| blob | eda7f3a2abfd4071be2120c6b8dbb1db6eec95e5 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx | FileCheck %s
4 %structTy = type { i8, i32, i32 }
6 @e = common global %structTy zeroinitializer, align 4
8 ;; Ensure that MergeConsecutiveStores doesn't incorrectly reorder
9 ;; store operations. The first test stores in increasing address
10 ;; order, the second in decreasing -- but in both cases should have
11 ;; the same result in memory in the end.
13 define void @redundant_stores_merging() {
14 ; CHECK-LABEL: redundant_stores_merging:
15 ; CHECK: # %bb.0:
16 ; CHECK-NEXT: movabsq $1958505086977, %rax # imm = 0x1C800000001
17 ; CHECK-NEXT: movq %rax, e+{{.*}}(%rip)
18 ; CHECK-NEXT: retq
19 store i32 1, i32* getelementptr inbounds (%structTy, %structTy* @e, i64 0, i32 1), align 4
20 store i32 123, i32* getelementptr inbounds (%structTy, %structTy* @e, i64 0, i32 2), align 4
21 store i32 456, i32* getelementptr inbounds (%structTy, %structTy* @e, i64 0, i32 2), align 4
22 ret void
23 }
25 ;; This variant tests PR25154.
26 define void @redundant_stores_merging_reverse() {
27 ; CHECK-LABEL: redundant_stores_merging_reverse:
28 ; CHECK: # %bb.0:
29 ; CHECK-NEXT: movabsq $528280977409, %rax # imm = 0x7B00000001
30 ; CHECK-NEXT: movq %rax, e+{{.*}}(%rip)
31 ; CHECK-NEXT: movl $456, e+{{.*}}(%rip) # imm = 0x1C8
32 ; CHECK-NEXT: retq
33 store i32 123, i32* getelementptr inbounds (%structTy, %structTy* @e, i64 0, i32 2), align 4
34 store i32 456, i32* getelementptr inbounds (%structTy, %structTy* @e, i64 0, i32 2), align 4
35 store i32 1, i32* getelementptr inbounds (%structTy, %structTy* @e, i64 0, i32 1), align 4
36 ret void
37 }
39 @b = common global [8 x i8] zeroinitializer, align 2
41 ;; The 2-byte store to offset 3 overlaps the 2-byte store to offset 2;
42 ;; these must not be reordered in MergeConsecutiveStores such that the
43 ;; store to 3 comes first (e.g. by merging the stores to 0 and 2 into
44 ;; a movl, after the store to 3).
46 define void @overlapping_stores_merging() {
47 ; CHECK-LABEL: overlapping_stores_merging:
48 ; CHECK: # %bb.0:
49 ; CHECK-NEXT: movl $1, {{.*}}(%rip)
50 ; CHECK-NEXT: movw $2, b+{{.*}}(%rip)
51 ; CHECK-NEXT: retq
52 store i16 0, i16* bitcast (i8* getelementptr inbounds ([8 x i8], [8 x i8]* @b, i64 0, i64 2) to i16*), align 2
53 store i16 2, i16* bitcast (i8* getelementptr inbounds ([8 x i8], [8 x i8]* @b, i64 0, i64 3) to i16*), align 1
54 store i16 1, i16* bitcast (i8* getelementptr inbounds ([8 x i8], [8 x i8]* @b, i64 0, i64 0) to i16*), align 2
55 ret void
56 }
58 define void @extract_vector_store_16_consecutive_bytes(<2 x i64> %v, i8* %ptr) #0 {
59 ; CHECK-LABEL: extract_vector_store_16_consecutive_bytes:
60 ; CHECK: # %bb.0:
61 ; CHECK-NEXT: vmovups %xmm0, (%rdi)
62 ; CHECK-NEXT: retq
63 %bc = bitcast <2 x i64> %v to <16 x i8>
64 %ext00 = extractelement <16 x i8> %bc, i32 0
65 %ext01 = extractelement <16 x i8> %bc, i32 1
66 %ext02 = extractelement <16 x i8> %bc, i32 2
67 %ext03 = extractelement <16 x i8> %bc, i32 3
68 %ext04 = extractelement <16 x i8> %bc, i32 4
69 %ext05 = extractelement <16 x i8> %bc, i32 5
70 %ext06 = extractelement <16 x i8> %bc, i32 6
71 %ext07 = extractelement <16 x i8> %bc, i32 7
72 %ext08 = extractelement <16 x i8> %bc, i32 8
73 %ext09 = extractelement <16 x i8> %bc, i32 9
74 %ext10 = extractelement <16 x i8> %bc, i32 10
75 %ext11 = extractelement <16 x i8> %bc, i32 11
76 %ext12 = extractelement <16 x i8> %bc, i32 12
77 %ext13 = extractelement <16 x i8> %bc, i32 13
78 %ext14 = extractelement <16 x i8> %bc, i32 14
79 %ext15 = extractelement <16 x i8> %bc, i32 15
80 %gep00 = getelementptr inbounds i8, i8* %ptr, i64 0
81 %gep01 = getelementptr inbounds i8, i8* %ptr, i64 1
82 %gep02 = getelementptr inbounds i8, i8* %ptr, i64 2
83 %gep03 = getelementptr inbounds i8, i8* %ptr, i64 3
84 %gep04 = getelementptr inbounds i8, i8* %ptr, i64 4
85 %gep05 = getelementptr inbounds i8, i8* %ptr, i64 5
86 %gep06 = getelementptr inbounds i8, i8* %ptr, i64 6
87 %gep07 = getelementptr inbounds i8, i8* %ptr, i64 7
88 %gep08 = getelementptr inbounds i8, i8* %ptr, i64 8
89 %gep09 = getelementptr inbounds i8, i8* %ptr, i64 9
90 %gep10 = getelementptr inbounds i8, i8* %ptr, i64 10
91 %gep11 = getelementptr inbounds i8, i8* %ptr, i64 11
92 %gep12 = getelementptr inbounds i8, i8* %ptr, i64 12
93 %gep13 = getelementptr inbounds i8, i8* %ptr, i64 13
94 %gep14 = getelementptr inbounds i8, i8* %ptr, i64 14
95 %gep15 = getelementptr inbounds i8, i8* %ptr, i64 15
96 store i8 %ext00, i8* %gep00, align 1
97 store i8 %ext01, i8* %gep01, align 1
98 store i8 %ext02, i8* %gep02, align 1
99 store i8 %ext03, i8* %gep03, align 1
100 store i8 %ext04, i8* %gep04, align 1
101 store i8 %ext05, i8* %gep05, align 1
102 store i8 %ext06, i8* %gep06, align 1
103 store i8 %ext07, i8* %gep07, align 1
104 store i8 %ext08, i8* %gep08, align 1
105 store i8 %ext09, i8* %gep09, align 1
106 store i8 %ext10, i8* %gep10, align 1
107 store i8 %ext11, i8* %gep11, align 1
108 store i8 %ext12, i8* %gep12, align 1
109 store i8 %ext13, i8* %gep13, align 1
110 store i8 %ext14, i8* %gep14, align 1
111 store i8 %ext15, i8* %gep15, align 1
112 ret void
113 }
115 ; PR34217 - https://bugs.llvm.org/show_bug.cgi?id=34217
117 define void @extract_vector_store_32_consecutive_bytes(<4 x i64> %v, i8* %ptr) #0 {
118 ; CHECK-LABEL: extract_vector_store_32_consecutive_bytes:
119 ; CHECK: # %bb.0:
120 ; CHECK-NEXT: vmovups %ymm0, (%rdi)
121 ; CHECK-NEXT: vzeroupper
122 ; CHECK-NEXT: retq
123 %bc = bitcast <4 x i64> %v to <32 x i8>
124 %ext00 = extractelement <32 x i8> %bc, i32 0
125 %ext01 = extractelement <32 x i8> %bc, i32 1
126 %ext02 = extractelement <32 x i8> %bc, i32 2
127 %ext03 = extractelement <32 x i8> %bc, i32 3
128 %ext04 = extractelement <32 x i8> %bc, i32 4
129 %ext05 = extractelement <32 x i8> %bc, i32 5
130 %ext06 = extractelement <32 x i8> %bc, i32 6
131 %ext07 = extractelement <32 x i8> %bc, i32 7
132 %ext08 = extractelement <32 x i8> %bc, i32 8
133 %ext09 = extractelement <32 x i8> %bc, i32 9
134 %ext10 = extractelement <32 x i8> %bc, i32 10
135 %ext11 = extractelement <32 x i8> %bc, i32 11
136 %ext12 = extractelement <32 x i8> %bc, i32 12
137 %ext13 = extractelement <32 x i8> %bc, i32 13
138 %ext14 = extractelement <32 x i8> %bc, i32 14
139 %ext15 = extractelement <32 x i8> %bc, i32 15
140 %ext16 = extractelement <32 x i8> %bc, i32 16
141 %ext17 = extractelement <32 x i8> %bc, i32 17
142 %ext18 = extractelement <32 x i8> %bc, i32 18
143 %ext19 = extractelement <32 x i8> %bc, i32 19
144 %ext20 = extractelement <32 x i8> %bc, i32 20
145 %ext21 = extractelement <32 x i8> %bc, i32 21
146 %ext22 = extractelement <32 x i8> %bc, i32 22
147 %ext23 = extractelement <32 x i8> %bc, i32 23
148 %ext24 = extractelement <32 x i8> %bc, i32 24
149 %ext25 = extractelement <32 x i8> %bc, i32 25
150 %ext26 = extractelement <32 x i8> %bc, i32 26
151 %ext27 = extractelement <32 x i8> %bc, i32 27
152 %ext28 = extractelement <32 x i8> %bc, i32 28
153 %ext29 = extractelement <32 x i8> %bc, i32 29
154 %ext30 = extractelement <32 x i8> %bc, i32 30
155 %ext31 = extractelement <32 x i8> %bc, i32 31
156 %gep00 = getelementptr inbounds i8, i8* %ptr, i64 0
157 %gep01 = getelementptr inbounds i8, i8* %ptr, i64 1
158 %gep02 = getelementptr inbounds i8, i8* %ptr, i64 2
159 %gep03 = getelementptr inbounds i8, i8* %ptr, i64 3
160 %gep04 = getelementptr inbounds i8, i8* %ptr, i64 4
161 %gep05 = getelementptr inbounds i8, i8* %ptr, i64 5
162 %gep06 = getelementptr inbounds i8, i8* %ptr, i64 6
163 %gep07 = getelementptr inbounds i8, i8* %ptr, i64 7
164 %gep08 = getelementptr inbounds i8, i8* %ptr, i64 8
165 %gep09 = getelementptr inbounds i8, i8* %ptr, i64 9
166 %gep10 = getelementptr inbounds i8, i8* %ptr, i64 10
167 %gep11 = getelementptr inbounds i8, i8* %ptr, i64 11
168 %gep12 = getelementptr inbounds i8, i8* %ptr, i64 12
169 %gep13 = getelementptr inbounds i8, i8* %ptr, i64 13
170 %gep14 = getelementptr inbounds i8, i8* %ptr, i64 14
171 %gep15 = getelementptr inbounds i8, i8* %ptr, i64 15
172 %gep16 = getelementptr inbounds i8, i8* %ptr, i64 16
173 %gep17 = getelementptr inbounds i8, i8* %ptr, i64 17
174 %gep18 = getelementptr inbounds i8, i8* %ptr, i64 18
175 %gep19 = getelementptr inbounds i8, i8* %ptr, i64 19
176 %gep20 = getelementptr inbounds i8, i8* %ptr, i64 20
177 %gep21 = getelementptr inbounds i8, i8* %ptr, i64 21
178 %gep22 = getelementptr inbounds i8, i8* %ptr, i64 22
179 %gep23 = getelementptr inbounds i8, i8* %ptr, i64 23
180 %gep24 = getelementptr inbounds i8, i8* %ptr, i64 24
181 %gep25 = getelementptr inbounds i8, i8* %ptr, i64 25
182 %gep26 = getelementptr inbounds i8, i8* %ptr, i64 26
183 %gep27 = getelementptr inbounds i8, i8* %ptr, i64 27
184 %gep28 = getelementptr inbounds i8, i8* %ptr, i64 28
185 %gep29 = getelementptr inbounds i8, i8* %ptr, i64 29
186 %gep30 = getelementptr inbounds i8, i8* %ptr, i64 30
187 %gep31 = getelementptr inbounds i8, i8* %ptr, i64 31
188 store i8 %ext00, i8* %gep00, align 1
189 store i8 %ext01, i8* %gep01, align 1
190 store i8 %ext02, i8* %gep02, align 1
191 store i8 %ext03, i8* %gep03, align 1
192 store i8 %ext04, i8* %gep04, align 1
193 store i8 %ext05, i8* %gep05, align 1
194 store i8 %ext06, i8* %gep06, align 1
195 store i8 %ext07, i8* %gep07, align 1
196 store i8 %ext08, i8* %gep08, align 1
197 store i8 %ext09, i8* %gep09, align 1
198 store i8 %ext10, i8* %gep10, align 1
199 store i8 %ext11, i8* %gep11, align 1
200 store i8 %ext12, i8* %gep12, align 1
201 store i8 %ext13, i8* %gep13, align 1
202 store i8 %ext14, i8* %gep14, align 1
203 store i8 %ext15, i8* %gep15, align 1
204 store i8 %ext16, i8* %gep16, align 1
205 store i8 %ext17, i8* %gep17, align 1
206 store i8 %ext18, i8* %gep18, align 1
207 store i8 %ext19, i8* %gep19, align 1
208 store i8 %ext20, i8* %gep20, align 1
209 store i8 %ext21, i8* %gep21, align 1
210 store i8 %ext22, i8* %gep22, align 1
211 store i8 %ext23, i8* %gep23, align 1
212 store i8 %ext24, i8* %gep24, align 1
213 store i8 %ext25, i8* %gep25, align 1
214 store i8 %ext26, i8* %gep26, align 1
215 store i8 %ext27, i8* %gep27, align 1
216 store i8 %ext28, i8* %gep28, align 1
217 store i8 %ext29, i8* %gep29, align 1
218 store i8 %ext30, i8* %gep30, align 1
219 store i8 %ext31, i8* %gep31, align 1
220 ret void
221 }