[AArch64] Incorrect result for vector conversion with -O2

[llvmbot]

Bugzilla Link	45541
Version	10.0
OS	Linux
Attachments	Test case
Reporter	LLVM Bugzilla Contributor
CC	@Arnaud-de-Grandmaison-ARM,@smithp35

Extended Description

The attached test case works on without optimization and on GCC, but fails with -O2 (clang-10 -v -target aarch64-linux-gnu -O2 -o mm_cvtepu8_epi16 mm_cvtepu8_epi16.c to be precise).

I've added a FOO preprocessor macro (see line 27) to switch between making simde__m128i a typedef to int64x2_t and int32x4_t. Depending on which is used different functions fail. It shouldn't actually matter since the code uses __builtin_memcpy to convert to/from simde__m128i_private and simde__m128i.

[mr-c]

I was able to reproduce the provided test case with clang 13 and 14; it looks like this was fixed in clang-15. Therefore I think this issue can be closed

[sjoerdmeijer]

I agree, I can't reproduce this with a trunk build, closing this.