35 lines · plain
1; RUN: llc < %s -mtriple=arm64-none-linux-gnu -mattr=+neon -O2 | FileCheck %s2 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.6 7define 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]13entry:14 %temp = alloca [2 x i8], align 415 %vectors.coerce.fca.0.extract = extractvalue [2 x <8 x i8>] %vectors.coerce, 016 %vectors.coerce.fca.1.extract = extractvalue [2 x <8 x i8>] %vectors.coerce, 117 call void @llvm.lifetime.start.p0(i64 2, ptr nonnull %temp) #418 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 420 %vget_lane = extractelement <8 x i8> %vectors.coerce.fca.0.extract, i64 621 %cmp8.not = icmp ne i8 %0, %vget_lane22 %arrayidx3.1 = getelementptr inbounds [2 x i8], ptr %temp, i64 0, i64 123 %1 = load i8, ptr %arrayidx3.1, align 124 %vget_lane.1 = extractelement <8 x i8> %vectors.coerce.fca.1.extract, i64 625 %cmp8.not.1 = icmp ne i8 %1, %vget_lane.126 %or.cond = select i1 %cmp8.not, i1 true, i1 %cmp8.not.127 %cmp.lcssa = zext i1 %or.cond to i3228 call void @llvm.lifetime.end.p0(i64 2, ptr nonnull %temp) #429 ret i32 %cmp.lcssa30}31 32declare void @llvm.lifetime.start.p0(i64 immarg, ptr nocapture) #233declare void @llvm.aarch64.neon.st2lane.v8i8.p0(<8 x i8>, <8 x i8>, i64, ptr nocapture) #234declare void @llvm.lifetime.end.p0(i64 immarg, ptr nocapture) #235