196 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avxvnni | FileCheck %s --check-prefix=AVX3; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx512vnni,+avx512vl,+avx512bw | FileCheck %s --check-prefix=AVX5124; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx512vnni,+avx512vl,+avx512bw,+avxvnni | FileCheck %s --check-prefix=AVX5 6define <4 x i32> @test_pmaddwd_v8i16_add_v4i32(<4 x i32> %a0, <8 x i16> %a1, <8 x i16> %a2) {7; AVX-LABEL: test_pmaddwd_v8i16_add_v4i32:8; AVX: # %bb.0:9; AVX-NEXT: {vex} vpdpwssd %xmm2, %xmm1, %xmm010; AVX-NEXT: retq11;12; AVX512-LABEL: test_pmaddwd_v8i16_add_v4i32:13; AVX512: # %bb.0:14; AVX512-NEXT: vpdpwssd %xmm2, %xmm1, %xmm015; AVX512-NEXT: retq16 %1 = call <4 x i32> @llvm.x86.sse2.pmadd.wd(<8 x i16> %a1, <8 x i16> %a2)17 %2 = add <4 x i32> %1, %a018 ret <4 x i32> %219}20 21define <4 x i32> @test_pmaddwd_v8i16_add_v4i32_commute(<4 x i32> %a0, <8 x i16> %a1, <8 x i16> %a2) {22; AVX-LABEL: test_pmaddwd_v8i16_add_v4i32_commute:23; AVX: # %bb.0:24; AVX-NEXT: {vex} vpdpwssd %xmm2, %xmm1, %xmm025; AVX-NEXT: retq26;27; AVX512-LABEL: test_pmaddwd_v8i16_add_v4i32_commute:28; AVX512: # %bb.0:29; AVX512-NEXT: vpdpwssd %xmm2, %xmm1, %xmm030; AVX512-NEXT: retq31 %1 = call <4 x i32> @llvm.x86.sse2.pmadd.wd(<8 x i16> %a1, <8 x i16> %a2)32 %2 = add <4 x i32> %a0, %133 ret <4 x i32> %234}35 36define <4 x i32> @test_pmaddwd_v8i16_add_v4i32_load1(<4 x i32> %a0, ptr %p1, <8 x i16> %a2) {37; AVX-LABEL: test_pmaddwd_v8i16_add_v4i32_load1:38; AVX: # %bb.0:39; AVX-NEXT: {vex} vpdpwssd (%rdi), %xmm1, %xmm040; AVX-NEXT: retq41;42; AVX512-LABEL: test_pmaddwd_v8i16_add_v4i32_load1:43; AVX512: # %bb.0:44; AVX512-NEXT: vpdpwssd (%rdi), %xmm1, %xmm045; AVX512-NEXT: retq46 %a1 = load <8 x i16>, ptr %p147 %1 = call <4 x i32> @llvm.x86.sse2.pmadd.wd(<8 x i16> %a1, <8 x i16> %a2)48 %2 = add <4 x i32> %1, %a049 ret <4 x i32> %250}51 52define <4 x i32> @test_pmaddwd_v8i16_add_v4i32_load2(<4 x i32> %a0, <8 x i16> %a1, ptr %p2) {53; AVX-LABEL: test_pmaddwd_v8i16_add_v4i32_load2:54; AVX: # %bb.0:55; AVX-NEXT: {vex} vpdpwssd (%rdi), %xmm1, %xmm056; AVX-NEXT: retq57;58; AVX512-LABEL: test_pmaddwd_v8i16_add_v4i32_load2:59; AVX512: # %bb.0:60; AVX512-NEXT: vpdpwssd (%rdi), %xmm1, %xmm061; AVX512-NEXT: retq62 %a2 = load <8 x i16>, ptr %p263 %1 = call <4 x i32> @llvm.x86.sse2.pmadd.wd(<8 x i16> %a1, <8 x i16> %a2)64 %2 = add <4 x i32> %1, %a065 ret <4 x i32> %266}67 68define <4 x i32> @test_pmaddwd_v8i16_add_v4i32_commute_load1(<4 x i32> %a0, ptr %p1, <8 x i16> %a2) {69; AVX-LABEL: test_pmaddwd_v8i16_add_v4i32_commute_load1:70; AVX: # %bb.0:71; AVX-NEXT: {vex} vpdpwssd (%rdi), %xmm1, %xmm072; AVX-NEXT: retq73;74; AVX512-LABEL: test_pmaddwd_v8i16_add_v4i32_commute_load1:75; AVX512: # %bb.0:76; AVX512-NEXT: vpdpwssd (%rdi), %xmm1, %xmm077; AVX512-NEXT: retq78 %a1 = load <8 x i16>, ptr %p179 %1 = call <4 x i32> @llvm.x86.sse2.pmadd.wd(<8 x i16> %a1, <8 x i16> %a2)80 %2 = add <4 x i32> %a0, %181 ret <4 x i32> %282}83 84define <4 x i32> @test_pmaddwd_v8i16_add_v4i32_commute_load2(<4 x i32> %a0, <8 x i16> %a1, ptr %p2) {85; AVX-LABEL: test_pmaddwd_v8i16_add_v4i32_commute_load2:86; AVX: # %bb.0:87; AVX-NEXT: {vex} vpdpwssd (%rdi), %xmm1, %xmm088; AVX-NEXT: retq89;90; AVX512-LABEL: test_pmaddwd_v8i16_add_v4i32_commute_load2:91; AVX512: # %bb.0:92; AVX512-NEXT: vpdpwssd (%rdi), %xmm1, %xmm093; AVX512-NEXT: retq94 %a2 = load <8 x i16>, ptr %p295 %1 = call <4 x i32> @llvm.x86.sse2.pmadd.wd(<8 x i16> %a1, <8 x i16> %a2)96 %2 = add <4 x i32> %a0, %197 ret <4 x i32> %298}99 100define <8 x i32> @test_pmaddwd_v16i16_add_v8i32(<8 x i32> %a0, <16 x i16> %a1, <16 x i16> %a2) {101; AVX-LABEL: test_pmaddwd_v16i16_add_v8i32:102; AVX: # %bb.0:103; AVX-NEXT: {vex} vpdpwssd %ymm2, %ymm1, %ymm0104; AVX-NEXT: retq105;106; AVX512-LABEL: test_pmaddwd_v16i16_add_v8i32:107; AVX512: # %bb.0:108; AVX512-NEXT: vpdpwssd %ymm2, %ymm1, %ymm0109; AVX512-NEXT: retq110 %1 = call <8 x i32> @llvm.x86.avx2.pmadd.wd(<16 x i16> %a1, <16 x i16> %a2)111 %2 = add <8 x i32> %1, %a0112 ret <8 x i32> %2113}114 115define <8 x i32> @test_pmaddwd_v16i16_add_v8i32_commute(<8 x i32> %a0, <16 x i16> %a1, <16 x i16> %a2) {116; AVX-LABEL: test_pmaddwd_v16i16_add_v8i32_commute:117; AVX: # %bb.0:118; AVX-NEXT: {vex} vpdpwssd %ymm2, %ymm1, %ymm0119; AVX-NEXT: retq120;121; AVX512-LABEL: test_pmaddwd_v16i16_add_v8i32_commute:122; AVX512: # %bb.0:123; AVX512-NEXT: vpdpwssd %ymm2, %ymm1, %ymm0124; AVX512-NEXT: retq125 %1 = call <8 x i32> @llvm.x86.avx2.pmadd.wd(<16 x i16> %a1, <16 x i16> %a2)126 %2 = add <8 x i32> %a0, %1127 ret <8 x i32> %2128}129 130define <8 x i32> @test_pmaddwd_v16i16_add_v8i32_load1(<8 x i32> %a0, ptr %p1, <16 x i16> %a2) {131; AVX-LABEL: test_pmaddwd_v16i16_add_v8i32_load1:132; AVX: # %bb.0:133; AVX-NEXT: {vex} vpdpwssd (%rdi), %ymm1, %ymm0134; AVX-NEXT: retq135;136; AVX512-LABEL: test_pmaddwd_v16i16_add_v8i32_load1:137; AVX512: # %bb.0:138; AVX512-NEXT: vpdpwssd (%rdi), %ymm1, %ymm0139; AVX512-NEXT: retq140 %a1 = load <16 x i16>, ptr %p1141 %1 = call <8 x i32> @llvm.x86.avx2.pmadd.wd(<16 x i16> %a1, <16 x i16> %a2)142 %2 = add <8 x i32> %1, %a0143 ret <8 x i32> %2144}145 146define <8 x i32> @test_pmaddwd_v16i16_add_v8i32_load2(<8 x i32> %a0, <16 x i16> %a1, ptr %p2) {147; AVX-LABEL: test_pmaddwd_v16i16_add_v8i32_load2:148; AVX: # %bb.0:149; AVX-NEXT: {vex} vpdpwssd (%rdi), %ymm1, %ymm0150; AVX-NEXT: retq151;152; AVX512-LABEL: test_pmaddwd_v16i16_add_v8i32_load2:153; AVX512: # %bb.0:154; AVX512-NEXT: vpdpwssd (%rdi), %ymm1, %ymm0155; AVX512-NEXT: retq156 %a2 = load <16 x i16>, ptr %p2157 %1 = call <8 x i32> @llvm.x86.avx2.pmadd.wd(<16 x i16> %a1, <16 x i16> %a2)158 %2 = add <8 x i32> %1, %a0159 ret <8 x i32> %2160}161 162define <8 x i32> @test_pmaddwd_v16i16_add_v8i32_commute_load1(<8 x i32> %a0, ptr %p1, <16 x i16> %a2) {163; AVX-LABEL: test_pmaddwd_v16i16_add_v8i32_commute_load1:164; AVX: # %bb.0:165; AVX-NEXT: {vex} vpdpwssd (%rdi), %ymm1, %ymm0166; AVX-NEXT: retq167;168; AVX512-LABEL: test_pmaddwd_v16i16_add_v8i32_commute_load1:169; AVX512: # %bb.0:170; AVX512-NEXT: vpdpwssd (%rdi), %ymm1, %ymm0171; AVX512-NEXT: retq172 %a1 = load <16 x i16>, ptr %p1173 %1 = call <8 x i32> @llvm.x86.avx2.pmadd.wd(<16 x i16> %a1, <16 x i16> %a2)174 %2 = add <8 x i32> %a0, %1175 ret <8 x i32> %2176}177 178define <8 x i32> @test_pmaddwd_v16i16_add_v8i32_commute_load2(<8 x i32> %a0, <16 x i16> %a1, ptr %p2) {179; AVX-LABEL: test_pmaddwd_v16i16_add_v8i32_commute_load2:180; AVX: # %bb.0:181; AVX-NEXT: {vex} vpdpwssd (%rdi), %ymm1, %ymm0182; AVX-NEXT: retq183;184; AVX512-LABEL: test_pmaddwd_v16i16_add_v8i32_commute_load2:185; AVX512: # %bb.0:186; AVX512-NEXT: vpdpwssd (%rdi), %ymm1, %ymm0187; AVX512-NEXT: retq188 %a2 = load <16 x i16>, ptr %p2189 %1 = call <8 x i32> @llvm.x86.avx2.pmadd.wd(<16 x i16> %a1, <16 x i16> %a2)190 %2 = add <8 x i32> %a0, %1191 ret <8 x i32> %2192}193 194declare <4 x i32> @llvm.x86.sse2.pmadd.wd(<8 x i16>, <8 x i16>)195declare <8 x i32> @llvm.x86.avx2.pmadd.wd(<16 x i16>, <16 x i16>)196