1; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl -mattr=+avx512bw| FileCheck %s 2 3; CHECK-LABEL: vpaddb512_test 4; CHECK: vpaddb %zmm{{.*}} 5; CHECK: ret 6define <64 x i8> @vpaddb512_test(<64 x i8> %i, <64 x i8> %j) nounwind readnone { 7 %x = add <64 x i8> %i, %j 8 ret <64 x i8> %x 9} 10 11; CHECK-LABEL: vpaddb512_fold_test 12; CHECK: vpaddb (%rdi), %zmm{{.*}} 13; CHECK: ret 14define <64 x i8> @vpaddb512_fold_test(<64 x i8> %i, <64 x i8>* %j) nounwind { 15 %tmp = load <64 x i8>, <64 x i8>* %j, align 4 16 %x = add <64 x i8> %i, %tmp 17 ret <64 x i8> %x 18} 19 20; CHECK-LABEL: vpaddw512_test 21; CHECK: vpaddw %zmm{{.*}} 22; CHECK: ret 23define <32 x i16> @vpaddw512_test(<32 x i16> %i, <32 x i16> %j) nounwind readnone { 24 %x = add <32 x i16> %i, %j 25 ret <32 x i16> %x 26} 27 28; CHECK-LABEL: vpaddw512_fold_test 29; CHECK: vpaddw (%rdi), %zmm{{.*}} 30; CHECK: ret 31define <32 x i16> @vpaddw512_fold_test(<32 x i16> %i, <32 x i16>* %j) nounwind { 32 %tmp = load <32 x i16>, <32 x i16>* %j, align 4 33 %x = add <32 x i16> %i, %tmp 34 ret <32 x i16> %x 35} 36 37; CHECK-LABEL: vpaddw512_mask_test 38; CHECK: vpaddw %zmm{{.*%k[1-7].*}} 39; CHECK: ret 40define <32 x i16> @vpaddw512_mask_test(<32 x i16> %i, <32 x i16> %j, <32 x i16> %mask1) nounwind readnone { 41 %mask = icmp ne <32 x i16> %mask1, zeroinitializer 42 %x = add <32 x i16> %i, %j 43 %r = select <32 x i1> %mask, <32 x i16> %x, <32 x i16> %i 44 ret <32 x i16> %r 45} 46 47; CHECK-LABEL: vpaddw512_maskz_test 48; CHECK: vpaddw %zmm{{.*{%k[1-7]} {z}.*}} 49; CHECK: ret 50define <32 x i16> @vpaddw512_maskz_test(<32 x i16> %i, <32 x i16> %j, <32 x i16> %mask1) nounwind readnone { 51 %mask = icmp ne <32 x i16> %mask1, zeroinitializer 52 %x = add <32 x i16> %i, %j 53 %r = select <32 x i1> %mask, <32 x i16> %x, <32 x i16> zeroinitializer 54 ret <32 x i16> %r 55} 56 57; CHECK-LABEL: vpaddw512_mask_fold_test 58; CHECK: vpaddw (%rdi), %zmm{{.*%k[1-7]}} 59; CHECK: ret 60define <32 x i16> @vpaddw512_mask_fold_test(<32 x i16> %i, <32 x i16>* %j.ptr, <32 x i16> %mask1) nounwind readnone { 61 %mask = icmp ne <32 x i16> %mask1, zeroinitializer 62 %j = load <32 x i16>, <32 x i16>* %j.ptr 63 %x = add <32 x i16> %i, %j 64 %r = select <32 x i1> %mask, <32 x i16> %x, <32 x i16> %i 65 ret <32 x i16> %r 66} 67 68; CHECK-LABEL: vpaddw512_maskz_fold_test 69; CHECK: vpaddw (%rdi), %zmm{{.*{%k[1-7]} {z}}} 70; CHECK: ret 71define <32 x i16> @vpaddw512_maskz_fold_test(<32 x i16> %i, <32 x i16>* %j.ptr, <32 x i16> %mask1) nounwind readnone { 72 %mask = icmp ne <32 x i16> %mask1, zeroinitializer 73 %j = load <32 x i16>, <32 x i16>* %j.ptr 74 %x = add <32 x i16> %i, %j 75 %r = select <32 x i1> %mask, <32 x i16> %x, <32 x i16> zeroinitializer 76 ret <32 x i16> %r 77} 78 79; CHECK-LABEL: vpsubb512_test 80; CHECK: vpsubb %zmm{{.*}} 81; CHECK: ret 82define <64 x i8> @vpsubb512_test(<64 x i8> %i, <64 x i8> %j) nounwind readnone { 83 %x = sub <64 x i8> %i, %j 84 ret <64 x i8> %x 85} 86 87; CHECK-LABEL: vpsubw512_test 88; CHECK: vpsubw %zmm{{.*}} 89; CHECK: ret 90define <32 x i16> @vpsubw512_test(<32 x i16> %i, <32 x i16> %j) nounwind readnone { 91 %x = sub <32 x i16> %i, %j 92 ret <32 x i16> %x 93} 94 95; CHECK-LABEL: vpmullw512_test 96; CHECK: vpmullw %zmm{{.*}} 97; CHECK: ret 98define <32 x i16> @vpmullw512_test(<32 x i16> %i, <32 x i16> %j) { 99 %x = mul <32 x i16> %i, %j 100 ret <32 x i16> %x 101} 102 103