85 lines · plain
1; RUN: opt -passes=loop-vectorize -mtriple=arm64-apple-ios -S -mcpu=cyclone -enable-interleaved-mem-accesses=false -neon-nonconst-stride-overhead=5 < %s | FileCheck %s2target 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-n32:64-S128"3 4@kernel = global [512 x float] zeroinitializer, align 165@kernel2 = global [512 x float] zeroinitializer, align 166@kernel3 = global [512 x float] zeroinitializer, align 167@kernel4 = global [512 x float] zeroinitializer, align 168@src_data = global [1536 x float] zeroinitializer, align 169@r_ = global i8 0, align 110@g_ = global i8 0, align 111@b_ = global i8 0, align 112 13; vectorize loop when we lower the cost of gather load14; Make sure we vectorize it, VF = 415; CHECK: <4 x float>16 17define void @_Z4testmm(i64 %size, i64 %offset) {18entry:19 %cmp53 = icmp eq i64 %size, 020 br i1 %cmp53, label %for.end, label %for.body.lr.ph21 22for.body.lr.ph:23 br label %for.body24 25for.body:26 %r.057 = phi float [ 0.000000e+00, %for.body.lr.ph ], [ %add10, %for.body ]27 %g.056 = phi float [ 0.000000e+00, %for.body.lr.ph ], [ %add20, %for.body ]28 %v.055 = phi i64 [ 0, %for.body.lr.ph ], [ %inc, %for.body ]29 %b.054 = phi float [ 0.000000e+00, %for.body.lr.ph ], [ %add30, %for.body ]30 %add = add i64 %v.055, %offset31 %mul = mul i64 %add, 332 %arrayidx = getelementptr inbounds [1536 x float], ptr @src_data, i64 0, i64 %mul33 %0 = load float, ptr %arrayidx, align 434 %arrayidx2 = getelementptr inbounds [512 x float], ptr @kernel, i64 0, i64 %v.05535 %1 = load float, ptr %arrayidx2, align 436 %mul3 = fmul fast float %0, %137 %arrayidx4 = getelementptr inbounds [512 x float], ptr @kernel2, i64 0, i64 %v.05538 %2 = load float, ptr %arrayidx4, align 439 %mul5 = fmul fast float %mul3, %240 %arrayidx6 = getelementptr inbounds [512 x float], ptr @kernel3, i64 0, i64 %v.05541 %3 = load float, ptr %arrayidx6, align 442 %mul7 = fmul fast float %mul5, %343 %arrayidx8 = getelementptr inbounds [512 x float], ptr @kernel4, i64 0, i64 %v.05544 %4 = load float, ptr %arrayidx8, align 445 %mul9 = fmul fast float %mul7, %446 %add10 = fadd fast float %r.057, %mul947 %arrayidx.sum = add i64 %mul, 148 %arrayidx11 = getelementptr inbounds [1536 x float], ptr @src_data, i64 0, i64 %arrayidx.sum49 %5 = load float, ptr %arrayidx11, align 450 %mul13 = fmul fast float %1, %551 %mul15 = fmul fast float %2, %mul1352 %mul17 = fmul fast float %3, %mul1553 %mul19 = fmul fast float %4, %mul1754 %add20 = fadd fast float %g.056, %mul1955 %arrayidx.sum52 = add i64 %mul, 256 %arrayidx21 = getelementptr inbounds [1536 x float], ptr @src_data, i64 0, i64 %arrayidx.sum5257 %6 = load float, ptr %arrayidx21, align 458 %mul23 = fmul fast float %1, %659 %mul25 = fmul fast float %2, %mul2360 %mul27 = fmul fast float %3, %mul2561 %mul29 = fmul fast float %4, %mul2762 %add30 = fadd fast float %b.054, %mul2963 %inc = add i64 %v.055, 164 %exitcond = icmp ne i64 %inc, %size65 br i1 %exitcond, label %for.body, label %for.cond.for.end_crit_edge66 67for.cond.for.end_crit_edge:68 %add30.lcssa = phi float [ %add30, %for.body ]69 %add20.lcssa = phi float [ %add20, %for.body ]70 %add10.lcssa = phi float [ %add10, %for.body ]71 %phitmp = fptoui float %add10.lcssa to i872 %phitmp60 = fptoui float %add20.lcssa to i873 %phitmp61 = fptoui float %add30.lcssa to i874 br label %for.end75 76for.end:77 %r.0.lcssa = phi i8 [ %phitmp, %for.cond.for.end_crit_edge ], [ 0, %entry ]78 %g.0.lcssa = phi i8 [ %phitmp60, %for.cond.for.end_crit_edge ], [ 0, %entry ]79 %b.0.lcssa = phi i8 [ %phitmp61, %for.cond.for.end_crit_edge ], [ 0, %entry ]80 store i8 %r.0.lcssa, ptr @r_, align 181 store i8 %g.0.lcssa, ptr @g_, align 182 store i8 %b.0.lcssa, ptr @b_, align 183 ret void84}85