This is an archive of the discontinued LLVM Phabricator instance.

Revert earlier change that flagged ds permute operations as not affecting lgkm counter.
ClosedPublic

Authored by kanarayan on Apr 19 2017, 4:02 PM.

Reviewers

Commits

Summary

https://bugs.llvm.org/show_bug.cgi?id=32666 found a case with memory operands. Need to study this better and correct the new pass.

LGTM but a test would be good

I will add the test from the bug report as a lit test separately.

In D32254#731305, @kanarayan wrote:

I will add the test from the bug report as a lit test separately.

All that is probably required is that there is a waitcnt lgkmcnt in the existing test for the bpermute intrinsic

This revision was automatically updated to reflect the committed changes.

Path

Size

llvm/

trunk/

lib/

Target/

AMDGPU/

2 lines

	Show First 20 Lines • Show All 249 Lines • ▼ Show 20 Lines
	class DS_1A1D_PERMUTE <string opName, SDPatternOperator node = null_frag>			class DS_1A1D_PERMUTE <string opName, SDPatternOperator node = null_frag>
	: DS_Pseudo<opName,			: DS_Pseudo<opName,
	(outs VGPR_32:$vdst),			(outs VGPR_32:$vdst),
	(ins VGPR_32:$addr, VGPR_32:$data0, offset:$offset),			(ins VGPR_32:$addr, VGPR_32:$data0, offset:$offset),
	"$vdst, $addr, $data0$offset",			"$vdst, $addr, $data0$offset",
	[(set i32:$vdst,			[(set i32:$vdst,
	(node (DS1Addr1Offset i32:$addr, i16:$offset), i32:$data0))] > {			(node (DS1Addr1Offset i32:$addr, i16:$offset), i32:$data0))] > {

	let LGKM_CNT = 0;

	let mayLoad = 0;			let mayLoad = 0;
	let mayStore = 0;			let mayStore = 0;
	let isConvergent = 1;			let isConvergent = 1;

	let has_data1 = 0;			let has_data1 = 0;
	let has_gds = 0;			let has_gds = 0;
	}			}

	▲ Show 20 Lines • Show All 691 Lines • Show Last 20 Lines