blob: a333e1520e87b6fe6b739cd431a729d5070e6e04 [file] [log] [blame]
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefixes=AVX512,AVX512F
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vl | FileCheck %s --check-prefixes=AVX512,AVX512VL
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefixes=AVX512,AVX512BW
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vbmi2 | FileCheck %s --check-prefixes=AVX512,AVX512VBMI2
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw,+avx512vl | FileCheck %s --check-prefixes=AVX512,AVX512VLBW
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vbmi2,+avx512vl | FileCheck %s --check-prefixes=AVX512,AVX512VLVBMI2
declare <8 x i64> @llvm.fshr.v8i64(<8 x i64>, <8 x i64>, <8 x i64>)
declare <16 x i32> @llvm.fshr.v16i32(<16 x i32>, <16 x i32>, <16 x i32>)
declare <32 x i16> @llvm.fshr.v32i16(<32 x i16>, <32 x i16>, <32 x i16>)
declare <64 x i8> @llvm.fshr.v64i8(<64 x i8>, <64 x i8>, <64 x i8>)
;
; Variable Shifts
;
define <8 x i64> @var_funnnel_v8i64(<8 x i64> %x, <8 x i64> %y, <8 x i64> %amt) nounwind {
; AVX512F-LABEL: var_funnnel_v8i64:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpbroadcastq {{.*#+}} zmm3 = [63,63,63,63,63,63,63,63]
; AVX512F-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512F-NEXT: vpsrlvq %zmm4, %zmm1, %zmm5
; AVX512F-NEXT: vpbroadcastq {{.*#+}} zmm6 = [64,64,64,64,64,64,64,64]
; AVX512F-NEXT: vpsubq %zmm4, %zmm6, %zmm4
; AVX512F-NEXT: vpsllvq %zmm4, %zmm0, %zmm0
; AVX512F-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512F-NEXT: vptestnmq %zmm3, %zmm2, %k1
; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: var_funnnel_v8i64:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpbroadcastq {{.*#+}} zmm3 = [63,63,63,63,63,63,63,63]
; AVX512VL-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512VL-NEXT: vpsrlvq %zmm4, %zmm1, %zmm5
; AVX512VL-NEXT: vpbroadcastq {{.*#+}} zmm6 = [64,64,64,64,64,64,64,64]
; AVX512VL-NEXT: vpsubq %zmm4, %zmm6, %zmm4
; AVX512VL-NEXT: vpsllvq %zmm4, %zmm0, %zmm0
; AVX512VL-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512VL-NEXT: vptestnmq %zmm3, %zmm2, %k1
; AVX512VL-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: var_funnnel_v8i64:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpbroadcastq {{.*#+}} zmm3 = [63,63,63,63,63,63,63,63]
; AVX512BW-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512BW-NEXT: vpsrlvq %zmm4, %zmm1, %zmm5
; AVX512BW-NEXT: vpbroadcastq {{.*#+}} zmm6 = [64,64,64,64,64,64,64,64]
; AVX512BW-NEXT: vpsubq %zmm4, %zmm6, %zmm4
; AVX512BW-NEXT: vpsllvq %zmm4, %zmm0, %zmm0
; AVX512BW-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512BW-NEXT: vptestnmq %zmm3, %zmm2, %k1
; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512BW-NEXT: retq
;
; AVX512VBMI2-LABEL: var_funnnel_v8i64:
; AVX512VBMI2: # %bb.0:
; AVX512VBMI2-NEXT: vpshrdvq %zmm2, %zmm0, %zmm1
; AVX512VBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VBMI2-NEXT: retq
;
; AVX512VLBW-LABEL: var_funnnel_v8i64:
; AVX512VLBW: # %bb.0:
; AVX512VLBW-NEXT: vpbroadcastq {{.*#+}} zmm3 = [63,63,63,63,63,63,63,63]
; AVX512VLBW-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512VLBW-NEXT: vpsrlvq %zmm4, %zmm1, %zmm5
; AVX512VLBW-NEXT: vpbroadcastq {{.*#+}} zmm6 = [64,64,64,64,64,64,64,64]
; AVX512VLBW-NEXT: vpsubq %zmm4, %zmm6, %zmm4
; AVX512VLBW-NEXT: vpsllvq %zmm4, %zmm0, %zmm0
; AVX512VLBW-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512VLBW-NEXT: vptestnmq %zmm3, %zmm2, %k1
; AVX512VLBW-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512VLBW-NEXT: retq
;
; AVX512VLVBMI2-LABEL: var_funnnel_v8i64:
; AVX512VLVBMI2: # %bb.0:
; AVX512VLVBMI2-NEXT: vpshrdvq %zmm2, %zmm0, %zmm1
; AVX512VLVBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VLVBMI2-NEXT: retq
%res = call <8 x i64> @llvm.fshr.v8i64(<8 x i64> %x, <8 x i64> %y, <8 x i64> %amt)
ret <8 x i64> %res
}
define <16 x i32> @var_funnnel_v16i32(<16 x i32> %x, <16 x i32> %y, <16 x i32> %amt) nounwind {
; AVX512F-LABEL: var_funnnel_v16i32:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpbroadcastd {{.*#+}} zmm3 = [31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31]
; AVX512F-NEXT: vpandd %zmm3, %zmm2, %zmm4
; AVX512F-NEXT: vpsrlvd %zmm4, %zmm1, %zmm5
; AVX512F-NEXT: vpbroadcastd {{.*#+}} zmm6 = [32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32]
; AVX512F-NEXT: vpsubd %zmm4, %zmm6, %zmm4
; AVX512F-NEXT: vpsllvd %zmm4, %zmm0, %zmm0
; AVX512F-NEXT: vpord %zmm5, %zmm0, %zmm0
; AVX512F-NEXT: vptestnmd %zmm3, %zmm2, %k1
; AVX512F-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: var_funnnel_v16i32:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpbroadcastd {{.*#+}} zmm3 = [31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31]
; AVX512VL-NEXT: vpandd %zmm3, %zmm2, %zmm4
; AVX512VL-NEXT: vpsrlvd %zmm4, %zmm1, %zmm5
; AVX512VL-NEXT: vpbroadcastd {{.*#+}} zmm6 = [32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32]
; AVX512VL-NEXT: vpsubd %zmm4, %zmm6, %zmm4
; AVX512VL-NEXT: vpsllvd %zmm4, %zmm0, %zmm0
; AVX512VL-NEXT: vpord %zmm5, %zmm0, %zmm0
; AVX512VL-NEXT: vptestnmd %zmm3, %zmm2, %k1
; AVX512VL-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: var_funnnel_v16i32:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpbroadcastd {{.*#+}} zmm3 = [31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31]
; AVX512BW-NEXT: vpandd %zmm3, %zmm2, %zmm4
; AVX512BW-NEXT: vpsrlvd %zmm4, %zmm1, %zmm5
; AVX512BW-NEXT: vpbroadcastd {{.*#+}} zmm6 = [32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32]
; AVX512BW-NEXT: vpsubd %zmm4, %zmm6, %zmm4
; AVX512BW-NEXT: vpsllvd %zmm4, %zmm0, %zmm0
; AVX512BW-NEXT: vpord %zmm5, %zmm0, %zmm0
; AVX512BW-NEXT: vptestnmd %zmm3, %zmm2, %k1
; AVX512BW-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
; AVX512BW-NEXT: retq
;
; AVX512VBMI2-LABEL: var_funnnel_v16i32:
; AVX512VBMI2: # %bb.0:
; AVX512VBMI2-NEXT: vpshrdvd %zmm2, %zmm0, %zmm1
; AVX512VBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VBMI2-NEXT: retq
;
; AVX512VLBW-LABEL: var_funnnel_v16i32:
; AVX512VLBW: # %bb.0:
; AVX512VLBW-NEXT: vpbroadcastd {{.*#+}} zmm3 = [31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31]
; AVX512VLBW-NEXT: vpandd %zmm3, %zmm2, %zmm4
; AVX512VLBW-NEXT: vpsrlvd %zmm4, %zmm1, %zmm5
; AVX512VLBW-NEXT: vpbroadcastd {{.*#+}} zmm6 = [32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32]
; AVX512VLBW-NEXT: vpsubd %zmm4, %zmm6, %zmm4
; AVX512VLBW-NEXT: vpsllvd %zmm4, %zmm0, %zmm0
; AVX512VLBW-NEXT: vpord %zmm5, %zmm0, %zmm0
; AVX512VLBW-NEXT: vptestnmd %zmm3, %zmm2, %k1
; AVX512VLBW-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
; AVX512VLBW-NEXT: retq
;
; AVX512VLVBMI2-LABEL: var_funnnel_v16i32:
; AVX512VLVBMI2: # %bb.0:
; AVX512VLVBMI2-NEXT: vpshrdvd %zmm2, %zmm0, %zmm1
; AVX512VLVBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VLVBMI2-NEXT: retq
%res = call <16 x i32> @llvm.fshr.v16i32(<16 x i32> %x, <16 x i32> %y, <16 x i32> %amt)
ret <16 x i32> %res
}
define <32 x i16> @var_funnnel_v32i16(<32 x i16> %x, <32 x i16> %y, <32 x i16> %amt) nounwind {
; AVX512F-LABEL: var_funnnel_v32i16:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm6 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX512F-NEXT: vpand %ymm6, %ymm4, %ymm4
; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm7 = ymm4[0],zero,ymm4[1],zero,ymm4[2],zero,ymm4[3],zero,ymm4[4],zero,ymm4[5],zero,ymm4[6],zero,ymm4[7],zero,ymm4[8],zero,ymm4[9],zero,ymm4[10],zero,ymm4[11],zero,ymm4[12],zero,ymm4[13],zero,ymm4[14],zero,ymm4[15],zero
; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm8 = ymm2[0],zero,ymm2[1],zero,ymm2[2],zero,ymm2[3],zero,ymm2[4],zero,ymm2[5],zero,ymm2[6],zero,ymm2[7],zero,ymm2[8],zero,ymm2[9],zero,ymm2[10],zero,ymm2[11],zero,ymm2[12],zero,ymm2[13],zero,ymm2[14],zero,ymm2[15],zero
; AVX512F-NEXT: vpsrlvd %zmm7, %zmm8, %zmm7
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm8 = [16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16]
; AVX512F-NEXT: vpsubw %ymm4, %ymm8, %ymm9
; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm9 = ymm9[0],zero,ymm9[1],zero,ymm9[2],zero,ymm9[3],zero,ymm9[4],zero,ymm9[5],zero,ymm9[6],zero,ymm9[7],zero,ymm9[8],zero,ymm9[9],zero,ymm9[10],zero,ymm9[11],zero,ymm9[12],zero,ymm9[13],zero,ymm9[14],zero,ymm9[15],zero
; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
; AVX512F-NEXT: vpsllvd %zmm9, %zmm0, %zmm0
; AVX512F-NEXT: vpord %zmm7, %zmm0, %zmm0
; AVX512F-NEXT: vpmovdw %zmm0, %ymm0
; AVX512F-NEXT: vpxor %xmm7, %xmm7, %xmm7
; AVX512F-NEXT: vpcmpeqw %ymm7, %ymm4, %ymm4
; AVX512F-NEXT: vpblendvb %ymm4, %ymm2, %ymm0, %ymm0
; AVX512F-NEXT: vpand %ymm6, %ymm5, %ymm2
; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm4 = ymm2[0],zero,ymm2[1],zero,ymm2[2],zero,ymm2[3],zero,ymm2[4],zero,ymm2[5],zero,ymm2[6],zero,ymm2[7],zero,ymm2[8],zero,ymm2[9],zero,ymm2[10],zero,ymm2[11],zero,ymm2[12],zero,ymm2[13],zero,ymm2[14],zero,ymm2[15],zero
; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm5 = ymm3[0],zero,ymm3[1],zero,ymm3[2],zero,ymm3[3],zero,ymm3[4],zero,ymm3[5],zero,ymm3[6],zero,ymm3[7],zero,ymm3[8],zero,ymm3[9],zero,ymm3[10],zero,ymm3[11],zero,ymm3[12],zero,ymm3[13],zero,ymm3[14],zero,ymm3[15],zero
; AVX512F-NEXT: vpsrlvd %zmm4, %zmm5, %zmm4
; AVX512F-NEXT: vpsubw %ymm2, %ymm8, %ymm5
; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm5 = ymm5[0],zero,ymm5[1],zero,ymm5[2],zero,ymm5[3],zero,ymm5[4],zero,ymm5[5],zero,ymm5[6],zero,ymm5[7],zero,ymm5[8],zero,ymm5[9],zero,ymm5[10],zero,ymm5[11],zero,ymm5[12],zero,ymm5[13],zero,ymm5[14],zero,ymm5[15],zero
; AVX512F-NEXT: vpmovzxwd {{.*#+}} zmm1 = ymm1[0],zero,ymm1[1],zero,ymm1[2],zero,ymm1[3],zero,ymm1[4],zero,ymm1[5],zero,ymm1[6],zero,ymm1[7],zero,ymm1[8],zero,ymm1[9],zero,ymm1[10],zero,ymm1[11],zero,ymm1[12],zero,ymm1[13],zero,ymm1[14],zero,ymm1[15],zero
; AVX512F-NEXT: vpsllvd %zmm5, %zmm1, %zmm1
; AVX512F-NEXT: vpord %zmm4, %zmm1, %zmm1
; AVX512F-NEXT: vpmovdw %zmm1, %ymm1
; AVX512F-NEXT: vpcmpeqw %ymm7, %ymm2, %ymm2
; AVX512F-NEXT: vpblendvb %ymm2, %ymm3, %ymm1, %ymm1
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: var_funnnel_v32i16:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm6 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX512VL-NEXT: vpand %ymm6, %ymm4, %ymm4
; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm7 = ymm4[0],zero,ymm4[1],zero,ymm4[2],zero,ymm4[3],zero,ymm4[4],zero,ymm4[5],zero,ymm4[6],zero,ymm4[7],zero,ymm4[8],zero,ymm4[9],zero,ymm4[10],zero,ymm4[11],zero,ymm4[12],zero,ymm4[13],zero,ymm4[14],zero,ymm4[15],zero
; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm8 = ymm2[0],zero,ymm2[1],zero,ymm2[2],zero,ymm2[3],zero,ymm2[4],zero,ymm2[5],zero,ymm2[6],zero,ymm2[7],zero,ymm2[8],zero,ymm2[9],zero,ymm2[10],zero,ymm2[11],zero,ymm2[12],zero,ymm2[13],zero,ymm2[14],zero,ymm2[15],zero
; AVX512VL-NEXT: vpsrlvd %zmm7, %zmm8, %zmm7
; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm8 = [16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16]
; AVX512VL-NEXT: vpsubw %ymm4, %ymm8, %ymm9
; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm9 = ymm9[0],zero,ymm9[1],zero,ymm9[2],zero,ymm9[3],zero,ymm9[4],zero,ymm9[5],zero,ymm9[6],zero,ymm9[7],zero,ymm9[8],zero,ymm9[9],zero,ymm9[10],zero,ymm9[11],zero,ymm9[12],zero,ymm9[13],zero,ymm9[14],zero,ymm9[15],zero
; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
; AVX512VL-NEXT: vpsllvd %zmm9, %zmm0, %zmm0
; AVX512VL-NEXT: vpord %zmm7, %zmm0, %zmm0
; AVX512VL-NEXT: vpmovdw %zmm0, %ymm0
; AVX512VL-NEXT: vpxor %xmm7, %xmm7, %xmm7
; AVX512VL-NEXT: vpcmpeqw %ymm7, %ymm4, %ymm4
; AVX512VL-NEXT: vpblendvb %ymm4, %ymm2, %ymm0, %ymm0
; AVX512VL-NEXT: vpand %ymm6, %ymm5, %ymm2
; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm4 = ymm2[0],zero,ymm2[1],zero,ymm2[2],zero,ymm2[3],zero,ymm2[4],zero,ymm2[5],zero,ymm2[6],zero,ymm2[7],zero,ymm2[8],zero,ymm2[9],zero,ymm2[10],zero,ymm2[11],zero,ymm2[12],zero,ymm2[13],zero,ymm2[14],zero,ymm2[15],zero
; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm5 = ymm3[0],zero,ymm3[1],zero,ymm3[2],zero,ymm3[3],zero,ymm3[4],zero,ymm3[5],zero,ymm3[6],zero,ymm3[7],zero,ymm3[8],zero,ymm3[9],zero,ymm3[10],zero,ymm3[11],zero,ymm3[12],zero,ymm3[13],zero,ymm3[14],zero,ymm3[15],zero
; AVX512VL-NEXT: vpsrlvd %zmm4, %zmm5, %zmm4
; AVX512VL-NEXT: vpsubw %ymm2, %ymm8, %ymm5
; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm5 = ymm5[0],zero,ymm5[1],zero,ymm5[2],zero,ymm5[3],zero,ymm5[4],zero,ymm5[5],zero,ymm5[6],zero,ymm5[7],zero,ymm5[8],zero,ymm5[9],zero,ymm5[10],zero,ymm5[11],zero,ymm5[12],zero,ymm5[13],zero,ymm5[14],zero,ymm5[15],zero
; AVX512VL-NEXT: vpmovzxwd {{.*#+}} zmm1 = ymm1[0],zero,ymm1[1],zero,ymm1[2],zero,ymm1[3],zero,ymm1[4],zero,ymm1[5],zero,ymm1[6],zero,ymm1[7],zero,ymm1[8],zero,ymm1[9],zero,ymm1[10],zero,ymm1[11],zero,ymm1[12],zero,ymm1[13],zero,ymm1[14],zero,ymm1[15],zero
; AVX512VL-NEXT: vpsllvd %zmm5, %zmm1, %zmm1
; AVX512VL-NEXT: vpord %zmm4, %zmm1, %zmm1
; AVX512VL-NEXT: vpmovdw %zmm1, %ymm1
; AVX512VL-NEXT: vpcmpeqw %ymm7, %ymm2, %ymm2
; AVX512VL-NEXT: vpblendvb %ymm2, %ymm3, %ymm1, %ymm1
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: var_funnnel_v32i16:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX512BW-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512BW-NEXT: vpsrlvw %zmm4, %zmm1, %zmm5
; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm6 = [16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16]
; AVX512BW-NEXT: vpsubw %zmm4, %zmm6, %zmm4
; AVX512BW-NEXT: vpsllvw %zmm4, %zmm0, %zmm0
; AVX512BW-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512BW-NEXT: vptestnmw %zmm3, %zmm2, %k1
; AVX512BW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k1}
; AVX512BW-NEXT: retq
;
; AVX512VBMI2-LABEL: var_funnnel_v32i16:
; AVX512VBMI2: # %bb.0:
; AVX512VBMI2-NEXT: vpshrdvw %zmm2, %zmm0, %zmm1
; AVX512VBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VBMI2-NEXT: retq
;
; AVX512VLBW-LABEL: var_funnnel_v32i16:
; AVX512VLBW: # %bb.0:
; AVX512VLBW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX512VLBW-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512VLBW-NEXT: vpsrlvw %zmm4, %zmm1, %zmm5
; AVX512VLBW-NEXT: vmovdqa64 {{.*#+}} zmm6 = [16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16]
; AVX512VLBW-NEXT: vpsubw %zmm4, %zmm6, %zmm4
; AVX512VLBW-NEXT: vpsllvw %zmm4, %zmm0, %zmm0
; AVX512VLBW-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512VLBW-NEXT: vptestnmw %zmm3, %zmm2, %k1
; AVX512VLBW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k1}
; AVX512VLBW-NEXT: retq
;
; AVX512VLVBMI2-LABEL: var_funnnel_v32i16:
; AVX512VLVBMI2: # %bb.0:
; AVX512VLVBMI2-NEXT: vpshrdvw %zmm2, %zmm0, %zmm1
; AVX512VLVBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VLVBMI2-NEXT: retq
%res = call <32 x i16> @llvm.fshr.v32i16(<32 x i16> %x, <32 x i16> %y, <32 x i16> %amt)
ret <32 x i16> %res
}
define <64 x i8> @var_funnnel_v64i8(<64 x i8> %x, <64 x i8> %y, <64 x i8> %amt) nounwind {
; AVX512F-LABEL: var_funnnel_v64i8:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpsrlw $4, %ymm2, %ymm7
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm6 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX512F-NEXT: vpand %ymm6, %ymm7, %ymm8
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm7 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7]
; AVX512F-NEXT: vpand %ymm7, %ymm4, %ymm9
; AVX512F-NEXT: vpsllw $5, %ymm9, %ymm10
; AVX512F-NEXT: vpblendvb %ymm10, %ymm8, %ymm2, %ymm8
; AVX512F-NEXT: vpsrlw $2, %ymm8, %ymm11
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm4 = [63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63]
; AVX512F-NEXT: vpand %ymm4, %ymm11, %ymm11
; AVX512F-NEXT: vpaddb %ymm10, %ymm10, %ymm10
; AVX512F-NEXT: vpblendvb %ymm10, %ymm11, %ymm8, %ymm11
; AVX512F-NEXT: vpsrlw $1, %ymm11, %ymm12
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm8 = [127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127]
; AVX512F-NEXT: vpand %ymm8, %ymm12, %ymm12
; AVX512F-NEXT: vpaddb %ymm10, %ymm10, %ymm10
; AVX512F-NEXT: vpblendvb %ymm10, %ymm12, %ymm11, %ymm10
; AVX512F-NEXT: vpsllw $4, %ymm0, %ymm11
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm12 = [240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240]
; AVX512F-NEXT: vpand %ymm12, %ymm11, %ymm11
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm13 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8]
; AVX512F-NEXT: vpsubb %ymm9, %ymm13, %ymm14
; AVX512F-NEXT: vpsllw $5, %ymm14, %ymm14
; AVX512F-NEXT: vpblendvb %ymm14, %ymm11, %ymm0, %ymm0
; AVX512F-NEXT: vpsllw $2, %ymm0, %ymm11
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm15 = [252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252]
; AVX512F-NEXT: vpand %ymm15, %ymm11, %ymm11
; AVX512F-NEXT: vpaddb %ymm14, %ymm14, %ymm14
; AVX512F-NEXT: vpblendvb %ymm14, %ymm11, %ymm0, %ymm0
; AVX512F-NEXT: vpaddb %ymm0, %ymm0, %ymm11
; AVX512F-NEXT: vpaddb %ymm14, %ymm14, %ymm14
; AVX512F-NEXT: vpblendvb %ymm14, %ymm11, %ymm0, %ymm0
; AVX512F-NEXT: vpor %ymm10, %ymm0, %ymm0
; AVX512F-NEXT: vpxor %xmm10, %xmm10, %xmm10
; AVX512F-NEXT: vpcmpeqb %ymm10, %ymm9, %ymm9
; AVX512F-NEXT: vpblendvb %ymm9, %ymm2, %ymm0, %ymm0
; AVX512F-NEXT: vpsrlw $4, %ymm3, %ymm2
; AVX512F-NEXT: vpand %ymm6, %ymm2, %ymm2
; AVX512F-NEXT: vpand %ymm7, %ymm5, %ymm5
; AVX512F-NEXT: vpsllw $5, %ymm5, %ymm6
; AVX512F-NEXT: vpblendvb %ymm6, %ymm2, %ymm3, %ymm2
; AVX512F-NEXT: vpsrlw $2, %ymm2, %ymm7
; AVX512F-NEXT: vpand %ymm4, %ymm7, %ymm4
; AVX512F-NEXT: vpaddb %ymm6, %ymm6, %ymm6
; AVX512F-NEXT: vpblendvb %ymm6, %ymm4, %ymm2, %ymm2
; AVX512F-NEXT: vpsrlw $1, %ymm2, %ymm4
; AVX512F-NEXT: vpand %ymm8, %ymm4, %ymm4
; AVX512F-NEXT: vpaddb %ymm6, %ymm6, %ymm6
; AVX512F-NEXT: vpblendvb %ymm6, %ymm4, %ymm2, %ymm2
; AVX512F-NEXT: vpsllw $4, %ymm1, %ymm4
; AVX512F-NEXT: vpand %ymm12, %ymm4, %ymm4
; AVX512F-NEXT: vpsubb %ymm5, %ymm13, %ymm6
; AVX512F-NEXT: vpsllw $5, %ymm6, %ymm6
; AVX512F-NEXT: vpblendvb %ymm6, %ymm4, %ymm1, %ymm1
; AVX512F-NEXT: vpsllw $2, %ymm1, %ymm4
; AVX512F-NEXT: vpand %ymm15, %ymm4, %ymm4
; AVX512F-NEXT: vpaddb %ymm6, %ymm6, %ymm6
; AVX512F-NEXT: vpblendvb %ymm6, %ymm4, %ymm1, %ymm1
; AVX512F-NEXT: vpaddb %ymm1, %ymm1, %ymm4
; AVX512F-NEXT: vpaddb %ymm6, %ymm6, %ymm6
; AVX512F-NEXT: vpblendvb %ymm6, %ymm4, %ymm1, %ymm1
; AVX512F-NEXT: vpor %ymm2, %ymm1, %ymm1
; AVX512F-NEXT: vpcmpeqb %ymm10, %ymm5, %ymm2
; AVX512F-NEXT: vpblendvb %ymm2, %ymm3, %ymm1, %ymm1
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: var_funnnel_v64i8:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpsrlw $4, %ymm2, %ymm6
; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm7 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX512VL-NEXT: vpand %ymm7, %ymm6, %ymm6
; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm8 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7]
; AVX512VL-NEXT: vpand %ymm8, %ymm4, %ymm4
; AVX512VL-NEXT: vpsllw $5, %ymm4, %ymm9
; AVX512VL-NEXT: vpblendvb %ymm9, %ymm6, %ymm2, %ymm6
; AVX512VL-NEXT: vpsrlw $2, %ymm6, %ymm10
; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm11 = [63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63,63]
; AVX512VL-NEXT: vpand %ymm11, %ymm10, %ymm10
; AVX512VL-NEXT: vpaddb %ymm9, %ymm9, %ymm9
; AVX512VL-NEXT: vpblendvb %ymm9, %ymm10, %ymm6, %ymm6
; AVX512VL-NEXT: vpsrlw $1, %ymm6, %ymm10
; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm12 = [127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127]
; AVX512VL-NEXT: vpand %ymm12, %ymm10, %ymm10
; AVX512VL-NEXT: vpaddb %ymm9, %ymm9, %ymm9
; AVX512VL-NEXT: vpblendvb %ymm9, %ymm10, %ymm6, %ymm6
; AVX512VL-NEXT: vpsllw $4, %ymm0, %ymm9
; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm10 = [240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240]
; AVX512VL-NEXT: vpand %ymm10, %ymm9, %ymm9
; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm13 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8]
; AVX512VL-NEXT: vpsubb %ymm4, %ymm13, %ymm14
; AVX512VL-NEXT: vpsllw $5, %ymm14, %ymm14
; AVX512VL-NEXT: vpblendvb %ymm14, %ymm9, %ymm0, %ymm0
; AVX512VL-NEXT: vpsllw $2, %ymm0, %ymm9
; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm15 = [252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252]
; AVX512VL-NEXT: vpand %ymm15, %ymm9, %ymm9
; AVX512VL-NEXT: vpaddb %ymm14, %ymm14, %ymm14
; AVX512VL-NEXT: vpblendvb %ymm14, %ymm9, %ymm0, %ymm0
; AVX512VL-NEXT: vpaddb %ymm0, %ymm0, %ymm9
; AVX512VL-NEXT: vpaddb %ymm14, %ymm14, %ymm14
; AVX512VL-NEXT: vpblendvb %ymm14, %ymm9, %ymm0, %ymm0
; AVX512VL-NEXT: vpor %ymm6, %ymm0, %ymm0
; AVX512VL-NEXT: vpxor %xmm6, %xmm6, %xmm6
; AVX512VL-NEXT: vpcmpeqb %ymm6, %ymm4, %ymm4
; AVX512VL-NEXT: vpblendvb %ymm4, %ymm2, %ymm0, %ymm0
; AVX512VL-NEXT: vpsrlw $4, %ymm3, %ymm2
; AVX512VL-NEXT: vpand %ymm7, %ymm2, %ymm2
; AVX512VL-NEXT: vpand %ymm8, %ymm5, %ymm4
; AVX512VL-NEXT: vpsllw $5, %ymm4, %ymm5
; AVX512VL-NEXT: vpblendvb %ymm5, %ymm2, %ymm3, %ymm2
; AVX512VL-NEXT: vpsrlw $2, %ymm2, %ymm7
; AVX512VL-NEXT: vpand %ymm11, %ymm7, %ymm7
; AVX512VL-NEXT: vpaddb %ymm5, %ymm5, %ymm5
; AVX512VL-NEXT: vpblendvb %ymm5, %ymm7, %ymm2, %ymm2
; AVX512VL-NEXT: vpsrlw $1, %ymm2, %ymm7
; AVX512VL-NEXT: vpand %ymm12, %ymm7, %ymm7
; AVX512VL-NEXT: vpaddb %ymm5, %ymm5, %ymm5
; AVX512VL-NEXT: vpblendvb %ymm5, %ymm7, %ymm2, %ymm2
; AVX512VL-NEXT: vpsllw $4, %ymm1, %ymm5
; AVX512VL-NEXT: vpand %ymm10, %ymm5, %ymm5
; AVX512VL-NEXT: vpsubb %ymm4, %ymm13, %ymm7
; AVX512VL-NEXT: vpsllw $5, %ymm7, %ymm7
; AVX512VL-NEXT: vpblendvb %ymm7, %ymm5, %ymm1, %ymm1
; AVX512VL-NEXT: vpsllw $2, %ymm1, %ymm5
; AVX512VL-NEXT: vpand %ymm15, %ymm5, %ymm5
; AVX512VL-NEXT: vpaddb %ymm7, %ymm7, %ymm7
; AVX512VL-NEXT: vpblendvb %ymm7, %ymm5, %ymm1, %ymm1
; AVX512VL-NEXT: vpaddb %ymm1, %ymm1, %ymm5
; AVX512VL-NEXT: vpaddb %ymm7, %ymm7, %ymm7
; AVX512VL-NEXT: vpblendvb %ymm7, %ymm5, %ymm1, %ymm1
; AVX512VL-NEXT: vpor %ymm2, %ymm1, %ymm1
; AVX512VL-NEXT: vpcmpeqb %ymm6, %ymm4, %ymm2
; AVX512VL-NEXT: vpblendvb %ymm2, %ymm3, %ymm1, %ymm1
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: var_funnnel_v64i8:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7]
; AVX512BW-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512BW-NEXT: vpsllw $5, %zmm4, %zmm5
; AVX512BW-NEXT: vpaddb %zmm5, %zmm5, %zmm6
; AVX512BW-NEXT: vpmovb2m %zmm6, %k1
; AVX512BW-NEXT: vpmovb2m %zmm5, %k2
; AVX512BW-NEXT: vpsrlw $4, %zmm1, %zmm5
; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm5, %zmm5
; AVX512BW-NEXT: vpblendmb %zmm5, %zmm1, %zmm5 {%k2}
; AVX512BW-NEXT: vpsrlw $2, %zmm5, %zmm7
; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm7, %zmm7
; AVX512BW-NEXT: vmovdqu8 %zmm7, %zmm5 {%k1}
; AVX512BW-NEXT: vpsrlw $1, %zmm5, %zmm7
; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm7, %zmm7
; AVX512BW-NEXT: vpaddb %zmm6, %zmm6, %zmm6
; AVX512BW-NEXT: vpmovb2m %zmm6, %k1
; AVX512BW-NEXT: vmovdqu8 %zmm7, %zmm5 {%k1}
; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm6 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8]
; AVX512BW-NEXT: vpsubb %zmm4, %zmm6, %zmm4
; AVX512BW-NEXT: vpsllw $5, %zmm4, %zmm4
; AVX512BW-NEXT: vpaddb %zmm4, %zmm4, %zmm6
; AVX512BW-NEXT: vpmovb2m %zmm6, %k1
; AVX512BW-NEXT: vpmovb2m %zmm4, %k2
; AVX512BW-NEXT: vpsllw $4, %zmm0, %zmm4
; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm4, %zmm4
; AVX512BW-NEXT: vmovdqu8 %zmm4, %zmm0 {%k2}
; AVX512BW-NEXT: vpsllw $2, %zmm0, %zmm4
; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm4, %zmm4
; AVX512BW-NEXT: vmovdqu8 %zmm4, %zmm0 {%k1}
; AVX512BW-NEXT: vpaddb %zmm6, %zmm6, %zmm4
; AVX512BW-NEXT: vpmovb2m %zmm4, %k1
; AVX512BW-NEXT: vpaddb %zmm0, %zmm0, %zmm0 {%k1}
; AVX512BW-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512BW-NEXT: vptestnmb %zmm3, %zmm2, %k1
; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
; AVX512BW-NEXT: retq
;
; AVX512VBMI2-LABEL: var_funnnel_v64i8:
; AVX512VBMI2: # %bb.0:
; AVX512VBMI2-NEXT: vmovdqa64 {{.*#+}} zmm3 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7]
; AVX512VBMI2-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512VBMI2-NEXT: vpsllw $5, %zmm4, %zmm5
; AVX512VBMI2-NEXT: vpaddb %zmm5, %zmm5, %zmm6
; AVX512VBMI2-NEXT: vpmovb2m %zmm6, %k1
; AVX512VBMI2-NEXT: vpmovb2m %zmm5, %k2
; AVX512VBMI2-NEXT: vpsrlw $4, %zmm1, %zmm5
; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm5, %zmm5
; AVX512VBMI2-NEXT: vpblendmb %zmm5, %zmm1, %zmm5 {%k2}
; AVX512VBMI2-NEXT: vpsrlw $2, %zmm5, %zmm7
; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm7, %zmm7
; AVX512VBMI2-NEXT: vmovdqu8 %zmm7, %zmm5 {%k1}
; AVX512VBMI2-NEXT: vpsrlw $1, %zmm5, %zmm7
; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm7, %zmm7
; AVX512VBMI2-NEXT: vpaddb %zmm6, %zmm6, %zmm6
; AVX512VBMI2-NEXT: vpmovb2m %zmm6, %k1
; AVX512VBMI2-NEXT: vmovdqu8 %zmm7, %zmm5 {%k1}
; AVX512VBMI2-NEXT: vmovdqa64 {{.*#+}} zmm6 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8]
; AVX512VBMI2-NEXT: vpsubb %zmm4, %zmm6, %zmm4
; AVX512VBMI2-NEXT: vpsllw $5, %zmm4, %zmm4
; AVX512VBMI2-NEXT: vpaddb %zmm4, %zmm4, %zmm6
; AVX512VBMI2-NEXT: vpmovb2m %zmm6, %k1
; AVX512VBMI2-NEXT: vpmovb2m %zmm4, %k2
; AVX512VBMI2-NEXT: vpsllw $4, %zmm0, %zmm4
; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm4, %zmm4
; AVX512VBMI2-NEXT: vmovdqu8 %zmm4, %zmm0 {%k2}
; AVX512VBMI2-NEXT: vpsllw $2, %zmm0, %zmm4
; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm4, %zmm4
; AVX512VBMI2-NEXT: vmovdqu8 %zmm4, %zmm0 {%k1}
; AVX512VBMI2-NEXT: vpaddb %zmm6, %zmm6, %zmm4
; AVX512VBMI2-NEXT: vpmovb2m %zmm4, %k1
; AVX512VBMI2-NEXT: vpaddb %zmm0, %zmm0, %zmm0 {%k1}
; AVX512VBMI2-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512VBMI2-NEXT: vptestnmb %zmm3, %zmm2, %k1
; AVX512VBMI2-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
; AVX512VBMI2-NEXT: retq
;
; AVX512VLBW-LABEL: var_funnnel_v64i8:
; AVX512VLBW: # %bb.0:
; AVX512VLBW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7]
; AVX512VLBW-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512VLBW-NEXT: vpsllw $5, %zmm4, %zmm5
; AVX512VLBW-NEXT: vpaddb %zmm5, %zmm5, %zmm6
; AVX512VLBW-NEXT: vpmovb2m %zmm6, %k1
; AVX512VLBW-NEXT: vpmovb2m %zmm5, %k2
; AVX512VLBW-NEXT: vpsrlw $4, %zmm1, %zmm5
; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm5, %zmm5
; AVX512VLBW-NEXT: vpblendmb %zmm5, %zmm1, %zmm5 {%k2}
; AVX512VLBW-NEXT: vpsrlw $2, %zmm5, %zmm7
; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm7, %zmm7
; AVX512VLBW-NEXT: vmovdqu8 %zmm7, %zmm5 {%k1}
; AVX512VLBW-NEXT: vpsrlw $1, %zmm5, %zmm7
; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm7, %zmm7
; AVX512VLBW-NEXT: vpaddb %zmm6, %zmm6, %zmm6
; AVX512VLBW-NEXT: vpmovb2m %zmm6, %k1
; AVX512VLBW-NEXT: vmovdqu8 %zmm7, %zmm5 {%k1}
; AVX512VLBW-NEXT: vmovdqa64 {{.*#+}} zmm6 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8]
; AVX512VLBW-NEXT: vpsubb %zmm4, %zmm6, %zmm4
; AVX512VLBW-NEXT: vpsllw $5, %zmm4, %zmm4
; AVX512VLBW-NEXT: vpaddb %zmm4, %zmm4, %zmm6
; AVX512VLBW-NEXT: vpmovb2m %zmm6, %k1
; AVX512VLBW-NEXT: vpmovb2m %zmm4, %k2
; AVX512VLBW-NEXT: vpsllw $4, %zmm0, %zmm4
; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm4, %zmm4
; AVX512VLBW-NEXT: vmovdqu8 %zmm4, %zmm0 {%k2}
; AVX512VLBW-NEXT: vpsllw $2, %zmm0, %zmm4
; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm4, %zmm4
; AVX512VLBW-NEXT: vmovdqu8 %zmm4, %zmm0 {%k1}
; AVX512VLBW-NEXT: vpaddb %zmm6, %zmm6, %zmm4
; AVX512VLBW-NEXT: vpmovb2m %zmm4, %k1
; AVX512VLBW-NEXT: vpaddb %zmm0, %zmm0, %zmm0 {%k1}
; AVX512VLBW-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512VLBW-NEXT: vptestnmb %zmm3, %zmm2, %k1
; AVX512VLBW-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
; AVX512VLBW-NEXT: retq
;
; AVX512VLVBMI2-LABEL: var_funnnel_v64i8:
; AVX512VLVBMI2: # %bb.0:
; AVX512VLVBMI2-NEXT: vmovdqa64 {{.*#+}} zmm3 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7]
; AVX512VLVBMI2-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512VLVBMI2-NEXT: vpsllw $5, %zmm4, %zmm5
; AVX512VLVBMI2-NEXT: vpaddb %zmm5, %zmm5, %zmm6
; AVX512VLVBMI2-NEXT: vpmovb2m %zmm6, %k1
; AVX512VLVBMI2-NEXT: vpmovb2m %zmm5, %k2
; AVX512VLVBMI2-NEXT: vpsrlw $4, %zmm1, %zmm5
; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm5, %zmm5
; AVX512VLVBMI2-NEXT: vpblendmb %zmm5, %zmm1, %zmm5 {%k2}
; AVX512VLVBMI2-NEXT: vpsrlw $2, %zmm5, %zmm7
; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm7, %zmm7
; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm7, %zmm5 {%k1}
; AVX512VLVBMI2-NEXT: vpsrlw $1, %zmm5, %zmm7
; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm7, %zmm7
; AVX512VLVBMI2-NEXT: vpaddb %zmm6, %zmm6, %zmm6
; AVX512VLVBMI2-NEXT: vpmovb2m %zmm6, %k1
; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm7, %zmm5 {%k1}
; AVX512VLVBMI2-NEXT: vmovdqa64 {{.*#+}} zmm6 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8]
; AVX512VLVBMI2-NEXT: vpsubb %zmm4, %zmm6, %zmm4
; AVX512VLVBMI2-NEXT: vpsllw $5, %zmm4, %zmm4
; AVX512VLVBMI2-NEXT: vpaddb %zmm4, %zmm4, %zmm6
; AVX512VLVBMI2-NEXT: vpmovb2m %zmm6, %k1
; AVX512VLVBMI2-NEXT: vpmovb2m %zmm4, %k2
; AVX512VLVBMI2-NEXT: vpsllw $4, %zmm0, %zmm4
; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm4, %zmm4
; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm4, %zmm0 {%k2}
; AVX512VLVBMI2-NEXT: vpsllw $2, %zmm0, %zmm4
; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm4, %zmm4
; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm4, %zmm0 {%k1}
; AVX512VLVBMI2-NEXT: vpaddb %zmm6, %zmm6, %zmm4
; AVX512VLVBMI2-NEXT: vpmovb2m %zmm4, %k1
; AVX512VLVBMI2-NEXT: vpaddb %zmm0, %zmm0, %zmm0 {%k1}
; AVX512VLVBMI2-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512VLVBMI2-NEXT: vptestnmb %zmm3, %zmm2, %k1
; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
; AVX512VLVBMI2-NEXT: retq
%res = call <64 x i8> @llvm.fshr.v64i8(<64 x i8> %x, <64 x i8> %y, <64 x i8> %amt)
ret <64 x i8> %res
}
;
; Uniform Variable Shifts
;
define <8 x i64> @splatvar_funnnel_v8i64(<8 x i64> %x, <8 x i64> %y, <8 x i64> %amt) nounwind {
; AVX512F-LABEL: splatvar_funnnel_v8i64:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpbroadcastq %xmm2, %zmm2
; AVX512F-NEXT: vpbroadcastq {{.*#+}} zmm3 = [63,63,63,63,63,63,63,63]
; AVX512F-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512F-NEXT: vpsrlq %xmm4, %zmm1, %zmm5
; AVX512F-NEXT: vmovdqa {{.*#+}} xmm6 = [64,64]
; AVX512F-NEXT: vpsubq %xmm4, %xmm6, %xmm4
; AVX512F-NEXT: vpsllq %xmm4, %zmm0, %zmm0
; AVX512F-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512F-NEXT: vptestnmq %zmm3, %zmm2, %k1
; AVX512F-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: splatvar_funnnel_v8i64:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpbroadcastq %xmm2, %zmm2
; AVX512VL-NEXT: vpbroadcastq {{.*#+}} zmm3 = [63,63,63,63,63,63,63,63]
; AVX512VL-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512VL-NEXT: vpsrlq %xmm4, %zmm1, %zmm5
; AVX512VL-NEXT: vmovdqa {{.*#+}} xmm6 = [64,64]
; AVX512VL-NEXT: vpsubq %xmm4, %xmm6, %xmm4
; AVX512VL-NEXT: vpsllq %xmm4, %zmm0, %zmm0
; AVX512VL-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512VL-NEXT: vptestnmq %zmm3, %zmm2, %k1
; AVX512VL-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: splatvar_funnnel_v8i64:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpbroadcastq %xmm2, %zmm2
; AVX512BW-NEXT: vpbroadcastq {{.*#+}} zmm3 = [63,63,63,63,63,63,63,63]
; AVX512BW-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512BW-NEXT: vpsrlq %xmm4, %zmm1, %zmm5
; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm6 = [64,64]
; AVX512BW-NEXT: vpsubq %xmm4, %xmm6, %xmm4
; AVX512BW-NEXT: vpsllq %xmm4, %zmm0, %zmm0
; AVX512BW-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512BW-NEXT: vptestnmq %zmm3, %zmm2, %k1
; AVX512BW-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512BW-NEXT: retq
;
; AVX512VBMI2-LABEL: splatvar_funnnel_v8i64:
; AVX512VBMI2: # %bb.0:
; AVX512VBMI2-NEXT: vpbroadcastq %xmm2, %zmm2
; AVX512VBMI2-NEXT: vpshrdvq %zmm2, %zmm0, %zmm1
; AVX512VBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VBMI2-NEXT: retq
;
; AVX512VLBW-LABEL: splatvar_funnnel_v8i64:
; AVX512VLBW: # %bb.0:
; AVX512VLBW-NEXT: vpbroadcastq %xmm2, %zmm2
; AVX512VLBW-NEXT: vpbroadcastq {{.*#+}} zmm3 = [63,63,63,63,63,63,63,63]
; AVX512VLBW-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512VLBW-NEXT: vpsrlq %xmm4, %zmm1, %zmm5
; AVX512VLBW-NEXT: vmovdqa {{.*#+}} xmm6 = [64,64]
; AVX512VLBW-NEXT: vpsubq %xmm4, %xmm6, %xmm4
; AVX512VLBW-NEXT: vpsllq %xmm4, %zmm0, %zmm0
; AVX512VLBW-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512VLBW-NEXT: vptestnmq %zmm3, %zmm2, %k1
; AVX512VLBW-NEXT: vmovdqa64 %zmm1, %zmm0 {%k1}
; AVX512VLBW-NEXT: retq
;
; AVX512VLVBMI2-LABEL: splatvar_funnnel_v8i64:
; AVX512VLVBMI2: # %bb.0:
; AVX512VLVBMI2-NEXT: vpbroadcastq %xmm2, %zmm2
; AVX512VLVBMI2-NEXT: vpshrdvq %zmm2, %zmm0, %zmm1
; AVX512VLVBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VLVBMI2-NEXT: retq
%splat = shufflevector <8 x i64> %amt, <8 x i64> undef, <8 x i32> zeroinitializer
%res = call <8 x i64> @llvm.fshr.v8i64(<8 x i64> %x, <8 x i64> %y, <8 x i64> %splat)
ret <8 x i64> %res
}
define <16 x i32> @splatvar_funnnel_v16i32(<16 x i32> %x, <16 x i32> %y, <16 x i32> %amt) nounwind {
; AVX512F-LABEL: splatvar_funnnel_v16i32:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpbroadcastd %xmm2, %zmm2
; AVX512F-NEXT: vpbroadcastd {{.*#+}} zmm3 = [31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31]
; AVX512F-NEXT: vpandd %zmm3, %zmm2, %zmm4
; AVX512F-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero
; AVX512F-NEXT: vpsrld %xmm5, %zmm1, %zmm5
; AVX512F-NEXT: vpbroadcastd {{.*#+}} xmm6 = [32,32,32,32]
; AVX512F-NEXT: vpsubd %xmm4, %xmm6, %xmm4
; AVX512F-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero
; AVX512F-NEXT: vpslld %xmm4, %zmm0, %zmm0
; AVX512F-NEXT: vpord %zmm5, %zmm0, %zmm0
; AVX512F-NEXT: vptestnmd %zmm3, %zmm2, %k1
; AVX512F-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: splatvar_funnnel_v16i32:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpbroadcastd %xmm2, %zmm2
; AVX512VL-NEXT: vpbroadcastd {{.*#+}} zmm3 = [31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31]
; AVX512VL-NEXT: vpandd %zmm3, %zmm2, %zmm4
; AVX512VL-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero
; AVX512VL-NEXT: vpsrld %xmm5, %zmm1, %zmm5
; AVX512VL-NEXT: vpbroadcastd {{.*#+}} xmm6 = [32,32,32,32]
; AVX512VL-NEXT: vpsubd %xmm4, %xmm6, %xmm4
; AVX512VL-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero
; AVX512VL-NEXT: vpslld %xmm4, %zmm0, %zmm0
; AVX512VL-NEXT: vpord %zmm5, %zmm0, %zmm0
; AVX512VL-NEXT: vptestnmd %zmm3, %zmm2, %k1
; AVX512VL-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: splatvar_funnnel_v16i32:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpbroadcastd %xmm2, %zmm2
; AVX512BW-NEXT: vpbroadcastd {{.*#+}} zmm3 = [31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31]
; AVX512BW-NEXT: vpandd %zmm3, %zmm2, %zmm4
; AVX512BW-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero
; AVX512BW-NEXT: vpsrld %xmm5, %zmm1, %zmm5
; AVX512BW-NEXT: vpbroadcastd {{.*#+}} xmm6 = [32,32,32,32]
; AVX512BW-NEXT: vpsubd %xmm4, %xmm6, %xmm4
; AVX512BW-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero
; AVX512BW-NEXT: vpslld %xmm4, %zmm0, %zmm0
; AVX512BW-NEXT: vpord %zmm5, %zmm0, %zmm0
; AVX512BW-NEXT: vptestnmd %zmm3, %zmm2, %k1
; AVX512BW-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
; AVX512BW-NEXT: retq
;
; AVX512VBMI2-LABEL: splatvar_funnnel_v16i32:
; AVX512VBMI2: # %bb.0:
; AVX512VBMI2-NEXT: vpbroadcastd %xmm2, %zmm2
; AVX512VBMI2-NEXT: vpshrdvd %zmm2, %zmm0, %zmm1
; AVX512VBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VBMI2-NEXT: retq
;
; AVX512VLBW-LABEL: splatvar_funnnel_v16i32:
; AVX512VLBW: # %bb.0:
; AVX512VLBW-NEXT: vpbroadcastd %xmm2, %zmm2
; AVX512VLBW-NEXT: vpbroadcastd {{.*#+}} zmm3 = [31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31]
; AVX512VLBW-NEXT: vpandd %zmm3, %zmm2, %zmm4
; AVX512VLBW-NEXT: vpmovzxdq {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero
; AVX512VLBW-NEXT: vpsrld %xmm5, %zmm1, %zmm5
; AVX512VLBW-NEXT: vpbroadcastd {{.*#+}} xmm6 = [32,32,32,32]
; AVX512VLBW-NEXT: vpsubd %xmm4, %xmm6, %xmm4
; AVX512VLBW-NEXT: vpmovzxdq {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero
; AVX512VLBW-NEXT: vpslld %xmm4, %zmm0, %zmm0
; AVX512VLBW-NEXT: vpord %zmm5, %zmm0, %zmm0
; AVX512VLBW-NEXT: vptestnmd %zmm3, %zmm2, %k1
; AVX512VLBW-NEXT: vmovdqa32 %zmm1, %zmm0 {%k1}
; AVX512VLBW-NEXT: retq
;
; AVX512VLVBMI2-LABEL: splatvar_funnnel_v16i32:
; AVX512VLVBMI2: # %bb.0:
; AVX512VLVBMI2-NEXT: vpbroadcastd %xmm2, %zmm2
; AVX512VLVBMI2-NEXT: vpshrdvd %zmm2, %zmm0, %zmm1
; AVX512VLVBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VLVBMI2-NEXT: retq
%splat = shufflevector <16 x i32> %amt, <16 x i32> undef, <16 x i32> zeroinitializer
%res = call <16 x i32> @llvm.fshr.v16i32(<16 x i32> %x, <16 x i32> %y, <16 x i32> %splat)
ret <16 x i32> %res
}
define <32 x i16> @splatvar_funnnel_v32i16(<32 x i16> %x, <32 x i16> %y, <32 x i16> %amt) nounwind {
; AVX512F-LABEL: splatvar_funnnel_v32i16:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpbroadcastw %xmm4, %ymm4
; AVX512F-NEXT: vpand {{.*}}(%rip), %ymm4, %ymm4
; AVX512F-NEXT: vpmovzxwq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX512F-NEXT: vpsrlw %xmm5, %ymm2, %ymm6
; AVX512F-NEXT: vmovdqa {{.*#+}} xmm7 = [16,16,16,16,16,16,16,16]
; AVX512F-NEXT: vpsubw %xmm4, %xmm7, %xmm7
; AVX512F-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
; AVX512F-NEXT: vpsllw %xmm7, %ymm0, %ymm0
; AVX512F-NEXT: vpor %ymm6, %ymm0, %ymm0
; AVX512F-NEXT: vpxor %xmm6, %xmm6, %xmm6
; AVX512F-NEXT: vpcmpeqw %ymm6, %ymm4, %ymm4
; AVX512F-NEXT: vpblendvb %ymm4, %ymm2, %ymm0, %ymm0
; AVX512F-NEXT: vpsrlw %xmm5, %ymm3, %ymm2
; AVX512F-NEXT: vpsllw %xmm7, %ymm1, %ymm1
; AVX512F-NEXT: vpor %ymm2, %ymm1, %ymm1
; AVX512F-NEXT: vpblendvb %ymm4, %ymm3, %ymm1, %ymm1
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: splatvar_funnnel_v32i16:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpbroadcastw %xmm4, %ymm4
; AVX512VL-NEXT: vpand {{.*}}(%rip), %ymm4, %ymm4
; AVX512VL-NEXT: vpmovzxwq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX512VL-NEXT: vpsrlw %xmm5, %ymm2, %ymm6
; AVX512VL-NEXT: vmovdqa {{.*#+}} xmm7 = [16,16,16,16,16,16,16,16]
; AVX512VL-NEXT: vpsubw %xmm4, %xmm7, %xmm7
; AVX512VL-NEXT: vpmovzxwq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,xmm7[1],zero,zero,zero
; AVX512VL-NEXT: vpsllw %xmm7, %ymm0, %ymm0
; AVX512VL-NEXT: vpor %ymm6, %ymm0, %ymm0
; AVX512VL-NEXT: vpxor %xmm6, %xmm6, %xmm6
; AVX512VL-NEXT: vpcmpeqw %ymm6, %ymm4, %ymm4
; AVX512VL-NEXT: vpblendvb %ymm4, %ymm2, %ymm0, %ymm0
; AVX512VL-NEXT: vpsrlw %xmm5, %ymm3, %ymm2
; AVX512VL-NEXT: vpsllw %xmm7, %ymm1, %ymm1
; AVX512VL-NEXT: vpor %ymm2, %ymm1, %ymm1
; AVX512VL-NEXT: vpblendvb %ymm4, %ymm3, %ymm1, %ymm1
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: splatvar_funnnel_v32i16:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpbroadcastw %xmm2, %zmm2
; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX512BW-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX512BW-NEXT: vpsrlw %xmm5, %zmm1, %zmm5
; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm6 = [16,16,16,16,16,16,16,16]
; AVX512BW-NEXT: vpsubw %xmm4, %xmm6, %xmm4
; AVX512BW-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX512BW-NEXT: vpsllw %xmm4, %zmm0, %zmm0
; AVX512BW-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512BW-NEXT: vptestnmw %zmm3, %zmm2, %k1
; AVX512BW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k1}
; AVX512BW-NEXT: retq
;
; AVX512VBMI2-LABEL: splatvar_funnnel_v32i16:
; AVX512VBMI2: # %bb.0:
; AVX512VBMI2-NEXT: vpbroadcastw %xmm2, %zmm2
; AVX512VBMI2-NEXT: vpshrdvw %zmm2, %zmm0, %zmm1
; AVX512VBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VBMI2-NEXT: retq
;
; AVX512VLBW-LABEL: splatvar_funnnel_v32i16:
; AVX512VLBW: # %bb.0:
; AVX512VLBW-NEXT: vpbroadcastw %xmm2, %zmm2
; AVX512VLBW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX512VLBW-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512VLBW-NEXT: vpmovzxwq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX512VLBW-NEXT: vpsrlw %xmm5, %zmm1, %zmm5
; AVX512VLBW-NEXT: vmovdqa {{.*#+}} xmm6 = [16,16,16,16,16,16,16,16]
; AVX512VLBW-NEXT: vpsubw %xmm4, %xmm6, %xmm4
; AVX512VLBW-NEXT: vpmovzxwq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,xmm4[1],zero,zero,zero
; AVX512VLBW-NEXT: vpsllw %xmm4, %zmm0, %zmm0
; AVX512VLBW-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512VLBW-NEXT: vptestnmw %zmm3, %zmm2, %k1
; AVX512VLBW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k1}
; AVX512VLBW-NEXT: retq
;
; AVX512VLVBMI2-LABEL: splatvar_funnnel_v32i16:
; AVX512VLVBMI2: # %bb.0:
; AVX512VLVBMI2-NEXT: vpbroadcastw %xmm2, %zmm2
; AVX512VLVBMI2-NEXT: vpshrdvw %zmm2, %zmm0, %zmm1
; AVX512VLVBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VLVBMI2-NEXT: retq
%splat = shufflevector <32 x i16> %amt, <32 x i16> undef, <32 x i32> zeroinitializer
%res = call <32 x i16> @llvm.fshr.v32i16(<32 x i16> %x, <32 x i16> %y, <32 x i16> %splat)
ret <32 x i16> %res
}
define <64 x i8> @splatvar_funnnel_v64i8(<64 x i8> %x, <64 x i8> %y, <64 x i8> %amt) nounwind {
; AVX512F-LABEL: splatvar_funnnel_v64i8:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpbroadcastb %xmm4, %ymm4
; AVX512F-NEXT: vpand {{.*}}(%rip), %ymm4, %ymm4
; AVX512F-NEXT: vpmovzxbq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero
; AVX512F-NEXT: vpsrlw %xmm5, %ymm2, %ymm6
; AVX512F-NEXT: vpcmpeqd %ymm9, %ymm9, %ymm9
; AVX512F-NEXT: vpsrlw %xmm5, %ymm9, %ymm8
; AVX512F-NEXT: vpsrlw $8, %ymm8, %ymm8
; AVX512F-NEXT: vpbroadcastb %xmm8, %ymm8
; AVX512F-NEXT: vpand %ymm8, %ymm6, %ymm6
; AVX512F-NEXT: vmovdqa {{.*#+}} xmm7 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8]
; AVX512F-NEXT: vpsubb %xmm4, %xmm7, %xmm7
; AVX512F-NEXT: vpmovzxbq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,zero,zero,zero,zero,xmm7[1],zero,zero,zero,zero,zero,zero,zero
; AVX512F-NEXT: vpsllw %xmm7, %ymm0, %ymm0
; AVX512F-NEXT: vpsllw %xmm7, %ymm9, %ymm9
; AVX512F-NEXT: vpbroadcastb %xmm9, %ymm9
; AVX512F-NEXT: vpand %ymm9, %ymm0, %ymm0
; AVX512F-NEXT: vpor %ymm6, %ymm0, %ymm0
; AVX512F-NEXT: vpxor %xmm6, %xmm6, %xmm6
; AVX512F-NEXT: vpcmpeqb %ymm6, %ymm4, %ymm4
; AVX512F-NEXT: vpblendvb %ymm4, %ymm2, %ymm0, %ymm0
; AVX512F-NEXT: vpsrlw %xmm5, %ymm3, %ymm2
; AVX512F-NEXT: vpand %ymm8, %ymm2, %ymm2
; AVX512F-NEXT: vpsllw %xmm7, %ymm1, %ymm1
; AVX512F-NEXT: vpand %ymm9, %ymm1, %ymm1
; AVX512F-NEXT: vpor %ymm2, %ymm1, %ymm1
; AVX512F-NEXT: vpblendvb %ymm4, %ymm3, %ymm1, %ymm1
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: splatvar_funnnel_v64i8:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpbroadcastb %xmm4, %ymm4
; AVX512VL-NEXT: vpand {{.*}}(%rip), %ymm4, %ymm4
; AVX512VL-NEXT: vpmovzxbq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero
; AVX512VL-NEXT: vpsrlw %xmm5, %ymm2, %ymm6
; AVX512VL-NEXT: vpcmpeqd %ymm9, %ymm9, %ymm9
; AVX512VL-NEXT: vpsrlw %xmm5, %ymm9, %ymm8
; AVX512VL-NEXT: vpsrlw $8, %ymm8, %ymm8
; AVX512VL-NEXT: vpbroadcastb %xmm8, %ymm8
; AVX512VL-NEXT: vpand %ymm8, %ymm6, %ymm6
; AVX512VL-NEXT: vmovdqa {{.*#+}} xmm7 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8]
; AVX512VL-NEXT: vpsubb %xmm4, %xmm7, %xmm7
; AVX512VL-NEXT: vpmovzxbq {{.*#+}} xmm7 = xmm7[0],zero,zero,zero,zero,zero,zero,zero,xmm7[1],zero,zero,zero,zero,zero,zero,zero
; AVX512VL-NEXT: vpsllw %xmm7, %ymm0, %ymm0
; AVX512VL-NEXT: vpsllw %xmm7, %ymm9, %ymm9
; AVX512VL-NEXT: vpbroadcastb %xmm9, %ymm9
; AVX512VL-NEXT: vpand %ymm9, %ymm0, %ymm0
; AVX512VL-NEXT: vpor %ymm6, %ymm0, %ymm0
; AVX512VL-NEXT: vpxor %xmm6, %xmm6, %xmm6
; AVX512VL-NEXT: vpcmpeqb %ymm6, %ymm4, %ymm4
; AVX512VL-NEXT: vpblendvb %ymm4, %ymm2, %ymm0, %ymm0
; AVX512VL-NEXT: vpsrlw %xmm5, %ymm3, %ymm2
; AVX512VL-NEXT: vpand %ymm8, %ymm2, %ymm2
; AVX512VL-NEXT: vpsllw %xmm7, %ymm1, %ymm1
; AVX512VL-NEXT: vpand %ymm9, %ymm1, %ymm1
; AVX512VL-NEXT: vpor %ymm2, %ymm1, %ymm1
; AVX512VL-NEXT: vpblendvb %ymm4, %ymm3, %ymm1, %ymm1
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: splatvar_funnnel_v64i8:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpbroadcastb %xmm2, %zmm2
; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7]
; AVX512BW-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512BW-NEXT: vpmovzxbq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero
; AVX512BW-NEXT: vpsrlw %xmm5, %zmm1, %zmm6
; AVX512BW-NEXT: vpternlogd $255, %zmm7, %zmm7, %zmm7
; AVX512BW-NEXT: vpsrlw %xmm5, %zmm7, %zmm5
; AVX512BW-NEXT: vpsrlw $8, %zmm5, %zmm5
; AVX512BW-NEXT: vpbroadcastb %xmm5, %zmm5
; AVX512BW-NEXT: vpandq %zmm5, %zmm6, %zmm5
; AVX512BW-NEXT: vmovdqa {{.*#+}} xmm6 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8]
; AVX512BW-NEXT: vpsubb %xmm4, %xmm6, %xmm4
; AVX512BW-NEXT: vpmovzxbq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero
; AVX512BW-NEXT: vpsllw %xmm4, %zmm0, %zmm0
; AVX512BW-NEXT: vpsllw %xmm4, %zmm7, %zmm4
; AVX512BW-NEXT: vpbroadcastb %xmm4, %zmm4
; AVX512BW-NEXT: vpandq %zmm4, %zmm0, %zmm0
; AVX512BW-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512BW-NEXT: vptestnmb %zmm3, %zmm2, %k1
; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
; AVX512BW-NEXT: retq
;
; AVX512VBMI2-LABEL: splatvar_funnnel_v64i8:
; AVX512VBMI2: # %bb.0:
; AVX512VBMI2-NEXT: vpbroadcastb %xmm2, %zmm2
; AVX512VBMI2-NEXT: vmovdqa64 {{.*#+}} zmm3 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7]
; AVX512VBMI2-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512VBMI2-NEXT: vpmovzxbq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero
; AVX512VBMI2-NEXT: vpsrlw %xmm5, %zmm1, %zmm6
; AVX512VBMI2-NEXT: vpternlogd $255, %zmm7, %zmm7, %zmm7
; AVX512VBMI2-NEXT: vpsrlw %xmm5, %zmm7, %zmm5
; AVX512VBMI2-NEXT: vpsrlw $8, %zmm5, %zmm5
; AVX512VBMI2-NEXT: vpbroadcastb %xmm5, %zmm5
; AVX512VBMI2-NEXT: vpandq %zmm5, %zmm6, %zmm5
; AVX512VBMI2-NEXT: vmovdqa {{.*#+}} xmm6 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8]
; AVX512VBMI2-NEXT: vpsubb %xmm4, %xmm6, %xmm4
; AVX512VBMI2-NEXT: vpmovzxbq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero
; AVX512VBMI2-NEXT: vpsllw %xmm4, %zmm0, %zmm0
; AVX512VBMI2-NEXT: vpsllw %xmm4, %zmm7, %zmm4
; AVX512VBMI2-NEXT: vpbroadcastb %xmm4, %zmm4
; AVX512VBMI2-NEXT: vpandq %zmm4, %zmm0, %zmm0
; AVX512VBMI2-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512VBMI2-NEXT: vptestnmb %zmm3, %zmm2, %k1
; AVX512VBMI2-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
; AVX512VBMI2-NEXT: retq
;
; AVX512VLBW-LABEL: splatvar_funnnel_v64i8:
; AVX512VLBW: # %bb.0:
; AVX512VLBW-NEXT: vpbroadcastb %xmm2, %zmm2
; AVX512VLBW-NEXT: vmovdqa64 {{.*#+}} zmm3 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7]
; AVX512VLBW-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512VLBW-NEXT: vpmovzxbq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero
; AVX512VLBW-NEXT: vpsrlw %xmm5, %zmm1, %zmm6
; AVX512VLBW-NEXT: vpternlogd $255, %zmm7, %zmm7, %zmm7
; AVX512VLBW-NEXT: vpsrlw %xmm5, %zmm7, %zmm5
; AVX512VLBW-NEXT: vpsrlw $8, %zmm5, %zmm5
; AVX512VLBW-NEXT: vpbroadcastb %xmm5, %zmm5
; AVX512VLBW-NEXT: vpandq %zmm5, %zmm6, %zmm5
; AVX512VLBW-NEXT: vmovdqa {{.*#+}} xmm6 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8]
; AVX512VLBW-NEXT: vpsubb %xmm4, %xmm6, %xmm4
; AVX512VLBW-NEXT: vpmovzxbq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero
; AVX512VLBW-NEXT: vpsllw %xmm4, %zmm0, %zmm0
; AVX512VLBW-NEXT: vpsllw %xmm4, %zmm7, %zmm4
; AVX512VLBW-NEXT: vpbroadcastb %xmm4, %zmm4
; AVX512VLBW-NEXT: vpandq %zmm4, %zmm0, %zmm0
; AVX512VLBW-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512VLBW-NEXT: vptestnmb %zmm3, %zmm2, %k1
; AVX512VLBW-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
; AVX512VLBW-NEXT: retq
;
; AVX512VLVBMI2-LABEL: splatvar_funnnel_v64i8:
; AVX512VLVBMI2: # %bb.0:
; AVX512VLVBMI2-NEXT: vpbroadcastb %xmm2, %zmm2
; AVX512VLVBMI2-NEXT: vmovdqa64 {{.*#+}} zmm3 = [7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7]
; AVX512VLVBMI2-NEXT: vpandq %zmm3, %zmm2, %zmm4
; AVX512VLVBMI2-NEXT: vpmovzxbq {{.*#+}} xmm5 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero
; AVX512VLVBMI2-NEXT: vpsrlw %xmm5, %zmm1, %zmm6
; AVX512VLVBMI2-NEXT: vpternlogd $255, %zmm7, %zmm7, %zmm7
; AVX512VLVBMI2-NEXT: vpsrlw %xmm5, %zmm7, %zmm5
; AVX512VLVBMI2-NEXT: vpsrlw $8, %zmm5, %zmm5
; AVX512VLVBMI2-NEXT: vpbroadcastb %xmm5, %zmm5
; AVX512VLVBMI2-NEXT: vpandq %zmm5, %zmm6, %zmm5
; AVX512VLVBMI2-NEXT: vmovdqa {{.*#+}} xmm6 = [8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8]
; AVX512VLVBMI2-NEXT: vpsubb %xmm4, %xmm6, %xmm4
; AVX512VLVBMI2-NEXT: vpmovzxbq {{.*#+}} xmm4 = xmm4[0],zero,zero,zero,zero,zero,zero,zero,xmm4[1],zero,zero,zero,zero,zero,zero,zero
; AVX512VLVBMI2-NEXT: vpsllw %xmm4, %zmm0, %zmm0
; AVX512VLVBMI2-NEXT: vpsllw %xmm4, %zmm7, %zmm4
; AVX512VLVBMI2-NEXT: vpbroadcastb %xmm4, %zmm4
; AVX512VLVBMI2-NEXT: vpandq %zmm4, %zmm0, %zmm0
; AVX512VLVBMI2-NEXT: vporq %zmm5, %zmm0, %zmm0
; AVX512VLVBMI2-NEXT: vptestnmb %zmm3, %zmm2, %k1
; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
; AVX512VLVBMI2-NEXT: retq
%splat = shufflevector <64 x i8> %amt, <64 x i8> undef, <64 x i32> zeroinitializer
%res = call <64 x i8> @llvm.fshr.v64i8(<64 x i8> %x, <64 x i8> %y, <64 x i8> %splat)
ret <64 x i8> %res
}
;
; Constant Shifts
;
define <8 x i64> @constant_funnnel_v8i64(<8 x i64> %x, <8 x i64> %y) nounwind {
; AVX512F-LABEL: constant_funnnel_v8i64:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpsrlvq {{.*}}(%rip), %zmm1, %zmm1
; AVX512F-NEXT: vpsllvq {{.*}}(%rip), %zmm0, %zmm0
; AVX512F-NEXT: vporq %zmm1, %zmm0, %zmm0
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: constant_funnnel_v8i64:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpsrlvq {{.*}}(%rip), %zmm1, %zmm1
; AVX512VL-NEXT: vpsllvq {{.*}}(%rip), %zmm0, %zmm0
; AVX512VL-NEXT: vporq %zmm1, %zmm0, %zmm0
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: constant_funnnel_v8i64:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpsrlvq {{.*}}(%rip), %zmm1, %zmm1
; AVX512BW-NEXT: vpsllvq {{.*}}(%rip), %zmm0, %zmm0
; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0
; AVX512BW-NEXT: retq
;
; AVX512VBMI2-LABEL: constant_funnnel_v8i64:
; AVX512VBMI2: # %bb.0:
; AVX512VBMI2-NEXT: vpshrdvq {{.*}}(%rip), %zmm0, %zmm1
; AVX512VBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VBMI2-NEXT: retq
;
; AVX512VLBW-LABEL: constant_funnnel_v8i64:
; AVX512VLBW: # %bb.0:
; AVX512VLBW-NEXT: vpsrlvq {{.*}}(%rip), %zmm1, %zmm1
; AVX512VLBW-NEXT: vpsllvq {{.*}}(%rip), %zmm0, %zmm0
; AVX512VLBW-NEXT: vporq %zmm1, %zmm0, %zmm0
; AVX512VLBW-NEXT: retq
;
; AVX512VLVBMI2-LABEL: constant_funnnel_v8i64:
; AVX512VLVBMI2: # %bb.0:
; AVX512VLVBMI2-NEXT: vpshrdvq {{.*}}(%rip), %zmm0, %zmm1
; AVX512VLVBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VLVBMI2-NEXT: retq
%res = call <8 x i64> @llvm.fshr.v8i64(<8 x i64> %x, <8 x i64> %y, <8 x i64> <i64 4, i64 14, i64 50, i64 60, i64 4, i64 14, i64 50, i64 60>)
ret <8 x i64> %res
}
define <16 x i32> @constant_funnnel_v16i32(<16 x i32> %x, <16 x i32> %y) nounwind {
; AVX512F-LABEL: constant_funnnel_v16i32:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpsrlvd {{.*}}(%rip), %zmm1, %zmm1
; AVX512F-NEXT: vpsllvd {{.*}}(%rip), %zmm0, %zmm0
; AVX512F-NEXT: vpord %zmm1, %zmm0, %zmm0
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: constant_funnnel_v16i32:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpsrlvd {{.*}}(%rip), %zmm1, %zmm1
; AVX512VL-NEXT: vpsllvd {{.*}}(%rip), %zmm0, %zmm0
; AVX512VL-NEXT: vpord %zmm1, %zmm0, %zmm0
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: constant_funnnel_v16i32:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpsrlvd {{.*}}(%rip), %zmm1, %zmm1
; AVX512BW-NEXT: vpsllvd {{.*}}(%rip), %zmm0, %zmm0
; AVX512BW-NEXT: vpord %zmm1, %zmm0, %zmm0
; AVX512BW-NEXT: retq
;
; AVX512VBMI2-LABEL: constant_funnnel_v16i32:
; AVX512VBMI2: # %bb.0:
; AVX512VBMI2-NEXT: vpshrdvd {{.*}}(%rip), %zmm0, %zmm1
; AVX512VBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VBMI2-NEXT: retq
;
; AVX512VLBW-LABEL: constant_funnnel_v16i32:
; AVX512VLBW: # %bb.0:
; AVX512VLBW-NEXT: vpsrlvd {{.*}}(%rip), %zmm1, %zmm1
; AVX512VLBW-NEXT: vpsllvd {{.*}}(%rip), %zmm0, %zmm0
; AVX512VLBW-NEXT: vpord %zmm1, %zmm0, %zmm0
; AVX512VLBW-NEXT: retq
;
; AVX512VLVBMI2-LABEL: constant_funnnel_v16i32:
; AVX512VLVBMI2: # %bb.0:
; AVX512VLVBMI2-NEXT: vpshrdvd {{.*}}(%rip), %zmm0, %zmm1
; AVX512VLVBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VLVBMI2-NEXT: retq
%res = call <16 x i32> @llvm.fshr.v16i32(<16 x i32> %x, <16 x i32> %y, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>)
ret <16 x i32> %res
}
define <32 x i16> @constant_funnnel_v32i16(<32 x i16> %x, <32 x i16> %y) nounwind {
; AVX512F-LABEL: constant_funnnel_v32i16:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm4 = <u,32768,16384,8192,4096,2048,1024,512,256,128,64,32,16,8,4,2>
; AVX512F-NEXT: vpmulhuw %ymm4, %ymm2, %ymm5
; AVX512F-NEXT: vpblendw {{.*#+}} ymm6 = ymm2[0],ymm5[1,2,3,4,5,6,7],ymm2[8],ymm5[9,10,11,12,13,14,15]
; AVX512F-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
; AVX512F-NEXT: vpmullw %ymm4, %ymm0, %ymm0
; AVX512F-NEXT: vpor %ymm5, %ymm0, %ymm0
; AVX512F-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm0[1,2,3,4,5,6,7],ymm2[8],ymm0[9,10,11,12,13,14,15]
; AVX512F-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
; AVX512F-NEXT: vpmulhuw %ymm4, %ymm3, %ymm2
; AVX512F-NEXT: vpblendw {{.*#+}} ymm5 = ymm3[0],ymm2[1,2,3,4,5,6,7],ymm3[8],ymm2[9,10,11,12,13,14,15]
; AVX512F-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
; AVX512F-NEXT: vpmullw %ymm4, %ymm1, %ymm1
; AVX512F-NEXT: vpor %ymm2, %ymm1, %ymm1
; AVX512F-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm1[1,2,3,4,5,6,7],ymm3[8],ymm1[9,10,11,12,13,14,15]
; AVX512F-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: constant_funnnel_v32i16:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm4 = <u,32768,16384,8192,4096,2048,1024,512,256,128,64,32,16,8,4,2>
; AVX512VL-NEXT: vpmulhuw %ymm4, %ymm2, %ymm5
; AVX512VL-NEXT: vpblendw {{.*#+}} ymm6 = ymm2[0],ymm5[1,2,3,4,5,6,7],ymm2[8],ymm5[9,10,11,12,13,14,15]
; AVX512VL-NEXT: vpblendd {{.*#+}} ymm5 = ymm6[0,1,2,3],ymm5[4,5,6,7]
; AVX512VL-NEXT: vpmullw %ymm4, %ymm0, %ymm0
; AVX512VL-NEXT: vpor %ymm5, %ymm0, %ymm0
; AVX512VL-NEXT: vpblendw {{.*#+}} ymm2 = ymm2[0],ymm0[1,2,3,4,5,6,7],ymm2[8],ymm0[9,10,11,12,13,14,15]
; AVX512VL-NEXT: vpblendd {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
; AVX512VL-NEXT: vpmulhuw %ymm4, %ymm3, %ymm2
; AVX512VL-NEXT: vpblendw {{.*#+}} ymm5 = ymm3[0],ymm2[1,2,3,4,5,6,7],ymm3[8],ymm2[9,10,11,12,13,14,15]
; AVX512VL-NEXT: vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
; AVX512VL-NEXT: vpmullw %ymm4, %ymm1, %ymm1
; AVX512VL-NEXT: vpor %ymm2, %ymm1, %ymm1
; AVX512VL-NEXT: vpblendw {{.*#+}} ymm2 = ymm3[0],ymm1[1,2,3,4,5,6,7],ymm3[8],ymm1[9,10,11,12,13,14,15]
; AVX512VL-NEXT: vpblendd {{.*#+}} ymm1 = ymm2[0,1,2,3],ymm1[4,5,6,7]
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: constant_funnnel_v32i16:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpsrlvw {{.*}}(%rip), %zmm1, %zmm2
; AVX512BW-NEXT: vpsllvw {{.*}}(%rip), %zmm0, %zmm0
; AVX512BW-NEXT: vporq %zmm2, %zmm0, %zmm0
; AVX512BW-NEXT: movl $65537, %eax # imm = 0x10001
; AVX512BW-NEXT: kmovd %eax, %k1
; AVX512BW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k1}
; AVX512BW-NEXT: retq
;
; AVX512VBMI2-LABEL: constant_funnnel_v32i16:
; AVX512VBMI2: # %bb.0:
; AVX512VBMI2-NEXT: vpshrdvw {{.*}}(%rip), %zmm0, %zmm1
; AVX512VBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VBMI2-NEXT: retq
;
; AVX512VLBW-LABEL: constant_funnnel_v32i16:
; AVX512VLBW: # %bb.0:
; AVX512VLBW-NEXT: vpsrlvw {{.*}}(%rip), %zmm1, %zmm2
; AVX512VLBW-NEXT: vpsllvw {{.*}}(%rip), %zmm0, %zmm0
; AVX512VLBW-NEXT: vporq %zmm2, %zmm0, %zmm0
; AVX512VLBW-NEXT: movl $65537, %eax # imm = 0x10001
; AVX512VLBW-NEXT: kmovd %eax, %k1
; AVX512VLBW-NEXT: vmovdqu16 %zmm1, %zmm0 {%k1}
; AVX512VLBW-NEXT: retq
;
; AVX512VLVBMI2-LABEL: constant_funnnel_v32i16:
; AVX512VLVBMI2: # %bb.0:
; AVX512VLVBMI2-NEXT: vpshrdvw {{.*}}(%rip), %zmm0, %zmm1
; AVX512VLVBMI2-NEXT: vmovdqa64 %zmm1, %zmm0
; AVX512VLVBMI2-NEXT: retq
%res = call <32 x i16> @llvm.fshr.v32i16(<32 x i16> %x, <32 x i16> %y, <32 x i16> <i16 0, i16 1, i16 2, i16 3, i16 4, i16 5, i16 6, i16 7, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15, i16 0, i16 1, i16 2, i16 3, i16 4, i16 5, i16 6, i16 7, i16 8, i16 9, i16 10, i16 11, i16 12, i16 13, i16 14, i16 15>)
ret <32 x i16> %res
}
define <64 x i8> @constant_funnnel_v64i8(<64 x i8> %x, <64 x i8> %y) nounwind {
; AVX512F-LABEL: constant_funnnel_v64i8:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpsllw $4, %ymm0, %ymm4
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm5 = [240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240]
; AVX512F-NEXT: vpand %ymm5, %ymm4, %ymm4
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm6 = [57600,41152,24704,8256,8448,24640,41088,57536,57600,41152,24704,8256,8448,24640,41088,57536]
; AVX512F-NEXT: vpblendvb %ymm6, %ymm4, %ymm0, %ymm0
; AVX512F-NEXT: vpsllw $2, %ymm0, %ymm4
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm7 = [252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252]
; AVX512F-NEXT: vpand %ymm7, %ymm4, %ymm4
; AVX512F-NEXT: vpaddb %ymm6, %ymm6, %ymm8
; AVX512F-NEXT: vpblendvb %ymm8, %ymm4, %ymm0, %ymm0
; AVX512F-NEXT: vpaddb %ymm0, %ymm0, %ymm4
; AVX512F-NEXT: vpaddb %ymm8, %ymm8, %ymm9
; AVX512F-NEXT: vpblendvb %ymm9, %ymm4, %ymm0, %ymm0
; AVX512F-NEXT: vpxor %xmm4, %xmm4, %xmm4
; AVX512F-NEXT: vpunpckhbw {{.*#+}} ymm10 = ymm2[8],ymm4[8],ymm2[9],ymm4[9],ymm2[10],ymm4[10],ymm2[11],ymm4[11],ymm2[12],ymm4[12],ymm2[13],ymm4[13],ymm2[14],ymm4[14],ymm2[15],ymm4[15],ymm2[24],ymm4[24],ymm2[25],ymm4[25],ymm2[26],ymm4[26],ymm2[27],ymm4[27],ymm2[28],ymm4[28],ymm2[29],ymm4[29],ymm2[30],ymm4[30],ymm2[31],ymm4[31]
; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm11 = [256,2,4,8,16,32,64,128,256,2,4,8,16,32,64,128]
; AVX512F-NEXT: # ymm11 = mem[0,1,0,1]
; AVX512F-NEXT: vpmullw %ymm11, %ymm10, %ymm10
; AVX512F-NEXT: vpsrlw $8, %ymm10, %ymm10
; AVX512F-NEXT: vpunpcklbw {{.*#+}} ymm12 = ymm2[0],ymm4[0],ymm2[1],ymm4[1],ymm2[2],ymm4[2],ymm2[3],ymm4[3],ymm2[4],ymm4[4],ymm2[5],ymm4[5],ymm2[6],ymm4[6],ymm2[7],ymm4[7],ymm2[16],ymm4[16],ymm2[17],ymm4[17],ymm2[18],ymm4[18],ymm2[19],ymm4[19],ymm2[20],ymm4[20],ymm2[21],ymm4[21],ymm2[22],ymm4[22],ymm2[23],ymm4[23]
; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm13 = [256,128,64,32,16,8,4,2,256,128,64,32,16,8,4,2]
; AVX512F-NEXT: # ymm13 = mem[0,1,0,1]
; AVX512F-NEXT: vpmullw %ymm13, %ymm12, %ymm12
; AVX512F-NEXT: vpsrlw $8, %ymm12, %ymm12
; AVX512F-NEXT: vpackuswb %ymm10, %ymm12, %ymm10
; AVX512F-NEXT: vpor %ymm10, %ymm0, %ymm0
; AVX512F-NEXT: vpbroadcastq {{.*#+}} ymm10 = [18446744073709551360,18446744073709551360,18446744073709551360,18446744073709551360]
; AVX512F-NEXT: vpblendvb %ymm10, %ymm0, %ymm2, %ymm0
; AVX512F-NEXT: vpsllw $4, %ymm1, %ymm2
; AVX512F-NEXT: vpand %ymm5, %ymm2, %ymm2
; AVX512F-NEXT: vpblendvb %ymm6, %ymm2, %ymm1, %ymm1
; AVX512F-NEXT: vpsllw $2, %ymm1, %ymm2
; AVX512F-NEXT: vpand %ymm7, %ymm2, %ymm2
; AVX512F-NEXT: vpblendvb %ymm8, %ymm2, %ymm1, %ymm1
; AVX512F-NEXT: vpaddb %ymm1, %ymm1, %ymm2
; AVX512F-NEXT: vpblendvb %ymm9, %ymm2, %ymm1, %ymm1
; AVX512F-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm3[8],ymm4[8],ymm3[9],ymm4[9],ymm3[10],ymm4[10],ymm3[11],ymm4[11],ymm3[12],ymm4[12],ymm3[13],ymm4[13],ymm3[14],ymm4[14],ymm3[15],ymm4[15],ymm3[24],ymm4[24],ymm3[25],ymm4[25],ymm3[26],ymm4[26],ymm3[27],ymm4[27],ymm3[28],ymm4[28],ymm3[29],ymm4[29],ymm3[30],ymm4[30],ymm3[31],ymm4[31]
; AVX512F-NEXT: vpmullw %ymm11, %ymm2, %ymm2
; AVX512F-NEXT: vpsrlw $8, %ymm2, %ymm2
; AVX512F-NEXT: vpunpcklbw {{.*#+}} ymm4 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[4],ymm4[4],ymm3[5],ymm4[5],ymm3[6],ymm4[6],ymm3[7],ymm4[7],ymm3[16],ymm4[16],ymm3[17],ymm4[17],ymm3[18],ymm4[18],ymm3[19],ymm4[19],ymm3[20],ymm4[20],ymm3[21],ymm4[21],ymm3[22],ymm4[22],ymm3[23],ymm4[23]
; AVX512F-NEXT: vpmullw %ymm13, %ymm4, %ymm4
; AVX512F-NEXT: vpsrlw $8, %ymm4, %ymm4
; AVX512F-NEXT: vpackuswb %ymm2, %ymm4, %ymm2
; AVX512F-NEXT: vpor %ymm2, %ymm1, %ymm1
; AVX512F-NEXT: vpblendvb %ymm10, %ymm1, %ymm3, %ymm1
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: constant_funnnel_v64i8:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpsllw $4, %ymm0, %ymm4
; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm5 = [240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240]
; AVX512VL-NEXT: vpand %ymm5, %ymm4, %ymm4
; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm6 = [57600,41152,24704,8256,8448,24640,41088,57536,57600,41152,24704,8256,8448,24640,41088,57536]
; AVX512VL-NEXT: vpblendvb %ymm6, %ymm4, %ymm0, %ymm0
; AVX512VL-NEXT: vpsllw $2, %ymm0, %ymm4
; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm7 = [252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252,252]
; AVX512VL-NEXT: vpand %ymm7, %ymm4, %ymm4
; AVX512VL-NEXT: vpaddb %ymm6, %ymm6, %ymm8
; AVX512VL-NEXT: vpblendvb %ymm8, %ymm4, %ymm0, %ymm0
; AVX512VL-NEXT: vpaddb %ymm0, %ymm0, %ymm4
; AVX512VL-NEXT: vpaddb %ymm8, %ymm8, %ymm9
; AVX512VL-NEXT: vpblendvb %ymm9, %ymm4, %ymm0, %ymm0
; AVX512VL-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
; AVX512VL-NEXT: vpsrlw $8, %ymm4, %ymm4
; AVX512VL-NEXT: vbroadcasti128 {{.*#+}} ymm10 = [256,2,4,8,16,32,64,128,256,2,4,8,16,32,64,128]
; AVX512VL-NEXT: # ymm10 = mem[0,1,0,1]
; AVX512VL-NEXT: vpmullw %ymm10, %ymm4, %ymm4
; AVX512VL-NEXT: vpsrlw $8, %ymm4, %ymm4
; AVX512VL-NEXT: vpunpcklbw {{.*#+}} ymm11 = ymm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
; AVX512VL-NEXT: vpsrlw $8, %ymm11, %ymm11
; AVX512VL-NEXT: vbroadcasti128 {{.*#+}} ymm12 = [256,128,64,32,16,8,4,2,256,128,64,32,16,8,4,2]
; AVX512VL-NEXT: # ymm12 = mem[0,1,0,1]
; AVX512VL-NEXT: vpmullw %ymm12, %ymm11, %ymm11
; AVX512VL-NEXT: vpsrlw $8, %ymm11, %ymm11
; AVX512VL-NEXT: vpackuswb %ymm4, %ymm11, %ymm4
; AVX512VL-NEXT: vpor %ymm4, %ymm0, %ymm0
; AVX512VL-NEXT: vpbroadcastq {{.*#+}} ymm4 = [18446744073709551360,18446744073709551360,18446744073709551360,18446744073709551360]
; AVX512VL-NEXT: vpblendvb %ymm4, %ymm0, %ymm2, %ymm0
; AVX512VL-NEXT: vpsllw $4, %ymm1, %ymm2
; AVX512VL-NEXT: vpand %ymm5, %ymm2, %ymm2
; AVX512VL-NEXT: vpblendvb %ymm6, %ymm2, %ymm1, %ymm1
; AVX512VL-NEXT: vpsllw $2, %ymm1, %ymm2
; AVX512VL-NEXT: vpand %ymm7, %ymm2, %ymm2
; AVX512VL-NEXT: vpblendvb %ymm8, %ymm2, %ymm1, %ymm1
; AVX512VL-NEXT: vpaddb %ymm1, %ymm1, %ymm2
; AVX512VL-NEXT: vpblendvb %ymm9, %ymm2, %ymm1, %ymm1
; AVX512VL-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
; AVX512VL-NEXT: vpsrlw $8, %ymm2, %ymm2
; AVX512VL-NEXT: vpmullw %ymm10, %ymm2, %ymm2
; AVX512VL-NEXT: vpsrlw $8, %ymm2, %ymm2
; AVX512VL-NEXT: vpunpcklbw {{.*#+}} ymm5 = ymm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
; AVX512VL-NEXT: vpsrlw $8, %ymm5, %ymm5
; AVX512VL-NEXT: vpmullw %ymm12, %ymm5, %ymm5
; AVX512VL-NEXT: vpsrlw $8, %ymm5, %ymm5
; AVX512VL-NEXT: vpackuswb %ymm2, %ymm5, %ymm2
; AVX512VL-NEXT: vpor %ymm2, %ymm1, %ymm1
; AVX512VL-NEXT: vpblendvb %ymm4, %ymm1, %ymm3, %ymm1
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: constant_funnnel_v64i8:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [57600,41152,24704,8256,8448,24640,41088,57536,57600,41152,24704,8256,8448,24640,41088,57536,57600,41152,24704,8256,8448,24640,41088,57536,57600,41152,24704,8256,8448,24640,41088,57536]
; AVX512BW-NEXT: vpmovb2m %zmm2, %k1
; AVX512BW-NEXT: vpsllw $4, %zmm0, %zmm3
; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm3, %zmm3
; AVX512BW-NEXT: vmovdqu8 %zmm3, %zmm0 {%k1}
; AVX512BW-NEXT: vpsllw $2, %zmm0, %zmm3
; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm3, %zmm3
; AVX512BW-NEXT: vpaddb %zmm2, %zmm2, %zmm2
; AVX512BW-NEXT: vpmovb2m %zmm2, %k1
; AVX512BW-NEXT: vmovdqu8 %zmm3, %zmm0 {%k1}
; AVX512BW-NEXT: vpaddb %zmm2, %zmm2, %zmm2
; AVX512BW-NEXT: vpmovb2m %zmm2, %k1
; AVX512BW-NEXT: vpaddb %zmm0, %zmm0, %zmm0 {%k1}
; AVX512BW-NEXT: vpunpckhbw {{.*#+}} zmm2 = zmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31,40,40,41,41,42,42,43,43,44,44,45,45,46,46,47,47,56,56,57,57,58,58,59,59,60,60,61,61,62,62,63,63]
; AVX512BW-NEXT: vpsrlw $8, %zmm2, %zmm2
; AVX512BW-NEXT: vpsllvw {{.*}}(%rip), %zmm2, %zmm2
; AVX512BW-NEXT: vpsrlw $8, %zmm2, %zmm2
; AVX512BW-NEXT: vpunpcklbw {{.*#+}} zmm3 = zmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23,32,32,33,33,34,34,35,35,36,36,37,37,38,38,39,39,48,48,49,49,50,50,51,51,52,52,53,53,54,54,55,55]
; AVX512BW-NEXT: vpsrlw $8, %zmm3, %zmm3
; AVX512BW-NEXT: vpsllvw {{.*}}(%rip), %zmm3, %zmm3
; AVX512BW-NEXT: vpsrlw $8, %zmm3, %zmm3
; AVX512BW-NEXT: vpackuswb %zmm2, %zmm3, %zmm2
; AVX512BW-NEXT: vporq %zmm2, %zmm0, %zmm0
; AVX512BW-NEXT: movabsq $72340172838076673, %rax # imm = 0x101010101010101
; AVX512BW-NEXT: kmovq %rax, %k1
; AVX512BW-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
; AVX512BW-NEXT: retq
;
; AVX512VBMI2-LABEL: constant_funnnel_v64i8:
; AVX512VBMI2: # %bb.0:
; AVX512VBMI2-NEXT: vmovdqa64 {{.*#+}} zmm2 = [57600,41152,24704,8256,8448,24640,41088,57536,57600,41152,24704,8256,8448,24640,41088,57536,57600,41152,24704,8256,8448,24640,41088,57536,57600,41152,24704,8256,8448,24640,41088,57536]
; AVX512VBMI2-NEXT: vpmovb2m %zmm2, %k1
; AVX512VBMI2-NEXT: vpsllw $4, %zmm0, %zmm3
; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm3, %zmm3
; AVX512VBMI2-NEXT: vmovdqu8 %zmm3, %zmm0 {%k1}
; AVX512VBMI2-NEXT: vpsllw $2, %zmm0, %zmm3
; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm3, %zmm3
; AVX512VBMI2-NEXT: vpaddb %zmm2, %zmm2, %zmm2
; AVX512VBMI2-NEXT: vpmovb2m %zmm2, %k1
; AVX512VBMI2-NEXT: vmovdqu8 %zmm3, %zmm0 {%k1}
; AVX512VBMI2-NEXT: vpaddb %zmm2, %zmm2, %zmm2
; AVX512VBMI2-NEXT: vpmovb2m %zmm2, %k1
; AVX512VBMI2-NEXT: vpaddb %zmm0, %zmm0, %zmm0 {%k1}
; AVX512VBMI2-NEXT: vpunpckhbw {{.*#+}} zmm2 = zmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31,40,40,41,41,42,42,43,43,44,44,45,45,46,46,47,47,56,56,57,57,58,58,59,59,60,60,61,61,62,62,63,63]
; AVX512VBMI2-NEXT: vpsrlw $8, %zmm2, %zmm2
; AVX512VBMI2-NEXT: vpsllvw {{.*}}(%rip), %zmm2, %zmm2
; AVX512VBMI2-NEXT: vpsrlw $8, %zmm2, %zmm2
; AVX512VBMI2-NEXT: vpunpcklbw {{.*#+}} zmm3 = zmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23,32,32,33,33,34,34,35,35,36,36,37,37,38,38,39,39,48,48,49,49,50,50,51,51,52,52,53,53,54,54,55,55]
; AVX512VBMI2-NEXT: vpsrlw $8, %zmm3, %zmm3
; AVX512VBMI2-NEXT: vpsllvw {{.*}}(%rip), %zmm3, %zmm3
; AVX512VBMI2-NEXT: vpsrlw $8, %zmm3, %zmm3
; AVX512VBMI2-NEXT: vpackuswb %zmm2, %zmm3, %zmm2
; AVX512VBMI2-NEXT: vporq %zmm2, %zmm0, %zmm0
; AVX512VBMI2-NEXT: movabsq $72340172838076673, %rax # imm = 0x101010101010101
; AVX512VBMI2-NEXT: kmovq %rax, %k1
; AVX512VBMI2-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
; AVX512VBMI2-NEXT: retq
;
; AVX512VLBW-LABEL: constant_funnnel_v64i8:
; AVX512VLBW: # %bb.0:
; AVX512VLBW-NEXT: vmovdqa64 {{.*#+}} zmm2 = [57600,41152,24704,8256,8448,24640,41088,57536,57600,41152,24704,8256,8448,24640,41088,57536,57600,41152,24704,8256,8448,24640,41088,57536,57600,41152,24704,8256,8448,24640,41088,57536]
; AVX512VLBW-NEXT: vpmovb2m %zmm2, %k1
; AVX512VLBW-NEXT: vpsllw $4, %zmm0, %zmm3
; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm3, %zmm3
; AVX512VLBW-NEXT: vmovdqu8 %zmm3, %zmm0 {%k1}
; AVX512VLBW-NEXT: vpsllw $2, %zmm0, %zmm3
; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm3, %zmm3
; AVX512VLBW-NEXT: vpaddb %zmm2, %zmm2, %zmm2
; AVX512VLBW-NEXT: vpmovb2m %zmm2, %k1
; AVX512VLBW-NEXT: vmovdqu8 %zmm3, %zmm0 {%k1}
; AVX512VLBW-NEXT: vpaddb %zmm2, %zmm2, %zmm2
; AVX512VLBW-NEXT: vpmovb2m %zmm2, %k1
; AVX512VLBW-NEXT: vpaddb %zmm0, %zmm0, %zmm0 {%k1}
; AVX512VLBW-NEXT: vpunpckhbw {{.*#+}} zmm2 = zmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31,40,40,41,41,42,42,43,43,44,44,45,45,46,46,47,47,56,56,57,57,58,58,59,59,60,60,61,61,62,62,63,63]
; AVX512VLBW-NEXT: vpsrlw $8, %zmm2, %zmm2
; AVX512VLBW-NEXT: vpsllvw {{.*}}(%rip), %zmm2, %zmm2
; AVX512VLBW-NEXT: vpsrlw $8, %zmm2, %zmm2
; AVX512VLBW-NEXT: vpunpcklbw {{.*#+}} zmm3 = zmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23,32,32,33,33,34,34,35,35,36,36,37,37,38,38,39,39,48,48,49,49,50,50,51,51,52,52,53,53,54,54,55,55]
; AVX512VLBW-NEXT: vpsrlw $8, %zmm3, %zmm3
; AVX512VLBW-NEXT: vpsllvw {{.*}}(%rip), %zmm3, %zmm3
; AVX512VLBW-NEXT: vpsrlw $8, %zmm3, %zmm3
; AVX512VLBW-NEXT: vpackuswb %zmm2, %zmm3, %zmm2
; AVX512VLBW-NEXT: vporq %zmm2, %zmm0, %zmm0
; AVX512VLBW-NEXT: movabsq $72340172838076673, %rax # imm = 0x101010101010101
; AVX512VLBW-NEXT: kmovq %rax, %k1
; AVX512VLBW-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
; AVX512VLBW-NEXT: retq
;
; AVX512VLVBMI2-LABEL: constant_funnnel_v64i8:
; AVX512VLVBMI2: # %bb.0:
; AVX512VLVBMI2-NEXT: vmovdqa64 {{.*#+}} zmm2 = [57600,41152,24704,8256,8448,24640,41088,57536,57600,41152,24704,8256,8448,24640,41088,57536,57600,41152,24704,8256,8448,24640,41088,57536,57600,41152,24704,8256,8448,24640,41088,57536]
; AVX512VLVBMI2-NEXT: vpmovb2m %zmm2, %k1
; AVX512VLVBMI2-NEXT: vpsllw $4, %zmm0, %zmm3
; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm3, %zmm3
; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm3, %zmm0 {%k1}
; AVX512VLVBMI2-NEXT: vpsllw $2, %zmm0, %zmm3
; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm3, %zmm3
; AVX512VLVBMI2-NEXT: vpaddb %zmm2, %zmm2, %zmm2
; AVX512VLVBMI2-NEXT: vpmovb2m %zmm2, %k1
; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm3, %zmm0 {%k1}
; AVX512VLVBMI2-NEXT: vpaddb %zmm2, %zmm2, %zmm2
; AVX512VLVBMI2-NEXT: vpmovb2m %zmm2, %k1
; AVX512VLVBMI2-NEXT: vpaddb %zmm0, %zmm0, %zmm0 {%k1}
; AVX512VLVBMI2-NEXT: vpunpckhbw {{.*#+}} zmm2 = zmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31,40,40,41,41,42,42,43,43,44,44,45,45,46,46,47,47,56,56,57,57,58,58,59,59,60,60,61,61,62,62,63,63]
; AVX512VLVBMI2-NEXT: vpsrlw $8, %zmm2, %zmm2
; AVX512VLVBMI2-NEXT: vpsllvw {{.*}}(%rip), %zmm2, %zmm2
; AVX512VLVBMI2-NEXT: vpsrlw $8, %zmm2, %zmm2
; AVX512VLVBMI2-NEXT: vpunpcklbw {{.*#+}} zmm3 = zmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23,32,32,33,33,34,34,35,35,36,36,37,37,38,38,39,39,48,48,49,49,50,50,51,51,52,52,53,53,54,54,55,55]
; AVX512VLVBMI2-NEXT: vpsrlw $8, %zmm3, %zmm3
; AVX512VLVBMI2-NEXT: vpsllvw {{.*}}(%rip), %zmm3, %zmm3
; AVX512VLVBMI2-NEXT: vpsrlw $8, %zmm3, %zmm3
; AVX512VLVBMI2-NEXT: vpackuswb %zmm2, %zmm3, %zmm2
; AVX512VLVBMI2-NEXT: vporq %zmm2, %zmm0, %zmm0
; AVX512VLVBMI2-NEXT: movabsq $72340172838076673, %rax # imm = 0x101010101010101
; AVX512VLVBMI2-NEXT: kmovq %rax, %k1
; AVX512VLVBMI2-NEXT: vmovdqu8 %zmm1, %zmm0 {%k1}
; AVX512VLVBMI2-NEXT: retq
%res = call <64 x i8> @llvm.fshr.v64i8(<64 x i8> %x, <64 x i8> %y, <64 x i8> <i8 0, i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 8, i8 7, i8 6, i8 5, i8 4, i8 3, i8 2, i8 1, i8 0, i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 8, i8 7, i8 6, i8 5, i8 4, i8 3, i8 2, i8 1, i8 0, i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 8, i8 7, i8 6, i8 5, i8 4, i8 3, i8 2, i8 1, i8 0, i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 8, i8 7, i8 6, i8 5, i8 4, i8 3, i8 2, i8 1>)
ret <64 x i8> %res
}
;
; Uniform Constant Shifts
;
define <8 x i64> @splatconstant_funnnel_v8i64(<8 x i64> %x, <8 x i64> %y) nounwind {
; AVX512F-LABEL: splatconstant_funnnel_v8i64:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpsrlq $14, %zmm1, %zmm1
; AVX512F-NEXT: vpsllq $50, %zmm0, %zmm0
; AVX512F-NEXT: vporq %zmm1, %zmm0, %zmm0
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: splatconstant_funnnel_v8i64:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpsrlq $14, %zmm1, %zmm1
; AVX512VL-NEXT: vpsllq $50, %zmm0, %zmm0
; AVX512VL-NEXT: vporq %zmm1, %zmm0, %zmm0
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: splatconstant_funnnel_v8i64:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpsrlq $14, %zmm1, %zmm1
; AVX512BW-NEXT: vpsllq $50, %zmm0, %zmm0
; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0
; AVX512BW-NEXT: retq
;
; AVX512VBMI2-LABEL: splatconstant_funnnel_v8i64:
; AVX512VBMI2: # %bb.0:
; AVX512VBMI2-NEXT: vpshrdq $14, %zmm0, %zmm1, %zmm0
; AVX512VBMI2-NEXT: retq
;
; AVX512VLBW-LABEL: splatconstant_funnnel_v8i64:
; AVX512VLBW: # %bb.0:
; AVX512VLBW-NEXT: vpsrlq $14, %zmm1, %zmm1
; AVX512VLBW-NEXT: vpsllq $50, %zmm0, %zmm0
; AVX512VLBW-NEXT: vporq %zmm1, %zmm0, %zmm0
; AVX512VLBW-NEXT: retq
;
; AVX512VLVBMI2-LABEL: splatconstant_funnnel_v8i64:
; AVX512VLVBMI2: # %bb.0:
; AVX512VLVBMI2-NEXT: vpshrdq $14, %zmm0, %zmm1, %zmm0
; AVX512VLVBMI2-NEXT: retq
%res = call <8 x i64> @llvm.fshr.v8i64(<8 x i64> %x, <8 x i64> %y, <8 x i64> <i64 14, i64 14, i64 14, i64 14, i64 14, i64 14, i64 14, i64 14>)
ret <8 x i64> %res
}
define <16 x i32> @splatconstant_funnnel_v16i32(<16 x i32> %x, <16 x i32> %y) nounwind {
; AVX512F-LABEL: splatconstant_funnnel_v16i32:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpsrld $4, %zmm1, %zmm1
; AVX512F-NEXT: vpslld $28, %zmm0, %zmm0
; AVX512F-NEXT: vpord %zmm1, %zmm0, %zmm0
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: splatconstant_funnnel_v16i32:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpsrld $4, %zmm1, %zmm1
; AVX512VL-NEXT: vpslld $28, %zmm0, %zmm0
; AVX512VL-NEXT: vpord %zmm1, %zmm0, %zmm0
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: splatconstant_funnnel_v16i32:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpsrld $4, %zmm1, %zmm1
; AVX512BW-NEXT: vpslld $28, %zmm0, %zmm0
; AVX512BW-NEXT: vpord %zmm1, %zmm0, %zmm0
; AVX512BW-NEXT: retq
;
; AVX512VBMI2-LABEL: splatconstant_funnnel_v16i32:
; AVX512VBMI2: # %bb.0:
; AVX512VBMI2-NEXT: vpshrdd $4, %zmm0, %zmm1, %zmm0
; AVX512VBMI2-NEXT: retq
;
; AVX512VLBW-LABEL: splatconstant_funnnel_v16i32:
; AVX512VLBW: # %bb.0:
; AVX512VLBW-NEXT: vpsrld $4, %zmm1, %zmm1
; AVX512VLBW-NEXT: vpslld $28, %zmm0, %zmm0
; AVX512VLBW-NEXT: vpord %zmm1, %zmm0, %zmm0
; AVX512VLBW-NEXT: retq
;
; AVX512VLVBMI2-LABEL: splatconstant_funnnel_v16i32:
; AVX512VLVBMI2: # %bb.0:
; AVX512VLVBMI2-NEXT: vpshrdd $4, %zmm0, %zmm1, %zmm0
; AVX512VLVBMI2-NEXT: retq
%res = call <16 x i32> @llvm.fshr.v16i32(<16 x i32> %x, <16 x i32> %y, <16 x i32> <i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4>)
ret <16 x i32> %res
}
define <32 x i16> @splatconstant_funnnel_v32i16(<32 x i16> %x, <32 x i16> %y) nounwind {
; AVX512F-LABEL: splatconstant_funnnel_v32i16:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpsrlw $7, %ymm2, %ymm2
; AVX512F-NEXT: vpsllw $9, %ymm0, %ymm0
; AVX512F-NEXT: vpor %ymm2, %ymm0, %ymm0
; AVX512F-NEXT: vpsrlw $7, %ymm3, %ymm2
; AVX512F-NEXT: vpsllw $9, %ymm1, %ymm1
; AVX512F-NEXT: vpor %ymm2, %ymm1, %ymm1
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: splatconstant_funnnel_v32i16:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpsrlw $7, %ymm2, %ymm2
; AVX512VL-NEXT: vpsllw $9, %ymm0, %ymm0
; AVX512VL-NEXT: vpor %ymm2, %ymm0, %ymm0
; AVX512VL-NEXT: vpsrlw $7, %ymm3, %ymm2
; AVX512VL-NEXT: vpsllw $9, %ymm1, %ymm1
; AVX512VL-NEXT: vpor %ymm2, %ymm1, %ymm1
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: splatconstant_funnnel_v32i16:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpsrlw $7, %zmm1, %zmm1
; AVX512BW-NEXT: vpsllw $9, %zmm0, %zmm0
; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0
; AVX512BW-NEXT: retq
;
; AVX512VBMI2-LABEL: splatconstant_funnnel_v32i16:
; AVX512VBMI2: # %bb.0:
; AVX512VBMI2-NEXT: vpshrdw $7, %zmm0, %zmm1, %zmm0
; AVX512VBMI2-NEXT: retq
;
; AVX512VLBW-LABEL: splatconstant_funnnel_v32i16:
; AVX512VLBW: # %bb.0:
; AVX512VLBW-NEXT: vpsrlw $7, %zmm1, %zmm1
; AVX512VLBW-NEXT: vpsllw $9, %zmm0, %zmm0
; AVX512VLBW-NEXT: vporq %zmm1, %zmm0, %zmm0
; AVX512VLBW-NEXT: retq
;
; AVX512VLVBMI2-LABEL: splatconstant_funnnel_v32i16:
; AVX512VLVBMI2: # %bb.0:
; AVX512VLVBMI2-NEXT: vpshrdw $7, %zmm0, %zmm1, %zmm0
; AVX512VLVBMI2-NEXT: retq
%res = call <32 x i16> @llvm.fshr.v32i16(<32 x i16> %x, <32 x i16> %y, <32 x i16> <i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7>)
ret <32 x i16> %res
}
define <64 x i8> @splatconstant_funnnel_v64i8(<64 x i8> %x, <64 x i8> %y) nounwind {
; AVX512F-LABEL: splatconstant_funnnel_v64i8:
; AVX512F: # %bb.0:
; AVX512F-NEXT: vpsrlw $4, %ymm2, %ymm2
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX512F-NEXT: vpand %ymm4, %ymm2, %ymm2
; AVX512F-NEXT: vpsllw $4, %ymm0, %ymm0
; AVX512F-NEXT: vmovdqa {{.*#+}} ymm5 = [240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240]
; AVX512F-NEXT: vpand %ymm5, %ymm0, %ymm0
; AVX512F-NEXT: vpor %ymm2, %ymm0, %ymm0
; AVX512F-NEXT: vpsrlw $4, %ymm3, %ymm2
; AVX512F-NEXT: vpand %ymm4, %ymm2, %ymm2
; AVX512F-NEXT: vpsllw $4, %ymm1, %ymm1
; AVX512F-NEXT: vpand %ymm5, %ymm1, %ymm1
; AVX512F-NEXT: vpor %ymm2, %ymm1, %ymm1
; AVX512F-NEXT: retq
;
; AVX512VL-LABEL: splatconstant_funnnel_v64i8:
; AVX512VL: # %bb.0:
; AVX512VL-NEXT: vpsrlw $4, %ymm2, %ymm2
; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
; AVX512VL-NEXT: vpand %ymm4, %ymm2, %ymm2
; AVX512VL-NEXT: vpsllw $4, %ymm0, %ymm0
; AVX512VL-NEXT: vmovdqa {{.*#+}} ymm5 = [240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240,240]
; AVX512VL-NEXT: vpand %ymm5, %ymm0, %ymm0
; AVX512VL-NEXT: vpor %ymm2, %ymm0, %ymm0
; AVX512VL-NEXT: vpsrlw $4, %ymm3, %ymm2
; AVX512VL-NEXT: vpand %ymm4, %ymm2, %ymm2
; AVX512VL-NEXT: vpsllw $4, %ymm1, %ymm1
; AVX512VL-NEXT: vpand %ymm5, %ymm1, %ymm1
; AVX512VL-NEXT: vpor %ymm2, %ymm1, %ymm1
; AVX512VL-NEXT: retq
;
; AVX512BW-LABEL: splatconstant_funnnel_v64i8:
; AVX512BW: # %bb.0:
; AVX512BW-NEXT: vpsrlw $4, %zmm1, %zmm1
; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm1, %zmm1
; AVX512BW-NEXT: vpsllw $4, %zmm0, %zmm0
; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm0, %zmm0
; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0
; AVX512BW-NEXT: retq
;
; AVX512VBMI2-LABEL: splatconstant_funnnel_v64i8:
; AVX512VBMI2: # %bb.0:
; AVX512VBMI2-NEXT: vpsrlw $4, %zmm1, %zmm1
; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm1, %zmm1
; AVX512VBMI2-NEXT: vpsllw $4, %zmm0, %zmm0
; AVX512VBMI2-NEXT: vpandq {{.*}}(%rip), %zmm0, %zmm0
; AVX512VBMI2-NEXT: vporq %zmm1, %zmm0, %zmm0
; AVX512VBMI2-NEXT: retq
;
; AVX512VLBW-LABEL: splatconstant_funnnel_v64i8:
; AVX512VLBW: # %bb.0:
; AVX512VLBW-NEXT: vpsrlw $4, %zmm1, %zmm1
; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm1, %zmm1
; AVX512VLBW-NEXT: vpsllw $4, %zmm0, %zmm0
; AVX512VLBW-NEXT: vpandq {{.*}}(%rip), %zmm0, %zmm0
; AVX512VLBW-NEXT: vporq %zmm1, %zmm0, %zmm0
; AVX512VLBW-NEXT: retq
;
; AVX512VLVBMI2-LABEL: splatconstant_funnnel_v64i8:
; AVX512VLVBMI2: # %bb.0:
; AVX512VLVBMI2-NEXT: vpsrlw $4, %zmm1, %zmm1
; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm1, %zmm1
; AVX512VLVBMI2-NEXT: vpsllw $4, %zmm0, %zmm0
; AVX512VLVBMI2-NEXT: vpandq {{.*}}(%rip), %zmm0, %zmm0
; AVX512VLVBMI2-NEXT: vporq %zmm1, %zmm0, %zmm0
; AVX512VLVBMI2-NEXT: retq
%res = call <64 x i8> @llvm.fshr.v64i8(<64 x i8> %x, <64 x i8> %y, <64 x i8> <i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4, i8 4>)
ret <64 x i8> %res
}