This is an archive of the discontinued LLVM Phabricator instance.

Differential D33576

[AMDGPU] Require waitcnt before barrier for all targets; adjust tests.
ClosedPublic

Authored by msearles on May 25 2017, 4:14 PM.

Download Raw Diff

Details

Reviewers

Group Reviewers

Restricted Project

Summary

[AMDGPU] Require waitcnt before barrier for all targets; adjust tests.

Diff Detail

Event Timeline

msearles created this revision.May 25 2017, 4:14 PM

Herald added subscribers: t-tye, tpr, dstuttard and 5 others. · View Herald TranscriptMay 25 2017, 4:14 PM

msearles added a reviewer: Restricted Project.May 25 2017, 4:16 PM

LGTM

rampitec accepted this revision.May 25 2017, 4:23 PM

This revision is now accepted and ready to land.May 25 2017, 4:23 PM

msearles added a subscriber: llvm-commits.May 25 2017, 4:45 PM

You can also combine gfx8/gfx9 s_waitcnt check into gcn check

Update test - combine GFX8/GFX9 s_waitcnt check into GCN check

Missing the source itself ;)

Doh! Add lib/Target/AMDGPU/AMDGPUSubtarget.h back into the diff.

Maybe this should be a full subtarget feature so it can be disabled

Removing content of this comment as it pertained to a different revision. Sorry for the noise.

Herald added subscribers: qcolombet, MatzeB. · View Herald TranscriptMay 26 2017, 3:15 PM

Ug. Fix revision; added wrong set of diffs; correcting.

r304217

Revision Contents

Path

Size

lib/

Target/

AMDGPU/

AMDGPUSubtarget.h

2 lines

test/

CodeGen/

AMDGPU/

llvm.amdgcn.s.barrier.ll

3 lines

Diff 100481

lib/Target/AMDGPU/AMDGPUSubtarget.h

Context not available.
	/// \returns True if waitcnt instruction is needed before barrier instruction,			/// \returns True if waitcnt instruction is needed before barrier instruction,
	/// false otherwise.			/// false otherwise.
	bool needWaitcntBeforeBarrier() const {			bool needWaitcntBeforeBarrier() const {
	return getGeneration() < GFX9;			return true;
	}			}

	/// \returns true if the flat_scratch register should be initialized with the			/// \returns true if the flat_scratch register should be initialized with the
Context not available.

test/CodeGen/AMDGPU/llvm.amdgcn.s.barrier.ll

Context not available.

	; GCN-LABEL: {{^}}test_barrier:			; GCN-LABEL: {{^}}test_barrier:
	; GFX8: buffer_store_dword			; GFX8: buffer_store_dword
	; GFX8: s_waitcnt
	; GFX9: flat_store_dword			; GFX9: flat_store_dword
	; GFX9-NOT: s_waitcnt			; GCN: s_waitcnt
	; GCN: s_barrier			; GCN: s_barrier
	define amdgpu_kernel void @test_barrier(i32 addrspace(1)* %out, i32 %size) #0 {			define amdgpu_kernel void @test_barrier(i32 addrspace(1)* %out, i32 %size) #0 {
	entry:			entry:
Context not available.