Skip to content

Commit dec74a8

Browse files
authored
[AMDGPU] Fix VS_CNT overflow assertion (#77935)
Always set the upper bound for VS_CNT higher than the lower bound. Before #77439 this code was only executed on function entry where the lower bound was 0 so it was not a problem. Fixes #77931
1 parent 108bede commit dec74a8

File tree

2 files changed

+61
-1
lines changed

2 files changed

+61
-1
lines changed

llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -293,7 +293,7 @@ class WaitcntBrackets {
293293
}
294294

295295
void setStateOnFunctionEntryOrReturn() {
296-
setScoreUB(VS_CNT, getWaitCountMax(VS_CNT));
296+
setScoreUB(VS_CNT, getScoreLB(VS_CNT) + getWaitCountMax(VS_CNT));
297297
PendingEvents |= WaitEventMaskForInst[VS_CNT];
298298
}
299299

Lines changed: 60 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,60 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 4
2+
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 -stop-after=si-insert-waitcnts -verify-machineinstrs < %s | FileCheck %s
3+
4+
declare fastcc void @bar()
5+
6+
define fastcc i32 @foo() {
7+
; CHECK-LABEL: name: foo
8+
; CHECK: bb.0 (%ir-block.0):
9+
; CHECK-NEXT: successors: %bb.1(0x80000000)
10+
; CHECK-NEXT: liveins: $sgpr12, $sgpr13, $sgpr14, $sgpr15, $sgpr30, $sgpr31, $vgpr31, $vgpr40, $sgpr4_sgpr5, $sgpr6_sgpr7, $sgpr8_sgpr9, $sgpr10_sgpr11
11+
; CHECK-NEXT: {{ $}}
12+
; CHECK-NEXT: S_WAITCNT 0
13+
; CHECK-NEXT: $sgpr16 = S_MOV_B32 $sgpr33
14+
; CHECK-NEXT: $sgpr33 = S_MOV_B32 $sgpr32
15+
; CHECK-NEXT: $sgpr17 = S_OR_SAVEEXEC_B32 -1, implicit-def $exec, implicit-def dead $scc, implicit $exec
16+
; CHECK-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr40, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr33, 0, 0, 0, implicit $exec :: (store (s32) into %stack.2, addrspace 5)
17+
; CHECK-NEXT: $exec_lo = S_MOV_B32 killed $sgpr17
18+
; CHECK-NEXT: $sgpr32 = frame-setup S_ADDK_I32 $sgpr32, 512, implicit-def dead $scc
19+
; CHECK-NEXT: $vgpr40 = V_WRITELANE_B32 killed $sgpr16, 2, undef $vgpr40
20+
; CHECK-NEXT: BUNDLE implicit-def $sgpr16_sgpr17, implicit-def $sgpr16, implicit-def $sgpr16_lo16, implicit-def $sgpr16_hi16, implicit-def $sgpr17, implicit-def $sgpr17_lo16, implicit-def $sgpr17_hi16, implicit-def $scc {
21+
; CHECK-NEXT: $sgpr16_sgpr17 = S_GETPC_B64
22+
; CHECK-NEXT: $sgpr16 = S_ADD_U32 internal $sgpr16, target-flags(amdgpu-gotprel32-lo) @bar + 4, implicit-def $scc
23+
; CHECK-NEXT: $sgpr17 = S_ADDC_U32 internal $sgpr17, target-flags(amdgpu-gotprel32-hi) @bar + 12, implicit-def $scc, implicit internal $scc
24+
; CHECK-NEXT: }
25+
; CHECK-NEXT: S_WAITCNT_VSCNT undef $sgpr_null, 0
26+
; CHECK-NEXT: BUFFER_GL0_INV implicit $exec
27+
; CHECK-NEXT: BUFFER_GL1_INV implicit $exec
28+
; CHECK-NEXT: renamable $sgpr16_sgpr17 = S_LOAD_DWORDX2_IMM killed renamable $sgpr16_sgpr17, 0, 0 :: (dereferenceable invariant load (s64) from got, addrspace 4)
29+
; CHECK-NEXT: $vgpr40 = V_WRITELANE_B32 killed $sgpr30, 0, $vgpr40
30+
; CHECK-NEXT: $vgpr40 = V_WRITELANE_B32 killed $sgpr31, 1, $vgpr40
31+
; CHECK-NEXT: S_WAITCNT 49279
32+
; CHECK-NEXT: dead $sgpr30_sgpr31 = SI_CALL killed renamable $sgpr16_sgpr17, @bar, csr_amdgpu, implicit killed $sgpr4_sgpr5, implicit killed $sgpr6_sgpr7, implicit killed $sgpr8_sgpr9, implicit killed $sgpr10_sgpr11, implicit killed $sgpr12, implicit killed $sgpr13, implicit killed $sgpr14, implicit killed $sgpr15, implicit killed $vgpr31, implicit $sgpr0_sgpr1_sgpr2_sgpr3
33+
; CHECK-NEXT: $vcc_lo = S_MOV_B32 $exec_lo
34+
; CHECK-NEXT: {{ $}}
35+
; CHECK-NEXT: bb.1 (%ir-block.1):
36+
; CHECK-NEXT: successors: %bb.2(0x04000000), %bb.1(0x7c000000)
37+
; CHECK-NEXT: liveins: $vcc_lo, $vgpr40
38+
; CHECK-NEXT: {{ $}}
39+
; CHECK-NEXT: S_CBRANCH_VCCNZ %bb.1, implicit $vcc_lo
40+
; CHECK-NEXT: {{ $}}
41+
; CHECK-NEXT: bb.2.DummyReturnBlock:
42+
; CHECK-NEXT: liveins: $vgpr40
43+
; CHECK-NEXT: {{ $}}
44+
; CHECK-NEXT: $sgpr31 = V_READLANE_B32 $vgpr40, 1
45+
; CHECK-NEXT: $sgpr30 = V_READLANE_B32 $vgpr40, 0
46+
; CHECK-NEXT: $sgpr4 = V_READLANE_B32 $vgpr40, 2
47+
; CHECK-NEXT: $sgpr5 = S_OR_SAVEEXEC_B32 -1, implicit-def $exec, implicit-def dead $scc, implicit $exec
48+
; CHECK-NEXT: $vgpr40 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr33, 0, 0, 0, implicit $exec :: (load (s32) from %stack.2, addrspace 5)
49+
; CHECK-NEXT: $exec_lo = S_MOV_B32 killed $sgpr5
50+
; CHECK-NEXT: $sgpr32 = frame-destroy S_ADDK_I32 $sgpr32, -512, implicit-def dead $scc
51+
; CHECK-NEXT: $sgpr33 = S_MOV_B32 killed $sgpr4
52+
; CHECK-NEXT: S_WAITCNT 16240
53+
; CHECK-NEXT: S_SETPC_B64_return undef $sgpr30_sgpr31, implicit undef $vgpr0
54+
fence acquire
55+
call fastcc void @bar()
56+
br label %1
57+
58+
1:
59+
br label %1
60+
}

0 commit comments

Comments
 (0)