503 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.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.vpdpbusd.512(<16 x i32>, <64 x i8>, <64 x i8>)14 15define <16 x i32> @test_int_x86_avx512_ask_vpdpbusd_512(<16 x i32> %x0, <64 x i8> %x1, <64 x i8> %x2) sanitize_memory {16; CHECK-LABEL: define <16 x i32> @test_int_x86_avx512_ask_vpdpbusd_512(17; CHECK-SAME: <16 x i32> [[X0:%.*]], <64 x i8> [[X1:%.*]], <64 x i8> [[X2:%.*]]) #[[ATTR1:[0-9]+]] {18; CHECK-NEXT: [[TMP24:%.*]] = load <64 x i8>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 64), align 819; CHECK-NEXT: [[TMP5:%.*]] = load <64 x i8>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 128), align 820; CHECK-NEXT: [[TMP23:%.*]] = load <16 x i32>, ptr @__msan_param_tls, align 821; CHECK-NEXT: call void @llvm.donothing()22; CHECK-NEXT: [[TMP10:%.*]] = icmp ne <64 x i8> [[TMP24]], zeroinitializer23; CHECK-NEXT: [[TMP11:%.*]] = icmp ne <64 x i8> [[TMP5]], zeroinitializer24; CHECK-NEXT: [[TMP6:%.*]] = icmp ne <64 x i8> [[X1]], zeroinitializer25; CHECK-NEXT: [[TMP7:%.*]] = icmp ne <64 x i8> [[X2]], zeroinitializer26; CHECK-NEXT: [[TMP12:%.*]] = and <64 x i1> [[TMP10]], [[TMP11]]27; CHECK-NEXT: [[TMP13:%.*]] = and <64 x i1> [[TMP6]], [[TMP11]]28; CHECK-NEXT: [[TMP14:%.*]] = and <64 x i1> [[TMP10]], [[TMP7]]29; CHECK-NEXT: [[TMP15:%.*]] = or <64 x i1> [[TMP12]], [[TMP13]]30; CHECK-NEXT: [[TMP16:%.*]] = or <64 x i1> [[TMP15]], [[TMP14]]31; CHECK-NEXT: [[TMP17:%.*]] = sext <64 x i1> [[TMP16]] to <64 x i8>32; CHECK-NEXT: [[TMP18:%.*]] = bitcast <64 x i8> [[TMP17]] to <16 x i32>33; CHECK-NEXT: [[TMP19:%.*]] = icmp ne <16 x i32> [[TMP18]], zeroinitializer34; CHECK-NEXT: [[TMP22:%.*]] = sext <16 x i1> [[TMP19]] to <16 x i32>35; CHECK-NEXT: [[_MSPROP1:%.*]] = or <16 x i32> [[TMP22]], [[TMP23]]36; CHECK-NEXT: [[TMP4:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpbusd.512(<16 x i32> [[X0]], <64 x i8> [[X1]], <64 x i8> [[X2]])37; CHECK-NEXT: store <16 x i32> [[_MSPROP1]], ptr @__msan_retval_tls, align 838; CHECK-NEXT: ret <16 x i32> [[TMP4]]39;40 %1 = call <16 x i32> @llvm.x86.avx512.vpdpbusd.512(<16 x i32> %x0, <64 x i8> %x1, <64 x i8> %x2)41 ret <16 x i32> %142}43 44define { <16 x i32>, <16 x i32> } @test_int_x86_avx512_mask_vpdpbusd_512(<16 x i32> %x0, <64 x i8> %x1, ptr %x2p, <64 x i8> %x4, i16 %x3) sanitize_memory {45; CHECK-LABEL: define { <16 x i32>, <16 x i32> } @test_int_x86_avx512_mask_vpdpbusd_512(46; CHECK-SAME: <16 x i32> [[X0:%.*]], <64 x i8> [[X1:%.*]], ptr [[X2P:%.*]], <64 x i8> [[X4:%.*]], i16 [[X3:%.*]]) #[[ATTR1]] {47; CHECK-NEXT: [[TMP1:%.*]] = load i64, ptr getelementptr (i8, ptr @__msan_param_tls, i64 128), align 848; CHECK-NEXT: [[TMP33:%.*]] = load <64 x i8>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 64), align 849; CHECK-NEXT: [[TMP2:%.*]] = load <16 x i32>, ptr @__msan_param_tls, align 850; CHECK-NEXT: [[TMP4:%.*]] = load i16, ptr getelementptr (i8, ptr @__msan_param_tls, i64 200), align 851; CHECK-NEXT: [[TMP40:%.*]] = load <64 x i8>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 136), align 852; CHECK-NEXT: call void @llvm.donothing()53; CHECK-NEXT: [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 054; CHECK-NEXT: br i1 [[_MSCMP]], label %[[BB6:.*]], label %[[BB7:.*]], !prof [[PROF1:![0-9]+]]55; CHECK: [[BB6]]:56; CHECK-NEXT: call void @__msan_warning_noreturn() #[[ATTR4:[0-9]+]]57; CHECK-NEXT: unreachable58; CHECK: [[BB7]]:59; CHECK-NEXT: [[TMP31:%.*]] = load <64 x i8>, ptr [[X2P]], align 6460; CHECK-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[X2P]] to i6461; CHECK-NEXT: [[TMP9:%.*]] = xor i64 [[TMP8]], 8796093022208062; CHECK-NEXT: [[TMP10:%.*]] = inttoptr i64 [[TMP9]] to ptr63; CHECK-NEXT: [[TMP30:%.*]] = load <64 x i8>, ptr [[TMP10]], align 6464; CHECK-NEXT: [[TMP35:%.*]] = icmp ne <64 x i8> [[TMP33]], zeroinitializer65; CHECK-NEXT: [[TMP36:%.*]] = icmp ne <64 x i8> [[TMP30]], zeroinitializer66; CHECK-NEXT: [[TMP32:%.*]] = icmp ne <64 x i8> [[X1]], zeroinitializer67; CHECK-NEXT: [[TMP37:%.*]] = icmp ne <64 x i8> [[TMP31]], zeroinitializer68; CHECK-NEXT: [[TMP38:%.*]] = and <64 x i1> [[TMP35]], [[TMP36]]69; CHECK-NEXT: [[TMP58:%.*]] = and <64 x i1> [[TMP32]], [[TMP36]]70; CHECK-NEXT: [[TMP59:%.*]] = and <64 x i1> [[TMP35]], [[TMP37]]71; CHECK-NEXT: [[TMP60:%.*]] = or <64 x i1> [[TMP38]], [[TMP58]]72; CHECK-NEXT: [[TMP61:%.*]] = or <64 x i1> [[TMP60]], [[TMP59]]73; CHECK-NEXT: [[TMP62:%.*]] = sext <64 x i1> [[TMP61]] to <64 x i8>74; CHECK-NEXT: [[TMP53:%.*]] = bitcast <64 x i8> [[TMP62]] to <16 x i32>75; CHECK-NEXT: [[TMP54:%.*]] = icmp ne <16 x i32> [[TMP53]], zeroinitializer76; CHECK-NEXT: [[TMP29:%.*]] = sext <16 x i1> [[TMP54]] to <16 x i32>77; CHECK-NEXT: [[_MSPROP1:%.*]] = or <16 x i32> [[TMP29]], [[TMP2]]78; CHECK-NEXT: [[TMP11:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpbusd.512(<16 x i32> [[X0]], <64 x i8> [[X1]], <64 x i8> [[TMP31]])79; CHECK-NEXT: [[TMP12:%.*]] = bitcast i16 [[TMP4]] to <16 x i1>80; CHECK-NEXT: [[TMP13:%.*]] = bitcast i16 [[X3]] to <16 x i1>81; CHECK-NEXT: [[TMP14:%.*]] = select <16 x i1> [[TMP13]], <16 x i32> [[_MSPROP1]], <16 x i32> [[TMP2]]82; CHECK-NEXT: [[TMP15:%.*]] = xor <16 x i32> [[TMP11]], [[X0]]83; CHECK-NEXT: [[TMP16:%.*]] = or <16 x i32> [[TMP15]], [[_MSPROP1]]84; CHECK-NEXT: [[TMP17:%.*]] = or <16 x i32> [[TMP16]], [[TMP2]]85; CHECK-NEXT: [[_MSPROP_SELECT:%.*]] = select <16 x i1> [[TMP12]], <16 x i32> [[TMP17]], <16 x i32> [[TMP14]]86; CHECK-NEXT: [[TMP18:%.*]] = select <16 x i1> [[TMP13]], <16 x i32> [[TMP11]], <16 x i32> [[X0]]87; CHECK-NEXT: [[TMP39:%.*]] = icmp ne <64 x i8> [[TMP33]], zeroinitializer88; CHECK-NEXT: [[TMP46:%.*]] = icmp ne <64 x i8> [[TMP40]], zeroinitializer89; CHECK-NEXT: [[TMP41:%.*]] = icmp ne <64 x i8> [[X1]], zeroinitializer90; CHECK-NEXT: [[TMP42:%.*]] = icmp ne <64 x i8> [[X4]], zeroinitializer91; CHECK-NEXT: [[TMP47:%.*]] = and <64 x i1> [[TMP39]], [[TMP46]]92; CHECK-NEXT: [[TMP48:%.*]] = and <64 x i1> [[TMP41]], [[TMP46]]93; CHECK-NEXT: [[TMP49:%.*]] = and <64 x i1> [[TMP39]], [[TMP42]]94; CHECK-NEXT: [[TMP50:%.*]] = or <64 x i1> [[TMP47]], [[TMP48]]95; CHECK-NEXT: [[TMP51:%.*]] = or <64 x i1> [[TMP50]], [[TMP49]]96; CHECK-NEXT: [[TMP52:%.*]] = sext <64 x i1> [[TMP51]] to <64 x i8>97; CHECK-NEXT: [[TMP55:%.*]] = bitcast <64 x i8> [[TMP52]] to <16 x i32>98; CHECK-NEXT: [[TMP56:%.*]] = icmp ne <16 x i32> [[TMP55]], zeroinitializer99; CHECK-NEXT: [[TMP57:%.*]] = sext <16 x i1> [[TMP56]] to <16 x i32>100; CHECK-NEXT: [[_MSPROP3:%.*]] = or <16 x i32> [[TMP57]], [[TMP2]]101; CHECK-NEXT: [[TMP19:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpbusd.512(<16 x i32> [[X0]], <64 x i8> [[X1]], <64 x i8> [[X4]])102; CHECK-NEXT: [[TMP20:%.*]] = bitcast i16 [[TMP4]] to <16 x i1>103; CHECK-NEXT: [[TMP21:%.*]] = bitcast i16 [[X3]] to <16 x i1>104; CHECK-NEXT: [[TMP22:%.*]] = select <16 x i1> [[TMP21]], <16 x i32> [[_MSPROP3]], <16 x i32> zeroinitializer105; CHECK-NEXT: [[TMP23:%.*]] = xor <16 x i32> [[TMP19]], zeroinitializer106; CHECK-NEXT: [[TMP24:%.*]] = or <16 x i32> [[TMP23]], [[_MSPROP3]]107; CHECK-NEXT: [[TMP25:%.*]] = or <16 x i32> [[TMP24]], zeroinitializer108; CHECK-NEXT: [[_MSPROP_SELECT4:%.*]] = select <16 x i1> [[TMP20]], <16 x i32> [[TMP25]], <16 x i32> [[TMP22]]109; CHECK-NEXT: [[TMP26:%.*]] = select <16 x i1> [[TMP21]], <16 x i32> [[TMP19]], <16 x i32> zeroinitializer110; 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]], 0111; CHECK-NEXT: [[RES1:%.*]] = insertvalue { <16 x i32>, <16 x i32> } poison, <16 x i32> [[TMP18]], 0112; CHECK-NEXT: [[TMP28:%.*]] = insertvalue { <16 x i32>, <16 x i32> } [[TMP27]], <16 x i32> [[_MSPROP_SELECT4]], 1113; CHECK-NEXT: [[RES2:%.*]] = insertvalue { <16 x i32>, <16 x i32> } [[RES1]], <16 x i32> [[TMP26]], 1114; CHECK-NEXT: store { <16 x i32>, <16 x i32> } [[TMP28]], ptr @__msan_retval_tls, align 8115; CHECK-NEXT: ret { <16 x i32>, <16 x i32> } [[RES2]]116;117 %x2 = load <64 x i8>, ptr %x2p118 %1 = call <16 x i32> @llvm.x86.avx512.vpdpbusd.512(<16 x i32> %x0, <64 x i8> %x1, <64 x i8> %x2)119 %2 = bitcast i16 %x3 to <16 x i1>120 %3 = select <16 x i1> %2, <16 x i32> %1, <16 x i32> %x0121 %4 = call <16 x i32> @llvm.x86.avx512.vpdpbusd.512(<16 x i32> %x0, <64 x i8> %x1, <64 x i8> %x4)122 %5 = bitcast i16 %x3 to <16 x i1>123 %6 = select <16 x i1> %5, <16 x i32> %4, <16 x i32> zeroinitializer124 %res1 = insertvalue { <16 x i32>, <16 x i32> } poison, <16 x i32> %3, 0125 %res2 = insertvalue { <16 x i32>, <16 x i32> } %res1, <16 x i32> %6, 1126 ret { <16 x i32>, <16 x i32> } %res2127}128 129declare <16 x i32> @llvm.x86.avx512.vpdpbusds.512(<16 x i32>, <64 x i8>, <64 x i8>)130 131define <16 x i32>@test_int_x86_avx512_vpdpbusds_512(<16 x i32> %x0, <64 x i8> %x1, <64 x i8> %x2) sanitize_memory {132; CHECK-LABEL: define <16 x i32> @test_int_x86_avx512_vpdpbusds_512(133; CHECK-SAME: <16 x i32> [[X0:%.*]], <64 x i8> [[X1:%.*]], <64 x i8> [[X2:%.*]]) #[[ATTR1]] {134; CHECK-NEXT: [[TMP24:%.*]] = load <64 x i8>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 64), align 8135; CHECK-NEXT: [[TMP5:%.*]] = load <64 x i8>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 128), align 8136; CHECK-NEXT: [[TMP23:%.*]] = load <16 x i32>, ptr @__msan_param_tls, align 8137; CHECK-NEXT: call void @llvm.donothing()138; CHECK-NEXT: [[TMP10:%.*]] = icmp ne <64 x i8> [[TMP24]], zeroinitializer139; CHECK-NEXT: [[TMP11:%.*]] = icmp ne <64 x i8> [[TMP5]], zeroinitializer140; CHECK-NEXT: [[TMP6:%.*]] = icmp ne <64 x i8> [[X1]], zeroinitializer141; CHECK-NEXT: [[TMP7:%.*]] = icmp ne <64 x i8> [[X2]], zeroinitializer142; CHECK-NEXT: [[TMP12:%.*]] = and <64 x i1> [[TMP10]], [[TMP11]]143; CHECK-NEXT: [[TMP13:%.*]] = and <64 x i1> [[TMP6]], [[TMP11]]144; CHECK-NEXT: [[TMP14:%.*]] = and <64 x i1> [[TMP10]], [[TMP7]]145; CHECK-NEXT: [[TMP15:%.*]] = or <64 x i1> [[TMP12]], [[TMP13]]146; CHECK-NEXT: [[TMP16:%.*]] = or <64 x i1> [[TMP15]], [[TMP14]]147; CHECK-NEXT: [[TMP17:%.*]] = sext <64 x i1> [[TMP16]] to <64 x i8>148; CHECK-NEXT: [[TMP18:%.*]] = bitcast <64 x i8> [[TMP17]] to <16 x i32>149; CHECK-NEXT: [[TMP19:%.*]] = icmp ne <16 x i32> [[TMP18]], zeroinitializer150; CHECK-NEXT: [[TMP22:%.*]] = sext <16 x i1> [[TMP19]] to <16 x i32>151; CHECK-NEXT: [[_MSPROP1:%.*]] = or <16 x i32> [[TMP22]], [[TMP23]]152; CHECK-NEXT: [[TMP4:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpbusds.512(<16 x i32> [[X0]], <64 x i8> [[X1]], <64 x i8> [[X2]])153; CHECK-NEXT: store <16 x i32> [[_MSPROP1]], ptr @__msan_retval_tls, align 8154; CHECK-NEXT: ret <16 x i32> [[TMP4]]155;156 %1 = call <16 x i32> @llvm.x86.avx512.vpdpbusds.512(<16 x i32> %x0, <64 x i8> %x1, <64 x i8> %x2)157 ret <16 x i32> %1158}159 160define { <16 x i32>, <16 x i32> } @test_int_x86_avx512_mask_vpdpbusds_512(<16 x i32> %x0, <64 x i8> %x1, ptr %x2p, <64 x i8> %x4, i16 %x3) sanitize_memory {161; CHECK-LABEL: define { <16 x i32>, <16 x i32> } @test_int_x86_avx512_mask_vpdpbusds_512(162; CHECK-SAME: <16 x i32> [[X0:%.*]], <64 x i8> [[X1:%.*]], ptr [[X2P:%.*]], <64 x i8> [[X4:%.*]], i16 [[X3:%.*]]) #[[ATTR1]] {163; CHECK-NEXT: [[TMP1:%.*]] = load i64, ptr getelementptr (i8, ptr @__msan_param_tls, i64 128), align 8164; CHECK-NEXT: [[TMP33:%.*]] = load <64 x i8>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 64), align 8165; CHECK-NEXT: [[TMP2:%.*]] = load <16 x i32>, ptr @__msan_param_tls, align 8166; CHECK-NEXT: [[TMP4:%.*]] = load i16, ptr getelementptr (i8, ptr @__msan_param_tls, i64 200), align 8167; CHECK-NEXT: [[TMP40:%.*]] = load <64 x i8>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 136), align 8168; CHECK-NEXT: call void @llvm.donothing()169; CHECK-NEXT: [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0170; CHECK-NEXT: br i1 [[_MSCMP]], label %[[BB6:.*]], label %[[BB7:.*]], !prof [[PROF1]]171; CHECK: [[BB6]]:172; CHECK-NEXT: call void @__msan_warning_noreturn() #[[ATTR4]]173; CHECK-NEXT: unreachable174; CHECK: [[BB7]]:175; CHECK-NEXT: [[TMP31:%.*]] = load <64 x i8>, ptr [[X2P]], align 64176; CHECK-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[X2P]] to i64177; CHECK-NEXT: [[TMP9:%.*]] = xor i64 [[TMP8]], 87960930222080178; CHECK-NEXT: [[TMP10:%.*]] = inttoptr i64 [[TMP9]] to ptr179; CHECK-NEXT: [[TMP30:%.*]] = load <64 x i8>, ptr [[TMP10]], align 64180; CHECK-NEXT: [[TMP35:%.*]] = icmp ne <64 x i8> [[TMP33]], zeroinitializer181; CHECK-NEXT: [[TMP36:%.*]] = icmp ne <64 x i8> [[TMP30]], zeroinitializer182; CHECK-NEXT: [[TMP32:%.*]] = icmp ne <64 x i8> [[X1]], zeroinitializer183; CHECK-NEXT: [[TMP37:%.*]] = icmp ne <64 x i8> [[TMP31]], zeroinitializer184; CHECK-NEXT: [[TMP38:%.*]] = and <64 x i1> [[TMP35]], [[TMP36]]185; CHECK-NEXT: [[TMP58:%.*]] = and <64 x i1> [[TMP32]], [[TMP36]]186; CHECK-NEXT: [[TMP59:%.*]] = and <64 x i1> [[TMP35]], [[TMP37]]187; CHECK-NEXT: [[TMP60:%.*]] = or <64 x i1> [[TMP38]], [[TMP58]]188; CHECK-NEXT: [[TMP61:%.*]] = or <64 x i1> [[TMP60]], [[TMP59]]189; CHECK-NEXT: [[TMP62:%.*]] = sext <64 x i1> [[TMP61]] to <64 x i8>190; CHECK-NEXT: [[TMP53:%.*]] = bitcast <64 x i8> [[TMP62]] to <16 x i32>191; CHECK-NEXT: [[TMP54:%.*]] = icmp ne <16 x i32> [[TMP53]], zeroinitializer192; CHECK-NEXT: [[TMP29:%.*]] = sext <16 x i1> [[TMP54]] to <16 x i32>193; CHECK-NEXT: [[_MSPROP1:%.*]] = or <16 x i32> [[TMP29]], [[TMP2]]194; CHECK-NEXT: [[TMP11:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpbusds.512(<16 x i32> [[X0]], <64 x i8> [[X1]], <64 x i8> [[TMP31]])195; CHECK-NEXT: [[TMP12:%.*]] = bitcast i16 [[TMP4]] to <16 x i1>196; CHECK-NEXT: [[TMP13:%.*]] = bitcast i16 [[X3]] to <16 x i1>197; CHECK-NEXT: [[TMP14:%.*]] = select <16 x i1> [[TMP13]], <16 x i32> [[_MSPROP1]], <16 x i32> [[TMP2]]198; CHECK-NEXT: [[TMP15:%.*]] = xor <16 x i32> [[TMP11]], [[X0]]199; CHECK-NEXT: [[TMP16:%.*]] = or <16 x i32> [[TMP15]], [[_MSPROP1]]200; CHECK-NEXT: [[TMP17:%.*]] = or <16 x i32> [[TMP16]], [[TMP2]]201; CHECK-NEXT: [[_MSPROP_SELECT:%.*]] = select <16 x i1> [[TMP12]], <16 x i32> [[TMP17]], <16 x i32> [[TMP14]]202; CHECK-NEXT: [[TMP18:%.*]] = select <16 x i1> [[TMP13]], <16 x i32> [[TMP11]], <16 x i32> [[X0]]203; CHECK-NEXT: [[TMP39:%.*]] = icmp ne <64 x i8> [[TMP33]], zeroinitializer204; CHECK-NEXT: [[TMP46:%.*]] = icmp ne <64 x i8> [[TMP40]], zeroinitializer205; CHECK-NEXT: [[TMP41:%.*]] = icmp ne <64 x i8> [[X1]], zeroinitializer206; CHECK-NEXT: [[TMP42:%.*]] = icmp ne <64 x i8> [[X4]], zeroinitializer207; CHECK-NEXT: [[TMP47:%.*]] = and <64 x i1> [[TMP39]], [[TMP46]]208; CHECK-NEXT: [[TMP48:%.*]] = and <64 x i1> [[TMP41]], [[TMP46]]209; CHECK-NEXT: [[TMP49:%.*]] = and <64 x i1> [[TMP39]], [[TMP42]]210; CHECK-NEXT: [[TMP50:%.*]] = or <64 x i1> [[TMP47]], [[TMP48]]211; CHECK-NEXT: [[TMP51:%.*]] = or <64 x i1> [[TMP50]], [[TMP49]]212; CHECK-NEXT: [[TMP52:%.*]] = sext <64 x i1> [[TMP51]] to <64 x i8>213; CHECK-NEXT: [[TMP55:%.*]] = bitcast <64 x i8> [[TMP52]] to <16 x i32>214; CHECK-NEXT: [[TMP56:%.*]] = icmp ne <16 x i32> [[TMP55]], zeroinitializer215; CHECK-NEXT: [[TMP57:%.*]] = sext <16 x i1> [[TMP56]] to <16 x i32>216; CHECK-NEXT: [[_MSPROP3:%.*]] = or <16 x i32> [[TMP57]], [[TMP2]]217; CHECK-NEXT: [[TMP19:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpbusds.512(<16 x i32> [[X0]], <64 x i8> [[X1]], <64 x i8> [[X4]])218; CHECK-NEXT: [[TMP20:%.*]] = bitcast i16 [[TMP4]] to <16 x i1>219; CHECK-NEXT: [[TMP21:%.*]] = bitcast i16 [[X3]] to <16 x i1>220; CHECK-NEXT: [[TMP22:%.*]] = select <16 x i1> [[TMP21]], <16 x i32> [[_MSPROP3]], <16 x i32> zeroinitializer221; CHECK-NEXT: [[TMP23:%.*]] = xor <16 x i32> [[TMP19]], zeroinitializer222; CHECK-NEXT: [[TMP24:%.*]] = or <16 x i32> [[TMP23]], [[_MSPROP3]]223; CHECK-NEXT: [[TMP25:%.*]] = or <16 x i32> [[TMP24]], zeroinitializer224; CHECK-NEXT: [[_MSPROP_SELECT4:%.*]] = select <16 x i1> [[TMP20]], <16 x i32> [[TMP25]], <16 x i32> [[TMP22]]225; CHECK-NEXT: [[TMP26:%.*]] = select <16 x i1> [[TMP21]], <16 x i32> [[TMP19]], <16 x i32> zeroinitializer226; 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]], 0227; CHECK-NEXT: [[RES1:%.*]] = insertvalue { <16 x i32>, <16 x i32> } poison, <16 x i32> [[TMP18]], 0228; CHECK-NEXT: [[TMP28:%.*]] = insertvalue { <16 x i32>, <16 x i32> } [[TMP27]], <16 x i32> [[_MSPROP_SELECT4]], 1229; CHECK-NEXT: [[RES2:%.*]] = insertvalue { <16 x i32>, <16 x i32> } [[RES1]], <16 x i32> [[TMP26]], 1230; CHECK-NEXT: store { <16 x i32>, <16 x i32> } [[TMP28]], ptr @__msan_retval_tls, align 8231; CHECK-NEXT: ret { <16 x i32>, <16 x i32> } [[RES2]]232;233 %x2 = load <64 x i8>, ptr %x2p234 %1 = call <16 x i32> @llvm.x86.avx512.vpdpbusds.512(<16 x i32> %x0, <64 x i8> %x1, <64 x i8> %x2)235 %2 = bitcast i16 %x3 to <16 x i1>236 %3 = select <16 x i1> %2, <16 x i32> %1, <16 x i32> %x0237 %4 = call <16 x i32> @llvm.x86.avx512.vpdpbusds.512(<16 x i32> %x0, <64 x i8> %x1, <64 x i8> %x4)238 %5 = bitcast i16 %x3 to <16 x i1>239 %6 = select <16 x i1> %5, <16 x i32> %4, <16 x i32> zeroinitializer240 %res1 = insertvalue { <16 x i32>, <16 x i32> } poison, <16 x i32> %3, 0241 %res2 = insertvalue { <16 x i32>, <16 x i32> } %res1, <16 x i32> %6, 1242 ret { <16 x i32>, <16 x i32> } %res2243}244 245declare <16 x i32> @llvm.x86.avx512.vpdpwssd.512(<16 x i32>, <16 x i32>, <16 x i32>)246 247define <16 x i32>@test_int_x86_avx512_vpdpwssd_512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2) sanitize_memory {248; CHECK-LABEL: define <16 x i32> @test_int_x86_avx512_vpdpwssd_512(249; CHECK-SAME: <16 x i32> [[X0:%.*]], <16 x i32> [[X1:%.*]], <16 x i32> [[X2:%.*]]) #[[ATTR1]] {250; CHECK-NEXT: [[TMP2:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 64), align 8251; CHECK-NEXT: [[TMP3:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 128), align 8252; CHECK-NEXT: [[TMP21:%.*]] = load <16 x i32>, ptr @__msan_param_tls, align 8253; CHECK-NEXT: call void @llvm.donothing()254; CHECK-NEXT: [[TMP22:%.*]] = bitcast <16 x i32> [[X1]] to <32 x i16>255; CHECK-NEXT: [[TMP5:%.*]] = bitcast <16 x i32> [[X2]] to <32 x i16>256; CHECK-NEXT: [[TMP6:%.*]] = bitcast <16 x i32> [[TMP2]] to <32 x i16>257; CHECK-NEXT: [[TMP7:%.*]] = bitcast <16 x i32> [[TMP3]] to <32 x i16>258; CHECK-NEXT: [[TMP8:%.*]] = icmp ne <32 x i16> [[TMP6]], zeroinitializer259; CHECK-NEXT: [[TMP9:%.*]] = icmp ne <32 x i16> [[TMP7]], zeroinitializer260; CHECK-NEXT: [[TMP10:%.*]] = icmp ne <32 x i16> [[TMP22]], zeroinitializer261; CHECK-NEXT: [[TMP11:%.*]] = icmp ne <32 x i16> [[TMP5]], zeroinitializer262; CHECK-NEXT: [[TMP12:%.*]] = and <32 x i1> [[TMP8]], [[TMP9]]263; CHECK-NEXT: [[TMP13:%.*]] = and <32 x i1> [[TMP10]], [[TMP9]]264; CHECK-NEXT: [[TMP14:%.*]] = and <32 x i1> [[TMP8]], [[TMP11]]265; CHECK-NEXT: [[TMP15:%.*]] = or <32 x i1> [[TMP12]], [[TMP13]]266; CHECK-NEXT: [[TMP16:%.*]] = or <32 x i1> [[TMP15]], [[TMP14]]267; CHECK-NEXT: [[TMP17:%.*]] = sext <32 x i1> [[TMP16]] to <32 x i16>268; CHECK-NEXT: [[TMP18:%.*]] = bitcast <32 x i16> [[TMP17]] to <16 x i32>269; CHECK-NEXT: [[TMP19:%.*]] = icmp ne <16 x i32> [[TMP18]], zeroinitializer270; CHECK-NEXT: [[TMP20:%.*]] = sext <16 x i1> [[TMP19]] to <16 x i32>271; CHECK-NEXT: [[_MSPROP1:%.*]] = or <16 x i32> [[TMP20]], [[TMP21]]272; CHECK-NEXT: [[TMP4:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpwssd.512(<16 x i32> [[X0]], <16 x i32> [[X1]], <16 x i32> [[X2]])273; CHECK-NEXT: store <16 x i32> [[_MSPROP1]], ptr @__msan_retval_tls, align 8274; CHECK-NEXT: ret <16 x i32> [[TMP4]]275;276 %1 = call <16 x i32> @llvm.x86.avx512.vpdpwssd.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2)277 ret <16 x i32> %1278}279 280define { <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 {281; CHECK-LABEL: define { <16 x i32>, <16 x i32> } @test_int_x86_avx512_mask_vpdpwssd_512(282; CHECK-SAME: <16 x i32> [[X0:%.*]], <16 x i32> [[X1:%.*]], ptr [[X2P:%.*]], <16 x i32> [[X4:%.*]], i16 [[X3:%.*]]) #[[ATTR1]] {283; CHECK-NEXT: [[TMP1:%.*]] = load i64, ptr getelementptr (i8, ptr @__msan_param_tls, i64 128), align 8284; CHECK-NEXT: [[TMP3:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 64), align 8285; CHECK-NEXT: [[TMP2:%.*]] = load <16 x i32>, ptr @__msan_param_tls, align 8286; CHECK-NEXT: [[TMP4:%.*]] = load i16, ptr getelementptr (i8, ptr @__msan_param_tls, i64 200), align 8287; CHECK-NEXT: [[TMP5:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 136), align 8288; CHECK-NEXT: call void @llvm.donothing()289; CHECK-NEXT: [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0290; CHECK-NEXT: br i1 [[_MSCMP]], label %[[BB6:.*]], label %[[BB7:.*]], !prof [[PROF1]]291; CHECK: [[BB6]]:292; CHECK-NEXT: call void @__msan_warning_noreturn() #[[ATTR4]]293; CHECK-NEXT: unreachable294; CHECK: [[BB7]]:295; CHECK-NEXT: [[X2:%.*]] = load <16 x i32>, ptr [[X2P]], align 64296; CHECK-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[X2P]] to i64297; CHECK-NEXT: [[TMP9:%.*]] = xor i64 [[TMP8]], 87960930222080298; CHECK-NEXT: [[TMP10:%.*]] = inttoptr i64 [[TMP9]] to ptr299; CHECK-NEXT: [[_MSLD:%.*]] = load <16 x i32>, ptr [[TMP10]], align 64300; CHECK-NEXT: [[TMP29:%.*]] = bitcast <16 x i32> [[X1]] to <32 x i16>301; CHECK-NEXT: [[TMP30:%.*]] = bitcast <16 x i32> [[X2]] to <32 x i16>302; CHECK-NEXT: [[TMP31:%.*]] = bitcast <16 x i32> [[TMP3]] to <32 x i16>303; CHECK-NEXT: [[TMP32:%.*]] = bitcast <16 x i32> [[_MSLD]] to <32 x i16>304; CHECK-NEXT: [[TMP33:%.*]] = icmp ne <32 x i16> [[TMP31]], zeroinitializer305; CHECK-NEXT: [[TMP34:%.*]] = icmp ne <32 x i16> [[TMP32]], zeroinitializer306; CHECK-NEXT: [[TMP35:%.*]] = icmp ne <32 x i16> [[TMP29]], zeroinitializer307; CHECK-NEXT: [[TMP36:%.*]] = icmp ne <32 x i16> [[TMP30]], zeroinitializer308; CHECK-NEXT: [[TMP54:%.*]] = and <32 x i1> [[TMP33]], [[TMP34]]309; CHECK-NEXT: [[TMP55:%.*]] = and <32 x i1> [[TMP35]], [[TMP34]]310; CHECK-NEXT: [[TMP56:%.*]] = and <32 x i1> [[TMP33]], [[TMP36]]311; CHECK-NEXT: [[TMP57:%.*]] = or <32 x i1> [[TMP54]], [[TMP55]]312; CHECK-NEXT: [[TMP58:%.*]] = or <32 x i1> [[TMP57]], [[TMP56]]313; CHECK-NEXT: [[TMP59:%.*]] = sext <32 x i1> [[TMP58]] to <32 x i16>314; CHECK-NEXT: [[TMP60:%.*]] = bitcast <32 x i16> [[TMP59]] to <16 x i32>315; CHECK-NEXT: [[TMP61:%.*]] = icmp ne <16 x i32> [[TMP60]], zeroinitializer316; CHECK-NEXT: [[TMP62:%.*]] = sext <16 x i1> [[TMP61]] to <16 x i32>317; CHECK-NEXT: [[_MSPROP1:%.*]] = or <16 x i32> [[TMP62]], [[TMP2]]318; CHECK-NEXT: [[TMP11:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpwssd.512(<16 x i32> [[X0]], <16 x i32> [[X1]], <16 x i32> [[X2]])319; CHECK-NEXT: [[TMP12:%.*]] = bitcast i16 [[TMP4]] to <16 x i1>320; CHECK-NEXT: [[TMP13:%.*]] = bitcast i16 [[X3]] to <16 x i1>321; CHECK-NEXT: [[TMP14:%.*]] = select <16 x i1> [[TMP13]], <16 x i32> [[_MSPROP1]], <16 x i32> [[TMP2]]322; CHECK-NEXT: [[TMP15:%.*]] = xor <16 x i32> [[TMP11]], [[X0]]323; CHECK-NEXT: [[TMP16:%.*]] = or <16 x i32> [[TMP15]], [[_MSPROP1]]324; CHECK-NEXT: [[TMP17:%.*]] = or <16 x i32> [[TMP16]], [[TMP2]]325; CHECK-NEXT: [[_MSPROP_SELECT:%.*]] = select <16 x i1> [[TMP12]], <16 x i32> [[TMP17]], <16 x i32> [[TMP14]]326; CHECK-NEXT: [[TMP18:%.*]] = select <16 x i1> [[TMP13]], <16 x i32> [[TMP11]], <16 x i32> [[X0]]327; CHECK-NEXT: [[TMP37:%.*]] = bitcast <16 x i32> [[X1]] to <32 x i16>328; CHECK-NEXT: [[TMP38:%.*]] = bitcast <16 x i32> [[X4]] to <32 x i16>329; CHECK-NEXT: [[TMP39:%.*]] = bitcast <16 x i32> [[TMP3]] to <32 x i16>330; CHECK-NEXT: [[TMP40:%.*]] = bitcast <16 x i32> [[TMP5]] to <32 x i16>331; CHECK-NEXT: [[TMP41:%.*]] = icmp ne <32 x i16> [[TMP39]], zeroinitializer332; CHECK-NEXT: [[TMP42:%.*]] = icmp ne <32 x i16> [[TMP40]], zeroinitializer333; CHECK-NEXT: [[TMP43:%.*]] = icmp ne <32 x i16> [[TMP37]], zeroinitializer334; CHECK-NEXT: [[TMP44:%.*]] = icmp ne <32 x i16> [[TMP38]], zeroinitializer335; CHECK-NEXT: [[TMP45:%.*]] = and <32 x i1> [[TMP41]], [[TMP42]]336; CHECK-NEXT: [[TMP46:%.*]] = and <32 x i1> [[TMP43]], [[TMP42]]337; CHECK-NEXT: [[TMP47:%.*]] = and <32 x i1> [[TMP41]], [[TMP44]]338; CHECK-NEXT: [[TMP48:%.*]] = or <32 x i1> [[TMP45]], [[TMP46]]339; CHECK-NEXT: [[TMP49:%.*]] = or <32 x i1> [[TMP48]], [[TMP47]]340; CHECK-NEXT: [[TMP50:%.*]] = sext <32 x i1> [[TMP49]] to <32 x i16>341; CHECK-NEXT: [[TMP51:%.*]] = bitcast <32 x i16> [[TMP50]] to <16 x i32>342; CHECK-NEXT: [[TMP52:%.*]] = icmp ne <16 x i32> [[TMP51]], zeroinitializer343; CHECK-NEXT: [[TMP53:%.*]] = sext <16 x i1> [[TMP52]] to <16 x i32>344; CHECK-NEXT: [[_MSPROP3:%.*]] = or <16 x i32> [[TMP53]], [[TMP2]]345; CHECK-NEXT: [[TMP19:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpwssd.512(<16 x i32> [[X0]], <16 x i32> [[X1]], <16 x i32> [[X4]])346; CHECK-NEXT: [[TMP20:%.*]] = bitcast i16 [[TMP4]] to <16 x i1>347; CHECK-NEXT: [[TMP21:%.*]] = bitcast i16 [[X3]] to <16 x i1>348; CHECK-NEXT: [[TMP22:%.*]] = select <16 x i1> [[TMP21]], <16 x i32> [[_MSPROP3]], <16 x i32> zeroinitializer349; CHECK-NEXT: [[TMP23:%.*]] = xor <16 x i32> [[TMP19]], zeroinitializer350; CHECK-NEXT: [[TMP24:%.*]] = or <16 x i32> [[TMP23]], [[_MSPROP3]]351; CHECK-NEXT: [[TMP25:%.*]] = or <16 x i32> [[TMP24]], zeroinitializer352; CHECK-NEXT: [[_MSPROP_SELECT4:%.*]] = select <16 x i1> [[TMP20]], <16 x i32> [[TMP25]], <16 x i32> [[TMP22]]353; CHECK-NEXT: [[TMP26:%.*]] = select <16 x i1> [[TMP21]], <16 x i32> [[TMP19]], <16 x i32> zeroinitializer354; 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]], 0355; CHECK-NEXT: [[RES1:%.*]] = insertvalue { <16 x i32>, <16 x i32> } poison, <16 x i32> [[TMP18]], 0356; CHECK-NEXT: [[TMP28:%.*]] = insertvalue { <16 x i32>, <16 x i32> } [[TMP27]], <16 x i32> [[_MSPROP_SELECT4]], 1357; CHECK-NEXT: [[RES2:%.*]] = insertvalue { <16 x i32>, <16 x i32> } [[RES1]], <16 x i32> [[TMP26]], 1358; CHECK-NEXT: store { <16 x i32>, <16 x i32> } [[TMP28]], ptr @__msan_retval_tls, align 8359; CHECK-NEXT: ret { <16 x i32>, <16 x i32> } [[RES2]]360;361 %x2 = load <16 x i32>, ptr %x2p362 %1 = call <16 x i32> @llvm.x86.avx512.vpdpwssd.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2)363 %2 = bitcast i16 %x3 to <16 x i1>364 %3 = select <16 x i1> %2, <16 x i32> %1, <16 x i32> %x0365 %4 = call <16 x i32> @llvm.x86.avx512.vpdpwssd.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x4)366 %5 = bitcast i16 %x3 to <16 x i1>367 %6 = select <16 x i1> %5, <16 x i32> %4, <16 x i32> zeroinitializer368 %res1 = insertvalue { <16 x i32>, <16 x i32> } poison, <16 x i32> %3, 0369 %res2 = insertvalue { <16 x i32>, <16 x i32> } %res1, <16 x i32> %6, 1370 ret { <16 x i32>, <16 x i32> } %res2371}372 373declare <16 x i32> @llvm.x86.avx512.vpdpwssds.512(<16 x i32>, <16 x i32>, <16 x i32>)374 375define <16 x i32>@test_int_x86_avx512_ask_vpdpwssds_512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2) sanitize_memory {376; CHECK-LABEL: define <16 x i32> @test_int_x86_avx512_ask_vpdpwssds_512(377; CHECK-SAME: <16 x i32> [[X0:%.*]], <16 x i32> [[X1:%.*]], <16 x i32> [[X2:%.*]]) #[[ATTR1]] {378; CHECK-NEXT: [[TMP2:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 64), align 8379; CHECK-NEXT: [[TMP3:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 128), align 8380; CHECK-NEXT: [[TMP21:%.*]] = load <16 x i32>, ptr @__msan_param_tls, align 8381; CHECK-NEXT: call void @llvm.donothing()382; CHECK-NEXT: [[TMP22:%.*]] = bitcast <16 x i32> [[X1]] to <32 x i16>383; CHECK-NEXT: [[TMP5:%.*]] = bitcast <16 x i32> [[X2]] to <32 x i16>384; CHECK-NEXT: [[TMP6:%.*]] = bitcast <16 x i32> [[TMP2]] to <32 x i16>385; CHECK-NEXT: [[TMP7:%.*]] = bitcast <16 x i32> [[TMP3]] to <32 x i16>386; CHECK-NEXT: [[TMP8:%.*]] = icmp ne <32 x i16> [[TMP6]], zeroinitializer387; CHECK-NEXT: [[TMP9:%.*]] = icmp ne <32 x i16> [[TMP7]], zeroinitializer388; CHECK-NEXT: [[TMP10:%.*]] = icmp ne <32 x i16> [[TMP22]], zeroinitializer389; CHECK-NEXT: [[TMP11:%.*]] = icmp ne <32 x i16> [[TMP5]], zeroinitializer390; CHECK-NEXT: [[TMP12:%.*]] = and <32 x i1> [[TMP8]], [[TMP9]]391; CHECK-NEXT: [[TMP13:%.*]] = and <32 x i1> [[TMP10]], [[TMP9]]392; CHECK-NEXT: [[TMP14:%.*]] = and <32 x i1> [[TMP8]], [[TMP11]]393; CHECK-NEXT: [[TMP15:%.*]] = or <32 x i1> [[TMP12]], [[TMP13]]394; CHECK-NEXT: [[TMP16:%.*]] = or <32 x i1> [[TMP15]], [[TMP14]]395; CHECK-NEXT: [[TMP17:%.*]] = sext <32 x i1> [[TMP16]] to <32 x i16>396; CHECK-NEXT: [[TMP18:%.*]] = bitcast <32 x i16> [[TMP17]] to <16 x i32>397; CHECK-NEXT: [[TMP19:%.*]] = icmp ne <16 x i32> [[TMP18]], zeroinitializer398; CHECK-NEXT: [[TMP20:%.*]] = sext <16 x i1> [[TMP19]] to <16 x i32>399; CHECK-NEXT: [[_MSPROP1:%.*]] = or <16 x i32> [[TMP20]], [[TMP21]]400; CHECK-NEXT: [[TMP4:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpwssds.512(<16 x i32> [[X0]], <16 x i32> [[X1]], <16 x i32> [[X2]])401; CHECK-NEXT: store <16 x i32> [[_MSPROP1]], ptr @__msan_retval_tls, align 8402; CHECK-NEXT: ret <16 x i32> [[TMP4]]403;404 %1 = call <16 x i32> @llvm.x86.avx512.vpdpwssds.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2)405 ret <16 x i32> %1406}407 408define { <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 {409; CHECK-LABEL: define { <16 x i32>, <16 x i32> } @test_int_x86_avx512_mask_vpdpwssds_512(410; CHECK-SAME: <16 x i32> [[X0:%.*]], <16 x i32> [[X1:%.*]], ptr [[X2P:%.*]], <16 x i32> [[X4:%.*]], i16 [[X3:%.*]]) #[[ATTR1]] {411; CHECK-NEXT: [[TMP1:%.*]] = load i64, ptr getelementptr (i8, ptr @__msan_param_tls, i64 128), align 8412; CHECK-NEXT: [[TMP3:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 64), align 8413; CHECK-NEXT: [[TMP2:%.*]] = load <16 x i32>, ptr @__msan_param_tls, align 8414; CHECK-NEXT: [[TMP4:%.*]] = load i16, ptr getelementptr (i8, ptr @__msan_param_tls, i64 200), align 8415; CHECK-NEXT: [[TMP5:%.*]] = load <16 x i32>, ptr getelementptr (i8, ptr @__msan_param_tls, i64 136), align 8416; CHECK-NEXT: call void @llvm.donothing()417; CHECK-NEXT: [[_MSCMP:%.*]] = icmp ne i64 [[TMP1]], 0418; CHECK-NEXT: br i1 [[_MSCMP]], label %[[BB6:.*]], label %[[BB7:.*]], !prof [[PROF1]]419; CHECK: [[BB6]]:420; CHECK-NEXT: call void @__msan_warning_noreturn() #[[ATTR4]]421; CHECK-NEXT: unreachable422; CHECK: [[BB7]]:423; CHECK-NEXT: [[X2:%.*]] = load <16 x i32>, ptr [[X2P]], align 64424; CHECK-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[X2P]] to i64425; CHECK-NEXT: [[TMP9:%.*]] = xor i64 [[TMP8]], 87960930222080426; CHECK-NEXT: [[TMP10:%.*]] = inttoptr i64 [[TMP9]] to ptr427; CHECK-NEXT: [[_MSLD:%.*]] = load <16 x i32>, ptr [[TMP10]], align 64428; CHECK-NEXT: [[TMP29:%.*]] = bitcast <16 x i32> [[X1]] to <32 x i16>429; CHECK-NEXT: [[TMP30:%.*]] = bitcast <16 x i32> [[X2]] to <32 x i16>430; CHECK-NEXT: [[TMP31:%.*]] = bitcast <16 x i32> [[TMP3]] to <32 x i16>431; CHECK-NEXT: [[TMP32:%.*]] = bitcast <16 x i32> [[_MSLD]] to <32 x i16>432; CHECK-NEXT: [[TMP33:%.*]] = icmp ne <32 x i16> [[TMP31]], zeroinitializer433; CHECK-NEXT: [[TMP34:%.*]] = icmp ne <32 x i16> [[TMP32]], zeroinitializer434; CHECK-NEXT: [[TMP35:%.*]] = icmp ne <32 x i16> [[TMP29]], zeroinitializer435; CHECK-NEXT: [[TMP36:%.*]] = icmp ne <32 x i16> [[TMP30]], zeroinitializer436; CHECK-NEXT: [[TMP54:%.*]] = and <32 x i1> [[TMP33]], [[TMP34]]437; CHECK-NEXT: [[TMP55:%.*]] = and <32 x i1> [[TMP35]], [[TMP34]]438; CHECK-NEXT: [[TMP56:%.*]] = and <32 x i1> [[TMP33]], [[TMP36]]439; CHECK-NEXT: [[TMP57:%.*]] = or <32 x i1> [[TMP54]], [[TMP55]]440; CHECK-NEXT: [[TMP58:%.*]] = or <32 x i1> [[TMP57]], [[TMP56]]441; CHECK-NEXT: [[TMP59:%.*]] = sext <32 x i1> [[TMP58]] to <32 x i16>442; CHECK-NEXT: [[TMP60:%.*]] = bitcast <32 x i16> [[TMP59]] to <16 x i32>443; CHECK-NEXT: [[TMP61:%.*]] = icmp ne <16 x i32> [[TMP60]], zeroinitializer444; CHECK-NEXT: [[TMP62:%.*]] = sext <16 x i1> [[TMP61]] to <16 x i32>445; CHECK-NEXT: [[_MSPROP1:%.*]] = or <16 x i32> [[TMP62]], [[TMP2]]446; CHECK-NEXT: [[TMP11:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpwssds.512(<16 x i32> [[X0]], <16 x i32> [[X1]], <16 x i32> [[X2]])447; CHECK-NEXT: [[TMP12:%.*]] = bitcast i16 [[TMP4]] to <16 x i1>448; CHECK-NEXT: [[TMP13:%.*]] = bitcast i16 [[X3]] to <16 x i1>449; CHECK-NEXT: [[TMP14:%.*]] = select <16 x i1> [[TMP13]], <16 x i32> [[_MSPROP1]], <16 x i32> [[TMP2]]450; CHECK-NEXT: [[TMP15:%.*]] = xor <16 x i32> [[TMP11]], [[X0]]451; CHECK-NEXT: [[TMP16:%.*]] = or <16 x i32> [[TMP15]], [[_MSPROP1]]452; CHECK-NEXT: [[TMP17:%.*]] = or <16 x i32> [[TMP16]], [[TMP2]]453; CHECK-NEXT: [[_MSPROP_SELECT:%.*]] = select <16 x i1> [[TMP12]], <16 x i32> [[TMP17]], <16 x i32> [[TMP14]]454; CHECK-NEXT: [[TMP18:%.*]] = select <16 x i1> [[TMP13]], <16 x i32> [[TMP11]], <16 x i32> [[X0]]455; CHECK-NEXT: [[TMP37:%.*]] = bitcast <16 x i32> [[X1]] to <32 x i16>456; CHECK-NEXT: [[TMP38:%.*]] = bitcast <16 x i32> [[X4]] to <32 x i16>457; CHECK-NEXT: [[TMP39:%.*]] = bitcast <16 x i32> [[TMP3]] to <32 x i16>458; CHECK-NEXT: [[TMP40:%.*]] = bitcast <16 x i32> [[TMP5]] to <32 x i16>459; CHECK-NEXT: [[TMP41:%.*]] = icmp ne <32 x i16> [[TMP39]], zeroinitializer460; CHECK-NEXT: [[TMP42:%.*]] = icmp ne <32 x i16> [[TMP40]], zeroinitializer461; CHECK-NEXT: [[TMP43:%.*]] = icmp ne <32 x i16> [[TMP37]], zeroinitializer462; CHECK-NEXT: [[TMP44:%.*]] = icmp ne <32 x i16> [[TMP38]], zeroinitializer463; CHECK-NEXT: [[TMP45:%.*]] = and <32 x i1> [[TMP41]], [[TMP42]]464; CHECK-NEXT: [[TMP46:%.*]] = and <32 x i1> [[TMP43]], [[TMP42]]465; CHECK-NEXT: [[TMP47:%.*]] = and <32 x i1> [[TMP41]], [[TMP44]]466; CHECK-NEXT: [[TMP48:%.*]] = or <32 x i1> [[TMP45]], [[TMP46]]467; CHECK-NEXT: [[TMP49:%.*]] = or <32 x i1> [[TMP48]], [[TMP47]]468; CHECK-NEXT: [[TMP50:%.*]] = sext <32 x i1> [[TMP49]] to <32 x i16>469; CHECK-NEXT: [[TMP51:%.*]] = bitcast <32 x i16> [[TMP50]] to <16 x i32>470; CHECK-NEXT: [[TMP52:%.*]] = icmp ne <16 x i32> [[TMP51]], zeroinitializer471; CHECK-NEXT: [[TMP53:%.*]] = sext <16 x i1> [[TMP52]] to <16 x i32>472; CHECK-NEXT: [[_MSPROP3:%.*]] = or <16 x i32> [[TMP53]], [[TMP2]]473; CHECK-NEXT: [[TMP19:%.*]] = call <16 x i32> @llvm.x86.avx512.vpdpwssds.512(<16 x i32> [[X0]], <16 x i32> [[X1]], <16 x i32> [[X4]])474; CHECK-NEXT: [[TMP20:%.*]] = bitcast i16 [[TMP4]] to <16 x i1>475; CHECK-NEXT: [[TMP21:%.*]] = bitcast i16 [[X3]] to <16 x i1>476; CHECK-NEXT: [[TMP22:%.*]] = select <16 x i1> [[TMP21]], <16 x i32> [[_MSPROP3]], <16 x i32> zeroinitializer477; CHECK-NEXT: [[TMP23:%.*]] = xor <16 x i32> [[TMP19]], zeroinitializer478; CHECK-NEXT: [[TMP24:%.*]] = or <16 x i32> [[TMP23]], [[_MSPROP3]]479; CHECK-NEXT: [[TMP25:%.*]] = or <16 x i32> [[TMP24]], zeroinitializer480; CHECK-NEXT: [[_MSPROP_SELECT4:%.*]] = select <16 x i1> [[TMP20]], <16 x i32> [[TMP25]], <16 x i32> [[TMP22]]481; CHECK-NEXT: [[TMP26:%.*]] = select <16 x i1> [[TMP21]], <16 x i32> [[TMP19]], <16 x i32> zeroinitializer482; 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]], 0483; CHECK-NEXT: [[RES1:%.*]] = insertvalue { <16 x i32>, <16 x i32> } poison, <16 x i32> [[TMP18]], 0484; CHECK-NEXT: [[TMP28:%.*]] = insertvalue { <16 x i32>, <16 x i32> } [[TMP27]], <16 x i32> [[_MSPROP_SELECT4]], 1485; CHECK-NEXT: [[RES2:%.*]] = insertvalue { <16 x i32>, <16 x i32> } [[RES1]], <16 x i32> [[TMP26]], 1486; CHECK-NEXT: store { <16 x i32>, <16 x i32> } [[TMP28]], ptr @__msan_retval_tls, align 8487; CHECK-NEXT: ret { <16 x i32>, <16 x i32> } [[RES2]]488;489 %x2 = load <16 x i32>, ptr %x2p490 %1 = call <16 x i32> @llvm.x86.avx512.vpdpwssds.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2)491 %2 = bitcast i16 %x3 to <16 x i1>492 %3 = select <16 x i1> %2, <16 x i32> %1, <16 x i32> %x0493 %4 = call <16 x i32> @llvm.x86.avx512.vpdpwssds.512(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x4)494 %5 = bitcast i16 %x3 to <16 x i1>495 %6 = select <16 x i1> %5, <16 x i32> %4, <16 x i32> zeroinitializer496 %res1 = insertvalue { <16 x i32>, <16 x i32> } poison, <16 x i32> %3, 0497 %res2 = insertvalue { <16 x i32>, <16 x i32> } %res1, <16 x i32> %6, 1498 ret { <16 x i32>, <16 x i32> } %res2499}500;.501; CHECK: [[PROF1]] = !{!"branch_weights", i32 1, i32 1048575}502;.503