runtime: rewrite channel implementation

This rewrite simplifies the channel implementation considerably, with
34% less LOC. Perhaps the most important change is the removal of the
channel state, which made sense when we had only send and receive
operations but only makes things more compliated when multiple select
operations can be pending on a single channel.

I did this rewrite originally to make it possible to make channels
parallelism-safe. The current implementation is not parallelism-safe,
but it will be easy to make it so (the main additions will be a channel
lock, a global select lock, and an atomic compare-and-swap in
chanQueue.pop).
This commit is contained in:
Ayke van Laethem
2024-10-27 08:44:19 +01:00
committed by Ron Evans
parent b7a3fd8d2f
commit 5ebda89d78
5 changed files with 408 additions and 587 deletions
+20 -17
View File
@@ -41,17 +41,17 @@ func (b *builder) createChanSend(instr *ssa.Send) {
b.CreateStore(chanValue, valueAlloca)
}
// Allocate blockedlist buffer.
channelBlockedList := b.getLLVMRuntimeType("channelBlockedList")
channelBlockedListAlloca, channelBlockedListAllocaSize := b.createTemporaryAlloca(channelBlockedList, "chan.blockedList")
// Allocate buffer for the channel operation.
channelOp := b.getLLVMRuntimeType("channelOp")
channelOpAlloca, channelOpAllocaSize := b.createTemporaryAlloca(channelOp, "chan.op")
// Do the send.
b.createRuntimeCall("chanSend", []llvm.Value{ch, valueAlloca, channelBlockedListAlloca}, "")
b.createRuntimeCall("chanSend", []llvm.Value{ch, valueAlloca, channelOpAlloca}, "")
// End the lifetime of the allocas.
// This also works around a bug in CoroSplit, at least in LLVM 8:
// https://bugs.llvm.org/show_bug.cgi?id=41742
b.emitLifetimeEnd(channelBlockedListAlloca, channelBlockedListAllocaSize)
b.emitLifetimeEnd(channelOpAlloca, channelOpAllocaSize)
if !isZeroSize {
b.emitLifetimeEnd(valueAlloca, valueAllocaSize)
}
@@ -72,12 +72,12 @@ func (b *builder) createChanRecv(unop *ssa.UnOp) llvm.Value {
valueAlloca, valueAllocaSize = b.createTemporaryAlloca(valueType, "chan.value")
}
// Allocate blockedlist buffer.
channelBlockedList := b.getLLVMRuntimeType("channelBlockedList")
channelBlockedListAlloca, channelBlockedListAllocaSize := b.createTemporaryAlloca(channelBlockedList, "chan.blockedList")
// Allocate buffer for the channel operation.
channelOp := b.getLLVMRuntimeType("channelOp")
channelOpAlloca, channelOpAllocaSize := b.createTemporaryAlloca(channelOp, "chan.op")
// Do the receive.
commaOk := b.createRuntimeCall("chanRecv", []llvm.Value{ch, valueAlloca, channelBlockedListAlloca}, "")
commaOk := b.createRuntimeCall("chanRecv", []llvm.Value{ch, valueAlloca, channelOpAlloca}, "")
var received llvm.Value
if isZeroSize {
received = llvm.ConstNull(valueType)
@@ -85,7 +85,7 @@ func (b *builder) createChanRecv(unop *ssa.UnOp) llvm.Value {
received = b.CreateLoad(valueType, valueAlloca, "chan.received")
b.emitLifetimeEnd(valueAlloca, valueAllocaSize)
}
b.emitLifetimeEnd(channelBlockedListAlloca, channelBlockedListAllocaSize)
b.emitLifetimeEnd(channelOpAlloca, channelOpAllocaSize)
if unop.CommaOk {
tuple := llvm.Undef(b.ctx.StructType([]llvm.Type{valueType, b.ctx.Int1Type()}, false))
@@ -198,10 +198,10 @@ func (b *builder) createSelect(expr *ssa.Select) llvm.Value {
if expr.Blocking {
// Stack-allocate operation structures.
// If these were simply created as a slice, they would heap-allocate.
chBlockAllocaType := llvm.ArrayType(b.getLLVMRuntimeType("channelBlockedList"), len(selectStates))
chBlockAlloca, chBlockSize := b.createTemporaryAlloca(chBlockAllocaType, "select.block.alloca")
chBlockLen := llvm.ConstInt(b.uintptrType, uint64(len(selectStates)), false)
chBlockPtr := b.CreateGEP(chBlockAllocaType, chBlockAlloca, []llvm.Value{
opsAllocaType := llvm.ArrayType(b.getLLVMRuntimeType("channelOp"), len(selectStates))
opsAlloca, opsSize := b.createTemporaryAlloca(opsAllocaType, "select.block.alloca")
opsLen := llvm.ConstInt(b.uintptrType, uint64(len(selectStates)), false)
opsPtr := b.CreateGEP(opsAllocaType, opsAlloca, []llvm.Value{
llvm.ConstInt(b.ctx.Int32Type(), 0, false),
llvm.ConstInt(b.ctx.Int32Type(), 0, false),
}, "select.block")
@@ -209,15 +209,18 @@ func (b *builder) createSelect(expr *ssa.Select) llvm.Value {
results = b.createRuntimeCall("chanSelect", []llvm.Value{
recvbuf,
statesPtr, statesLen, statesLen, // []chanSelectState
chBlockPtr, chBlockLen, chBlockLen, // []channelBlockList
opsPtr, opsLen, opsLen, // []channelOp
}, "select.result")
// Terminate the lifetime of the operation structures.
b.emitLifetimeEnd(chBlockAlloca, chBlockSize)
b.emitLifetimeEnd(opsAlloca, opsSize)
} else {
results = b.createRuntimeCall("tryChanSelect", []llvm.Value{
opsPtr := llvm.ConstNull(b.dataPtrType)
opsLen := llvm.ConstInt(b.uintptrType, 0, false)
results = b.createRuntimeCall("chanSelect", []llvm.Value{
recvbuf,
statesPtr, statesLen, statesLen, // []chanSelectState
opsPtr, opsLen, opsLen, // []channelOp (nil slice)
}, "select.result")
}
+26 -26
View File
@@ -3,7 +3,7 @@ source_filename = "channel.go"
target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-n32:64-S128-ni:1:10:20"
target triple = "wasm32-unknown-wasi"
%runtime.channelBlockedList = type { ptr, ptr, ptr, { ptr, i32, i32 } }
%runtime.channelOp = type { ptr, ptr, i32, ptr }
%runtime.chanSelectState = type { ptr, ptr }
; Function Attrs: allockind("alloc,zeroed") allocsize(0)
@@ -18,15 +18,15 @@ entry:
}
; Function Attrs: nounwind
define hidden void @main.chanIntSend(ptr dereferenceable_or_null(32) %ch, ptr %context) unnamed_addr #2 {
define hidden void @main.chanIntSend(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #2 {
entry:
%chan.blockedList = alloca %runtime.channelBlockedList, align 8
%chan.op = alloca %runtime.channelOp, align 8
%chan.value = alloca i32, align 4
call void @llvm.lifetime.start.p0(i64 4, ptr nonnull %chan.value)
store i32 3, ptr %chan.value, align 4
call void @llvm.lifetime.start.p0(i64 24, ptr nonnull %chan.blockedList)
call void @runtime.chanSend(ptr %ch, ptr nonnull %chan.value, ptr nonnull %chan.blockedList, ptr undef) #4
call void @llvm.lifetime.end.p0(i64 24, ptr nonnull %chan.blockedList)
call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %chan.op)
call void @runtime.chanSend(ptr %ch, ptr nonnull %chan.value, ptr nonnull %chan.op, ptr undef) #4
call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %chan.op)
call void @llvm.lifetime.end.p0(i64 4, ptr nonnull %chan.value)
ret void
}
@@ -34,48 +34,48 @@ entry:
; Function Attrs: nocallback nofree nosync nounwind willreturn memory(argmem: readwrite)
declare void @llvm.lifetime.start.p0(i64 immarg, ptr nocapture) #3
declare void @runtime.chanSend(ptr dereferenceable_or_null(32), ptr, ptr dereferenceable_or_null(24), ptr) #1
declare void @runtime.chanSend(ptr dereferenceable_or_null(36), ptr, ptr dereferenceable_or_null(16), ptr) #1
; Function Attrs: nocallback nofree nosync nounwind willreturn memory(argmem: readwrite)
declare void @llvm.lifetime.end.p0(i64 immarg, ptr nocapture) #3
; Function Attrs: nounwind
define hidden void @main.chanIntRecv(ptr dereferenceable_or_null(32) %ch, ptr %context) unnamed_addr #2 {
define hidden void @main.chanIntRecv(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #2 {
entry:
%chan.blockedList = alloca %runtime.channelBlockedList, align 8
%chan.op = alloca %runtime.channelOp, align 8
%chan.value = alloca i32, align 4
call void @llvm.lifetime.start.p0(i64 4, ptr nonnull %chan.value)
call void @llvm.lifetime.start.p0(i64 24, ptr nonnull %chan.blockedList)
%0 = call i1 @runtime.chanRecv(ptr %ch, ptr nonnull %chan.value, ptr nonnull %chan.blockedList, ptr undef) #4
call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %chan.op)
%0 = call i1 @runtime.chanRecv(ptr %ch, ptr nonnull %chan.value, ptr nonnull %chan.op, ptr undef) #4
call void @llvm.lifetime.end.p0(i64 4, ptr nonnull %chan.value)
call void @llvm.lifetime.end.p0(i64 24, ptr nonnull %chan.blockedList)
call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %chan.op)
ret void
}
declare i1 @runtime.chanRecv(ptr dereferenceable_or_null(32), ptr, ptr dereferenceable_or_null(24), ptr) #1
declare i1 @runtime.chanRecv(ptr dereferenceable_or_null(36), ptr, ptr dereferenceable_or_null(16), ptr) #1
; Function Attrs: nounwind
define hidden void @main.chanZeroSend(ptr dereferenceable_or_null(32) %ch, ptr %context) unnamed_addr #2 {
define hidden void @main.chanZeroSend(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #2 {
entry:
%chan.blockedList = alloca %runtime.channelBlockedList, align 8
call void @llvm.lifetime.start.p0(i64 24, ptr nonnull %chan.blockedList)
call void @runtime.chanSend(ptr %ch, ptr null, ptr nonnull %chan.blockedList, ptr undef) #4
call void @llvm.lifetime.end.p0(i64 24, ptr nonnull %chan.blockedList)
%chan.op = alloca %runtime.channelOp, align 8
call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %chan.op)
call void @runtime.chanSend(ptr %ch, ptr null, ptr nonnull %chan.op, ptr undef) #4
call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %chan.op)
ret void
}
; Function Attrs: nounwind
define hidden void @main.chanZeroRecv(ptr dereferenceable_or_null(32) %ch, ptr %context) unnamed_addr #2 {
define hidden void @main.chanZeroRecv(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #2 {
entry:
%chan.blockedList = alloca %runtime.channelBlockedList, align 8
call void @llvm.lifetime.start.p0(i64 24, ptr nonnull %chan.blockedList)
%0 = call i1 @runtime.chanRecv(ptr %ch, ptr null, ptr nonnull %chan.blockedList, ptr undef) #4
call void @llvm.lifetime.end.p0(i64 24, ptr nonnull %chan.blockedList)
%chan.op = alloca %runtime.channelOp, align 8
call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %chan.op)
%0 = call i1 @runtime.chanRecv(ptr %ch, ptr null, ptr nonnull %chan.op, ptr undef) #4
call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %chan.op)
ret void
}
; Function Attrs: nounwind
define hidden void @main.selectZeroRecv(ptr dereferenceable_or_null(32) %ch1, ptr dereferenceable_or_null(32) %ch2, ptr %context) unnamed_addr #2 {
define hidden void @main.selectZeroRecv(ptr dereferenceable_or_null(36) %ch1, ptr dereferenceable_or_null(36) %ch2, ptr %context) unnamed_addr #2 {
entry:
%select.states.alloca = alloca [2 x %runtime.chanSelectState], align 8
%select.send.value = alloca i32, align 4
@@ -88,7 +88,7 @@ entry:
store ptr %ch2, ptr %0, align 4
%.repack3 = getelementptr inbounds [2 x %runtime.chanSelectState], ptr %select.states.alloca, i32 0, i32 1, i32 1
store ptr null, ptr %.repack3, align 4
%select.result = call { i32, i1 } @runtime.tryChanSelect(ptr undef, ptr nonnull %select.states.alloca, i32 2, i32 2, ptr undef) #4
%select.result = call { i32, i1 } @runtime.chanSelect(ptr undef, ptr nonnull %select.states.alloca, i32 2, i32 2, ptr null, i32 0, i32 0, ptr undef) #4
call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %select.states.alloca)
%1 = extractvalue { i32, i1 } %select.result, 0
%2 = icmp eq i32 %1, 0
@@ -105,7 +105,7 @@ select.body: ; preds = %select.next
br label %select.done
}
declare { i32, i1 } @runtime.tryChanSelect(ptr, ptr, i32, i32, ptr) #1
declare { i32, i1 } @runtime.chanSelect(ptr, ptr, i32, i32, ptr, i32, i32, ptr) #1
attributes #0 = { allockind("alloc,zeroed") allocsize(0) "alloc-family"="runtime.alloc" "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext" }
attributes #1 = { "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext" }
+2 -2
View File
@@ -135,13 +135,13 @@ entry:
declare i32 @runtime.sliceCopy(ptr nocapture writeonly, ptr nocapture readonly, i32, i32, i32, ptr) #2
; Function Attrs: nounwind
define hidden void @main.closeBuiltinGoroutine(ptr dereferenceable_or_null(32) %ch, ptr %context) unnamed_addr #1 {
define hidden void @main.closeBuiltinGoroutine(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #1 {
entry:
call void @runtime.chanClose(ptr %ch, ptr undef) #9
ret void
}
declare void @runtime.chanClose(ptr dereferenceable_or_null(32), ptr) #2
declare void @runtime.chanClose(ptr dereferenceable_or_null(36), ptr) #2
; Function Attrs: nounwind
define hidden void @main.startInterfaceMethod(ptr %itf.typecode, ptr %itf.value, ptr %context) unnamed_addr #1 {
+2 -2
View File
@@ -144,13 +144,13 @@ entry:
declare i32 @runtime.sliceCopy(ptr nocapture writeonly, ptr nocapture readonly, i32, i32, i32, ptr) #1
; Function Attrs: nounwind
define hidden void @main.closeBuiltinGoroutine(ptr dereferenceable_or_null(32) %ch, ptr %context) unnamed_addr #2 {
define hidden void @main.closeBuiltinGoroutine(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #2 {
entry:
call void @runtime.chanClose(ptr %ch, ptr undef) #9
ret void
}
declare void @runtime.chanClose(ptr dereferenceable_or_null(32), ptr) #1
declare void @runtime.chanClose(ptr dereferenceable_or_null(36), ptr) #1
; Function Attrs: nounwind
define hidden void @main.startInterfaceMethod(ptr %itf.typecode, ptr %itf.value, ptr %context) unnamed_addr #2 {
+358 -540
View File
File diff suppressed because it is too large Load Diff