1 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx,+slow-unaligned-mem-32 | FileCheck %s --check-prefix=AVXSLOW
2 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx,-slow-unaligned-mem-32 | FileCheck %s --check-prefix=AVXFAST
3 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx2 | FileCheck %s --check-prefix=AVX2
5 ; Don't generate an unaligned 32-byte load on this test if that is slower than two 16-byte loads.
7 define <8 x float> @load32bytes(<8 x float>* %Ap) {
8 ; AVXSLOW-LABEL: load32bytes:
10 ; AVXSLOW-NEXT: vmovaps (%rdi), %xmm0
11 ; AVXSLOW-NEXT: vinsertf128 $1, 16(%rdi), %ymm0, %ymm0
14 ; AVXFAST-LABEL: load32bytes:
16 ; AVXFAST-NEXT: vmovups (%rdi), %ymm0
19 ; AVX2-LABEL: load32bytes:
21 ; AVX2-NEXT: vmovups (%rdi), %ymm0
23 %A = load <8 x float>, <8 x float>* %Ap, align 16
27 ; Don't generate an unaligned 32-byte store on this test if that is slower than two 16-byte loads.
29 define void @store32bytes(<8 x float> %A, <8 x float>* %P) {
30 ; AVXSLOW-LABEL: store32bytes:
32 ; AVXSLOW-NEXT: vextractf128 $1, %ymm0, 16(%rdi)
33 ; AVXSLOW-NEXT: vmovaps %xmm0, (%rdi)
34 ; AVXSLOW-NEXT: vzeroupper
37 ; AVXFAST-LABEL: store32bytes:
39 ; AVXFAST-NEXT: vmovups %ymm0, (%rdi)
40 ; AVXFAST-NEXT: vzeroupper
43 ; AVX2-LABEL: store32bytes:
45 ; AVX2-NEXT: vmovups %ymm0, (%rdi)
46 ; AVX2-NEXT: vzeroupper
48 store <8 x float> %A, <8 x float>* %P, align 16
52 ; Merge two consecutive 16-byte subvector loads into a single 32-byte load if it's faster.
54 define <8 x float> @combine_16_byte_loads_no_intrinsic(<4 x float>* %ptr) {
55 ; AVXSLOW-LABEL: combine_16_byte_loads_no_intrinsic:
57 ; AVXSLOW-NEXT: vmovups 48(%rdi), %xmm0
58 ; AVXSLOW-NEXT: vinsertf128 $1, 64(%rdi), %ymm0, %ymm0
61 ; AVXFAST-LABEL: combine_16_byte_loads_no_intrinsic:
63 ; AVXFAST-NEXT: vmovups 48(%rdi), %ymm0
66 ; AVX2-LABEL: combine_16_byte_loads_no_intrinsic:
68 ; AVX2-NEXT: vmovups 48(%rdi), %ymm0
70 %ptr1 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 3
71 %ptr2 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 4
72 %v1 = load <4 x float>, <4 x float>* %ptr1, align 1
73 %v2 = load <4 x float>, <4 x float>* %ptr2, align 1
74 %v3 = shufflevector <4 x float> %v1, <4 x float> %v2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
78 ; If the first load is 32-byte aligned, then the loads should be merged in all cases.
80 define <8 x float> @combine_16_byte_loads_aligned(<4 x float>* %ptr) {
81 ; AVXSLOW-LABEL: combine_16_byte_loads_aligned:
83 ; AVXSLOW-NEXT: vmovaps 48(%rdi), %ymm0
86 ; AVXFAST-LABEL: combine_16_byte_loads_aligned:
88 ; AVXFAST-NEXT: vmovaps 48(%rdi), %ymm0
91 ; AVX2-LABEL: combine_16_byte_loads_aligned:
93 ; AVX2-NEXT: vmovaps 48(%rdi), %ymm0
95 %ptr1 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 3
96 %ptr2 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 4
97 %v1 = load <4 x float>, <4 x float>* %ptr1, align 32
98 %v2 = load <4 x float>, <4 x float>* %ptr2, align 1
99 %v3 = shufflevector <4 x float> %v1, <4 x float> %v2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
103 ; Swap the order of the shufflevector operands to ensure that the pattern still matches.
105 define <8 x float> @combine_16_byte_loads_no_intrinsic_swap(<4 x float>* %ptr) {
106 ; AVXSLOW-LABEL: combine_16_byte_loads_no_intrinsic_swap:
108 ; AVXSLOW-NEXT: vmovups 64(%rdi), %xmm0
109 ; AVXSLOW-NEXT: vinsertf128 $1, 80(%rdi), %ymm0, %ymm0
112 ; AVXFAST-LABEL: combine_16_byte_loads_no_intrinsic_swap:
114 ; AVXFAST-NEXT: vmovups 64(%rdi), %ymm0
117 ; AVX2-LABEL: combine_16_byte_loads_no_intrinsic_swap:
119 ; AVX2-NEXT: vmovups 64(%rdi), %ymm0
121 %ptr1 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 4
122 %ptr2 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 5
123 %v1 = load <4 x float>, <4 x float>* %ptr1, align 1
124 %v2 = load <4 x float>, <4 x float>* %ptr2, align 1
125 %v3 = shufflevector <4 x float> %v2, <4 x float> %v1, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 0, i32 1, i32 2, i32 3>
129 ; Check each element type other than float to make sure it is handled correctly.
130 ; Use the loaded values with an 'add' to make sure we're using the correct load type.
131 ; Don't generate 32-byte loads for integer ops unless we have AVX2.
133 define <4 x i64> @combine_16_byte_loads_i64(<2 x i64>* %ptr, <4 x i64> %x) {
134 ; AVXSLOW-LABEL: combine_16_byte_loads_i64:
136 ; AVXSLOW-NEXT: vextractf128 $1, %ymm0, %xmm1
137 ; AVXSLOW-NEXT: vpaddq 96(%rdi), %xmm1, %xmm1
138 ; AVXSLOW-NEXT: vpaddq 80(%rdi), %xmm0, %xmm0
139 ; AVXSLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
142 ; AVXFAST-LABEL: combine_16_byte_loads_i64:
144 ; AVXFAST-NEXT: vextractf128 $1, %ymm0, %xmm1
145 ; AVXFAST-NEXT: vpaddq 96(%rdi), %xmm1, %xmm1
146 ; AVXFAST-NEXT: vpaddq 80(%rdi), %xmm0, %xmm0
147 ; AVXFAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
150 ; AVX2-LABEL: combine_16_byte_loads_i64:
152 ; AVX2-NEXT: vpaddq 80(%rdi), %ymm0, %ymm0
154 %ptr1 = getelementptr inbounds <2 x i64>, <2 x i64>* %ptr, i64 5
155 %ptr2 = getelementptr inbounds <2 x i64>, <2 x i64>* %ptr, i64 6
156 %v1 = load <2 x i64>, <2 x i64>* %ptr1, align 1
157 %v2 = load <2 x i64>, <2 x i64>* %ptr2, align 1
158 %v3 = shufflevector <2 x i64> %v1, <2 x i64> %v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
159 %v4 = add <4 x i64> %v3, %x
163 define <8 x i32> @combine_16_byte_loads_i32(<4 x i32>* %ptr, <8 x i32> %x) {
164 ; AVXSLOW-LABEL: combine_16_byte_loads_i32:
166 ; AVXSLOW-NEXT: vextractf128 $1, %ymm0, %xmm1
167 ; AVXSLOW-NEXT: vpaddd 112(%rdi), %xmm1, %xmm1
168 ; AVXSLOW-NEXT: vpaddd 96(%rdi), %xmm0, %xmm0
169 ; AVXSLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
172 ; AVXFAST-LABEL: combine_16_byte_loads_i32:
174 ; AVXFAST-NEXT: vextractf128 $1, %ymm0, %xmm1
175 ; AVXFAST-NEXT: vpaddd 112(%rdi), %xmm1, %xmm1
176 ; AVXFAST-NEXT: vpaddd 96(%rdi), %xmm0, %xmm0
177 ; AVXFAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
180 ; AVX2-LABEL: combine_16_byte_loads_i32:
182 ; AVX2-NEXT: vpaddd 96(%rdi), %ymm0, %ymm0
184 %ptr1 = getelementptr inbounds <4 x i32>, <4 x i32>* %ptr, i64 6
185 %ptr2 = getelementptr inbounds <4 x i32>, <4 x i32>* %ptr, i64 7
186 %v1 = load <4 x i32>, <4 x i32>* %ptr1, align 1
187 %v2 = load <4 x i32>, <4 x i32>* %ptr2, align 1
188 %v3 = shufflevector <4 x i32> %v1, <4 x i32> %v2, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
189 %v4 = add <8 x i32> %v3, %x
193 define <16 x i16> @combine_16_byte_loads_i16(<8 x i16>* %ptr, <16 x i16> %x) {
194 ; AVXSLOW-LABEL: combine_16_byte_loads_i16:
196 ; AVXSLOW-NEXT: vextractf128 $1, %ymm0, %xmm1
197 ; AVXSLOW-NEXT: vpaddw 128(%rdi), %xmm1, %xmm1
198 ; AVXSLOW-NEXT: vpaddw 112(%rdi), %xmm0, %xmm0
199 ; AVXSLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
202 ; AVXFAST-LABEL: combine_16_byte_loads_i16:
204 ; AVXFAST-NEXT: vextractf128 $1, %ymm0, %xmm1
205 ; AVXFAST-NEXT: vpaddw 128(%rdi), %xmm1, %xmm1
206 ; AVXFAST-NEXT: vpaddw 112(%rdi), %xmm0, %xmm0
207 ; AVXFAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
210 ; AVX2-LABEL: combine_16_byte_loads_i16:
212 ; AVX2-NEXT: vpaddw 112(%rdi), %ymm0, %ymm0
214 %ptr1 = getelementptr inbounds <8 x i16>, <8 x i16>* %ptr, i64 7
215 %ptr2 = getelementptr inbounds <8 x i16>, <8 x i16>* %ptr, i64 8
216 %v1 = load <8 x i16>, <8 x i16>* %ptr1, align 1
217 %v2 = load <8 x i16>, <8 x i16>* %ptr2, align 1
218 %v3 = shufflevector <8 x i16> %v1, <8 x i16> %v2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
219 %v4 = add <16 x i16> %v3, %x
223 define <32 x i8> @combine_16_byte_loads_i8(<16 x i8>* %ptr, <32 x i8> %x) {
224 ; AVXSLOW-LABEL: combine_16_byte_loads_i8:
226 ; AVXSLOW-NEXT: vextractf128 $1, %ymm0, %xmm1
227 ; AVXSLOW-NEXT: vpaddb 144(%rdi), %xmm1, %xmm1
228 ; AVXSLOW-NEXT: vpaddb 128(%rdi), %xmm0, %xmm0
229 ; AVXSLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
232 ; AVXFAST-LABEL: combine_16_byte_loads_i8:
234 ; AVXFAST-NEXT: vextractf128 $1, %ymm0, %xmm1
235 ; AVXFAST-NEXT: vpaddb 144(%rdi), %xmm1, %xmm1
236 ; AVXFAST-NEXT: vpaddb 128(%rdi), %xmm0, %xmm0
237 ; AVXFAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
240 ; AVX2-LABEL: combine_16_byte_loads_i8:
242 ; AVX2-NEXT: vpaddb 128(%rdi), %ymm0, %ymm0
244 %ptr1 = getelementptr inbounds <16 x i8>, <16 x i8>* %ptr, i64 8
245 %ptr2 = getelementptr inbounds <16 x i8>, <16 x i8>* %ptr, i64 9
246 %v1 = load <16 x i8>, <16 x i8>* %ptr1, align 1
247 %v2 = load <16 x i8>, <16 x i8>* %ptr2, align 1
248 %v3 = shufflevector <16 x i8> %v1, <16 x i8> %v2, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
249 %v4 = add <32 x i8> %v3, %x
253 define <4 x double> @combine_16_byte_loads_double(<2 x double>* %ptr, <4 x double> %x) {
254 ; AVXSLOW-LABEL: combine_16_byte_loads_double:
256 ; AVXSLOW-NEXT: vmovupd 144(%rdi), %xmm1
257 ; AVXSLOW-NEXT: vinsertf128 $1, 160(%rdi), %ymm1, %ymm1
258 ; AVXSLOW-NEXT: vaddpd %ymm0, %ymm1, %ymm0
261 ; AVXFAST-LABEL: combine_16_byte_loads_double:
263 ; AVXFAST-NEXT: vaddpd 144(%rdi), %ymm0, %ymm0
266 ; AVX2-LABEL: combine_16_byte_loads_double:
268 ; AVX2-NEXT: vaddpd 144(%rdi), %ymm0, %ymm0
270 %ptr1 = getelementptr inbounds <2 x double>, <2 x double>* %ptr, i64 9
271 %ptr2 = getelementptr inbounds <2 x double>, <2 x double>* %ptr, i64 10
272 %v1 = load <2 x double>, <2 x double>* %ptr1, align 1
273 %v2 = load <2 x double>, <2 x double>* %ptr2, align 1
274 %v3 = shufflevector <2 x double> %v1, <2 x double> %v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
275 %v4 = fadd <4 x double> %v3, %x