164 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt < %s -passes=slp-vectorizer -S | FileCheck %s --check-prefix=DEFAULT3; RUN: opt < %s -slp-schedule-budget=0 -slp-min-tree-size=0 -slp-threshold=-30 -passes=slp-vectorizer -S | FileCheck %s --check-prefix=GATHER4; RUN: opt < %s -slp-schedule-budget=0 -slp-threshold=-30 -passes=slp-vectorizer -S | FileCheck %s --check-prefix=MAX-COST5 6target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"7target triple = "aarch64--linux-gnu"8 9@a = common global [80 x i8] zeroinitializer, align 1610 11define void @PR28330(i32 %n) {12; DEFAULT-LABEL: @PR28330(13; DEFAULT-NEXT: entry:14; DEFAULT-NEXT: [[TMP0:%.*]] = load <8 x i8>, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 1), align 115; DEFAULT-NEXT: [[TMP1:%.*]] = icmp eq <8 x i8> [[TMP0]], zeroinitializer16; DEFAULT-NEXT: br label [[FOR_BODY:%.*]]17; DEFAULT: for.body:18; DEFAULT-NEXT: [[P17:%.*]] = phi i32 [ [[OP_RDX:%.*]], [[FOR_BODY]] ], [ 0, [[ENTRY:%.*]] ]19; DEFAULT-NEXT: [[TMP2:%.*]] = select <8 x i1> [[TMP1]], <8 x i32> splat (i32 -720), <8 x i32> splat (i32 -80)20; DEFAULT-NEXT: [[TMP3:%.*]] = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[TMP2]])21; DEFAULT-NEXT: [[OP_RDX]] = add i32 [[TMP3]], [[P17]]22; DEFAULT-NEXT: br label [[FOR_BODY]]23;24; GATHER-LABEL: @PR28330(25; GATHER-NEXT: entry:26; GATHER-NEXT: [[TMP0:%.*]] = load <8 x i8>, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 1), align 127; GATHER-NEXT: [[TMP1:%.*]] = icmp eq <8 x i8> [[TMP0]], zeroinitializer28; GATHER-NEXT: br label [[FOR_BODY:%.*]]29; GATHER: for.body:30; GATHER-NEXT: [[P17:%.*]] = phi i32 [ [[OP_RDX:%.*]], [[FOR_BODY]] ], [ 0, [[ENTRY:%.*]] ]31; GATHER-NEXT: [[TMP2:%.*]] = select <8 x i1> [[TMP1]], <8 x i32> splat (i32 -720), <8 x i32> splat (i32 -80)32; GATHER-NEXT: [[TMP3:%.*]] = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[TMP2]])33; GATHER-NEXT: [[OP_RDX]] = add i32 [[TMP3]], [[P17]]34; GATHER-NEXT: br label [[FOR_BODY]]35;36; MAX-COST-LABEL: @PR28330(37; MAX-COST-NEXT: entry:38; MAX-COST-NEXT: [[TMP0:%.*]] = load <8 x i8>, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 1), align 139; MAX-COST-NEXT: [[TMP1:%.*]] = icmp eq <8 x i8> [[TMP0]], zeroinitializer40; MAX-COST-NEXT: br label [[FOR_BODY:%.*]]41; MAX-COST: for.body:42; MAX-COST-NEXT: [[P17:%.*]] = phi i32 [ [[OP_RDX:%.*]], [[FOR_BODY]] ], [ 0, [[ENTRY:%.*]] ]43; MAX-COST-NEXT: [[TMP2:%.*]] = select <8 x i1> [[TMP1]], <8 x i32> splat (i32 -720), <8 x i32> splat (i32 -80)44; MAX-COST-NEXT: [[TMP3:%.*]] = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[TMP2]])45; MAX-COST-NEXT: [[OP_RDX]] = add i32 [[TMP3]], [[P17]]46; MAX-COST-NEXT: br label [[FOR_BODY]]47;48entry:49 %p0 = load i8, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 1), align 150 %p1 = icmp eq i8 %p0, 051 %p2 = load i8, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 2), align 252 %p3 = icmp eq i8 %p2, 053 %p4 = load i8, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 3), align 154 %p5 = icmp eq i8 %p4, 055 %p6 = load i8, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 4), align 456 %p7 = icmp eq i8 %p6, 057 %p8 = load i8, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 5), align 158 %p9 = icmp eq i8 %p8, 059 %p10 = load i8, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 6), align 260 %p11 = icmp eq i8 %p10, 061 %p12 = load i8, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 7), align 162 %p13 = icmp eq i8 %p12, 063 %p14 = load i8, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 8), align 864 %p15 = icmp eq i8 %p14, 065 br label %for.body66 67for.body:68 %p17 = phi i32 [ %p34, %for.body ], [ 0, %entry ]69 %p19 = select i1 %p1, i32 -720, i32 -8070 %p20 = add i32 %p17, %p1971 %p21 = select i1 %p3, i32 -720, i32 -8072 %p22 = add i32 %p20, %p2173 %p23 = select i1 %p5, i32 -720, i32 -8074 %p24 = add i32 %p22, %p2375 %p25 = select i1 %p7, i32 -720, i32 -8076 %p26 = add i32 %p24, %p2577 %p27 = select i1 %p9, i32 -720, i32 -8078 %p28 = add i32 %p26, %p2779 %p29 = select i1 %p11, i32 -720, i32 -8080 %p30 = add i32 %p28, %p2981 %p31 = select i1 %p13, i32 -720, i32 -8082 %p32 = add i32 %p30, %p3183 %p33 = select i1 %p15, i32 -720, i32 -8084 %p34 = add i32 %p32, %p3385 br label %for.body86}87 88define void @PR32038(i32 %n) {89; DEFAULT-LABEL: @PR32038(90; DEFAULT-NEXT: entry:91; DEFAULT-NEXT: [[TMP0:%.*]] = load <8 x i8>, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 1), align 192; DEFAULT-NEXT: [[TMP1:%.*]] = icmp eq <8 x i8> [[TMP0]], zeroinitializer93; DEFAULT-NEXT: br label [[FOR_BODY:%.*]]94; DEFAULT: for.body:95; DEFAULT-NEXT: [[P17:%.*]] = phi i32 [ [[OP_RDX:%.*]], [[FOR_BODY]] ], [ 0, [[ENTRY:%.*]] ]96; DEFAULT-NEXT: [[TMP2:%.*]] = select <8 x i1> [[TMP1]], <8 x i32> splat (i32 -720), <8 x i32> splat (i32 -80)97; DEFAULT-NEXT: [[TMP3:%.*]] = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[TMP2]])98; DEFAULT-NEXT: [[OP_RDX]] = add i32 [[TMP3]], -599; DEFAULT-NEXT: br label [[FOR_BODY]]100;101; GATHER-LABEL: @PR32038(102; GATHER-NEXT: entry:103; GATHER-NEXT: [[TMP0:%.*]] = load <8 x i8>, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 1), align 1104; GATHER-NEXT: [[TMP1:%.*]] = icmp eq <8 x i8> [[TMP0]], zeroinitializer105; GATHER-NEXT: br label [[FOR_BODY:%.*]]106; GATHER: for.body:107; GATHER-NEXT: [[P17:%.*]] = phi i32 [ [[OP_RDX:%.*]], [[FOR_BODY]] ], [ 0, [[ENTRY:%.*]] ]108; GATHER-NEXT: [[TMP2:%.*]] = select <8 x i1> [[TMP1]], <8 x i32> splat (i32 -720), <8 x i32> splat (i32 -80)109; GATHER-NEXT: [[TMP3:%.*]] = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[TMP2]])110; GATHER-NEXT: [[OP_RDX]] = add i32 [[TMP3]], -5111; GATHER-NEXT: br label [[FOR_BODY]]112;113; MAX-COST-LABEL: @PR32038(114; MAX-COST-NEXT: entry:115; MAX-COST-NEXT: [[TMP0:%.*]] = load <8 x i8>, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 1), align 1116; MAX-COST-NEXT: [[TMP1:%.*]] = icmp eq <8 x i8> [[TMP0]], zeroinitializer117; MAX-COST-NEXT: br label [[FOR_BODY:%.*]]118; MAX-COST: for.body:119; MAX-COST-NEXT: [[P17:%.*]] = phi i32 [ [[OP_RDX:%.*]], [[FOR_BODY]] ], [ 0, [[ENTRY:%.*]] ]120; MAX-COST-NEXT: [[TMP2:%.*]] = select <8 x i1> [[TMP1]], <8 x i32> splat (i32 -720), <8 x i32> splat (i32 -80)121; MAX-COST-NEXT: [[TMP3:%.*]] = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[TMP2]])122; MAX-COST-NEXT: [[OP_RDX]] = add i32 [[TMP3]], -5123; MAX-COST-NEXT: br label [[FOR_BODY]]124;125entry:126 %p0 = load i8, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 1), align 1127 %p1 = icmp eq i8 %p0, 0128 %p2 = load i8, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 2), align 2129 %p3 = icmp eq i8 %p2, 0130 %p4 = load i8, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 3), align 1131 %p5 = icmp eq i8 %p4, 0132 %p6 = load i8, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 4), align 4133 %p7 = icmp eq i8 %p6, 0134 %p8 = load i8, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 5), align 1135 %p9 = icmp eq i8 %p8, 0136 %p10 = load i8, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 6), align 2137 %p11 = icmp eq i8 %p10, 0138 %p12 = load i8, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 7), align 1139 %p13 = icmp eq i8 %p12, 0140 %p14 = load i8, ptr getelementptr inbounds ([80 x i8], ptr @a, i64 0, i64 8), align 8141 %p15 = icmp eq i8 %p14, 0142 br label %for.body143 144for.body:145 %p17 = phi i32 [ %p34, %for.body ], [ 0, %entry ]146 %p19 = select i1 %p1, i32 -720, i32 -80147 %p20 = add i32 -5, %p19148 %p21 = select i1 %p3, i32 -720, i32 -80149 %p22 = add i32 %p20, %p21150 %p23 = select i1 %p5, i32 -720, i32 -80151 %p24 = add i32 %p22, %p23152 %p25 = select i1 %p7, i32 -720, i32 -80153 %p26 = add i32 %p24, %p25154 %p27 = select i1 %p9, i32 -720, i32 -80155 %p28 = add i32 %p26, %p27156 %p29 = select i1 %p11, i32 -720, i32 -80157 %p30 = add i32 %p28, %p29158 %p31 = select i1 %p13, i32 -720, i32 -80159 %p32 = add i32 %p30, %p31160 %p33 = select i1 %p15, i32 -720, i32 -80161 %p34 = add i32 %p32, %p33162 br label %for.body163}164