compiler: avoid zero-sized alloca in channel operations

This works around a bug in LLVM
(https://bugs.llvm.org/show_bug.cgi?id=49916) but seems like a good
change in general.

5 files changed, 169 insertions, 10 deletions

diff --git a/compiler/channel.go b/compiler/channel.go
index a036fbf45..a9886e100 100644
--- a/compiler/channel.go
+++ b/compiler/channel.go
@@ -32,8 +32,15 @@ func (b *builder) createChanSend(instr *ssa.Send) {
 
 	// store value-to-send
 	valueType := b.getLLVMType(instr.X.Type())
-	valueAlloca, valueAllocaCast, valueAllocaSize := b.createTemporaryAlloca(valueType, "chan.value")
-	b.CreateStore(chanValue, valueAlloca)
+	isZeroSize := b.targetData.TypeAllocSize(valueType) == 0
+	var valueAlloca, valueAllocaCast, valueAllocaSize llvm.Value
+	if isZeroSize {
+		valueAlloca = llvm.ConstNull(llvm.PointerType(valueType, 0))
+		valueAllocaCast = llvm.ConstNull(b.i8ptrType)
+	} else {
+		valueAlloca, valueAllocaCast, valueAllocaSize = b.createTemporaryAlloca(valueType, "chan.value")
+		b.CreateStore(chanValue, valueAlloca)
+	}
 
 	// Allocate blockedlist buffer.
 	channelBlockedList := b.mod.GetTypeByName("runtime.channelBlockedList")
@@ -46,7 +53,9 @@ func (b *builder) createChanSend(instr *ssa.Send) {
 	// This also works around a bug in CoroSplit, at least in LLVM 8:
 	// https://bugs.llvm.org/show_bug.cgi?id=41742
 	b.emitLifetimeEnd(channelBlockedListAllocaCast, channelBlockedListAllocaSize)
-	b.emitLifetimeEnd(valueAllocaCast, valueAllocaSize)
+	if !isZeroSize {
+		b.emitLifetimeEnd(valueAllocaCast, valueAllocaSize)
+	}
 }
 
 // createChanRecv emits a pseudo chan receive operation. It is lowered to the
@@ -56,7 +65,14 @@ func (b *builder) createChanRecv(unop *ssa.UnOp) llvm.Value {
 	ch := b.getValue(unop.X)
 
 	// Allocate memory to receive into.
-	valueAlloca, valueAllocaCast, valueAllocaSize := b.createTemporaryAlloca(valueType, "chan.value")
+	isZeroSize := b.targetData.TypeAllocSize(valueType) == 0
+	var valueAlloca, valueAllocaCast, valueAllocaSize llvm.Value
+	if isZeroSize {
+		valueAlloca = llvm.ConstNull(llvm.PointerType(valueType, 0))
+		valueAllocaCast = llvm.ConstNull(b.i8ptrType)
+	} else {
+		valueAlloca, valueAllocaCast, valueAllocaSize = b.createTemporaryAlloca(valueType, "chan.value")
+	}
 
 	// Allocate blockedlist buffer.
 	channelBlockedList := b.mod.GetTypeByName("runtime.channelBlockedList")
@@ -64,9 +80,14 @@ func (b *builder) createChanRecv(unop *ssa.UnOp) llvm.Value {
 
 	// Do the receive.
 	commaOk := b.createRuntimeCall("chanRecv", []llvm.Value{ch, valueAllocaCast, channelBlockedListAlloca}, "")
-	received := b.CreateLoad(valueAlloca, "chan.received")
+	var received llvm.Value
+	if isZeroSize {
+		received = llvm.ConstNull(valueType)
+	} else {
+		received = b.CreateLoad(valueAlloca, "chan.received")
+		b.emitLifetimeEnd(valueAllocaCast, valueAllocaSize)
+	}
 	b.emitLifetimeEnd(channelBlockedListAllocaCast, channelBlockedListAllocaSize)
-	b.emitLifetimeEnd(valueAllocaCast, valueAllocaSize)
 
 	if unop.CommaOk {
 		tuple := llvm.Undef(b.ctx.StructType([]llvm.Type{valueType, b.ctx.Int1Type()}, false))
@@ -116,7 +137,6 @@ func (b *builder) createSelect(expr *ssa.Select) llvm.Value {
 	// determine the receive buffer size and alignment.
 	recvbufSize := uint64(0)
 	recvbufAlign := 0
-	hasReceives := false
 	var selectStates []llvm.Value
 	chanSelectStateType := b.getLLVMRuntimeType("chanSelectState")
 	for _, state := range expr.States {
@@ -133,7 +153,6 @@ func (b *builder) createSelect(expr *ssa.Select) llvm.Value {
 			if align := b.targetData.ABITypeAlignment(llvmType); align > recvbufAlign {
 				recvbufAlign = align
 			}
-			hasReceives = true
 		case types.SendOnly:
 			// Store this value in an alloca and put a pointer to this alloca
 			// in the send state.
@@ -150,7 +169,7 @@ func (b *builder) createSelect(expr *ssa.Select) llvm.Value {
 
 	// Create a receive buffer, where the received value will be stored.
 	recvbuf := llvm.Undef(b.i8ptrType)
-	if hasReceives {
+	if recvbufSize != 0 {
 		allocaType := llvm.ArrayType(b.ctx.Int8Type(), int(recvbufSize))
 		recvbufAlloca, _, _ := b.createTemporaryAlloca(allocaType, "select.recvbuf.alloca")
 		recvbufAlloca.SetAlignment(recvbufAlign)
diff --git a/compiler/compiler.go b/compiler/compiler.go
index c0ad39d46..1ca798d19 100644
--- a/compiler/compiler.go
+++ b/compiler/compiler.go
@@ -23,7 +23,7 @@ import (
 // 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 = 18 // last change: fix duplicated named structs
+const Version = 19 // last change: fix channel ops with zero values
 
 func init() {
 	llvm.InitializeAllTargets()
diff --git a/compiler/compiler_test.go b/compiler/compiler_test.go
index 0fb9fa913..a4ac0f7b5 100644
--- a/compiler/compiler_test.go
+++ b/compiler/compiler_test.go
@@ -51,6 +51,7 @@ func TestCompiler(t *testing.T) {
 		{"pragma.go", ""},
 		{"goroutine.go", "wasm"},
 		{"goroutine.go", "cortex-m-qemu"},
+		{"channel.go", ""},
 		{"intrinsics.go", "cortex-m-qemu"},
 		{"intrinsics.go", "wasm"},
 	}
diff --git a/compiler/testdata/channel.go b/compiler/testdata/channel.go
new file mode 100644
index 000000000..ecc837a71
--- /dev/null
+++ b/compiler/testdata/channel.go
@@ -0,0 +1,25 @@
+package main
+
+func chanIntSend(ch chan int) {
+	ch <- 3
+}
+
+func chanIntRecv(ch chan int) {
+	<-ch
+}
+
+func chanZeroSend(ch chan struct{}) {
+	ch <- struct{}{}
+}
+
+func chanZeroRecv(ch chan struct{}) {
+	<-ch
+}
+
+func selectZeroRecv(ch1 chan int, ch2 chan struct{}) {
+	select {
+	case ch1 <- 1:
+	case <-ch2:
+	default:
+	}
+}
diff --git a/compiler/testdata/channel.ll b/compiler/testdata/channel.ll
new file mode 100644
index 000000000..0d33d330b
--- /dev/null
+++ b/compiler/testdata/channel.ll
@@ -0,0 +1,114 @@
+; ModuleID = 'channel.go'
+source_filename = "channel.go"
+target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
+target triple = "wasm32--wasi"
+
+%runtime.channel = type { i32, i32, i8, %runtime.channelBlockedList*, i32, i32, i32, i8* }
+%runtime.channelBlockedList = type { %runtime.channelBlockedList*, %"internal/task.Task"*, %runtime.chanSelectState*, { %runtime.channelBlockedList*, i32, i32 } }
+%"internal/task.Task" = type { %"internal/task.Task"*, i8*, i64, %"internal/task.state" }
+%"internal/task.state" = type { i8* }
+%runtime.chanSelectState = type { %runtime.channel*, i8* }
+
+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.chanIntSend(%runtime.channel* dereferenceable_or_null(32) %ch, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %chan.blockedList = alloca %runtime.channelBlockedList, align 8
+  %chan.value = alloca i32, align 4
+  %chan.value.bitcast = bitcast i32* %chan.value to i8*
+  call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %chan.value.bitcast)
+  store i32 3, i32* %chan.value, align 4
+  %chan.blockedList.bitcast = bitcast %runtime.channelBlockedList* %chan.blockedList to i8*
+  call void @llvm.lifetime.start.p0i8(i64 24, i8* nonnull %chan.blockedList.bitcast)
+  call void @runtime.chanSend(%runtime.channel* %ch, i8* nonnull %chan.value.bitcast, %runtime.channelBlockedList* nonnull %chan.blockedList, i8* undef, i8* null)
+  call void @llvm.lifetime.end.p0i8(i64 24, i8* nonnull %chan.blockedList.bitcast)
+  call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %chan.value.bitcast)
+  ret void
+}
+
+; Function Attrs: argmemonly nounwind willreturn
+declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #0
+
+declare void @runtime.chanSend(%runtime.channel* dereferenceable_or_null(32), i8*, %runtime.channelBlockedList* dereferenceable_or_null(24), i8*, i8*)
+
+; Function Attrs: argmemonly nounwind willreturn
+declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #0
+
+define hidden void @main.chanIntRecv(%runtime.channel* dereferenceable_or_null(32) %ch, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %chan.blockedList = alloca %runtime.channelBlockedList, align 8
+  %chan.value = alloca i32, align 4
+  %chan.value.bitcast = bitcast i32* %chan.value to i8*
+  call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull %chan.value.bitcast)
+  %chan.blockedList.bitcast = bitcast %runtime.channelBlockedList* %chan.blockedList to i8*
+  call void @llvm.lifetime.start.p0i8(i64 24, i8* nonnull %chan.blockedList.bitcast)
+  %0 = call i1 @runtime.chanRecv(%runtime.channel* %ch, i8* nonnull %chan.value.bitcast, %runtime.channelBlockedList* nonnull %chan.blockedList, i8* undef, i8* null)
+  call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull %chan.value.bitcast)
+  call void @llvm.lifetime.end.p0i8(i64 24, i8* nonnull %chan.blockedList.bitcast)
+  ret void
+}
+
+declare i1 @runtime.chanRecv(%runtime.channel* dereferenceable_or_null(32), i8*, %runtime.channelBlockedList* dereferenceable_or_null(24), i8*, i8*)
+
+define hidden void @main.chanZeroSend(%runtime.channel* dereferenceable_or_null(32) %ch, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %chan.blockedList = alloca %runtime.channelBlockedList, align 8
+  %chan.blockedList.bitcast = bitcast %runtime.channelBlockedList* %chan.blockedList to i8*
+  call void @llvm.lifetime.start.p0i8(i64 24, i8* nonnull %chan.blockedList.bitcast)
+  call void @runtime.chanSend(%runtime.channel* %ch, i8* null, %runtime.channelBlockedList* nonnull %chan.blockedList, i8* undef, i8* null)
+  call void @llvm.lifetime.end.p0i8(i64 24, i8* nonnull %chan.blockedList.bitcast)
+  ret void
+}
+
+define hidden void @main.chanZeroRecv(%runtime.channel* dereferenceable_or_null(32) %ch, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %chan.blockedList = alloca %runtime.channelBlockedList, align 8
+  %chan.blockedList.bitcast = bitcast %runtime.channelBlockedList* %chan.blockedList to i8*
+  call void @llvm.lifetime.start.p0i8(i64 24, i8* nonnull %chan.blockedList.bitcast)
+  %0 = call i1 @runtime.chanRecv(%runtime.channel* %ch, i8* null, %runtime.channelBlockedList* nonnull %chan.blockedList, i8* undef, i8* null)
+  call void @llvm.lifetime.end.p0i8(i64 24, i8* nonnull %chan.blockedList.bitcast)
+  ret void
+}
+
+define hidden void @main.selectZeroRecv(%runtime.channel* dereferenceable_or_null(32) %ch1, %runtime.channel* dereferenceable_or_null(32) %ch2, i8* %context, i8* %parentHandle) unnamed_addr {
+entry:
+  %select.states.alloca = alloca [2 x %runtime.chanSelectState], align 8
+  %select.send.value = alloca i32, align 4
+  store i32 1, i32* %select.send.value, align 4
+  %select.states.alloca.bitcast = bitcast [2 x %runtime.chanSelectState]* %select.states.alloca to i8*
+  call void @llvm.lifetime.start.p0i8(i64 16, i8* nonnull %select.states.alloca.bitcast)
+  %.repack = getelementptr inbounds [2 x %runtime.chanSelectState], [2 x %runtime.chanSelectState]* %select.states.alloca, i32 0, i32 0, i32 0
+  store %runtime.channel* %ch1, %runtime.channel** %.repack, align 8
+  %.repack1 = getelementptr inbounds [2 x %runtime.chanSelectState], [2 x %runtime.chanSelectState]* %select.states.alloca, i32 0, i32 0, i32 1
+  %0 = bitcast i8** %.repack1 to i32**
+  store i32* %select.send.value, i32** %0, align 4
+  %.repack3 = getelementptr inbounds [2 x %runtime.chanSelectState], [2 x %runtime.chanSelectState]* %select.states.alloca, i32 0, i32 1, i32 0
+  store %runtime.channel* %ch2, %runtime.channel** %.repack3, align 8
+  %.repack4 = getelementptr inbounds [2 x %runtime.chanSelectState], [2 x %runtime.chanSelectState]* %select.states.alloca, i32 0, i32 1, i32 1
+  store i8* null, i8** %.repack4, align 4
+  %select.states = getelementptr inbounds [2 x %runtime.chanSelectState], [2 x %runtime.chanSelectState]* %select.states.alloca, i32 0, i32 0
+  %select.result = call { i32, i1 } @runtime.tryChanSelect(i8* undef, %runtime.chanSelectState* nonnull %select.states, i32 2, i32 2, i8* undef, i8* null)
+  call void @llvm.lifetime.end.p0i8(i64 16, i8* nonnull %select.states.alloca.bitcast)
+  %1 = extractvalue { i32, i1 } %select.result, 0
+  %2 = icmp eq i32 %1, 0
+  br i1 %2, label %select.done, label %select.next
+
+select.done:                                      ; preds = %select.body, %select.next, %entry
+  ret void
+
+select.next:                                      ; preds = %entry
+  %3 = icmp eq i32 %1, 1
+  br i1 %3, label %select.body, label %select.done
+
+select.body:                                      ; preds = %select.next
+  br label %select.done
+}
+
+declare { i32, i1 } @runtime.tryChanSelect(i8*, %runtime.chanSelectState*, i32, i32, i8*, i8*)
+
+attributes #0 = { argmemonly nounwind willreturn }