Skip to content

Commit 30007a5

Browse files
arsenmarsenm
authored
AMDGPU: Fix crash in rewrite AGPR copy MFMA pass on dead valnos (#153915)
Test isn't great, probably won't reliably reproduce the problem in the future.
1 parent fc6f235 commit 30007a5

File tree

2 files changed

+214
-0
lines changed

2 files changed

+214
-0
lines changed

‎llvm/lib/Target/AMDGPU/AMDGPURewriteAGPRCopyMFMA.cpp

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -147,6 +147,9 @@ bool AMDGPURewriteAGPRCopyMFMAImpl::run(MachineFunction &MF) const {
147147

148148
// TODO: Test multiple uses
149149
for (VNInfo *VNI : LI.vnis()) {
150+
if (VNI->isPHIDef() || VNI->isUnused())
151+
continue;
152+
150153
MachineInstr *DefMI = LIS.getInstructionFromIndex(VNI->def);
151154

152155
// TODO: Handle SplitKit produced copy bundles for partially defined

‎llvm/test/CodeGen/AMDGPU/av-split-dead-valno-crash.ll

Lines changed: 211 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,211 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
2+
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -stress-regalloc=18 < %s | FileCheck %s
3+
4+
define amdgpu_kernel void @vgpr_mfma_pass_av_split_crash(double %arg1, i1 %arg2, i1 %cond.i.i.i2295, ptr addrspace(1) %ptr, ptr %ptr1) #0 {
5+
; CHECK-LABEL: vgpr_mfma_pass_av_split_crash:
6+
; CHECK: ; %bb.0: ; %entry
7+
; CHECK-NEXT: s_load_dword s0, s[4:5], 0x8
8+
; CHECK-NEXT: s_load_dwordx2 s[10:11], s[4:5], 0x0
9+
; CHECK-NEXT: s_load_dwordx4 s[12:15], s[4:5], 0x10
10+
; CHECK-NEXT: v_mov_b32_e32 v1, 0x3e21eeb6
11+
; CHECK-NEXT: v_mov_b32_e32 v2, 0xa17f65f6
12+
; CHECK-NEXT: s_waitcnt lgkmcnt(0)
13+
; CHECK-NEXT: s_bitcmp1_b32 s0, 0
14+
; CHECK-NEXT: s_cselect_b64 s[16:17], -1, 0
15+
; CHECK-NEXT: s_xor_b64 s[18:19], s[16:17], -1
16+
; CHECK-NEXT: s_bitcmp1_b32 s0, 8
17+
; CHECK-NEXT: s_cselect_b64 s[2:3], -1, 0
18+
; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[2:3]
19+
; CHECK-NEXT: s_xor_b64 s[20:21], s[2:3], -1
20+
; CHECK-NEXT: v_cmp_ne_u32_e64 s[0:1], 1, v0
21+
; CHECK-NEXT: s_and_b64 s[2:3], exec, s[2:3]
22+
; CHECK-NEXT: v_mov_b32_e32 v0, 0x9037ab78
23+
; CHECK-NEXT: v_mov_b32_e32 v3, 0xbe927e4f
24+
; CHECK-NEXT: v_mov_b32_e32 v4, 0x19f4ec90
25+
; CHECK-NEXT: v_mov_b32_e32 v5, 0x3efa01a0
26+
; CHECK-NEXT: v_mov_b32_e32 v6, 0x16c16967
27+
; CHECK-NEXT: v_mov_b32_e32 v7, 0xbf56c16c
28+
; CHECK-NEXT: v_mov_b32_e32 v8, 0x69efb384
29+
; CHECK-NEXT: v_mov_b32_e32 v9, 0x3f4b2bb0
30+
; CHECK-NEXT: v_mov_b32_e32 v10, 0xa57d9582
31+
; CHECK-NEXT: v_mov_b32_e32 v11, 0xbf8c6ea4
32+
; CHECK-NEXT: v_mov_b32_e32 v12, 0xe82d3ff0
33+
; CHECK-NEXT: v_mov_b32_e32 v13, 0xbfa59976
34+
; CHECK-NEXT: v_mov_b32_e32 v14, 0x8427b883
35+
; CHECK-NEXT: v_mov_b32_e32 v15, 0x3fae1bb4
36+
; CHECK-NEXT: s_mov_b64 s[22:23], 0
37+
; CHECK-NEXT: v_mov_b32_e32 v16, 0x57b87036
38+
; CHECK-NEXT: v_mov_b32_e32 v17, 0x3fb3b136
39+
; CHECK-NEXT: s_and_b64 s[4:5], exec, s[16:17]
40+
; CHECK-NEXT: v_mov_b32_e32 v18, 0x55555523
41+
; CHECK-NEXT: v_mov_b32_e32 v19, 0xbfd55555
42+
; CHECK-NEXT: s_and_b64 s[6:7], exec, s[18:19]
43+
; CHECK-NEXT: v_mov_b32_e32 v20, 0
44+
; CHECK-NEXT: ; implicit-def: $agpr0_agpr1
45+
; CHECK-NEXT: ; implicit-def: $vgpr22_vgpr23
46+
; CHECK-NEXT: s_branch .LBB0_2
47+
; CHECK-NEXT: .LBB0_1: ; %Flow9
48+
; CHECK-NEXT: ; in Loop: Header=BB0_2 Depth=1
49+
; CHECK-NEXT: s_andn2_b64 vcc, exec, s[24:25]
50+
; CHECK-NEXT: s_cbranch_vccz .LBB0_17
51+
; CHECK-NEXT: .LBB0_2: ; %._crit_edge1942.i.i.i3548
52+
; CHECK-NEXT: ; =>This Loop Header: Depth=1
53+
; CHECK-NEXT: ; Child Loop BB0_6 Depth 2
54+
; CHECK-NEXT: s_and_b64 vcc, exec, s[0:1]
55+
; CHECK-NEXT: s_cbranch_vccnz .LBB0_9
56+
; CHECK-NEXT: ; %bb.3: ; %.preheader1868.i.i.i3244
57+
; CHECK-NEXT: ; in Loop: Header=BB0_2 Depth=1
58+
; CHECK-NEXT: s_mov_b64 vcc, s[4:5]
59+
; CHECK-NEXT: s_cbranch_vccz .LBB0_10
60+
; CHECK-NEXT: ; %bb.4: ; %.preheader1855.i.i.i3329.preheader
61+
; CHECK-NEXT: ; in Loop: Header=BB0_2 Depth=1
62+
; CHECK-NEXT: v_mov_b64_e32 v[24:25], s[14:15]
63+
; CHECK-NEXT: flat_load_dwordx2 v[24:25], v[24:25]
64+
; CHECK-NEXT: v_mov_b64_e32 v[26:27], v[0:1]
65+
; CHECK-NEXT: v_mov_b64_e32 v[28:29], v[2:3]
66+
; CHECK-NEXT: v_mov_b64_e32 v[30:31], v[16:17]
67+
; CHECK-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
68+
; CHECK-NEXT: v_fmac_f64_e32 v[26:27], 0, v[24:25]
69+
; CHECK-NEXT: v_fmac_f64_e32 v[28:29], 0, v[26:27]
70+
; CHECK-NEXT: v_mov_b64_e32 v[26:27], v[4:5]
71+
; CHECK-NEXT: v_fmac_f64_e32 v[26:27], 0, v[28:29]
72+
; CHECK-NEXT: v_mov_b64_e32 v[28:29], v[6:7]
73+
; CHECK-NEXT: v_fmac_f64_e32 v[28:29], 0, v[26:27]
74+
; CHECK-NEXT: v_mov_b64_e32 v[26:27], v[8:9]
75+
; CHECK-NEXT: v_fmac_f64_e32 v[26:27], 0, v[28:29]
76+
; CHECK-NEXT: v_mov_b64_e32 v[28:29], v[10:11]
77+
; CHECK-NEXT: v_fmac_f64_e32 v[28:29], 0, v[26:27]
78+
; CHECK-NEXT: v_mov_b64_e32 v[26:27], v[12:13]
79+
; CHECK-NEXT: v_fmac_f64_e32 v[26:27], 0, v[28:29]
80+
; CHECK-NEXT: v_mov_b64_e32 v[28:29], v[14:15]
81+
; CHECK-NEXT: v_fmac_f64_e32 v[28:29], 0, v[26:27]
82+
; CHECK-NEXT: v_fmac_f64_e32 v[30:31], 0, v[28:29]
83+
; CHECK-NEXT: v_mov_b64_e32 v[26:27], v[18:19]
84+
; CHECK-NEXT: v_fmac_f64_e32 v[26:27], 0, v[30:31]
85+
; CHECK-NEXT: v_mov_b64_e32 v[30:31], 0
86+
; CHECK-NEXT: s_branch .LBB0_6
87+
; CHECK-NEXT: .LBB0_5: ; %Flow
88+
; CHECK-NEXT: ; in Loop: Header=BB0_6 Depth=2
89+
; CHECK-NEXT: s_and_b64 vcc, exec, s[8:9]
90+
; CHECK-NEXT: s_cbranch_vccnz .LBB0_11
91+
; CHECK-NEXT: .LBB0_6: ; %.preheader1855.i.i.i3329
92+
; CHECK-NEXT: ; Parent Loop BB0_2 Depth=1
93+
; CHECK-NEXT: ; => This Inner Loop Header: Depth=2
94+
; CHECK-NEXT: v_mov_b64_e32 v[28:29], v[30:31]
95+
; CHECK-NEXT: s_mov_b64 s[24:25], -1
96+
; CHECK-NEXT: s_mov_b64 s[8:9], -1
97+
; CHECK-NEXT: s_mov_b64 vcc, s[2:3]
98+
; CHECK-NEXT: ; implicit-def: $vgpr30_vgpr31
99+
; CHECK-NEXT: s_cbranch_vccz .LBB0_5
100+
; CHECK-NEXT: ; %bb.7: ; %.lr.ph2070.i.i.i3291
101+
; CHECK-NEXT: ; in Loop: Header=BB0_6 Depth=2
102+
; CHECK-NEXT: v_accvgpr_read_b32 v31, a1
103+
; CHECK-NEXT: v_accvgpr_read_b32 v30, a0
104+
; CHECK-NEXT: s_mov_b64 s[8:9], s[18:19]
105+
; CHECK-NEXT: s_mov_b64 vcc, s[6:7]
106+
; CHECK-NEXT: s_cbranch_vccz .LBB0_5
107+
; CHECK-NEXT: ; %bb.8: ; %.preheader1856.preheader.i.i.i3325
108+
; CHECK-NEXT: ; in Loop: Header=BB0_6 Depth=2
109+
; CHECK-NEXT: s_mov_b64 s[24:25], 0
110+
; CHECK-NEXT: v_mov_b64_e32 v[30:31], v[26:27]
111+
; CHECK-NEXT: s_mov_b64 s[8:9], 0
112+
; CHECK-NEXT: s_branch .LBB0_5
113+
; CHECK-NEXT: .LBB0_9: ; in Loop: Header=BB0_2 Depth=1
114+
; CHECK-NEXT: v_mov_b64_e32 v[24:25], s[10:11]
115+
; CHECK-NEXT: v_accvgpr_write_b32 a0, v24
116+
; CHECK-NEXT: s_mov_b64 s[22:23], 0
117+
; CHECK-NEXT: v_accvgpr_write_b32 a1, v25
118+
; CHECK-NEXT: s_mov_b64 s[8:9], s[20:21]
119+
; CHECK-NEXT: s_branch .LBB0_15
120+
; CHECK-NEXT: .LBB0_10: ; in Loop: Header=BB0_2 Depth=1
121+
; CHECK-NEXT: s_mov_b64 s[8:9], -1
122+
; CHECK-NEXT: v_mov_b64_e32 v[22:23], 0
123+
; CHECK-NEXT: s_branch .LBB0_15
124+
; CHECK-NEXT: .LBB0_11: ; %loop.exit.guard
125+
; CHECK-NEXT: ; in Loop: Header=BB0_2 Depth=1
126+
; CHECK-NEXT: s_and_b64 vcc, exec, s[24:25]
127+
; CHECK-NEXT: s_cbranch_vccz .LBB0_13
128+
; CHECK-NEXT: ; %bb.12: ; %._crit_edge2105.i.i.i2330.loopexit
129+
; CHECK-NEXT: ; in Loop: Header=BB0_2 Depth=1
130+
; CHECK-NEXT: v_cmp_nlg_f64_e64 s[8:9], 0, v[28:29]
131+
; CHECK-NEXT: v_accvgpr_write_b32 a0, v24
132+
; CHECK-NEXT: v_cndmask_b32_e64 v23, v23, 0, s[16:17]
133+
; CHECK-NEXT: v_cndmask_b32_e64 v26, 0, 1, s[8:9]
134+
; CHECK-NEXT: v_mov_b32_e32 v27, v26
135+
; CHECK-NEXT: s_and_b64 s[8:9], exec, s[16:17]
136+
; CHECK-NEXT: v_cndmask_b32_e64 v22, v22, 0, s[16:17]
137+
; CHECK-NEXT: global_store_dwordx2 v20, v[26:27], s[12:13]
138+
; CHECK-NEXT: s_cselect_b32 s23, s23, 0
139+
; CHECK-NEXT: s_cselect_b32 s22, s22, 0
140+
; CHECK-NEXT: s_mov_b64 s[8:9], -1
141+
; CHECK-NEXT: s_branch .LBB0_14
142+
; CHECK-NEXT: .LBB0_13: ; in Loop: Header=BB0_2 Depth=1
143+
; CHECK-NEXT: v_accvgpr_write_b32 a0, v24
144+
; CHECK-NEXT: s_mov_b64 s[8:9], 0
145+
; CHECK-NEXT: v_mov_b64_e32 v[22:23], 0
146+
; CHECK-NEXT: .LBB0_14: ; %Flow6
147+
; CHECK-NEXT: ; in Loop: Header=BB0_2 Depth=1
148+
; CHECK-NEXT: v_accvgpr_write_b32 a1, v25
149+
; CHECK-NEXT: .LBB0_15: ; %Flow6
150+
; CHECK-NEXT: ; in Loop: Header=BB0_2 Depth=1
151+
; CHECK-NEXT: s_mov_b64 s[24:25], -1
152+
; CHECK-NEXT: s_and_b64 vcc, exec, s[8:9]
153+
; CHECK-NEXT: s_cbranch_vccz .LBB0_1
154+
; CHECK-NEXT: ; %bb.16: ; %._crit_edge2105.i.i.i2330
155+
; CHECK-NEXT: ; in Loop: Header=BB0_2 Depth=1
156+
; CHECK-NEXT: v_mov_b32_e32 v21, v20
157+
; CHECK-NEXT: s_mov_b64 s[24:25], 0
158+
; CHECK-NEXT: global_store_dwordx2 v20, v[20:21], s[12:13]
159+
; CHECK-NEXT: s_branch .LBB0_1
160+
; CHECK-NEXT: .LBB0_17: ; %DummyReturnBlock
161+
; CHECK-NEXT: s_endpgm
162+
entry:
163+
br label %._crit_edge1942.i.i.i3548
164+
165+
._crit_edge1942.i.i.i3548: ; preds = %._crit_edge2105.i.i.i2330, %entry
166+
%.sroa.02591.0.i.i.i226323 = phi double [ poison, %entry ], [ %.sroa.02591.3.i.i.i2301, %._crit_edge2105.i.i.i2330 ]
167+
%.sroa.3.0.i.i.i2270 = phi double [ poison, %entry ], [ %.sroa.3.3.i.i.i2308, %._crit_edge2105.i.i.i2330 ]
168+
%.014942244.i.i.i2280 = phi double [ 0.000000e+00, %entry ], [ %.31497.i.i.i2317, %._crit_edge2105.i.i.i2330 ]
169+
br i1 %cond.i.i.i2295, label %.preheader1868.i.i.i3244, label %._crit_edge2105.i.i.i2330
170+
171+
.preheader1868.i.i.i3244: ; preds = %._crit_edge1942.i.i.i3548
172+
%i = load double, ptr %ptr1, align 8
173+
%i3 = call double @llvm.fma.f64(double %i, double 0.000000e+00, double 0x3E21EEB69037AB78)
174+
%i4 = call double @llvm.fma.f64(double 0.000000e+00, double %i3, double 0xBE927E4FA17F65F6)
175+
%i5 = call double @llvm.fma.f64(double 0.000000e+00, double %i4, double 0x3EFA01A019F4EC90)
176+
%i6 = call double @llvm.fma.f64(double 0.000000e+00, double %i5, double 0xBF56C16C16C16967)
177+
%spec.select.i.i.i3288 = select i1 %arg2, double 0.000000e+00, double %.sroa.3.0.i.i.i2270
178+
br i1 %arg2, label %.preheader1855.i.i.i3329, label %._crit_edge2105.i.i.i2330
179+
180+
.lr.ph2070.i.i.i3291: ; preds = %.preheader1855.i.i.i3329
181+
br i1 %arg2, label %.preheader1855.i.i.i3329, label %.preheader1856.preheader.i.i.i3325
182+
183+
.preheader1856.preheader.i.i.i3325: ; preds = %.lr.ph2070.i.i.i3291
184+
%i11 = call double @llvm.fma.f64(double 0.000000e+00, double %i6, double 0x3F4B2BB069EFB384)
185+
%i14 = call double @llvm.fma.f64(double 0.000000e+00, double %i11, double 0xBF8C6EA4A57D9582)
186+
%i18 = call double @llvm.fma.f64(double 0.000000e+00, double %i14, double 0xBFA59976E82D3FF0)
187+
%i21 = call double @llvm.fma.f64(double 0.000000e+00, double %i18, double 0x3FAE1BB48427B883)
188+
%i23 = call double @llvm.fma.f64(double 0.000000e+00, double %i21, double 0x3FB3B13657B87036)
189+
%i28 = call double @llvm.fma.f64(double 0.000000e+00, double %i23, double 0xBFD5555555555523)
190+
br label %.preheader1855.i.i.i3329
191+
192+
.preheader1855.i.i.i3329: ; preds = %.preheader1856.preheader.i.i.i3325, %.lr.ph2070.i.i.i3291, %.preheader1868.i.i.i3244
193+
%.sroa.02591.4.i.i.i3335 = phi double [ %i28, %.preheader1856.preheader.i.i.i3325 ], [ %.sroa.02591.0.i.i.i226323, %.lr.ph2070.i.i.i3291 ], [ 0.000000e+00, %.preheader1868.i.i.i3244 ]
194+
%.21496.ph.i.i.i3348 = select i1 %arg2, double %.014942244.i.i.i2280, double 0.000000e+00
195+
%i31 = fcmp one double %.sroa.02591.4.i.i.i3335, 0.000000e+00
196+
%i32 = select i1 %i31, <2 x i32> zeroinitializer, <2 x i32> splat (i32 1)
197+
store <2 x i32> %i32, ptr addrspace(1) %ptr, align 8
198+
br i1 %cond.i.i.i2295, label %.lr.ph2070.i.i.i3291, label %._crit_edge2105.i.i.i2330
199+
200+
._crit_edge2105.i.i.i2330: ; preds = %.preheader1855.i.i.i3329, %.preheader1868.i.i.i3244, %._crit_edge1942.i.i.i3548
201+
%.sroa.02591.3.i.i.i2301 = phi double [ %.sroa.02591.0.i.i.i226323, %.preheader1868.i.i.i3244 ], [ %arg1, %._crit_edge1942.i.i.i3548 ], [ %i, %.preheader1855.i.i.i3329 ]
202+
%.sroa.3.3.i.i.i2308 = phi double [ 0.000000e+00, %.preheader1868.i.i.i3244 ], [ %.sroa.3.0.i.i.i2270, %._crit_edge1942.i.i.i3548 ], [ %spec.select.i.i.i3288, %.preheader1855.i.i.i3329 ]
203+
%.31497.i.i.i2317 = phi double [ %.014942244.i.i.i2280, %.preheader1868.i.i.i3244 ], [ 0.000000e+00, %._crit_edge1942.i.i.i3548 ], [ %.21496.ph.i.i.i3348, %.preheader1855.i.i.i3329 ]
204+
store double 0.000000e+00, ptr addrspace(1) %ptr, align 8
205+
br label %._crit_edge1942.i.i.i3548
206+
}
207+
208+
declare double @llvm.fma.f64(double, double, double) #1
209+
210+
attributes #0 = { "amdgpu-waves-per-eu"="8,8" "target-cpu"="gfx942" }
211+
attributes #1 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }

0 commit comments

Comments
 (0)