1 ; RUN: llc < %s -mtriple=arm64-none-linux-gnu -mattr=+neon -O2 | FileCheck %s
3 ; st2 must before two ldrb.
4 ; The situation that put one ldrb before st2 because of the conservative memVT set for st2lane,
5 ; which lead to basic-aa goes wrong.
7 define dso_local i32 @test_vst2_lane_u8([2 x <8 x i8>] %vectors.coerce) local_unnamed_addr {
8 ; CHECK-LABEL: test_vst2_lane_u8:
9 ; CHECK: st2 { v[[V1:[0-9]+]].b, v[[V2:[0-9]+]].b }[6], [x8]
10 ; CHECK-NEXT: umov w[[W1:[0-9]+]], v[[V12:[0-9]+]].b[6]
11 ; CHECK-NEXT: ldrb w[[W2:[0-9]+]], [sp, #12]
12 ; CHECK-NEXT: ldrb w[[W2:[0-9]+]], [sp, #13]
14 %temp = alloca [2 x i8], align 4
15 %vectors.coerce.fca.0.extract = extractvalue [2 x <8 x i8>] %vectors.coerce, 0
16 %vectors.coerce.fca.1.extract = extractvalue [2 x <8 x i8>] %vectors.coerce, 1
17 call void @llvm.lifetime.start.p0(i64 2, ptr nonnull %temp) #4
18 call void @llvm.aarch64.neon.st2lane.v8i8.p0(<8 x i8> %vectors.coerce.fca.0.extract, <8 x i8> %vectors.coerce.fca.1.extract, i64 6, ptr nonnull %temp)
19 %0 = load i8, ptr %temp, align 4
20 %vget_lane = extractelement <8 x i8> %vectors.coerce.fca.0.extract, i64 6
21 %cmp8.not = icmp ne i8 %0, %vget_lane
22 %arrayidx3.1 = getelementptr inbounds [2 x i8], ptr %temp, i64 0, i64 1
23 %1 = load i8, ptr %arrayidx3.1, align 1
24 %vget_lane.1 = extractelement <8 x i8> %vectors.coerce.fca.1.extract, i64 6
25 %cmp8.not.1 = icmp ne i8 %1, %vget_lane.1
26 %or.cond = select i1 %cmp8.not, i1 true, i1 %cmp8.not.1
27 %cmp.lcssa = zext i1 %or.cond to i32
28 call void @llvm.lifetime.end.p0(i64 2, ptr nonnull %temp) #4
32 declare void @llvm.lifetime.start.p0(i64 immarg, ptr nocapture) #2
33 declare void @llvm.aarch64.neon.st2lane.v8i8.p0(<8 x i8>, <8 x i8>, i64, ptr nocapture) #2
34 declare void @llvm.lifetime.end.p0(i64 immarg, ptr nocapture) #2