This is an archive of the discontinued LLVM Phabricator instance.

Differential D117527

AMDGPU: add test to show wwm register overwrite issue
ClosedPublic

Authored by ruiling on Jan 17 2022, 5:03 PM.

Details

Reviewers
arsenm
Commits
rG2f4d44bcd4a1: AMDGPU: add test to show wwm register overwrite issue
Summary

Pre-commit the test to make the diff easy to read later.

Diff Detail

Event Timeline

ruiling created this revision.Jan 17 2022, 5:03 PM
Herald added subscribers: kerbowa, t-tye, tpr and 5 others. · View Herald TranscriptJan 17 2022, 5:03 PM
ruiling requested review of this revision.Jan 17 2022, 5:03 PM
Herald added a project: Restricted Project. · View Herald TranscriptJan 17 2022, 5:03 PM
Herald added subscribers: llvm-commits, wdng. · View Herald Transcript
arsenm accepted this revision.Jan 17 2022, 5:05 PM
arsenm added a subscriber: arsenm.
arsenm added inline comments.
llvm/test/CodeGen/AMDGPU/set-inactive-wwm-overwrite.ll
5

Don't need the dllexports

This revision is now accepted and ready to land.Jan 17 2022, 5:05 PM
ruiling mentioned this in D117482: AMDGPU: Don't clobber source register for V_SET_INACTIVE_*.Jan 17 2022, 5:05 PM
ruiling added a child revision: D117482: AMDGPU: Don't clobber source register for V_SET_INACTIVE_*.Jan 17 2022, 5:07 PM
Harbormaster completed remote builds in B143895: Diff 400678.Jan 17 2022, 5:40 PM
This revision was landed with ongoing or failed builds.Feb 5 2022, 8:38 PM
Closed by commit rG2f4d44bcd4a1: AMDGPU: add test to show wwm register overwrite issue (authored by ruiling). · Explain Why
This revision was automatically updated to reflect the committed changes.
ruiling added a commit: rG2f4d44bcd4a1: AMDGPU: add test to show wwm register overwrite issue.

Revision Contents

PathSize
llvm/
test/
CodeGen/
AMDGPU/
136 lines

Diff 406224

llvm/test/CodeGen/AMDGPU/set-inactive-wwm-overwrite.ll

  • This file was added.
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck %s -check-prefix=GCN
define amdgpu_cs void @if_then(<4 x i32> inreg %input, <4 x i32> inreg %output, <3 x i32> %LocalInvocationId) {
; GCN-LABEL: if_then:
arsenmUnsubmitted
Not Done
ReplyInline Actions

Don't need the dllexports

arsenm: Don't need the dllexports
; GCN: ; %bb.0: ; %.entry
; GCN-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
; GCN-NEXT: v_mov_b32_e32 v1, 0
; GCN-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GCN-NEXT: ; %bb.1: ; %.bb0
; GCN-NEXT: v_mov_b32_e32 v1, 1
; GCN-NEXT: ; %bb.2: ; %.merge
; GCN-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GCN-NEXT: v_cmp_lt_u32_e32 vcc_lo, 3, v0
; GCN-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GCN-NEXT: s_cbranch_execz .LBB0_4
; GCN-NEXT: ; %bb.3: ; %.then
; GCN-NEXT: s_not_b32 exec_lo, exec_lo
; GCN-NEXT: v_mov_b32_e32 v1, 0
; GCN-NEXT: s_not_b32 exec_lo, exec_lo
; GCN-NEXT: s_or_saveexec_b32 s1, -1
; GCN-NEXT: v_mov_b32_e32 v2, 0
; GCN-NEXT: v_mov_b32_dpp v2, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GCN-NEXT: s_mov_b32 exec_lo, s1
; GCN-NEXT: v_mov_b32_e32 v0, v2
; GCN-NEXT: v_mov_b32_e32 v3, -1
; GCN-NEXT: v_mov_b32_e32 v1, 0
; GCN-NEXT: buffer_store_dword v3, v0, s[4:7], 0 offen
; GCN-NEXT: .LBB0_4: ; %.end
; GCN-NEXT: s_waitcnt_depctr 0xffe3
; GCN-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GCN-NEXT: v_mov_b32_e32 v0, -1
; GCN-NEXT: buffer_store_dword v0, v1, s[4:7], 0 offen
; GCN-NEXT: s_endpgm
.entry:
%LocalInvocationId.i0 = extractelement <3 x i32> %LocalInvocationId, i32 0
%.not10002 = icmp eq i32 %LocalInvocationId.i0, 0
%i530 = icmp ult i32 %LocalInvocationId.i0, 4
br i1 %.not10002, label %.merge, label %.bb0
.bb0:
br label %.merge
.merge:
%src = phi i32 [ 0, %.entry ], [ 1, %.bb0 ]
br i1 %i530, label %.end, label %.then
.then:
%i562 = call i32 @llvm.amdgcn.set.inactive.i32(i32 %src, i32 0)
%i563 = call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %i562, i32 273, i32 15, i32 15, i1 false)
%i564 = call i32 @llvm.amdgcn.strict.wwm.i32(i32 %i563)
call void @llvm.amdgcn.raw.buffer.store.i32(i32 -1, <4 x i32> %output, i32 %i564, i32 0, i32 0)
br label %.end
.end:
%idx = phi i32 [ 0, %.then ], [ %src, %.merge ]
call void @llvm.amdgcn.raw.buffer.store.i32(i32 -1, <4 x i32> %output, i32 %idx, i32 0, i32 0)
ret void
}
define amdgpu_cs void @if_else_vgpr_opt(<4 x i32> inreg %input, <4 x i32> inreg %output, <3 x i32> %LocalInvocationId) {
; GCN-LABEL: if_else_vgpr_opt:
; GCN: ; %bb.0: ; %.entry
; GCN-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
; GCN-NEXT: v_mov_b32_e32 v2, 0
; GCN-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GCN-NEXT: ; %bb.1: ; %.bb0
; GCN-NEXT: v_mov_b32_e32 v2, 1
; GCN-NEXT: ; %bb.2: ; %.merge
; GCN-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GCN-NEXT: v_cmp_lt_u32_e32 vcc_lo, 3, v0
; GCN-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GCN-NEXT: s_xor_b32 s0, exec_lo, s0
; GCN-NEXT: s_cbranch_execz .LBB1_4
; GCN-NEXT: ; %bb.3: ; %.else
; GCN-NEXT: s_or_saveexec_b32 s1, -1
; GCN-NEXT: v_mov_b32_e32 v1, 0
; GCN-NEXT: s_mov_b32 exec_lo, s1
; GCN-NEXT: s_not_b32 exec_lo, exec_lo
; GCN-NEXT: v_mov_b32_e32 v2, 0
; GCN-NEXT: s_not_b32 exec_lo, exec_lo
; GCN-NEXT: s_or_saveexec_b32 s1, -1
; GCN-NEXT: v_mov_b32_dpp v1, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GCN-NEXT: s_mov_b32 exec_lo, s1
; GCN-NEXT: v_mov_b32_e32 v0, v1
; GCN-NEXT: v_mov_b32_e32 v3, -1
; GCN-NEXT: ; implicit-def: $vgpr2
; GCN-NEXT: buffer_store_dword v3, v0, s[4:7], 0 offen
; GCN-NEXT: .LBB1_4: ; %Flow
; GCN-NEXT: s_or_saveexec_b32 s0, s0
; GCN-NEXT: s_waitcnt_depctr 0xffe3
; GCN-NEXT: s_xor_b32 exec_lo, exec_lo, s0
; GCN-NEXT: s_cbranch_execz .LBB1_6
; GCN-NEXT: ; %bb.5: ; %.then
; GCN-NEXT: v_mov_b32_e32 v0, -1
; GCN-NEXT: buffer_store_dword v0, v2, s[4:7], 0 offen
; GCN-NEXT: .LBB1_6: ; %.end
; GCN-NEXT: s_endpgm
.entry:
%LocalInvocationId.i0 = extractelement <3 x i32> %LocalInvocationId, i32 0
%.not10002 = icmp eq i32 %LocalInvocationId.i0, 0
%i530 = icmp ult i32 %LocalInvocationId.i0, 4
br i1 %.not10002, label %.merge, label %.bb0
.bb0:
br label %.merge
.merge:
%src = phi i32 [ 0, %.entry ], [ 1, %.bb0 ]
br i1 %i530, label %.then, label %.else
.then:
call void @llvm.amdgcn.raw.buffer.store.i32(i32 -1, <4 x i32> %output, i32 %src, i32 0, i32 0)
br label %.end
.else:
%i562 = call i32 @llvm.amdgcn.set.inactive.i32(i32 %src, i32 0)
%i563 = call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %i562, i32 273, i32 15, i32 15, i1 false)
%i564 = call i32 @llvm.amdgcn.strict.wwm.i32(i32 %i563)
call void @llvm.amdgcn.raw.buffer.store.i32(i32 -1, <4 x i32> %output, i32 %i564, i32 0, i32 0)
br label %.end
.end:
ret void
}
declare i32 @llvm.amdgcn.set.inactive.i32(i32, i32) #0
declare i32 @llvm.amdgcn.update.dpp.i32(i32, i32, i32, i32, i32, i1) #0
declare i32 @llvm.amdgcn.strict.wwm.i32(i32) #1
declare void @llvm.amdgcn.raw.buffer.store.i32(i32, <4 x i32>, i32, i32 immarg, i32 immarg) #2
attributes #0 = { convergent nounwind readnone willreturn }
attributes #1 = { convergent nounwind readnone speculatable willreturn }
attributes #2 = { nounwind willreturn writeonly }