60 lines · plain
1; RUN: opt -passes='loop-vectorize,loop-mssa(licm)' -force-vector-width=2 \2; RUN: -force-vector-interleave=1 -S < %s | FileCheck %s3 4target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"5 6; In order to vectorize the inner loop, it needs to be versioned with7; memchecks between {A} x {B, C} first:8;9; for (i = 0; i < n; i++)10; for (j = 0; j < m; j++)11; A[j] += B[i] + C[j];12;13; Since in the versioned vector loop A and B can no longer alias, B[i] can be14; LICM'ed from the inner loop.15 16 17define void @f(ptr %a, ptr %b, ptr %c) {18entry:19 br label %outer20 21outer:22 %i.2 = phi i64 [ 0, %entry ], [ %i, %inner.end ]23 %arrayidxB = getelementptr inbounds i32, ptr %b, i64 %i.224 br label %inner.ph25 26inner.ph:27; CHECK: vector.ph:28; CHECK: load i32, ptr %arrayidxB,29; CHECK: br label %vector.body30 br label %inner31 32inner:33 %j.2 = phi i64 [ 0, %inner.ph ], [ %j, %inner ]34 35 %arrayidxA = getelementptr inbounds i32, ptr %a, i64 %j.236 %loadA = load i32, ptr %arrayidxA, align 437 38 %loadB = load i32, ptr %arrayidxB, align 439 40 %arrayidxC = getelementptr inbounds i32, ptr %c, i64 %j.241 %loadC = load i32, ptr %arrayidxC, align 442 43 %add = add nuw i32 %loadA, %loadB44 %add2 = add nuw i32 %add, %loadC45 46 store i32 %add2, ptr %arrayidxA, align 447 48 %j = add nuw nsw i64 %j.2, 149 %cond1 = icmp eq i64 %j, 2050 br i1 %cond1, label %inner.end, label %inner51 52inner.end:53 %i = add nuw nsw i64 %i.2, 154 %cond2 = icmp eq i64 %i, 3055 br i1 %cond2, label %outer.end, label %outer56 57outer.end:58 ret void59}60