1 ; RUN: opt < %s -force-vector-interleave=1 -store-to-load-forwarding-conflict-detection=false -loop-vectorize -dce -instcombine -S | FileCheck %s
3 target datalayout = "e-m:e-i64:64-i128:128-n32:64-S128"
4 target triple = "aarch64--linux-gnu"
6 %struct.pair = type { i32, i32 }
8 ; Check vectorization of interleaved access groups with positive dependence
9 ; distances. In this test, the maximum safe dependence distance for
10 ; vectorization is 16 bytes. Normally, this would lead to a maximum VF of 4.
11 ; However, for interleaved groups, the effective VF is VF * IF, where IF is the
12 ; interleave factor. Here, the maximum safe dependence distance is recomputed
13 ; as 16 / IF bytes, resulting in VF=2. Since IF=2, we should generate <4 x i32>
14 ; loads and stores instead of <8 x i32> accesses.
16 ; Note: LAA's conflict detection optimization has to be disabled for this test
24 ; void max_vf(struct pair *restrict p) {
25 ; for (int i = 0; i < 1000; i++) {
31 ; CHECK-LABEL: @max_vf
32 ; CHECK: load <4 x i32>
33 ; CHECK: store <4 x i32>
35 define void @max_vf(%struct.pair* noalias nocapture %p) {
40 %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
41 %0 = add nuw nsw i64 %i, 2
42 %p_i.x = getelementptr inbounds %struct.pair, %struct.pair* %p, i64 %i, i32 0
43 %p_i_plus_2.x = getelementptr inbounds %struct.pair, %struct.pair* %p, i64 %0, i32 0
44 %1 = load i32, i32* %p_i.x, align 4
45 store i32 %1, i32* %p_i_plus_2.x, align 4
46 %p_i.y = getelementptr inbounds %struct.pair, %struct.pair* %p, i64 %i, i32 1
47 %p_i_plus_2.y = getelementptr inbounds %struct.pair, %struct.pair* %p, i64 %0, i32 1
48 %2 = load i32, i32* %p_i.y, align 4
49 store i32 %2, i32* %p_i_plus_2.y, align 4
50 %i.next = add nuw nsw i64 %i, 1
51 %cond = icmp eq i64 %i.next, 1000
52 br i1 %cond, label %for.exit, label %for.body