36 lines · plain
1; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s2 3@e = dso_local global [8 x i32] [i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8], align 164@d = dso_local global [8 x i32] [i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1], align 165 6; The global 'e' has 16 byte alignment, so make sure we don't generate an7; aligned 32-byte load instruction when we combine the load+insert sequence.8 9define dso_local i32 @subb() nounwind ssp {10; CHECK-LABEL: subb:11; CHECK: vmovups e(%rip), %ymm12entry:13 %0 = load i32, ptr getelementptr inbounds ([8 x i32], ptr @e, i64 0, i64 7), align 414 %1 = load i32, ptr getelementptr inbounds ([8 x i32], ptr @e, i64 0, i64 6), align 815 %2 = load i32, ptr getelementptr inbounds ([8 x i32], ptr @e, i64 0, i64 5), align 416 %3 = load i32, ptr getelementptr inbounds ([8 x i32], ptr @e, i64 0, i64 4), align 1617 %4 = load i32, ptr getelementptr inbounds ([8 x i32], ptr @e, i64 0, i64 3), align 418 %5 = load i32, ptr getelementptr inbounds ([8 x i32], ptr @e, i64 0, i64 2), align 819 %6 = load i32, ptr getelementptr inbounds ([8 x i32], ptr @e, i64 0, i64 1), align 420 %7 = load i32, ptr @e, align 1621 %vecinit.i = insertelement <8 x i32> undef, i32 %7, i32 022 %vecinit1.i = insertelement <8 x i32> %vecinit.i, i32 %6, i32 123 %vecinit2.i = insertelement <8 x i32> %vecinit1.i, i32 %5, i32 224 %vecinit3.i = insertelement <8 x i32> %vecinit2.i, i32 %4, i32 325 %vecinit4.i = insertelement <8 x i32> %vecinit3.i, i32 %3, i32 426 %vecinit5.i = insertelement <8 x i32> %vecinit4.i, i32 %2, i32 527 %vecinit6.i = insertelement <8 x i32> %vecinit5.i, i32 %1, i32 628 %vecinit7.i = insertelement <8 x i32> %vecinit6.i, i32 %0, i32 729 %8 = bitcast <8 x i32> %vecinit7.i to <32 x i8>30 tail call void @llvm.x86.avx.storeu.dq.256(ptr @d, <32 x i8> %8)31 ret i32 032}33 34declare void @llvm.x86.avx.storeu.dq.256(ptr, <32 x i8>) nounwind35 36