Skip to content

Commit f9da4c6

Browse files
alexey-bataevalexey-bataev
committed
[SLP][NFC]Add a test with extending the types for vectorized
stores/insertelement instructions, NFC.
1 parent 60caa8e commit f9da4c6

File tree

1 file changed

+109
-0
lines changed

1 file changed

+109
-0
lines changed

‎llvm/test/Transforms/SLPVectorizer/X86/store-insertelement-minbitwidth.ll

Lines changed: 109 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,109 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2+
; RUN: opt < %s -S -mtriple=x86_64-unknown -passes=slp-vectorizer -mattr=+avx -pass-remarks-output=%t | FileCheck %s
3+
; RUN: FileCheck --input-file=%t --check-prefix=YAML %s
4+
5+
; YAML-LABEL: --- !Passed
6+
; YAML-NEXT: Pass: slp-vectorizer
7+
; YAML-NEXT: Name: StoresVectorized
8+
; YAML-NEXT: Function: stores
9+
; YAML-NEXT: Args:
10+
; YAML-NEXT: - String: 'Stores SLP vectorized with cost '
11+
; YAML-NEXT: - Cost: '-3'
12+
; YAML-NEXT: - String: ' and with tree size '
13+
; YAML-NEXT: - TreeSize: '6'
14+
define void @stores(ptr noalias %in, ptr noalias %inn, ptr noalias %out) {
15+
; CHECK-LABEL: @stores(
16+
; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i8>, ptr [[IN:%.*]], align 1
17+
; CHECK-NEXT: [[TMP2:%.*]] = load <4 x i8>, ptr [[INN:%.*]], align 1
18+
; CHECK-NEXT: [[TMP3:%.*]] = zext <4 x i8> [[TMP1]] to <4 x i64>
19+
; CHECK-NEXT: [[TMP4:%.*]] = zext <4 x i8> [[TMP2]] to <4 x i64>
20+
; CHECK-NEXT: [[TMP5:%.*]] = add <4 x i64> [[TMP3]], [[TMP4]]
21+
; CHECK-NEXT: store <4 x i64> [[TMP5]], ptr [[OUT:%.*]], align 4
22+
; CHECK-NEXT: ret void
23+
;
24+
%load.1 = load i8, ptr %in, align 1
25+
%gep.1 = getelementptr inbounds i8, ptr %in, i64 1
26+
%load.2 = load i8, ptr %gep.1, align 1
27+
%gep.2 = getelementptr inbounds i8, ptr %in, i64 2
28+
%load.3 = load i8, ptr %gep.2, align 1
29+
%gep.3 = getelementptr inbounds i8, ptr %in, i64 3
30+
%load.4 = load i8, ptr %gep.3, align 1
31+
%load.5 = load i8, ptr %inn, align 1
32+
%gep.4 = getelementptr inbounds i8, ptr %inn, i64 1
33+
%load.6 = load i8, ptr %gep.4, align 1
34+
%gep.5 = getelementptr inbounds i8, ptr %inn, i64 2
35+
%load.7 = load i8, ptr %gep.5, align 1
36+
%gep.6 = getelementptr inbounds i8, ptr %inn, i64 3
37+
%load.8 = load i8, ptr %gep.6, align 1
38+
%z1 = zext i8 %load.1 to i64
39+
%z2 = zext i8 %load.2 to i64
40+
%z3 = zext i8 %load.3 to i64
41+
%z4 = zext i8 %load.4 to i64
42+
%z5 = zext i8 %load.5 to i64
43+
%z6 = zext i8 %load.6 to i64
44+
%z7 = zext i8 %load.7 to i64
45+
%z8 = zext i8 %load.8 to i64
46+
%add1 = add i64 %z1, %z5
47+
%add2 = add i64 %z2, %z6
48+
%add3 = add i64 %z3, %z7
49+
%add4 = add i64 %z4, %z8
50+
%gep.8 = getelementptr inbounds i64, ptr %out, i64 1
51+
%gep.9 = getelementptr inbounds i64, ptr %out, i64 2
52+
%gep.10 = getelementptr inbounds i64, ptr %out, i64 3
53+
store i64 %add1, ptr %out, align 4
54+
store i64 %add2, ptr %gep.8, align 4
55+
store i64 %add3, ptr %gep.9, align 4
56+
store i64 %add4, ptr %gep.10, align 4
57+
ret void
58+
}
59+
60+
; YAML-LABEL: --- !Passed
61+
; YAML-NEXT: Pass: slp-vectorizer
62+
; YAML-NEXT: Name: VectorizedList
63+
; YAML-NEXT: Function: insertelems
64+
; YAML-NEXT: Args:
65+
; YAML-NEXT: - String: 'SLP vectorized with cost '
66+
; YAML-NEXT: - Cost: '-5'
67+
; YAML-NEXT: - String: ' and with tree size '
68+
; YAML-NEXT: - TreeSize: '6'
69+
define <4 x i64> @insertelems(ptr noalias %in, ptr noalias %inn) {
70+
; CHECK-LABEL: @insertelems(
71+
; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i8>, ptr [[IN:%.*]], align 1
72+
; CHECK-NEXT: [[TMP2:%.*]] = load <4 x i8>, ptr [[INN:%.*]], align 1
73+
; CHECK-NEXT: [[TMP3:%.*]] = zext <4 x i8> [[TMP1]] to <4 x i64>
74+
; CHECK-NEXT: [[TMP4:%.*]] = zext <4 x i8> [[TMP2]] to <4 x i64>
75+
; CHECK-NEXT: [[TMP5:%.*]] = add <4 x i64> [[TMP3]], [[TMP4]]
76+
; CHECK-NEXT: ret <4 x i64> [[TMP5]]
77+
;
78+
%load.1 = load i8, ptr %in, align 1
79+
%gep.1 = getelementptr inbounds i8, ptr %in, i64 1
80+
%load.2 = load i8, ptr %gep.1, align 1
81+
%gep.2 = getelementptr inbounds i8, ptr %in, i64 2
82+
%load.3 = load i8, ptr %gep.2, align 1
83+
%gep.3 = getelementptr inbounds i8, ptr %in, i64 3
84+
%load.4 = load i8, ptr %gep.3, align 1
85+
%load.5 = load i8, ptr %inn, align 1
86+
%gep.4 = getelementptr inbounds i8, ptr %inn, i64 1
87+
%load.6 = load i8, ptr %gep.4, align 1
88+
%gep.5 = getelementptr inbounds i8, ptr %inn, i64 2
89+
%load.7 = load i8, ptr %gep.5, align 1
90+
%gep.6 = getelementptr inbounds i8, ptr %inn, i64 3
91+
%load.8 = load i8, ptr %gep.6, align 1
92+
%z1 = zext i8 %load.1 to i64
93+
%z2 = zext i8 %load.2 to i64
94+
%z3 = zext i8 %load.3 to i64
95+
%z4 = zext i8 %load.4 to i64
96+
%z5 = zext i8 %load.5 to i64
97+
%z6 = zext i8 %load.6 to i64
98+
%z7 = zext i8 %load.7 to i64
99+
%z8 = zext i8 %load.8 to i64
100+
%add1 = add i64 %z1, %z5
101+
%add2 = add i64 %z2, %z6
102+
%add3 = add i64 %z3, %z7
103+
%add4 = add i64 %z4, %z8
104+
%ins1 = insertelement <4 x i64> poison, i64 %add1, i32 0
105+
%ins2 = insertelement <4 x i64> %ins1, i64 %add2, i32 1
106+
%ins3 = insertelement <4 x i64> %ins2, i64 %add3, i32 2
107+
%ins4 = insertelement <4 x i64> %ins3, i64 %add4, i32 3
108+
ret <4 x i64> %ins4
109+
}

0 commit comments

Comments
 (0)