Skip to content

Commit c80d495

Browse files
arsenmarsenm
authored
GlobalISel: Adjust insert point when expanding G_[SU]DIVREM
 (#160683) The insert point management is messy here. We probably should have an insert point guard, and not have ths dest operand utilities modify the insert point. Fixes #159716
1 parent c1f8dbb commit c80d495

File tree

2 files changed

+127
-0
lines changed

2 files changed

+127
-0
lines changed

‎llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2935,6 +2935,7 @@ LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) {
29352935
widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT);
29362936
widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_SEXT);
29372937
widenScalarDst(MI, WideTy);
2938+
MIRBuilder.setInsertPt(MIRBuilder.getMBB(), --MIRBuilder.getInsertPt());
29382939
widenScalarDst(MI, WideTy, 1);
29392940
Observer.changedInstr(MI);
29402941
return Legalized;
@@ -2972,6 +2973,7 @@ LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) {
29722973
widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT);
29732974
widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ZEXT);
29742975
widenScalarDst(MI, WideTy);
2976+
MIRBuilder.setInsertPt(MIRBuilder.getMBB(), --MIRBuilder.getInsertPt());
29752977
widenScalarDst(MI, WideTy, 1);
29762978
Observer.changedInstr(MI);
29772979
return Legalized;

‎llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-divrem.mir‎

Lines changed: 125 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,125 @@
1+
# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 6
2+
# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx803 -run-pass=legalizer -o - %s | FileCheck -check-prefix=GFX8 %s
3+
4+
---
5+
name: test_sdivrem_s16
6+
body: |
7+
bb.0:
8+
liveins: $vgpr0, $vgpr1
9+
10+
; GFX8-LABEL: name: test_sdivrem_s16
11+
; GFX8: liveins: $vgpr0, $vgpr1
12+
; GFX8-NEXT: {{ $}}
13+
; GFX8-NEXT: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
14+
; GFX8-NEXT: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
15+
; GFX8-NEXT: [[SEXT_INREG:%[0-9]+]]:_(s32) = G_SEXT_INREG [[COPY]], 16
16+
; GFX8-NEXT: [[SEXT_INREG1:%[0-9]+]]:_(s32) = G_SEXT_INREG [[COPY1]], 16
17+
; GFX8-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 31
18+
; GFX8-NEXT: [[ASHR:%[0-9]+]]:_(s32) = G_ASHR [[SEXT_INREG]], [[C]](s32)
19+
; GFX8-NEXT: [[ASHR1:%[0-9]+]]:_(s32) = G_ASHR [[SEXT_INREG1]], [[C]](s32)
20+
; GFX8-NEXT: [[ADD:%[0-9]+]]:_(s32) = G_ADD [[SEXT_INREG]], [[ASHR]]
21+
; GFX8-NEXT: [[ADD1:%[0-9]+]]:_(s32) = G_ADD [[SEXT_INREG1]], [[ASHR1]]
22+
; GFX8-NEXT: [[XOR:%[0-9]+]]:_(s32) = G_XOR [[ADD]], [[ASHR]]
23+
; GFX8-NEXT: [[XOR1:%[0-9]+]]:_(s32) = G_XOR [[ADD1]], [[ASHR1]]
24+
; GFX8-NEXT: [[UITOFP:%[0-9]+]]:_(s32) = G_UITOFP [[XOR1]](s32)
25+
; GFX8-NEXT: [[AMDGPU_RCP_IFLAG:%[0-9]+]]:_(s32) = G_AMDGPU_RCP_IFLAG [[UITOFP]](s32)
26+
; GFX8-NEXT: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float 0x41EFFFFFC0000000
27+
; GFX8-NEXT: [[FMUL:%[0-9]+]]:_(s32) = G_FMUL [[AMDGPU_RCP_IFLAG]], [[C1]]
28+
; GFX8-NEXT: [[FPTOUI:%[0-9]+]]:_(s32) = G_FPTOUI [[FMUL]](s32)
29+
; GFX8-NEXT: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
30+
; GFX8-NEXT: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[C2]], [[XOR1]]
31+
; GFX8-NEXT: [[MUL:%[0-9]+]]:_(s32) = G_MUL [[SUB]], [[FPTOUI]]
32+
; GFX8-NEXT: [[UMULH:%[0-9]+]]:_(s32) = G_UMULH [[FPTOUI]], [[MUL]]
33+
; GFX8-NEXT: [[ADD2:%[0-9]+]]:_(s32) = G_ADD [[FPTOUI]], [[UMULH]]
34+
; GFX8-NEXT: [[UMULH1:%[0-9]+]]:_(s32) = G_UMULH [[XOR]], [[ADD2]]
35+
; GFX8-NEXT: [[MUL1:%[0-9]+]]:_(s32) = G_MUL [[UMULH1]], [[XOR1]]
36+
; GFX8-NEXT: [[SUB1:%[0-9]+]]:_(s32) = G_SUB [[XOR]], [[MUL1]]
37+
; GFX8-NEXT: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
38+
; GFX8-NEXT: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(uge), [[SUB1]](s32), [[XOR1]]
39+
; GFX8-NEXT: [[ADD3:%[0-9]+]]:_(s32) = G_ADD [[UMULH1]], [[C3]]
40+
; GFX8-NEXT: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[ICMP]](s1), [[ADD3]], [[UMULH1]]
41+
; GFX8-NEXT: [[SUB2:%[0-9]+]]:_(s32) = G_SUB [[SUB1]], [[XOR1]]
42+
; GFX8-NEXT: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[ICMP]](s1), [[SUB2]], [[SUB1]]
43+
; GFX8-NEXT: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(uge), [[SELECT1]](s32), [[XOR1]]
44+
; GFX8-NEXT: [[ADD4:%[0-9]+]]:_(s32) = G_ADD [[SELECT]], [[C3]]
45+
; GFX8-NEXT: [[SELECT2:%[0-9]+]]:_(s32) = G_SELECT [[ICMP1]](s1), [[ADD4]], [[SELECT]]
46+
; GFX8-NEXT: [[SUB3:%[0-9]+]]:_(s32) = G_SUB [[SELECT1]], [[XOR1]]
47+
; GFX8-NEXT: [[SELECT3:%[0-9]+]]:_(s32) = G_SELECT [[ICMP1]](s1), [[SUB3]], [[SELECT1]]
48+
; GFX8-NEXT: [[XOR2:%[0-9]+]]:_(s32) = G_XOR [[ASHR]], [[ASHR1]]
49+
; GFX8-NEXT: [[XOR3:%[0-9]+]]:_(s32) = G_XOR [[SELECT2]], [[XOR2]]
50+
; GFX8-NEXT: [[SUB4:%[0-9]+]]:_(s32) = G_SUB [[XOR3]], [[XOR2]]
51+
; GFX8-NEXT: [[XOR4:%[0-9]+]]:_(s32) = G_XOR [[SELECT3]], [[ASHR]]
52+
; GFX8-NEXT: [[SUB5:%[0-9]+]]:_(s32) = G_SUB [[XOR4]], [[ASHR]]
53+
; GFX8-NEXT: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[SUB4]](s32)
54+
; GFX8-NEXT: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SUB5]](s32)
55+
; GFX8-NEXT: S_NOP 0, implicit [[TRUNC]](s16), implicit [[TRUNC1]](s16)
56+
; GFX8-NEXT: $vgpr0 = COPY [[SUB4]](s32)
57+
; GFX8-NEXT: $vgpr0 = COPY [[SUB5]](s32)
58+
%0:_(s32) = COPY $vgpr0
59+
%1:_(s32) = COPY $vgpr1
60+
%2:_(s16) = G_TRUNC %0
61+
%3:_(s16) = G_TRUNC %1
62+
%4:_(s16), %5:_(s16) = G_SDIVREM %2, %3
63+
S_NOP 0, implicit %4, implicit %5
64+
%6:_(s32) = G_ANYEXT %4
65+
%7:_(s32) = G_ANYEXT %5
66+
$vgpr0 = COPY %6
67+
$vgpr0 = COPY %7
68+
69+
...
70+
71+
---
72+
name: test_udivrem_s16
73+
body: |
74+
bb.0:
75+
liveins: $vgpr0, $vgpr1
76+
77+
; GFX8-LABEL: name: test_udivrem_s16
78+
; GFX8: liveins: $vgpr0, $vgpr1
79+
; GFX8-NEXT: {{ $}}
80+
; GFX8-NEXT: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
81+
; GFX8-NEXT: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
82+
; GFX8-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 65535
83+
; GFX8-NEXT: [[AND:%[0-9]+]]:_(s32) = G_AND [[COPY]], [[C]]
84+
; GFX8-NEXT: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY1]], [[C]]
85+
; GFX8-NEXT: [[UITOFP:%[0-9]+]]:_(s32) = G_UITOFP [[AND1]](s32)
86+
; GFX8-NEXT: [[AMDGPU_RCP_IFLAG:%[0-9]+]]:_(s32) = G_AMDGPU_RCP_IFLAG [[UITOFP]](s32)
87+
; GFX8-NEXT: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float 0x41EFFFFFC0000000
88+
; GFX8-NEXT: [[FMUL:%[0-9]+]]:_(s32) = G_FMUL [[AMDGPU_RCP_IFLAG]], [[C1]]
89+
; GFX8-NEXT: [[FPTOUI:%[0-9]+]]:_(s32) = G_FPTOUI [[FMUL]](s32)
90+
; GFX8-NEXT: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
91+
; GFX8-NEXT: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[C2]], [[AND1]]
92+
; GFX8-NEXT: [[MUL:%[0-9]+]]:_(s32) = G_MUL [[SUB]], [[FPTOUI]]
93+
; GFX8-NEXT: [[UMULH:%[0-9]+]]:_(s32) = G_UMULH [[FPTOUI]], [[MUL]]
94+
; GFX8-NEXT: [[ADD:%[0-9]+]]:_(s32) = G_ADD [[FPTOUI]], [[UMULH]]
95+
; GFX8-NEXT: [[UMULH1:%[0-9]+]]:_(s32) = G_UMULH [[AND]], [[ADD]]
96+
; GFX8-NEXT: [[MUL1:%[0-9]+]]:_(s32) = G_MUL [[UMULH1]], [[AND1]]
97+
; GFX8-NEXT: [[SUB1:%[0-9]+]]:_(s32) = G_SUB [[AND]], [[MUL1]]
98+
; GFX8-NEXT: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 1
99+
; GFX8-NEXT: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(uge), [[SUB1]](s32), [[AND1]]
100+
; GFX8-NEXT: [[ADD1:%[0-9]+]]:_(s32) = G_ADD [[UMULH1]], [[C3]]
101+
; GFX8-NEXT: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[ICMP]](s1), [[ADD1]], [[UMULH1]]
102+
; GFX8-NEXT: [[SUB2:%[0-9]+]]:_(s32) = G_SUB [[SUB1]], [[AND1]]
103+
; GFX8-NEXT: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[ICMP]](s1), [[SUB2]], [[SUB1]]
104+
; GFX8-NEXT: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(uge), [[SELECT1]](s32), [[AND1]]
105+
; GFX8-NEXT: [[ADD2:%[0-9]+]]:_(s32) = G_ADD [[SELECT]], [[C3]]
106+
; GFX8-NEXT: [[SELECT2:%[0-9]+]]:_(s32) = G_SELECT [[ICMP1]](s1), [[ADD2]], [[SELECT]]
107+
; GFX8-NEXT: [[SUB3:%[0-9]+]]:_(s32) = G_SUB [[SELECT1]], [[AND1]]
108+
; GFX8-NEXT: [[SELECT3:%[0-9]+]]:_(s32) = G_SELECT [[ICMP1]](s1), [[SUB3]], [[SELECT1]]
109+
; GFX8-NEXT: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[SELECT2]](s32)
110+
; GFX8-NEXT: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SELECT3]](s32)
111+
; GFX8-NEXT: S_NOP 0, implicit [[TRUNC]](s16), implicit [[TRUNC1]](s16)
112+
; GFX8-NEXT: $vgpr0 = COPY [[SELECT2]](s32)
113+
; GFX8-NEXT: $vgpr0 = COPY [[SELECT3]](s32)
114+
%0:_(s32) = COPY $vgpr0
115+
%1:_(s32) = COPY $vgpr1
116+
%2:_(s16) = G_TRUNC %0
117+
%3:_(s16) = G_TRUNC %1
118+
%4:_(s16), %5:_(s16) = G_UDIVREM %2, %3
119+
S_NOP 0, implicit %4, implicit %5
120+
%6:_(s32) = G_ANYEXT %4
121+
%7:_(s32) = G_ANYEXT %5
122+
$vgpr0 = COPY %6
123+
$vgpr0 = COPY %7
124+
125+
...

0 commit comments

Comments
 (0)