151 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -passes=loop-vectorize -mtriple=x86_64-apple-macosx -S -mcpu=corei7-avx -enable-interleaved-mem-accesses=false < %s | FileCheck %s3target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"4 5@kernel = global [512 x float] zeroinitializer, align 166@kernel2 = global [512 x float] zeroinitializer, align 167@kernel3 = global [512 x float] zeroinitializer, align 168@kernel4 = global [512 x float] zeroinitializer, align 169@src_data = global [1536 x float] zeroinitializer, align 1610@r_ = global i8 0, align 111@g_ = global i8 0, align 112@b_ = global i8 0, align 113 14; We don't want to vectorize most loops containing gathers because they are15; expensive. This function represents a point where vectorization starts to16; become beneficial.17; Make sure we are conservative and don't vectorize it.18 19define void @_Z4testmm(i64 %size, i64 %offset) {20; CHECK-LABEL: @_Z4testmm(21; CHECK-NEXT: entry:22; CHECK-NEXT: [[CMP53:%.*]] = icmp eq i64 [[SIZE:%.*]], 023; CHECK-NEXT: br i1 [[CMP53]], label [[FOR_END:%.*]], label [[FOR_BODY_LR_PH:%.*]]24; CHECK: for.body.lr.ph:25; CHECK-NEXT: br label [[FOR_BODY:%.*]]26; CHECK: for.body:27; CHECK-NEXT: [[R_057:%.*]] = phi float [ 0.000000e+00, [[FOR_BODY_LR_PH]] ], [ [[ADD10:%.*]], [[FOR_BODY]] ]28; CHECK-NEXT: [[G_056:%.*]] = phi float [ 0.000000e+00, [[FOR_BODY_LR_PH]] ], [ [[ADD20:%.*]], [[FOR_BODY]] ]29; CHECK-NEXT: [[V_055:%.*]] = phi i64 [ 0, [[FOR_BODY_LR_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ]30; CHECK-NEXT: [[B_054:%.*]] = phi float [ 0.000000e+00, [[FOR_BODY_LR_PH]] ], [ [[ADD30:%.*]], [[FOR_BODY]] ]31; CHECK-NEXT: [[ADD:%.*]] = add i64 [[V_055]], [[OFFSET:%.*]]32; CHECK-NEXT: [[MUL:%.*]] = mul i64 [[ADD]], 333; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [1536 x float], ptr @src_data, i64 0, i64 [[MUL]]34; CHECK-NEXT: [[TMP0:%.*]] = load float, ptr [[ARRAYIDX]], align 435; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds [512 x float], ptr @kernel, i64 0, i64 [[V_055]]36; CHECK-NEXT: [[TMP1:%.*]] = load float, ptr [[ARRAYIDX2]], align 437; CHECK-NEXT: [[MUL3:%.*]] = fmul fast float [[TMP0]], [[TMP1]]38; CHECK-NEXT: [[ARRAYIDX4:%.*]] = getelementptr inbounds [512 x float], ptr @kernel2, i64 0, i64 [[V_055]]39; CHECK-NEXT: [[TMP2:%.*]] = load float, ptr [[ARRAYIDX4]], align 440; CHECK-NEXT: [[MUL5:%.*]] = fmul fast float [[MUL3]], [[TMP2]]41; CHECK-NEXT: [[ARRAYIDX6:%.*]] = getelementptr inbounds [512 x float], ptr @kernel3, i64 0, i64 [[V_055]]42; CHECK-NEXT: [[TMP3:%.*]] = load float, ptr [[ARRAYIDX6]], align 443; CHECK-NEXT: [[MUL7:%.*]] = fmul fast float [[MUL5]], [[TMP3]]44; CHECK-NEXT: [[ARRAYIDX8:%.*]] = getelementptr inbounds [512 x float], ptr @kernel4, i64 0, i64 [[V_055]]45; CHECK-NEXT: [[TMP4:%.*]] = load float, ptr [[ARRAYIDX8]], align 446; CHECK-NEXT: [[MUL9:%.*]] = fmul fast float [[MUL7]], [[TMP4]]47; CHECK-NEXT: [[ADD10]] = fadd fast float [[R_057]], [[MUL9]]48; CHECK-NEXT: [[ARRAYIDX_SUM:%.*]] = add i64 [[MUL]], 149; CHECK-NEXT: [[ARRAYIDX11:%.*]] = getelementptr inbounds [1536 x float], ptr @src_data, i64 0, i64 [[ARRAYIDX_SUM]]50; CHECK-NEXT: [[TMP5:%.*]] = load float, ptr [[ARRAYIDX11]], align 451; CHECK-NEXT: [[MUL13:%.*]] = fmul fast float [[TMP1]], [[TMP5]]52; CHECK-NEXT: [[MUL15:%.*]] = fmul fast float [[TMP2]], [[MUL13]]53; CHECK-NEXT: [[MUL17:%.*]] = fmul fast float [[TMP3]], [[MUL15]]54; CHECK-NEXT: [[MUL19:%.*]] = fmul fast float [[TMP4]], [[MUL17]]55; CHECK-NEXT: [[ADD20]] = fadd fast float [[G_056]], [[MUL19]]56; CHECK-NEXT: [[ARRAYIDX_SUM52:%.*]] = add i64 [[MUL]], 257; CHECK-NEXT: [[ARRAYIDX21:%.*]] = getelementptr inbounds [1536 x float], ptr @src_data, i64 0, i64 [[ARRAYIDX_SUM52]]58; CHECK-NEXT: [[TMP6:%.*]] = load float, ptr [[ARRAYIDX21]], align 459; CHECK-NEXT: [[MUL23:%.*]] = fmul fast float [[TMP1]], [[TMP6]]60; CHECK-NEXT: [[MUL25:%.*]] = fmul fast float [[TMP2]], [[MUL23]]61; CHECK-NEXT: [[MUL27:%.*]] = fmul fast float [[TMP3]], [[MUL25]]62; CHECK-NEXT: [[MUL29:%.*]] = fmul fast float [[TMP4]], [[MUL27]]63; CHECK-NEXT: [[ADD30]] = fadd fast float [[B_054]], [[MUL29]]64; CHECK-NEXT: [[INC]] = add i64 [[V_055]], 165; CHECK-NEXT: [[EXITCOND:%.*]] = icmp ne i64 [[INC]], [[SIZE]]66; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_BODY]], label [[FOR_COND_FOR_END_CRIT_EDGE:%.*]]67; CHECK: for.cond.for.end_crit_edge:68; CHECK-NEXT: [[ADD30_LCSSA:%.*]] = phi float [ [[ADD30]], [[FOR_BODY]] ]69; CHECK-NEXT: [[ADD20_LCSSA:%.*]] = phi float [ [[ADD20]], [[FOR_BODY]] ]70; CHECK-NEXT: [[ADD10_LCSSA:%.*]] = phi float [ [[ADD10]], [[FOR_BODY]] ]71; CHECK-NEXT: [[PHITMP:%.*]] = fptoui float [[ADD10_LCSSA]] to i872; CHECK-NEXT: [[PHITMP60:%.*]] = fptoui float [[ADD20_LCSSA]] to i873; CHECK-NEXT: [[PHITMP61:%.*]] = fptoui float [[ADD30_LCSSA]] to i874; CHECK-NEXT: br label [[FOR_END]]75; CHECK: for.end:76; CHECK-NEXT: [[R_0_LCSSA:%.*]] = phi i8 [ [[PHITMP]], [[FOR_COND_FOR_END_CRIT_EDGE]] ], [ 0, [[ENTRY:%.*]] ]77; CHECK-NEXT: [[G_0_LCSSA:%.*]] = phi i8 [ [[PHITMP60]], [[FOR_COND_FOR_END_CRIT_EDGE]] ], [ 0, [[ENTRY]] ]78; CHECK-NEXT: [[B_0_LCSSA:%.*]] = phi i8 [ [[PHITMP61]], [[FOR_COND_FOR_END_CRIT_EDGE]] ], [ 0, [[ENTRY]] ]79; CHECK-NEXT: store i8 [[R_0_LCSSA]], ptr @r_, align 180; CHECK-NEXT: store i8 [[G_0_LCSSA]], ptr @g_, align 181; CHECK-NEXT: store i8 [[B_0_LCSSA]], ptr @b_, align 182; CHECK-NEXT: ret void83;84entry:85 %cmp53 = icmp eq i64 %size, 086 br i1 %cmp53, label %for.end, label %for.body.lr.ph87 88for.body.lr.ph:89 br label %for.body90 91for.body:92 %r.057 = phi float [ 0.000000e+00, %for.body.lr.ph ], [ %add10, %for.body ]93 %g.056 = phi float [ 0.000000e+00, %for.body.lr.ph ], [ %add20, %for.body ]94 %v.055 = phi i64 [ 0, %for.body.lr.ph ], [ %inc, %for.body ]95 %b.054 = phi float [ 0.000000e+00, %for.body.lr.ph ], [ %add30, %for.body ]96 %add = add i64 %v.055, %offset97 %mul = mul i64 %add, 398 %arrayidx = getelementptr inbounds [1536 x float], ptr @src_data, i64 0, i64 %mul99 %0 = load float, ptr %arrayidx, align 4100 %arrayidx2 = getelementptr inbounds [512 x float], ptr @kernel, i64 0, i64 %v.055101 %1 = load float, ptr %arrayidx2, align 4102 %mul3 = fmul fast float %0, %1103 %arrayidx4 = getelementptr inbounds [512 x float], ptr @kernel2, i64 0, i64 %v.055104 %2 = load float, ptr %arrayidx4, align 4105 %mul5 = fmul fast float %mul3, %2106 %arrayidx6 = getelementptr inbounds [512 x float], ptr @kernel3, i64 0, i64 %v.055107 %3 = load float, ptr %arrayidx6, align 4108 %mul7 = fmul fast float %mul5, %3109 %arrayidx8 = getelementptr inbounds [512 x float], ptr @kernel4, i64 0, i64 %v.055110 %4 = load float, ptr %arrayidx8, align 4111 %mul9 = fmul fast float %mul7, %4112 %add10 = fadd fast float %r.057, %mul9113 %arrayidx.sum = add i64 %mul, 1114 %arrayidx11 = getelementptr inbounds [1536 x float], ptr @src_data, i64 0, i64 %arrayidx.sum115 %5 = load float, ptr %arrayidx11, align 4116 %mul13 = fmul fast float %1, %5117 %mul15 = fmul fast float %2, %mul13118 %mul17 = fmul fast float %3, %mul15119 %mul19 = fmul fast float %4, %mul17120 %add20 = fadd fast float %g.056, %mul19121 %arrayidx.sum52 = add i64 %mul, 2122 %arrayidx21 = getelementptr inbounds [1536 x float], ptr @src_data, i64 0, i64 %arrayidx.sum52123 %6 = load float, ptr %arrayidx21, align 4124 %mul23 = fmul fast float %1, %6125 %mul25 = fmul fast float %2, %mul23126 %mul27 = fmul fast float %3, %mul25127 %mul29 = fmul fast float %4, %mul27128 %add30 = fadd fast float %b.054, %mul29129 %inc = add i64 %v.055, 1130 %exitcond = icmp ne i64 %inc, %size131 br i1 %exitcond, label %for.body, label %for.cond.for.end_crit_edge132 133for.cond.for.end_crit_edge:134 %add30.lcssa = phi float [ %add30, %for.body ]135 %add20.lcssa = phi float [ %add20, %for.body ]136 %add10.lcssa = phi float [ %add10, %for.body ]137 %phitmp = fptoui float %add10.lcssa to i8138 %phitmp60 = fptoui float %add20.lcssa to i8139 %phitmp61 = fptoui float %add30.lcssa to i8140 br label %for.end141 142for.end:143 %r.0.lcssa = phi i8 [ %phitmp, %for.cond.for.end_crit_edge ], [ 0, %entry ]144 %g.0.lcssa = phi i8 [ %phitmp60, %for.cond.for.end_crit_edge ], [ 0, %entry ]145 %b.0.lcssa = phi i8 [ %phitmp61, %for.cond.for.end_crit_edge ], [ 0, %entry ]146 store i8 %r.0.lcssa, ptr @r_, align 1147 store i8 %g.0.lcssa, ptr @g_, align 1148 store i8 %b.0.lcssa, ptr @b_, align 1149 ret void150}151