170 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512vbmi --show-mc-encoding | FileCheck %s --check-prefixes=CHECK,X863; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vbmi --show-mc-encoding | FileCheck %s --check-prefixes=CHECK,X644 5declare <64 x i8> @llvm.x86.avx512.mask.permvar.qi.512(<64 x i8>, <64 x i8>, <64 x i8>, i64)6 7define <64 x i8>@test_int_x86_avx512_permvar_qi_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2) {8; CHECK-LABEL: test_int_x86_avx512_permvar_qi_512:9; CHECK: # %bb.0:10; CHECK-NEXT: vpermb %zmm0, %zmm1, %zmm0 # encoding: [0x62,0xf2,0x75,0x48,0x8d,0xc0]11; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]12 %res = call <64 x i8> @llvm.x86.avx512.mask.permvar.qi.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 -1)13 ret <64 x i8> %res14}15 16define <64 x i8>@test_int_x86_avx512_mask_permvar_qi_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3) {17; X86-LABEL: test_int_x86_avx512_mask_permvar_qi_512:18; X86: # %bb.0:19; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x04]20; X86-NEXT: vpermb %zmm0, %zmm1, %zmm2 {%k1} # encoding: [0x62,0xf2,0x75,0x49,0x8d,0xd0]21; X86-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]22; X86-NEXT: retl # encoding: [0xc3]23;24; X64-LABEL: test_int_x86_avx512_mask_permvar_qi_512:25; X64: # %bb.0:26; X64-NEXT: kmovq %rdi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xcf]27; X64-NEXT: vpermb %zmm0, %zmm1, %zmm2 {%k1} # encoding: [0x62,0xf2,0x75,0x49,0x8d,0xd0]28; X64-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]29; X64-NEXT: retq # encoding: [0xc3]30 %res = call <64 x i8> @llvm.x86.avx512.mask.permvar.qi.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3)31 ret <64 x i8> %res32}33 34define <64 x i8>@test_int_x86_avx512_maskz_permvar_qi_512(<64 x i8> %x0, <64 x i8> %x1, i64 %x3) {35; X86-LABEL: test_int_x86_avx512_maskz_permvar_qi_512:36; X86: # %bb.0:37; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x04]38; X86-NEXT: vpermb %zmm0, %zmm1, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0xc9,0x8d,0xc0]39; X86-NEXT: retl # encoding: [0xc3]40;41; X64-LABEL: test_int_x86_avx512_maskz_permvar_qi_512:42; X64: # %bb.0:43; X64-NEXT: kmovq %rdi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xcf]44; X64-NEXT: vpermb %zmm0, %zmm1, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0xc9,0x8d,0xc0]45; X64-NEXT: retq # encoding: [0xc3]46 %res = call <64 x i8> @llvm.x86.avx512.mask.permvar.qi.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> zeroinitializer, i64 %x3)47 ret <64 x i8> %res48}49 50declare <64 x i8> @llvm.x86.avx512.mask.pmultishift.qb.512(<64 x i8>, <64 x i8>, <64 x i8>, i64)51 52define <64 x i8>@test_int_x86_avx512_pmultishift_qb_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2) {53; CHECK-LABEL: test_int_x86_avx512_pmultishift_qb_512:54; CHECK: # %bb.0:55; CHECK-NEXT: vpmultishiftqb %zmm1, %zmm0, %zmm0 # encoding: [0x62,0xf2,0xfd,0x48,0x83,0xc1]56; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]57 %res = call <64 x i8> @llvm.x86.avx512.mask.pmultishift.qb.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 -1)58 ret <64 x i8> %res59}60 61define <64 x i8>@test_int_x86_avx512_mask_pmultishift_qb_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3) {62; X86-LABEL: test_int_x86_avx512_mask_pmultishift_qb_512:63; X86: # %bb.0:64; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x04]65; X86-NEXT: vpmultishiftqb %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x49,0x83,0xd1]66; X86-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]67; X86-NEXT: retl # encoding: [0xc3]68;69; X64-LABEL: test_int_x86_avx512_mask_pmultishift_qb_512:70; X64: # %bb.0:71; X64-NEXT: kmovq %rdi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xcf]72; X64-NEXT: vpmultishiftqb %zmm1, %zmm0, %zmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x49,0x83,0xd1]73; X64-NEXT: vmovdqa64 %zmm2, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc2]74; X64-NEXT: retq # encoding: [0xc3]75 %res = call <64 x i8> @llvm.x86.avx512.mask.pmultishift.qb.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3)76 ret <64 x i8> %res77}78 79define <64 x i8>@test_int_x86_avx512_maskz_pmultishift_qb_512(<64 x i8> %x0, <64 x i8> %x1, i64 %x3) {80; X86-LABEL: test_int_x86_avx512_maskz_pmultishift_qb_512:81; X86: # %bb.0:82; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x04]83; X86-NEXT: vpmultishiftqb %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xc9,0x83,0xc1]84; X86-NEXT: retl # encoding: [0xc3]85;86; X64-LABEL: test_int_x86_avx512_maskz_pmultishift_qb_512:87; X64: # %bb.0:88; X64-NEXT: kmovq %rdi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xcf]89; X64-NEXT: vpmultishiftqb %zmm1, %zmm0, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0xfd,0xc9,0x83,0xc1]90; X64-NEXT: retq # encoding: [0xc3]91 %res = call <64 x i8> @llvm.x86.avx512.mask.pmultishift.qb.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> zeroinitializer, i64 %x3)92 ret <64 x i8> %res93}94 95declare <64 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.512(<64 x i8>, <64 x i8>, <64 x i8>, i64)96 97define <64 x i8>@test_int_x86_avx512_vpermi2var_qi_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2) {98; CHECK-LABEL: test_int_x86_avx512_vpermi2var_qi_512:99; CHECK: # %bb.0:100; CHECK-NEXT: vpermt2b %zmm2, %zmm1, %zmm0 # encoding: [0x62,0xf2,0x75,0x48,0x7d,0xc2]101; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]102 %res = call <64 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 -1)103 ret <64 x i8> %res104}105 106define <64 x i8>@test_int_x86_avx512_mask_vpermi2var_qi_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3) {107; X86-LABEL: test_int_x86_avx512_mask_vpermi2var_qi_512:108; X86: # %bb.0:109; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x04]110; X86-NEXT: vpermi2b %zmm2, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x75,0xca]111; X86-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]112; X86-NEXT: retl # encoding: [0xc3]113;114; X64-LABEL: test_int_x86_avx512_mask_vpermi2var_qi_512:115; X64: # %bb.0:116; X64-NEXT: kmovq %rdi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xcf]117; X64-NEXT: vpermi2b %zmm2, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x75,0xca]118; X64-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]119; X64-NEXT: retq # encoding: [0xc3]120 %res = call <64 x i8> @llvm.x86.avx512.mask.vpermi2var.qi.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3)121 ret <64 x i8> %res122}123 124declare <64 x i8> @llvm.x86.avx512.mask.vpermt2var.qi.512(<64 x i8>, <64 x i8>, <64 x i8>, i64)125 126define <64 x i8>@test_int_x86_avx512_vpermt2var_qi_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2) {127; CHECK-LABEL: test_int_x86_avx512_vpermt2var_qi_512:128; CHECK: # %bb.0:129; CHECK-NEXT: vpermi2b %zmm2, %zmm1, %zmm0 # encoding: [0x62,0xf2,0x75,0x48,0x75,0xc2]130; CHECK-NEXT: ret{{[l|q]}} # encoding: [0xc3]131 %res = call <64 x i8> @llvm.x86.avx512.mask.vpermt2var.qi.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 -1)132 ret <64 x i8> %res133}134 135define <64 x i8>@test_int_x86_avx512_mask_vpermt2var_qi_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3) {136; X86-LABEL: test_int_x86_avx512_mask_vpermt2var_qi_512:137; X86: # %bb.0:138; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x04]139; X86-NEXT: vpermt2b %zmm2, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x7d,0xca]140; X86-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]141; X86-NEXT: retl # encoding: [0xc3]142;143; X64-LABEL: test_int_x86_avx512_mask_vpermt2var_qi_512:144; X64: # %bb.0:145; X64-NEXT: kmovq %rdi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xcf]146; X64-NEXT: vpermt2b %zmm2, %zmm0, %zmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x49,0x7d,0xca]147; X64-NEXT: vmovdqa64 %zmm1, %zmm0 # encoding: [0x62,0xf1,0xfd,0x48,0x6f,0xc1]148; X64-NEXT: retq # encoding: [0xc3]149 %res = call <64 x i8> @llvm.x86.avx512.mask.vpermt2var.qi.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3)150 ret <64 x i8> %res151}152 153declare <64 x i8> @llvm.x86.avx512.maskz.vpermt2var.qi.512(<64 x i8>, <64 x i8>, <64 x i8>, i64)154 155define <64 x i8>@test_int_x86_avx512_maskz_vpermt2var_qi_512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3) {156; X86-LABEL: test_int_x86_avx512_maskz_vpermt2var_qi_512:157; X86: # %bb.0:158; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x04]159; X86-NEXT: vpermi2b %zmm2, %zmm1, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0xc9,0x75,0xc2]160; X86-NEXT: retl # encoding: [0xc3]161;162; X64-LABEL: test_int_x86_avx512_maskz_vpermt2var_qi_512:163; X64: # %bb.0:164; X64-NEXT: kmovq %rdi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xcf]165; X64-NEXT: vpermi2b %zmm2, %zmm1, %zmm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0xc9,0x75,0xc2]166; X64-NEXT: retq # encoding: [0xc3]167 %res = call <64 x i8> @llvm.x86.avx512.maskz.vpermt2var.qi.512(<64 x i8> %x0, <64 x i8> %x1, <64 x i8> %x2, i64 %x3)168 ret <64 x i8> %res169}170