171 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -mtriple=s390x-unknown-linux -mcpu=z16 -passes=slp-vectorizer -S -slp-revec %s | FileCheck %s3 4@g_155 = external dso_local global i64, align 85@g_855 = external dso_local global i8, align 26@g_3_1_0 = external dso_local global i32, align 47@g_7 = external dso_local global [5 x i32], align 48 9; Function Attrs: nofree norecurse noreturn nounwind memory(readwrite, argmem: none)10define void @func_1() {11; CHECK-LABEL: @func_1(12; CHECK-NEXT: [[DOTPRE:%.*]] = load i32, ptr @g_7, align 413; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i32>, ptr getelementptr inbounds nuw (i8, ptr @g_7, i64 4), align 414; CHECK-NEXT: br label [[DOTLOOPEXIT:%.*]]15; CHECK: .loopexit:16; CHECK-NEXT: [[TMP2:%.*]] = phi i32 [ [[OP_RDX15:%.*]], [[DOTLOOPEXIT]] ], [ [[DOTPRE]], [[TMP0:%.*]] ]17; CHECK-NEXT: [[TMP3:%.*]] = phi <4 x i32> [ [[TMP71:%.*]], [[DOTLOOPEXIT]] ], [ [[TMP1]], [[TMP0]] ]18; CHECK-NEXT: [[TMP4:%.*]] = load volatile i32, ptr @g_3_1_0, align 419; CHECK-NEXT: [[TMP5:%.*]] = load volatile i32, ptr @g_3_1_0, align 420; CHECK-NEXT: [[TMP6:%.*]] = load volatile i32, ptr @g_3_1_0, align 421; CHECK-NEXT: [[TMP7:%.*]] = load volatile i32, ptr @g_3_1_0, align 422; CHECK-NEXT: [[TMP8:%.*]] = load volatile i32, ptr @g_3_1_0, align 423; CHECK-NEXT: [[TMP9:%.*]] = load volatile i32, ptr @g_3_1_0, align 424; CHECK-NEXT: [[TMP10:%.*]] = load volatile i32, ptr @g_3_1_0, align 425; CHECK-NEXT: [[TMP11:%.*]] = load volatile i32, ptr @g_3_1_0, align 426; CHECK-NEXT: [[TMP12:%.*]] = load volatile i32, ptr @g_3_1_0, align 427; CHECK-NEXT: [[TMP13:%.*]] = load volatile i32, ptr @g_3_1_0, align 428; CHECK-NEXT: [[TMP14:%.*]] = load volatile i32, ptr @g_3_1_0, align 429; CHECK-NEXT: [[TMP15:%.*]] = load volatile i32, ptr @g_3_1_0, align 430; CHECK-NEXT: [[TMP16:%.*]] = load volatile i32, ptr @g_3_1_0, align 431; CHECK-NEXT: [[TMP17:%.*]] = load volatile i32, ptr @g_3_1_0, align 432; CHECK-NEXT: [[TMP18:%.*]] = load volatile i32, ptr @g_3_1_0, align 433; CHECK-NEXT: [[TMP19:%.*]] = load volatile i32, ptr @g_3_1_0, align 434; CHECK-NEXT: [[TMP20:%.*]] = insertelement <4 x i32> poison, i32 [[TMP4]], i32 035; CHECK-NEXT: [[TMP21:%.*]] = insertelement <4 x i32> [[TMP20]], i32 [[TMP9]], i32 136; CHECK-NEXT: [[TMP22:%.*]] = insertelement <4 x i32> [[TMP21]], i32 [[TMP14]], i32 237; CHECK-NEXT: [[TMP23:%.*]] = insertelement <4 x i32> [[TMP22]], i32 [[TMP19]], i32 338; CHECK-NEXT: [[TMP24:%.*]] = icmp eq <4 x i32> [[TMP23]], zeroinitializer39; CHECK-NEXT: [[TMP25:%.*]] = load volatile i32, ptr @g_3_1_0, align 440; CHECK-NEXT: [[TMP26:%.*]] = load volatile i32, ptr @g_3_1_0, align 441; CHECK-NEXT: [[TMP27:%.*]] = load volatile i32, ptr @g_3_1_0, align 442; CHECK-NEXT: [[TMP28:%.*]] = load volatile i32, ptr @g_3_1_0, align 443; CHECK-NEXT: [[TMP29:%.*]] = load volatile i32, ptr @g_3_1_0, align 444; CHECK-NEXT: [[DOTNOT2_410:%.*]] = icmp eq i32 [[TMP29]], 045; CHECK-NEXT: [[TMP30:%.*]] = zext i1 [[DOTNOT2_410]] to i3246; CHECK-NEXT: [[TMP31:%.*]] = zext <4 x i1> [[TMP24]] to <4 x i32>47; CHECK-NEXT: [[TMP32:%.*]] = call i32 @llvm.vector.reduce.xor.v4i32(<4 x i32> [[TMP31]])48; CHECK-NEXT: [[OP_RDX:%.*]] = xor i32 [[TMP32]], [[TMP30]]49; CHECK-NEXT: [[OP_RDX15]] = xor i32 [[OP_RDX]], [[TMP2]]50; CHECK-NEXT: [[TMP33:%.*]] = load volatile i32, ptr @g_3_1_0, align 451; CHECK-NEXT: [[TMP34:%.*]] = load volatile i32, ptr @g_3_1_0, align 452; CHECK-NEXT: [[TMP35:%.*]] = load volatile i32, ptr @g_3_1_0, align 453; CHECK-NEXT: [[TMP36:%.*]] = load volatile i32, ptr @g_3_1_0, align 454; CHECK-NEXT: [[TMP37:%.*]] = insertelement <4 x i32> poison, i32 [[TMP5]], i32 055; CHECK-NEXT: [[TMP38:%.*]] = insertelement <4 x i32> [[TMP37]], i32 [[TMP6]], i32 156; CHECK-NEXT: [[TMP39:%.*]] = insertelement <4 x i32> [[TMP38]], i32 [[TMP7]], i32 257; CHECK-NEXT: [[TMP40:%.*]] = insertelement <4 x i32> [[TMP39]], i32 [[TMP8]], i32 358; CHECK-NEXT: [[TMP41:%.*]] = icmp eq <4 x i32> [[TMP40]], zeroinitializer59; CHECK-NEXT: [[TMP42:%.*]] = zext <4 x i1> [[TMP41]] to <4 x i32>60; CHECK-NEXT: [[TMP43:%.*]] = xor <4 x i32> [[TMP3]], [[TMP42]]61; CHECK-NEXT: [[TMP44:%.*]] = insertelement <4 x i32> poison, i32 [[TMP10]], i32 062; CHECK-NEXT: [[TMP45:%.*]] = insertelement <4 x i32> [[TMP44]], i32 [[TMP11]], i32 163; CHECK-NEXT: [[TMP46:%.*]] = insertelement <4 x i32> [[TMP45]], i32 [[TMP12]], i32 264; CHECK-NEXT: [[TMP47:%.*]] = insertelement <4 x i32> [[TMP46]], i32 [[TMP13]], i32 365; CHECK-NEXT: [[TMP48:%.*]] = icmp eq <4 x i32> [[TMP47]], zeroinitializer66; CHECK-NEXT: [[TMP49:%.*]] = zext <4 x i1> [[TMP48]] to <4 x i32>67; CHECK-NEXT: [[TMP50:%.*]] = xor <4 x i32> [[TMP43]], [[TMP49]]68; CHECK-NEXT: [[TMP51:%.*]] = insertelement <4 x i32> poison, i32 [[TMP15]], i32 069; CHECK-NEXT: [[TMP52:%.*]] = insertelement <4 x i32> [[TMP51]], i32 [[TMP16]], i32 170; CHECK-NEXT: [[TMP53:%.*]] = insertelement <4 x i32> [[TMP52]], i32 [[TMP17]], i32 271; CHECK-NEXT: [[TMP54:%.*]] = insertelement <4 x i32> [[TMP53]], i32 [[TMP18]], i32 372; CHECK-NEXT: [[TMP55:%.*]] = icmp eq <4 x i32> [[TMP54]], zeroinitializer73; CHECK-NEXT: [[TMP56:%.*]] = zext <4 x i1> [[TMP55]] to <4 x i32>74; CHECK-NEXT: [[TMP57:%.*]] = xor <4 x i32> [[TMP50]], [[TMP56]]75; CHECK-NEXT: [[TMP58:%.*]] = insertelement <4 x i32> poison, i32 [[TMP25]], i32 076; CHECK-NEXT: [[TMP59:%.*]] = insertelement <4 x i32> [[TMP58]], i32 [[TMP26]], i32 177; CHECK-NEXT: [[TMP60:%.*]] = insertelement <4 x i32> [[TMP59]], i32 [[TMP27]], i32 278; CHECK-NEXT: [[TMP61:%.*]] = insertelement <4 x i32> [[TMP60]], i32 [[TMP28]], i32 379; CHECK-NEXT: [[TMP62:%.*]] = icmp eq <4 x i32> [[TMP61]], zeroinitializer80; CHECK-NEXT: [[TMP63:%.*]] = zext <4 x i1> [[TMP62]] to <4 x i32>81; CHECK-NEXT: [[TMP64:%.*]] = xor <4 x i32> [[TMP57]], [[TMP63]]82; CHECK-NEXT: [[TMP65:%.*]] = insertelement <4 x i32> poison, i32 [[TMP33]], i32 083; CHECK-NEXT: [[TMP66:%.*]] = insertelement <4 x i32> [[TMP65]], i32 [[TMP34]], i32 184; CHECK-NEXT: [[TMP67:%.*]] = insertelement <4 x i32> [[TMP66]], i32 [[TMP35]], i32 285; CHECK-NEXT: [[TMP68:%.*]] = insertelement <4 x i32> [[TMP67]], i32 [[TMP36]], i32 386; CHECK-NEXT: [[TMP69:%.*]] = icmp eq <4 x i32> [[TMP68]], zeroinitializer87; CHECK-NEXT: [[TMP70:%.*]] = zext <4 x i1> [[TMP69]] to <4 x i32>88; CHECK-NEXT: [[TMP71]] = xor <4 x i32> [[TMP64]], [[TMP70]]89; CHECK-NEXT: br label [[DOTLOOPEXIT]]90;91 %.pre = load i32, ptr @g_7, align 492 %1 = load <4 x i32>, ptr getelementptr inbounds nuw (i8, ptr @g_7, i64 4), align 493 br label %.loopexit94 95.loopexit: ; preds = %.loopexit, %096 %2 = phi i32 [ %op.rdx15, %.loopexit ], [ %.pre, %0 ]97 %3 = phi <4 x i32> [ %71, %.loopexit ], [ %1, %0 ]98 %4 = load volatile i32, ptr @g_3_1_0, align 499 %5 = load volatile i32, ptr @g_3_1_0, align 4100 %6 = load volatile i32, ptr @g_3_1_0, align 4101 %7 = load volatile i32, ptr @g_3_1_0, align 4102 %8 = load volatile i32, ptr @g_3_1_0, align 4103 %9 = load volatile i32, ptr @g_3_1_0, align 4104 %10 = load volatile i32, ptr @g_3_1_0, align 4105 %11 = load volatile i32, ptr @g_3_1_0, align 4106 %12 = load volatile i32, ptr @g_3_1_0, align 4107 %13 = load volatile i32, ptr @g_3_1_0, align 4108 %14 = load volatile i32, ptr @g_3_1_0, align 4109 %15 = load volatile i32, ptr @g_3_1_0, align 4110 %16 = load volatile i32, ptr @g_3_1_0, align 4111 %17 = load volatile i32, ptr @g_3_1_0, align 4112 %18 = load volatile i32, ptr @g_3_1_0, align 4113 %19 = load volatile i32, ptr @g_3_1_0, align 4114 %20 = insertelement <4 x i32> poison, i32 %4, i32 0115 %21 = insertelement <4 x i32> %20, i32 %9, i32 1116 %22 = insertelement <4 x i32> %21, i32 %14, i32 2117 %23 = insertelement <4 x i32> %22, i32 %19, i32 3118 %24 = icmp eq <4 x i32> %23, zeroinitializer119 %25 = load volatile i32, ptr @g_3_1_0, align 4120 %26 = load volatile i32, ptr @g_3_1_0, align 4121 %27 = load volatile i32, ptr @g_3_1_0, align 4122 %28 = load volatile i32, ptr @g_3_1_0, align 4123 %29 = load volatile i32, ptr @g_3_1_0, align 4124 %.not2.410 = icmp eq i32 %29, 0125 %30 = zext i1 %.not2.410 to i32126 %31 = zext <4 x i1> %24 to <4 x i32>127 %32 = call i32 @llvm.vector.reduce.xor.v4i32(<4 x i32> %31)128 %op.rdx = xor i32 %32, %30129 %op.rdx15 = xor i32 %op.rdx, %2130 %33 = load volatile i32, ptr @g_3_1_0, align 4131 %34 = load volatile i32, ptr @g_3_1_0, align 4132 %35 = load volatile i32, ptr @g_3_1_0, align 4133 %36 = load volatile i32, ptr @g_3_1_0, align 4134 %37 = insertelement <4 x i32> poison, i32 %5, i32 0135 %38 = insertelement <4 x i32> %37, i32 %6, i32 1136 %39 = insertelement <4 x i32> %38, i32 %7, i32 2137 %40 = insertelement <4 x i32> %39, i32 %8, i32 3138 %41 = icmp eq <4 x i32> %40, zeroinitializer139 %42 = zext <4 x i1> %41 to <4 x i32>140 %43 = xor <4 x i32> %3, %42141 %44 = insertelement <4 x i32> poison, i32 %10, i32 0142 %45 = insertelement <4 x i32> %44, i32 %11, i32 1143 %46 = insertelement <4 x i32> %45, i32 %12, i32 2144 %47 = insertelement <4 x i32> %46, i32 %13, i32 3145 %48 = icmp eq <4 x i32> %47, zeroinitializer146 %49 = zext <4 x i1> %48 to <4 x i32>147 %50 = xor <4 x i32> %43, %49148 %51 = insertelement <4 x i32> poison, i32 %15, i32 0149 %52 = insertelement <4 x i32> %51, i32 %16, i32 1150 %53 = insertelement <4 x i32> %52, i32 %17, i32 2151 %54 = insertelement <4 x i32> %53, i32 %18, i32 3152 %55 = icmp eq <4 x i32> %54, zeroinitializer153 %56 = zext <4 x i1> %55 to <4 x i32>154 %57 = xor <4 x i32> %50, %56155 %58 = insertelement <4 x i32> poison, i32 %25, i32 0156 %59 = insertelement <4 x i32> %58, i32 %26, i32 1157 %60 = insertelement <4 x i32> %59, i32 %27, i32 2158 %61 = insertelement <4 x i32> %60, i32 %28, i32 3159 %62 = icmp eq <4 x i32> %61, zeroinitializer160 %63 = zext <4 x i1> %62 to <4 x i32>161 %64 = xor <4 x i32> %57, %63162 %65 = insertelement <4 x i32> poison, i32 %33, i32 0163 %66 = insertelement <4 x i32> %65, i32 %34, i32 1164 %67 = insertelement <4 x i32> %66, i32 %35, i32 2165 %68 = insertelement <4 x i32> %67, i32 %36, i32 3166 %69 = icmp eq <4 x i32> %68, zeroinitializer167 %70 = zext <4 x i1> %69 to <4 x i32>168 %71 = xor <4 x i32> %64, %70169 br label %.loopexit170}171