86 lines · plain
1; RUN: opt -passes=loop-vectorize -mtriple=arm64-apple-ios -S -mcpu=cyclone -enable-interleaved-mem-accesses=false < %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; We don't want to vectorize most loops containing gathers because they are14; expensive.15; Make sure we don't vectorize it.16; CHECK-NOT: x float>17 18define void @_Z4testmm(i64 %size, i64 %offset) {19entry:20 %cmp53 = icmp eq i64 %size, 021 br i1 %cmp53, label %for.end, label %for.body.lr.ph22 23for.body.lr.ph:24 br label %for.body25 26for.body:27 %r.057 = phi float [ 0.000000e+00, %for.body.lr.ph ], [ %add10, %for.body ]28 %g.056 = phi float [ 0.000000e+00, %for.body.lr.ph ], [ %add20, %for.body ]29 %v.055 = phi i64 [ 0, %for.body.lr.ph ], [ %inc, %for.body ]30 %b.054 = phi float [ 0.000000e+00, %for.body.lr.ph ], [ %add30, %for.body ]31 %add = add i64 %v.055, %offset32 %mul = mul i64 %add, 333 %arrayidx = getelementptr inbounds [1536 x float], ptr @src_data, i64 0, i64 %mul34 %0 = load float, ptr %arrayidx, align 435 %arrayidx2 = getelementptr inbounds [512 x float], ptr @kernel, i64 0, i64 %v.05536 %1 = load float, ptr %arrayidx2, align 437 %mul3 = fmul fast float %0, %138 %arrayidx4 = getelementptr inbounds [512 x float], ptr @kernel2, i64 0, i64 %v.05539 %2 = load float, ptr %arrayidx4, align 440 %mul5 = fmul fast float %mul3, %241 %arrayidx6 = getelementptr inbounds [512 x float], ptr @kernel3, i64 0, i64 %v.05542 %3 = load float, ptr %arrayidx6, align 443 %mul7 = fmul fast float %mul5, %344 %arrayidx8 = getelementptr inbounds [512 x float], ptr @kernel4, i64 0, i64 %v.05545 %4 = load float, ptr %arrayidx8, align 446 %mul9 = fmul fast float %mul7, %447 %add10 = fadd fast float %r.057, %mul948 %arrayidx.sum = add i64 %mul, 149 %arrayidx11 = getelementptr inbounds [1536 x float], ptr @src_data, i64 0, i64 %arrayidx.sum50 %5 = load float, ptr %arrayidx11, align 451 %mul13 = fmul fast float %1, %552 %mul15 = fmul fast float %2, %mul1353 %mul17 = fmul fast float %3, %mul1554 %mul19 = fmul fast float %4, %mul1755 %add20 = fadd fast float %g.056, %mul1956 %arrayidx.sum52 = add i64 %mul, 257 %arrayidx21 = getelementptr inbounds [1536 x float], ptr @src_data, i64 0, i64 %arrayidx.sum5258 %6 = load float, ptr %arrayidx21, align 459 %mul23 = fmul fast float %1, %660 %mul25 = fmul fast float %2, %mul2361 %mul27 = fmul fast float %3, %mul2562 %mul29 = fmul fast float %4, %mul2763 %add30 = fadd fast float %b.054, %mul2964 %inc = add i64 %v.055, 165 %exitcond = icmp ne i64 %inc, %size66 br i1 %exitcond, label %for.body, label %for.cond.for.end_crit_edge67 68for.cond.for.end_crit_edge:69 %add30.lcssa = phi float [ %add30, %for.body ]70 %add20.lcssa = phi float [ %add20, %for.body ]71 %add10.lcssa = phi float [ %add10, %for.body ]72 %phitmp = fptoui float %add10.lcssa to i873 %phitmp60 = fptoui float %add20.lcssa to i874 %phitmp61 = fptoui float %add30.lcssa to i875 br label %for.end76 77for.end:78 %r.0.lcssa = phi i8 [ %phitmp, %for.cond.for.end_crit_edge ], [ 0, %entry ]79 %g.0.lcssa = phi i8 [ %phitmp60, %for.cond.for.end_crit_edge ], [ 0, %entry ]80 %b.0.lcssa = phi i8 [ %phitmp61, %for.cond.for.end_crit_edge ], [ 0, %entry ]81 store i8 %r.0.lcssa, ptr @r_, align 182 store i8 %g.0.lcssa, ptr @g_, align 183 store i8 %b.0.lcssa, ptr @b_, align 184 ret void85}86