Skip to content

Commit 07e5f57

Browse files
fhahnfhahn
committed
[LV] Add tests for #60831.
Also contains an extra test mentioned in D144434.
1 parent ab21bea commit 07e5f57

File tree

1 file changed

+231
-0
lines changed

1 file changed

+231
-0
lines changed

‎llvm/test/Transforms/LoopVectorize/AArch64/pr60831-sve-inv-store-crash.ll

Lines changed: 231 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,231 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2+
; RUN: opt -passes=loop-vectorize -S -prefer-predicate-over-epilogue=scalar-epilogue < %s | FileCheck %s
3+
4+
target triple = "aarch64-unknown-linux-gnu"
5+
6+
; REQUIRES: asserts
7+
; XFAIL: *
8+
9+
; Test cases for PR60831.
10+
11+
define void @test_invar_gep(ptr %dst) #0 {
12+
; CHECK-LABEL: @test_invar_gep(
13+
; CHECK-NEXT: entry:
14+
; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
15+
; CHECK-NEXT: [[TMP1:%.*]] = mul i64 [[TMP0]], 4
16+
; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 100, [[TMP1]]
17+
; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
18+
; CHECK: vector.ph:
19+
; CHECK-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
20+
; CHECK-NEXT: [[TMP3:%.*]] = mul i64 [[TMP2]], 4
21+
; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 100, [[TMP3]]
22+
; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 100, [[N_MOD_VF]]
23+
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
24+
; CHECK: vector.body:
25+
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
26+
; CHECK-NEXT: [[TMP4:%.*]] = call <vscale x 4 x i64> @llvm.experimental.stepvector.nxv4i64()
27+
; CHECK-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[INDEX]], i64 0
28+
; CHECK-NEXT: [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i64> [[DOTSPLATINSERT]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
29+
; CHECK-NEXT: [[TMP5:%.*]] = add <vscale x 4 x i64> zeroinitializer, [[TMP4]]
30+
; CHECK-NEXT: [[TMP6:%.*]] = mul <vscale x 4 x i64> [[TMP5]], shufflevector (<vscale x 4 x i64> insertelement (<vscale x 4 x i64> poison, i64 1, i64 0), <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer)
31+
; CHECK-NEXT: [[TMP7:%.*]] = add <vscale x 4 x i64> [[DOTSPLAT]], [[TMP6]]
32+
; CHECK-NEXT: [[TMP8:%.*]] = add i64 [[INDEX]], 0
33+
; CHECK-NEXT: [[TMP9:%.*]] = add i64 [[INDEX]], 1
34+
; CHECK-NEXT: [[TMP10:%.*]] = add i64 [[INDEX]], 2
35+
; CHECK-NEXT: [[TMP11:%.*]] = add i64 [[INDEX]], 3
36+
; CHECK-NEXT: [[TMP12:%.*]] = getelementptr i8, ptr [[DST:%.*]], i64 0
37+
; CHECK-NEXT: [[TMP13:%.*]] = call i32 @llvm.vscale.i32()
38+
; CHECK-NEXT: [[TMP14:%.*]] = mul i32 [[TMP13]], 4
39+
; CHECK-NEXT: [[TMP15:%.*]] = sub i32 [[TMP14]], 1
40+
; CHECK-NEXT: [[TMP16:%.*]] = extractelement <vscale x 4 x i64> [[TMP7]], i32 [[TMP15]]
41+
; CHECK-NEXT: store i64 [[TMP16]], ptr [[TMP12]], align 1
42+
; CHECK-NEXT: [[TMP17:%.*]] = call i64 @llvm.vscale.i64()
43+
; CHECK-NEXT: [[TMP18:%.*]] = mul i64 [[TMP17]], 4
44+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP18]]
45+
; CHECK-NEXT: [[TMP19:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
46+
; CHECK-NEXT: br i1 [[TMP19]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
47+
; CHECK: middle.block:
48+
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 100, [[N_VEC]]
49+
; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
50+
; CHECK: scalar.ph:
51+
; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
52+
; CHECK-NEXT: br label [[LOOP:%.*]]
53+
; CHECK: loop:
54+
; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ]
55+
; CHECK-NEXT: [[GEP_INVAR:%.*]] = getelementptr i8, ptr [[DST]], i64 0
56+
; CHECK-NEXT: store i64 [[IV]], ptr [[GEP_INVAR]], align 1
57+
; CHECK-NEXT: [[IV_NEXT]] = add nsw i64 [[IV]], 1
58+
; CHECK-NEXT: [[EC:%.*]] = icmp eq i64 [[IV_NEXT]], 100
59+
; CHECK-NEXT: br i1 [[EC]], label [[EXIT]], label [[LOOP]], !llvm.loop [[LOOP4:![0-9]+]]
60+
; CHECK: exit:
61+
; CHECK-NEXT: ret void
62+
;
63+
entry:
64+
br label %loop
65+
66+
loop:
67+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
68+
%gep.invar = getelementptr i8, ptr %dst, i64 0
69+
store i64 %iv, ptr %gep.invar, align 1
70+
%iv.next = add nsw i64 %iv, 1
71+
%ec = icmp eq i64 %iv.next, 100
72+
br i1 %ec, label %exit, label %loop, !llvm.loop !0
73+
74+
exit:
75+
ret void
76+
}
77+
78+
define void @test_loop2(i64 %n, ptr %dst) {
79+
; CHECK-LABEL: @test_loop2(
80+
; CHECK-NEXT: iter.check:
81+
; CHECK-NEXT: br i1 false, label [[VEC_EPILOG_SCALAR_PH:%.*]], label [[VECTOR_MAIN_LOOP_ITER_CHECK:%.*]]
82+
; CHECK: vector.main.loop.iter.check:
83+
; CHECK-NEXT: br i1 false, label [[VEC_EPILOG_PH:%.*]], label [[VECTOR_PH:%.*]]
84+
; CHECK: vector.ph:
85+
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
86+
; CHECK: vector.body:
87+
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
88+
; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[INDEX]], 0
89+
; CHECK-NEXT: [[TMP1:%.*]] = add i64 [[INDEX]], 1
90+
; CHECK-NEXT: [[TMP2:%.*]] = add i64 [[INDEX]], 2
91+
; CHECK-NEXT: [[TMP3:%.*]] = add i64 [[INDEX]], 3
92+
; CHECK-NEXT: [[TMP4:%.*]] = add i64 [[INDEX]], 4
93+
; CHECK-NEXT: [[TMP5:%.*]] = add i64 [[INDEX]], 5
94+
; CHECK-NEXT: [[TMP6:%.*]] = add i64 [[INDEX]], 6
95+
; CHECK-NEXT: [[TMP7:%.*]] = add i64 [[INDEX]], 7
96+
; CHECK-NEXT: [[TMP8:%.*]] = add i64 [[INDEX]], 8
97+
; CHECK-NEXT: [[TMP9:%.*]] = add i64 [[INDEX]], 9
98+
; CHECK-NEXT: [[TMP10:%.*]] = add i64 [[INDEX]], 10
99+
; CHECK-NEXT: [[TMP11:%.*]] = add i64 [[INDEX]], 11
100+
; CHECK-NEXT: [[TMP12:%.*]] = add i64 [[INDEX]], 12
101+
; CHECK-NEXT: [[TMP13:%.*]] = add i64 [[INDEX]], 13
102+
; CHECK-NEXT: [[TMP14:%.*]] = add i64 [[INDEX]], 14
103+
; CHECK-NEXT: [[TMP15:%.*]] = add i64 [[INDEX]], 15
104+
; CHECK-NEXT: [[TMP16:%.*]] = sub nsw i64 [[N:%.*]], [[TMP0]]
105+
; CHECK-NEXT: [[TMP17:%.*]] = sub nsw i64 [[N]], [[TMP1]]
106+
; CHECK-NEXT: [[TMP18:%.*]] = sub nsw i64 [[N]], [[TMP2]]
107+
; CHECK-NEXT: [[TMP19:%.*]] = sub nsw i64 [[N]], [[TMP3]]
108+
; CHECK-NEXT: [[TMP20:%.*]] = sub nsw i64 [[N]], [[TMP4]]
109+
; CHECK-NEXT: [[TMP21:%.*]] = sub nsw i64 [[N]], [[TMP5]]
110+
; CHECK-NEXT: [[TMP22:%.*]] = sub nsw i64 [[N]], [[TMP6]]
111+
; CHECK-NEXT: [[TMP23:%.*]] = sub nsw i64 [[N]], [[TMP7]]
112+
; CHECK-NEXT: [[TMP24:%.*]] = sub nsw i64 [[N]], [[TMP8]]
113+
; CHECK-NEXT: [[TMP25:%.*]] = sub nsw i64 [[N]], [[TMP9]]
114+
; CHECK-NEXT: [[TMP26:%.*]] = sub nsw i64 [[N]], [[TMP10]]
115+
; CHECK-NEXT: [[TMP27:%.*]] = sub nsw i64 [[N]], [[TMP11]]
116+
; CHECK-NEXT: [[TMP28:%.*]] = sub nsw i64 [[N]], [[TMP12]]
117+
; CHECK-NEXT: [[TMP29:%.*]] = sub nsw i64 [[N]], [[TMP13]]
118+
; CHECK-NEXT: [[TMP30:%.*]] = sub nsw i64 [[N]], [[TMP14]]
119+
; CHECK-NEXT: [[TMP31:%.*]] = sub nsw i64 [[N]], [[TMP15]]
120+
; CHECK-NEXT: [[TMP32:%.*]] = insertelement <16 x i64> poison, i64 [[TMP16]], i32 0
121+
; CHECK-NEXT: [[TMP33:%.*]] = insertelement <16 x i64> [[TMP32]], i64 [[TMP17]], i32 1
122+
; CHECK-NEXT: [[TMP34:%.*]] = insertelement <16 x i64> [[TMP33]], i64 [[TMP18]], i32 2
123+
; CHECK-NEXT: [[TMP35:%.*]] = insertelement <16 x i64> [[TMP34]], i64 [[TMP19]], i32 3
124+
; CHECK-NEXT: [[TMP36:%.*]] = insertelement <16 x i64> [[TMP35]], i64 [[TMP20]], i32 4
125+
; CHECK-NEXT: [[TMP37:%.*]] = insertelement <16 x i64> [[TMP36]], i64 [[TMP21]], i32 5
126+
; CHECK-NEXT: [[TMP38:%.*]] = insertelement <16 x i64> [[TMP37]], i64 [[TMP22]], i32 6
127+
; CHECK-NEXT: [[TMP39:%.*]] = insertelement <16 x i64> [[TMP38]], i64 [[TMP23]], i32 7
128+
; CHECK-NEXT: [[TMP40:%.*]] = insertelement <16 x i64> [[TMP39]], i64 [[TMP24]], i32 8
129+
; CHECK-NEXT: [[TMP41:%.*]] = insertelement <16 x i64> [[TMP40]], i64 [[TMP25]], i32 9
130+
; CHECK-NEXT: [[TMP42:%.*]] = insertelement <16 x i64> [[TMP41]], i64 [[TMP26]], i32 10
131+
; CHECK-NEXT: [[TMP43:%.*]] = insertelement <16 x i64> [[TMP42]], i64 [[TMP27]], i32 11
132+
; CHECK-NEXT: [[TMP44:%.*]] = insertelement <16 x i64> [[TMP43]], i64 [[TMP28]], i32 12
133+
; CHECK-NEXT: [[TMP45:%.*]] = insertelement <16 x i64> [[TMP44]], i64 [[TMP29]], i32 13
134+
; CHECK-NEXT: [[TMP46:%.*]] = insertelement <16 x i64> [[TMP45]], i64 [[TMP30]], i32 14
135+
; CHECK-NEXT: [[TMP47:%.*]] = insertelement <16 x i64> [[TMP46]], i64 [[TMP31]], i32 15
136+
; CHECK-NEXT: [[TMP48:%.*]] = trunc <16 x i64> [[TMP47]] to <16 x i8>
137+
; CHECK-NEXT: [[TMP49:%.*]] = add i64 [[TMP0]], [[TMP16]]
138+
; CHECK-NEXT: [[TMP50:%.*]] = getelementptr i8, ptr [[DST:%.*]], i64 [[TMP49]]
139+
; CHECK-NEXT: [[TMP51:%.*]] = extractelement <16 x i8> [[TMP48]], i32 15
140+
; CHECK-NEXT: store i8 [[TMP51]], ptr [[TMP50]], align 1
141+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 16
142+
; CHECK-NEXT: [[TMP52:%.*]] = icmp eq i64 [[INDEX_NEXT]], 992
143+
; CHECK-NEXT: br i1 [[TMP52]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
144+
; CHECK: middle.block:
145+
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 1001, 992
146+
; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[VEC_EPILOG_ITER_CHECK:%.*]]
147+
; CHECK: vec.epilog.iter.check:
148+
; CHECK-NEXT: br i1 false, label [[VEC_EPILOG_SCALAR_PH]], label [[VEC_EPILOG_PH]]
149+
; CHECK: vec.epilog.ph:
150+
; CHECK-NEXT: [[VEC_EPILOG_RESUME_VAL:%.*]] = phi i64 [ 992, [[VEC_EPILOG_ITER_CHECK]] ], [ 0, [[VECTOR_MAIN_LOOP_ITER_CHECK]] ]
151+
; CHECK-NEXT: br label [[VEC_EPILOG_VECTOR_BODY:%.*]]
152+
; CHECK: vec.epilog.vector.body:
153+
; CHECK-NEXT: [[INDEX2:%.*]] = phi i64 [ [[VEC_EPILOG_RESUME_VAL]], [[VEC_EPILOG_PH]] ], [ [[INDEX_NEXT3:%.*]], [[VEC_EPILOG_VECTOR_BODY]] ]
154+
; CHECK-NEXT: [[TMP53:%.*]] = add i64 [[INDEX2]], 0
155+
; CHECK-NEXT: [[TMP54:%.*]] = add i64 [[INDEX2]], 1
156+
; CHECK-NEXT: [[TMP55:%.*]] = add i64 [[INDEX2]], 2
157+
; CHECK-NEXT: [[TMP56:%.*]] = add i64 [[INDEX2]], 3
158+
; CHECK-NEXT: [[TMP57:%.*]] = add i64 [[INDEX2]], 4
159+
; CHECK-NEXT: [[TMP58:%.*]] = add i64 [[INDEX2]], 5
160+
; CHECK-NEXT: [[TMP59:%.*]] = add i64 [[INDEX2]], 6
161+
; CHECK-NEXT: [[TMP60:%.*]] = add i64 [[INDEX2]], 7
162+
; CHECK-NEXT: [[TMP61:%.*]] = sub nsw i64 [[N]], [[TMP53]]
163+
; CHECK-NEXT: [[TMP62:%.*]] = sub nsw i64 [[N]], [[TMP54]]
164+
; CHECK-NEXT: [[TMP63:%.*]] = sub nsw i64 [[N]], [[TMP55]]
165+
; CHECK-NEXT: [[TMP64:%.*]] = sub nsw i64 [[N]], [[TMP56]]
166+
; CHECK-NEXT: [[TMP65:%.*]] = sub nsw i64 [[N]], [[TMP57]]
167+
; CHECK-NEXT: [[TMP66:%.*]] = sub nsw i64 [[N]], [[TMP58]]
168+
; CHECK-NEXT: [[TMP67:%.*]] = sub nsw i64 [[N]], [[TMP59]]
169+
; CHECK-NEXT: [[TMP68:%.*]] = sub nsw i64 [[N]], [[TMP60]]
170+
; CHECK-NEXT: [[TMP69:%.*]] = insertelement <8 x i64> poison, i64 [[TMP61]], i32 0
171+
; CHECK-NEXT: [[TMP70:%.*]] = insertelement <8 x i64> [[TMP69]], i64 [[TMP62]], i32 1
172+
; CHECK-NEXT: [[TMP71:%.*]] = insertelement <8 x i64> [[TMP70]], i64 [[TMP63]], i32 2
173+
; CHECK-NEXT: [[TMP72:%.*]] = insertelement <8 x i64> [[TMP71]], i64 [[TMP64]], i32 3
174+
; CHECK-NEXT: [[TMP73:%.*]] = insertelement <8 x i64> [[TMP72]], i64 [[TMP65]], i32 4
175+
; CHECK-NEXT: [[TMP74:%.*]] = insertelement <8 x i64> [[TMP73]], i64 [[TMP66]], i32 5
176+
; CHECK-NEXT: [[TMP75:%.*]] = insertelement <8 x i64> [[TMP74]], i64 [[TMP67]], i32 6
177+
; CHECK-NEXT: [[TMP76:%.*]] = insertelement <8 x i64> [[TMP75]], i64 [[TMP68]], i32 7
178+
; CHECK-NEXT: [[TMP77:%.*]] = trunc <8 x i64> [[TMP76]] to <8 x i8>
179+
; CHECK-NEXT: [[TMP78:%.*]] = add i64 [[TMP53]], [[TMP61]]
180+
; CHECK-NEXT: [[TMP79:%.*]] = getelementptr i8, ptr [[DST]], i64 [[TMP78]]
181+
; CHECK-NEXT: [[TMP80:%.*]] = extractelement <8 x i8> [[TMP77]], i32 7
182+
; CHECK-NEXT: store i8 [[TMP80]], ptr [[TMP79]], align 1
183+
; CHECK-NEXT: [[INDEX_NEXT3]] = add nuw i64 [[INDEX2]], 8
184+
; CHECK-NEXT: [[TMP81:%.*]] = icmp eq i64 [[INDEX_NEXT3]], 1000
185+
; CHECK-NEXT: br i1 [[TMP81]], label [[VEC_EPILOG_MIDDLE_BLOCK:%.*]], label [[VEC_EPILOG_VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
186+
; CHECK: vec.epilog.middle.block:
187+
; CHECK-NEXT: [[CMP_N1:%.*]] = icmp eq i64 1001, 1000
188+
; CHECK-NEXT: br i1 [[CMP_N1]], label [[EXIT]], label [[VEC_EPILOG_SCALAR_PH]]
189+
; CHECK: vec.epilog.scalar.ph:
190+
; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ 1000, [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ 992, [[VEC_EPILOG_ITER_CHECK]] ], [ 0, [[ITER_CHECK:%.*]] ]
191+
; CHECK-NEXT: br label [[LOOP:%.*]]
192+
; CHECK: loop:
193+
; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[VEC_EPILOG_SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ]
194+
; CHECK-NEXT: [[SUB_N:%.*]] = sub nsw i64 [[N]], [[IV]]
195+
; CHECK-NEXT: [[SUB_N_TRUNC:%.*]] = trunc i64 [[SUB_N]] to i8
196+
; CHECK-NEXT: [[ADD:%.*]] = add i64 [[IV]], [[SUB_N]]
197+
; CHECK-NEXT: [[GEP:%.*]] = getelementptr i8, ptr [[DST]], i64 [[ADD]]
198+
; CHECK-NEXT: store i8 [[SUB_N_TRUNC]], ptr [[GEP]], align 1
199+
; CHECK-NEXT: [[IV_NEXT]] = add nsw i64 [[IV]], 1
200+
; CHECK-NEXT: [[C:%.*]] = icmp sle i64 [[IV_NEXT]], 1000
201+
; CHECK-NEXT: br i1 [[C]], label [[LOOP]], label [[EXIT]], !llvm.loop [[LOOP7:![0-9]+]]
202+
; CHECK: exit:
203+
; CHECK-NEXT: ret void
204+
;
205+
entry:
206+
br label %loop
207+
208+
loop:
209+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
210+
%sub.n = sub nsw i64 %n, %iv
211+
%sub.n.trunc = trunc i64 %sub.n to i8
212+
%add = add i64 %iv, %sub.n
213+
%gep = getelementptr i8, ptr %dst, i64 %add
214+
store i8 %sub.n.trunc, ptr %gep, align 1
215+
%iv.next = add nsw i64 %iv, 1
216+
%c = icmp sle i64 %iv.next, 1000
217+
br i1 %c, label %loop, label %exit
218+
219+
exit:
220+
ret void
221+
}
222+
223+
attributes #0 = { "target-features"="+neon,+sve" vscale_range(1, 16) }
224+
225+
!0 = distinct !{!0, !1, !2, !3, !4, !5}
226+
!1 = !{!"llvm.loop.mustprogress"}
227+
!2 = !{!"llvm.loop.vectorize.width", i32 4}
228+
!3 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
229+
!4 = !{!"llvm.loop.vectorize.enable", i1 true}
230+
!5 = !{!"llvm.loop.interleave.count", i32 1}
231+

0 commit comments

Comments
 (0)