LLVM doesn't know that psrad xmm,xmm ignores the high part of rhs

[Alcaro]

#include <immintrin.h>

__m128i a(__m128i __ix, __m128i __iy) {
	return _mm_sra_epi32(__ix, _mm_unpacklo_epi32(__iy, _mm_setzero_si128()));
}
__m128i b(__m128i __ix, __m128i __iy) {
	return _mm_sra_epi32(__ix, __iy);
}

__m128i c(__m128i __ix, __m128i __iy) {
	return _mm_sra_epi32(__ix, _mm_unpackhi_epi32(__iy, _mm_setzero_si128()));
}
__m128i d(__m128i __ix, __m128i __iy) {
	return _mm_sra_epi32(__ix, _mm_unpackhi_epi32(__iy, __iy));
}

Expected:

a(long long vector[2], long long vector[2]):
        psrad   xmm0, xmm1
        ret

b(long long vector[2], long long vector[2]):
        psrad   xmm0, xmm1
        ret

c(long long vector[2], long long vector[2]):
        pshufd  xmm1, xmm1, 250
        psrad   xmm0, xmm1
        ret

d(long long vector[2], long long vector[2]):
        pshufd  xmm1, xmm1, 250
        psrad   xmm0, xmm1
        ret

Actual:

a(long long vector[2], long long vector[2]):
        xorps   xmm2, xmm2
        movss   xmm2, xmm1
        psrad   xmm0, xmm2
        ret

b(long long vector[2], long long vector[2]):
        psrad   xmm0, xmm1
        ret

c(long long vector[2], long long vector[2]):
        pxor    xmm2, xmm2
        punpckhdq       xmm1, xmm2
        psrad   xmm0, xmm1
        ret

d(long long vector[2], long long vector[2]):
        pshufd  xmm1, xmm1, 250
        psrad   xmm0, xmm1
        ret

https://godbolt.org/z/1j5nWoW6q

Originally found in libstdc++ std::experimental::simd https://godbolt.org/z/d5Gdnsqch (they should probably add corresponding fixes to their header)

(See also #141475, though that one's about the shift instructions' lhs/ret, not rhs)

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: None (Alcaro)

```c++ #include <immintrin.h>

__m128i a(__m128i __ix, __m128i __iy) {
return _mm_sra_epi32(__ix, _mm_unpacklo_epi32(__iy, _mm_setzero_si128()));
}
__m128i b(__m128i __ix, __m128i __iy) {
return _mm_sra_epi32(__ix, __iy);
}

__m128i c(__m128i __ix, __m128i __iy) {
return _mm_sra_epi32(__ix, _mm_unpackhi_epi32(__iy, _mm_setzero_si128()));
}
__m128i d(__m128i __ix, __m128i __iy) {
return _mm_sra_epi32(__ix, _mm_unpackhi_epi32(__iy, __iy));
}

Expected:

a(long long vector[2], long long vector[2]):
psrad xmm0, xmm1
ret

b(long long vector[2], long long vector[2]):
psrad xmm0, xmm1
ret

c(long long vector[2], long long vector[2]):
pshufd xmm1, xmm1, 250
psrad xmm0, xmm1
ret

d(long long vector[2], long long vector[2]):
pshufd xmm1, xmm1, 250
psrad xmm0, xmm1
ret

Actual:

a(long long vector[2], long long vector[2]):
xorps xmm2, xmm2
movss xmm2, xmm1
psrad xmm0, xmm2
ret

b(long long vector[2], long long vector[2]):
psrad xmm0, xmm1
ret

c(long long vector[2], long long vector[2]):
pxor xmm2, xmm2
punpckhdq xmm1, xmm2
psrad xmm0, xmm1
ret

d(long long vector[2], long long vector[2]):
pshufd xmm1, xmm1, 250
psrad xmm0, xmm1
ret

https://godbolt.org/z/1j5nWoW6q

Originally found in libstdc++ std::experimental::simd https://godbolt.org/z/d5Gdnsqch (they should probably add corresponding fixes to their header)

(See also #<!-- -->141475, though that one's about the shift instructions' lhs/ret, not rhs)
</details>

[RKSimon]

https://www.felixcloutier.com/x86/psraw:psrad:psraq

SSE shifts use the bottom 64 bits of the shift amount operand to create the uniform shift amount, not just the bottom element of the result type. LLVM will discard the upper 64 bits.

[Alcaro]

Ah, you're right, I screwed up. Sorry, thanks.