38 lines · plain
1; RUN: opt -passes='default<Oz>' -S < %s | FileCheck %s2 3; Forcing vectorization should allow for more aggressive loop-rotation with4; -Oz, because LV requires rotated loops. Make sure the loop in @foo is5; vectorized with -Oz.6 7target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"8target triple = "arm64-apple-ios5.0.0"9 10define void @foo(ptr noalias nocapture %ptrA, ptr noalias nocapture readonly %ptrB, i64 %size) {11; CHECK-LABEL: @foo(12; CHECK: fmul <4 x float>13;14entry:15 br label %for.cond16 17for.cond: ; preds = %for.body, %entry18 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]19 %exitcond = icmp eq i64 %indvars.iv, %size20 br i1 %exitcond, label %for.cond.cleanup, label %for.body21 22for.body: ; preds = %for.cond23 %arrayidx = getelementptr inbounds float, ptr %ptrB, i64 %indvars.iv24 %0 = load float, ptr %arrayidx, align 425 %arrayidx2 = getelementptr inbounds float, ptr %ptrA, i64 %indvars.iv26 %1 = load float, ptr %arrayidx2, align 427 %mul3 = fmul float %0, %128 store float %mul3, ptr %arrayidx2, align 429 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 130 br label %for.cond, !llvm.loop !031 32for.cond.cleanup: ; preds = %for.cond33 ret void34}35 36!0 = distinct !{!0, !1}37!1 = !{!"llvm.loop.vectorize.enable", i1 true}38