[PATCH] R600/SI: Insert s_waitcnt before s_barrier instructions.
Matt Arsenault
Matthew.Arsenault at amd.com
Tue Jan 6 08:14:50 PST 2015
On 12/29/2014 03:14 PM, Tom Stellard wrote:
> This ensures that all memory operations are complete when all threads
> reach the barrier.
> ---
> lib/Target/R600/SIInsertWaits.cpp | 6 +++++-
> test/CodeGen/R600/llvm.AMDGPU.barrier.global.ll | 4 +++-
> test/CodeGen/R600/llvm.AMDGPU.barrier.local.ll | 5 ++++-
> 3 files changed, 12 insertions(+), 3 deletions(-)
LGTM
>
> diff --git a/lib/Target/R600/SIInsertWaits.cpp b/lib/Target/R600/SIInsertWaits.cpp
> index 2e56508..181b116 100644
> --- a/lib/Target/R600/SIInsertWaits.cpp
> +++ b/lib/Target/R600/SIInsertWaits.cpp
> @@ -428,7 +428,11 @@ bool SIInsertWaits::runOnMachineFunction(MachineFunction &MF) {
> for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
> I != E; ++I) {
>
> - Changes |= insertWait(MBB, I, handleOperands(*I));
> + // Wait for everything before a barrier.
> + if (I->getOpcode() == AMDGPU::S_BARRIER)
> + Changes |= insertWait(MBB, I, LastIssued);
> + else
> + Changes |= insertWait(MBB, I, handleOperands(*I));
> pushInstruction(MBB, I);
> }
>
> diff --git a/test/CodeGen/R600/llvm.AMDGPU.barrier.global.ll b/test/CodeGen/R600/llvm.AMDGPU.barrier.global.ll
> index 98f6695..b20457e 100644
> --- a/test/CodeGen/R600/llvm.AMDGPU.barrier.global.ll
> +++ b/test/CodeGen/R600/llvm.AMDGPU.barrier.global.ll
> @@ -3,7 +3,9 @@
>
> ; FUNC-LABEL: {{^}}test_barrier_global:
> ; EG: GROUP_BARRIER
> -; SI: s_barrier
> +; SI: buffer_store_dword
> +; SI: s_waitcnt
> +; SI-NEXT: s_barrier
>
> define void @test_barrier_global(i32 addrspace(1)* %out) {
> entry:
> diff --git a/test/CodeGen/R600/llvm.AMDGPU.barrier.local.ll b/test/CodeGen/R600/llvm.AMDGPU.barrier.local.ll
> index 92fe9f2..ca3c8b8 100644
> --- a/test/CodeGen/R600/llvm.AMDGPU.barrier.local.ll
> +++ b/test/CodeGen/R600/llvm.AMDGPU.barrier.local.ll
> @@ -3,7 +3,10 @@
>
> ; FUNC-LABEL: {{^}}test_barrier_local:
> ; EG: GROUP_BARRIER
> -; SI: s_barrier
> +
> +; SI: buffer_store_dword
> +; SI: s_waitcnt
> +; SI-NEXT: s_barrier
>
> define void @test_barrier_local(i32 addrspace(1)* %out) {
> entry:
More information about the llvm-commits
mailing list