25 lines · plain
1; RUN: opt -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a < %s -passes=loop-vectorize -S | FileCheck -check-prefix=GFX90A %s2 3; GFX90A-LABEL: @vectorize_v2f32_loop(4; GFX90A-COUNT-2: load <2 x float>5; GFX90A-COUNT-2: fadd fast <2 x float>6 7define float @vectorize_v2f32_loop(ptr addrspace(1) noalias %s) {8entry:9 br label %for.body10 11for.body:12 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]13 %q.04 = phi float [ 0.0, %entry ], [ %add, %for.body ]14 %arrayidx = getelementptr inbounds float, ptr addrspace(1) %s, i64 %indvars.iv15 %load = load float, ptr addrspace(1) %arrayidx, align 416 %add = fadd fast float %q.04, %load17 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 118 %exitcond = icmp eq i64 %indvars.iv.next, 25619 br i1 %exitcond, label %for.end, label %for.body20 21for.end:22 %add.lcssa = phi float [ %add, %for.body ]23 ret float %add.lcssa24}25