Skip to content

Commit 57ba58d

Browse files
fhahnfhahn
committed
[LV] Modernize version-mem-access.ll tests.
Auto-generate CHECK lines and simplify tests a bit.
1 parent 279a81e commit 57ba58d

File tree

1 file changed

+96
-66
lines changed

1 file changed

+96
-66
lines changed

‎llvm/test/Transforms/LoopVectorize/version-mem-access.ll‎

Lines changed: 96 additions & 66 deletions
Original file line numberDiff line numberDiff line change
@@ -1,94 +1,124 @@
1-
; RUN: opt -passes=loop-vectorize -enable-mem-access-versioning -force-vector-width=2 -force-vector-interleave=1 < %s -S | FileCheck %s
2-
3-
target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --filter-out-after "scalar.ph:" --version 6
2+
; RUN: opt -passes=loop-vectorize -enable-mem-access-versioning -force-vector-width=2 -force-vector-interleave=1 -S %s | FileCheck %s
43

54
; Check that we version this loop with speculating the value 1 for symbolic
65
; strides. This also checks that the symbolic stride information is correctly
76
; propagated to the memcheck generation. Without this the loop wouldn't
87
; vectorize because we couldn't determine the array bounds for the required
98
; memchecks.
109

11-
; CHECK-LABEL: test
12-
define void @test(ptr %A, i64 %AStride,
13-
ptr %B, i32 %BStride,
14-
ptr %C, i64 %CStride, i32 %N) {
10+
define void @test(ptr noalias %A, i64 %AStride, ptr noalias %B, i32 %BStride, ptr noalias %C, i64 %CStride) {
11+
; CHECK-LABEL: define void @test(
12+
; CHECK-SAME: ptr noalias [[A:%.*]], i64 [[ASTRIDE:%.*]], ptr noalias [[B:%.*]], i32 [[BSTRIDE:%.*]], ptr noalias [[C:%.*]], i64 [[CSTRIDE:%.*]]) {
13+
; CHECK-NEXT: [[ENTRY:.*:]]
14+
; CHECK-NEXT: br label %[[VECTOR_SCEVCHECK:.*]]
15+
; CHECK: [[VECTOR_SCEVCHECK]]:
16+
; CHECK-NEXT: [[IDENT_CHECK1:%.*]] = icmp ne i32 [[BSTRIDE]], 1
17+
; CHECK-NEXT: [[IDENT_CHECK2:%.*]] = icmp ne i64 [[CSTRIDE]], 1
18+
; CHECK-NEXT: [[DIFF_CHECK6:%.*]] = icmp ne i64 [[ASTRIDE]], 1
19+
; CHECK-NEXT: [[DIFF_CHECK:%.*]] = or i1 [[IDENT_CHECK1]], [[IDENT_CHECK2]]
20+
; CHECK-NEXT: [[CONFLICT_RDX:%.*]] = or i1 [[DIFF_CHECK]], [[DIFF_CHECK6]]
21+
; CHECK-NEXT: br i1 [[CONFLICT_RDX]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
22+
; CHECK: [[VECTOR_PH]]:
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:%.*]] = trunc i64 [[INDEX]] to i32
27+
; CHECK-NEXT: [[TMP5:%.*]] = zext i32 [[TMP4]] to i64
28+
; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[TMP5]]
29+
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, ptr [[TMP6]], align 4
30+
; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[INDEX]]
31+
; CHECK-NEXT: [[WIDE_LOAD7:%.*]] = load <2 x i32>, ptr [[TMP7]], align 4
32+
; CHECK-NEXT: [[TMP8:%.*]] = mul nsw <2 x i32> [[WIDE_LOAD7]], [[WIDE_LOAD]]
33+
; CHECK-NEXT: [[TMP9:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDEX]]
34+
; CHECK-NEXT: store <2 x i32> [[TMP8]], ptr [[TMP9]], align 4
35+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
36+
; CHECK-NEXT: [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], 100
37+
; CHECK-NEXT: br i1 [[TMP10]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
38+
; CHECK: [[MIDDLE_BLOCK]]:
39+
; CHECK-NEXT: br [[EXIT:label %.*]]
40+
; CHECK: [[SCALAR_PH]]:
41+
;
1542
entry:
16-
%cmp13 = icmp eq i32 %N, 0
17-
br i1 %cmp13, label %for.end, label %for.body.preheader
18-
19-
; CHECK-DAG: icmp ne i64 %AStride, 1
20-
; CHECK-DAG: icmp ne i32 %BStride, 1
21-
; CHECK-DAG: icmp ne i64 %CStride, 1
22-
; CHECK: or
23-
; CHECK: or
24-
; CHECK: br
25-
26-
; CHECK: vector.body
27-
; CHECK: load <2 x i32>
43+
br label %loop
2844

29-
for.body.preheader:
30-
br label %for.body
31-
32-
for.body:
33-
%indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %for.body.preheader ]
34-
%iv.trunc = trunc i64 %indvars.iv to i32
45+
loop:
46+
%iv = phi i64 [ %iv.next, %loop ], [ 0, %entry ]
47+
%iv.trunc = trunc i64 %iv to i32
3548
%mul = mul i32 %iv.trunc, %BStride
3649
%mul64 = zext i32 %mul to i64
37-
%arrayidx = getelementptr inbounds i32, ptr %B, i64 %mul64
38-
%0 = load i32, ptr %arrayidx, align 4
39-
%mul2 = mul nsw i64 %indvars.iv, %CStride
40-
%arrayidx3 = getelementptr inbounds i32, ptr %C, i64 %mul2
41-
%1 = load i32, ptr %arrayidx3, align 4
50+
%gep.x = getelementptr inbounds i32, ptr %B, i64 %mul64
51+
%0 = load i32, ptr %gep.x, align 4
52+
%mul2 = mul nsw i64 %iv, %CStride
53+
%gep.c = getelementptr inbounds i32, ptr %C, i64 %mul2
54+
%1 = load i32, ptr %gep.c, align 4
4255
%mul4 = mul nsw i32 %1, %0
43-
%mul3 = mul nsw i64 %indvars.iv, %AStride
44-
%arrayidx7 = getelementptr inbounds i32, ptr %A, i64 %mul3
45-
store i32 %mul4, ptr %arrayidx7, align 4
46-
%indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
47-
%lftr.wideiv = trunc i64 %indvars.iv.next to i32
48-
%exitcond = icmp eq i32 %lftr.wideiv, %N
49-
br i1 %exitcond, label %for.end.loopexit, label %for.body
50-
51-
for.end.loopexit:
52-
br label %for.end
53-
54-
for.end:
56+
%mul3 = mul nsw i64 %iv, %AStride
57+
%gep.a = getelementptr inbounds i32, ptr %A, i64 %mul3
58+
store i32 %mul4, ptr %gep.a, align 4
59+
%iv.next = add nuw nsw i64 %iv, 1
60+
%ec = icmp eq i64 %iv.next, 100
61+
br i1 %ec, label %exit, label %loop
62+
63+
exit:
5564
ret void
5665
}
5766

5867
; We used to crash on this function because we removed the fptosi cast when
5968
; replacing the symbolic stride '%conv'.
6069
; PR18480
6170

62-
; CHECK-LABEL: fn1
63-
; CHECK: load <2 x double>
64-
6571
define void @fn1(ptr noalias %x, ptr noalias %c, double %a) {
72+
; CHECK-LABEL: define void @fn1(
73+
; CHECK-SAME: ptr noalias [[X:%.*]], ptr noalias [[C:%.*]], double [[A:%.*]]) {
74+
; CHECK-NEXT: [[ENTRY:.*:]]
75+
; CHECK-NEXT: [[CONV:%.*]] = fptosi double [[A]] to i32
76+
; CHECK-NEXT: [[CONV2:%.*]] = add i32 [[CONV]], 4
77+
; CHECK-NEXT: [[CMP8:%.*]] = icmp sgt i32 [[CONV2]], 0
78+
; CHECK-NEXT: br i1 [[CMP8]], label %[[LOOP_PREHEADER:.*]], [[EXIT:label %.*]]
79+
; CHECK: [[LOOP_PREHEADER]]:
80+
; CHECK-NEXT: [[TMP0:%.*]] = zext i32 [[CONV2]] to i64
81+
; CHECK-NEXT: br label %[[VECTOR_SCEVCHECK:.*]]
82+
; CHECK: [[VECTOR_SCEVCHECK]]:
83+
; CHECK-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i32 [[CONV]], 1
84+
; CHECK-NEXT: br i1 [[IDENT_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
85+
; CHECK: [[VECTOR_PH]]:
86+
; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
87+
; CHECK: [[VECTOR_BODY]]:
88+
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
89+
; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[INDEX]] to i32
90+
; CHECK-NEXT: [[TMP2:%.*]] = sext i32 [[TMP1]] to i64
91+
; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds double, ptr [[X]], i64 [[TMP2]]
92+
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x double>, ptr [[TMP3]], align 8
93+
; CHECK-NEXT: [[TMP4:%.*]] = getelementptr inbounds double, ptr [[C]], i64 [[INDEX]]
94+
; CHECK-NEXT: store <2 x double> [[WIDE_LOAD]], ptr [[TMP4]], align 8
95+
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
96+
; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], 4
97+
; CHECK-NEXT: br i1 [[TMP5]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
98+
; CHECK: [[MIDDLE_BLOCK]]:
99+
; CHECK-NEXT: br label %[[SCALAR_PH]]
100+
; CHECK: [[SCALAR_PH]]:
101+
;
66102
entry:
67103
%conv = fptosi double %a to i32
68104
%conv2 = add i32 %conv, 4
69105
%cmp8 = icmp sgt i32 %conv2, 0
70-
br i1 %cmp8, label %for.body.preheader, label %for.end
71-
72-
for.body.preheader:
73-
br label %for.body
106+
br i1 %cmp8, label %loop, label %exit
74107

75-
for.body:
76-
%indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %for.body.preheader ]
77-
%0 = trunc i64 %indvars.iv to i32
108+
loop:
109+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
110+
%0 = trunc i64 %iv to i32
78111
%mul = mul nsw i32 %0, %conv
79-
%idxprom = sext i32 %mul to i64
80-
%arrayidx = getelementptr inbounds double, ptr %x, i64 %idxprom
81-
%1 = load double, ptr %arrayidx, align 8
82-
%arrayidx3 = getelementptr inbounds double, ptr %c, i64 %indvars.iv
83-
store double %1, ptr %arrayidx3, align 8
84-
%indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
85-
%lftr.wideiv = trunc i64 %indvars.iv.next to i32
86-
%exitcond = icmp eq i32 %lftr.wideiv, %conv2
87-
br i1 %exitcond, label %for.end.loopexit, label %for.body
88-
89-
for.end.loopexit:
90-
br label %for.end
91-
92-
for.end:
112+
%mul.ext = sext i32 %mul to i64
113+
%gep.x = getelementptr inbounds double, ptr %x, i64 %mul.ext
114+
%1 = load double, ptr %gep.x, align 8
115+
%gep.c = getelementptr inbounds double, ptr %c, i64 %iv
116+
store double %1, ptr %gep.c, align 8
117+
%iv.next = add nuw nsw i64 %iv, 1
118+
%iv.trunc = trunc i64 %iv.next to i32
119+
%ec = icmp eq i32 %iv.trunc, %conv2
120+
br i1 %ec, label %exit, label %loop
121+
122+
exit:
93123
ret void
94124
}

0 commit comments

Comments
 (0)