compiler: add support for atomic operations

compiler/atomic.go

@@ -0,0 +1,57 @@
+package compiler
+
+import (
+	"golang.org/x/tools/go/ssa"
+	"tinygo.org/x/go-llvm"
+)
+
+// createAtomicOp lowers an atomic library call by lowering it as an LLVM atomic
+// operation. It returns the result of the operation and true if the call could
+// be lowered inline, and false otherwise.
+func (b *builder) createAtomicOp(call *ssa.CallCommon) (llvm.Value, bool) {
+	name := call.Value.(*ssa.Function).Name()
+	switch name {
+	case "AddInt32", "AddInt64", "AddUint32", "AddUint64", "AddUintptr":
+		ptr := b.getValue(call.Args[0])
+		val := b.getValue(call.Args[1])
+		oldVal := b.CreateAtomicRMW(llvm.AtomicRMWBinOpAdd, ptr, val, llvm.AtomicOrderingSequentiallyConsistent, true)
+		// Return the new value, not the original value returned by atomicrmw.
+		return b.CreateAdd(oldVal, val, ""), true
+	case "SwapInt32", "SwapInt64", "SwapUint32", "SwapUint64", "SwapUintptr", "SwapPointer":
+		ptr := b.getValue(call.Args[0])
+		val := b.getValue(call.Args[1])
+		isPointer := val.Type().TypeKind() == llvm.PointerTypeKind
+		if isPointer {
+			// atomicrmw only supports integers, so cast to an integer.
+			val = b.CreatePtrToInt(val, b.uintptrType, "")
+			ptr = b.CreateBitCast(ptr, llvm.PointerType(val.Type(), 0), "")
+		}
+		oldVal := b.CreateAtomicRMW(llvm.AtomicRMWBinOpXchg, ptr, val, llvm.AtomicOrderingSequentiallyConsistent, true)
+		if isPointer {
+			oldVal = b.CreateIntToPtr(oldVal, b.i8ptrType, "")
+		}
+		return oldVal, true
+	case "CompareAndSwapInt32", "CompareAndSwapInt64", "CompareAndSwapUint32", "CompareAndSwapUint64", "CompareAndSwapUintptr", "CompareAndSwapPointer":
+		ptr := b.getValue(call.Args[0])
+		old := b.getValue(call.Args[1])
+		newVal := b.getValue(call.Args[2])
+		tuple := b.CreateAtomicCmpXchg(ptr, old, newVal, llvm.AtomicOrderingSequentiallyConsistent, llvm.AtomicOrderingSequentiallyConsistent, true)
+		swapped := b.CreateExtractValue(tuple, 1, "")
+		return swapped, true
+	case "LoadInt32", "LoadInt64", "LoadUint32", "LoadUint64", "LoadUintptr", "LoadPointer":
+		ptr := b.getValue(call.Args[0])
+		val := b.CreateLoad(ptr, "")
+		val.SetOrdering(llvm.AtomicOrderingSequentiallyConsistent)
+		val.SetAlignment(b.targetData.PrefTypeAlignment(val.Type())) // required
+		return val, true
+	case "StoreInt32", "StoreInt64", "StoreUint32", "StoreUint64", "StoreUintptr", "StorePointer":
+		ptr := b.getValue(call.Args[0])
+		val := b.getValue(call.Args[1])
+		store := b.CreateStore(val, ptr)
+		store.SetOrdering(llvm.AtomicOrderingSequentiallyConsistent)
+		store.SetAlignment(b.targetData.PrefTypeAlignment(val.Type())) // required
+		return store, true
+	default:
+		return llvm.Value{}, false
+	}
+}

compiler/compiler.go

@@ -1323,6 +1323,14 @@ func (b *builder) createFunctionCall(instr *ssa.CallCommon) (llvm.Value, error)
 			return b.createVolatileLoad(instr)
 		case strings.HasPrefix(name, "runtime/volatile.Store"):
 			return b.createVolatileStore(instr)
+		case strings.HasPrefix(name, "sync/atomic."):
+			val, ok := b.createAtomicOp(instr)
+			if ok {
+				// This call could be lowered as an atomic operation.
+				return val, nil
+			}
+			// This call couldn't be lowered as an atomic operation, it's
+			// probably something else. Continue as usual.
 		case name == "runtime/interrupt.New":
 			return b.createInterruptGlobal(instr)
 		}

src/device/riscv/riscv.go

@@ -19,3 +19,19 @@ func Asm(asm string)
 // You can use {} in the asm string (which expands to a register) to set the
 // return value.
 func AsmFull(asm string, regs map[string]interface{}) uintptr
+
+// DisableInterrupts disables all interrupts, and returns the old interrupt
+// state.
+func DisableInterrupts() uintptr {
+	// Note: this can be optimized with a CSRRW instruction, which atomically
+	// swaps the value and returns the old value.
+	mask := MIE.Get()
+	MIE.Set(0)
+	return mask
+}
+
+// EnableInterrupts enables all interrupts again. The value passed in must be
+// the mask returned by DisableInterrupts.
+func EnableInterrupts(mask uintptr) {
+	MIE.Set(mask)
+}

src/runtime/arch_cortexm.go

@@ -19,3 +19,86 @@ func align(ptr uintptr) uintptr {
 func getCurrentStackPointer() uintptr {
 	return arm.AsmFull("mov {}, sp", nil)
 }
+
+// Documentation:
+// * https://llvm.org/docs/Atomics.html
+// * https://gcc.gnu.org/onlinedocs/gcc/_005f_005fsync-Builtins.html
+//
+// In the case of Cortex-M, some atomic operations are emitted inline while
+// others are emitted as libcalls. How many are emitted as libcalls depends on
+// the MCU core variant (M3 and higher support some 32-bit atomic operations
+// while M0 and M0+ do not).
+
+//export __sync_fetch_and_add_4
+func __sync_fetch_and_add_4(ptr *uint32, value uint32) uint32 {
+	mask := arm.DisableInterrupts()
+	oldValue := *ptr
+	*ptr = oldValue + value
+	arm.EnableInterrupts(mask)
+	return oldValue
+}
+
+//export __sync_fetch_and_add_8
+func __sync_fetch_and_add_8(ptr *uint64, value uint64) uint64 {
+	mask := arm.DisableInterrupts()
+	oldValue := *ptr
+	*ptr = oldValue + value
+	arm.EnableInterrupts(mask)
+	return oldValue
+}
+
+//export __sync_lock_test_and_set_4
+func __sync_lock_test_and_set_4(ptr *uint32, value uint32) uint32 {
+	mask := arm.DisableInterrupts()
+	oldValue := *ptr
+	*ptr = value
+	arm.EnableInterrupts(mask)
+	return oldValue
+}
+
+//export __sync_lock_test_and_set_8
+func __sync_lock_test_and_set_8(ptr *uint64, value uint64) uint64 {
+	mask := arm.DisableInterrupts()
+	oldValue := *ptr
+	*ptr = value
+	arm.EnableInterrupts(mask)
+	return oldValue
+}
+
+//export __sync_val_compare_and_swap_4
+func __sync_val_compare_and_swap_4(ptr *uint32, expected, desired uint32) uint32 {
+	mask := arm.DisableInterrupts()
+	oldValue := *ptr
+	if oldValue == expected {
+		*ptr = desired
+	}
+	arm.EnableInterrupts(mask)
+	return oldValue
+}
+
+//export __sync_val_compare_and_swap_8
+func __sync_val_compare_and_swap_8(ptr *uint64, expected, desired uint64) uint64 {
+	mask := arm.DisableInterrupts()
+	oldValue := *ptr
+	if oldValue == expected {
+		*ptr = desired
+	}
+	arm.EnableInterrupts(mask)
+	return oldValue
+}
+
+// The safest thing to do here would just be to disable interrupts for
+// procPin/procUnpin. Note that a global variable is safe in this case, as any
+// access to procPinnedMask will happen with interrupts disabled.
+
+var procPinnedMask uintptr
+
+//go:linkname procPin sync/atomic.runtime_procPin
+func procPin() {
+	procPinnedMask = arm.DisableInterrupts()
+}
+
+//go:linkname procUnpin sync/atomic.runtime_procUnpin
+func procUnpin() {
+	arm.EnableInterrupts(procPinnedMask)
+}

src/runtime/arch_tinygoriscv.go

@@ -17,3 +17,76 @@ func align(ptr uintptr) uintptr {
 func getCurrentStackPointer() uintptr {
 	return riscv.AsmFull("mv {}, sp", nil)
 }
+
+// Documentation:
+// * https://llvm.org/docs/Atomics.html
+// * https://gcc.gnu.org/onlinedocs/gcc/_005f_005fsync-Builtins.html
+//
+// In the case of RISC-V, some operations may be implemented with libcalls if
+// the operation is too big to be handled by assembly. Officially, these calls
+// should be implemented with a lock-free algorithm but as (as of this time) all
+// supported RISC-V chips have a single hart, we can simply disable interrupts
+// to get the same behavior.
+
+//export __atomic_load_8
+func __atomic_load_8(ptr *uint64, ordering int32) uint64 {
+	mask := riscv.DisableInterrupts()
+	value := *ptr
+	riscv.EnableInterrupts(mask)
+	return value
+}
+
+//export __atomic_store_8
+func __atomic_store_8(ptr *uint64, value uint64, ordering int32) {
+	mask := riscv.DisableInterrupts()
+	*ptr = value
+	riscv.EnableInterrupts(mask)
+}
+
+//export __atomic_exchange_8
+func __atomic_exchange_8(ptr *uint64, value uint64, ordering int32) uint64 {
+	mask := riscv.DisableInterrupts()
+	oldValue := *ptr
+	*ptr = value
+	riscv.EnableInterrupts(mask)
+	return oldValue
+}
+
+//export __atomic_compare_exchange_8
+func __atomic_compare_exchange_8(ptr, expected *uint64, desired uint64, success_ordering, failure_ordering int32) bool {
+	mask := riscv.DisableInterrupts()
+	oldValue := *ptr
+	success := oldValue == *expected
+	if success {
+		*ptr = desired
+	} else {
+		*expected = oldValue
+	}
+	riscv.EnableInterrupts(mask)
+	return success
+}
+
+//export __atomic_fetch_add_8
+func __atomic_fetch_add_8(ptr *uint64, value uint64, ordering int32) uint64 {
+	mask := riscv.DisableInterrupts()
+	oldValue := *ptr
+	*ptr = oldValue + value
+	riscv.EnableInterrupts(mask)
+	return oldValue
+}
+
+// The safest thing to do here would just be to disable interrupts for
+// procPin/procUnpin. Note that a global variable is safe in this case, as any
+// access to procPinnedMask will happen with interrupts disabled.
+
+var procPinnedMask uintptr
+
+//go:linkname procPin sync/atomic.runtime_procPin
+func procPin() {
+	procPinnedMask = riscv.DisableInterrupts()
+}
+
+//go:linkname procUnpin sync/atomic.runtime_procUnpin
+func procUnpin() {
+	riscv.EnableInterrupts(procPinnedMask)
+}

src/runtime/runtime_unix.go

@@ -94,3 +94,14 @@ func extalloc(size uintptr) unsafe.Pointer {
 
 //export free
 func extfree(ptr unsafe.Pointer)
+
+// TinyGo does not yet support any form of parallelism on an OS, so these can be
+// left empty.
+
+//go:linkname procPin sync/atomic.runtime_procPin
+func procPin() {
+}
+
+//go:linkname procUnpin sync/atomic.runtime_procUnpin
+func procUnpin() {
+}

src/runtime/runtime_wasm.go

@@ -91,3 +91,14 @@ func ticks() timeUnit
 func abort() {
 	trap()
 }
+
+// TinyGo does not yet support any form of parallelism on WebAssembly, so these
+// can be left empty.
+
+//go:linkname procPin sync/atomic.runtime_procPin
+func procPin() {
+}
+
+//go:linkname procUnpin sync/atomic.runtime_procUnpin
+func procUnpin() {
+}