516 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 52; RUN: opt < %s -mattr=+avx512vnni -passes=msan -S | FileCheck %s3 4; Forked from llvm/test/CodeGen/X86/avx512vnni-intrinsics-upgrade.ll5;6; Handled strictly: (none)7;8; Handled heuristically: (none)9 10target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"11target triple = "x86_64-unknown-linux-gnu"12 13declare <16 x i32> @llvm.x86.avx512.mask.vpdpbusd.512(<16 x i32>, <16 x i32>, <16 x i32>, i16)14declare <16 x i32> @llvm.x86.avx512.maskz.vpdpbusd.512(<16 x i32>, <16 x i32>, <16 x i32>, i16)15 16define <16 x i32>@test_int_x86_avx512_vpdpbusd_512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2) sanitize_memory {17; CHECK-LABEL: define <16 x i32> @test_int_x86_avx512_vpdpbusd_512(18; CHECK-SAME: <16 x i32> [[X0:%.*]], <16 x i32> [[X1:%.*]], <16 x i32> [[X2:%.*]]) #[[ATTR0:[0-9]+]] {19; CHECK-NEXT: [[TMP2:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 64), align 820; CHECK-NEXT: [[TMP3:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 128), align 821; CHECK-NEXT: [[TMP23:%.*]] = load <16 x i32>, ptr @__msan_param_tls, align 822; CHECK-NEXT: call void @llvm.donothing()23; CHECK-NEXT: [[TMP6:%.*]] = bitcast <16 x i32> [[TMP2]] to <64 x i8>24; CHECK-NEXT: [[TMP24:%.*]] = bitcast <16 x i32> [[X1]] to <64 x i8>25; CHECK-NEXT: [[TMP7:%.*]] = bitcast <16 x i32> [[TMP3]] to <64 x i8>26; CHECK-NEXT: [[TMP5:%.*]] = bitcast <16 x i32> [[X2]] to <64 x i8>27; CHECK-NEXT: [[TMP8:%.*]] = icmp ne <64 x i8> [[TMP6]], zeroinitializer28; CHECK-NEXT: [[TMP9:%.*]] = icmp ne <64 x i8> [[TMP7]], zeroinitializer29; CHECK-NEXT: [[TMP10:%.*]] = icmp ne <64 x i8> [[TMP24]], zeroinitializer30; CHECK-NEXT: [[TMP11:%.*]] = icmp ne <64 x i8> [[TMP5]], zeroinitializer31; CHECK-NEXT: [[TMP12:%.*]] = and <64 x i1> [[TMP8]], [[TMP9]]32; CHECK-NEXT: [[TMP13:%.*]] = and <64 x i1> [[TMP10]], [[TMP9]]33; CHECK-NEXT: [[TMP14:%.*]] = and <64 x i1> [[TMP8]], [[TMP11]]34; CHECK-NEXT: [[TMP15:%.*]] = or <64 x i1> [[TMP12]], [[TMP13]]35; CHECK-NEXT: [[TMP16:%.*]] = or <64 x i1> [[TMP15]], [[TMP14]]36; CHECK-NEXT: [[TMP17:%.*]] = sext <64 x i1> [[TMP16]] to <64 x i8>37; CHECK-NEXT: [[TMP18:%.*]] = bitcast <64 x i8> [[TMP17]] to <16 x i32>38; CHECK-NEXT: [[TMP19:%.*]] = icmp ne <16 x i32> [[TMP18]], zeroinitializer39; CHECK-NEXT: [[TMP22:%.*]] = sext <16 x i1> [[TMP19]] to <16 x i32>40; CHECK-NEXT: [[_MSPROP1:%.*]] = or <16 x i32> [[TMP22]], [[TMP23]]41; CHECK-NEXT: [[TMP4:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpbusd.512(<16 x i32> [[X0]], <64 x i8> [[TMP24]], <64 x i8> [[TMP5]])42; CHECK-NEXT: store <16 x i32> [[_MSPROP1]], ptr @__msan_retval_tls, align 843; CHECK-NEXT: ret <16 x i32> [[TMP4]]44;45 %res = call <16 x i32> @llvm.x86.avx512.mask.vpdpbusd.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2, i16 -1)46 ret <16 x i32> %res47}48 49define { <16 x i32>, <16 x i32> } @test_int_x86_avx512_mask_vpdpbusd_512(<16 x i32> %x0, <16 x i32> %x1, ptr %x2p, <16 x i32> %x4, i16 %x3) sanitize_memory {50; CHECK-LABEL: define { <16 x i32>, <16 x i32> } @test_int_x86_avx512_mask_vpdpbusd_512(51; CHECK-SAME: <16 x i32> [[X0:%.*]], <16 x i32> [[X1:%.*]], ptr [[X2P:%.*]], <16 x i32> [[X4:%.*]], i16 [[X3:%.*]]) #[[ATTR0]] {52; CHECK-NEXT: [[TMP1:%.*]] = load i64, ptr getelementptr (i8, ptr @__msan_param_tls, i64 128), align 853; CHECK-NEXT: [[TMP3:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 64), align 854; CHECK-NEXT: [[TMP2:%.*]] = load <16 x i32>, ptr @__msan_param_tls, align 855; CHECK-NEXT: [[TMP4:%.*]] = load i16, ptr getelementptr (i8, ptr @__msan_param_tls, i64 200), align 856; CHECK-NEXT: [[TMP5:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 136), align 857; CHECK-NEXT: call void @llvm.donothing()58; CHECK-NEXT: [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 059; CHECK-NEXT: br i1 [[_MSCMP]], label %[[BB6:.*]], label %[[BB7:.*]], !prof [[PROF1:![0-9]+]]60; CHECK: [[BB6]]:61; CHECK-NEXT: call void @__msan_warning_noreturn() #[[ATTR4:[0-9]+]]62; CHECK-NEXT: unreachable63; CHECK: [[BB7]]:64; CHECK-NEXT: [[X2:%.*]] = load <16 x i32>, ptr [[X2P]], align 6465; CHECK-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[X2P]] to i6466; CHECK-NEXT: [[TMP9:%.*]] = xor i64 [[TMP8]], 8796093022208067; CHECK-NEXT: [[TMP10:%.*]] = inttoptr i64 [[TMP9]] to ptr68; CHECK-NEXT: [[_MSLD:%.*]] = load <16 x i32>, ptr [[TMP10]], align 6469; CHECK-NEXT: [[TMP32:%.*]] = bitcast <16 x i32> [[TMP3]] to <64 x i8>70; CHECK-NEXT: [[TMP30:%.*]] = bitcast <16 x i32> [[X1]] to <64 x i8>71; CHECK-NEXT: [[TMP33:%.*]] = bitcast <16 x i32> [[_MSLD]] to <64 x i8>72; CHECK-NEXT: [[TMP31:%.*]] = bitcast <16 x i32> [[X2]] to <64 x i8>73; CHECK-NEXT: [[TMP34:%.*]] = icmp ne <64 x i8> [[TMP32]], zeroinitializer74; CHECK-NEXT: [[TMP35:%.*]] = icmp ne <64 x i8> [[TMP33]], zeroinitializer75; CHECK-NEXT: [[TMP36:%.*]] = icmp ne <64 x i8> [[TMP30]], zeroinitializer76; CHECK-NEXT: [[TMP37:%.*]] = icmp ne <64 x i8> [[TMP31]], zeroinitializer77; CHECK-NEXT: [[TMP38:%.*]] = and <64 x i1> [[TMP34]], [[TMP35]]78; CHECK-NEXT: [[TMP58:%.*]] = and <64 x i1> [[TMP36]], [[TMP35]]79; CHECK-NEXT: [[TMP59:%.*]] = and <64 x i1> [[TMP34]], [[TMP37]]80; CHECK-NEXT: [[TMP60:%.*]] = or <64 x i1> [[TMP38]], [[TMP58]]81; CHECK-NEXT: [[TMP61:%.*]] = or <64 x i1> [[TMP60]], [[TMP59]]82; CHECK-NEXT: [[TMP62:%.*]] = sext <64 x i1> [[TMP61]] to <64 x i8>83; CHECK-NEXT: [[TMP53:%.*]] = bitcast <64 x i8> [[TMP62]] to <16 x i32>84; CHECK-NEXT: [[TMP54:%.*]] = icmp ne <16 x i32> [[TMP53]], zeroinitializer85; CHECK-NEXT: [[TMP29:%.*]] = sext <16 x i1> [[TMP54]] to <16 x i32>86; CHECK-NEXT: [[_MSPROP1:%.*]] = or <16 x i32> [[TMP29]], [[TMP2]]87; CHECK-NEXT: [[TMP11:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpbusd.512(<16 x i32> [[X0]], <64 x i8> [[TMP30]], <64 x i8> [[TMP31]])88; CHECK-NEXT: [[TMP12:%.*]] = bitcast i16 [[TMP4]] to <16 x i1>89; CHECK-NEXT: [[TMP13:%.*]] = bitcast i16 [[X3]] to <16 x i1>90; CHECK-NEXT: [[TMP14:%.*]] = select <16 x i1> [[TMP13]], <16 x i32> [[_MSPROP1]], <16 x i32> [[TMP2]]91; CHECK-NEXT: [[TMP15:%.*]] = xor <16 x i32> [[TMP11]], [[X0]]92; CHECK-NEXT: [[TMP16:%.*]] = or <16 x i32> [[TMP15]], [[_MSPROP1]]93; CHECK-NEXT: [[TMP17:%.*]] = or <16 x i32> [[TMP16]], [[TMP2]]94; CHECK-NEXT: [[_MSPROP_SELECT:%.*]] = select <16 x i1> [[TMP12]], <16 x i32> [[TMP17]], <16 x i32> [[TMP14]]95; CHECK-NEXT: [[TMP18:%.*]] = select <16 x i1> [[TMP13]], <16 x i32> [[TMP11]], <16 x i32> [[X0]]96; CHECK-NEXT: [[TMP41:%.*]] = bitcast <16 x i32> [[TMP3]] to <64 x i8>97; CHECK-NEXT: [[TMP39:%.*]] = bitcast <16 x i32> [[X1]] to <64 x i8>98; CHECK-NEXT: [[TMP42:%.*]] = bitcast <16 x i32> [[TMP5]] to <64 x i8>99; CHECK-NEXT: [[TMP40:%.*]] = bitcast <16 x i32> [[X4]] to <64 x i8>100; CHECK-NEXT: [[TMP43:%.*]] = icmp ne <64 x i8> [[TMP41]], zeroinitializer101; CHECK-NEXT: [[TMP44:%.*]] = icmp ne <64 x i8> [[TMP42]], zeroinitializer102; CHECK-NEXT: [[TMP45:%.*]] = icmp ne <64 x i8> [[TMP39]], zeroinitializer103; CHECK-NEXT: [[TMP46:%.*]] = icmp ne <64 x i8> [[TMP40]], zeroinitializer104; CHECK-NEXT: [[TMP47:%.*]] = and <64 x i1> [[TMP43]], [[TMP44]]105; CHECK-NEXT: [[TMP48:%.*]] = and <64 x i1> [[TMP45]], [[TMP44]]106; CHECK-NEXT: [[TMP49:%.*]] = and <64 x i1> [[TMP43]], [[TMP46]]107; CHECK-NEXT: [[TMP50:%.*]] = or <64 x i1> [[TMP47]], [[TMP48]]108; CHECK-NEXT: [[TMP51:%.*]] = or <64 x i1> [[TMP50]], [[TMP49]]109; CHECK-NEXT: [[TMP52:%.*]] = sext <64 x i1> [[TMP51]] to <64 x i8>110; CHECK-NEXT: [[TMP55:%.*]] = bitcast <64 x i8> [[TMP52]] to <16 x i32>111; CHECK-NEXT: [[TMP56:%.*]] = icmp ne <16 x i32> [[TMP55]], zeroinitializer112; CHECK-NEXT: [[TMP57:%.*]] = sext <16 x i1> [[TMP56]] to <16 x i32>113; CHECK-NEXT: [[_MSPROP3:%.*]] = or <16 x i32> [[TMP57]], [[TMP2]]114; CHECK-NEXT: [[TMP19:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpbusd.512(<16 x i32> [[X0]], <64 x i8> [[TMP39]], <64 x i8> [[TMP40]])115; CHECK-NEXT: [[TMP20:%.*]] = bitcast i16 [[TMP4]] to <16 x i1>116; CHECK-NEXT: [[TMP21:%.*]] = bitcast i16 [[X3]] to <16 x i1>117; CHECK-NEXT: [[TMP22:%.*]] = select <16 x i1> [[TMP21]], <16 x i32> [[_MSPROP3]], <16 x i32> zeroinitializer118; CHECK-NEXT: [[TMP23:%.*]] = xor <16 x i32> [[TMP19]], zeroinitializer119; CHECK-NEXT: [[TMP24:%.*]] = or <16 x i32> [[TMP23]], [[_MSPROP3]]120; CHECK-NEXT: [[TMP25:%.*]] = or <16 x i32> [[TMP24]], zeroinitializer121; CHECK-NEXT: [[_MSPROP_SELECT4:%.*]] = select <16 x i1> [[TMP20]], <16 x i32> [[TMP25]], <16 x i32> [[TMP22]]122; CHECK-NEXT: [[TMP26:%.*]] = select <16 x i1> [[TMP21]], <16 x i32> [[TMP19]], <16 x i32> zeroinitializer123; CHECK-NEXT: [[TMP27:%.*]] = insertvalue { <16 x i32>, <16 x i32> } { <16 x i32> splat (i32 -1), <16 x i32> splat (i32 -1) }, <16 x i32> [[_MSPROP_SELECT]], 0124; CHECK-NEXT: [[RES2:%.*]] = insertvalue { <16 x i32>, <16 x i32> } poison, <16 x i32> [[TMP18]], 0125; CHECK-NEXT: [[TMP28:%.*]] = insertvalue { <16 x i32>, <16 x i32> } [[TMP27]], <16 x i32> [[_MSPROP_SELECT4]], 1126; CHECK-NEXT: [[RES3:%.*]] = insertvalue { <16 x i32>, <16 x i32> } [[RES2]], <16 x i32> [[TMP26]], 1127; CHECK-NEXT: store { <16 x i32>, <16 x i32> } [[TMP28]], ptr @__msan_retval_tls, align 8128; CHECK-NEXT: ret { <16 x i32>, <16 x i32> } [[RES3]]129;130 %x2 = load <16 x i32>, ptr %x2p131 %res0 = call <16 x i32> @llvm.x86.avx512.mask.vpdpbusd.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2, i16 %x3)132 %res1 = call <16 x i32> @llvm.x86.avx512.maskz.vpdpbusd.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x4, i16 %x3)133 %res2 = insertvalue { <16 x i32>, <16 x i32> } poison, <16 x i32> %res0, 0134 %res3 = insertvalue { <16 x i32>, <16 x i32> } %res2, <16 x i32> %res1, 1135 ret { <16 x i32>, <16 x i32> } %res3136}137 138declare <16 x i32> @llvm.x86.avx512.mask.vpdpbusds.512(<16 x i32>, <16 x i32>, <16 x i32>, i16)139declare <16 x i32> @llvm.x86.avx512.maskz.vpdpbusds.512(<16 x i32>, <16 x i32>, <16 x i32>, i16)140 141define <16 x i32>@test_int_x86_avx512_vpdpbusds_512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2) sanitize_memory {142; CHECK-LABEL: define <16 x i32> @test_int_x86_avx512_vpdpbusds_512(143; CHECK-SAME: <16 x i32> [[X0:%.*]], <16 x i32> [[X1:%.*]], <16 x i32> [[X2:%.*]]) #[[ATTR0]] {144; CHECK-NEXT: [[TMP2:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 64), align 8145; CHECK-NEXT: [[TMP3:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 128), align 8146; CHECK-NEXT: [[TMP23:%.*]] = load <16 x i32>, ptr @__msan_param_tls, align 8147; CHECK-NEXT: call void @llvm.donothing()148; CHECK-NEXT: [[TMP6:%.*]] = bitcast <16 x i32> [[TMP2]] to <64 x i8>149; CHECK-NEXT: [[TMP24:%.*]] = bitcast <16 x i32> [[X1]] to <64 x i8>150; CHECK-NEXT: [[TMP7:%.*]] = bitcast <16 x i32> [[TMP3]] to <64 x i8>151; CHECK-NEXT: [[TMP5:%.*]] = bitcast <16 x i32> [[X2]] to <64 x i8>152; CHECK-NEXT: [[TMP8:%.*]] = icmp ne <64 x i8> [[TMP6]], zeroinitializer153; CHECK-NEXT: [[TMP9:%.*]] = icmp ne <64 x i8> [[TMP7]], zeroinitializer154; CHECK-NEXT: [[TMP10:%.*]] = icmp ne <64 x i8> [[TMP24]], zeroinitializer155; CHECK-NEXT: [[TMP11:%.*]] = icmp ne <64 x i8> [[TMP5]], zeroinitializer156; CHECK-NEXT: [[TMP12:%.*]] = and <64 x i1> [[TMP8]], [[TMP9]]157; CHECK-NEXT: [[TMP13:%.*]] = and <64 x i1> [[TMP10]], [[TMP9]]158; CHECK-NEXT: [[TMP14:%.*]] = and <64 x i1> [[TMP8]], [[TMP11]]159; CHECK-NEXT: [[TMP15:%.*]] = or <64 x i1> [[TMP12]], [[TMP13]]160; CHECK-NEXT: [[TMP16:%.*]] = or <64 x i1> [[TMP15]], [[TMP14]]161; CHECK-NEXT: [[TMP17:%.*]] = sext <64 x i1> [[TMP16]] to <64 x i8>162; CHECK-NEXT: [[TMP18:%.*]] = bitcast <64 x i8> [[TMP17]] to <16 x i32>163; CHECK-NEXT: [[TMP19:%.*]] = icmp ne <16 x i32> [[TMP18]], zeroinitializer164; CHECK-NEXT: [[TMP22:%.*]] = sext <16 x i1> [[TMP19]] to <16 x i32>165; CHECK-NEXT: [[_MSPROP1:%.*]] = or <16 x i32> [[TMP22]], [[TMP23]]166; CHECK-NEXT: [[TMP4:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpbusds.512(<16 x i32> [[X0]], <64 x i8> [[TMP24]], <64 x i8> [[TMP5]])167; CHECK-NEXT: store <16 x i32> [[_MSPROP1]], ptr @__msan_retval_tls, align 8168; CHECK-NEXT: ret <16 x i32> [[TMP4]]169;170 %res = call <16 x i32> @llvm.x86.avx512.mask.vpdpbusds.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2, i16 -1)171 ret <16 x i32> %res172}173 174define { <16 x i32>, <16 x i32> } @test_int_x86_avx512_mask_vpdpbusds_512(<16 x i32> %x0, <16 x i32> %x1, ptr %x2p, <16 x i32> %x4, i16 %x3) sanitize_memory {175; CHECK-LABEL: define { <16 x i32>, <16 x i32> } @test_int_x86_avx512_mask_vpdpbusds_512(176; CHECK-SAME: <16 x i32> [[X0:%.*]], <16 x i32> [[X1:%.*]], ptr [[X2P:%.*]], <16 x i32> [[X4:%.*]], i16 [[X3:%.*]]) #[[ATTR0]] {177; CHECK-NEXT: [[TMP1:%.*]] = load i64, ptr getelementptr (i8, ptr @__msan_param_tls, i64 128), align 8178; CHECK-NEXT: [[TMP3:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 64), align 8179; CHECK-NEXT: [[TMP2:%.*]] = load <16 x i32>, ptr @__msan_param_tls, align 8180; CHECK-NEXT: [[TMP4:%.*]] = load i16, ptr getelementptr (i8, ptr @__msan_param_tls, i64 200), align 8181; CHECK-NEXT: [[TMP5:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 136), align 8182; CHECK-NEXT: call void @llvm.donothing()183; CHECK-NEXT: [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0184; CHECK-NEXT: br i1 [[_MSCMP]], label %[[BB6:.*]], label %[[BB7:.*]], !prof [[PROF1]]185; CHECK: [[BB6]]:186; CHECK-NEXT: call void @__msan_warning_noreturn() #[[ATTR4]]187; CHECK-NEXT: unreachable188; CHECK: [[BB7]]:189; CHECK-NEXT: [[X2:%.*]] = load <16 x i32>, ptr [[X2P]], align 64190; CHECK-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[X2P]] to i64191; CHECK-NEXT: [[TMP9:%.*]] = xor i64 [[TMP8]], 87960930222080192; CHECK-NEXT: [[TMP10:%.*]] = inttoptr i64 [[TMP9]] to ptr193; CHECK-NEXT: [[_MSLD:%.*]] = load <16 x i32>, ptr [[TMP10]], align 64194; CHECK-NEXT: [[TMP32:%.*]] = bitcast <16 x i32> [[TMP3]] to <64 x i8>195; CHECK-NEXT: [[TMP30:%.*]] = bitcast <16 x i32> [[X1]] to <64 x i8>196; CHECK-NEXT: [[TMP33:%.*]] = bitcast <16 x i32> [[_MSLD]] to <64 x i8>197; CHECK-NEXT: [[TMP31:%.*]] = bitcast <16 x i32> [[X2]] to <64 x i8>198; CHECK-NEXT: [[TMP34:%.*]] = icmp ne <64 x i8> [[TMP32]], zeroinitializer199; CHECK-NEXT: [[TMP35:%.*]] = icmp ne <64 x i8> [[TMP33]], zeroinitializer200; CHECK-NEXT: [[TMP36:%.*]] = icmp ne <64 x i8> [[TMP30]], zeroinitializer201; CHECK-NEXT: [[TMP37:%.*]] = icmp ne <64 x i8> [[TMP31]], zeroinitializer202; CHECK-NEXT: [[TMP38:%.*]] = and <64 x i1> [[TMP34]], [[TMP35]]203; CHECK-NEXT: [[TMP58:%.*]] = and <64 x i1> [[TMP36]], [[TMP35]]204; CHECK-NEXT: [[TMP59:%.*]] = and <64 x i1> [[TMP34]], [[TMP37]]205; CHECK-NEXT: [[TMP60:%.*]] = or <64 x i1> [[TMP38]], [[TMP58]]206; CHECK-NEXT: [[TMP61:%.*]] = or <64 x i1> [[TMP60]], [[TMP59]]207; CHECK-NEXT: [[TMP62:%.*]] = sext <64 x i1> [[TMP61]] to <64 x i8>208; CHECK-NEXT: [[TMP53:%.*]] = bitcast <64 x i8> [[TMP62]] to <16 x i32>209; CHECK-NEXT: [[TMP54:%.*]] = icmp ne <16 x i32> [[TMP53]], zeroinitializer210; CHECK-NEXT: [[TMP29:%.*]] = sext <16 x i1> [[TMP54]] to <16 x i32>211; CHECK-NEXT: [[_MSPROP1:%.*]] = or <16 x i32> [[TMP29]], [[TMP2]]212; CHECK-NEXT: [[TMP11:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpbusds.512(<16 x i32> [[X0]], <64 x i8> [[TMP30]], <64 x i8> [[TMP31]])213; CHECK-NEXT: [[TMP12:%.*]] = bitcast i16 [[TMP4]] to <16 x i1>214; CHECK-NEXT: [[TMP13:%.*]] = bitcast i16 [[X3]] to <16 x i1>215; CHECK-NEXT: [[TMP14:%.*]] = select <16 x i1> [[TMP13]], <16 x i32> [[_MSPROP1]], <16 x i32> [[TMP2]]216; CHECK-NEXT: [[TMP15:%.*]] = xor <16 x i32> [[TMP11]], [[X0]]217; CHECK-NEXT: [[TMP16:%.*]] = or <16 x i32> [[TMP15]], [[_MSPROP1]]218; CHECK-NEXT: [[TMP17:%.*]] = or <16 x i32> [[TMP16]], [[TMP2]]219; CHECK-NEXT: [[_MSPROP_SELECT:%.*]] = select <16 x i1> [[TMP12]], <16 x i32> [[TMP17]], <16 x i32> [[TMP14]]220; CHECK-NEXT: [[TMP18:%.*]] = select <16 x i1> [[TMP13]], <16 x i32> [[TMP11]], <16 x i32> [[X0]]221; CHECK-NEXT: [[TMP41:%.*]] = bitcast <16 x i32> [[TMP3]] to <64 x i8>222; CHECK-NEXT: [[TMP39:%.*]] = bitcast <16 x i32> [[X1]] to <64 x i8>223; CHECK-NEXT: [[TMP42:%.*]] = bitcast <16 x i32> [[TMP5]] to <64 x i8>224; CHECK-NEXT: [[TMP40:%.*]] = bitcast <16 x i32> [[X4]] to <64 x i8>225; CHECK-NEXT: [[TMP43:%.*]] = icmp ne <64 x i8> [[TMP41]], zeroinitializer226; CHECK-NEXT: [[TMP44:%.*]] = icmp ne <64 x i8> [[TMP42]], zeroinitializer227; CHECK-NEXT: [[TMP45:%.*]] = icmp ne <64 x i8> [[TMP39]], zeroinitializer228; CHECK-NEXT: [[TMP46:%.*]] = icmp ne <64 x i8> [[TMP40]], zeroinitializer229; CHECK-NEXT: [[TMP47:%.*]] = and <64 x i1> [[TMP43]], [[TMP44]]230; CHECK-NEXT: [[TMP48:%.*]] = and <64 x i1> [[TMP45]], [[TMP44]]231; CHECK-NEXT: [[TMP49:%.*]] = and <64 x i1> [[TMP43]], [[TMP46]]232; CHECK-NEXT: [[TMP50:%.*]] = or <64 x i1> [[TMP47]], [[TMP48]]233; CHECK-NEXT: [[TMP51:%.*]] = or <64 x i1> [[TMP50]], [[TMP49]]234; CHECK-NEXT: [[TMP52:%.*]] = sext <64 x i1> [[TMP51]] to <64 x i8>235; CHECK-NEXT: [[TMP55:%.*]] = bitcast <64 x i8> [[TMP52]] to <16 x i32>236; CHECK-NEXT: [[TMP56:%.*]] = icmp ne <16 x i32> [[TMP55]], zeroinitializer237; CHECK-NEXT: [[TMP57:%.*]] = sext <16 x i1> [[TMP56]] to <16 x i32>238; CHECK-NEXT: [[_MSPROP3:%.*]] = or <16 x i32> [[TMP57]], [[TMP2]]239; CHECK-NEXT: [[TMP19:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpbusds.512(<16 x i32> [[X0]], <64 x i8> [[TMP39]], <64 x i8> [[TMP40]])240; CHECK-NEXT: [[TMP20:%.*]] = bitcast i16 [[TMP4]] to <16 x i1>241; CHECK-NEXT: [[TMP21:%.*]] = bitcast i16 [[X3]] to <16 x i1>242; CHECK-NEXT: [[TMP22:%.*]] = select <16 x i1> [[TMP21]], <16 x i32> [[_MSPROP3]], <16 x i32> zeroinitializer243; CHECK-NEXT: [[TMP23:%.*]] = xor <16 x i32> [[TMP19]], zeroinitializer244; CHECK-NEXT: [[TMP24:%.*]] = or <16 x i32> [[TMP23]], [[_MSPROP3]]245; CHECK-NEXT: [[TMP25:%.*]] = or <16 x i32> [[TMP24]], zeroinitializer246; CHECK-NEXT: [[_MSPROP_SELECT4:%.*]] = select <16 x i1> [[TMP20]], <16 x i32> [[TMP25]], <16 x i32> [[TMP22]]247; CHECK-NEXT: [[TMP26:%.*]] = select <16 x i1> [[TMP21]], <16 x i32> [[TMP19]], <16 x i32> zeroinitializer248; CHECK-NEXT: [[TMP27:%.*]] = insertvalue { <16 x i32>, <16 x i32> } { <16 x i32> splat (i32 -1), <16 x i32> splat (i32 -1) }, <16 x i32> [[_MSPROP_SELECT]], 0249; CHECK-NEXT: [[RES2:%.*]] = insertvalue { <16 x i32>, <16 x i32> } poison, <16 x i32> [[TMP18]], 0250; CHECK-NEXT: [[TMP28:%.*]] = insertvalue { <16 x i32>, <16 x i32> } [[TMP27]], <16 x i32> [[_MSPROP_SELECT4]], 1251; CHECK-NEXT: [[RES3:%.*]] = insertvalue { <16 x i32>, <16 x i32> } [[RES2]], <16 x i32> [[TMP26]], 1252; CHECK-NEXT: store { <16 x i32>, <16 x i32> } [[TMP28]], ptr @__msan_retval_tls, align 8253; CHECK-NEXT: ret { <16 x i32>, <16 x i32> } [[RES3]]254;255 %x2 = load <16 x i32>, ptr %x2p256 %res0 = call <16 x i32> @llvm.x86.avx512.mask.vpdpbusds.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2, i16 %x3)257 %res1 = call <16 x i32> @llvm.x86.avx512.maskz.vpdpbusds.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x4, i16 %x3)258 %res2 = insertvalue { <16 x i32>, <16 x i32> } poison, <16 x i32> %res0, 0259 %res3 = insertvalue { <16 x i32>, <16 x i32> } %res2, <16 x i32> %res1, 1260 ret { <16 x i32>, <16 x i32> } %res3261}262 263declare <16 x i32> @llvm.x86.avx512.mask.vpdpwssd.512(<16 x i32>, <16 x i32>, <16 x i32>, i16)264declare <16 x i32> @llvm.x86.avx512.maskz.vpdpwssd.512(<16 x i32>, <16 x i32>, <16 x i32>, i16)265 266define <16 x i32>@test_int_x86_avx512_vpdpwssd_512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2) sanitize_memory {267; CHECK-LABEL: define <16 x i32> @test_int_x86_avx512_vpdpwssd_512(268; CHECK-SAME: <16 x i32> [[X0:%.*]], <16 x i32> [[X1:%.*]], <16 x i32> [[X2:%.*]]) #[[ATTR0]] {269; CHECK-NEXT: [[TMP2:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 64), align 8270; CHECK-NEXT: [[TMP3:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 128), align 8271; CHECK-NEXT: [[TMP21:%.*]] = load <16 x i32>, ptr @__msan_param_tls, align 8272; CHECK-NEXT: call void @llvm.donothing()273; CHECK-NEXT: [[TMP22:%.*]] = bitcast <16 x i32> [[X1]] to <32 x i16>274; CHECK-NEXT: [[TMP5:%.*]] = bitcast <16 x i32> [[X2]] to <32 x i16>275; CHECK-NEXT: [[TMP6:%.*]] = bitcast <16 x i32> [[TMP2]] to <32 x i16>276; CHECK-NEXT: [[TMP7:%.*]] = bitcast <16 x i32> [[TMP3]] to <32 x i16>277; CHECK-NEXT: [[TMP8:%.*]] = icmp ne <32 x i16> [[TMP6]], zeroinitializer278; CHECK-NEXT: [[TMP9:%.*]] = icmp ne <32 x i16> [[TMP7]], zeroinitializer279; CHECK-NEXT: [[TMP10:%.*]] = icmp ne <32 x i16> [[TMP22]], zeroinitializer280; CHECK-NEXT: [[TMP11:%.*]] = icmp ne <32 x i16> [[TMP5]], zeroinitializer281; CHECK-NEXT: [[TMP12:%.*]] = and <32 x i1> [[TMP8]], [[TMP9]]282; CHECK-NEXT: [[TMP13:%.*]] = and <32 x i1> [[TMP10]], [[TMP9]]283; CHECK-NEXT: [[TMP14:%.*]] = and <32 x i1> [[TMP8]], [[TMP11]]284; CHECK-NEXT: [[TMP15:%.*]] = or <32 x i1> [[TMP12]], [[TMP13]]285; CHECK-NEXT: [[TMP16:%.*]] = or <32 x i1> [[TMP15]], [[TMP14]]286; CHECK-NEXT: [[TMP17:%.*]] = sext <32 x i1> [[TMP16]] to <32 x i16>287; CHECK-NEXT: [[TMP18:%.*]] = bitcast <32 x i16> [[TMP17]] to <16 x i32>288; CHECK-NEXT: [[TMP19:%.*]] = icmp ne <16 x i32> [[TMP18]], zeroinitializer289; CHECK-NEXT: [[TMP20:%.*]] = sext <16 x i1> [[TMP19]] to <16 x i32>290; CHECK-NEXT: [[_MSPROP1:%.*]] = or <16 x i32> [[TMP20]], [[TMP21]]291; CHECK-NEXT: [[TMP4:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpwssd.512(<16 x i32> [[X0]], <16 x i32> [[X1]], <16 x i32> [[X2]])292; CHECK-NEXT: store <16 x i32> [[_MSPROP1]], ptr @__msan_retval_tls, align 8293; CHECK-NEXT: ret <16 x i32> [[TMP4]]294;295 %res = call <16 x i32> @llvm.x86.avx512.mask.vpdpwssd.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2, i16 -1)296 ret <16 x i32> %res297}298 299define { <16 x i32>, <16 x i32> } @test_int_x86_avx512_mask_vpdpwssd_512(<16 x i32> %x0, <16 x i32> %x1, ptr %x2p, <16 x i32> %x4, i16 %x3) sanitize_memory {300; CHECK-LABEL: define { <16 x i32>, <16 x i32> } @test_int_x86_avx512_mask_vpdpwssd_512(301; CHECK-SAME: <16 x i32> [[X0:%.*]], <16 x i32> [[X1:%.*]], ptr [[X2P:%.*]], <16 x i32> [[X4:%.*]], i16 [[X3:%.*]]) #[[ATTR0]] {302; CHECK-NEXT: [[TMP1:%.*]] = load i64, ptr getelementptr (i8, ptr @__msan_param_tls, i64 128), align 8303; CHECK-NEXT: [[TMP3:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 64), align 8304; CHECK-NEXT: [[TMP2:%.*]] = load <16 x i32>, ptr @__msan_param_tls, align 8305; CHECK-NEXT: [[TMP4:%.*]] = load i16, ptr getelementptr (i8, ptr @__msan_param_tls, i64 200), align 8306; CHECK-NEXT: [[TMP5:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 136), align 8307; CHECK-NEXT: call void @llvm.donothing()308; CHECK-NEXT: [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0309; CHECK-NEXT: br i1 [[_MSCMP]], label %[[BB6:.*]], label %[[BB7:.*]], !prof [[PROF1]]310; CHECK: [[BB6]]:311; CHECK-NEXT: call void @__msan_warning_noreturn() #[[ATTR4]]312; CHECK-NEXT: unreachable313; CHECK: [[BB7]]:314; CHECK-NEXT: [[X2:%.*]] = load <16 x i32>, ptr [[X2P]], align 64315; CHECK-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[X2P]] to i64316; CHECK-NEXT: [[TMP9:%.*]] = xor i64 [[TMP8]], 87960930222080317; CHECK-NEXT: [[TMP10:%.*]] = inttoptr i64 [[TMP9]] to ptr318; CHECK-NEXT: [[_MSLD:%.*]] = load <16 x i32>, ptr [[TMP10]], align 64319; CHECK-NEXT: [[TMP29:%.*]] = bitcast <16 x i32> [[X1]] to <32 x i16>320; CHECK-NEXT: [[TMP30:%.*]] = bitcast <16 x i32> [[X2]] to <32 x i16>321; CHECK-NEXT: [[TMP31:%.*]] = bitcast <16 x i32> [[TMP3]] to <32 x i16>322; CHECK-NEXT: [[TMP32:%.*]] = bitcast <16 x i32> [[_MSLD]] to <32 x i16>323; CHECK-NEXT: [[TMP33:%.*]] = icmp ne <32 x i16> [[TMP31]], zeroinitializer324; CHECK-NEXT: [[TMP34:%.*]] = icmp ne <32 x i16> [[TMP32]], zeroinitializer325; CHECK-NEXT: [[TMP35:%.*]] = icmp ne <32 x i16> [[TMP29]], zeroinitializer326; CHECK-NEXT: [[TMP36:%.*]] = icmp ne <32 x i16> [[TMP30]], zeroinitializer327; CHECK-NEXT: [[TMP54:%.*]] = and <32 x i1> [[TMP33]], [[TMP34]]328; CHECK-NEXT: [[TMP55:%.*]] = and <32 x i1> [[TMP35]], [[TMP34]]329; CHECK-NEXT: [[TMP56:%.*]] = and <32 x i1> [[TMP33]], [[TMP36]]330; CHECK-NEXT: [[TMP57:%.*]] = or <32 x i1> [[TMP54]], [[TMP55]]331; CHECK-NEXT: [[TMP58:%.*]] = or <32 x i1> [[TMP57]], [[TMP56]]332; CHECK-NEXT: [[TMP59:%.*]] = sext <32 x i1> [[TMP58]] to <32 x i16>333; CHECK-NEXT: [[TMP60:%.*]] = bitcast <32 x i16> [[TMP59]] to <16 x i32>334; CHECK-NEXT: [[TMP61:%.*]] = icmp ne <16 x i32> [[TMP60]], zeroinitializer335; CHECK-NEXT: [[TMP62:%.*]] = sext <16 x i1> [[TMP61]] to <16 x i32>336; CHECK-NEXT: [[_MSPROP1:%.*]] = or <16 x i32> [[TMP62]], [[TMP2]]337; CHECK-NEXT: [[TMP11:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpwssd.512(<16 x i32> [[X0]], <16 x i32> [[X1]], <16 x i32> [[X2]])338; CHECK-NEXT: [[TMP12:%.*]] = bitcast i16 [[TMP4]] to <16 x i1>339; CHECK-NEXT: [[TMP13:%.*]] = bitcast i16 [[X3]] to <16 x i1>340; CHECK-NEXT: [[TMP14:%.*]] = select <16 x i1> [[TMP13]], <16 x i32> [[_MSPROP1]], <16 x i32> [[TMP2]]341; CHECK-NEXT: [[TMP15:%.*]] = xor <16 x i32> [[TMP11]], [[X0]]342; CHECK-NEXT: [[TMP16:%.*]] = or <16 x i32> [[TMP15]], [[_MSPROP1]]343; CHECK-NEXT: [[TMP17:%.*]] = or <16 x i32> [[TMP16]], [[TMP2]]344; CHECK-NEXT: [[_MSPROP_SELECT:%.*]] = select <16 x i1> [[TMP12]], <16 x i32> [[TMP17]], <16 x i32> [[TMP14]]345; CHECK-NEXT: [[TMP18:%.*]] = select <16 x i1> [[TMP13]], <16 x i32> [[TMP11]], <16 x i32> [[X0]]346; CHECK-NEXT: [[TMP37:%.*]] = bitcast <16 x i32> [[X1]] to <32 x i16>347; CHECK-NEXT: [[TMP38:%.*]] = bitcast <16 x i32> [[X4]] to <32 x i16>348; CHECK-NEXT: [[TMP39:%.*]] = bitcast <16 x i32> [[TMP3]] to <32 x i16>349; CHECK-NEXT: [[TMP40:%.*]] = bitcast <16 x i32> [[TMP5]] to <32 x i16>350; CHECK-NEXT: [[TMP41:%.*]] = icmp ne <32 x i16> [[TMP39]], zeroinitializer351; CHECK-NEXT: [[TMP42:%.*]] = icmp ne <32 x i16> [[TMP40]], zeroinitializer352; CHECK-NEXT: [[TMP43:%.*]] = icmp ne <32 x i16> [[TMP37]], zeroinitializer353; CHECK-NEXT: [[TMP44:%.*]] = icmp ne <32 x i16> [[TMP38]], zeroinitializer354; CHECK-NEXT: [[TMP45:%.*]] = and <32 x i1> [[TMP41]], [[TMP42]]355; CHECK-NEXT: [[TMP46:%.*]] = and <32 x i1> [[TMP43]], [[TMP42]]356; CHECK-NEXT: [[TMP47:%.*]] = and <32 x i1> [[TMP41]], [[TMP44]]357; CHECK-NEXT: [[TMP48:%.*]] = or <32 x i1> [[TMP45]], [[TMP46]]358; CHECK-NEXT: [[TMP49:%.*]] = or <32 x i1> [[TMP48]], [[TMP47]]359; CHECK-NEXT: [[TMP50:%.*]] = sext <32 x i1> [[TMP49]] to <32 x i16>360; CHECK-NEXT: [[TMP51:%.*]] = bitcast <32 x i16> [[TMP50]] to <16 x i32>361; CHECK-NEXT: [[TMP52:%.*]] = icmp ne <16 x i32> [[TMP51]], zeroinitializer362; CHECK-NEXT: [[TMP53:%.*]] = sext <16 x i1> [[TMP52]] to <16 x i32>363; CHECK-NEXT: [[_MSPROP3:%.*]] = or <16 x i32> [[TMP53]], [[TMP2]]364; CHECK-NEXT: [[TMP19:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpwssd.512(<16 x i32> [[X0]], <16 x i32> [[X1]], <16 x i32> [[X4]])365; CHECK-NEXT: [[TMP20:%.*]] = bitcast i16 [[TMP4]] to <16 x i1>366; CHECK-NEXT: [[TMP21:%.*]] = bitcast i16 [[X3]] to <16 x i1>367; CHECK-NEXT: [[TMP22:%.*]] = select <16 x i1> [[TMP21]], <16 x i32> [[_MSPROP3]], <16 x i32> zeroinitializer368; CHECK-NEXT: [[TMP23:%.*]] = xor <16 x i32> [[TMP19]], zeroinitializer369; CHECK-NEXT: [[TMP24:%.*]] = or <16 x i32> [[TMP23]], [[_MSPROP3]]370; CHECK-NEXT: [[TMP25:%.*]] = or <16 x i32> [[TMP24]], zeroinitializer371; CHECK-NEXT: [[_MSPROP_SELECT4:%.*]] = select <16 x i1> [[TMP20]], <16 x i32> [[TMP25]], <16 x i32> [[TMP22]]372; CHECK-NEXT: [[TMP26:%.*]] = select <16 x i1> [[TMP21]], <16 x i32> [[TMP19]], <16 x i32> zeroinitializer373; CHECK-NEXT: [[TMP27:%.*]] = insertvalue { <16 x i32>, <16 x i32> } { <16 x i32> splat (i32 -1), <16 x i32> splat (i32 -1) }, <16 x i32> [[_MSPROP_SELECT]], 0374; CHECK-NEXT: [[RES2:%.*]] = insertvalue { <16 x i32>, <16 x i32> } poison, <16 x i32> [[TMP18]], 0375; CHECK-NEXT: [[TMP28:%.*]] = insertvalue { <16 x i32>, <16 x i32> } [[TMP27]], <16 x i32> [[_MSPROP_SELECT4]], 1376; CHECK-NEXT: [[RES3:%.*]] = insertvalue { <16 x i32>, <16 x i32> } [[RES2]], <16 x i32> [[TMP26]], 1377; CHECK-NEXT: store { <16 x i32>, <16 x i32> } [[TMP28]], ptr @__msan_retval_tls, align 8378; CHECK-NEXT: ret { <16 x i32>, <16 x i32> } [[RES3]]379;380 %x2 = load <16 x i32>, ptr %x2p381 %res0 = call <16 x i32> @llvm.x86.avx512.mask.vpdpwssd.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2, i16 %x3)382 %res1 = call <16 x i32> @llvm.x86.avx512.maskz.vpdpwssd.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x4, i16 %x3)383 %res2 = insertvalue { <16 x i32>, <16 x i32> } poison, <16 x i32> %res0, 0384 %res3 = insertvalue { <16 x i32>, <16 x i32> } %res2, <16 x i32> %res1, 1385 ret { <16 x i32>, <16 x i32> } %res3386}387 388declare <16 x i32> @llvm.x86.avx512.mask.vpdpwssds.512(<16 x i32>, <16 x i32>, <16 x i32>, i16)389declare <16 x i32> @llvm.x86.avx512.maskz.vpdpwssds.512(<16 x i32>, <16 x i32>, <16 x i32>, i16)390 391define <16 x i32>@test_int_x86_avx512_vpdpwssds_512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2) sanitize_memory {392; CHECK-LABEL: define <16 x i32> @test_int_x86_avx512_vpdpwssds_512(393; CHECK-SAME: <16 x i32> [[X0:%.*]], <16 x i32> [[X1:%.*]], <16 x i32> [[X2:%.*]]) #[[ATTR0]] {394; CHECK-NEXT: [[TMP2:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 64), align 8395; CHECK-NEXT: [[TMP3:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 128), align 8396; CHECK-NEXT: [[TMP21:%.*]] = load <16 x i32>, ptr @__msan_param_tls, align 8397; CHECK-NEXT: call void @llvm.donothing()398; CHECK-NEXT: [[TMP22:%.*]] = bitcast <16 x i32> [[X1]] to <32 x i16>399; CHECK-NEXT: [[TMP5:%.*]] = bitcast <16 x i32> [[X2]] to <32 x i16>400; CHECK-NEXT: [[TMP6:%.*]] = bitcast <16 x i32> [[TMP2]] to <32 x i16>401; CHECK-NEXT: [[TMP7:%.*]] = bitcast <16 x i32> [[TMP3]] to <32 x i16>402; CHECK-NEXT: [[TMP8:%.*]] = icmp ne <32 x i16> [[TMP6]], zeroinitializer403; CHECK-NEXT: [[TMP9:%.*]] = icmp ne <32 x i16> [[TMP7]], zeroinitializer404; CHECK-NEXT: [[TMP10:%.*]] = icmp ne <32 x i16> [[TMP22]], zeroinitializer405; CHECK-NEXT: [[TMP11:%.*]] = icmp ne <32 x i16> [[TMP5]], zeroinitializer406; CHECK-NEXT: [[TMP12:%.*]] = and <32 x i1> [[TMP8]], [[TMP9]]407; CHECK-NEXT: [[TMP13:%.*]] = and <32 x i1> [[TMP10]], [[TMP9]]408; CHECK-NEXT: [[TMP14:%.*]] = and <32 x i1> [[TMP8]], [[TMP11]]409; CHECK-NEXT: [[TMP15:%.*]] = or <32 x i1> [[TMP12]], [[TMP13]]410; CHECK-NEXT: [[TMP16:%.*]] = or <32 x i1> [[TMP15]], [[TMP14]]411; CHECK-NEXT: [[TMP17:%.*]] = sext <32 x i1> [[TMP16]] to <32 x i16>412; CHECK-NEXT: [[TMP18:%.*]] = bitcast <32 x i16> [[TMP17]] to <16 x i32>413; CHECK-NEXT: [[TMP19:%.*]] = icmp ne <16 x i32> [[TMP18]], zeroinitializer414; CHECK-NEXT: [[TMP20:%.*]] = sext <16 x i1> [[TMP19]] to <16 x i32>415; CHECK-NEXT: [[_MSPROP1:%.*]] = or <16 x i32> [[TMP20]], [[TMP21]]416; CHECK-NEXT: [[TMP4:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpwssds.512(<16 x i32> [[X0]], <16 x i32> [[X1]], <16 x i32> [[X2]])417; CHECK-NEXT: store <16 x i32> [[_MSPROP1]], ptr @__msan_retval_tls, align 8418; CHECK-NEXT: ret <16 x i32> [[TMP4]]419;420 %res = call <16 x i32> @llvm.x86.avx512.mask.vpdpwssds.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2, i16 -1)421 ret <16 x i32> %res422}423 424define { <16 x i32>, <16 x i32> } @test_int_x86_avx512_mask_vpdpwssds_512(<16 x i32> %x0, <16 x i32> %x1, ptr %x2p, <16 x i32> %x4, i16 %x3) sanitize_memory {425; CHECK-LABEL: define { <16 x i32>, <16 x i32> } @test_int_x86_avx512_mask_vpdpwssds_512(426; CHECK-SAME: <16 x i32> [[X0:%.*]], <16 x i32> [[X1:%.*]], ptr [[X2P:%.*]], <16 x i32> [[X4:%.*]], i16 [[X3:%.*]]) #[[ATTR0]] {427; CHECK-NEXT: [[TMP1:%.*]] = load i64, ptr getelementptr (i8, ptr @__msan_param_tls, i64 128), align 8428; CHECK-NEXT: [[TMP3:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 64), align 8429; CHECK-NEXT: [[TMP2:%.*]] = load <16 x i32>, ptr @__msan_param_tls, align 8430; CHECK-NEXT: [[TMP4:%.*]] = load i16, ptr getelementptr (i8, ptr @__msan_param_tls, i64 200), align 8431; CHECK-NEXT: [[TMP5:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 136), align 8432; CHECK-NEXT: call void @llvm.donothing()433; CHECK-NEXT: [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0434; CHECK-NEXT: br i1 [[_MSCMP]], label %[[BB6:.*]], label %[[BB7:.*]], !prof [[PROF1]]435; CHECK: [[BB6]]:436; CHECK-NEXT: call void @__msan_warning_noreturn() #[[ATTR4]]437; CHECK-NEXT: unreachable438; CHECK: [[BB7]]:439; CHECK-NEXT: [[X2:%.*]] = load <16 x i32>, ptr [[X2P]], align 64440; CHECK-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[X2P]] to i64441; CHECK-NEXT: [[TMP9:%.*]] = xor i64 [[TMP8]], 87960930222080442; CHECK-NEXT: [[TMP10:%.*]] = inttoptr i64 [[TMP9]] to ptr443; CHECK-NEXT: [[_MSLD:%.*]] = load <16 x i32>, ptr [[TMP10]], align 64444; CHECK-NEXT: [[TMP29:%.*]] = bitcast <16 x i32> [[X1]] to <32 x i16>445; CHECK-NEXT: [[TMP30:%.*]] = bitcast <16 x i32> [[X2]] to <32 x i16>446; CHECK-NEXT: [[TMP31:%.*]] = bitcast <16 x i32> [[TMP3]] to <32 x i16>447; CHECK-NEXT: [[TMP32:%.*]] = bitcast <16 x i32> [[_MSLD]] to <32 x i16>448; CHECK-NEXT: [[TMP33:%.*]] = icmp ne <32 x i16> [[TMP31]], zeroinitializer449; CHECK-NEXT: [[TMP34:%.*]] = icmp ne <32 x i16> [[TMP32]], zeroinitializer450; CHECK-NEXT: [[TMP35:%.*]] = icmp ne <32 x i16> [[TMP29]], zeroinitializer451; CHECK-NEXT: [[TMP36:%.*]] = icmp ne <32 x i16> [[TMP30]], zeroinitializer452; CHECK-NEXT: [[TMP54:%.*]] = and <32 x i1> [[TMP33]], [[TMP34]]453; CHECK-NEXT: [[TMP55:%.*]] = and <32 x i1> [[TMP35]], [[TMP34]]454; CHECK-NEXT: [[TMP56:%.*]] = and <32 x i1> [[TMP33]], [[TMP36]]455; CHECK-NEXT: [[TMP57:%.*]] = or <32 x i1> [[TMP54]], [[TMP55]]456; CHECK-NEXT: [[TMP58:%.*]] = or <32 x i1> [[TMP57]], [[TMP56]]457; CHECK-NEXT: [[TMP59:%.*]] = sext <32 x i1> [[TMP58]] to <32 x i16>458; CHECK-NEXT: [[TMP60:%.*]] = bitcast <32 x i16> [[TMP59]] to <16 x i32>459; CHECK-NEXT: [[TMP61:%.*]] = icmp ne <16 x i32> [[TMP60]], zeroinitializer460; CHECK-NEXT: [[TMP62:%.*]] = sext <16 x i1> [[TMP61]] to <16 x i32>461; CHECK-NEXT: [[_MSPROP1:%.*]] = or <16 x i32> [[TMP62]], [[TMP2]]462; CHECK-NEXT: [[TMP11:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpwssds.512(<16 x i32> [[X0]], <16 x i32> [[X1]], <16 x i32> [[X2]])463; CHECK-NEXT: [[TMP12:%.*]] = bitcast i16 [[TMP4]] to <16 x i1>464; CHECK-NEXT: [[TMP13:%.*]] = bitcast i16 [[X3]] to <16 x i1>465; CHECK-NEXT: [[TMP14:%.*]] = select <16 x i1> [[TMP13]], <16 x i32> [[_MSPROP1]], <16 x i32> [[TMP2]]466; CHECK-NEXT: [[TMP15:%.*]] = xor <16 x i32> [[TMP11]], [[X0]]467; CHECK-NEXT: [[TMP16:%.*]] = or <16 x i32> [[TMP15]], [[_MSPROP1]]468; CHECK-NEXT: [[TMP17:%.*]] = or <16 x i32> [[TMP16]], [[TMP2]]469; CHECK-NEXT: [[_MSPROP_SELECT:%.*]] = select <16 x i1> [[TMP12]], <16 x i32> [[TMP17]], <16 x i32> [[TMP14]]470; CHECK-NEXT: [[TMP18:%.*]] = select <16 x i1> [[TMP13]], <16 x i32> [[TMP11]], <16 x i32> [[X0]]471; CHECK-NEXT: [[TMP37:%.*]] = bitcast <16 x i32> [[X1]] to <32 x i16>472; CHECK-NEXT: [[TMP38:%.*]] = bitcast <16 x i32> [[X4]] to <32 x i16>473; CHECK-NEXT: [[TMP39:%.*]] = bitcast <16 x i32> [[TMP3]] to <32 x i16>474; CHECK-NEXT: [[TMP40:%.*]] = bitcast <16 x i32> [[TMP5]] to <32 x i16>475; CHECK-NEXT: [[TMP41:%.*]] = icmp ne <32 x i16> [[TMP39]], zeroinitializer476; CHECK-NEXT: [[TMP42:%.*]] = icmp ne <32 x i16> [[TMP40]], zeroinitializer477; CHECK-NEXT: [[TMP43:%.*]] = icmp ne <32 x i16> [[TMP37]], zeroinitializer478; CHECK-NEXT: [[TMP44:%.*]] = icmp ne <32 x i16> [[TMP38]], zeroinitializer479; CHECK-NEXT: [[TMP45:%.*]] = and <32 x i1> [[TMP41]], [[TMP42]]480; CHECK-NEXT: [[TMP46:%.*]] = and <32 x i1> [[TMP43]], [[TMP42]]481; CHECK-NEXT: [[TMP47:%.*]] = and <32 x i1> [[TMP41]], [[TMP44]]482; CHECK-NEXT: [[TMP48:%.*]] = or <32 x i1> [[TMP45]], [[TMP46]]483; CHECK-NEXT: [[TMP49:%.*]] = or <32 x i1> [[TMP48]], [[TMP47]]484; CHECK-NEXT: [[TMP50:%.*]] = sext <32 x i1> [[TMP49]] to <32 x i16>485; CHECK-NEXT: [[TMP51:%.*]] = bitcast <32 x i16> [[TMP50]] to <16 x i32>486; CHECK-NEXT: [[TMP52:%.*]] = icmp ne <16 x i32> [[TMP51]], zeroinitializer487; CHECK-NEXT: [[TMP53:%.*]] = sext <16 x i1> [[TMP52]] to <16 x i32>488; CHECK-NEXT: [[_MSPROP3:%.*]] = or <16 x i32> [[TMP53]], [[TMP2]]489; CHECK-NEXT: [[TMP19:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpwssds.512(<16 x i32> [[X0]], <16 x i32> [[X1]], <16 x i32> [[X4]])490; CHECK-NEXT: [[TMP20:%.*]] = bitcast i16 [[TMP4]] to <16 x i1>491; CHECK-NEXT: [[TMP21:%.*]] = bitcast i16 [[X3]] to <16 x i1>492; CHECK-NEXT: [[TMP22:%.*]] = select <16 x i1> [[TMP21]], <16 x i32> [[_MSPROP3]], <16 x i32> zeroinitializer493; CHECK-NEXT: [[TMP23:%.*]] = xor <16 x i32> [[TMP19]], zeroinitializer494; CHECK-NEXT: [[TMP24:%.*]] = or <16 x i32> [[TMP23]], [[_MSPROP3]]495; CHECK-NEXT: [[TMP25:%.*]] = or <16 x i32> [[TMP24]], zeroinitializer496; CHECK-NEXT: [[_MSPROP_SELECT4:%.*]] = select <16 x i1> [[TMP20]], <16 x i32> [[TMP25]], <16 x i32> [[TMP22]]497; CHECK-NEXT: [[TMP26:%.*]] = select <16 x i1> [[TMP21]], <16 x i32> [[TMP19]], <16 x i32> zeroinitializer498; CHECK-NEXT: [[TMP27:%.*]] = insertvalue { <16 x i32>, <16 x i32> } { <16 x i32> splat (i32 -1), <16 x i32> splat (i32 -1) }, <16 x i32> [[_MSPROP_SELECT]], 0499; CHECK-NEXT: [[RES2:%.*]] = insertvalue { <16 x i32>, <16 x i32> } poison, <16 x i32> [[TMP18]], 0500; CHECK-NEXT: [[TMP28:%.*]] = insertvalue { <16 x i32>, <16 x i32> } [[TMP27]], <16 x i32> [[_MSPROP_SELECT4]], 1501; CHECK-NEXT: [[RES3:%.*]] = insertvalue { <16 x i32>, <16 x i32> } [[RES2]], <16 x i32> [[TMP26]], 1502; CHECK-NEXT: store { <16 x i32>, <16 x i32> } [[TMP28]], ptr @__msan_retval_tls, align 8503; CHECK-NEXT: ret { <16 x i32>, <16 x i32> } [[RES3]]504;505 %x2 = load <16 x i32>, ptr %x2p506 %res0 = call <16 x i32> @llvm.x86.avx512.mask.vpdpwssds.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2, i16 %x3)507 %res1 = call <16 x i32> @llvm.x86.avx512.maskz.vpdpwssds.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x4, i16 %x3)508 %res2 = insertvalue { <16 x i32>, <16 x i32> } poison, <16 x i32> %res0, 0509 %res3 = insertvalue { <16 x i32>, <16 x i32> } %res2, <16 x i32> %res1, 1510 ret { <16 x i32>, <16 x i32> } %res3511}512 513;.514; CHECK: [[PROF1]] = !{!"branch_weights", i32 1, i32 1048575}515;.516