`GroupMemoryBarrierWithGroupSync` is required to be marked as convergent so that it can't generate duplicate calls or be moved to identical control flow. Without it, we generate undefined behaviour during optimization. For instance: https://godbolt.org/z/9j3vsq1h3. Testing that the convergent attribute is added is sufficient. There already exists testing, [here](https://github.com/inbelic/llvm-project/blob/main/llvm/test/Transforms/SimplifyCFG/attr-convergent.ll), that it will not be moved as described in the above link.
21 lines
1014 B
HLSL
21 lines
1014 B
HLSL
// RUN: %clang_cc1 -finclude-default-header -x hlsl -triple \
|
|
// RUN: dxil-pc-shadermodel6.3-library %s \
|
|
// RUN: -emit-llvm -disable-llvm-passes -o - | FileCheck %s \
|
|
// RUN: -DTARGET=dx -check-prefixes=CHECK,CHECK-DXIL
|
|
// RUN: %clang_cc1 -finclude-default-header -x hlsl -triple \
|
|
// RUN: spirv-unknown-vulkan-compute %s \
|
|
// RUN: -emit-llvm -disable-llvm-passes -o - | FileCheck %s \
|
|
// RUN: -DTARGET=spv -check-prefixes=CHECK,CHECK-SPIRV
|
|
|
|
// CHECK-DXIL: define hidden void @
|
|
// CHECK-SPIRV: define hidden spir_func void @
|
|
void test_GroupMemoryBarrierWithGroupSync() {
|
|
// CHECK-DXIL: call void @llvm.[[TARGET]].group.memory.barrier.with.group.sync()
|
|
// CHECK-SPIRV: call spir_func void @llvm.[[TARGET]].group.memory.barrier.with.group.sync()
|
|
GroupMemoryBarrierWithGroupSync();
|
|
}
|
|
|
|
// CHECK: declare void @llvm.[[TARGET]].group.memory.barrier.with.group.sync() #[[ATTRS:[0-9]+]]
|
|
// CHECK-NOT: attributes #[[ATTRS]] = {{.+}}memory(none){{.+}}
|
|
// CHECK: attributes #[[ATTRS]] = {{.+}}convergent{{.+}}
|