[AArch64] incorrect results from vcvt* functions with negative inputs when optimization is enabled

[llvmbot]

Bugzilla Link	46844
Version	10.0
OS	Linux
Reporter	LLVM Bugzilla Contributor
CC	@efriedma-quic,@smithp35

Extended Description

Several vcvt* functions return garbage instead of 0 when the input are < 0. Specifically:

• vcvts_u32_f32
• vcvtd_u64_f64
• vcvt_u32_f32
• vcvt_u64_f64
• vcvtq_u32_f32
• vcvtq_u64_f64

Here is a simple test case, compile without optimizations and the result is 0 (which is correct), but with -O1 or higher it is incorrect.

#include <arm_neon.h>
#include <stdio.h>
#include <inttypes.h>

uint32_t foo(float a) {
return vgetq_lane_u32(vcvtq_u32_f32(vdupq_n_f32(a)), 0);
}

int main(void) {
printf("%" PRIu32 "\n", foo(-3.14));

return 0;

}

[efriedma-quic]

Looking

[efriedma-quic]

Posted https://reviews.llvm.org/D84703

[efriedma-quic]

8dfb5d7 fixes 64-bit.

I'm not planning to write the fix for the equivalent issue for 32-bit.