280 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx,+slow-unaligned-mem-32 | FileCheck %s --check-prefix=AVXSLOW3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx,-slow-unaligned-mem-32 | FileCheck %s --check-prefix=AVXFAST4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx2 | FileCheck %s --check-prefix=AVX25 6; Don't generate an unaligned 32-byte load on this test if that is slower than two 16-byte loads.7 8define <8 x float> @load32bytes(ptr %Ap) {9; AVXSLOW-LABEL: load32bytes:10; AVXSLOW: # %bb.0:11; AVXSLOW-NEXT: vmovaps (%rdi), %xmm012; AVXSLOW-NEXT: vinsertf128 $1, 16(%rdi), %ymm0, %ymm013; AVXSLOW-NEXT: retq14;15; AVXFAST-LABEL: load32bytes:16; AVXFAST: # %bb.0:17; AVXFAST-NEXT: vmovups (%rdi), %ymm018; AVXFAST-NEXT: retq19;20; AVX2-LABEL: load32bytes:21; AVX2: # %bb.0:22; AVX2-NEXT: vmovups (%rdi), %ymm023; AVX2-NEXT: retq24 %A = load <8 x float>, ptr %Ap, align 1625 ret <8 x float> %A26}27 28; Don't generate an unaligned 32-byte store on this test if that is slower than two 16-byte loads.29 30define void @store32bytes(<8 x float> %A, ptr %P) {31; AVXSLOW-LABEL: store32bytes:32; AVXSLOW: # %bb.0:33; AVXSLOW-NEXT: vextractf128 $1, %ymm0, 16(%rdi)34; AVXSLOW-NEXT: vmovaps %xmm0, (%rdi)35; AVXSLOW-NEXT: vzeroupper36; AVXSLOW-NEXT: retq37;38; AVXFAST-LABEL: store32bytes:39; AVXFAST: # %bb.0:40; AVXFAST-NEXT: vmovups %ymm0, (%rdi)41; AVXFAST-NEXT: vzeroupper42; AVXFAST-NEXT: retq43;44; AVX2-LABEL: store32bytes:45; AVX2: # %bb.0:46; AVX2-NEXT: vmovups %ymm0, (%rdi)47; AVX2-NEXT: vzeroupper48; AVX2-NEXT: retq49 store <8 x float> %A, ptr %P, align 1650 ret void51}52 53; Merge two consecutive 16-byte subvector loads into a single 32-byte load if it's faster.54 55define <8 x float> @combine_16_byte_loads_no_intrinsic(ptr %ptr) {56; AVXSLOW-LABEL: combine_16_byte_loads_no_intrinsic:57; AVXSLOW: # %bb.0:58; AVXSLOW-NEXT: vmovups 48(%rdi), %xmm059; AVXSLOW-NEXT: vinsertf128 $1, 64(%rdi), %ymm0, %ymm060; AVXSLOW-NEXT: retq61;62; AVXFAST-LABEL: combine_16_byte_loads_no_intrinsic:63; AVXFAST: # %bb.0:64; AVXFAST-NEXT: vmovups 48(%rdi), %ymm065; AVXFAST-NEXT: retq66;67; AVX2-LABEL: combine_16_byte_loads_no_intrinsic:68; AVX2: # %bb.0:69; AVX2-NEXT: vmovups 48(%rdi), %ymm070; AVX2-NEXT: retq71 %ptr1 = getelementptr inbounds <4 x float>, ptr %ptr, i64 372 %ptr2 = getelementptr inbounds <4 x float>, ptr %ptr, i64 473 %v1 = load <4 x float>, ptr %ptr1, align 174 %v2 = load <4 x float>, ptr %ptr2, align 175 %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>76 ret <8 x float> %v377}78 79; If the first load is 32-byte aligned, then the loads should be merged in all cases.80 81define <8 x float> @combine_16_byte_loads_aligned(ptr %ptr) {82; AVXSLOW-LABEL: combine_16_byte_loads_aligned:83; AVXSLOW: # %bb.0:84; AVXSLOW-NEXT: vmovaps 48(%rdi), %ymm085; AVXSLOW-NEXT: retq86;87; AVXFAST-LABEL: combine_16_byte_loads_aligned:88; AVXFAST: # %bb.0:89; AVXFAST-NEXT: vmovaps 48(%rdi), %ymm090; AVXFAST-NEXT: retq91;92; AVX2-LABEL: combine_16_byte_loads_aligned:93; AVX2: # %bb.0:94; AVX2-NEXT: vmovaps 48(%rdi), %ymm095; AVX2-NEXT: retq96 %ptr1 = getelementptr inbounds <4 x float>, ptr %ptr, i64 397 %ptr2 = getelementptr inbounds <4 x float>, ptr %ptr, i64 498 %v1 = load <4 x float>, ptr %ptr1, align 3299 %v2 = load <4 x float>, ptr %ptr2, align 1100 %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>101 ret <8 x float> %v3102}103 104; Swap the order of the shufflevector operands to ensure that the pattern still matches.105 106define <8 x float> @combine_16_byte_loads_no_intrinsic_swap(ptr %ptr) {107; AVXSLOW-LABEL: combine_16_byte_loads_no_intrinsic_swap:108; AVXSLOW: # %bb.0:109; AVXSLOW-NEXT: vmovups 64(%rdi), %xmm0110; AVXSLOW-NEXT: vinsertf128 $1, 80(%rdi), %ymm0, %ymm0111; AVXSLOW-NEXT: retq112;113; AVXFAST-LABEL: combine_16_byte_loads_no_intrinsic_swap:114; AVXFAST: # %bb.0:115; AVXFAST-NEXT: vmovups 64(%rdi), %ymm0116; AVXFAST-NEXT: retq117;118; AVX2-LABEL: combine_16_byte_loads_no_intrinsic_swap:119; AVX2: # %bb.0:120; AVX2-NEXT: vmovups 64(%rdi), %ymm0121; AVX2-NEXT: retq122 %ptr1 = getelementptr inbounds <4 x float>, ptr %ptr, i64 4123 %ptr2 = getelementptr inbounds <4 x float>, ptr %ptr, i64 5124 %v1 = load <4 x float>, ptr %ptr1, align 1125 %v2 = load <4 x float>, ptr %ptr2, align 1126 %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>127 ret <8 x float> %v3128}129 130; Check each element type other than float to make sure it is handled correctly.131; Use the loaded values with an 'add' to make sure we're using the correct load type.132; Don't generate 32-byte loads for integer ops unless we have AVX2.133 134define <4 x i64> @combine_16_byte_loads_i64(ptr %ptr, <4 x i64> %x) {135; AVXSLOW-LABEL: combine_16_byte_loads_i64:136; AVXSLOW: # %bb.0:137; AVXSLOW-NEXT: vextractf128 $1, %ymm0, %xmm1138; AVXSLOW-NEXT: vpaddq 96(%rdi), %xmm1, %xmm1139; AVXSLOW-NEXT: vpaddq 80(%rdi), %xmm0, %xmm0140; AVXSLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0141; AVXSLOW-NEXT: retq142;143; AVXFAST-LABEL: combine_16_byte_loads_i64:144; AVXFAST: # %bb.0:145; AVXFAST-NEXT: vextractf128 $1, %ymm0, %xmm1146; AVXFAST-NEXT: vpaddq 96(%rdi), %xmm1, %xmm1147; AVXFAST-NEXT: vpaddq 80(%rdi), %xmm0, %xmm0148; AVXFAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0149; AVXFAST-NEXT: retq150;151; AVX2-LABEL: combine_16_byte_loads_i64:152; AVX2: # %bb.0:153; AVX2-NEXT: vpaddq 80(%rdi), %ymm0, %ymm0154; AVX2-NEXT: retq155 %ptr1 = getelementptr inbounds <2 x i64>, ptr %ptr, i64 5156 %ptr2 = getelementptr inbounds <2 x i64>, ptr %ptr, i64 6157 %v1 = load <2 x i64>, ptr %ptr1, align 1158 %v2 = load <2 x i64>, ptr %ptr2, align 1159 %v3 = shufflevector <2 x i64> %v1, <2 x i64> %v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>160 %v4 = add <4 x i64> %v3, %x161 ret <4 x i64> %v4162}163 164define <8 x i32> @combine_16_byte_loads_i32(ptr %ptr, <8 x i32> %x) {165; AVXSLOW-LABEL: combine_16_byte_loads_i32:166; AVXSLOW: # %bb.0:167; AVXSLOW-NEXT: vextractf128 $1, %ymm0, %xmm1168; AVXSLOW-NEXT: vpaddd 112(%rdi), %xmm1, %xmm1169; AVXSLOW-NEXT: vpaddd 96(%rdi), %xmm0, %xmm0170; AVXSLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0171; AVXSLOW-NEXT: retq172;173; AVXFAST-LABEL: combine_16_byte_loads_i32:174; AVXFAST: # %bb.0:175; AVXFAST-NEXT: vextractf128 $1, %ymm0, %xmm1176; AVXFAST-NEXT: vpaddd 112(%rdi), %xmm1, %xmm1177; AVXFAST-NEXT: vpaddd 96(%rdi), %xmm0, %xmm0178; AVXFAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0179; AVXFAST-NEXT: retq180;181; AVX2-LABEL: combine_16_byte_loads_i32:182; AVX2: # %bb.0:183; AVX2-NEXT: vpaddd 96(%rdi), %ymm0, %ymm0184; AVX2-NEXT: retq185 %ptr1 = getelementptr inbounds <4 x i32>, ptr %ptr, i64 6186 %ptr2 = getelementptr inbounds <4 x i32>, ptr %ptr, i64 7187 %v1 = load <4 x i32>, ptr %ptr1, align 1188 %v2 = load <4 x i32>, ptr %ptr2, align 1189 %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>190 %v4 = add <8 x i32> %v3, %x191 ret <8 x i32> %v4192}193 194define <16 x i16> @combine_16_byte_loads_i16(ptr %ptr, <16 x i16> %x) {195; AVXSLOW-LABEL: combine_16_byte_loads_i16:196; AVXSLOW: # %bb.0:197; AVXSLOW-NEXT: vextractf128 $1, %ymm0, %xmm1198; AVXSLOW-NEXT: vpaddw 128(%rdi), %xmm1, %xmm1199; AVXSLOW-NEXT: vpaddw 112(%rdi), %xmm0, %xmm0200; AVXSLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0201; AVXSLOW-NEXT: retq202;203; AVXFAST-LABEL: combine_16_byte_loads_i16:204; AVXFAST: # %bb.0:205; AVXFAST-NEXT: vextractf128 $1, %ymm0, %xmm1206; AVXFAST-NEXT: vpaddw 128(%rdi), %xmm1, %xmm1207; AVXFAST-NEXT: vpaddw 112(%rdi), %xmm0, %xmm0208; AVXFAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0209; AVXFAST-NEXT: retq210;211; AVX2-LABEL: combine_16_byte_loads_i16:212; AVX2: # %bb.0:213; AVX2-NEXT: vpaddw 112(%rdi), %ymm0, %ymm0214; AVX2-NEXT: retq215 %ptr1 = getelementptr inbounds <8 x i16>, ptr %ptr, i64 7216 %ptr2 = getelementptr inbounds <8 x i16>, ptr %ptr, i64 8217 %v1 = load <8 x i16>, ptr %ptr1, align 1218 %v2 = load <8 x i16>, ptr %ptr2, align 1219 %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>220 %v4 = add <16 x i16> %v3, %x221 ret <16 x i16> %v4222}223 224define <32 x i8> @combine_16_byte_loads_i8(ptr %ptr, <32 x i8> %x) {225; AVXSLOW-LABEL: combine_16_byte_loads_i8:226; AVXSLOW: # %bb.0:227; AVXSLOW-NEXT: vextractf128 $1, %ymm0, %xmm1228; AVXSLOW-NEXT: vpaddb 144(%rdi), %xmm1, %xmm1229; AVXSLOW-NEXT: vpaddb 128(%rdi), %xmm0, %xmm0230; AVXSLOW-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0231; AVXSLOW-NEXT: retq232;233; AVXFAST-LABEL: combine_16_byte_loads_i8:234; AVXFAST: # %bb.0:235; AVXFAST-NEXT: vextractf128 $1, %ymm0, %xmm1236; AVXFAST-NEXT: vpaddb 144(%rdi), %xmm1, %xmm1237; AVXFAST-NEXT: vpaddb 128(%rdi), %xmm0, %xmm0238; AVXFAST-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0239; AVXFAST-NEXT: retq240;241; AVX2-LABEL: combine_16_byte_loads_i8:242; AVX2: # %bb.0:243; AVX2-NEXT: vpaddb 128(%rdi), %ymm0, %ymm0244; AVX2-NEXT: retq245 %ptr1 = getelementptr inbounds <16 x i8>, ptr %ptr, i64 8246 %ptr2 = getelementptr inbounds <16 x i8>, ptr %ptr, i64 9247 %v1 = load <16 x i8>, ptr %ptr1, align 1248 %v2 = load <16 x i8>, ptr %ptr2, align 1249 %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>250 %v4 = add <32 x i8> %v3, %x251 ret <32 x i8> %v4252}253 254define <4 x double> @combine_16_byte_loads_double(ptr %ptr, <4 x double> %x) {255; AVXSLOW-LABEL: combine_16_byte_loads_double:256; AVXSLOW: # %bb.0:257; AVXSLOW-NEXT: vmovups 144(%rdi), %xmm1258; AVXSLOW-NEXT: vinsertf128 $1, 160(%rdi), %ymm1, %ymm1259; AVXSLOW-NEXT: vaddpd %ymm0, %ymm1, %ymm0260; AVXSLOW-NEXT: retq261;262; AVXFAST-LABEL: combine_16_byte_loads_double:263; AVXFAST: # %bb.0:264; AVXFAST-NEXT: vaddpd 144(%rdi), %ymm0, %ymm0265; AVXFAST-NEXT: retq266;267; AVX2-LABEL: combine_16_byte_loads_double:268; AVX2: # %bb.0:269; AVX2-NEXT: vaddpd 144(%rdi), %ymm0, %ymm0270; AVX2-NEXT: retq271 %ptr1 = getelementptr inbounds <2 x double>, ptr %ptr, i64 9272 %ptr2 = getelementptr inbounds <2 x double>, ptr %ptr, i64 10273 %v1 = load <2 x double>, ptr %ptr1, align 1274 %v2 = load <2 x double>, ptr %ptr2, align 1275 %v3 = shufflevector <2 x double> %v1, <2 x double> %v2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>276 %v4 = fadd <4 x double> %v3, %x277 ret <4 x double> %v4278}279 280