DumpSSA       bool
-	VerifyIR      bool
-	PrintCommands bool
-	Debug         bool
-	PrintSizes    string
-	PrintStacks   bool
-	CFlags        []string
-	LDFlags       []string
-	Tags          string
-	WasmAbi       string
-	TestConfig    TestConfig
-	Programmer    string
+	Target          string
+	Opt             string
+	GC              string
+	PanicStrategy   string
+	Scheduler       string
+	PrintIR         bool
+	DumpSSA         bool
+	VerifyIR        bool
+	PrintCommands   bool
+	Debug           bool
+	PrintSizes      string
+	PrintAllocs     *regexp.Regexp // regexp string
+	PrintStacks     bool
+	Tags            string
+	WasmAbi         string
+	GlobalValues    map[string]map[string]string // map[pkgpath]map[varname]value
+	TestConfig      TestConfig
+	Programmer      string
+	OpenOCDCommands []string
+	LLVMFeatures    string
 }

 // Verify performs a validation on the given options, raising an error if options are not valid.
@@ -73,6 +77,12 @@ func (o *Options) Verify() error {
 		}
 	}

+	if o.Opt != "" {
+		if !isInArray(validOptOptions, o.Opt) {
+			return fmt.Errorf("invalid -opt=%s: valid values are %s", o.Opt, strings.Join(validOptOptions, ", "))
+		}
+	}
+
 	return nil
 }

@@ -7,6 +7,7 @@ import (
 	"errors"
 	"io"
 	"os"
+	"os/exec"
 	"path/filepath"
 	"reflect"
 	"runtime"
@@ -30,7 +31,6 @@ type TargetSpec struct {
 	BuildTags        []string `json:"build-tags"`
 	GC               string   `json:"gc"`
 	Scheduler        string   `json:"scheduler"`
-	Compiler         string   `json:"compiler"`
 	Linker           string   `json:"linker"`
 	RTLib            string   `json:"rtlib"` // compiler runtime library (libgcc, compiler-rt)
 	Libc             string   `json:"libc"`
@@ -42,7 +42,7 @@ type TargetSpec struct {
 	ExtraFiles       []string `json:"extra-files"`
 	Emulator         []string `json:"emulator" override:"copy"` // inherited Emulator must not be append
 	FlashCommand     string   `json:"flash-command"`
-	GDB              string   `json:"gdb"`
+	GDB              []string `json:"gdb"`
 	PortReset        string   `json:"flash-1200-bps-reset"`
 	FlashMethod      string   `json:"flash-method"`
 	FlashVolume      string   `json:"msd-volume-name"`
@@ -229,31 +229,40 @@ func LoadTarget(target string) (*TargetSpec, error) {
 	}
 }

+// WindowsBuildNotSupportedErr is being thrown, when goos is windows and no target has been specified.
+var WindowsBuildNotSupportedErr = errors.New("Building Windows binaries is currently not supported. Try specifying a different target")
+
 func defaultTarget(goos, goarch, triple string) (*TargetSpec, error) {
+	if goos == "windows" {
+		return nil, WindowsBuildNotSupportedErr
+	}
 	// No target spec available. Use the default one, useful on most systems
 	// with a regular OS.
 	spec := TargetSpec{
-		Triple:    triple,
-		GOOS:      goos,
-		GOARCH:    goarch,
-		BuildTags: []string{goos, goarch},
-		Compiler:  "clang",
-		Linker:    "cc",
-		CFlags:    []string{"--target=" + triple},
-		GDB:       "gdb",
-		PortReset: "false",
+		Triple:           triple,
+		GOOS:             goos,
+		GOARCH:           goarch,
+		BuildTags:        []string{goos, goarch},
+		Scheduler:        "tasks",
+		Linker:           "cc",
+		DefaultStackSize: 1024 * 64, // 64kB
+		CFlags:           []string{"--target=" + triple},
+		GDB:              []string{"gdb"},
+		PortReset:        "false",
 	}
 	if goos == "darwin" {
+		spec.CFlags = append(spec.CFlags, "-isysroot", "/Library/Developer/CommandLineTools/SDKs/MacOSX.sdk")
 		spec.LDFlags = append(spec.LDFlags, "-Wl,-dead_strip")
 	} else {
 		spec.LDFlags = append(spec.LDFlags, "-no-pie", "-Wl,--gc-sections") // WARNING: clang < 5.0 requires -nopie
 	}
 	if goarch != "wasm" {
 		spec.ExtraFiles = append(spec.ExtraFiles, "src/runtime/gc_"+goarch+".S")
+		spec.ExtraFiles = append(spec.ExtraFiles, "src/internal/task/task_stack_"+goarch+".S")
 	}
 	if goarch != runtime.GOARCH {
 		// Some educated guesses as to how to invoke helper programs.
-		spec.GDB = "gdb-multiarch"
+		spec.GDB = []string{"gdb-multiarch"}
 		if goarch == "arm" && goos == "linux" {
 			spec.CFlags = append(spec.CFlags, "--sysroot=/usr/arm-linux-gnueabihf")
 			spec.Linker = "arm-linux-gnueabihf-gcc"
@@ -271,3 +280,17 @@ func defaultTarget(goos, goarch, triple string) (*TargetSpec, error) {
 	}
 	return &spec, nil
 }
+
+// LookupGDB looks up a gdb executable.
+func (spec *TargetSpec) LookupGDB() (string, error) {
+	if len(spec.GDB) == 0 {
+		return "", errors.New("gdb not configured in the target specification")
+	}
+	for _, d := range spec.GDB {
+		_, err := exec.LookPath(d)
+		if err == nil {
+			return d, nil
+		}
+	}
+	return "", errors.New("no gdb found configured in the target specification (" + strings.Join(spec.GDB, ", ") + ")")
+}
@@ -159,6 +159,9 @@ func (b *builder) createNilCheck(inst ssa.Value, ptr llvm.Value, blockPrefix str
 	}

 	switch inst := inst.(type) {
+	case *ssa.Alloc:
+		// An alloc is never nil.
+		return
 	case *ssa.IndexAddr:
 		// This pointer is the result of an index operation into a slice or
 		// array. Such slices/arrays are already bounds checked so the pointer
@@ -58,10 +58,10 @@ func (b *builder) createCall(fn llvm.Value, args []llvm.Value, name string) llvm

 // Expand an argument type to a list that can be used in a function call
 // parameter list.
-func expandFormalParamType(t llvm.Type, name string, goType types.Type) []paramInfo {
+func (c *compilerContext) expandFormalParamType(t llvm.Type, name string, goType types.Type) []paramInfo {
 	switch t.TypeKind() {
 	case llvm.StructTypeKind:
-		fieldInfos := flattenAggregateType(t, name, goType)
+		fieldInfos := c.flattenAggregateType(t, name, goType)
 		if len(fieldInfos) <= maxFieldsPerParam {
 			return fieldInfos
 		} else {
@@ -105,7 +105,7 @@ func (b *builder) expandFormalParamOffsets(t llvm.Type) []uint64 {
 func (b *builder) expandFormalParam(v llvm.Value) []llvm.Value {
 	switch v.Type().TypeKind() {
 	case llvm.StructTypeKind:
-		fieldInfos := flattenAggregateType(v.Type(), "", nil)
+		fieldInfos := b.flattenAggregateType(v.Type(), "", nil)
 		if len(fieldInfos) <= maxFieldsPerParam {
 			fields := b.flattenAggregate(v)
 			if len(fields) != len(fieldInfos) {
@@ -124,12 +124,15 @@ func (b *builder) expandFormalParam(v llvm.Value) []llvm.Value {

 // Try to flatten a struct type to a list of types. Returns a 1-element slice
 // with the passed in type if this is not possible.
-func flattenAggregateType(t llvm.Type, name string, goType types.Type) []paramInfo {
+func (c *compilerContext) flattenAggregateType(t llvm.Type, name string, goType types.Type) []paramInfo {
 	typeFlags := getTypeFlags(goType)
 	switch t.TypeKind() {
 	case llvm.StructTypeKind:
-		paramInfos := make([]paramInfo, 0, t.StructElementTypesCount())
+		var paramInfos []paramInfo
 		for i, subfield := range t.StructElementTypes() {
+			if c.targetData.TypeAllocSize(subfield) == 0 {
+				continue
+			}
 			suffix := strconv.Itoa(i)
 			if goType != nil {
 				// Try to come up with a good suffix for this struct field,
@@ -152,7 +155,7 @@ func flattenAggregateType(t llvm.Type, name string, goType types.Type) []paramIn
 					suffix = []string{"context", "funcptr"}[i]
 				}
 			}
-			subInfos := flattenAggregateType(subfield, name+"."+suffix, extractSubfield(goType, i))
+			subInfos := c.flattenAggregateType(subfield, name+"."+suffix, extractSubfield(goType, i))
 			for i := range subInfos {
 				subInfos[i].flags |= typeFlags
 			}
@@ -218,8 +221,11 @@ func extractSubfield(t types.Type, field int) types.Type {
 func (c *compilerContext) flattenAggregateTypeOffsets(t llvm.Type) []uint64 {
 	switch t.TypeKind() {
 	case llvm.StructTypeKind:
-		fields := make([]uint64, 0, t.StructElementTypesCount())
+		var fields []uint64
 		for fieldIndex, field := range t.StructElementTypes() {
+			if c.targetData.TypeAllocSize(field) == 0 {
+				continue
+			}
 			suboffsets := c.flattenAggregateTypeOffsets(field)
 			offset := c.targetData.ElementOffset(t, fieldIndex)
 			for i := range suboffsets {
@@ -238,8 +244,11 @@ func (c *compilerContext) flattenAggregateTypeOffsets(t llvm.Type) []uint64 {
 func (b *builder) flattenAggregate(v llvm.Value) []llvm.Value {
 	switch v.Type().TypeKind() {
 	case llvm.StructTypeKind:
-		fields := make([]llvm.Value, 0, v.Type().StructElementTypesCount())
-		for i := range v.Type().StructElementTypes() {
+		var fields []llvm.Value
+		for i, field := range v.Type().StructElementTypes() {
+			if b.targetData.TypeAllocSize(field) == 0 {
+				continue
+			}
 			subfield := b.CreateExtractValue(v, i, "")
 			subfields := b.flattenAggregate(subfield)
 			fields = append(fields, subfields...)
@@ -266,10 +275,13 @@ func (b *builder) collapseFormalParam(t llvm.Type, fields []llvm.Value) llvm.Val
 func (b *builder) collapseFormalParamInternal(t llvm.Type, fields []llvm.Value) (llvm.Value, []llvm.Value) {
 	switch t.TypeKind() {
 	case llvm.StructTypeKind:
-		flattened := flattenAggregateType(t, "", nil)
+		flattened := b.flattenAggregateType(t, "", nil)
 		if len(flattened) <= maxFieldsPerParam {
 			value := llvm.ConstNull(t)
 			for i, subtyp := range t.StructElementTypes() {
+				if b.targetData.TypeAllocSize(subtyp) == 0 {
+					continue
+				}
 				structField, remaining := b.collapseFormalParamInternal(subtyp, fields)
 				fields = remaining
 				value = b.CreateInsertValue(value, structField, i, "")
@@ -20,6 +20,11 @@ import (
 	"tinygo.org/x/go-llvm"
 )

+// Version of the compiler pacakge. Must be incremented each time the compiler
+// package changes in a way that affects the generated LLVM module.
+// This version is independent of the TinyGo version number.
+const Version = 9 // last change: implement reflect.New()
+
 func init() {
 	llvm.InitializeAllTargets()
 	llvm.InitializeAllTargetMCs()
@@ -54,6 +59,7 @@ type Config struct {
 	DefaultStackSize   uint64
 	NeedsStackObjects  bool
 	Debug              bool // Whether to emit debug information in the LLVM module.
+	LLVMFeatures       string
 }

 // compilerContext contains function-independent data that should still be
@@ -84,12 +90,13 @@ type compilerContext struct {
 // importantly with a newly created LLVM context and module.
 func newCompilerContext(moduleName string, machine llvm.TargetMachine, config *Config, dumpSSA bool) *compilerContext {
 	c := &compilerContext{
-		Config:     config,
-		DumpSSA:    dumpSSA,
-		difiles:    make(map[string]llvm.Metadata),
-		ditypes:    make(map[types.Type]llvm.Metadata),
-		machine:    machine,
-		targetData: machine.CreateTargetData(),
+		Config:      config,
+		DumpSSA:     dumpSSA,
+		difiles:     make(map[string]llvm.Metadata),
+		ditypes:     make(map[types.Type]llvm.Metadata),
+		machine:     machine,
+		targetData:  machine.CreateTargetData(),
+		astComments: map[string]*ast.CommentGroup{},
 	}

 	c.ctx = llvm.NewContext()
@@ -179,7 +186,12 @@ func NewTargetMachine(config *Config) (llvm.TargetMachine, error) {
 	if err != nil {
 		return llvm.TargetMachine{}, err
 	}
-	features := strings.Join(config.Features, ",")
+
+	feat := config.Features
+	if len(config.LLVMFeatures) > 0 {
+		feat = append(feat, config.LLVMFeatures)
+	}
+	features := strings.Join(feat, ",")

 	var codeModel llvm.CodeModel
 	var relocationModel llvm.RelocMode
@@ -241,21 +253,14 @@ func Sizes(machine llvm.TargetMachine) types.Sizes {
 	}
 }

-// CompileProgram compiles the given package path or .go file path. Return an
-// error when this fails (in any stage). If successful it returns the LLVM
-// module. If not, one or more errors will be returned.
-func CompileProgram(lprogram *loader.Program, machine llvm.TargetMachine, config *Config, dumpSSA bool) (llvm.Module, []error) {
-	c := newCompilerContext("", machine, config, dumpSSA)
+// CompilePackage compiles a single package to a LLVM module.
+func CompilePackage(moduleName string, pkg *loader.Package, ssaPkg *ssa.Package, machine llvm.TargetMachine, config *Config, dumpSSA bool) (llvm.Module, []error) {
+	c := newCompilerContext(moduleName, machine, config, dumpSSA)
+	c.runtimePkg = ssaPkg.Prog.ImportedPackage("runtime").Pkg
+	c.program = ssaPkg.Prog

-	c.program = lprogram.LoadSSA()
-	c.program.Build()
-	c.runtimePkg = c.program.ImportedPackage("runtime").Pkg
-
-	// Run a simple dead code elimination pass.
-	functions, err := c.simpleDCE(lprogram)
-	if err != nil {
-		return llvm.Module{}, []error{err}
-	}
+	// Convert AST to SSA.
+	ssaPkg.Build()

 	// Initialize debug information.
 	if c.Debug {
@@ -268,62 +273,31 @@ func CompileProgram(lprogram *loader.Program, machine llvm.TargetMachine, config
 		})
 	}

-	c.loadASTComments(lprogram)
+	// Load comments such as //go:extern on globals.
+	c.loadASTComments(pkg)

 	// Predeclare the runtime.alloc function, which is used by the wordpack
 	// functionality.
 	c.getFunction(c.program.ImportedPackage("runtime").Members["alloc"].(*ssa.Function))

-	// Add definitions to declarations.
-	var initFuncs []llvm.Value
+	// Compile all functions, methods, and global variables in this package.
 	irbuilder := c.ctx.NewBuilder()
 	defer irbuilder.Dispose()
-	for _, f := range functions {
-		if f.Synthetic == "package initializer" {
-			initFuncs = append(initFuncs, c.getFunction(f))
-		}
-		if f.Blocks == nil {
-			continue // external function
-		}
-
-		// Create the function definition.
-		b := newBuilder(c, irbuilder, f)
-		b.createFunction()
-	}
-
-	// After all packages are imported, add a synthetic initializer function
-	// that calls the initializer of each package.
-	initFn := c.program.ImportedPackage("runtime").Members["initAll"].(*ssa.Function)
-	llvmInitFn := c.getFunction(initFn)
-	llvmInitFn.SetLinkage(llvm.InternalLinkage)
-	llvmInitFn.SetUnnamedAddr(true)
-	if c.Debug {
-		difunc := c.attachDebugInfo(initFn)
-		pos := c.program.Fset.Position(initFn.Pos())
-		irbuilder.SetCurrentDebugLocation(uint(pos.Line), uint(pos.Column), difunc, llvm.Metadata{})
-	}
-	llvmInitFn.Param(0).SetName("context")
-	llvmInitFn.Param(1).SetName("parentHandle")
-	block := c.ctx.AddBasicBlock(llvmInitFn, "entry")
-	irbuilder.SetInsertPointAtEnd(block)
-	for _, fn := range initFuncs {
-		irbuilder.CreateCall(fn, []llvm.Value{llvm.Undef(c.i8ptrType), llvm.Undef(c.i8ptrType)}, "")
-	}
-	irbuilder.CreateRetVoid()
+	c.createPackage(irbuilder, ssaPkg)

 	// see: https://reviews.llvm.org/D18355
 	if c.Debug {
 		c.mod.AddNamedMetadataOperand("llvm.module.flags",
 			c.ctx.MDNode([]llvm.Metadata{
 				llvm.ConstInt(c.ctx.Int32Type(), 1, false).ConstantAsMetadata(), // Error on mismatch
-				llvm.GlobalContext().MDString("Debug Info Version"),
+				c.ctx.MDString("Debug Info Version"),
 				llvm.ConstInt(c.ctx.Int32Type(), 3, false).ConstantAsMetadata(), // DWARF version
 			}),
 		)
 		c.mod.AddNamedMetadataOperand("llvm.module.flags",
 			c.ctx.MDNode([]llvm.Metadata{
 				llvm.ConstInt(c.ctx.Int32Type(), 1, false).ConstantAsMetadata(),
-				llvm.GlobalContext().MDString("Dwarf Version"),
+				c.ctx.MDString("Dwarf Version"),
 				llvm.ConstInt(c.ctx.Int32Type(), 4, false).ConstantAsMetadata(),
 			}),
 		)
@@ -333,50 +307,6 @@ func CompileProgram(lprogram *loader.Program, machine llvm.TargetMachine, config
 	return c.mod, c.diagnostics
 }

-// CompilePackage compiles a single package to a LLVM module.
-func CompilePackage(moduleName string, pkg *loader.Package, machine llvm.TargetMachine, config *Config, dumpSSA bool) (llvm.Module, []error) {
-	c := newCompilerContext(moduleName, machine, config, dumpSSA)
-
-	// Build SSA from AST.
-	ssaPkg := pkg.LoadSSA()
-	ssaPkg.Build()
-
-	// Sort by position, so that the order of the functions in the IR matches
-	// the order of functions in the source file. This is useful for testing,
-	// for example.
-	var members []string
-	for name := range ssaPkg.Members {
-		members = append(members, name)
-	}
-	sort.Slice(members, func(i, j int) bool {
-		iPos := ssaPkg.Members[members[i]].Pos()
-		jPos := ssaPkg.Members[members[j]].Pos()
-		if i == j {
-			// Cannot sort by pos, so do it by name.
-			return members[i] < members[j]
-		}
-		return iPos < jPos
-	})
-
-	// Define all functions.
-	irbuilder := c.ctx.NewBuilder()
-	defer irbuilder.Dispose()
-	for _, name := range members {
-		member := ssaPkg.Members[name]
-		switch member := member.(type) {
-		case *ssa.Function:
-			if member.Blocks == nil {
-				continue // external function
-			}
-			// Create the function definition.
-			b := newBuilder(c, irbuilder, member)
-			b.createFunction()
-		}
-	}
-
-	return c.mod, nil
-}
-
 // getLLVMRuntimeType obtains a named type from the runtime package and returns
 // it as a LLVM type, creating it if necessary. It is a shorthand for
 // getLLVMType(getRuntimeType(name)).
@@ -720,8 +650,9 @@ func (c *compilerContext) attachDebugInfo(f *ssa.Function) llvm.Metadata {
 // debug info is added to the function.
 func (c *compilerContext) attachDebugInfoRaw(f *ssa.Function, llvmFn llvm.Value, suffix, filename string, line int) llvm.Metadata {
 	// Debug info for this function.
-	diparams := make([]llvm.Metadata, 0, len(f.Params))
-	for _, param := range f.Params {
+	params := getParams(f.Signature)
+	diparams := make([]llvm.Metadata, 0, len(params))
+	for _, param := range params {
 		diparams = append(diparams, c.getDIType(param.Type()))
 	}
 	diFuncType := c.dibuilder.CreateSubroutineType(llvm.DISubroutineType{
@@ -759,6 +690,83 @@ func (c *compilerContext) getDIFile(filename string) llvm.Metadata {
 	return c.difiles[filename]
 }

+// createPackage builds the LLVM IR for all types, methods, and global variables
+// in the given package.
+func (c *compilerContext) createPackage(irbuilder llvm.Builder, pkg *ssa.Package) {
+	// Sort by position, so that the order of the functions in the IR matches
+	// the order of functions in the source file. This is useful for testing,
+	// for example.
+	var members []string
+	for name := range pkg.Members {
+		members = append(members, name)
+	}
+	sort.Slice(members, func(i, j int) bool {
+		iPos := pkg.Members[members[i]].Pos()
+		jPos := pkg.Members[members[j]].Pos()
+		if i == j {
+			// Cannot sort by pos, so do it by name.
+			return members[i] < members[j]
+		}
+		return iPos < jPos
+	})
+
+	// Define all functions.
+	for _, name := range members {
+		member := pkg.Members[name]
+		switch member := member.(type) {
+		case *ssa.Function:
+			if member.Blocks == nil {
+				continue // external function
+			}
+			// Create the function definition.
+			b := newBuilder(c, irbuilder, member)
+			b.createFunction()
+		case *ssa.Type:
+			if types.IsInterface(member.Type()) {
+				// Interfaces don't have concrete methods.
+				continue
+			}
+
+			// Named type. We should make sure all methods are created.
+			// This includes both functions with pointer receivers and those
+			// without.
+			methods := getAllMethods(pkg.Prog, member.Type())
+			methods = append(methods, getAllMethods(pkg.Prog, types.NewPointer(member.Type()))...)
+			for _, method := range methods {
+				// Parse this method.
+				fn := pkg.Prog.MethodValue(method)
+				if fn.Blocks == nil {
+					continue // external function
+				}
+				if member.Type().String() != member.String() {
+					// This is a member on a type alias. Do not build such a
+					// function.
+					continue
+				}
+				if fn.Synthetic != "" && fn.Synthetic != "package initializer" {
+					// This function is a kind of wrapper function (created by
+					// the ssa package, not appearing in the source code) that
+					// is created by the getFunction method as needed.
+					// Therefore, don't build it here to avoid "function
+					// redeclared" errors.
+					continue
+				}
+				// Create the function definition.
+				b := newBuilder(c, irbuilder, fn)
+				b.createFunction()
+			}
+		case *ssa.Global:
+			// Global variable.
+			info := c.getGlobalInfo(member)
+			if !info.extern {
+				global := c.getGlobal(member)
+				global.SetInitializer(llvm.ConstNull(global.Type().ElementType()))
+				global.SetVisibility(llvm.HiddenVisibility)
+			}
+		}
+	}
+}
+
 // createFunction builds the LLVM IR implementation for this function. The
 // function must not yet be defined, otherwise this function will create a
 // diagnostic.
@@ -767,7 +775,7 @@ func (b *builder) createFunction() {
 		fmt.Printf("\nfunc %s:\n", b.fn)
 	}
 	if !b.llvmFn.IsDeclaration() {
-		errValue := b.fn.Name() + " redeclared in this program"
+		errValue := b.llvmFn.Name() + " redeclared in this program"
 		fnPos := getPosition(b.llvmFn)
 		if fnPos.IsValid() {
 			errValue += "\n\tprevious declaration at " + fnPos.String()
@@ -776,9 +784,15 @@ func (b *builder) createFunction() {
 		return
 	}
 	if !b.info.exported {
-		b.llvmFn.SetLinkage(llvm.InternalLinkage)
+		b.llvmFn.SetVisibility(llvm.HiddenVisibility)
 		b.llvmFn.SetUnnamedAddr(true)
 	}
+	if b.info.exported && strings.HasPrefix(b.Triple, "wasm") {
+		// Set the exported name. This is necessary for WebAssembly because
+		// otherwise the function is not exported.
+		functionAttr := b.ctx.CreateStringAttribute("wasm-export-name", b.info.linkName)
+		b.llvmFn.AddFunctionAttr(functionAttr)
+	}

 	// Some functions have a pragma controlling the inlining level.
 	switch b.info.inline {
@@ -821,7 +835,7 @@ func (b *builder) createFunction() {
 	for _, param := range b.fn.Params {
 		llvmType := b.getLLVMType(param.Type())
 		fields := make([]llvm.Value, 0, 1)
-		for _, info := range expandFormalParamType(llvmType, param.Name(), param.Type()) {
+		for _, info := range b.expandFormalParamType(llvmType, param.Name(), param.Type()) {
 			param := b.llvmFn.Param(llvmParamIndex)
 			param.SetName(info.name)
 			fields = append(fields, param)
@@ -948,13 +962,67 @@ func (b *builder) createFunction() {
 			b.trackValue(phi.llvm)
 		}
 	}
+
+	// Create anonymous functions (closures etc.).
+	for _, sub := range b.fn.AnonFuncs {
+		b := newBuilder(b.compilerContext, b.Builder, sub)
+		b.createFunction()
+	}
+}
+
+// posser is an interface that's implemented by both ssa.Value and
+// ssa.Instruction. It is implemented by everything that has a Pos() method,
+// which is all that getPos() needs.
+type posser interface {
+	Pos() token.Pos
+}
+
+// getPos returns position information for a ssa.Value or ssa.Instruction.
+//
+// Not all instructions have position information, especially when they're
+// implicit (such as implicit casts or implicit returns at the end of a
+// function). In these cases, it makes sense to try a bit harder to guess what
+// the position really should be.
+func getPos(val posser) token.Pos {
+	pos := val.Pos()
+	if pos != token.NoPos {
+		// Easy: position is known.
+		return pos
+	}
+
+	// No position information is known.
+	switch val := val.(type) {
+	case *ssa.MakeInterface:
+		return getPos(val.X)
+	case *ssa.Return:
+		syntax := val.Parent().Syntax()
+		if syntax != nil {
+			// non-synthetic
+			return syntax.End()
+		}
+		return token.NoPos
+	case *ssa.FieldAddr:
+		return getPos(val.X)
+	case *ssa.IndexAddr:
+		return getPos(val.X)
+	case *ssa.Slice:
+		return getPos(val.X)
+	case *ssa.Store:
+		return getPos(val.Addr)
+	case *ssa.Extract:
+		return getPos(val.Tuple)
+	default:
+		// This is reachable, for example with *ssa.Const, *ssa.If, and
+		// *ssa.Jump. They might be implemented in some way in the future.
+		return token.NoPos
+	}
 }

 // createInstruction builds the LLVM IR equivalent instructions for the
 // particular Go SSA instruction.
 func (b *builder) createInstruction(instr ssa.Instruction) {
 	if b.Debug {
-		pos := b.program.Fset.Position(instr.Pos())
+		pos := b.program.Fset.Position(getPos(instr))
 		b.SetCurrentDebugLocation(uint(pos.Line), uint(pos.Column), b.difunc, llvm.Metadata{})
 	}

@@ -2244,14 +2312,20 @@ func (b *builder) createConst(prefix string, expr *ssa.Const) llvm.Value {
 		} else if typ.Info()&types.IsString != 0 {
 			str := constant.StringVal(expr.Value)
 			strLen := llvm.ConstInt(b.uintptrType, uint64(len(str)), false)
-			objname := prefix + "$string"
-			global := llvm.AddGlobal(b.mod, llvm.ArrayType(b.ctx.Int8Type(), len(str)), objname)
-			global.SetInitializer(b.ctx.ConstString(str, false))
-			global.SetLinkage(llvm.InternalLinkage)
-			global.SetGlobalConstant(true)
-			global.SetUnnamedAddr(true)
-			zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
-			strPtr := b.CreateInBoundsGEP(global, []llvm.Value{zero, zero}, "")
+			var strPtr llvm.Value
+			if str != "" {
+				objname := prefix + "$string"
+				global := llvm.AddGlobal(b.mod, llvm.ArrayType(b.ctx.Int8Type(), len(str)), objname)
+				global.SetInitializer(b.ctx.ConstString(str, false))
+				global.SetLinkage(llvm.InternalLinkage)
+				global.SetGlobalConstant(true)
+				global.SetUnnamedAddr(true)
+				global.SetAlignment(1)
+				zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
+				strPtr = b.CreateInBoundsGEP(global, []llvm.Value{zero, zero}, "")
+			} else {
+				strPtr = llvm.ConstNull(b.i8ptrType)
+			}
 			strObj := llvm.ConstNamedStruct(b.getLLVMRuntimeType("_string"), []llvm.Value{strPtr, strLen})
 			return strObj
 		} else if typ.Kind() == types.UnsafePointer {
@@ -4,7 +4,6 @@ import (
 	"flag"
 	"go/types"
 	"io/ioutil"
-	"regexp"
 	"strconv"
 	"strings"
 	"testing"
@@ -20,7 +19,20 @@ var flagUpdate = flag.Bool("update", false, "update tests based on test output")
 // Basic tests for the compiler. Build some Go files and compare the output with
 // the expected LLVM IR for regression testing.
 func TestCompiler(t *testing.T) {
-	target, err := compileopts.LoadTarget("i686--linux")
+	// Check LLVM version.
+	llvmMajor, err := strconv.Atoi(strings.SplitN(llvm.Version, ".", 2)[0])
+	if err != nil {
+		t.Fatal("could not parse LLVM version:", llvm.Version)
+	}
+	if llvmMajor < 11 {
+		// It is likely this version needs to be bumped in the future.
+		// The goal is to at least test the LLVM version that's used by default
+		// in TinyGo and (if possible without too many workarounds) also some
+		// previous versions.
+		t.Skip("compiler tests require LLVM 11 or above, got LLVM ", llvm.Version)
+	}
+
+	target, err := compileopts.LoadTarget("wasm")
 	if err != nil {
 		t.Fatal("failed to load target:", err)
 	}
@@ -47,7 +59,10 @@ func TestCompiler(t *testing.T) {
 		"basic.go",
 		"pointer.go",
 		"slice.go",
+		"string.go",
 		"float.go",
+		"interface.go",
+		"func.go",
 	}

 	for _, testCase := range tests {
@@ -65,15 +80,21 @@ func TestCompiler(t *testing.T) {
 			}

 			// Compile AST to IR.
+			program := lprogram.LoadSSA()
 			pkg := lprogram.MainPkg()
-			mod, errs := CompilePackage(testCase, pkg, machine, compilerConfig, false)
+			mod, errs := CompilePackage(testCase, pkg, program.Package(pkg.Pkg), machine, compilerConfig, false)
 			if errs != nil {
 				for _, err := range errs {
-					t.Log("error:", err)
+					t.Error(err)
 				}
 				return
 			}

+			err = llvm.VerifyModule(mod, llvm.PrintMessageAction)
+			if err != nil {
+				t.Error(err)
+			}
+
 			// Optimize IR a little.
 			funcPasses := llvm.NewFunctionPassManagerForModule(mod)
 			defer funcPasses.Dispose()
@@ -107,8 +128,6 @@ func TestCompiler(t *testing.T) {
 	}
 }

-var alignRegexp = regexp.MustCompile(", align [0-9]+$")
-
 // fuzzyEqualIR returns true if the two LLVM IR strings passed in are roughly
 // equal. That means, only relevant lines are compared (excluding comments
 // etc.).
@@ -120,15 +139,6 @@ func fuzzyEqualIR(s1, s2 string) bool {
 	}
 	for i, line1 := range lines1 {
 		line2 := lines2[i]
-		match1 := alignRegexp.MatchString(line1)
-		match2 := alignRegexp.MatchString(line2)
-		if match1 != match2 {
-			// Only one of the lines has the align keyword. Remove it.
-			// This is a change to make the test work in both LLVM 10 and LLVM
-			// 11 (LLVM 11 appears to automatically add alignment everywhere).
-			line1 = alignRegexp.ReplaceAllString(line1, "")
-			line2 = alignRegexp.ReplaceAllString(line2, "")
-		}
 		if line1 != line2 {
 			return false
 		}
@@ -142,12 +152,6 @@ func fuzzyEqualIR(s1, s2 string) bool {
 // stripped out.
 func filterIrrelevantIRLines(lines []string) []string {
 	var out []string
-	llvmVersion, err := strconv.Atoi(strings.Split(llvm.Version, ".")[0])
-	if err != nil {
-		// Note: this should never happen and if it does, it will always happen
-		// for a particular build because llvm.Version is a constant.
-		panic(err)
-	}
 	for _, line := range lines {
 		line = strings.Split(line, ";")[0]    // strip out comments/info
 		line = strings.TrimRight(line, "\r ") // drop '\r' on Windows and remove trailing spaces from comments
@@ -157,15 +161,6 @@ func filterIrrelevantIRLines(lines []string) []string {
 		if strings.HasPrefix(line, "source_filename = ") {
 			continue
 		}
-		if llvmVersion < 10 && strings.HasPrefix(line, "attributes ") {
-			// Ignore attribute groups. These may change between LLVM versions.
-			// Right now test outputs are for LLVM 10.
-			continue
-		}
-		if llvmVersion < 10 && strings.HasPrefix(line, "target datalayout ") {
-			// Ignore the target layout. This may change between LLVM versions.
-			continue
-		}
 		out = append(out, line)
 	}
 	return out
@@ -352,7 +352,7 @@ func (b *builder) createRunDefers() {

 			// Get the real defer struct type and cast to it.
 			valueTypes := []llvm.Type{b.uintptrType, llvm.PointerType(b.getLLVMRuntimeType("_defer"), 0)}
-			for _, param := range callback.Params {
+			for _, param := range getParams(callback.Signature) {
 				valueTypes = append(valueTypes, b.getLLVMType(param.Type()))
 			}
 			deferFrameType := b.ctx.StructType(valueTypes, false)
@@ -361,7 +361,7 @@ func (b *builder) createRunDefers() {
 			// Extract the params from the struct.
 			forwardParams := []llvm.Value{}
 			zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
-			for i := range callback.Params {
+			for i := range getParams(callback.Signature) {
 				gep := b.CreateInBoundsGEP(deferFramePtr, []llvm.Value{zero, llvm.ConstInt(b.ctx.Int32Type(), uint64(i+2), false)}, "gep")
 				forwardParam := b.CreateLoad(gep, "param")
 				forwardParams = append(forwardParams, forwardParam)
@@ -25,19 +25,18 @@ func (c *compilerContext) createFuncValue(builder llvm.Builder, funcPtr, context
 		// Closure is: {context, function pointer}
 		funcValueScalar = funcPtr
 	case "switch":
-		sigGlobal := c.getTypeCode(sig)
 		funcValueWithSignatureGlobalName := funcPtr.Name() + "$withSignature"
 		funcValueWithSignatureGlobal := c.mod.NamedGlobal(funcValueWithSignatureGlobalName)
 		if funcValueWithSignatureGlobal.IsNil() {
 			funcValueWithSignatureType := c.getLLVMRuntimeType("funcValueWithSignature")
 			funcValueWithSignature := llvm.ConstNamedStruct(funcValueWithSignatureType, []llvm.Value{
 				llvm.ConstPtrToInt(funcPtr, c.uintptrType),
-				sigGlobal,
+				c.getFuncSignatureID(sig),
 			})
 			funcValueWithSignatureGlobal = llvm.AddGlobal(c.mod, funcValueWithSignatureType, funcValueWithSignatureGlobalName)
 			funcValueWithSignatureGlobal.SetInitializer(funcValueWithSignature)
 			funcValueWithSignatureGlobal.SetGlobalConstant(true)
-			funcValueWithSignatureGlobal.SetLinkage(llvm.InternalLinkage)
+			funcValueWithSignatureGlobal.SetLinkage(llvm.LinkOnceODRLinkage)
 		}
 		funcValueScalar = llvm.ConstPtrToInt(funcValueWithSignatureGlobal, c.uintptrType)
 	default:
@@ -50,6 +49,19 @@ func (c *compilerContext) createFuncValue(builder llvm.Builder, funcPtr, context
 	return funcValue
 }

+// getFuncSignatureID returns a new external global for a given signature. This
+// global reference is not real, it is only used during func lowering to assign
+// signature types to functions and will then be removed.
+func (c *compilerContext) getFuncSignatureID(sig *types.Signature) llvm.Value {
+	sigGlobalName := "reflect/types.funcid:" + getTypeCodeName(sig)
+	sigGlobal := c.mod.NamedGlobal(sigGlobalName)
+	if sigGlobal.IsNil() {
+		sigGlobal = llvm.AddGlobal(c.mod, c.ctx.Int8Type(), sigGlobalName)
+		sigGlobal.SetGlobalConstant(true)
+	}
+	return sigGlobal
+}
+
 // extractFuncScalar returns some scalar that can be used in comparisons. It is
 // a cheap operation.
 func (b *builder) extractFuncScalar(funcValue llvm.Value) llvm.Value {
@@ -71,7 +83,7 @@ func (b *builder) decodeFuncValue(funcValue llvm.Value, sig *types.Signature) (f
 		funcPtr = b.CreateExtractValue(funcValue, 1, "")
 	case "switch":
 		llvmSig := b.getRawFuncType(sig)
-		sigGlobal := b.getTypeCode(sig)
+		sigGlobal := b.getFuncSignatureID(sig)
 		funcPtr = b.createRuntimeCall("getFuncPtr", []llvm.Value{funcValue, sigGlobal}, "")
 		funcPtr = b.CreateIntToPtr(funcPtr, llvmSig, "")
 	default:
@@ -124,13 +136,13 @@ func (c *compilerContext) getRawFuncType(typ *types.Signature) llvm.Type {
 			// The receiver is not an interface, but a i8* type.
 			recv = c.i8ptrType
 		}
-		for _, info := range expandFormalParamType(recv, "", nil) {
+		for _, info := range c.expandFormalParamType(recv, "", nil) {
 			paramTypes = append(paramTypes, info.llvmType)
 		}
 	}
 	for i := 0; i < typ.Params().Len(); i++ {
 		subType := c.getLLVMType(typ.Params().At(i).Type())
-		for _, info := range expandFormalParamType(subType, "", nil) {
+		for _, info := range c.expandFormalParamType(subType, "", nil) {
 			paramTypes = append(paramTypes, info.llvmType)
 		}
 	}
@@ -34,6 +34,9 @@ func (b *builder) createGoInstruction(funcPtr llvm.Value, params []llvm.Value, p
 		} else {
 			// The stack size is fixed at compile time. By emitting it here as a
 			// constant, it can be optimized.
+			if b.Scheduler == "tasks" && b.DefaultStackSize == 0 {
+				b.addError(pos, "default stack size for goroutines is not set")
+			}
 			stackSize = llvm.ConstInt(b.uintptrType, b.DefaultStackSize, false)
 		}
 	case "coroutines":
@@ -84,7 +87,7 @@ func (c *compilerContext) createGoroutineStartWrapper(fn llvm.Value, prefix stri
 		// Create the wrapper.
 		wrapperType := llvm.FunctionType(c.ctx.VoidType(), []llvm.Type{c.i8ptrType}, false)
 		wrapper = llvm.AddFunction(c.mod, name+"$gowrapper", wrapperType)
-		wrapper.SetLinkage(llvm.InternalLinkage)
+		wrapper.SetLinkage(llvm.LinkOnceODRLinkage)
 		wrapper.SetUnnamedAddr(true)
 		wrapper.AddAttributeAtIndex(-1, c.ctx.CreateStringAttribute("tinygo-gowrapper", name))
 		entry := c.ctx.AddBasicBlock(wrapper, "entry")
@@ -141,7 +144,7 @@ func (c *compilerContext) createGoroutineStartWrapper(fn llvm.Value, prefix stri
 		// Create the wrapper.
 		wrapperType := llvm.FunctionType(c.ctx.VoidType(), []llvm.Type{c.i8ptrType}, false)
 		wrapper = llvm.AddFunction(c.mod, prefix+".gowrapper", wrapperType)
-		wrapper.SetLinkage(llvm.InternalLinkage)
+		wrapper.SetLinkage(llvm.LinkOnceODRLinkage)
 		wrapper.SetUnnamedAddr(true)
 		wrapper.AddAttributeAtIndex(-1, c.ctx.CreateStringAttribute("tinygo-gowrapper", ""))
 		entry := c.ctx.AddBasicBlock(wrapper, "entry")
@@ -24,16 +24,7 @@ import (
 func (b *builder) createMakeInterface(val llvm.Value, typ types.Type, pos token.Pos) llvm.Value {
 	itfValue := b.emitPointerPack([]llvm.Value{val})
 	itfTypeCodeGlobal := b.getTypeCode(typ)
-	itfMethodSetGlobal := b.getTypeMethodSet(typ)
-	itfConcreteTypeGlobal := b.mod.NamedGlobal("typeInInterface:" + itfTypeCodeGlobal.Name())
-	if itfConcreteTypeGlobal.IsNil() {
-		typeInInterface := b.getLLVMRuntimeType("typeInInterface")
-		itfConcreteTypeGlobal = llvm.AddGlobal(b.mod, typeInInterface, "typeInInterface:"+itfTypeCodeGlobal.Name())
-		itfConcreteTypeGlobal.SetInitializer(llvm.ConstNamedStruct(typeInInterface, []llvm.Value{itfTypeCodeGlobal, itfMethodSetGlobal}))
-		itfConcreteTypeGlobal.SetGlobalConstant(true)
-		itfConcreteTypeGlobal.SetLinkage(llvm.PrivateLinkage)
-	}
-	itfTypeCode := b.CreatePtrToInt(itfConcreteTypeGlobal, b.uintptrType, "")
+	itfTypeCode := b.CreatePtrToInt(itfTypeCodeGlobal, b.uintptrType, "")
 	itf := llvm.Undef(b.getLLVMRuntimeType("_interface"))
 	itf = b.CreateInsertValue(itf, itfTypeCode, 0, "")
 	itf = b.CreateInsertValue(itf, itfValue, 1, "")
@@ -54,6 +45,8 @@ func (c *compilerContext) getTypeCode(typ types.Type) llvm.Value {
 		// reflect lowering simpler.
 		var references llvm.Value
 		var length int64
+		var methodSet llvm.Value
+		var ptrTo llvm.Value
 		switch typ := typ.(type) {
 		case *types.Named:
 			references = c.getTypeCode(typ.Underlying())
@@ -70,18 +63,32 @@ func (c *compilerContext) getTypeCode(typ types.Type) llvm.Value {
 			// Take a pointer to the typecodeID of the first field (if it exists).
 			structGlobal := c.makeStructTypeFields(typ)
 			references = llvm.ConstBitCast(structGlobal, global.Type())
+		case *types.Interface:
+			methodSetGlobal := c.getInterfaceMethodSet(typ)
+			references = llvm.ConstBitCast(methodSetGlobal, global.Type())
 		}
+		if _, ok := typ.Underlying().(*types.Interface); !ok {
+			methodSet = c.getTypeMethodSet(typ)
+		}
+		if _, ok := typ.Underlying().(*types.Pointer); !ok {
+			ptrTo = c.getTypeCode(types.NewPointer(typ))
+		}
+		globalValue := llvm.ConstNull(global.Type().ElementType())
 		if !references.IsNil() {
-			// Set the 'references' field of the runtime.typecodeID struct.
-			globalValue := llvm.ConstNull(global.Type().ElementType())
 			globalValue = llvm.ConstInsertValue(globalValue, references, []uint32{0})
-			if length != 0 {
-				lengthValue := llvm.ConstInt(c.uintptrType, uint64(length), false)
-				globalValue = llvm.ConstInsertValue(globalValue, lengthValue, []uint32{1})
-			}
-			global.SetInitializer(globalValue)
-			global.SetLinkage(llvm.PrivateLinkage)
 		}
+		if length != 0 {
+			lengthValue := llvm.ConstInt(c.uintptrType, uint64(length), false)
+			globalValue = llvm.ConstInsertValue(globalValue, lengthValue, []uint32{1})
+		}
+		if !methodSet.IsNil() {
+			globalValue = llvm.ConstInsertValue(globalValue, methodSet, []uint32{2})
+		}
+		if !ptrTo.IsNil() {
+			globalValue = llvm.ConstInsertValue(globalValue, ptrTo, []uint32{3})
+		}
+		global.SetInitializer(globalValue)
+		global.SetLinkage(llvm.LinkOnceODRLinkage)
 		global.SetGlobalConstant(true)
 	}
 	return global
@@ -103,8 +110,8 @@ func (c *compilerContext) makeStructTypeFields(typ *types.Struct) llvm.Value {
 		fieldName.SetLinkage(llvm.PrivateLinkage)
 		fieldName.SetUnnamedAddr(true)
 		fieldName = llvm.ConstGEP(fieldName, []llvm.Value{
-			llvm.ConstInt(llvm.Int32Type(), 0, false),
-			llvm.ConstInt(llvm.Int32Type(), 0, false),
+			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+			llvm.ConstInt(c.ctx.Int32Type(), 0, false),
 		})
 		fieldGlobalValue = llvm.ConstInsertValue(fieldGlobalValue, fieldName, []uint32{1})
 		if typ.Tag(i) != "" {
@@ -112,8 +119,8 @@ func (c *compilerContext) makeStructTypeFields(typ *types.Struct) llvm.Value {
 			fieldTag.SetLinkage(llvm.PrivateLinkage)
 			fieldTag.SetUnnamedAddr(true)
 			fieldTag = llvm.ConstGEP(fieldTag, []llvm.Value{
-				llvm.ConstInt(llvm.Int32Type(), 0, false),
-				llvm.ConstInt(llvm.Int32Type(), 0, false),
+				llvm.ConstInt(c.ctx.Int32Type(), 0, false),
+				llvm.ConstInt(c.ctx.Int32Type(), 0, false),
 			})
 			fieldGlobalValue = llvm.ConstInsertValue(fieldGlobalValue, fieldTag, []uint32{2})
 		}
@@ -186,7 +193,7 @@ func getTypeCodeName(t types.Type) string {
 	case *types.Interface:
 		methods := make([]string, t.NumMethods())
 		for i := 0; i < t.NumMethods(); i++ {
-			methods[i] = getTypeCodeName(t.Method(i).Type())
+			methods[i] = t.Method(i).Name() + ":" + getTypeCodeName(t.Method(i).Type())
 		}
 		return "interface:" + "{" + strings.Join(methods, ",") + "}"
 	case *types.Map:
@@ -264,7 +271,7 @@ func (c *compilerContext) getTypeMethodSet(typ types.Type) llvm.Value {
 	global = llvm.AddGlobal(c.mod, arrayType, typ.String()+"$methodset")
 	global.SetInitializer(value)
 	global.SetGlobalConstant(true)
-	global.SetLinkage(llvm.PrivateLinkage)
+	global.SetLinkage(llvm.LinkOnceODRLinkage)
 	return llvm.ConstGEP(global, []llvm.Value{zero, zero})
 }

@@ -295,7 +302,7 @@ func (c *compilerContext) getInterfaceMethodSet(typ types.Type) llvm.Value {
 	global = llvm.AddGlobal(c.mod, value.Type(), name+"$interface")
 	global.SetInitializer(value)
 	global.SetGlobalConstant(true)
-	global.SetLinkage(llvm.PrivateLinkage)
+	global.SetLinkage(llvm.LinkOnceODRLinkage)
 	return llvm.ConstGEP(global, []llvm.Value{zero, zero})
 }

@@ -338,10 +345,16 @@ func (b *builder) createTypeAssert(expr *ssa.TypeAssert) llvm.Value {
 		commaOk = b.createRuntimeCall("interfaceImplements", []llvm.Value{actualTypeNum, methodSet}, "")

 	} else {
+		globalName := "reflect/types.typeid:" + getTypeCodeName(expr.AssertedType)
+		assertedTypeCodeGlobal := b.mod.NamedGlobal(globalName)
+		if assertedTypeCodeGlobal.IsNil() {
+			// Create a new typecode global.
+			assertedTypeCodeGlobal = llvm.AddGlobal(b.mod, b.ctx.Int8Type(), globalName)
+			assertedTypeCodeGlobal.SetGlobalConstant(true)
+		}
 		// Type assert on concrete type.
 		// Call runtime.typeAssert, which will be lowered to a simple icmp or
 		// const false in the interface lowering pass.
-		assertedTypeCodeGlobal := b.getTypeCode(expr.AssertedType)
 		commaOk = b.createRuntimeCall("typeAssert", []llvm.Value{actualTypeNum, assertedTypeCodeGlobal}, "typecode")
 	}

@@ -445,9 +458,9 @@ func (c *compilerContext) getInterfaceInvokeWrapper(fn *ssa.Function, llvmFn llv
 	}

 	// Get the expanded receiver type.
-	receiverType := c.getLLVMType(fn.Params[0].Type())
+	receiverType := c.getLLVMType(fn.Signature.Recv().Type())
 	var expandedReceiverType []llvm.Type
-	for _, info := range expandFormalParamType(receiverType, "", nil) {
+	for _, info := range c.expandFormalParamType(receiverType, "", nil) {
 		expandedReceiverType = append(expandedReceiverType, info.llvmType)
 	}

@@ -467,7 +480,7 @@ func (c *compilerContext) getInterfaceInvokeWrapper(fn *ssa.Function, llvmFn llv
 	wrapper = llvm.AddFunction(c.mod, wrapperName, wrapFnType)
 	wrapper.LastParam().SetName("parentHandle")

-	wrapper.SetLinkage(llvm.InternalLinkage)
+	wrapper.SetLinkage(llvm.LinkOnceODRLinkage)
 	wrapper.SetUnnamedAddr(true)

 	// Create a new builder just to create this wrapper.
@@ -46,7 +46,7 @@ func (b *builder) createInterruptGlobal(instr *ssa.CallCommon) (llvm.Value, erro
 		return llvm.Value{}, b.makeError(instr.Pos(), "interrupt redeclared in this program")
 	}
 	global := llvm.AddGlobal(b.mod, globalLLVMType, globalName)
-	global.SetLinkage(llvm.PrivateLinkage)
+	global.SetVisibility(llvm.HiddenVisibility)
 	global.SetGlobalConstant(true)
 	global.SetUnnamedAddr(true)
 	initializer := llvm.ConstNull(globalLLVMType)
@@ -1,164 +0,0 @@
-package compiler
-
-// This file implements a simple reachability analysis, to reduce compile time.
-// This DCE pass used to be necessary for improving other passes but now it
-// isn't necessary anymore.
-
-import (
-	"errors"
-	"go/types"
-	"sort"
-
-	"github.com/tinygo-org/tinygo/loader"
-	"golang.org/x/tools/go/ssa"
-)
-
-type dceState struct {
-	*compilerContext
-	functions   []*dceFunction
-	functionMap map[*ssa.Function]*dceFunction
-}
-
-type dceFunction struct {
-	*ssa.Function
-	functionInfo
-	flag bool // used by dead code elimination
-}
-
-func (p *dceState) addFunction(ssaFn *ssa.Function) {
-	if _, ok := p.functionMap[ssaFn]; ok {
-		return
-	}
-	f := &dceFunction{Function: ssaFn}
-	f.functionInfo = p.getFunctionInfo(ssaFn)
-	p.functions = append(p.functions, f)
-	p.functionMap[ssaFn] = f
-
-	for _, anon := range ssaFn.AnonFuncs {
-		p.addFunction(anon)
-	}
-}
-
-// simpleDCE returns a list of alive functions in the program. Compiling only
-// these functions makes the compiler faster.
-//
-// This functionality will likely be replaced in the future with build caching.
-func (c *compilerContext) simpleDCE(lprogram *loader.Program) ([]*ssa.Function, error) {
-	mainPkg := c.program.Package(lprogram.MainPkg().Pkg)
-	if mainPkg == nil {
-		panic("could not find main package")
-	}
-	p := &dceState{
-		compilerContext: c,
-		functionMap:     make(map[*ssa.Function]*dceFunction),
-	}
-
-	for _, pkg := range lprogram.Sorted() {
-		pkg := c.program.Package(pkg.Pkg)
-		memberNames := make([]string, 0)
-		for name := range pkg.Members {
-			memberNames = append(memberNames, name)
-		}
-		sort.Strings(memberNames)
-
-		for _, name := range memberNames {
-			member := pkg.Members[name]
-			switch member := member.(type) {
-			case *ssa.Function:
-				p.addFunction(member)
-			case *ssa.Type:
-				methods := getAllMethods(pkg.Prog, member.Type())
-				if !types.IsInterface(member.Type()) {
-					// named type
-					for _, method := range methods {
-						p.addFunction(pkg.Prog.MethodValue(method))
-					}
-				}
-			case *ssa.Global:
-				// Ignore. Globals are not handled here.
-			case *ssa.NamedConst:
-				// Ignore: these are already resolved.
-			default:
-				panic("unknown member type: " + member.String())
-			}
-		}
-	}
-
-	// Initial set of live functions. Include main.main, *.init and runtime.*
-	// functions.
-	main, ok := mainPkg.Members["main"].(*ssa.Function)
-	if !ok {
-		if mainPkg.Members["main"] == nil {
-			return nil, errors.New("function main is undeclared in the main package")
-		} else {
-			return nil, errors.New("cannot declare main - must be func")
-		}
-	}
-	runtimePkg := c.program.ImportedPackage("runtime")
-	mathPkg := c.program.ImportedPackage("math")
-	taskPkg := c.program.ImportedPackage("internal/task")
-	p.functionMap[main].flag = true
-	worklist := []*ssa.Function{main}
-	for _, f := range p.functions {
-		if f.exported || f.Synthetic == "package initializer" || f.Pkg == runtimePkg || f.Pkg == taskPkg || (f.Pkg == mathPkg && f.Pkg != nil) {
-			if f.flag {
-				continue
-			}
-			f.flag = true
-			worklist = append(worklist, f.Function)
-		}
-	}
-
-	// Mark all called functions recursively.
-	for len(worklist) != 0 {
-		f := worklist[len(worklist)-1]
-		worklist = worklist[:len(worklist)-1]
-		for _, block := range f.Blocks {
-			for _, instr := range block.Instrs {
-				if instr, ok := instr.(*ssa.MakeInterface); ok {
-					for _, sel := range getAllMethods(c.program, instr.X.Type()) {
-						fn := c.program.MethodValue(sel)
-						callee := p.functionMap[fn]
-						if callee == nil {
-							// TODO: why is this necessary?
-							p.addFunction(fn)
-							callee = p.functionMap[fn]
-						}
-						if !callee.flag {
-							callee.flag = true
-							worklist = append(worklist, callee.Function)
-						}
-					}
-				}
-				for _, operand := range instr.Operands(nil) {
-					if operand == nil || *operand == nil {
-						continue
-					}
-					switch operand := (*operand).(type) {
-					case *ssa.Function:
-						f := p.functionMap[operand]
-						if f == nil {
-							// FIXME HACK: this function should have been
-							// discovered already. It is not for bound methods.
-							p.addFunction(operand)
-							f = p.functionMap[operand]
-						}
-						if !f.flag {
-							f.flag = true
-							worklist = append(worklist, operand)
-						}
-					}
-				}
-			}
-		}
-	}
-
-	// Return all live functions.
-	liveFunctions := []*ssa.Function{}
-	for _, f := range p.functions {
-		if f.flag {
-			liveFunctions = append(liveFunctions, f.Function)
-		}
-	}
-	return liveFunctions, nil
-}
@@ -150,6 +150,9 @@ func (s *stdSizes) Sizeof(T types.Type) int64 {
 		return s.PtrSize * 2
 	case *types.Pointer:
 		return s.PtrSize
+	case *types.Signature:
+		// Func values in TinyGo are two words in size.
+		return s.PtrSize * 2
 	default:
 		panic("unknown type: " + t.String())
 	}
@@ -72,9 +72,9 @@ func (c *compilerContext) getFunction(fn *ssa.Function) llvm.Value {
 	}

 	var paramInfos []paramInfo
-	for _, param := range fn.Params {
+	for _, param := range getParams(fn.Signature) {
 		paramType := c.getLLVMType(param.Type())
-		paramFragmentInfos := expandFormalParamType(paramType, param.Name(), param.Type())
+		paramFragmentInfos := c.expandFormalParamType(paramType, param.Name(), param.Type())
 		paramInfos = append(paramInfos, paramFragmentInfos...)
 	}

@@ -139,6 +139,17 @@ func (c *compilerContext) getFunction(fn *ssa.Function) llvm.Value {
 		for _, attrName := range []string{"noalias", "nonnull"} {
 			llvmFn.AddAttributeAtIndex(0, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(attrName), 0))
 		}
+	case "runtime.sliceAppend":
+		// Appending a slice will only read the to-be-appended slice, it won't
+		// be modified.
+		llvmFn.AddAttributeAtIndex(2, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("nocapture"), 0))
+		llvmFn.AddAttributeAtIndex(2, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("readonly"), 0))
+	case "runtime.sliceCopy":
+		// Copying a slice won't capture any of the parameters.
+		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("writeonly"), 0))
+		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("nocapture"), 0))
+		llvmFn.AddAttributeAtIndex(2, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("readonly"), 0))
+		llvmFn.AddAttributeAtIndex(2, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("nocapture"), 0))
 	case "runtime.trackPointer":
 		// This function is necessary for tracking pointers on the stack in a
 		// portable way (see gc_stack_portable.go). Indicate to the optimizer
@@ -172,6 +183,21 @@ func (c *compilerContext) getFunction(fn *ssa.Function) llvm.Value {
 		}
 	}

+	// Synthetic functions are functions that do not appear in the source code,
+	// they are artificially constructed. Usually they are wrapper functions
+	// that are not referenced anywhere except in a SSA call instruction so
+	// should be created right away.
+	// The exception is the package initializer, which does appear in the
+	// *ssa.Package members and so shouldn't be created here.
+	if fn.Synthetic != "" && fn.Synthetic != "package initializer" {
+		irbuilder := c.ctx.NewBuilder()
+		b := newBuilder(c, irbuilder, fn)
+		b.createFunction()
+		irbuilder.Dispose()
+		llvmFn.SetLinkage(llvm.LinkOnceODRLinkage)
+		llvmFn.SetUnnamedAddr(true)
+	}
+
 	return llvmFn
 }

@@ -278,6 +304,20 @@ func (info *functionInfo) parsePragmas(f *ssa.Function) {
 	}
 }

+// getParams returns the function parameters, including the receiver at the
+// start. This is an alternative to the Params member of *ssa.Function, which is
+// not yet populated when the package has not yet been built.
+func getParams(sig *types.Signature) []*types.Var {
+	params := []*types.Var{}
+	if sig.Recv() != nil {
+		params = append(params, sig.Recv())
+	}
+	for i := 0; i < sig.Params().Len(); i++ {
+		params = append(params, sig.Params().At(i))
+	}
+	return params
+}
+
 // globalInfo contains some information about a specific global. By default,
 // linkName is equal to .RelString(nil) on a global and extern is false, but for
 // some symbols this is different (due to //go:extern for example).
@@ -289,25 +329,22 @@ type globalInfo struct {

 // loadASTComments loads comments on globals from the AST, for use later in the
 // program. In particular, they are required for //go:extern pragmas on globals.
-func (c *compilerContext) loadASTComments(lprogram *loader.Program) {
-	c.astComments = map[string]*ast.CommentGroup{}
-	for _, pkgInfo := range lprogram.Sorted() {
-		for _, file := range pkgInfo.Files {
-			for _, decl := range file.Decls {
-				switch decl := decl.(type) {
-				case *ast.GenDecl:
-					switch decl.Tok {
-					case token.VAR:
-						if len(decl.Specs) != 1 {
-							continue
-						}
-						for _, spec := range decl.Specs {
-							switch spec := spec.(type) {
-							case *ast.ValueSpec: // decl.Tok == token.VAR
-								for _, name := range spec.Names {
-									id := pkgInfo.Pkg.Path() + "." + name.Name
-									c.astComments[id] = decl.Doc
-								}
+func (c *compilerContext) loadASTComments(pkg *loader.Package) {
+	for _, file := range pkg.Files {
+		for _, decl := range file.Decls {
+			switch decl := decl.(type) {
+			case *ast.GenDecl:
+				switch decl.Tok {
+				case token.VAR:
+					if len(decl.Specs) != 1 {
+						continue
+					}
+					for _, spec := range decl.Specs {
+						switch spec := spec.(type) {
+						case *ast.ValueSpec: // decl.Tok == token.VAR
+							for _, name := range spec.Names {
+								id := pkg.Pkg.Path() + "." + name.Name
+								c.astComments[id] = decl.Doc
 							}
 						}
 					}
@@ -326,29 +363,25 @@ func (c *compilerContext) getGlobal(g *ssa.Global) llvm.Value {
 		typ := g.Type().(*types.Pointer).Elem()
 		llvmType := c.getLLVMType(typ)
 		llvmGlobal = llvm.AddGlobal(c.mod, llvmType, info.linkName)
-		if !info.extern {
-			llvmGlobal.SetInitializer(llvm.ConstNull(llvmType))
-			llvmGlobal.SetLinkage(llvm.InternalLinkage)
-		}

 		// Set alignment from the //go:align comment.
 		var alignInBits uint32
-		if info.align < 0 || info.align&(info.align-1) != 0 {
+		alignment := c.targetData.ABITypeAlignment(llvmType)
+		if info.align > alignment {
+			alignment = info.align
+		}
+		if alignment <= 0 || alignment&(alignment-1) != 0 {
 			// Check for power-of-two (or 0).
 			// See: https://stackoverflow.com/a/108360
 			c.addError(g.Pos(), "global variable alignment must be a positive power of two")
 		} else {
 			// Set the alignment only when it is a power of two.
-			alignInBits = uint32(info.align) ^ uint32(info.align-1)
-			if info.align > c.targetData.ABITypeAlignment(llvmType) {
-				llvmGlobal.SetAlignment(info.align)
-			}
+			alignInBits = uint32(alignment) ^ uint32(alignment-1)
+			llvmGlobal.SetAlignment(alignment)
 		}

 		if c.Debug && !info.extern {
 			// Add debug info.
-			// TODO: this should be done for every global in the program, not just
-			// the ones that are referenced from some code.
 			pos := c.program.Fset.Position(g.Pos())
 			diglobal := c.dibuilder.CreateGlobalVariableExpression(c.difiles[pos.Filename], llvm.DIGlobalVariableExpression{
 				Name:        g.RelString(nil),
@@ -1,72 +1,74 @@
 ; ModuleID = 'basic.go'
 source_filename = "basic.go"
-target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-f64:32:64-f80:32-n8:16:32-S128"
-target triple = "i686--linux"
+target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
+target triple = "wasm32--wasi"

-define internal void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  ret void
 }

-define internal i32 @main.addInt(i32 %x, i32 %y, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i32 @main.addInt(i32 %x, i32 %y, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = add i32 %x, %y
  ret i32 %0
 }

-define internal i1 @main.equalInt(i32 %x, i32 %y, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i1 @main.equalInt(i32 %x, i32 %y, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = icmp eq i32 %x, %y
  ret i1 %0
 }

-define internal i1 @main.floatEQ(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i1 @main.floatEQ(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = fcmp oeq float %x, %y
  ret i1 %0
 }

-define internal i1 @main.floatNE(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i1 @main.floatNE(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = fcmp une float %x, %y
  ret i1 %0
 }

-define internal i1 @main.floatLower(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i1 @main.floatLower(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = fcmp olt float %x, %y
  ret i1 %0
 }

-define internal i1 @main.floatLowerEqual(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i1 @main.floatLowerEqual(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = fcmp ole float %x, %y
  ret i1 %0
 }

-define internal i1 @main.floatGreater(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i1 @main.floatGreater(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = fcmp ogt float %x, %y
  ret i1 %0
 }

-define internal i1 @main.floatGreaterEqual(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i1 @main.floatGreaterEqual(float %x, float %y, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = fcmp oge float %x, %y
  ret i1 %0
 }

-define internal float @main.complexReal(float %x.r, float %x.i, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden float @main.complexReal(float %x.r, float %x.i, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  ret float %x.r
 }

-define internal float @main.complexImag(float %x.r, float %x.i, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden float @main.complexImag(float %x.r, float %x.i, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  ret float %x.i
 }

-define internal { float, float } @main.complexAdd(float %x.r, float %x.i, float %y.r, float %y.i, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden { float, float } @main.complexAdd(float %x.r, float %x.i, float %y.r, float %y.i, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = fadd float %x.r, %y.r
  %1 = fadd float %x.i, %y.i
@@ -75,7 +77,7 @@ entry:
  ret { float, float } %3
 }

-define internal { float, float } @main.complexSub(float %x.r, float %x.i, float %y.r, float %y.i, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden { float, float } @main.complexSub(float %x.r, float %x.i, float %y.r, float %y.i, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = fsub float %x.r, %y.r
  %1 = fsub float %x.i, %y.i
@@ -84,7 +86,7 @@ entry:
  ret { float, float } %3
 }

-define internal { float, float } @main.complexMul(float %x.r, float %x.i, float %y.r, float %y.i, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden { float, float } @main.complexMul(float %x.r, float %x.i, float %y.r, float %y.i, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = fmul float %x.r, %y.r
  %1 = fmul float %x.i, %y.i
@@ -1,14 +1,16 @@
 ; ModuleID = 'float.go'
 source_filename = "float.go"
-target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-f64:32:64-f80:32-n8:16:32-S128"
-target triple = "i686--linux"
+target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
+target triple = "wasm32--wasi"

-define internal void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  ret void
 }

-define internal i32 @main.f32tou32(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i32 @main.f32tou32(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %positive = fcmp oge float %v, 0.000000e+00
  %withinmax = fcmp ole float %v, 0x41EFFFFFC0000000
@@ -19,22 +21,22 @@ entry:
  ret i32 %0
 }

-define internal float @main.maxu32f(i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden float @main.maxu32f(i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  ret float 0x41F0000000000000
 }

-define internal i32 @main.maxu32tof32(i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i32 @main.maxu32tof32(i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  ret i32 -1
 }

-define internal { i32, i32, i32, i32 } @main.inftoi32(i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden { i32, i32, i32, i32 } @main.inftoi32(i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  ret { i32, i32, i32, i32 } { i32 -1, i32 0, i32 2147483647, i32 -2147483648 }
 }

-define internal i32 @main.u32tof32tou32(i32 %v, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i32 @main.u32tof32tou32(i32 %v, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = uitofp i32 %v to float
  %withinmax = fcmp ole float %0, 0x41EFFFFFC0000000
@@ -43,7 +45,7 @@ entry:
  ret i32 %1
 }

-define internal float @main.f32tou32tof32(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden float @main.f32tou32tof32(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %positive = fcmp oge float %v, 0.000000e+00
  %withinmax = fcmp ole float %v, 0x41EFFFFFC0000000
@@ -55,7 +57,7 @@ entry:
  ret float %1
 }

-define internal i8 @main.f32tou8(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i8 @main.f32tou8(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %positive = fcmp oge float %v, 0.000000e+00
  %withinmax = fcmp ole float %v, 2.550000e+02
@@ -66,7 +68,7 @@ entry:
  ret i8 %0
 }

-define internal i8 @main.f32toi8(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i8 @main.f32toi8(float %v, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %abovemin = fcmp oge float %v, -1.280000e+02
  %belowmax = fcmp ole float %v, 1.270000e+02
@@ -0,0 +1,12 @@
+package main
+
+func foo(callback func(int)) {
+	callback(3)
+}
+
+func bar() {
+	foo(someFunc)
+}
+
+func someFunc(int) {
+}
@@ -0,0 +1,47 @@
+; ModuleID = 'func.go'
+source_filename = "func.go"
+target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
+target triple = "wasm32--wasi"
+
+%runtime.funcValueWithSignature = type { i32, i8* }
+
+@"reflect/types.funcid:func:{basic:int}{}" = external constant i8
+@"main.someFunc$withSignature" = linkonce_odr constant %runtime.funcValueWithSignature { i32 ptrtoint (void (i32, i8*, i8*)* @main.someFunc to i32), i8* @"reflect/types.funcid:func:{basic:int}{}" }
+
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret void
+}
+
+define hidden void @main.foo(i8* %callback.context, i32 %callback.funcptr, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = call i32 @runtime.getFuncPtr(i8* %callback.context, i32 %callback.funcptr, i8* nonnull @"reflect/types.funcid:func:{basic:int}{}", i8* undef, i8* null)
+  %1 = icmp eq i32 %0, 0
+  br i1 %1, label %fpcall.throw, label %fpcall.next
+
+fpcall.throw:                                     ; preds = %entry
+  call void @runtime.nilPanic(i8* undef, i8* null)
+  unreachable
+
+fpcall.next:                                      ; preds = %entry
+  %2 = inttoptr i32 %0 to void (i32, i8*, i8*)*
+  call void %2(i32 3, i8* %callback.context, i8* undef)
+  ret void
+}
+
+declare i32 @runtime.getFuncPtr(i8*, i32, i8* dereferenceable_or_null(1), i8*, i8*)
+
+declare void @runtime.nilPanic(i8*, i8*)
+
+define hidden void @main.bar(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  call void @main.foo(i8* undef, i32 ptrtoint (%runtime.funcValueWithSignature* @"main.someFunc$withSignature" to i32), i8* undef, i8* undef)
+  ret void
+}
+
+define hidden void @main.someFunc(i32 %arg0, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret void
+}
@@ -0,0 +1,54 @@
+// This file tests interface types and interface builtins.
+
+package main
+
+// Test interface construction.
+
+func simpleType() interface{} {
+	return 0
+}
+
+func pointerType() interface{} {
+	// Pointers have an element type, in this case int.
+	var v *int
+	return v
+}
+
+func interfaceType() interface{} {
+	// Interfaces can exist in interfaces, but only indirectly (through
+	// pointers).
+	var v *error
+	return v
+}
+
+func anonymousInterfaceType() interface{} {
+	var v *interface {
+		String() string
+	}
+	return v
+}
+
+// Test interface builtins.
+
+func isInt(itf interface{}) bool {
+	_, ok := itf.(int)
+	return ok
+}
+
+func isError(itf interface{}) bool {
+	// Interface assert on (builtin) named interface type.
+	_, ok := itf.(error)
+	return ok
+}
+
+func isStringer(itf interface{}) bool {
+	// Interface assert on anonymous interface type.
+	_, ok := itf.(interface {
+		String() string
+	})
+	return ok
+}
+
+func callErrorMethod(itf error) string {
+	return itf.Error()
+}
@@ -0,0 +1,101 @@
+; ModuleID = 'interface.go'
+source_filename = "interface.go"
+target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
+target triple = "wasm32--wasi"
+
+%runtime.typecodeID = type { %runtime.typecodeID*, i32, %runtime.interfaceMethodInfo*, %runtime.typecodeID* }
+%runtime.interfaceMethodInfo = type { i8*, i32 }
+%runtime._interface = type { i32, i8* }
+%runtime._string = type { i8*, i32 }
+
+@"reflect/types.type:basic:int" = linkonce_odr constant %runtime.typecodeID { %runtime.typecodeID* null, i32 0, %runtime.interfaceMethodInfo* null, %runtime.typecodeID* @"reflect/types.type:pointer:basic:int" }
+@"reflect/types.type:pointer:basic:int" = linkonce_odr constant %runtime.typecodeID { %runtime.typecodeID* @"reflect/types.type:basic:int", i32 0, %runtime.interfaceMethodInfo* null, %runtime.typecodeID* null }
+@"reflect/types.type:pointer:named:error" = linkonce_odr constant %runtime.typecodeID { %runtime.typecodeID* @"reflect/types.type:named:error", i32 0, %runtime.interfaceMethodInfo* null, %runtime.typecodeID* null }
+@"reflect/types.type:named:error" = linkonce_odr constant %runtime.typecodeID { %runtime.typecodeID* @"reflect/types.type:interface:{Error:func:{}{basic:string}}", i32 0, %runtime.interfaceMethodInfo* null, %runtime.typecodeID* @"reflect/types.type:pointer:named:error" }
+@"reflect/types.type:interface:{Error:func:{}{basic:string}}" = linkonce_odr constant %runtime.typecodeID { %runtime.typecodeID* bitcast ([1 x i8*]* @"reflect/types.interface:interface{Error() string}$interface" to %runtime.typecodeID*), i32 0, %runtime.interfaceMethodInfo* null, %runtime.typecodeID* @"reflect/types.type:pointer:interface:{Error:func:{}{basic:string}}" }
+@"func Error() string" = external constant i8
+@"reflect/types.interface:interface{Error() string}$interface" = linkonce_odr constant [1 x i8*] [i8* @"func Error() string"]
+@"reflect/types.type:pointer:interface:{Error:func:{}{basic:string}}" = linkonce_odr constant %runtime.typecodeID { %runtime.typecodeID* @"reflect/types.type:interface:{Error:func:{}{basic:string}}", i32 0, %runtime.interfaceMethodInfo* null, %runtime.typecodeID* null }
+@"reflect/types.type:pointer:interface:{String:func:{}{basic:string}}" = linkonce_odr constant %runtime.typecodeID { %runtime.typecodeID* @"reflect/types.type:interface:{String:func:{}{basic:string}}", i32 0, %runtime.interfaceMethodInfo* null, %runtime.typecodeID* null }
+@"reflect/types.type:interface:{String:func:{}{basic:string}}" = linkonce_odr constant %runtime.typecodeID { %runtime.typecodeID* bitcast ([1 x i8*]* @"reflect/types.interface:interface{String() string}$interface" to %runtime.typecodeID*), i32 0, %runtime.interfaceMethodInfo* null, %runtime.typecodeID* @"reflect/types.type:pointer:interface:{String:func:{}{basic:string}}" }
+@"func String() string" = external constant i8
+@"reflect/types.interface:interface{String() string}$interface" = linkonce_odr constant [1 x i8*] [i8* @"func String() string"]
+@"reflect/types.typeid:basic:int" = external constant i8
+@"error$interface" = linkonce_odr constant [1 x i8*] [i8* @"func Error() string"]
+
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret void
+}
+
+define hidden %runtime._interface @main.simpleType(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret %runtime._interface { i32 ptrtoint (%runtime.typecodeID* @"reflect/types.type:basic:int" to i32), i8* null }
+}
+
+define hidden %runtime._interface @main.pointerType(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret %runtime._interface { i32 ptrtoint (%runtime.typecodeID* @"reflect/types.type:pointer:basic:int" to i32), i8* null }
+}
+
+define hidden %runtime._interface @main.interfaceType(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret %runtime._interface { i32 ptrtoint (%runtime.typecodeID* @"reflect/types.type:pointer:named:error" to i32), i8* null }
+}
+
+define hidden %runtime._interface @main.anonymousInterfaceType(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret %runtime._interface { i32 ptrtoint (%runtime.typecodeID* @"reflect/types.type:pointer:interface:{String:func:{}{basic:string}}" to i32), i8* null }
+}
+
+define hidden i1 @main.isInt(i32 %itf.typecode, i8* %itf.value, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %typecode = call i1 @runtime.typeAssert(i32 %itf.typecode, i8* nonnull @"reflect/types.typeid:basic:int", i8* undef, i8* null)
+  br i1 %typecode, label %typeassert.ok, label %typeassert.next
+
+typeassert.ok:                                    ; preds = %entry
+  br label %typeassert.next
+
+typeassert.next:                                  ; preds = %typeassert.ok, %entry
+  ret i1 %typecode
+}
+
+declare i1 @runtime.typeAssert(i32, i8* dereferenceable_or_null(1), i8*, i8*)
+
+define hidden i1 @main.isError(i32 %itf.typecode, i8* %itf.value, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = call i1 @runtime.interfaceImplements(i32 %itf.typecode, i8** getelementptr inbounds ([1 x i8*], [1 x i8*]* @"error$interface", i32 0, i32 0), i8* undef, i8* null)
+  br i1 %0, label %typeassert.ok, label %typeassert.next
+
+typeassert.ok:                                    ; preds = %entry
+  br label %typeassert.next
+
+typeassert.next:                                  ; preds = %typeassert.ok, %entry
+  ret i1 %0
+}
+
+declare i1 @runtime.interfaceImplements(i32, i8** dereferenceable_or_null(4), i8*, i8*)
+
+define hidden i1 @main.isStringer(i32 %itf.typecode, i8* %itf.value, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = call i1 @runtime.interfaceImplements(i32 %itf.typecode, i8** getelementptr inbounds ([1 x i8*], [1 x i8*]* @"reflect/types.interface:interface{String() string}$interface", i32 0, i32 0), i8* undef, i8* null)
+  br i1 %0, label %typeassert.ok, label %typeassert.next
+
+typeassert.ok:                                    ; preds = %entry
+  br label %typeassert.next
+
+typeassert.next:                                  ; preds = %typeassert.ok, %entry
+  ret i1 %0
+}
+
+define hidden %runtime._string @main.callErrorMethod(i32 %itf.typecode, i8* %itf.value, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %invoke.func = call i32 @runtime.interfaceMethod(i32 %itf.typecode, i8** getelementptr inbounds ([1 x i8*], [1 x i8*]* @"error$interface", i32 0, i32 0), i8* nonnull @"func Error() string", i8* undef, i8* null)
+  %invoke.func.cast = inttoptr i32 %invoke.func to %runtime._string (i8*, i8*, i8*)*
+  %0 = call %runtime._string %invoke.func.cast(i8* %itf.value, i8* undef, i8* undef)
+  ret %runtime._string %0
+}
+
+declare i32 @runtime.interfaceMethod(i32, i8** dereferenceable_or_null(4), i8* dereferenceable_or_null(1), i8*, i8*)
@@ -1,48 +1,50 @@
 ; ModuleID = 'pointer.go'
 source_filename = "pointer.go"
-target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-f64:32:64-f80:32-n8:16:32-S128"
-target triple = "i686--linux"
+target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
+target triple = "wasm32--wasi"

-define internal void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  ret void
 }

-define internal [0 x i32] @main.pointerDerefZero([0 x i32]* %x, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden [0 x i32] @main.pointerDerefZero([0 x i32]* %x, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  ret [0 x i32] zeroinitializer
 }

-define internal i32* @main.pointerCastFromUnsafe(i8* %x, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i32* @main.pointerCastFromUnsafe(i8* %x, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = bitcast i8* %x to i32*
  ret i32* %0
 }

-define internal i8* @main.pointerCastToUnsafe(i32* dereferenceable_or_null(4) %x, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i8* @main.pointerCastToUnsafe(i32* dereferenceable_or_null(4) %x, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = bitcast i32* %x to i8*
  ret i8* %0
 }

-define internal i8* @main.pointerCastToUnsafeNoop(i8* dereferenceable_or_null(1) %x, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i8* @main.pointerCastToUnsafeNoop(i8* dereferenceable_or_null(1) %x, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  ret i8* %x
 }

-define internal i8* @main.pointerUnsafeGEPFixedOffset(i8* dereferenceable_or_null(1) %ptr, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i8* @main.pointerUnsafeGEPFixedOffset(i8* dereferenceable_or_null(1) %ptr, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = getelementptr inbounds i8, i8* %ptr, i32 10
  ret i8* %0
 }

-define internal i8* @main.pointerUnsafeGEPByteOffset(i8* dereferenceable_or_null(1) %ptr, i32 %offset, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i8* @main.pointerUnsafeGEPByteOffset(i8* dereferenceable_or_null(1) %ptr, i32 %offset, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = getelementptr inbounds i8, i8* %ptr, i32 %offset
  ret i8* %0
 }

-define internal i32* @main.pointerUnsafeGEPIntOffset(i32* dereferenceable_or_null(4) %ptr, i32 %offset, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i32* @main.pointerUnsafeGEPIntOffset(i32* dereferenceable_or_null(4) %ptr, i32 %offset, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  %0 = getelementptr i32, i32* %ptr, i32 %offset
  ret i32* %0
@@ -7,3 +7,19 @@ func sliceLen(ints []int) int {
 func sliceCap(ints []int) int {
 	return cap(ints)
 }
+
+func sliceElement(ints []int, index int) int {
+	return ints[index]
+}
+
+func sliceAppendValues(ints []int) []int {
+	return append(ints, 1, 2, 3)
+}
+
+func sliceAppendSlice(ints, added []int) []int {
+	return append(ints, added...)
+}
+
+func sliceCopy(dst, src []int) int {
+	return copy(dst, src)
+}
@@ -1,19 +1,88 @@
 ; ModuleID = 'slice.go'
 source_filename = "slice.go"
-target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-f64:32:64-f80:32-n8:16:32-S128"
-target triple = "i686--linux"
+target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
+target triple = "wasm32--wasi"

-define internal void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  ret void
 }

-define internal i32 @main.sliceLen(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i32 @main.sliceLen(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  ret i32 %ints.len
 }

-define internal i32 @main.sliceCap(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i8* %context, i8* %parentHandle) unnamed_addr {
+define hidden i32 @main.sliceCap(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i8* %context, i8* %parentHandle) unnamed_addr {
 entry:
  ret i32 %ints.cap
 }
+
+define hidden i32 @main.sliceElement(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i32 %index, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %.not = icmp ult i32 %index, %ints.len
+  br i1 %.not, label %lookup.next, label %lookup.throw
+
+lookup.throw:                                     ; preds = %entry
+  call void @runtime.lookupPanic(i8* undef, i8* null)
+  unreachable
+
+lookup.next:                                      ; preds = %entry
+  %0 = getelementptr inbounds i32, i32* %ints.data, i32 %index
+  %1 = load i32, i32* %0, align 4
+  ret i32 %1
+}
+
+declare void @runtime.lookupPanic(i8*, i8*)
+
+define hidden { i32*, i32, i32 } @main.sliceAppendValues(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %varargs = call i8* @runtime.alloc(i32 12, i8* undef, i8* null)
+  %0 = bitcast i8* %varargs to i32*
+  store i32 1, i32* %0, align 4
+  %1 = getelementptr inbounds i8, i8* %varargs, i32 4
+  %2 = bitcast i8* %1 to i32*
+  store i32 2, i32* %2, align 4
+  %3 = getelementptr inbounds i8, i8* %varargs, i32 8
+  %4 = bitcast i8* %3 to i32*
+  store i32 3, i32* %4, align 4
+  %append.srcPtr = bitcast i32* %ints.data to i8*
+  %append.new = call { i8*, i32, i32 } @runtime.sliceAppend(i8* %append.srcPtr, i8* nonnull %varargs, i32 %ints.len, i32 %ints.cap, i32 3, i32 4, i8* undef, i8* null)
+  %append.newPtr = extractvalue { i8*, i32, i32 } %append.new, 0
+  %append.newBuf = bitcast i8* %append.newPtr to i32*
+  %append.newLen = extractvalue { i8*, i32, i32 } %append.new, 1
+  %append.newCap = extractvalue { i8*, i32, i32 } %append.new, 2
+  %5 = insertvalue { i32*, i32, i32 } undef, i32* %append.newBuf, 0
+  %6 = insertvalue { i32*, i32, i32 } %5, i32 %append.newLen, 1
+  %7 = insertvalue { i32*, i32, i32 } %6, i32 %append.newCap, 2
+  ret { i32*, i32, i32 } %7
+}
+
+declare { i8*, i32, i32 } @runtime.sliceAppend(i8*, i8* nocapture readonly, i32, i32, i32, i32, i8*, i8*)
+
+define hidden { i32*, i32, i32 } @main.sliceAppendSlice(i32* %ints.data, i32 %ints.len, i32 %ints.cap, i32* %added.data, i32 %added.len, i32 %added.cap, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %append.srcPtr = bitcast i32* %ints.data to i8*
+  %append.srcPtr1 = bitcast i32* %added.data to i8*
+  %append.new = call { i8*, i32, i32 } @runtime.sliceAppend(i8* %append.srcPtr, i8* %append.srcPtr1, i32 %ints.len, i32 %ints.cap, i32 %added.len, i32 4, i8* undef, i8* null)
+  %append.newPtr = extractvalue { i8*, i32, i32 } %append.new, 0
+  %append.newBuf = bitcast i8* %append.newPtr to i32*
+  %append.newLen = extractvalue { i8*, i32, i32 } %append.new, 1
+  %append.newCap = extractvalue { i8*, i32, i32 } %append.new, 2
+  %0 = insertvalue { i32*, i32, i32 } undef, i32* %append.newBuf, 0
+  %1 = insertvalue { i32*, i32, i32 } %0, i32 %append.newLen, 1
+  %2 = insertvalue { i32*, i32, i32 } %1, i32 %append.newCap, 2
+  ret { i32*, i32, i32 } %2
+}
+
+define hidden i32 @main.sliceCopy(i32* %dst.data, i32 %dst.len, i32 %dst.cap, i32* %src.data, i32 %src.len, i32 %src.cap, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %copy.dstPtr = bitcast i32* %dst.data to i8*
+  %copy.srcPtr = bitcast i32* %src.data to i8*
+  %copy.n = call i32 @runtime.sliceCopy(i8* %copy.dstPtr, i8* %copy.srcPtr, i32 %dst.len, i32 %src.len, i32 4, i8* undef, i8* null)
+  ret i32 %copy.n
+}
+
+declare i32 @runtime.sliceCopy(i8* nocapture writeonly, i8* nocapture readonly, i32, i32, i32, i8*, i8*)
@@ -0,0 +1,29 @@
+package main
+
+func someString() string {
+	return "foo"
+}
+
+func zeroLengthString() string {
+	return ""
+}
+
+func stringLen(s string) int {
+	return len(s)
+}
+
+func stringIndex(s string, index int) byte {
+	return s[index]
+}
+
+func stringCompareEqual(s1, s2 string) bool {
+	return s1 == s2
+}
+
+func stringCompareUnequal(s1, s2 string) bool {
+	return s1 != s2
+}
+
+func stringCompareLarger(s1, s2 string) bool {
+	return s1 > s2
+}
@@ -0,0 +1,71 @@
+; ModuleID = 'string.go'
+source_filename = "string.go"
+target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
+target triple = "wasm32--wasi"
+
+%runtime._string = type { i8*, i32 }
+
+@"main.someString$string" = internal unnamed_addr constant [3 x i8] c"foo", align 1
+
+declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*)
+
+define hidden void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret void
+}
+
+define hidden %runtime._string @main.someString(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret %runtime._string { i8* getelementptr inbounds ([3 x i8], [3 x i8]* @"main.someString$string", i32 0, i32 0), i32 3 }
+}
+
+define hidden %runtime._string @main.zeroLengthString(i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret %runtime._string zeroinitializer
+}
+
+define hidden i32 @main.stringLen(i8* %s.data, i32 %s.len, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  ret i32 %s.len
+}
+
+define hidden i8 @main.stringIndex(i8* %s.data, i32 %s.len, i32 %index, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %.not = icmp ult i32 %index, %s.len
+  br i1 %.not, label %lookup.next, label %lookup.throw
+
+lookup.throw:                                     ; preds = %entry
+  call void @runtime.lookupPanic(i8* undef, i8* null)
+  unreachable
+
+lookup.next:                                      ; preds = %entry
+  %0 = getelementptr inbounds i8, i8* %s.data, i32 %index
+  %1 = load i8, i8* %0, align 1
+  ret i8 %1
+}
+
+declare void @runtime.lookupPanic(i8*, i8*)
+
+define hidden i1 @main.stringCompareEqual(i8* %s1.data, i32 %s1.len, i8* %s2.data, i32 %s2.len, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = call i1 @runtime.stringEqual(i8* %s1.data, i32 %s1.len, i8* %s2.data, i32 %s2.len, i8* undef, i8* null)
+  ret i1 %0
+}
+
+declare i1 @runtime.stringEqual(i8*, i32, i8*, i32, i8*, i8*)
+
+define hidden i1 @main.stringCompareUnequal(i8* %s1.data, i32 %s1.len, i8* %s2.data, i32 %s2.len, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = call i1 @runtime.stringEqual(i8* %s1.data, i32 %s1.len, i8* %s2.data, i32 %s2.len, i8* undef, i8* null)
+  %1 = xor i1 %0, true
+  ret i1 %1
+}
+
+define hidden i1 @main.stringCompareLarger(i8* %s1.data, i32 %s1.len, i8* %s2.data, i32 %s2.len, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %0 = call i1 @runtime.stringLess(i8* %s1.data, i32 %s1.len, i8* %s2.data, i32 %s2.len, i8* undef, i8* null)
+  %1 = xor i1 %0, true
+  ret i1 %1
+}
+
+declare i1 @runtime.stringLess(i8*, i32, i8*, i32, i8*, i8*)
@@ -12,5 +12,5 @@ require (
 	go.bug.st/serial v1.1.2
 	golang.org/x/sys v0.0.0-20210113181707-4bcb84eeeb78
 	golang.org/x/tools v0.0.0-20200216192241-b320d3a0f5a2
-	tinygo.org/x/go-llvm v0.0.0-20210206225315-7fe719483a0f
+	tinygo.org/x/go-llvm v0.0.0-20210325115028-e7b85195e81c
 )
@@ -57,5 +57,7 @@ golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898 h1:/atklqdjdhuosWIl6AIbO
 golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
 gopkg.in/yaml.v2 v2.2.2/go.mod h1:hI93XBmqTisBFMUTm0b8Fm+jr3Dg1NNxqwp+5A1VGuI=
-tinygo.org/x/go-llvm v0.0.0-20210206225315-7fe719483a0f h1:FP5Do5omlQ/dLQ3Hfy7oyJo69VS5Hn46rZw004r0lGU=
-tinygo.org/x/go-llvm v0.0.0-20210206225315-7fe719483a0f/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
+tinygo.org/x/go-llvm v0.0.0-20210308112806-9ef958b6bed4 h1:CMUHxVTb+UuUePuMf8vkWjZ3gTp9BBK91KrgOCwoNHs=
+tinygo.org/x/go-llvm v0.0.0-20210308112806-9ef958b6bed4/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
+tinygo.org/x/go-llvm v0.0.0-20210325115028-e7b85195e81c h1:vn9IPshzYmzZis10UEVrsIBRv9FpykADw6M3/tHHROg=
+tinygo.org/x/go-llvm v0.0.0-20210325115028-e7b85195e81c/go.mod h1:fv1F0BSNpxMfCL0zF3M4OPFbgYHnhtB6ST0HvUtu/LE=
@@ -171,8 +171,9 @@ func getGoroot() string {
 	switch runtime.GOOS {
 	case "linux":
 		candidates = []string{
-			"/usr/local/go", // manually installed
-			"/usr/lib/go",   // from the distribution
+			"/usr/local/go",     // manually installed
+			"/usr/lib/go",       // from the distribution
+			"/snap/go/current/", // installed using snap
 		}
 	case "darwin":
 		candidates = []string{
@@ -12,7 +12,7 @@ import (

 // Version of TinyGo.
 // Update this value before release of new version of software.
-const Version = "0.18.0-dev"
+const Version = "0.18.0"

 // GetGorootVersion returns the major and minor version for a given GOROOT path.
 // If the goroot cannot be determined, (0, 0) is returned.
@@ -25,6 +25,7 @@ type function struct {
 // basicBlock represents a LLVM basic block and contains a slice of
 // instructions. The last instruction must be a terminator instruction.
 type basicBlock struct {
+	phiNodes     []instruction
 	instructions []instruction
 }

@@ -135,6 +136,15 @@ func (r *runner) compileFunction(llvmFn llvm.Value) *function {
 				default:
 					panic("unknown number of operands")
 				}
+			case llvm.Switch:
+				// A switch is an array of (value, label) pairs, of which the
+				// first one indicates the to-switch value and the default
+				// label.
+				numOperands := llvmInst.OperandsCount()
+				for i := 0; i < numOperands; i += 2 {
+					inst.operands = append(inst.operands, r.getValue(llvmInst.Operand(i)))
+					inst.operands = append(inst.operands, literalValue{uint32(blockIndices[llvmInst.Operand(i+1)])})
+				}
 			case llvm.PHI:
 				inst.name = llvmInst.Name()
 				incomingCount := inst.llvmInst.IncomingCount()
@@ -337,7 +347,13 @@ func (r *runner) compileFunction(llvmFn llvm.Value) *function {
 				// 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)
+			if inst.opcode == llvm.PHI {
+				// PHI nodes need to be treated specially, see the comment in
+				// interpreter.go for an explanation.
+				bb.phiNodes = append(bb.phiNodes, inst)
+			} else {
+				bb.instructions = append(bb.instructions, inst)
+			}
 		}
 	}
 	return fn
@@ -20,6 +20,11 @@ var (
 	errMapAlreadyCreated      = errors.New("interp: map already created")
 )

+// This is one of the errors that can be returned from toLLVMValue when the
+// passed type does not fit the data to serialize. It is recoverable by
+// serializing without a type (using rawValue.rawLLVMValue).
+var errInvalidPtrToIntSize = errors.New("interp: ptrtoint integer size does not equal pointer size")
+
 func isRecoverableError(err error) bool {
 	return err == errIntegerAsPointer || err == errUnsupportedInst || err == errUnsupportedRuntimeInst || err == errMapAlreadyCreated
 }
@@ -11,6 +11,12 @@ import (
 	"tinygo.org/x/go-llvm"
 )

+// Version of the interp package. It must be incremented whenever the interp
+// package is changed in a way that affects the output so that cached package
+// builds will be invalidated.
+// This version is independent of the TinyGo version number.
+const Version = 1
+
 // Enable extra checks, which should be disabled by default.
 // This may help track down bugs by adding a few more sanity checks.
 const checks = true
@@ -32,9 +38,7 @@ type runner struct {
 	callsExecuted uint64
 }

-// Run evaluates runtime.initAll function as much as possible at compile time.
-// Set debug to true if it should print output while running.
-func Run(mod llvm.Module, debug bool) error {
+func newRunner(mod llvm.Module, debug bool) *runner {
 	r := runner{
 		mod:           mod,
 		targetData:    llvm.NewTargetData(mod.DataLayout()),
@@ -47,6 +51,13 @@ func Run(mod llvm.Module, debug bool) error {
 	r.pointerSize = uint32(r.targetData.PointerSize())
 	r.i8ptrType = llvm.PointerType(mod.Context().Int8Type(), 0)
 	r.maxAlign = r.targetData.PrefTypeAlignment(r.i8ptrType) // assume pointers are maximally aligned (this is not always the case)
+	return &r
+}
+
+// Run evaluates runtime.initAll function as much as possible at compile time.
+// Set debug to true if it should print output while running.
+func Run(mod llvm.Module, debug bool) error {
+	r := newRunner(mod, debug)

 	initAll := mod.NamedFunction("runtime.initAll")
 	bb := initAll.EntryBasicBlock()
@@ -117,7 +128,7 @@ func Run(mod llvm.Module, debug bool) error {
 	r.pkgName = ""

 	// Update all global variables in the LLVM module.
-	mem := memoryView{r: &r}
+	mem := memoryView{r: r}
 	for _, obj := range r.objects {
 		if obj.llvmGlobal.IsNil() {
 			continue
@@ -125,7 +136,34 @@ func Run(mod llvm.Module, debug bool) error {
 		if obj.buffer == nil {
 			continue
 		}
-		initializer := obj.buffer.toLLVMValue(obj.llvmGlobal.Type().ElementType(), &mem)
+		initializer, err := obj.buffer.toLLVMValue(obj.llvmGlobal.Type().ElementType(), &mem)
+		if err == errInvalidPtrToIntSize {
+			// This can happen when a previous interp run did not have the
+			// correct LLVM type for a global and made something up. In that
+			// case, some fields could be written out as a series of (null)
+			// bytes even though they actually contain a pointer value.
+			// As a fallback, use asRawValue to get something of the correct
+			// memory layout.
+			initializer, err := obj.buffer.asRawValue(r).rawLLVMValue(&mem)
+			if err != nil {
+				return err
+			}
+			initializerType := initializer.Type()
+			newGlobal := llvm.AddGlobal(mod, initializerType, obj.llvmGlobal.Name()+".tmp")
+			newGlobal.SetInitializer(initializer)
+			newGlobal.SetLinkage(obj.llvmGlobal.Linkage())
+			newGlobal.SetAlignment(obj.llvmGlobal.Alignment())
+			// TODO: copy debug info, unnamed_addr, ...
+			bitcast := llvm.ConstBitCast(newGlobal, obj.llvmGlobal.Type())
+			obj.llvmGlobal.ReplaceAllUsesWith(bitcast)
+			name := obj.llvmGlobal.Name()
+			obj.llvmGlobal.EraseFromParentAsGlobal()
+			newGlobal.SetName(name)
+			continue
+		}
+		if err != nil {
+			return err
+		}
 		if checks && initializer.Type() != obj.llvmGlobal.Type().ElementType() {
 			panic("initializer type mismatch")
 		}
@@ -135,6 +173,87 @@ func Run(mod llvm.Module, debug bool) error {
 	return nil
 }

+// RunFunc evaluates a single package initializer at compile time.
+// Set debug to true if it should print output while running.
+func RunFunc(fn llvm.Value, debug bool) error {
+	// Create and initialize *runner object.
+	mod := fn.GlobalParent()
+	r := newRunner(mod, debug)
+	initName := fn.Name()
+	if !strings.HasSuffix(initName, ".init") {
+		return errorAt(fn, "interp: unexpected function name (expected *.init)")
+	}
+	r.pkgName = initName[:len(initName)-len(".init")]
+
+	// Create new function with the interp result.
+	newFn := llvm.AddFunction(mod, fn.Name()+".tmp", fn.Type().ElementType())
+	newFn.SetLinkage(fn.Linkage())
+	newFn.SetVisibility(fn.Visibility())
+	entry := mod.Context().AddBasicBlock(newFn, "entry")
+
+	// Create a builder, to insert instructions that could not be evaluated at
+	// compile time.
+	r.builder = mod.Context().NewBuilder()
+	defer r.builder.Dispose()
+	r.builder.SetInsertPointAtEnd(entry)
+
+	// Copy debug information.
+	subprogram := fn.Subprogram()
+	if !subprogram.IsNil() {
+		newFn.SetSubprogram(subprogram)
+		r.builder.SetCurrentDebugLocation(subprogram.SubprogramLine(), 0, subprogram, llvm.Metadata{})
+	}
+
+	// Run the initializer, filling the .init.tmp function.
+	if r.debug {
+		fmt.Fprintln(os.Stderr, "interp:", fn.Name())
+	}
+	_, pkgMem, callErr := r.run(r.getFunction(fn), nil, nil, "    ")
+	if callErr != nil {
+		if isRecoverableError(callErr.Err) {
+			// Could not finish, but could recover from it.
+			if r.debug {
+				fmt.Fprintln(os.Stderr, "not interpreting", r.pkgName, "because of error:", callErr.Error())
+			}
+			newFn.EraseFromParentAsFunction()
+			return nil
+		}
+		return callErr
+	}
+	for index, obj := range pkgMem.objects {
+		r.objects[index] = obj
+	}
+
+	// Update globals with values determined while running the initializer above.
+	mem := memoryView{r: r}
+	for _, obj := range r.objects {
+		if obj.llvmGlobal.IsNil() {
+			continue
+		}
+		if obj.buffer == nil {
+			continue
+		}
+		initializer, err := obj.buffer.toLLVMValue(obj.llvmGlobal.Type().ElementType(), &mem)
+		if err != nil {
+			return err
+		}
+		if checks && initializer.Type() != obj.llvmGlobal.Type().ElementType() {
+			panic("initializer type mismatch")
+		}
+		obj.llvmGlobal.SetInitializer(initializer)
+	}
+
+	// Finalize: remove the old init function and replace it with the new
+	// (.init.tmp) function.
+	r.builder.CreateRetVoid()
+	fnName := fn.Name()
+	fn.ReplaceAllUsesWith(newFn)
+	fn.EraseFromParentAsFunction()
+	newFn.SetName(fnName)
+
+	return nil
+}
+
 // getFunction returns the compiled version of the given LLVM function. It
 // compiles the function if necessary and caches the result.
 func (r *runner) getFunction(llvmFn llvm.Value) *function {
@@ -13,9 +13,9 @@ import (
 func TestInterp(t *testing.T) {
 	for _, name := range []string{
 		"basic",
+		"phi",
 		"slice-copy",
 		"consteval",
-		"map",
 		"interface",
 	} {
 		name := name // make tc local to this closure
@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"math"
 	"os"
+	"strconv"
 	"strings"
 	"time"

@@ -33,6 +34,50 @@ func (r *runner) run(fn *function, params []value, parentMem *memoryView, indent
 	lastBB := -1 // last basic block is undefined, only defined after a branch
 	var operands []value
 	for instIndex := 0; instIndex < len(bb.instructions); instIndex++ {
+		if instIndex == 0 {
+			// This is the start of a new basic block.
+			// There may be PHI nodes that need to be resolved. Resolve all PHI
+			// nodes before continuing with regular instructions.
+			// PHI nodes need to be treated specially because they can have a
+			// mutual dependency:
+			//   for.loop:
+			//     %a = phi i8 [ 1, %entry ], [ %b, %for.loop ]
+			//     %b = phi i8 [ 3, %entry ], [ %a, %for.loop ]
+			// If these PHI nodes are processed like a regular instruction, %a
+			// and %b are both 3 on the second iteration of the loop because %b
+			// loads the value of %a from the second iteration, while it should
+			// load the value from the previous iteration. The correct behavior
+			// is that these two values swap each others place on each
+			// iteration.
+			var phiValues []value
+			var phiIndices []int
+			for _, inst := range bb.phiNodes {
+				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")
+				}
+				phiValues = append(phiValues, result)
+				phiIndices = append(phiIndices, inst.localIndex)
+			}
+			for i, value := range phiValues {
+				locals[phiIndices[i]] = value
+			}
+		}
+
 		inst := bb.instructions[instIndex]
 		operands = operands[:0]
 		isRuntimeInst := false
@@ -103,27 +148,24 @@ func (r *runner) run(fn *function, params []value, parentMem *memoryView, indent
 			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
-					}
+		case llvm.Switch:
+			// Switch statement: [value, defaultLabel, case0, label0, case1, label1, ...]
+			value := operands[0].Uint()
+			targetLabel := operands[1].Uint() // default label
+			// Do a lazy switch by iterating over all cases.
+			for i := 2; i < len(operands); i += 2 {
+				if value == operands[i].Uint() {
+					targetLabel = operands[i+1].Uint()
 					break
 				}
 			}
+			lastBB = currentBB
+			currentBB = int(targetLabel)
+			bb = fn.blocks[currentBB]
+			instIndex = -1 // start at 0 the next cycle
 			if r.debug {
-				fmt.Fprintln(os.Stderr, indent+"phi", inst.operands, "->", result)
+				fmt.Fprintln(os.Stderr, indent+"switch", operands, "->", currentBB)
 			}
-			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.
@@ -155,17 +197,17 @@ func (r *runner) run(fn *function, params []value, parentMem *memoryView, indent
 				// which case this call won't even get to this point but will
 				// already be emitted in initAll.
 				continue
-			case callFn.name == "(reflect.Type).Elem" || strings.HasPrefix(callFn.name, "runtime.print") || callFn.name == "runtime._panic" || callFn.name == "runtime.hashmapGet":
+			case strings.HasPrefix(callFn.name, "runtime.print") || callFn.name == "runtime._panic" || callFn.name == "runtime.hashmapGet" || callFn.name == "os.runtime_args":
 				// 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.
+				//   * os.runtime_args reads globals that are initialized outside
+				//     the view of the interp package so it always needs to be run
+				//     at runtime.
 				err := r.runAtRuntime(fn, inst, locals, &mem, indent)
 				if err != nil {
 					return nil, mem, err
@@ -280,26 +322,58 @@ func (r *runner) run(fn *function, params []value, parentMem *memoryView, indent
 				copy(dstBuf.buf[dst.offset():dst.offset()+nBytes], srcBuf.buf[src.offset():])
 				dstObj.buffer = dstBuf
 				mem.put(dst.index(), dstObj)
+			case callFn.name == "(reflect.rawType).elem":
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"call (reflect.rawType).elem:", operands[1:])
+				}
+				// Extract the type code global from the first parameter.
+				typecodeIDPtrToInt, err := operands[1].toLLVMValue(inst.llvmInst.Operand(0).Type(), &mem)
+				if err != nil {
+					return nil, mem, r.errorAt(inst, err)
+				}
+				typecodeID := typecodeIDPtrToInt.Operand(0)
+
+				// Get the type class.
+				// See also: getClassAndValueFromTypeCode in transform/reflect.go.
+				typecodeName := typecodeID.Name()
+				const prefix = "reflect/types.type:"
+				if !strings.HasPrefix(typecodeName, prefix) {
+					panic("unexpected typecode name: " + typecodeName)
+				}
+				id := typecodeName[len(prefix):]
+				class := id[:strings.IndexByte(id, ':')]
+				value := id[len(class)+1:]
+				if class == "named" {
+					// Get the underlying type.
+					class = value[:strings.IndexByte(value, ':')]
+					value = value[len(class)+1:]
+				}
+
+				// Elem() is only valid for certain type classes.
+				switch class {
+				case "chan", "pointer", "slice", "array":
+					elementType := llvm.ConstExtractValue(typecodeID.Initializer(), []uint32{0})
+					uintptrType := r.mod.Context().IntType(int(mem.r.pointerSize) * 8)
+					locals[inst.localIndex] = r.getValue(llvm.ConstPtrToInt(elementType, uintptrType))
+				default:
+					return nil, mem, r.errorAt(inst, fmt.Errorf("(reflect.Type).Elem() called on %s type", class))
+				}
 			case callFn.name == "runtime.typeAssert":
 				// This function must be implemented manually as it is normally
 				// implemented by the interface lowering pass.
 				if r.debug {
 					fmt.Fprintln(os.Stderr, indent+"typeassert:", operands[1:])
 				}
-				typeInInterfacePtr, err := operands[1].asPointer(r)
+				assertedType, err := operands[2].toLLVMValue(inst.llvmInst.Operand(1).Type(), &mem)
 				if err != nil {
 					return nil, mem, r.errorAt(inst, err)
 				}
-				actualType, err := mem.load(typeInInterfacePtr, r.pointerSize).asPointer(r)
+				actualTypePtrToInt, err := operands[1].toLLVMValue(inst.llvmInst.Operand(0).Type(), &mem)
 				if err != nil {
 					return nil, mem, r.errorAt(inst, err)
 				}
-				assertedType, err := operands[2].asPointer(r)
-				if err != nil {
-					return nil, mem, r.errorAt(inst, err)
-				}
-				result := assertedType.asRawValue(r).equal(actualType.asRawValue(r))
-				if result {
+				actualType := actualTypePtrToInt.Operand(0)
+				if strings.TrimPrefix(actualType.Name(), "reflect/types.type:") == strings.TrimPrefix(assertedType.Name(), "reflect/types.typeid:") {
 					locals[inst.localIndex] = literalValue{uint8(1)}
 				} else {
 					locals[inst.localIndex] = literalValue{uint8(0)}
@@ -310,11 +384,11 @@ func (r *runner) run(fn *function, params []value, parentMem *memoryView, indent
 				}

 				// Load various values for the interface implements check below.
-				typeInInterfacePtr, err := operands[1].asPointer(r)
+				typecodePtr, err := operands[1].asPointer(r)
 				if err != nil {
 					return nil, mem, r.errorAt(inst, err)
 				}
-				methodSetPtr, err := mem.load(typeInInterfacePtr.addOffset(r.pointerSize), r.pointerSize).asPointer(r)
+				methodSetPtr, err := mem.load(typecodePtr.addOffset(r.pointerSize*2), r.pointerSize).asPointer(r)
 				if err != nil {
 					return nil, mem, r.errorAt(inst, err)
 				}
@@ -349,74 +423,43 @@ func (r *runner) run(fn *function, params []value, parentMem *memoryView, indent
 				}
 				// If assertOk is still 1, the assertion succeeded.
 				locals[inst.localIndex] = literalValue{assertOk}
-			case callFn.name == "runtime.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),
+			case callFn.name == "runtime.interfaceMethod":
+				// This builtin returns the function (which may be a thunk) to
+				// invoke a method on an interface. It does not call the method.
+				if r.debug {
+					fmt.Fprintln(os.Stderr, indent+"interface method:", operands[1:])
 				}
-				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)
+				// Load the first param, which is the type code (ptrtoint of the
+				// type code global).
+				typecodeIDPtrToInt, err := operands[1].toLLVMValue(inst.llvmInst.Operand(0).Type(), &mem)
 				if err != nil {
 					return nil, mem, r.errorAt(inst, err)
 				}
-				m := mem.getWritable(mapPtr.index()).buffer.(*mapValue)
-				keyPtr, err := operands[2].asPointer(r)
-				if err != nil {
-					return nil, mem, r.errorAt(inst, err)
+				typecodeID := typecodeIDPtrToInt.Operand(0).Initializer()
+
+				// Load the method set, which is part of the typecodeID object.
+				methodSet := llvm.ConstExtractValue(typecodeID, []uint32{2}).Operand(0).Initializer()
+
+				// We don't need to load the interface method set.
+
+				// Load the signature of the to-be-called function.
+				signature := inst.llvmInst.Operand(2)
+
+				// Iterate through all methods, looking for the one method that
+				// should be returned.
+				numMethods := methodSet.Type().ArrayLength()
+				var method llvm.Value
+				for i := 0; i < numMethods; i++ {
+					methodSignature := llvm.ConstExtractValue(methodSet, []uint32{uint32(i), 0})
+					if methodSignature == signature {
+						method = llvm.ConstExtractValue(methodSet, []uint32{uint32(i), 1}).Operand(0)
+					}
 				}
-				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)
+				if method.IsNil() {
+					return nil, mem, r.errorAt(inst, errors.New("could not find method: "+signature.Name()))
 				}
+				locals[inst.localIndex] = r.getValue(method)
 			default:
 				if len(callFn.blocks) == 0 {
 					// Call to a function declaration without a definition
@@ -482,6 +525,13 @@ func (r *runner) run(fn *function, params []value, parentMem *memoryView, indent
 				continue
 			}
 			result := mem.load(ptr, uint32(size))
+			if result == nil {
+				err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+				if err != nil {
+					return nil, mem, err
+				}
+				continue
+			}
 			if r.debug {
 				fmt.Fprintln(os.Stderr, indent+"load:", ptr, "->", result)
 			}
@@ -504,7 +554,14 @@ func (r *runner) run(fn *function, params []value, parentMem *memoryView, indent
 			if r.debug {
 				fmt.Fprintln(os.Stderr, indent+"store:", val, ptr)
 			}
-			mem.store(val, ptr)
+			ok := mem.store(val, ptr)
+			if !ok {
+				// Could not store the value, do it at runtime.
+				err := r.runAtRuntime(fn, inst, locals, &mem, indent)
+				if err != nil {
+					return nil, mem, err
+				}
+			}
 		case llvm.Alloca:
 			// Alloca normally allocates some stack memory. In the interpreter,
 			// it allocates a global instead.
@@ -836,7 +893,11 @@ func (r *runner) runAtRuntime(fn *function, inst instruction, locals []value, me
 	for i := 0; i < numOperands; i++ {
 		operand := inst.llvmInst.Operand(i)
 		if !operand.IsAInstruction().IsNil() || !operand.IsAArgument().IsNil() {
-			operand = locals[fn.locals[operand]].toLLVMValue(operand.Type(), mem)
+			var err error
+			operand, err = locals[fn.locals[operand]].toLLVMValue(operand.Type(), mem)
+			if err != nil {
+				return r.errorAt(inst, err)
+			}
 		}
 		operands[i] = operand
 	}
@@ -870,16 +931,31 @@ func (r *runner) runAtRuntime(fn *function, inst instruction, locals []value, me
 		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")
+		// Note: the Go LLVM API doesn't support multiple indices, so simulate
+		// this operation with some extra extractvalue instructions. Hopefully
+		// this is optimized to a single instruction.
+		agg := operands[0]
+		for i := 0; i < len(indices)-1; i++ {
+			agg = r.builder.CreateExtractValue(agg, int(indices[i]), inst.name+".agg")
 		}
-		result = r.builder.CreateExtractValue(operands[0], int(indices[0]), inst.name)
+		result = r.builder.CreateExtractValue(agg, int(indices[len(indices)-1]), inst.name)
 	case llvm.InsertValue:
 		indices := inst.llvmInst.Indices()
-		if len(indices) != 1 {
-			panic("expected exactly one index")
+		// Similar to extractvalue, we're working around a limitation in the Go
+		// LLVM API here by splitting the insertvalue into multiple instructions
+		// if there is more than one operand.
+		agg := operands[0]
+		aggregates := []llvm.Value{agg}
+		for i := 0; i < len(indices)-1; i++ {
+			agg = r.builder.CreateExtractValue(agg, int(indices[i]), inst.name+".agg"+strconv.Itoa(i))
+			aggregates = append(aggregates, agg)
 		}
-		result = r.builder.CreateInsertValue(operands[0], operands[1], int(indices[0]), inst.name)
+		result = operands[1]
+		for i := len(indices) - 1; i >= 0; i-- {
+			agg := aggregates[i]
+			result = r.builder.CreateInsertValue(agg, result, int(indices[i]), inst.name+".insertvalue"+strconv.Itoa(i))
+		}
+
 	case llvm.Add:
 		result = r.builder.CreateAdd(operands[0], operands[1], inst.name)
 	case llvm.Sub:
@@ -259,12 +259,17 @@ func (mv *memoryView) put(index uint32, obj object) {
 	mv.objects[index] = obj
 }

-// Load the value behind the given pointer.
+// Load the value behind the given pointer. Returns nil if the pointer points to
+// an external global.
 func (mv *memoryView) load(p pointerValue, size uint32) value {
 	if checks && mv.hasExternalStore(p) {
 		panic("interp: load from object with external store")
 	}
 	obj := mv.get(p.index())
+	if obj.buffer == nil {
+		// External global, return nil.
+		return nil
+	}
 	if p.offset() == 0 && size == obj.size {
 		return obj.buffer.clone()
 	}
@@ -280,12 +285,17 @@ func (mv *memoryView) load(p pointerValue, size uint32) value {

 // Store to the value behind the given pointer. This overwrites the value in the
 // memory view, so that the changed value is discarded when the memory view is
-// reverted.
-func (mv *memoryView) store(v value, p pointerValue) {
+// reverted. Returns true on success, false if the object to store to is
+// external.
+func (mv *memoryView) store(v value, p pointerValue) bool {
 	if checks && mv.hasExternalLoadOrStore(p) {
 		panic("interp: store to object with external load/store")
 	}
 	obj := mv.get(p.index())
+	if obj.buffer == nil {
+		// External global, return false (for a failure).
+		return false
+	}
 	if checks && p.offset()+v.len(mv.r) > obj.size {
 		panic("interp: store out of bounds")
 	}
@@ -301,6 +311,7 @@ func (mv *memoryView) store(v value, p pointerValue) {
 		}
 	}
 	mv.put(p.index(), obj)
+	return true // success
 }

 // value is some sort of value, comparable to a LLVM constant. It can be
@@ -314,7 +325,7 @@ type value interface {
 	asRawValue(*runner) rawValue
 	Uint() uint64
 	Int() int64
-	toLLVMValue(llvm.Type, *memoryView) llvm.Value
+	toLLVMValue(llvm.Type, *memoryView) (llvm.Value, error)
 	String() string
 }

@@ -405,25 +416,25 @@ func (v literalValue) Int() int64 {
 	}
 }

-func (v literalValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
+func (v literalValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) (llvm.Value, error) {
 	switch llvmType.TypeKind() {
 	case llvm.IntegerTypeKind:
 		switch value := v.value.(type) {
 		case uint64:
-			return llvm.ConstInt(llvmType, value, false)
+			return llvm.ConstInt(llvmType, value, false), nil
 		case uint32:
-			return llvm.ConstInt(llvmType, uint64(value), false)
+			return llvm.ConstInt(llvmType, uint64(value), false), nil
 		case uint16:
-			return llvm.ConstInt(llvmType, uint64(value), false)
+			return llvm.ConstInt(llvmType, uint64(value), false), nil
 		case uint8:
-			return llvm.ConstInt(llvmType, uint64(value), false)
+			return llvm.ConstInt(llvmType, uint64(value), false), nil
 		default:
-			panic("inpterp: unknown literal type")
+			return llvm.Value{}, errors.New("interp: unknown literal type")
 		}
 	case llvm.DoubleTypeKind:
-		return llvm.ConstFloat(llvmType, math.Float64frombits(v.value.(uint64)))
+		return llvm.ConstFloat(llvmType, math.Float64frombits(v.value.(uint64))), nil
 	case llvm.FloatTypeKind:
-		return llvm.ConstFloat(llvmType, float64(math.Float32frombits(v.value.(uint32))))
+		return llvm.ConstFloat(llvmType, float64(math.Float32frombits(v.value.(uint32)))), nil
 	default:
 		return v.asRawValue(mem.r).toLLVMValue(llvmType, mem)
 	}
@@ -507,7 +518,7 @@ func (v pointerValue) llvmValue(mem *memoryView) llvm.Value {
 // toLLVMValue returns the LLVM value for this pointer, which may be a GEP or
 // bitcast. The llvm.Type parameter is optional, if omitted the pointer type may
 // be different than expected.
-func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
+func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) (llvm.Value, error) {
 	// Obtain the llvmValue, creating it if it doesn't exist yet.
 	llvmValue := v.llvmValue(mem)
 	if llvmValue.IsNil() {
@@ -518,7 +529,10 @@ func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Valu
 		if obj.llvmType.IsNil() {
 			// Create an initializer without knowing the global type.
 			// This is probably the result of a runtime.alloc call.
-			initializer := obj.buffer.asRawValue(mem.r).rawLLVMValue(mem)
+			initializer, err := obj.buffer.asRawValue(mem.r).rawLLVMValue(mem)
+			if err != nil {
+				return llvm.Value{}, err
+			}
 			globalType := initializer.Type()
 			llvmValue = llvm.AddGlobal(mem.r.mod, globalType, obj.globalName)
 			llvmValue.SetInitializer(initializer)
@@ -537,9 +551,12 @@ func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Valu
 			// Set the initializer for the global. Do this after creation to avoid
 			// infinite recursion between creating the global and creating the
 			// contents of the global (if the global contains itself).
-			initializer := obj.buffer.toLLVMValue(globalType, mem)
+			initializer, err := obj.buffer.toLLVMValue(globalType, mem)
+			if err != nil {
+				return llvm.Value{}, err
+			}
 			if checks && initializer.Type() != globalType {
-				panic("allocated value does not match allocated type")
+				return llvm.Value{}, errors.New("interp: allocated value does not match allocated type")
 			}
 			llvmValue.SetInitializer(initializer)
 		}
@@ -552,7 +569,7 @@ func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Valu
 	}

 	if llvmType.IsNil() {
-		return llvmValue
+		return llvmValue, nil
 	}

 	if llvmType.TypeKind() != llvm.PointerTypeKind {
@@ -564,7 +581,7 @@ func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Valu
 		// This can be worked around by simply converting to a raw value,
 		// rawValue knows how to create such structs.
 		if v.offset() != 0 {
-			panic("offset set without known pointer type")
+			return llvm.Value{}, errors.New("interp: offset set without known pointer type")
 		}
 		return v.asRawValue(mem.r).toLLVMValue(llvmType, mem)
 	}
@@ -575,14 +592,14 @@ func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Valu
 		if v.offset() != 0 {
 			// This should never happen, if offset is non-zero, the types
 			// shouldn't match.
-			panic("offset set while there is no way to convert the type")
+			return llvm.Value{}, errors.New("interp: offset set while there is no way to convert the type")
 		}
-		return llvmValue
+		return llvmValue, nil
 	}

 	if v.offset() == 0 {
 		// Offset is zero, so we can just bitcast to get a correct pointer.
-		return llvm.ConstBitCast(llvmValue, llvmType)
+		return llvm.ConstBitCast(llvmValue, llvmType), nil
 	}

 	// We need to make a constant GEP for pointer arithmetic.
@@ -606,11 +623,11 @@ func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Valu
 			offset -= int64(mem.r.targetData.ElementOffset(objectElementType, element))
 			objectElementType = objectElementType.StructElementTypes()[element]
 		default:
-			panic("pointer index with something other than a struct or array?")
+			return llvm.Value{}, errors.New("interp: pointer index with something other than a struct or array?")
 		}
 	}
 	if offset < 0 {
-		panic("offset has somehow gone negative, this should be impossible")
+		return llvm.Value{}, errors.New("interp: offset has somehow gone negative, this should be impossible")
 	}

 	// Finally do the gep, using the above computed indices.
@@ -618,283 +635,9 @@ func (v pointerValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Valu
 	// the bits of the pointer are now correct, just not the type).
 	gep := llvm.ConstInBoundsGEP(llvmValue, indices)
 	if gep.Type() != llvmType {
-		return llvm.ConstBitCast(gep, llvmType)
+		return llvm.ConstBitCast(gep, llvmType), nil
 	}
-	return gep
-}
-
-// mapValue implements a Go map which is created at compile time and stored as a
-// global variable.
-// The value itself is only used as part of an object (object.buffer). Maps are
-// reference types aka pointers, so it can only be used as a pointerValue, not
-// directly.
-type mapValue struct {
-	r           *runner
-	pkgName     string
-	size        uint32 // byte size of runtime.hashmap
-	hashmap     llvm.Value
-	keyIsString bool
-	keys        []interface{} // either rawValue (for binary key) or mapStringKey (for string key)
-	values      []rawValue
-	keySize     uint32
-	valueSize   uint32
-}
-
-type mapStringKey struct {
-	buf  pointerValue
-	size uint64
-	data []uint64
-}
-
-func newMapValue(r *runner, hashmapPointerType llvm.Type, keySize, valueSize uint32) *mapValue {
-	size := uint32(r.targetData.TypeAllocSize(hashmapPointerType.ElementType()))
-	return &mapValue{
-		r:         r,
-		pkgName:   r.pkgName,
-		size:      size,
-		keySize:   keySize,
-		valueSize: valueSize,
-	}
-}
-
-func (v *mapValue) len(r *runner) uint32 {
-	return v.size
-}
-
-func (v *mapValue) clone() value {
-	// Return a copy of mapValue.
-	clone := *v
-	clone.keys = append([]interface{}{}, clone.keys...)
-	clone.values = append([]rawValue{}, clone.values...)
-	return &clone
-}
-
-func (v *mapValue) asPointer(r *runner) (pointerValue, error) {
-	panic("interp: mapValue.asPointer")
-}
-
-func (v *mapValue) asRawValue(r *runner) rawValue {
-	panic("interp: mapValue.asRawValue")
-}
-
-func (v *mapValue) Uint() uint64 {
-	panic("interp: mapValue.Uint")
-}
-
-func (v *mapValue) Int() int64 {
-	panic("interp: mapValue.Int")
-}
-
-// Temporary struct to collect data before turning this into a hashmap bucket
-// LLVM value.
-type mapBucket struct {
-	m       *mapValue
-	tophash [8]uint8
-	keys    []rawValue // can have up to 8 keys
-	values  []rawValue // can have up to 8 values, len(keys) == len(values)
-}
-
-// create returns a (pointer to a) buffer structurally equivalent to
-// runtime.hashmapBucket.
-func (b *mapBucket) create(ctx llvm.Context, nextBucket llvm.Value, mem *memoryView) llvm.Value {
-	// Create tophash array.
-	int8Type := ctx.Int8Type()
-	tophashValues := make([]llvm.Value, 8)
-	for i := range tophashValues {
-		tophashValues[i] = llvm.ConstInt(int8Type, uint64(b.tophash[i]), false)
-	}
-	tophash := llvm.ConstArray(int8Type, tophashValues)
-
-	// Create next pointer (if not set).
-	if nextBucket.IsNil() {
-		nextBucket = llvm.ConstNull(llvm.PointerType(int8Type, 0))
-	}
-
-	// Create data for keys.
-	var keyValues []llvm.Value
-	for _, key := range b.keys {
-		keyValues = append(keyValues, key.rawLLVMValue(mem))
-	}
-	if len(b.keys) < 8 {
-		keyValues = append(keyValues, llvm.ConstNull(llvm.ArrayType(int8Type, int(b.m.keySize)*(8-len(b.keys)))))
-	}
-	keyValue := ctx.ConstStruct(keyValues, false)
-	if checks && uint32(b.m.r.targetData.TypeAllocSize(keyValue.Type())) != b.m.keySize*8 {
-		panic("key size invalid")
-	}
-
-	// Create data for values.
-	var valueValues []llvm.Value
-	for _, value := range b.values {
-		valueValues = append(valueValues, value.rawLLVMValue(mem))
-	}
-	if len(b.values) < 8 {
-		valueValues = append(valueValues, llvm.ConstNull(llvm.ArrayType(int8Type, int(b.m.valueSize)*(8-len(b.values)))))
-	}
-	valueValue := ctx.ConstStruct(valueValues, false)
-	if checks && uint32(b.m.r.targetData.TypeAllocSize(valueValue.Type())) != b.m.valueSize*8 {
-		panic("value size invalid")
-	}
-
-	// Create the bucket.
-	bucketInitializer := ctx.ConstStruct([]llvm.Value{
-		tophash,
-		nextBucket,
-		keyValue,
-		valueValue,
-	}, false)
-	bucket := llvm.AddGlobal(b.m.r.mod, bucketInitializer.Type(), b.m.pkgName+"$mapbucket")
-	bucket.SetInitializer(bucketInitializer)
-	bucket.SetLinkage(llvm.InternalLinkage)
-	bucket.SetUnnamedAddr(true)
-	return bucket
-}
-
-func (v *mapValue) toLLVMValue(hashmapType llvm.Type, mem *memoryView) llvm.Value {
-	if !v.hashmap.IsNil() {
-		return v.hashmap
-	}
-
-	// Create a slice of buckets with all the keys and values in the hashmap.
-	var buckets []*mapBucket
-	var bucket *mapBucket
-	for i, key := range v.keys {
-		var data []uint64
-		var keyValue rawValue
-		switch key := key.(type) {
-		case mapStringKey:
-			data = key.data
-			keyValue = newRawValue(v.keySize)
-			// runtime._string is {ptr, length}
-			for i := uint32(0); i < v.keySize/2; i++ {
-				keyValue.buf[i] = key.buf.pointer
-			}
-			copy(keyValue.buf[v.keySize/2:], literalValue{key.size}.asRawValue(v.r).buf)
-		case rawValue:
-			if key.hasPointer() {
-				panic("todo: map key with pointer")
-			}
-			data = key.buf
-			keyValue = key
-		default:
-			panic("unknown map key type")
-		}
-		buf := make([]byte, len(data))
-		for i, p := range data {
-			buf[i] = byte(p)
-		}
-		hash := v.hash(buf)
-
-		if i%8 == 0 {
-			bucket = &mapBucket{m: v}
-			buckets = append(buckets, bucket)
-		}
-		bucket.tophash[i%8] = v.topHash(hash)
-		bucket.keys = append(bucket.keys, keyValue)
-		bucket.values = append(bucket.values, v.values[i])
-	}
-
-	// Convert these buckets into LLVM global variables.
-	ctx := v.r.mod.Context()
-	var nextBucket llvm.Value
-	for i := len(buckets) - 1; i >= 0; i-- {
-		bucket = buckets[i]
-		bucketValue := bucket.create(ctx, nextBucket, mem)
-		nextBucket = bucketValue
-	}
-	firstBucket := nextBucket
-	if firstBucket.IsNil() {
-		firstBucket = llvm.ConstNull(mem.r.i8ptrType)
-	} else {
-		firstBucket = llvm.ConstBitCast(firstBucket, mem.r.i8ptrType)
-	}
-
-	// Create the hashmap itself, pointing to these buckets.
-	hashmapPointerType := llvm.PointerType(hashmapType, 0)
-	hashmap := llvm.ConstNamedStruct(hashmapType, []llvm.Value{
-		llvm.ConstPointerNull(hashmapPointerType), // next
-		firstBucket, // 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
-	})
-
-	v.hashmap = hashmap
-	return v.hashmap
-}
-
-// putString does a map assign operation, assuming that the map is of type
-// map[string]T.
-func (v *mapValue) putString(mem *memoryView, stringBuf pointerValue, stringLen uint64, valuePtr pointerValue) error {
-	if !v.hashmap.IsNil() {
-		return errMapAlreadyCreated
-	}
-
-	value := mem.load(valuePtr, v.valueSize)
-	stringValue := mem.load(stringBuf, uint32(stringLen)).asRawValue(v.r)
-	if stringValue.hasPointer() {
-		panic("interp: string contains pointer")
-	}
-
-	// TODO: avoid duplicate keys
-	v.keys = append(v.keys, mapStringKey{stringBuf, stringLen, stringValue.buf})
-	v.values = append(v.values, value.asRawValue(v.r))
-	v.keyIsString = true
-
-	return nil
-}
-
-// putBinary does a map assign operation for binary data (e.g. [3]int etc). The
-// key must not contain pointer values.
-func (v *mapValue) putBinary(mem *memoryView, keyPtr, valuePtr pointerValue) error {
-	if !v.hashmap.IsNil() {
-		return errMapAlreadyCreated
-	}
-
-	key := mem.load(keyPtr, v.keySize)
-	value := mem.load(valuePtr, v.valueSize)
-
-	// Sanity checks.
-	if v.keySize != key.len(mem.r) || v.valueSize != value.len(mem.r) {
-		// This is a bug (not unhandled input), so panic.
-		panic("interp: key or value size mismatch")
-	}
-	if v.keyIsString {
-		panic("cannot put binary keys in string map")
-	}
-
-	// TODO: avoid duplicate keys
-	v.keys = append(v.keys, key.asRawValue(v.r))
-	v.values = append(v.values, value.asRawValue(v.r))
-
-	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++
-	}
-	return tophash
-}
-
-func (v *mapValue) String() string {
-	return "<map keySize=" + strconv.Itoa(int(v.keySize)) + " valueSize=" + strconv.Itoa(int(v.valueSize)) + ">"
+	return gep, nil
 }

 // rawValue is a raw memory buffer that can store either pointers or regular
@@ -1018,7 +761,7 @@ func (v rawValue) equal(rhs rawValue) bool {
 // goes. The resulting value does not have a specified type, but it will be the
 // same size and have the same bytes if it was created with a provided LLVM type
 // (through toLLVMValue).
-func (v rawValue) rawLLVMValue(mem *memoryView) llvm.Value {
+func (v rawValue) rawLLVMValue(mem *memoryView) (llvm.Value, error) {
 	var structFields []llvm.Value
 	ctx := mem.r.mod.Context()
 	int8Type := ctx.Int8Type()
@@ -1042,13 +785,16 @@ func (v rawValue) rawLLVMValue(mem *memoryView) llvm.Value {
 	for i := uint32(0); i < uint32(len(v.buf)); {
 		if v.buf[i] > 255 {
 			addBytes()
-			field := pointerValue{v.buf[i]}.toLLVMValue(llvm.Type{}, mem)
+			field, err := pointerValue{v.buf[i]}.toLLVMValue(llvm.Type{}, mem)
+			if err != nil {
+				return llvm.Value{}, err
+			}
 			elementType := field.Type().ElementType()
 			if elementType.TypeKind() == llvm.StructTypeKind {
 				// There are some special pointer types that should be used as a
 				// ptrtoint, so that they can be used in certain optimizations.
 				name := elementType.StructName()
-				if name == "runtime.typeInInterface" || name == "runtime.funcValueWithSignature" {
+				if name == "runtime.typecodeID" || name == "runtime.funcValueWithSignature" {
 					uintptrType := ctx.IntType(int(mem.r.pointerSize) * 8)
 					field = llvm.ConstPtrToInt(field, uintptrType)
 				}
@@ -1065,12 +811,12 @@ func (v rawValue) rawLLVMValue(mem *memoryView) llvm.Value {

 	// Return the created data.
 	if len(structFields) == 1 {
-		return structFields[0]
+		return structFields[0], nil
 	}
-	return ctx.ConstStruct(structFields, false)
+	return ctx.ConstStruct(structFields, false), nil
 }

-func (v rawValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
+func (v rawValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) (llvm.Value, error) {
 	isZero := true
 	for _, p := range v.buf {
 		if p != 0 {
@@ -1079,7 +825,7 @@ func (v rawValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
 		}
 	}
 	if isZero {
-		return llvm.ConstNull(llvmType)
+		return llvm.ConstNull(llvmType), nil
 	}
 	switch llvmType.TypeKind() {
 	case llvm.IntegerTypeKind:
@@ -1088,7 +834,17 @@ func (v rawValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
 			if err != nil {
 				panic(err)
 			}
-			return llvm.ConstPtrToInt(ptr.toLLVMValue(llvm.Type{}, mem), llvmType)
+			if checks && mem.r.targetData.TypeAllocSize(llvmType) != mem.r.targetData.TypeAllocSize(mem.r.i8ptrType) {
+				// Probably trying to serialize a pointer to a byte array,
+				// perhaps as a result of rawLLVMValue() in a previous interp
+				// run.
+				return llvm.Value{}, errInvalidPtrToIntSize
+			}
+			v, err := ptr.toLLVMValue(llvm.Type{}, mem)
+			if err != nil {
+				return llvm.Value{}, err
+			}
+			return llvm.ConstPtrToInt(v, llvmType), nil
 		}
 		var n uint64
 		switch llvmType.IntTypeWidth() {
@@ -1108,7 +864,7 @@ func (v rawValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
 		default:
 			panic("unknown integer size")
 		}
-		return llvm.ConstInt(llvmType, n, false)
+		return llvm.ConstInt(llvmType, n, false), nil
 	case llvm.StructTypeKind:
 		fieldTypes := llvmType.StructElementTypes()
 		fields := make([]llvm.Value, len(fieldTypes))
@@ -1117,12 +873,16 @@ func (v rawValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
 			field := rawValue{
 				buf: v.buf[offset:],
 			}
-			fields[i] = field.toLLVMValue(fieldType, mem)
+			var err error
+			fields[i], err = field.toLLVMValue(fieldType, mem)
+			if err != nil {
+				return llvm.Value{}, err
+			}
 		}
 		if llvmType.StructName() != "" {
-			return llvm.ConstNamedStruct(llvmType, fields)
+			return llvm.ConstNamedStruct(llvmType, fields), nil
 		}
-		return llvmType.Context().ConstStruct(fields, false)
+		return llvmType.Context().ConstStruct(fields, false), nil
 	case llvm.ArrayTypeKind:
 		numElements := llvmType.ArrayLength()
 		childType := llvmType.ElementType()
@@ -1133,27 +893,34 @@ func (v rawValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
 			field := rawValue{
 				buf: v.buf[offset:],
 			}
-			fields[i] = field.toLLVMValue(childType, mem)
+			var err error
+			fields[i], err = field.toLLVMValue(childType, mem)
+			if err != nil {
+				return llvm.Value{}, err
+			}
 			if checks && fields[i].Type() != childType {
 				panic("child type doesn't match")
 			}
 		}
-		return llvm.ConstArray(childType, fields)
+		return llvm.ConstArray(childType, fields), nil
 	case llvm.PointerTypeKind:
 		if v.buf[0] > 255 {
 			// This is a regular pointer.
-			llvmValue := pointerValue{v.buf[0]}.toLLVMValue(llvm.Type{}, mem)
+			llvmValue, err := pointerValue{v.buf[0]}.toLLVMValue(llvm.Type{}, mem)
+			if err != nil {
+				return llvm.Value{}, err
+			}
 			if llvmValue.Type() != llvmType {
 				llvmValue = llvm.ConstBitCast(llvmValue, llvmType)
 			}
-			return llvmValue
+			return llvmValue, nil
 		}
 		// This is either a null pointer or a raw pointer for memory-mapped I/O
 		// (such as 0xe000ed00).
 		ptr := rawValue{v.buf[:mem.r.pointerSize]}.Uint()
 		if ptr == 0 {
 			// Null pointer.
-			return llvm.ConstNull(llvmType)
+			return llvm.ConstNull(llvmType), nil
 		}
 		var ptrValue llvm.Value // the underlying int
 		switch mem.r.pointerSize {
@@ -1164,19 +931,19 @@ func (v rawValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
 		case 2:
 			ptrValue = llvm.ConstInt(llvmType.Context().Int16Type(), ptr, false)
 		default:
-			panic("unknown pointer size")
+			return llvm.Value{}, errors.New("interp: unknown pointer size")
 		}
-		return llvm.ConstIntToPtr(ptrValue, llvmType)
+		return llvm.ConstIntToPtr(ptrValue, llvmType), nil
 	case llvm.DoubleTypeKind:
 		b := rawValue{v.buf[:8]}.Uint()
 		f := math.Float64frombits(b)
-		return llvm.ConstFloat(llvmType, f)
+		return llvm.ConstFloat(llvmType, f), nil
 	case llvm.FloatTypeKind:
 		b := uint32(rawValue{v.buf[:4]}.Uint())
 		f := math.Float32frombits(b)
-		return llvm.ConstFloat(llvmType, float64(f))
+		return llvm.ConstFloat(llvmType, float64(f)), nil
 	default:
-		panic("todo: raw value to LLVM value: " + llvmType.String())
+		return llvm.Value{}, errors.New("interp: todo: raw value to LLVM value: " + llvmType.String())
 	}
 }

@@ -1365,8 +1132,8 @@ func (v localValue) Int() int64 {
 	panic("interp: localValue.Int")
 }

-func (v localValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) llvm.Value {
-	return v.value
+func (v localValue) toLLVMValue(llvmType llvm.Type, mem *memoryView) (llvm.Value, error) {
+	return v.value, nil
 }

 func (r *runner) getValue(llvmValue llvm.Value) value {
@@ -8,6 +8,7 @@ target triple = "x86_64--linux"
@main.exportedValue = global [1 x i16*] [i16* @main.exposedValue1]
@main.exposedValue1 = global i16 0
@main.exposedValue2 = global i16 0
+@main.insertedValue = global {i8, i32, {float, {i64, i16}}} zeroinitializer

 declare void @runtime.printint64(i64) unnamed_addr

@@ -65,6 +66,19 @@ entry:
  call void @modifyExternal(i32* bitcast (void ()* @willModifyGlobal to i32*))
  store i16 7, i16* @main.exposedValue2

+  ; Test switch statement.
+  %switch1 = call i64 @testSwitch(i64 1) ; 1 returns 6
+  %switch2 = call i64 @testSwitch(i64 9) ; 9 returns the default value -1
+  call void @runtime.printint64(i64 %switch1)
+  call void @runtime.printint64(i64 %switch2)
+
+  ; Test extractvalue/insertvalue with multiple operands.
+  %agg = call {i8, i32, {float, {i64, i16}}} @nestedStruct()
+  %elt = extractvalue {i8, i32, {float, {i64, i16}}} %agg, 2, 1, 0
+  call void @runtime.printint64(i64 %elt)
+  %agg2 = insertvalue {i8, i32, {float, {i64, i16}}} %agg, i64 5, 2, 1, 0
+  store {i8, i32, {float, {i64, i16}}} %agg2, {i8, i32, {float, {i64, i16}}}* @main.insertedValue
+
  ret void
 }

@@ -87,3 +101,24 @@ entry:
  store i16 8, i16* @main.exposedValue2
  ret void
 }
+
+define i64 @testSwitch(i64 %val) {
+entry:
+  ; Test switch statement.
+  switch i64 %val, label %otherwise [ i64 0, label %zero
+                                      i64 1, label %one
+                                      i64 2, label %two ]
+zero:
+  ret i64 5
+
+one:
+  ret i64 6
+
+two:
+  ret i64 7
+
+otherwise:
+  ret i64 -1
+}
+
+declare {i8, i32, {float, {i64, i16}}} @nestedStruct()
@@ -7,6 +7,7 @@ target triple = "x86_64--linux"
@main.exportedValue = global [1 x i16*] [i16* @main.exposedValue1]
@main.exposedValue1 = global i16 0
@main.exposedValue2 = local_unnamed_addr global i16 0
+@main.insertedValue = local_unnamed_addr global { i8, i32, { float, { i64, i16 } } } zeroinitializer

 declare void @runtime.printint64(i64) unnamed_addr

@@ -25,6 +26,19 @@ entry:
  store i16 5, i16* @main.exposedValue1
  call void @modifyExternal(i32* bitcast (void ()* @willModifyGlobal to i32*))
  store i16 7, i16* @main.exposedValue2
+  call void @runtime.printint64(i64 6)
+  call void @runtime.printint64(i64 -1)
+  %agg = call { i8, i32, { float, { i64, i16 } } } @nestedStruct()
+  %elt.agg = extractvalue { i8, i32, { float, { i64, i16 } } } %agg, 2
+  %elt.agg1 = extractvalue { float, { i64, i16 } } %elt.agg, 1
+  %elt = extractvalue { i64, i16 } %elt.agg1, 0
+  call void @runtime.printint64(i64 %elt)
+  %agg2.agg0 = extractvalue { i8, i32, { float, { i64, i16 } } } %agg, 2
+  %agg2.agg1 = extractvalue { float, { i64, i16 } } %agg2.agg0, 1
+  %agg2.insertvalue2 = insertvalue { i64, i16 } %agg2.agg1, i64 5, 0
+  %agg2.insertvalue1 = insertvalue { float, { i64, i16 } } %agg2.agg0, { i64, i16 } %agg2.insertvalue2, 1
+  %agg2.insertvalue0 = insertvalue { i8, i32, { float, { i64, i16 } } } %agg, { float, { i64, i16 } } %agg2.insertvalue1, 2
+  store { i8, i32, { float, { i64, i16 } } } %agg2.insertvalue0, { i8, i32, { float, { i64, i16 } } }* @main.insertedValue
  ret void
 }

@@ -44,3 +58,26 @@ entry:
  store i16 8, i16* @main.exposedValue2
  ret void
 }
+
+define i64 @testSwitch(i64 %val) local_unnamed_addr {
+entry:
+  switch i64 %val, label %otherwise [
+    i64 0, label %zero
+    i64 1, label %one
+    i64 2, label %two
+  ]
+
+zero:                                             ; preds = %entry
+  ret i64 5
+
+one:                                              ; preds = %entry
+  ret i64 6
+
+two:                                              ; preds = %entry
+  ret i64 7
+
+otherwise:                                        ; preds = %entry
+  ret i64 -1
+}
+
+declare { i8, i32, { float, { i64, i16 } } } @nestedStruct() local_unnamed_addr
@@ -1,17 +1,16 @@
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64--linux"

-%runtime.typecodeID = type { %runtime.typecodeID*, i64 }
+%runtime.typecodeID = type { %runtime.typecodeID*, i64, %runtime.interfaceMethodInfo* }
 %runtime.interfaceMethodInfo = type { i8*, i64 }
-%runtime.typeInInterface = type { %runtime.typecodeID*, %runtime.interfaceMethodInfo* }

@main.v1 = global i1 0
-@"reflect/types.type:named:main.foo" = private constant %runtime.typecodeID { %runtime.typecodeID* @"reflect/types.type:basic:int", i64 0 }
+@"reflect/types.type:named:main.foo" = private constant %runtime.typecodeID { %runtime.typecodeID* @"reflect/types.type:basic:int", i64 0, %runtime.interfaceMethodInfo* null }
+@"reflect/types.typeid:named:main.foo" = external constant i8
@"reflect/types.type:basic:int" = external constant %runtime.typecodeID
-@"typeInInterface:reflect/types.type:named:main.foo" = private constant %runtime.typeInInterface { %runtime.typecodeID* @"reflect/types.type:named:main.foo", %runtime.interfaceMethodInfo* null }


-declare i1 @runtime.typeAssert(i64, %runtime.typecodeID*, i8*, i8*)
+declare i1 @runtime.typeAssert(i64, i8*, i8*, i8*)

 define void @runtime.initAll() unnamed_addr {
 entry:
@@ -22,7 +21,7 @@ entry:
 define internal void @main.init() unnamed_addr {
 entry:
  ; Test type asserts.
-  %typecode = call i1 @runtime.typeAssert(i64 ptrtoint (%runtime.typeInInterface* @"typeInInterface:reflect/types.type:named:main.foo" to i64), %runtime.typecodeID* @"reflect/types.type:named:main.foo", i8* undef, i8* null)
+  %typecode = call i1 @runtime.typeAssert(i64 ptrtoint (%runtime.typecodeID* @"reflect/types.type:named:main.foo" to i64), i8* @"reflect/types.typeid:named:main.foo", i8* undef, i8* null)
  store i1 %typecode, i1* @main.v1
  ret void
 }
@@ -1,74 +0,0 @@
-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._string = type { i8*, i32 }
-%runtime.hashmap = type { %runtime.hashmap*, i8*, i32, i8, i8, i8 }
-
-@main.m = global %runtime.hashmap* null
-@main.binaryMap = global %runtime.hashmap* null
-@main.stringMap = global %runtime.hashmap* null
-@main.init.string = internal unnamed_addr constant [7 x i8] c"CONNECT"
-
-declare %runtime.hashmap* @runtime.hashmapMake(i8, i8, i32, i8* %context, i8* %parentHandle)
-declare void @runtime.hashmapBinarySet(%runtime.hashmap*, i8*, i8*, i8* %context, i8* %parentHandle)
-declare void @runtime.hashmapStringSet(%runtime.hashmap*, i8*, i32, i8*, i8* %context, i8* %parentHandle)
-declare void @llvm.lifetime.end.p0i8(i64, i8*)
-declare void @llvm.lifetime.start.p0i8(i64, i8*)
-
-define void @runtime.initAll() unnamed_addr {
-entry:
-  call void @main.init(i8* undef, i8* null)
-  ret void
-}
-
-define internal void @main.init(i8* %context, i8* %parentHandle) unnamed_addr {
-entry:
-; Test that hashmap optimizations generally work (even with lifetimes).
-  %hashmap.key = alloca i8
-  %hashmap.value = alloca %runtime._string
-  %0 = call %runtime.hashmap* @runtime.hashmapMake(i8 1, i8 8, i32 1, i8* undef, i8* null)
-  %hashmap.value.bitcast = bitcast %runtime._string* %hashmap.value to i8*
-  call void @llvm.lifetime.start.p0i8(i64 8, i8* %hashmap.value.bitcast)
-  store %runtime._string { i8* getelementptr inbounds ([7 x i8], [7 x i8]* @main.init.string, i32 0, i32 0), i32 7 }, %runtime._string* %hashmap.value
-  call void @llvm.lifetime.start.p0i8(i64 1, i8* %hashmap.key)
-  store i8 1, i8* %hashmap.key
-  call void @runtime.hashmapBinarySet(%runtime.hashmap* %0, i8* %hashmap.key, i8* %hashmap.value.bitcast, i8* undef, i8* null)
-  call void @llvm.lifetime.end.p0i8(i64 1, i8* %hashmap.key)
-  call void @llvm.lifetime.end.p0i8(i64 8, i8* %hashmap.value.bitcast)
-  store %runtime.hashmap* %0, %runtime.hashmap** @main.m
-
-  ; Other tests, that can be done in a separate function.
-  call void @main.testNonConstantBinarySet()
-  call void @main.testNonConstantStringSet()
-  ret void
-}
-
-; Test that a map loaded from a global can still be used for mapassign
-; operations (with binary keys).
-define internal void @main.testNonConstantBinarySet() {
-  %hashmap.key = alloca i8
-  %hashmap.value = alloca i8
-  ; 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
-  ; Do the binary set to the newly loaded map.
-  store i8 1, i8* %hashmap.key
-  store i8 2, i8* %hashmap.value
-  call void @runtime.hashmapBinarySet(%runtime.hashmap* %map, i8* %hashmap.key, i8* %hashmap.value, i8* undef, i8* null)
-  ret void
-}
-
-; Test that a map loaded from a global can still be used for mapassign
-; operations (with string keys).
-define internal void @main.testNonConstantStringSet() {
-  %hashmap.value = alloca i8
-  ; 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
-  ; Do the string set to the newly loaded map.
-  store i8 2, i8* %hashmap.value
-  call void @runtime.hashmapStringSet(%runtime.hashmap* %map, i8* getelementptr inbounds ([7 x i8], [7 x i8]* @main.init.string, i32 0, i32 0), i32 7, i8* %hashmap.value, i8* undef, i8* null)
-  ret void
-}
@@ -1,20 +0,0 @@
-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 }
-
-@main.m = local_unnamed_addr global %runtime.hashmap* @"main$map"
-@main.binaryMap = local_unnamed_addr global %runtime.hashmap* @"main$map.1"
-@main.stringMap = local_unnamed_addr global %runtime.hashmap* @"main$map.3"
-@main.init.string = internal unnamed_addr constant [7 x i8] c"CONNECT"
-@"main$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:
-  ret void
-}
@@ -0,0 +1,31 @@
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64--linux"
+
+@main.phiNodesResultA = global i8 0
+@main.phiNodesResultB = global i8 0
+
+define void @runtime.initAll() {
+  call void @main.init()
+  ret void
+}
+
+; PHI nodes always use the value from the previous block, even in a loop. This
+; means that the loop below should swap the values %a and %b on each iteration.
+; Previously there was a bug which resulted in %b getting the value 3 on the
+; second iteration while it should have gotten 1 (from the first iteration of
+; %for.loop).
+define internal void @main.init() {
+entry:
+  br label %for.loop
+
+for.loop:
+  %a = phi i8 [ 1, %entry ], [ %b, %for.loop ]
+  %b = phi i8 [ 3, %entry ], [ %a, %for.loop ]
+  %icmp = icmp eq i8 %a, 3
+  br i1 %icmp, label %for.done, label %for.loop
+
+for.done:
+  store i8 %a, i8* @main.phiNodesResultA
+  store i8 %b, i8* @main.phiNodesResultB
+  ret void
+}
@@ -0,0 +1,9 @@
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64--linux"
+
+@main.phiNodesResultA = local_unnamed_addr global i8 3
+@main.phiNodesResultB = local_unnamed_addr global i8 1
+
+define void @runtime.initAll() local_unnamed_addr {
+  ret void
+}
@@ -23,3 +23,13 @@ type Error struct {
 func (e Error) Error() string {
 	return e.Err.Error()
 }
+
+// Error returned when loading a *Program for a test binary but no test files
+// are present.
+type NoTestFilesError struct {
+	ImportPath string
+}
+
+func (e NoTestFilesError) Error() string {
+	return "no test files"
+}
@@ -2,6 +2,7 @@ package loader

 import (
 	"bytes"
+	"crypto/sha512"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -11,6 +12,7 @@ import (
 	"go/token"
 	"go/types"
 	"io"
+	"io/ioutil"
 	"os"
 	"os/exec"
 	"path/filepath"
@@ -66,10 +68,12 @@ type PackageJSON struct {
 type Package struct {
 	PackageJSON

-	program *Program
-	Files   []*ast.File
-	Pkg     *types.Package
-	info    types.Info
+	program    *Program
+	Files      []*ast.File
+	FileHashes map[string][]byte
+	CFlags     []string // CFlags used during CGo preprocessing (only set if CGo is used)
+	Pkg        *types.Package
+	info       types.Info
 }

 // Load loads the given package with all dependencies (including the runtime
@@ -118,7 +122,8 @@ func Load(config *compileopts.Config, inputPkgs []string, clangHeaders string, t
 	decoder := json.NewDecoder(buf)
 	for {
 		pkg := &Package{
-			program: p,
+			program:    p,
+			FileHashes: make(map[string][]byte),
 			info: types.Info{
 				Types:      make(map[ast.Expr]types.TypeAndValue),
 				Defs:       make(map[*ast.Ident]types.Object),
@@ -205,6 +210,12 @@ func Load(config *compileopts.Config, inputPkgs []string, clangHeaders string, t
 		p.Packages[pkg.ImportPath] = pkg
 	}

+	if config.TestConfig.CompileTestBinary && !strings.HasSuffix(p.sorted[len(p.sorted)-1].ImportPath, ".test") {
+		// Trying to compile a test binary but there are no test files in this
+		// package.
+		return p, NoTestFilesError{p.sorted[len(p.sorted)-1].ImportPath}
+	}
+
 	return p, nil
 }

@@ -277,17 +288,15 @@ func (p *Program) Parse() error {
 }

 // parseFile is a wrapper around parser.ParseFile.
-func (p *Program) parseFile(path string, mode parser.Mode) (*ast.File, error) {
-	if p.fset == nil {
-		p.fset = token.NewFileSet()
-	}
-
-	rd, err := os.Open(path)
+func (p *Package) parseFile(path string, mode parser.Mode) (*ast.File, error) {
+	originalPath := p.program.getOriginalPath(path)
+	data, err := ioutil.ReadFile(path)
 	if err != nil {
 		return nil, err
 	}
-	defer rd.Close()
-	return parser.ParseFile(p.fset, p.getOriginalPath(path), rd, mode)
+	sum := sha512.Sum512_224(data)
+	p.FileHashes[originalPath] = sum[:]
+	return parser.ParseFile(p.program.fset, originalPath, data, mode)
 }

 // Parse parses and typechecks this package.
@@ -363,7 +372,7 @@ func (p *Package) parseFiles() ([]*ast.File, error) {
 		if !filepath.IsAbs(file) {
 			file = filepath.Join(p.Dir, file)
 		}
-		f, err := p.program.parseFile(file, parser.ParseComments)
+		f, err := p.parseFile(file, parser.ParseComments)
 		if err != nil {
 			fileErrs = append(fileErrs, err)
 			return
@@ -379,13 +388,17 @@ func (p *Package) parseFiles() ([]*ast.File, error) {

 	// Do CGo processing.
 	if len(p.CgoFiles) != 0 {
-		var cflags []string
-		cflags = append(cflags, p.program.config.CFlags()...)
-		cflags = append(cflags, "-I"+p.Dir)
+		var initialCFlags []string
+		initialCFlags = append(initialCFlags, p.program.config.CFlags()...)
+		initialCFlags = append(initialCFlags, "-I"+p.Dir)
 		if p.program.clangHeaders != "" {
-			cflags = append(cflags, "-Xclang", "-internal-isystem", "-Xclang", p.program.clangHeaders)
+			initialCFlags = append(initialCFlags, "-Xclang", "-internal-isystem", "-Xclang", p.program.clangHeaders)
+		}
+		generated, cflags, ldflags, accessedFiles, errs := cgo.Process(files, p.program.workingDir, p.program.fset, initialCFlags)
+		p.CFlags = append(initialCFlags, cflags...)
+		for path, hash := range accessedFiles {
+			p.FileHashes[path] = hash
 		}
-		generated, ldflags, errs := cgo.Process(files, p.program.workingDir, p.program.fset, cflags)
 		if errs != nil {
 			fileErrs = append(fileErrs, errs...)
 		}
@@ -13,11 +13,13 @@ import (
 	"os/exec"
 	"os/signal"
 	"path/filepath"
+	"regexp"
 	"runtime"
 	"strings"
 	"sync/atomic"
 	"time"

+	"github.com/google/shlex"
 	"github.com/mattn/go-colorable"
 	"github.com/tinygo-org/tinygo/builder"
 	"github.com/tinygo-org/tinygo/compileopts"
@@ -28,6 +30,7 @@ import (
 	"tinygo.org/x/go-llvm"

 	"go.bug.st/serial"
+	"go.bug.st/serial/enumerator"
 )

 var (
@@ -93,7 +96,7 @@ func copyFile(src, dst string) error {
 // executeCommand is a simple wrapper to exec.Cmd
 func executeCommand(options *compileopts.Options, name string, arg ...string) *exec.Cmd {
 	if options.PrintCommands {
-		fmt.Printf("%s %s\n ", name, strings.Join(arg, " "))
+		fmt.Printf("%s %s\n", name, strings.Join(arg, " "))
 	}
 	return exec.Command(name, arg...)
 }
@@ -133,15 +136,17 @@ func Build(pkgName, outpath string, options *compileopts.Options) error {
 	})
 }

-// Test runs the tests in the given package.
-func Test(pkgName string, options *compileopts.Options, testCompileOnly bool, outpath string) error {
+// Test runs the tests in the given package. Returns whether the test passed and
+// possibly an error if the test failed to run.
+func Test(pkgName string, options *compileopts.Options, testCompileOnly bool, outpath string) (bool, error) {
 	options.TestConfig.CompileTestBinary = true
 	config, err := builder.NewConfig(options)
 	if err != nil {
-		return err
+		return false, err
 	}

-	return builder.Build(pkgName, outpath, config, func(result builder.BuildResult) error {
+	var passed bool
+	err = builder.Build(pkgName, outpath, config, func(result builder.BuildResult) error {
 		if testCompileOnly || outpath != "" {
 			// Write test binary to the specified file name.
 			if outpath == "" {
@@ -155,48 +160,78 @@ func Test(pkgName string, options *compileopts.Options, testCompileOnly bool, ou
 			// Do not run the test.
 			return nil
 		}
-		if len(config.Target.Emulator) == 0 {
-			// Run directly.
-			cmd := executeCommand(config.Options, result.Binary)
-			cmd.Stdout = os.Stdout
-			cmd.Stderr = os.Stderr
-			cmd.Dir = result.MainDir
-			err := cmd.Run()
-			if err != nil {
-				// Propagate the exit code
-				if err, ok := err.(*exec.ExitError); ok {
-					os.Exit(err.ExitCode())
-				}
-				return &commandError{"failed to run compiled binary", result.Binary, err}
-			}
-			return nil
+
+		// Run the test.
+		start := time.Now()
+		var err error
+		passed, err = runPackageTest(config, result)
+		if err != nil {
+			return err
+		}
+		duration := time.Since(start)
+
+		// Print the result.
+		importPath := strings.TrimSuffix(result.ImportPath, ".test")
+		if passed {
+			fmt.Printf("ok  \t%s\t%.3fs\n", importPath, duration.Seconds())
 		} else {
-			// Run in an emulator.
-			args := append(config.Target.Emulator[1:], result.Binary)
-			cmd := executeCommand(config.Options, config.Target.Emulator[0], args...)
-			buf := &bytes.Buffer{}
-			w := io.MultiWriter(os.Stdout, buf)
-			cmd.Stdout = w
-			cmd.Stderr = os.Stderr
-			err := cmd.Run()
-			if err != nil {
-				if err, ok := err.(*exec.ExitError); !ok || !err.Exited() {
-					// Workaround for QEMU which always exits with an error.
-					return &commandError{"failed to run emulator with", result.Binary, err}
-				}
+			fmt.Printf("FAIL\t%s\t%.3fs\n", importPath, duration.Seconds())
+		}
+		return nil
+	})
+	if err, ok := err.(loader.NoTestFilesError); ok {
+		fmt.Printf("?   \t%s\t[no test files]\n", err.ImportPath)
+		// Pretend the test passed - it at least didn't fail.
+		return true, nil
+	}
+	return passed, err
+}
+
+// runPackageTest runs a test binary that was previously built. The return
+// values are whether the test passed and any errors encountered while trying to
+// run the binary.
+func runPackageTest(config *compileopts.Config, result builder.BuildResult) (bool, error) {
+	if len(config.Target.Emulator) == 0 {
+		// Run directly.
+		cmd := executeCommand(config.Options, result.Binary)
+		cmd.Stdout = os.Stdout
+		cmd.Stderr = os.Stderr
+		cmd.Dir = result.MainDir
+		err := cmd.Run()
+		if err != nil {
+			if _, ok := err.(*exec.ExitError); ok {
+				// Binary exited with a non-zero exit code, which means the test
+				// failed.
+				return false, nil
 			}
-			testOutput := string(buf.Bytes())
-			if testOutput == "PASS\n" || strings.HasSuffix(testOutput, "\nPASS\n") {
-				// Test passed.
-				return nil
-			} else {
-				// Test failed, either by ending with the word "FAIL" or with a
-				// panic of some sort.
-				os.Exit(1)
-				return nil // unreachable
+			return false, &commandError{"failed to run compiled binary", result.Binary, err}
+		}
+		return true, nil
+	} else {
+		// Run in an emulator.
+		args := append(config.Target.Emulator[1:], result.Binary)
+		cmd := executeCommand(config.Options, config.Target.Emulator[0], args...)
+		buf := &bytes.Buffer{}
+		w := io.MultiWriter(os.Stdout, buf)
+		cmd.Stdout = w
+		cmd.Stderr = os.Stderr
+		err := cmd.Run()
+		if err != nil {
+			if err, ok := err.(*exec.ExitError); !ok || !err.Exited() {
+				// Workaround for QEMU which always exits with an error.
+				return false, &commandError{"failed to run emulator with", result.Binary, err}
 			}
 		}
-	})
+		testOutput := string(buf.Bytes())
+		if testOutput == "PASS\n" || strings.HasSuffix(testOutput, "\nPASS\n") {
+			// Test passed.
+			return true, nil
+		} else {
+			// Test failed, either by ending with the word "FAIL" or with a
+			// panic of some sort.
+			return false, nil
+		}
+	}
 }

 // Flash builds and flashes the built binary to the given serial port.
@@ -240,15 +275,12 @@ func Flash(pkgName, port string, options *compileopts.Options) error {
 	return builder.Build(pkgName, fileExt, config, func(result builder.BuildResult) error {
 		// do we need port reset to put MCU into bootloader mode?
 		if config.Target.PortReset == "true" && flashMethod != "openocd" {
-			if port == "" {
-				var err error
-				port, err = getDefaultPort()
-				if err != nil {
-					return err
-				}
+			port, err := getDefaultPort(strings.FieldsFunc(port, func(c rune) bool { return c == ',' }))
+			if err != nil {
+				return err
 			}

-			err := touchSerialPortAt1200bps(port)
+			err = touchSerialPortAt1200bps(port)
 			if err != nil {
 				return &commandError{"failed to reset port", result.Binary, err}
 			}
@@ -264,9 +296,9 @@ func Flash(pkgName, port string, options *compileopts.Options) error {
 			fileToken := "{" + fileExt[1:] + "}"
 			flashCmd = strings.ReplaceAll(flashCmd, fileToken, result.Binary)

-			if port == "" && strings.Contains(flashCmd, "{port}") {
+			if strings.Contains(flashCmd, "{port}") {
 				var err error
-				port, err = getDefaultPort()
+				port, err = getDefaultPort(strings.FieldsFunc(port, func(c rune) bool { return c == ',' }))
 				if err != nil {
 					return err
 				}
@@ -344,8 +376,9 @@ func FlashGDB(pkgName string, ocdOutput bool, options *compileopts.Options) erro
 	if err != nil {
 		return err
 	}
-	if config.Target.GDB == "" {
-		return errors.New("gdb not configured in the target specification")
+	gdb, err := config.Target.LookupGDB()
+	if err != nil {
+		return err
 	}

 	return builder.Build(pkgName, "", config, func(result builder.BuildResult) error {
@@ -490,7 +523,7 @@ func FlashGDB(pkgName string, ocdOutput bool, options *compileopts.Options) erro
 		for _, cmd := range gdbCommands {
 			params = append(params, "-ex", cmd)
 		}
-		cmd := executeCommand(config.Options, config.Target.GDB, params...)
+		cmd := executeCommand(config.Options, gdb, params...)
 		cmd.Stdin = os.Stdin
 		cmd.Stdout = os.Stdout
 		cmd.Stderr = os.Stderr
@@ -664,41 +697,68 @@ func windowsFindUSBDrive(volume string, options *compileopts.Options) (string, e
 }

 // getDefaultPort returns the default serial port depending on the operating system.
-func getDefaultPort() (port string, err error) {
-	var portPath string
+func getDefaultPort(portCandidates []string) (port string, err error) {
+	if len(portCandidates) == 1 {
+		return portCandidates[0], nil
+	}
+
+	var ports []string
 	switch runtime.GOOS {
-	case "darwin":
-		portPath = "/dev/cu.usb*"
-	case "linux":
-		portPath = "/dev/ttyACM*"
 	case "freebsd":
-		portPath = "/dev/cuaU*"
-	case "windows":
-		ports, err := serial.GetPortsList()
+		ports, err = filepath.Glob("/dev/cuaU*")
+	case "darwin", "linux", "windows":
+		var portsList []*enumerator.PortDetails
+		portsList, err = enumerator.GetDetailedPortsList()
 		if err != nil {
 			return "", err
 		}

-		if len(ports) == 0 {
-			return "", errors.New("no serial ports available")
-		} else if len(ports) > 1 {
-			return "", errors.New("multiple serial ports available - use -port flag")
+		for _, p := range portsList {
+			if p.IsUSB {
+				ports = append(ports, p.Name)
+			}
 		}

-		return ports[0], nil
+		if ports == nil || len(ports) == 0 {
+			// fallback
+			switch runtime.GOOS {
+			case "darwin":
+				ports, err = filepath.Glob("/dev/cu.usb*")
+			case "linux":
+				ports, err = filepath.Glob("/dev/ttyACM*")
+			case "windows":
+				ports, err = serial.GetPortsList()
+			}
+		}
 	default:
 		return "", errors.New("unable to search for a default USB device to be flashed on this OS")
 	}

-	d, err := filepath.Glob(portPath)
 	if err != nil {
 		return "", err
-	}
-	if d == nil {
+	} else if ports == nil {
 		return "", errors.New("unable to locate a serial port")
+	} else if len(ports) == 0 {
+		return "", errors.New("no serial ports available")
 	}

-	return d[0], nil
+	if len(portCandidates) == 0 {
+		if len(ports) == 1 {
+			return ports[0], nil
+		} else {
+			return "", errors.New("multiple serial ports available - use -port flag, available ports are " + strings.Join(ports, ", "))
+		}
+	}
+
+	for _, ps := range portCandidates {
+		for _, p := range ports {
+			if p == ps {
+				return p, nil
+			}
+		}
+	}
+
+	return "", errors.New("port you specified '" + strings.Join(portCandidates, ",") + "' does not exist, available ports are " + strings.Join(ports, ", "))
 }

 func usage() {
@@ -811,6 +871,52 @@ func handleCompilerError(err error) {
 	}
 }

+// This is a special type for the -X flag to parse the pkgpath.Var=stringVal
+// format. It has to be a special type to allow multiple variables to be defined
+// this way.
+type globalValuesFlag map[string]map[string]string
+
+func (m globalValuesFlag) String() string {
+	return "pkgpath.Var=value"
+}
+
+func (m globalValuesFlag) Set(value string) error {
+	equalsIndex := strings.IndexByte(value, '=')
+	if equalsIndex < 0 {
+		return errors.New("expected format pkgpath.Var=value")
+	}
+	pathAndName := value[:equalsIndex]
+	pointIndex := strings.LastIndexByte(pathAndName, '.')
+	if pointIndex < 0 {
+		return errors.New("expected format pkgpath.Var=value")
+	}
+	path := pathAndName[:pointIndex]
+	name := pathAndName[pointIndex+1:]
+	stringValue := value[equalsIndex+1:]
+	if m[path] == nil {
+		m[path] = make(map[string]string)
+	}
+	m[path][name] = stringValue
+	return nil
+}
+
+// parseGoLinkFlag parses the -ldflags parameter. Its primary purpose right now
+// is the -X flag, for setting the value of global string variables.
+func parseGoLinkFlag(flagsString string) (map[string]map[string]string, error) {
+	set := flag.NewFlagSet("link", flag.ExitOnError)
+	globalVarValues := make(globalValuesFlag)
+	set.Var(globalVarValues, "X", "Set the value of the string variable to the given value.")
+	flags, err := shlex.Split(flagsString)
+	if err != nil {
+		return nil, err
+	}
+	err = set.Parse(flags)
+	if err != nil {
+		return nil, err
+	}
+	return map[string]map[string]string(globalVarValues), nil
+}
+
 func main() {
 	if len(os.Args) < 2 {
 		fmt.Fprintln(os.Stderr, "No command-line arguments supplied.")
@@ -830,14 +936,16 @@ func main() {
 	target := flag.String("target", "", "LLVM target | .json file with TargetSpec")
 	printSize := flag.String("size", "", "print sizes (none, short, full)")
 	printStacks := flag.Bool("print-stacks", false, "print stack sizes of goroutines")
+	printAllocsString := flag.String("print-allocs", "", "regular expression of functions for which heap allocations should be printed")
 	printCommands := flag.Bool("x", false, "Print commands")
 	nodebug := flag.Bool("no-debug", false, "disable DWARF debug symbol generation")
+	ocdCommandsString := flag.String("ocd-commands", "", "OpenOCD commands, overriding target spec (can specify multiple separated by commas)")
 	ocdOutput := flag.Bool("ocd-output", false, "print OCD daemon output during debug")
-	port := flag.String("port", "", "flash port")
+	port := flag.String("port", "", "flash port (can specify multiple candidates separated by commas)")
 	programmer := flag.String("programmer", "", "which hardware programmer to use")
-	cFlags := flag.String("cflags", "", "additional cflags for compiler")
-	ldFlags := flag.String("ldflags", "", "additional ldflags for linker")
+	ldflags := flag.String("ldflags", "", "Go link tool compatible ldflags")
 	wasmAbi := flag.String("wasm-abi", "", "WebAssembly ABI conventions: js (no i64 params) or generic")
+	llvmFeatures := flag.String("llvm-features", "", "comma separated LLVM features to enable")

 	var flagJSON, flagDeps *bool
 	if command == "help" || command == "list" {
@@ -866,35 +974,51 @@ func main() {
 	}

 	flag.CommandLine.Parse(os.Args[2:])
+	globalVarValues, err := parseGoLinkFlag(*ldflags)
+	if err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(1)
+	}
+
+	var printAllocs *regexp.Regexp
+	if *printAllocsString != "" {
+		printAllocs, err = regexp.Compile(*printAllocsString)
+		if err != nil {
+			fmt.Fprintln(os.Stderr, err)
+			os.Exit(1)
+		}
+	}
+
+	var ocdCommands []string
+	if *ocdCommandsString != "" {
+		ocdCommands = strings.Split(*ocdCommandsString, ",")
+	}
+
 	options := &compileopts.Options{
-		Target:        *target,
-		Opt:           *opt,
-		GC:            *gc,
-		PanicStrategy: *panicStrategy,
-		Scheduler:     *scheduler,
-		PrintIR:       *printIR,
-		DumpSSA:       *dumpSSA,
-		VerifyIR:      *verifyIR,
-		Debug:         !*nodebug,
-		PrintSizes:    *printSize,
-		PrintStacks:   *printStacks,
-		PrintCommands: *printCommands,
-		Tags:          *tags,
-		WasmAbi:       *wasmAbi,
-		Programmer:    *programmer,
-	}
-
-	if *cFlags != "" {
-		options.CFlags = strings.Split(*cFlags, " ")
-	}
-
-	if *ldFlags != "" {
-		options.LDFlags = strings.Split(*ldFlags, " ")
+		Target:          *target,
+		Opt:             *opt,
+		GC:              *gc,
+		PanicStrategy:   *panicStrategy,
+		Scheduler:       *scheduler,
+		PrintIR:         *printIR,
+		DumpSSA:         *dumpSSA,
+		VerifyIR:        *verifyIR,
+		Debug:           !*nodebug,
+		PrintSizes:      *printSize,
+		PrintStacks:     *printStacks,
+		PrintAllocs:     printAllocs,
+		PrintCommands:   *printCommands,
+		Tags:            *tags,
+		GlobalValues:    globalVarValues,
+		WasmAbi:         *wasmAbi,
+		Programmer:      *programmer,
+		OpenOCDCommands: ocdCommands,
+		LLVMFeatures:    *llvmFeatures,
 	}

 	os.Setenv("CC", "clang -target="+*target)

-	err := options.Verify()
+	err = options.Verify()
 	if err != nil {
 		fmt.Fprintln(os.Stderr, err.Error())
 		usage()
@@ -982,16 +1106,26 @@ func main() {
 		err := Run(pkgName, options)
 		handleCompilerError(err)
 	case "test":
-		pkgName := "."
-		if flag.NArg() == 1 {
-			pkgName = filepath.ToSlash(flag.Arg(0))
-		} else if flag.NArg() > 1 {
-			fmt.Fprintln(os.Stderr, "test only accepts a single positional argument: package name, but multiple were specified")
-			usage()
+		var pkgNames []string
+		for i := 0; i < flag.NArg(); i++ {
+			pkgNames = append(pkgNames, filepath.ToSlash(flag.Arg(i)))
+		}
+		if len(pkgNames) == 0 {
+			pkgNames = []string{"."}
+		}
+		allTestsPassed := true
+		for _, pkgName := range pkgNames {
+			// TODO: parallelize building the test binaries
+			passed, err := Test(pkgName, options, *testCompileOnlyFlag, outpath)
+			handleCompilerError(err)
+			if !passed {
+				allTestsPassed = false
+			}
+		}
+		if !allTestsPassed {
+			fmt.Println("FAIL")
 			os.Exit(1)
 		}
-		err := Test(pkgName, options, *testCompileOnlyFlag, outpath)
-		handleCompilerError(err)
 	case "targets":
 		dir := filepath.Join(goenv.Get("TINYGOROOT"), "targets")
 		entries, err := ioutil.ReadDir(dir)
@@ -13,7 +13,6 @@ import (
 	"os/exec"
 	"path/filepath"
 	"runtime"
-	"sort"
 	"strings"
 	"sync"
 	"testing"
@@ -29,38 +28,43 @@ const TESTDATA = "testdata"
 var testTarget = flag.String("target", "", "override test target")

 func TestCompiler(t *testing.T) {
-	matches, err := filepath.Glob(filepath.Join(TESTDATA, "*.go"))
-	if err != nil {
-		t.Fatal("could not read test files:", err)
+	tests := []string{
+		"alias.go",
+		"atomic.go",
+		"binop.go",
+		"calls.go",
+		"cgo/",
+		"channel.go",
+		"coroutines.go",
+		"float.go",
+		"gc.go",
+		"init.go",
+		"init_multi.go",
+		"interface.go",
+		"json.go",
+		"map.go",
+		"math.go",
+		"print.go",
+		"reflect.go",
+		"slice.go",
+		"sort.go",
+		"stdlib.go",
+		"string.go",
+		"structs.go",
+		"zeroalloc.go",
 	}

-	dirMatches, err := filepath.Glob(filepath.Join(TESTDATA, "*", "main.go"))
-	if err != nil {
-		t.Fatal("could not read test packages:", err)
-	}
-	if len(matches) == 0 || len(dirMatches) == 0 {
-		t.Fatal("no test files found")
-	}
-	for _, m := range dirMatches {
-		matches = append(matches, filepath.Dir(m)+string(filepath.Separator))
-	}
-
-	sort.Strings(matches)
-
 	if *testTarget != "" {
 		// This makes it possible to run one specific test (instead of all),
 		// which is especially useful to quickly check whether some changes
 		// affect a particular target architecture.
-		runPlatTests(*testTarget, matches, t)
+		runPlatTests(*testTarget, tests, t)
 		return
 	}

 	if runtime.GOOS != "windows" {
 		t.Run("Host", func(t *testing.T) {
-			runPlatTests("", matches, t)
-			if runtime.GOOS == "darwin" {
-				runTest("testdata/libc/env.go", "", t, []string{"ENV1=VALUE1", "ENV2=VALUE2"}...)
-			}
+			runPlatTests("", tests, t)
 		})
 	}

@@ -69,26 +73,26 @@ func TestCompiler(t *testing.T) {
 	}

 	t.Run("EmulatedCortexM3", func(t *testing.T) {
-		runPlatTests("cortex-m-qemu", matches, t)
+		runPlatTests("cortex-m-qemu", tests, t)
 	})

 	if runtime.GOOS == "windows" || runtime.GOOS == "darwin" {
 		// Note: running only on Windows and macOS because Linux (as of 2020)
 		// usually has an outdated QEMU version that doesn't support RISC-V yet.
 		t.Run("EmulatedRISCV", func(t *testing.T) {
-			runPlatTests("riscv-qemu", matches, t)
+			runPlatTests("riscv-qemu", tests, t)
 		})
 	}

 	if runtime.GOOS == "linux" {
 		t.Run("X86Linux", func(t *testing.T) {
-			runPlatTests("i386--linux-gnu", matches, t)
+			runPlatTests("i386--linux-gnu", tests, t)
 		})
 		t.Run("ARMLinux", func(t *testing.T) {
-			runPlatTests("arm--linux-gnueabihf", matches, t)
+			runPlatTests("arm--linux-gnueabihf", tests, t)
 		})
 		t.Run("ARM64Linux", func(t *testing.T) {
-			runPlatTests("aarch64--linux-gnu", matches, t)
+			runPlatTests("aarch64--linux-gnu", tests, t)
 		})
 		goVersion, err := goenv.GorootVersionString(goenv.Get("GOROOT"))
 		if err != nil {
@@ -101,25 +105,72 @@ func TestCompiler(t *testing.T) {
 			// below that are also not supported but still seem to pass, so
 			// include them in the tests for now.
 			t.Run("WebAssembly", func(t *testing.T) {
-				runPlatTests("wasm", matches, t)
+				runPlatTests("wasm", tests, t)
 			})
 		}

 		t.Run("WASI", func(t *testing.T) {
-			runPlatTests("wasi", matches, t)
-			runTest("testdata/libc/env.go", "wasi", t, []string{"ENV1=VALUE1", "ENV2=VALUE2"}...)
+			runPlatTests("wasi", tests, t)
 		})
 	}
+
+	// Test a few build options.
+	t.Run("build-options", func(t *testing.T) {
+		if runtime.GOOS == "windows" {
+			// These tests assume a host that is supported by TinyGo.
+			t.Skip("can't test build options on Windows")
+		}
+		t.Parallel()
+
+		// Test with few optimizations enabled (no inlining, etc).
+		t.Run("opt=1", func(t *testing.T) {
+			t.Parallel()
+			runTestWithConfig("stdlib.go", "", t, &compileopts.Options{
+				Opt: "1",
+			}, nil, nil)
+		})
+
+		// Test with only the bare minimum of optimizations enabled.
+		// TODO: fix this for stdlib.go, which currently fails.
+		t.Run("opt=0", func(t *testing.T) {
+			t.Parallel()
+			runTestWithConfig("print.go", "", t, &compileopts.Options{
+				Opt: "0",
+			}, nil, nil)
+		})
+
+		t.Run("ldflags", func(t *testing.T) {
+			t.Parallel()
+			runTestWithConfig("ldflags.go", "", t, &compileopts.Options{
+				Opt: "z",
+				GlobalValues: map[string]map[string]string{
+					"main": {
+						"someGlobal": "foobar",
+					},
+				},
+			}, nil, nil)
+		})
+	})
 }

-func runPlatTests(target string, matches []string, t *testing.T) {
+func runPlatTests(target string, tests []string, t *testing.T) {
 	t.Parallel()

-	for _, path := range matches {
-		path := path // redefine to avoid race condition
-		t.Run(filepath.Base(path), func(t *testing.T) {
+	for _, name := range tests {
+		name := name // redefine to avoid race condition
+		t.Run(name, func(t *testing.T) {
 			t.Parallel()
-			runTest(path, target, t)
+			runTest(name, target, t, nil, nil)
+		})
+	}
+	if target == "wasi" || target == "" {
+		t.Run("filesystem.go", func(t *testing.T) {
+			t.Parallel()
+			runTest("filesystem.go", target, t, nil, nil)
+		})
+		t.Run("env.go", func(t *testing.T) {
+			t.Parallel()
+			runTest("env.go", target, t, []string{"first", "second"}, []string{"ENV1=VALUE1", "ENV2=VALUE2"})
 		})
 	}
 }
@@ -136,10 +187,28 @@ func runBuild(src, out string, opts *compileopts.Options) error {
 	return Build(src, out, opts)
 }

-func runTest(path, target string, t *testing.T, environmentVars ...string) {
+func runTest(name, target string, t *testing.T, cmdArgs, environmentVars []string) {
+	options := &compileopts.Options{
+		Target:     target,
+		Opt:        "z",
+		PrintIR:    false,
+		DumpSSA:    false,
+		VerifyIR:   true,
+		Debug:      true,
+		PrintSizes: "",
+		WasmAbi:    "",
+	}
+	runTestWithConfig(name, target, t, options, cmdArgs, environmentVars)
+}
+
+func runTestWithConfig(name, target string, t *testing.T, options *compileopts.Options, cmdArgs, environmentVars []string) {
+	// Get the expected output for this test.
+	// Note: not using filepath.Join as it strips the path separator at the end
+	// of the path.
+	path := TESTDATA + "/" + name
 	// Get the expected output for this test.
 	txtpath := path[:len(path)-3] + ".txt"
-	if path[len(path)-1] == os.PathSeparator {
+	if path[len(path)-1] == '/' {
 		txtpath = path + "out.txt"
 	}
 	expected, err := ioutil.ReadFile(txtpath)
@@ -160,19 +229,8 @@ func runTest(path, target string, t *testing.T, environmentVars ...string) {
 	}()

 	// Build the test binary.
-	config := &compileopts.Options{
-		Target:     target,
-		Opt:        "z",
-		PrintIR:    false,
-		DumpSSA:    false,
-		VerifyIR:   true,
-		Debug:      true,
-		PrintSizes: "",
-		WasmAbi:    "",
-	}
-
 	binary := filepath.Join(tmpdir, "test")
-	err = runBuild("./"+path, binary, config)
+	err = runBuild("./"+path, binary, options)
 	if err != nil {
 		printCompilerError(t.Log, err)
 		t.Fail()
@@ -186,6 +244,7 @@ func runTest(path, target string, t *testing.T, environmentVars ...string) {
 	if target == "" {
 		cmd = exec.Command(binary)
 		cmd.Env = append(cmd.Env, environmentVars...)
+		cmd.Args = append(cmd.Args, cmdArgs...)
 	} else {
 		spec, err := compileopts.LoadTarget(target)
 		if err != nil {
@@ -199,9 +258,12 @@ func runTest(path, target string, t *testing.T, environmentVars ...string) {
 		}

 		if len(spec.Emulator) != 0 && spec.Emulator[0] == "wasmtime" {
+			// Allow reading from the current directory.
+			cmd.Args = append(cmd.Args, "--dir=.")
 			for _, v := range environmentVars {
 				cmd.Args = append(cmd.Args, "--env", v)
 			}
+			cmd.Args = append(cmd.Args, cmdArgs...)
 		} else {
 			cmd.Env = append(cmd.Env, environmentVars...)
 		}
@@ -26,12 +26,21 @@ type SCB_Type struct {
 	SHPR2 volatile.Register32 // 0xD1C: System Handler Priority Register 2
 	SHPR3 volatile.Register32 // 0xD20: System Handler Priority Register 3
 	// the following are only applicable for Cortex-M3/M33/M4/M7
-	SHCSR volatile.Register32 // 0xD24: System Handler Control and State Register
-	CFSR  volatile.Register32 // 0xD28: Configurable Fault Status Register
-	HFSR  volatile.Register32 // 0xD2C: HardFault Status Register
-	DFSR  volatile.Register32 // 0xD30: Debug Fault Status Register
-	MMFAR volatile.Register32 // 0xD34: MemManage Fault Address Register
-	BFAR  volatile.Register32 // 0xD38: BusFault Address Register
+	SHCSR volatile.Register32    // 0xD24: System Handler Control and State Register
+	CFSR  volatile.Register32    // 0xD28: Configurable Fault Status Register
+	HFSR  volatile.Register32    // 0xD2C: HardFault Status Register
+	DFSR  volatile.Register32    // 0xD30: Debug Fault Status Register
+	MMFAR volatile.Register32    // 0xD34: MemManage Fault Address Register
+	BFAR  volatile.Register32    // 0xD38: BusFault Address Register
+	AFSR  volatile.Register32    // 0xD3C: Auxiliary Fault Status Register
+	PFR   [2]volatile.Register32 // 0xD40: Processor Feature Register
+	DFR   volatile.Register32    // 0xD48: Debug Feature Register
+	ADR   volatile.Register32    // 0xD4C: Auxiliary Feature Register
+	MMFR  [4]volatile.Register32 // 0xD50: Memory Model Feature Register
+	ISAR  [5]volatile.Register32 // 0xD60: Instruction Set Attributes Register
+	_     [5]uint32              // reserved
+	CPACR volatile.Register32    // 0xD88: Coprocessor Access Control Register
+
 }

 var SCB = (*SCB_Type)(unsafe.Pointer(uintptr(SCB_BASE)))
@@ -0,0 +1,71 @@
+// Hand created file. DO NOT DELETE.
+// atsamd51x bitfield definitions that are not auto-generated by gen-device-svd.go
+
+// +build sam,atsame5x
+
+// These are the supported pchctrl function numberings on the atsamd51x
+// See http://ww1.microchip.com/downloads/en/DeviceDoc/SAM_D5xE5x_Family_Data_Sheet_DS60001507F.pdf
+// table 14-9
+
+package sam
+
+const (
+	PCHCTRL_GCLK_OSCCTRL_DFLL48     = 0 // DFLL48 input clock source
+	PCHCTRL_GCLK_OSCCTRL_FDPLL0     = 1 // Reference clock for FDPLL0
+	PCHCTRL_GCLK_OSCCTRL_FDPLL1     = 2 // Reference clock for FDPLL1
+	PCHCTRL_GCLK_OSCCTRL_FDPLL0_32K = 3 // FDPLL0 = 3 // 32KHz clock for internal lock timer
+	PCHCTRL_GCLK_OSCCTRL_FDPLL1_32K = 3 // FDPLL1 = 3 // 32KHz clock for internal lock timer
+	PCHCTRL_GCLK_SDHC0_SLOW         = 3 // SDHC0 = 3 // Slow
+	PCHCTRL_GCLK_SDHC1_SLOW         = 3 // SDHC1 = 3 // Slow
+	PCHCTRL_GCLK_SERCOMX_SLOW       = 3 // GCLK_SERCOM[0..7]_SLOW = 3
+	PCHCTRL_GCLK_EIC                = 4
+	PCHCTRL_GCLK_FREQM_MSR          = 5 // FREQM Measure
+	PCHCTRL_GCLK_FREQM_REF          = 6 // FREQM Reference
+	PCHCTRL_GCLK_SERCOM0_CORE       = 7 // SERCOM0 Core
+	PCHCTRL_GCLK_SERCOM1_CORE       = 8 // SERCOM1 Core
+	PCHCTRL_GCLK_TC0                = 9
+	PCHCTRL_GCLK_TC1                = 9  // TC0, TC1
+	PCHCTRL_GCLK_USB                = 10 // USB
+	PCHCTRL_GCLK_EVSYS0             = 11
+	PCHCTRL_GCLK_EVSYS1             = 12
+	PCHCTRL_GCLK_EVSYS2             = 13
+	PCHCTRL_GCLK_EVSYS3             = 14
+	PCHCTRL_GCLK_EVSYS4             = 15
+	PCHCTRL_GCLK_EVSYS5             = 16
+	PCHCTRL_GCLK_EVSYS6             = 17
+	PCHCTRL_GCLK_EVSYS7             = 18
+	PCHCTRL_GCLK_EVSYS8             = 19
+	PCHCTRL_GCLK_EVSYS9             = 20
+	PCHCTRL_GCLK_EVSYS10            = 21
+	PCHCTRL_GCLK_EVSYS11            = 22
+	PCHCTRL_GCLK_SERCOM2_CORE       = 23 // SERCOM2 Core
+	PCHCTRL_GCLK_SERCOM3_CORE       = 24 // SERCOM3 Core
+	PCHCTRL_GCLK_TCC0               = 25
+	PCHCTRL_GCLK_TCC1               = 25 // TCC0, TCC1
+	PCHCTRL_GCLK_TC2                = 26
+	PCHCTRL_GCLK_TC3                = 26 // TC2, TC3
+	PCHCTRL_GCLK_CAN0               = 27 // CAN0
+	PCHCTRL_GCLK_CAN1               = 28 // CAN1
+	PCHCTRL_GCLK_TCC2               = 29
+	PCHCTRL_GCLK_TCC3               = 29 // TCC2, TCC3
+	PCHCTRL_GCLK_TC4                = 30
+	PCHCTRL_GCLK_TC5                = 30 // TC4, TC5
+	PCHCTRL_GCLK_PDEC               = 31 // PDEC
+	PCHCTRL_GCLK_AC                 = 32 // AC
+	PCHCTRL_GCLK_CCL                = 33 // CCL
+	PCHCTRL_GCLK_SERCOM4_CORE       = 34 // SERCOM4 Core
+	PCHCTRL_GCLK_SERCOM5_CORE       = 35 // SERCOM5 Core
+	PCHCTRL_GCLK_SERCOM6_CORE       = 36 // SERCOM6 Core
+	PCHCTRL_GCLK_SERCOM7_CORE       = 37 // SERCOM7 Core
+	PCHCTRL_GCLK_TCC4               = 38 // TCC4
+	PCHCTRL_GCLK_TC6                = 39
+	PCHCTRL_GCLK_TC7                = 39 // TC6, TC7
+	PCHCTRL_GCLK_ADC0               = 40 // ADC0
+	PCHCTRL_GCLK_ADC1               = 41 // ADC1
+	PCHCTRL_GCLK_DAC                = 42 // DAC
+	PCHCTRL_GCLK_I2S0               = 43
+	PCHCTRL_GCLK_I2S1               = 44
+	PCHCTRL_GCLK_SDHC0              = 45 // SDHC0
+	PCHCTRL_GCLK_SDHC1              = 46 // SDHC1
+	PCHCTRL_GCLK_CM4_TRACE          = 47 // CM4 Trace
+)
@@ -0,0 +1,15 @@
+// +build feather_m4_can
+
+package main
+
+import (
+	"machine"
+)
+
+func init() {
+	// power on the CAN Transceiver
+	// https://learn.adafruit.com/adafruit-feather-m4-can-express/pinouts#can-bus-3078990-8
+	boost_en := machine.BOOST_EN
+	boost_en.Configure(machine.PinConfig{Mode: machine.PinOutput})
+	boost_en.High()
+}
@@ -0,0 +1,53 @@
+package main
+
+import (
+	"fmt"
+	"machine"
+	"time"
+)
+
+func main() {
+	can1 := machine.CAN1
+	can1.Configure(machine.CANConfig{
+		TransferRate:   machine.CANTransferRate500kbps,
+		TransferRateFD: machine.CANTransferRate1000kbps,
+		Rx:             machine.CAN1_RX,
+		Tx:             machine.CAN1_TX,
+		Standby:        machine.CAN1_STANDBY,
+	})
+
+	can0 := machine.CAN0
+	can0.Configure(machine.CANConfig{
+		TransferRate:   machine.CANTransferRate500kbps,
+		TransferRateFD: machine.CANTransferRate1000kbps,
+		Rx:             machine.CAN0_RX,
+		Tx:             machine.CAN0_TX,
+		Standby:        machine.NoPin,
+	})
+
+	rxMsg := machine.CANRxBufferElement{}
+
+	for {
+		can1.Tx(0x123, []byte{0x01, 0x23, 0x45, 0x67, 0x89, 0xAB, 0xCD, 0xEF}, false, false)
+		can1.Tx(0x789, []byte{0x02, 0x24, 0x46, 0x67, 0x89, 0xAB, 0xCD, 0xEF, 0x01, 0x23, 0x45, 0x67, 0x89, 0xAB, 0xCD, 0xEF}, true, false)
+		time.Sleep(time.Millisecond * 1000)
+
+		sz0 := can0.RxFifoSize()
+		if sz0 > 0 {
+			fmt.Printf("CAN0 %d\r\n", sz0)
+			for i := 0; i < sz0; i++ {
+				can0.RxRaw(&rxMsg)
+				fmt.Printf("-> %08X %X %#v\r\n", rxMsg.ID, rxMsg.DLC, rxMsg.Data())
+			}
+		}
+
+		sz1 := can1.RxFifoSize()
+		if sz1 > 0 {
+			fmt.Printf("CAN1 %d\r\n", sz1)
+			for i := 0; i < sz1; i++ {
+				can1.RxRaw(&rxMsg)
+				fmt.Printf("-> %08X %X %#v\r\n", rxMsg.ID, rxMsg.DLC, rxMsg.Data())
+			}
+		}
+	}
+}
@@ -0,0 +1,15 @@
+// +build feather_m4_can
+
+package main
+
+import (
+	"machine"
+)
+
+func init() {
+	// power on the CAN Transceiver
+	// https://learn.adafruit.com/adafruit-feather-m4-can-express/pinouts#can-bus-3078990-8
+	boost_en := machine.BOOST_EN
+	boost_en.Configure(machine.PinConfig{Mode: machine.PinOutput})
+	boost_en.High()
+}
@@ -0,0 +1,75 @@
+package main
+
+import (
+	"device/sam"
+	"fmt"
+	"machine"
+	"time"
+)
+
+type canMsg struct {
+	ch   byte
+	id   uint32
+	dlc  byte
+	data []byte
+}
+
+func main() {
+	ch := make(chan canMsg, 10)
+	go func() {
+		for {
+			select {
+			case m := <-ch:
+				fmt.Printf("%d %03X %X ", m.ch, m.id, m.dlc)
+				for _, d := range m.data {
+					fmt.Printf("%02X ", d)
+				}
+				fmt.Printf("\r\n")
+			}
+
+		}
+	}()
+
+	can1 := machine.CAN1
+	can1.Configure(machine.CANConfig{
+		TransferRate:   machine.CANTransferRate500kbps,
+		TransferRateFD: machine.CANTransferRate1000kbps,
+		Rx:             machine.CAN1_RX,
+		Tx:             machine.CAN1_TX,
+		Standby:        machine.CAN1_STANDBY,
+	})
+	// RF0NE : Rx FIFO 0 New Message Interrupt Enable
+	can1.SetInterrupt(sam.CAN_IE_RF0NE, func(*machine.CAN) {
+		rxMsg := machine.CANRxBufferElement{}
+		can1.RxRaw(&rxMsg)
+		m := canMsg{ch: 1, id: rxMsg.ID, dlc: rxMsg.DLC, data: rxMsg.Data()}
+		select {
+		case ch <- m:
+		}
+	})
+
+	can0 := machine.CAN0
+	can0.Configure(machine.CANConfig{
+		TransferRate:   machine.CANTransferRate500kbps,
+		TransferRateFD: machine.CANTransferRate1000kbps,
+		Rx:             machine.CAN0_RX,
+		Tx:             machine.CAN0_TX,
+		Standby:        machine.NoPin,
+	})
+	// RF0NE : Rx FIFO 0 New Message Interrupt Enable
+	can0.SetInterrupt(sam.CAN_IE_RF0NE, func(*machine.CAN) {
+		rxMsg := machine.CANRxBufferElement{}
+		can0.RxRaw(&rxMsg)
+		m := canMsg{ch: 2, id: rxMsg.ID, dlc: rxMsg.DLC, data: rxMsg.Data()}
+		select {
+		case ch <- m:
+		}
+	})
+
+	for {
+		can0.Tx(0x123, []byte{0x01, 0x23, 0x45, 0x67, 0x89, 0xAB, 0xCD, 0xEF}, false, false)
+		time.Sleep(time.Millisecond * 500)
+		can1.Tx(0x456, []byte{0xAA, 0xBB, 0xCC}, false, false)
+		time.Sleep(time.Millisecond * 1000)
+	}
+}
@@ -0,0 +1,12 @@
+// +build arduino_mega1280
+
+package main
+
+import "machine"
+
+var (
+	// Configuration on an Arduino Uno.
+	pwm  = machine.Timer3
+	pinA = machine.PH3 // pin 6 on the Mega
+	pinB = machine.PH4 // pin 7 on the Mega
+)
@@ -0,0 +1,12 @@
+// +build arduino
+
+package main
+
+import "machine"
+
+var (
+	// Configuration on an Arduino Uno.
+	pwm  = machine.Timer2
+	pinA = machine.PB3 // pin 11 on the Uno
+	pinB = machine.PD3 // pin 3 on the Uno
+)
@@ -0,0 +1,11 @@
+// +build feather_m4
+
+package main
+
+import "machine"
+
+var (
+	pwm  = machine.TCC0
+	pinA = machine.D12
+	pinB = machine.D13
+)
@@ -0,0 +1,11 @@
+// +build itsybitsy_m0
+
+package main
+
+import "machine"
+
+var (
+	pwm  = machine.TCC0
+	pinA = machine.D3
+	pinB = machine.D4
+)
@@ -0,0 +1,11 @@
+// +build itsybitsy_m4
+
+package main
+
+import "machine"
+
+var (
+	pwm  = machine.TCC0
+	pinA = machine.D12
+	pinB = machine.D13
+)
@@ -1,64 +1,74 @@
 package main

+// This example demonstrates some features of the PWM support.
+
 import (
 	"machine"
 	"time"
 )

-// This example assumes that an RGB LED is connected to pins 3, 5 and 6 on an Arduino.
-// Change the values below to use different pins.
-const (
-	redPin   = machine.D4
-	greenPin = machine.D5
-	bluePin  = machine.D6
-)
-
-// cycleColor is just a placeholder until math/rand or some equivalent is working.
-func cycleColor(color uint8) uint8 {
-	if color < 10 {
-		return color + 1
-	} else if color < 200 {
-		return color + 10
-	} else {
-		return 0
-	}
-}
+const delayBetweenPeriods = time.Second * 5

 func main() {
-	machine.InitPWM()
+	// Delay a bit on startup to easily catch the first messages.
+	time.Sleep(time.Second * 2)

-	red := machine.PWM{redPin}
-	err := red.Configure()
-	checkError(err, "failed to configure red pin")
+	// Configure the PWM with the given period.
+	err := pwm.Configure(machine.PWMConfig{
+		Period: 16384e3, // 16.384ms
+	})
+	if err != nil {
+		println("failed to configure PWM")
+		return
+	}

-	green := machine.PWM{greenPin}
-	err = green.Configure()
-	checkError(err, "failed to configure green pin")
+	// The top value is the highest value that can be passed to PWMChannel.Set.
+	// It is usually an even number.
+	println("top:", pwm.Top())

-	blue := machine.PWM{bluePin}
-	err = blue.Configure()
-	checkError(err, "failed to configure blue pin")
+	// Configure the two channels we'll use as outputs.
+	channelA, err := pwm.Channel(pinA)
+	if err != nil {
+		println("failed to configure channel A")
+		return
+	}
+	channelB, err := pwm.Channel(pinB)
+	if err != nil {
+		println("failed to configure channel B")
+		return
+	}

-	var rc uint8
-	var gc uint8 = 20
-	var bc uint8 = 30
+	// Invert one of the channels to demonstrate output polarity.
+	pwm.SetInverting(channelB, true)
+
+	// Test out various frequencies below, including some edge cases.
+
+	println("running at 0% duty cycle")
+	pwm.Set(channelA, 0)
+	pwm.Set(channelB, 0)
+	time.Sleep(delayBetweenPeriods)
+
+	println("running at 1")
+	pwm.Set(channelA, 1)
+	pwm.Set(channelB, 1)
+	time.Sleep(delayBetweenPeriods)
+
+	println("running at 25% duty cycle")
+	pwm.Set(channelA, pwm.Top()/4)
+	pwm.Set(channelB, pwm.Top()/4)
+	time.Sleep(delayBetweenPeriods)
+
+	println("running at top-1")
+	pwm.Set(channelA, pwm.Top()-1)
+	pwm.Set(channelB, pwm.Top()-1)
+	time.Sleep(delayBetweenPeriods)
+
+	println("running at 100% duty cycle")
+	pwm.Set(channelA, pwm.Top())
+	pwm.Set(channelB, pwm.Top())
+	time.Sleep(delayBetweenPeriods)

 	for {
-		rc = cycleColor(rc)
-		gc = cycleColor(gc)
-		bc = cycleColor(bc)
-
-		red.Set(uint16(rc) << 8)
-		green.Set(uint16(gc) << 8)
-		blue.Set(uint16(bc) << 8)
-
-		time.Sleep(time.Millisecond * 500)
-	}
-}
-
-func checkError(err error, msg string) {
-	if err != nil {
-		print(msg, ": ", err.Error())
-		println()
+		time.Sleep(time.Second)
 	}
 }
@@ -11,7 +11,7 @@ import (
 type rawState uint8

 //export llvm.coro.resume
-func (s *rawState) resume()
+func coroResume(*rawState)

 type state struct{ *rawState }

@@ -20,7 +20,7 @@ func noopState() *rawState

 // Resume the task until it pauses or completes.
 func (t *Task) Resume() {
-	t.state.resume()
+	coroResume(t.state.rawState)
 }

 // setState is used by the compiler to set the state of the function at the beginning of a function call.
@@ -77,22 +77,9 @@ func Current() *Task
 // This is implemented inside the compiler.
 func Pause()

-type taskHolder interface {
-	setState(*rawState) *rawState
-	returnTo(*rawState)
-	returnCurrent()
-	setReturnPtr(unsafe.Pointer)
-	getReturnPtr() unsafe.Pointer
-}
-
-// If there are no direct references to the task methods, they will not be discovered by the compiler, and this will trigger a compiler error.
-// Instantiating this interface forces discovery of these methods.
-var _ = taskHolder((*Task)(nil))
-
 func fake() {
 	// Hack to ensure intrinsics are discovered.
 	Current()
-	go func() {}()
 	Pause()
 }

@@ -0,0 +1,58 @@
+.section .text.tinygo_startTask
+.global  tinygo_startTask
+.type    tinygo_startTask, %function
+tinygo_startTask:
+    .cfi_startproc
+    // Small assembly stub for starting a goroutine. This is already run on the
+    // new stack, with the callee-saved registers already loaded.
+    // Most importantly, EBX contain the pc of the to-be-started function and
+    // ESI contain the only argument it is given. Multiple arguments are packed
+    // into one by storing them in a new allocation.
+
+    // Indicate to the unwinder that there is nothing to unwind, this is the
+    // root frame. It avoids bogus extra frames in GDB.
+    .cfi_undefined eip
+
+    // Set the first argument of the goroutine start wrapper, which contains all
+    // the arguments.
+    pushl %esi
+
+    // Branch to the "goroutine start" function.
+    calll *%ebx
+
+    // Rebalance the stack (to undo the above push).
+    addl $4, %esp
+
+    // After return, exit this goroutine. This is a tail call.
+    jmp tinygo_pause
+    .cfi_endproc
+
+.global tinygo_swapTask
+.type tinygo_swapTask, %function
+tinygo_swapTask:
+    // This function gets the following parameters:
+    movl 4(%esp), %eax // newStack uintptr
+    movl 8(%esp), %ecx // oldStack *uintptr
+    // More information on the calling convention:
+    // https://wiki.osdev.org/System_V_ABI#i386
+
+    // Save all callee-saved registers:
+    pushl %ebp
+    pushl %edi
+    pushl %esi
+    pushl %ebx
+
+    // Save the current stack pointer in oldStack.
+    movl %esp, (%ecx)
+
+    // Switch to the new stack pointer.
+    movl %eax, %esp
+
+    // Load saved register from the new stack.
+    popl %ebx
+    popl %esi
+    popl %edi
+    popl %ebp
+
+    // Return into the new task, as if tinygo_swapTask was a regular call.
+    ret
@@ -0,0 +1,59 @@
+// +build scheduler.tasks,386
+
+package task
+
+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_386.S that relies on the exact
+// layout of this struct.
+type calleeSavedRegs struct {
+	ebx uintptr
+	esi uintptr
+	edi uintptr
+	ebp uintptr
+
+	pc uintptr
+}
+
+// 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_386.S). This assembly code calls a function
+	// (passed in EBX) with a single argument (passed in ESI). After the
+	// function returns, it calls Pause().
+	r.pc = uintptr(unsafe.Pointer(&startTask))
+
+	// Pass the function to call in EBX.
+	// 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.ebx = fn
+
+	// Pass the pointer to the arguments struct in ESI.
+	r.esi = 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
+}
@@ -0,0 +1,74 @@
+#ifdef __MACH__ // Darwin
+.global  _tinygo_startTask
+_tinygo_startTask:
+#else // Linux etc
+.section .text.tinygo_startTask
+.global  tinygo_startTask
+tinygo_startTask:
+#endif
+    .cfi_startproc
+    // Small assembly stub for starting a goroutine. This is already run on the
+    // new stack, with the callee-saved registers already loaded.
+    // Most importantly, r12 contain the pc of the to-be-started function and
+    // r13 contain the only argument it is given. Multiple arguments are packed
+    // into one by storing them in a new allocation.
+
+    // Indicate to the unwinder that there is nothing to unwind, this is the
+    // root frame. It avoids bogus extra frames in GDB like here:
+    //     #10 0x00000000004277b6 in <goroutine wrapper> () at [...]
+    //     #11 0x00000000004278f3 in tinygo_startTask () at [...]
+    //     #12 0x0000000000002030 in ?? ()
+    //     #13 0x0000000000000071 in ?? ()
+    .cfi_undefined rip
+
+    // Set the first argument of the goroutine start wrapper, which contains all
+    // the arguments.
+    movq %r13, %rdi
+
+    // Branch to the "goroutine start" function.
+    callq *%r12
+
+    // After return, exit this goroutine. This is a tail call.
+    #ifdef __MACH__
+    jmp _tinygo_pause
+    #else
+    jmp tinygo_pause
+    #endif
+    .cfi_endproc
+
+#ifdef __MACH__ // Darwin
+.global _tinygo_swapTask
+_tinygo_swapTask:
+#else // Linux etc
+.global tinygo_swapTask
+.section .text.tinygo_swapTask
+tinygo_swapTask:
+#endif
+    // This function gets the following parameters:
+    // %rdi = newStack uintptr
+    // %rsi = oldStack *uintptr
+
+    // Save all callee-saved registers:
+    pushq %r15
+    pushq %r14
+    pushq %r13
+    pushq %r12
+    pushq %rbp
+    pushq %rbx
+
+    // Save the current stack pointer in oldStack.
+    movq %rsp, (%rsi)
+
+    // Switch to the new stack pointer.
+    movq %rdi, %rsp
+
+    // Load saved register from the new stack.
+    popq %rbx
+    popq %rbp
+    popq %r12
+    popq %r13
+    popq %r14
+    popq %r15
+
+    // Return into the new task, as if tinygo_swapTask was a regular call.
+    ret
@@ -0,0 +1,61 @@
+// +build scheduler.tasks,amd64
+
+package task
+
+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_amd64.S that relies on the exact
+// layout of this struct.
+type calleeSavedRegs struct {
+	rbx uintptr
+	rbp uintptr
+	r12 uintptr
+	r13 uintptr
+	r14 uintptr
+	r15 uintptr
+
+	pc uintptr
+}
+
+// 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_amd64.S). This assembly code calls a
+	// function (passed in r12) with a single argument (passed in r13). After
+	// the function returns, it calls Pause().
+	r.pc = uintptr(unsafe.Pointer(&startTask))
+
+	// Pass the function to call in r12.
+	// 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.r12 = fn
+
+	// Pass the pointer to the arguments struct in r13.
+	r.r13 = 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
+}
@@ -0,0 +1,51 @@
+// Only generate .debug_frame, don't generate .eh_frame.
+.cfi_sections .debug_frame
+
+.section .text.tinygo_startTask
+.global  tinygo_startTask
+.type    tinygo_startTask, %function
+tinygo_startTask:
+    .cfi_startproc
+    // 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.
+
+    // Indicate to the unwinder that there is nothing to unwind, this is the
+    // root frame. It avoids the following (bogus) error message in GDB:
+    //     Backtrace stopped: previous frame identical to this frame (corrupt stack?)
+    .cfi_undefined lr
+
+    // Set the first argument of the goroutine start wrapper, which contains all
+    // the arguments.
+    mov   r0, r5
+
+    // Branch to the "goroutine start" function. By using blx instead of bx,
+    // we'll return here instead of tail calling.
+    blx   r4
+
+    // After return, exit this goroutine. This is a tail call.
+    bl    tinygo_pause
+    .cfi_endproc
+.size tinygo_startTask, .-tinygo_startTask
+
+.global tinygo_swapTask
+.type tinygo_swapTask, %function
+tinygo_swapTask:
+    // This function gets the following parameters:
+    // r0 = newStack uintptr
+    // r1 = oldStack *uintptr
+
+    // Save all callee-saved registers:
+    push {r4-r11, lr}
+
+    // Save the current stack pointer in oldStack.
+    str sp, [r1]
+
+    // Switch to the new stack pointer.
+    mov sp, r0
+
+    // Load state from new task and branch to the previous position in the
+    // program.
+    pop {r4-r11, pc}
@@ -0,0 +1,61 @@
+// +build scheduler.tasks,arm,!cortexm,!avr,!xtensa
+
+package task
+
+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_arm.S that relies on the exact
+// layout of this struct.
+type calleeSavedRegs struct {
+	r4  uintptr
+	r5  uintptr
+	r6  uintptr
+	r7  uintptr
+	r8  uintptr
+	r9  uintptr
+	r10 uintptr
+	r11 uintptr
+
+	pc uintptr
+}
+
+// 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_arm.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))
+
+	// Pass the function to call in r4.
+	// 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.r4 = fn
+
+	// Pass the pointer to the arguments struct in r5.
+	r.r5 = 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
+}
@@ -0,0 +1,59 @@
+.section .text.tinygo_startTask
+.global  tinygo_startTask
+.type    tinygo_startTask, %function
+tinygo_startTask:
+    .cfi_startproc
+    // Small assembly stub for starting a goroutine. This is already run on the
+    // new stack, with the callee-saved registers already loaded.
+    // Most importantly, x19 contains the pc of the to-be-started function and
+    // x20 contains the only argument it is given. Multiple arguments are packed
+    // into one by storing them in a new allocation.
+
+    // Indicate to the unwinder that there is nothing to unwind, this is the
+    // root frame. It avoids the following (bogus) error message in GDB:
+    //     Backtrace stopped: previous frame identical to this frame (corrupt stack?)
+    .cfi_undefined lr
+
+    // Set the first argument of the goroutine start wrapper, which contains all
+    // the arguments.
+    mov   x0, x20
+
+    // Branch to the "goroutine start" function. By using blx instead of bx,
+    // we'll return here instead of tail calling.
+    blr   x19
+
+    // After return, exit this goroutine. This is a tail call.
+    b     tinygo_pause
+    .cfi_endproc
+.size tinygo_startTask, .-tinygo_startTask
+
+.global tinygo_swapTask
+.type tinygo_swapTask, %function
+tinygo_swapTask:
+    // This function gets the following parameters:
+    // x0 = newStack uintptr
+    // x1 = oldStack *uintptr
+
+    // Save all callee-saved registers:
+    stp     x19, x20, [sp, #-96]!
+    stp     x21, x22, [sp, #16]
+    stp     x23, x24, [sp, #32]
+    stp     x25, x26, [sp, #48]
+    stp     x27, x28, [sp, #64]
+    stp     x29, x30, [sp, #80]
+
+    // Save the current stack pointer in oldStack.
+    mov x8, sp
+    str x8, [x1]
+
+    // Switch to the new stack pointer.
+    mov sp, x0
+
+    // Restore stack state and return.
+    ldp     x29, x30, [sp, #80]
+    ldp     x27, x28, [sp, #64]
+    ldp     x25, x26, [sp, #48]
+    ldp     x23, x24, [sp, #32]
+    ldp     x21, x22, [sp, #16]
+    ldp     x19, x20, [sp], #96
+    ret
@@ -0,0 +1,64 @@
+// +build scheduler.tasks,arm64
+
+package task
+
+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_arm64.S that relies on the exact
+// layout of this struct.
+type calleeSavedRegs struct {
+	x19 uintptr
+	x20 uintptr
+	x21 uintptr
+	x22 uintptr
+	x23 uintptr
+	x24 uintptr
+	x25 uintptr
+	x26 uintptr
+	x27 uintptr
+	x28 uintptr
+	x29 uintptr
+
+	pc uintptr // aka x30 aka LR
+}
+
+// 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_arm64.S).
+	// This assembly code calls a function (passed in x19) with a single argument
+	// (passed in x20). After the function returns, it calls Pause().
+	r.pc = uintptr(unsafe.Pointer(&startTask))
+
+	// Pass the function to call in x19.
+	// 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.x19 = fn
+
+	// Pass the pointer to the arguments struct in x20.
+	r.x20 = 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
+}
@@ -2,6 +2,12 @@

 package task

+// Note that this is almost the same as task_stack_arm.go, but it uses the MSP
+// register to store the system stack pointer instead of a global variable. The
+// big advantage of this is that interrupts always execute with MSP (and not
+// PSP, which is used for goroutines) so that goroutines do not need extra stack
+// space for interrupts.
+
 import (
 	"device/arm"
 	"unsafe"
@@ -0,0 +1,104 @@
+// +build arduino_mega1280
+
+package machine
+
+// Return the current CPU frequency in hertz.
+func CPUFrequency() uint32 {
+	return 16000000
+}
+
+const (
+	AREF Pin = NoPin
+	LED  Pin = PB7
+
+	A0  Pin = PF0
+	A1  Pin = PF1
+	A2  Pin = PF2
+	A3  Pin = PF3
+	A4  Pin = PF4
+	A5  Pin = PF5
+	A6  Pin = PF6
+	A7  Pin = PF7
+	A8  Pin = PK0
+	A9  Pin = PK1
+	A10 Pin = PK2
+	A11 Pin = PK3
+	A12 Pin = PK4
+	A13 Pin = PK5
+	A14 Pin = PK6
+	A15 Pin = PK7
+
+	// Analog Input
+	ADC0  Pin = PF0
+	ADC1  Pin = PF1
+	ADC2  Pin = PF2
+	ADC3  Pin = PF3
+	ADC4  Pin = PF4
+	ADC5  Pin = PF5
+	ADC6  Pin = PF6
+	ADC7  Pin = PF7
+	ADC8  Pin = PK0
+	ADC9  Pin = PK1
+	ADC10 Pin = PK2
+	ADC11 Pin = PK3
+	ADC12 Pin = PK4
+	ADC13 Pin = PK5
+	ADC14 Pin = PK6
+	ADC15 Pin = PK7
+
+	// Digital pins
+	D0  Pin = PE0
+	D1  Pin = PE1
+	D2  Pin = PE4
+	D3  Pin = PE5
+	D4  Pin = PG5
+	D5  Pin = PE3
+	D6  Pin = PH3
+	D7  Pin = PH4
+	D8  Pin = PH5
+	D9  Pin = PH6
+	D10 Pin = PB4
+	D11 Pin = PB5
+	D12 Pin = PB6
+	D13 Pin = PB7
+	D14 Pin = PJ1
+	D15 Pin = PJ0
+	D16 Pin = PH1
+	D17 Pin = PH0
+	D18 Pin = PD3
+	D19 Pin = PD2
+	D20 Pin = PD1
+	D21 Pin = PD0
+	D22 Pin = PA0
+	D23 Pin = PA1
+	D24 Pin = PA2
+	D25 Pin = PA3
+	D26 Pin = PA4
+	D27 Pin = PA5
+	D28 Pin = PA6
+	D29 Pin = PA7
+	D30 Pin = PC7
+	D31 Pin = PC6
+	D32 Pin = PC5
+	D33 Pin = PC4
+	D34 Pin = PC3
+	D35 Pin = PC2
+	D36 Pin = PC1
+	D37 Pin = PC0
+	D38 Pin = PD7
+	D39 Pin = PG2
+	D40 Pin = PG1
+	D41 Pin = PG0
+	D42 Pin = PL7
+	D43 Pin = PL6
+	D44 Pin = PL5
+	D45 Pin = PL4
+	D46 Pin = PL3
+	D47 Pin = PL2
+	D48 Pin = PL1
+	D49 Pin = PL0
+	D50 Pin = PB3
+	D51 Pin = PB2
+	D52 Pin = PB1
+	D53 Pin = PB0
+)
@@ -32,7 +32,7 @@ func init() {

 // I2C on the Arduino Nano 33.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM4_I2CM,
 		SERCOM: 4,
 	}
@@ -0,0 +1,341 @@
+// +build atsame54_xpro
+
+package machine
+
+import (
+	"device/sam"
+	"runtime/interrupt"
+)
+
+// Definition for compatibility, but not used
+const RESET_MAGIC_VALUE = 0x00000000
+
+const (
+	LED    = PC18
+	BUTTON = PB31
+)
+
+const (
+	// https://ww1.microchip.com/downloads/en/DeviceDoc/70005321A.pdf
+
+	// Extension Header EXT1
+	EXT1_PIN3_ADC_P     = PB04
+	EXT1_PIN4_ADC_N     = PB05
+	EXT1_PIN5_GPIO1     = PA06
+	EXT1_PIN6_GPIO2     = PA07
+	EXT1_PIN7_PWM_P     = PB08
+	EXT1_PIN8_PWM_N     = PB09
+	EXT1_PIN9_IRQ       = PB07
+	EXT1_PIN9_GPIO      = PB07
+	EXT1_PIN10_SPI_SS_B = PA27
+	EXT1_PIN10_GPIO     = PA27
+	EXT1_PIN11_TWI_SDA  = PA22
+	EXT1_PIN12_TWI_SCL  = PA23
+	EXT1_PIN13_UART_RX  = PA05
+	EXT1_PIN14_UART_TX  = PA04
+	EXT1_PIN15_SPI_SS_A = PB28
+	EXT1_PIN16_SPI_SDO  = PB27
+	EXT1_PIN17_SPI_SDI  = PB29
+	EXT1_PIN18_SPI_SCK  = PB26
+
+	// Extension Header EXT2
+	EXT2_PIN3_ADC_P     = PB00
+	EXT2_PIN4_ADC_N     = PA03
+	EXT2_PIN5_GPIO1     = PB01
+	EXT2_PIN6_GPIO2     = PB06
+	EXT2_PIN7_PWM_P     = PB14
+	EXT2_PIN8_PWM_N     = PB15
+	EXT2_PIN9_IRQ       = PD00
+	EXT2_PIN9_GPIO      = PD00
+	EXT2_PIN10_SPI_SS_B = PB02
+	EXT2_PIN10_GPIO     = PB02
+	EXT2_PIN11_TWI_SDA  = PD08
+	EXT2_PIN12_TWI_SCL  = PD09
+	EXT2_PIN13_UART_RX  = PB17
+	EXT2_PIN14_UART_TX  = PB16
+	EXT2_PIN15_SPI_SS_A = PC06
+	EXT2_PIN16_SPI_SDO  = PC04
+	EXT2_PIN17_SPI_SDI  = PC07
+	EXT2_PIN18_SPI_SCK  = PC05
+
+	// Extension Header EXT3
+	EXT3_PIN3_ADC_P     = PC02
+	EXT3_PIN4_ADC_N     = PC03
+	EXT3_PIN5_GPIO1     = PC01
+	EXT3_PIN6_GPIO2     = PC10
+	EXT3_PIN7_PWM_P     = PD10
+	EXT3_PIN8_PWM_N     = PD11
+	EXT3_PIN9_IRQ       = PC30
+	EXT3_PIN9_GPIO      = PC30
+	EXT3_PIN10_SPI_SS_B = PC31
+	EXT3_PIN10_GPIO     = PC31
+	EXT3_PIN11_TWI_SDA  = PD08
+	EXT3_PIN12_TWI_SCL  = PD09
+	EXT3_PIN13_UART_RX  = PC23
+	EXT3_PIN14_UART_TX  = PC22
+	EXT3_PIN15_SPI_SS_A = PC14
+	EXT3_PIN16_SPI_SDO  = PC04
+	EXT3_PIN17_SPI_SDI  = PC07
+	EXT3_PIN18_SPI_SCK  = PC05
+
+	// SD_CARD
+	SD_CARD_MCDA0   = PB18
+	SD_CARD_MCDA1   = PB19
+	SD_CARD_MCDA2   = PB20
+	SD_CARD_MCDA3   = PB21
+	SD_CARD_MCCK    = PA21
+	SD_CARD_MCCDA   = PA20
+	SD_CARD_DETECT  = PD20
+	SD_CARD_PROTECT = PD21
+
+	// I2C
+	I2C_SDA = PD08
+	I2C_SCL = PD09
+
+	// CAN
+	CAN0_TX = PA22
+	CAN0_RX = PA23
+
+	CAN1_STANDBY = PC13
+	CAN1_TX      = PB12
+	CAN1_RX      = PB13
+
+	CAN_STANDBY = CAN1_STANDBY
+	CAN_TX      = CAN1_TX
+	CAN_RX      = CAN1_RX
+
+	// PDEC
+	PDEC_PHASE_A = PC16
+	PDEC_PHASE_B = PC17
+	PDEC_INDEX   = PC18
+
+	// PCC
+	PCC_I2C_SDA    = PD08
+	PCC_I2C_SCL    = PD09
+	PCC_VSYNC_DEN1 = PA12
+	PCC_HSYNC_DEN2 = PA13
+	PCC_CLK        = PA14
+	PCC_XCLK       = PA15
+	PCC_DATA00     = PA16
+	PCC_DATA01     = PA17
+	PCC_DATA02     = PA18
+	PCC_DATA03     = PA19
+	PCC_DATA04     = PA20
+	PCC_DATA05     = PA21
+	PCC_DATA06     = PA22
+	PCC_DATA07     = PA23
+	PCC_DATA08     = PB14
+	PCC_DATA09     = PB15
+	PCC_RESET      = PC12
+	PCC_PWDN       = PC11
+
+	// Ethernet
+	ETHERNET_TXCK  = PA14
+	ETHERNET_TXEN  = PA17
+	ETHERNET_TX0   = PA18
+	ETHERNET_TX1   = PA19
+	ETHERNET_RXER  = PA15
+	ETHERNET_RX0   = PA13
+	ETHERNET_RX1   = PA12
+	ETHERNET_RXDV  = PC20
+	ETHERNET_MDIO  = PC12
+	ETHERNET_MDC   = PC11
+	ETHERNET_INT   = PD12
+	ETHERNET_RESET = PC21
+
+	PIN_QT_BUTTON   = PA16
+	PIN_BTN0        = PB31
+	PIN_ETH_LED     = PC15
+	PIN_LED0        = PC18
+	PIN_ADC_DAC     = PA02
+	PIN_VBUS_DETECT = PC00
+	PIN_USB_ID      = PC19
+)
+
+// UART0 aka USBCDC pins
+const (
+	USBCDC_DM_PIN = PA24
+	USBCDC_DP_PIN = PA25
+)
+
+// UART pins
+const (
+	// Extension Header EXT1
+	UART_TX_PIN = PA04 // TX : SERCOM0/PAD[0]
+	UART_RX_PIN = PA05 // RX : SERCOM0/PAD[1]
+
+	// Extension Header EXT2
+	UART2_TX_PIN = PB16 // TX : SERCOM5/PAD[0]
+	UART2_RX_PIN = PB17 // RX : SERCOM5/PAD[1]
+
+	// Extension Header EXT3
+	UART3_TX_PIN = PC22 // TX : SERCOM1/PAD[0]
+	UART3_RX_PIN = PC23 // RX : SERCOM1/PAD[1]
+
+	// Virtual COM Port
+	UART4_TX_PIN = PB25 // TX : SERCOM2/PAD[0]
+	UART4_RX_PIN = PB24 // RX : SERCOM2/PAD[1]
+)
+
+// I2C pins
+const (
+	// Extension Header EXT1
+	SDA0_PIN = PA22 // SDA: SERCOM3/PAD[0]
+	SCL0_PIN = PA23 // SCL: SERCOM3/PAD[1]
+
+	// Extension Header EXT2
+	SDA1_PIN = PD08 // SDA: SERCOM7/PAD[0]
+	SCL1_PIN = PD09 // SCL: SERCOM7/PAD[1]
+
+	// Extension Header EXT3
+	SDA2_PIN = PD08 // SDA: SERCOM7/PAD[0]
+	SCL2_PIN = PD09 // SCL: SERCOM7/PAD[1]
+
+	// Data Gateway Interface
+	SDA_DGI_PIN = PD08 // SDA: SERCOM7/PAD[0]
+	SCL_DGI_PIN = PD09 // SCL: SERCOM7/PAD[1]
+
+	SDA_PIN = SDA0_PIN
+	SCL_PIN = SCL0_PIN
+)
+
+// SPI pins
+const (
+	// Extension Header EXT1
+	SPI0_SCK_PIN = PB26 // SCK: SERCOM4/PAD[1]
+	SPI0_SDO_PIN = PB27 // SDO: SERCOM4/PAD[0]
+	SPI0_SDI_PIN = PB29 // SDI: SERCOM4/PAD[3]
+	SPI0_SS_PIN  = PB28 // SS : SERCOM4/PAD[2]
+
+	// Extension Header EXT2
+	SPI1_SCK_PIN = PC05 // SCK: SERCOM6/PAD[1]
+	SPI1_SDO_PIN = PC04 // SDO: SERCOM6/PAD[0]
+	SPI1_SDI_PIN = PC07 // SDI: SERCOM6/PAD[3]
+	SPI1_SS_PIN  = PC06 // SS : SERCOM6/PAD[2]
+
+	// Extension Header EXT3
+	SPI2_SCK_PIN = PC05 // SCK: SERCOM6/PAD[1]
+	SPI2_SDO_PIN = PC04 // SDO: SERCOM6/PAD[0]
+	SPI2_SDI_PIN = PC07 // SDI: SERCOM6/PAD[3]
+	SPI2_SS_PIN  = PC14 // SS : GPIO
+
+	// Data Gateway Interface
+	SPI_DGI_SCK_PIN = PC05 // SCK: SERCOM6/PAD[1]
+	SPI_DGI_SDO_PIN = PC04 // SDO: SERCOM6/PAD[0]
+	SPI_DGI_SDI_PIN = PC07 // SDI: SERCOM6/PAD[3]
+	SPI_DGI_SS_PIN  = PD01 // SS : GPIO
+)
+
+// USB CDC identifiers
+const (
+	usb_STRING_PRODUCT      = "SAM E54 Xplained Pro"
+	usb_STRING_MANUFACTURER = "Atmel"
+)
+
+var (
+	usb_VID uint16 = 0x03EB
+	usb_PID uint16 = 0x2404
+)
+
+// UART on the SAM E54 Xplained Pro
+var (
+	// Extension Header EXT1
+	UART1 = UART{
+		Buffer: NewRingBuffer(),
+		Bus:    sam.SERCOM0_USART_INT,
+		SERCOM: 0,
+	}
+
+	// Extension Header EXT2
+	UART2 = UART{
+		Buffer: NewRingBuffer(),
+		Bus:    sam.SERCOM5_USART_INT,
+		SERCOM: 5,
+	}
+
+	// Extension Header EXT3
+	UART3 = UART{
+		Buffer: NewRingBuffer(),
+		Bus:    sam.SERCOM1_USART_INT,
+		SERCOM: 1,
+	}
+
+	// EDBG Virtual COM Port
+	UART4 = UART{
+		Buffer: NewRingBuffer(),
+		Bus:    sam.SERCOM2_USART_INT,
+		SERCOM: 2,
+	}
+)
+
+func init() {
+	UART1.Interrupt = interrupt.New(sam.IRQ_SERCOM0_2, UART1.handleInterrupt)
+	UART2.Interrupt = interrupt.New(sam.IRQ_SERCOM5_2, UART2.handleInterrupt)
+	UART3.Interrupt = interrupt.New(sam.IRQ_SERCOM1_2, UART3.handleInterrupt)
+	UART4.Interrupt = interrupt.New(sam.IRQ_SERCOM2_2, UART4.handleInterrupt)
+}
+
+// I2C on the SAM E54 Xplained Pro
+var (
+	// Extension Header EXT1
+	I2C0 = I2C{
+		Bus:    sam.SERCOM3_I2CM,
+		SERCOM: 3,
+	}
+
+	// Extension Header EXT2
+	I2C1 = I2C{
+		Bus:    sam.SERCOM7_I2CM,
+		SERCOM: 7,
+	}
+
+	// Extension Header EXT3
+	I2C2 = I2C{
+		Bus:    sam.SERCOM7_I2CM,
+		SERCOM: 7,
+	}
+
+	// Data Gateway Interface
+	I2C3 = I2C{
+		Bus:    sam.SERCOM7_I2CM,
+		SERCOM: 7,
+	}
+)
+
+// SPI on the SAM E54 Xplained Pro
+var (
+	// Extension Header EXT1
+	SPI0 = SPI{
+		Bus:    sam.SERCOM4_SPIM,
+		SERCOM: 4,
+	}
+
+	// Extension Header EXT2
+	SPI1 = SPI{
+		Bus:    sam.SERCOM6_SPIM,
+		SERCOM: 6,
+	}
+
+	// Extension Header EXT3
+	SPI2 = SPI{
+		Bus:    sam.SERCOM6_SPIM,
+		SERCOM: 6,
+	}
+
+	// Data Gateway Interface
+	SPI3 = SPI{
+		Bus:    sam.SERCOM6_SPIM,
+		SERCOM: 6,
+	}
+)
+
+// CAN on the SAM E54 Xplained Pro
+var (
+	CAN0 = CAN{
+		Bus: sam.CAN0,
+	}
+
+	CAN1 = CAN{
+		Bus: sam.CAN1,
+	}
+)
@@ -85,6 +85,6 @@ const (

 // I2C pins
 const (
-	SDA_PIN = PB7
-	SCL_PIN = PB6
+	I2C0_SDA_PIN = PB7
+	I2C0_SCL_PIN = PB6
 )
@@ -25,8 +25,8 @@ const (

 // Analog Pins
 const (
-	A0  = PA02 // PWM available, also ADC/AIN[0]
-	A1  = PA05 // ADC/AIN[5]
+	A0  = PA02 // ADC/AIN[0]
+	A1  = PA05 // PWM available, also ADC/AIN[5]
 	A2  = PA06 // PWM available, also ADC/AIN[6]
 	A3  = PA07 // PWM available, also ADC/AIN[7]
 	A4  = PB03 // PORTB
@@ -23,12 +23,12 @@ func init() {
 // I2C on the Circuit Playground Express.
 var (
 	// external device
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM5_I2CM,
 		SERCOM: 5,
 	}
 	// internal device
-	I2C1 = I2C{
+	I2C1 = &I2C{
 		Bus:    sam.SERCOM1_I2CM,
 		SERCOM: 1,
 	}
@@ -75,7 +75,7 @@ const (

 // I2C on the Feather M0.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM3_I2CM,
 		SERCOM: 3,
 	}
@@ -0,0 +1,153 @@
+// +build feather_m4_can
+
+package machine
+
+import (
+	"device/sam"
+	"runtime/interrupt"
+)
+
+// used to reset into bootloader
+const RESET_MAGIC_VALUE = 0xf01669ef
+
+// GPIO Pins
+const (
+	D0  = PB17 // UART0 RX/PWM available
+	D1  = PB16 // UART0 TX/PWM available
+	D4  = PA14 // PWM available
+	D5  = PA16 // PWM available
+	D6  = PA18 // PWM available
+	D7  = PB03 // neopixel power
+	D8  = PB02 // built-in neopixel
+	D9  = PA19 // PWM available
+	D10 = PA20 // can be used for PWM or UART1 TX
+	D11 = PA21 // can be used for PWM or UART1 RX
+	D12 = PA22 // PWM available
+	D13 = PA23 // PWM available
+	D21 = PA13 // PWM available
+	D22 = PA12 // PWM available
+	D23 = PB22 // PWM available
+	D24 = PB23 // PWM available
+	D25 = PA17 // PWM available
+)
+
+// Analog pins
+const (
+	A0 = PA02 // ADC/AIN[0]
+	A1 = PA05 // ADC/AIN[2]
+	A2 = PB08 // ADC/AIN[3]
+	A3 = PB09 // ADC/AIN[4]
+	A4 = PA04 // ADC/AIN[5]
+	A5 = PA06 // ADC/AIN[10]
+)
+
+const (
+	LED       = D13
+	NEOPIXELS = D8
+)
+
+// UART0 aka USBCDC pins
+const (
+	USBCDC_DM_PIN = PA24
+	USBCDC_DP_PIN = PA25
+)
+
+const (
+	UART_TX_PIN = D1
+	UART_RX_PIN = D0
+)
+
+const (
+	UART2_TX_PIN = A4
+	UART2_RX_PIN = A5
+)
+
+// I2C pins
+const (
+	SDA_PIN = D22 // SDA: SERCOM2/PAD[0]
+	SCL_PIN = D21 // SCL: SERCOM2/PAD[1]
+)
+
+// SPI pins
+const (
+	SPI0_SCK_PIN = D25 // SCK: SERCOM1/PAD[1]
+	SPI0_SDO_PIN = D24 // SDO: SERCOM1/PAD[3]
+	SPI0_SDI_PIN = D23 // SDI: SERCOM1/PAD[2]
+)
+
+// CAN pins
+const (
+	CAN0_TX = PA22
+	CAN0_RX = PA23
+
+	CAN1_STANDBY = PB12
+	CAN1_TX      = PB14
+	CAN1_RX      = PB15
+	BOOST_EN     = PB13 // power control of CAN1's TCAN1051HGV (H: enable)
+
+	CAN_STANDBY = CAN1_STANDBY
+	CAN_S       = CAN1_STANDBY
+	CAN_TX      = CAN1_TX
+	CAN_RX      = CAN1_RX
+)
+
+// USB CDC identifiers
+const (
+	usb_STRING_PRODUCT      = "Adafruit Feather M4 CAN"
+	usb_STRING_MANUFACTURER = "Adafruit"
+)
+
+var (
+	usb_VID uint16 = 0x239A
+	usb_PID uint16 = 0x80CD
+)
+
+var (
+	UART1 = UART{
+		Buffer: NewRingBuffer(),
+		Bus:    sam.SERCOM5_USART_INT,
+		SERCOM: 5,
+	}
+
+	UART2 = UART{
+		Buffer: NewRingBuffer(),
+		Bus:    sam.SERCOM0_USART_INT,
+		SERCOM: 0,
+	}
+)
+
+func init() {
+	UART1.Interrupt = interrupt.New(sam.IRQ_SERCOM5_2, UART1.handleInterrupt)
+	UART2.Interrupt = interrupt.New(sam.IRQ_SERCOM0_2, UART2.handleInterrupt)
+
+	// turn on neopixel
+	D7.Configure(PinConfig{Mode: PinOutput})
+	D7.High()
+}
+
+// I2C on the Feather M4 CAN.
+var (
+	I2C0 = &I2C{
+		Bus:    sam.SERCOM2_I2CM,
+		SERCOM: 2,
+	}
+)
+
+// SPI on the Feather M4 CAN.
+var (
+	SPI0 = SPI{
+		Bus:    sam.SERCOM1_SPIM,
+		SERCOM: 1,
+	}
+)
+
+// CAN on the Feather M4 CAN.
+var (
+	CAN0 = CAN{
+		Bus: sam.CAN0,
+	}
+
+	CAN1 = CAN{
+		Bus: sam.CAN1,
+	}
+)
@@ -28,7 +28,7 @@ func init() {

 // I2C on the Feather M4.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM2_I2CM,
 		SERCOM: 2,
 	}
@@ -120,19 +120,22 @@ const (

 var (
 	UART1 = UART{
-		Buffer:          NewRingBuffer(),
-		Bus:             stm32.USART3,
-		AltFuncSelector: AF7_USART1_2_3,
+		Buffer:            NewRingBuffer(),
+		Bus:               stm32.USART3,
+		TxAltFuncSelector: AF7_USART1_2_3,
+		RxAltFuncSelector: AF7_USART1_2_3,
 	}
 	UART2 = UART{
-		Buffer:          NewRingBuffer(),
-		Bus:             stm32.USART6,
-		AltFuncSelector: AF8_USART4_5_6,
+		Buffer:            NewRingBuffer(),
+		Bus:               stm32.USART6,
+		TxAltFuncSelector: AF8_USART4_5_6,
+		RxAltFuncSelector: AF8_USART4_5_6,
 	}
 	UART3 = UART{
-		Buffer:          NewRingBuffer(),
-		Bus:             stm32.USART1,
-		AltFuncSelector: AF7_USART1_2_3,
+		Buffer:            NewRingBuffer(),
+		Bus:               stm32.USART1,
+		TxAltFuncSelector: AF7_USART1_2_3,
+		RxAltFuncSelector: AF7_USART1_2_3,
 	}
 	UART0 = UART1
 )
@@ -227,15 +230,15 @@ const (
 )

 var (
-	I2C1 = I2C{
+	I2C1 = &I2C{
 		Bus:             stm32.I2C1,
 		AltFuncSelector: AF4_I2C1_2_3,
 	}
-	I2C2 = I2C{
+	I2C2 = &I2C{
 		Bus:             stm32.I2C2,
 		AltFuncSelector: AF4_I2C1_2_3,
 	}
-	I2C3 = I2C{
+	I2C3 = &I2C{
 		Bus:             stm32.I2C1,
 		AltFuncSelector: AF4_I2C1_2_3,
 	}
@@ -40,11 +40,11 @@ var (

 // I2C on the Grand Central M4
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM3_I2CM,
 		SERCOM: 3,
 	}
-	I2C1 = I2C{
+	I2C1 = &I2C{
 		Bus:    sam.SERCOM6_I2CM,
 		SERCOM: 6,
 	}
@@ -10,10 +10,3 @@ var (
 		Bus: sifive.QSPI1,
 	}
 )
-
-// I2C on the HiFive1 rev B.
-var (
-	I2C0 = I2C{
-		Bus: sifive.I2C0,
-	}
-)
@@ -75,7 +75,7 @@ const (

 // I2C on the ItsyBitsy M0.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM3_I2CM,
 		SERCOM: 3,
 	}
@@ -28,7 +28,7 @@ func init() {

 // I2C on the ItsyBitsy M4.
 var (
-	I2C0 = I2C{
+	I2C0 = &I2C{
 		Bus:    sam.SERCOM2_I2CM,
 		SERCOM: 2,
 	}
--- a/Show More
+++ b/Show More