69 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=+avx512vnni,+avx512vl,+avx512bw | FileCheck %s --check-prefixes=CHECK3 4define <16 x i32> @test_pmaddwd_v32i16_add_v16i32(<16 x i32> %a0, <32 x i16> %a1, <32 x i16> %a2) {5; CHECK-LABEL: test_pmaddwd_v32i16_add_v16i32:6; CHECK: # %bb.0:7; CHECK-NEXT: vpdpwssd %zmm2, %zmm1, %zmm08; CHECK-NEXT: retq9 %1 = call <16 x i32> @llvm.x86.avx512.pmaddw.d.512(<32 x i16> %a1, <32 x i16> %a2)10 %2 = add <16 x i32> %1, %a011 ret <16 x i32> %212}13 14define <16 x i32> @test_pmaddwd_v32i16_add_v16i32_commute(<16 x i32> %a0, <32 x i16> %a1, <32 x i16> %a2) {15; CHECK-LABEL: test_pmaddwd_v32i16_add_v16i32_commute:16; CHECK: # %bb.0:17; CHECK-NEXT: vpdpwssd %zmm2, %zmm1, %zmm018; CHECK-NEXT: retq19 %1 = call <16 x i32> @llvm.x86.avx512.pmaddw.d.512(<32 x i16> %a1, <32 x i16> %a2)20 %2 = add <16 x i32> %a0, %121 ret <16 x i32> %222}23 24define <16 x i32> @test_pmaddwd_v32i16_add_v16i32_load1(<16 x i32> %a0, ptr %p1, <32 x i16> %a2) {25; CHECK-LABEL: test_pmaddwd_v32i16_add_v16i32_load1:26; CHECK: # %bb.0:27; CHECK-NEXT: vpdpwssd (%rdi), %zmm1, %zmm028; CHECK-NEXT: retq29 %a1 = load <32 x i16>, ptr %p130 %1 = call <16 x i32> @llvm.x86.avx512.pmaddw.d.512(<32 x i16> %a1, <32 x i16> %a2)31 %2 = add <16 x i32> %1, %a032 ret <16 x i32> %233}34 35define <16 x i32> @test_pmaddwd_v32i16_add_v16i32_load2(<16 x i32> %a0, <32 x i16> %a1, ptr %p2) {36; CHECK-LABEL: test_pmaddwd_v32i16_add_v16i32_load2:37; CHECK: # %bb.0:38; CHECK-NEXT: vpdpwssd (%rdi), %zmm1, %zmm039; CHECK-NEXT: retq40 %a2 = load <32 x i16>, ptr %p241 %1 = call <16 x i32> @llvm.x86.avx512.pmaddw.d.512(<32 x i16> %a1, <32 x i16> %a2)42 %2 = add <16 x i32> %1, %a043 ret <16 x i32> %244}45 46define <16 x i32> @test_pmaddwd_v32i16_add_v16i32_commute_load1(<16 x i32> %a0, ptr %p1, <32 x i16> %a2) {47; CHECK-LABEL: test_pmaddwd_v32i16_add_v16i32_commute_load1:48; CHECK: # %bb.0:49; CHECK-NEXT: vpdpwssd (%rdi), %zmm1, %zmm050; CHECK-NEXT: retq51 %a1 = load <32 x i16>, ptr %p152 %1 = call <16 x i32> @llvm.x86.avx512.pmaddw.d.512(<32 x i16> %a1, <32 x i16> %a2)53 %2 = add <16 x i32> %a0, %154 ret <16 x i32> %255}56 57define <16 x i32> @test_pmaddwd_v32i16_add_v16i32_commute_load2(<16 x i32> %a0, <32 x i16> %a1, ptr %p2) {58; CHECK-LABEL: test_pmaddwd_v32i16_add_v16i32_commute_load2:59; CHECK: # %bb.0:60; CHECK-NEXT: vpdpwssd (%rdi), %zmm1, %zmm061; CHECK-NEXT: retq62 %a2 = load <32 x i16>, ptr %p263 %1 = call <16 x i32> @llvm.x86.avx512.pmaddw.d.512(<32 x i16> %a1, <32 x i16> %a2)64 %2 = add <16 x i32> %a0, %165 ret <16 x i32> %266}67 68declare <16 x i32> @llvm.x86.avx512.pmaddw.d.512(<32 x i16>, <32 x i16>)69