LLVM forgets that movd+pslld zeroes the high part

[Alcaro]

#include <emmintrin.h>
#include <stdint.h>
__m128i a(uint32_t in)
{
	__m128i val = _mm_cvtsi32_si128(in*2);
	val = _mm_shufflelo_epi16(val, _MM_SHUFFLE(0,0,0,0));
    return val;
}
__m128i b(uint32_t in)
{
	__m128i val = _mm_cvtsi32_si128(in);
	val = _mm_shufflelo_epi16(val, _MM_SHUFFLE(0,0,0,0));
    return val;
}

Expected:

a(unsigned int):
        movd    xmm0, edi
        pslld   xmm0, 1
        pshuflw xmm0, xmm0, 0
        ret
b(unsigned int):
        movd    xmm0, edi
        pshuflw xmm0, xmm0, 0
        ret

(or, like gcc, do a scalar shift, I have no opinion either way)

Actual:

a(unsigned int):
        movd    xmm0, edi
        pslld   xmm0, 1
        xorps   xmm1, xmm1
        movss   xmm1, xmm0
        pshuflw xmm0, xmm1, 0
        ret

b(unsigned int):
        movd    xmm0, edi
        pshuflw xmm0, xmm0, 0
        ret

https://godbolt.org/z/W96qqGKc6

The xorps+movss dance takes the low 32 bits from xmm0 into xmm1 and zeroes the rest of xmm1, which is completely unnecessary because the upper 96 bits of xmm0 are already zero.

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: None (Alcaro)

```c #include <emmintrin.h> #include <stdint.h> __m128i a(uint32_t in) { __m128i val = _mm_cvtsi32_si128(in*2); val = _mm_shufflelo_epi16(val, _MM_SHUFFLE(0,0,0,0)); return val; } __m128i b(uint32_t in) { __m128i val = _mm_cvtsi32_si128(in); val = _mm_shufflelo_epi16(val, _MM_SHUFFLE(0,0,0,0)); return val; } ``` Expected: ```asm a(unsigned int): movd xmm0, edi pslld xmm0, 1 pshuflw xmm0, xmm0, 0 ret b(unsigned int): movd xmm0, edi pshuflw xmm0, xmm0, 0 ret ``` (or, like gcc, do a scalar shift, I have no opinion either way)

Actual:

a(unsigned int):
        movd    xmm0, edi
        pslld   xmm0, 1
        xorps   xmm1, xmm1
        movss   xmm1, xmm0
        pshuflw xmm0, xmm1, 0
        ret

b(unsigned int):
        movd    xmm0, edi
        pshuflw xmm0, xmm0, 0
        ret

https://godbolt.org/z/W96qqGKc6

The xorps+movss dance takes the low 32 bits from xmm0 into xmm1 and zeroes the rest of xmm1, which is completely unnecessary because the upper 96 bits of xmm0 are already zero.

[RKSimon]

define <2 x i64> @a(i32 %in) {
entry:
  %mul = shl i32 %in, 1
  %vecinit3.i = insertelement <4 x i32> <i32 poison, i32 poison, i32 0, i32 0>, i32 %mul, i64 0
  %0 = bitcast <4 x i32> %vecinit3.i to <8 x i16>
  %pshuflw = shufflevector <8 x i16> %0, <8 x i16> poison, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 4, i32 5, i32 6, i32 7>
  %1 = bitcast <8 x i16> %pshuflw to <2 x i64>
  ret <2 x i64> %1
}

define <2 x i64> @b(i32 %in) {
entry:
  %vecinit3.i = insertelement <4 x i32> <i32 poison, i32 poison, i32 0, i32 0>, i32 %in, i64 0
  %0 = bitcast <4 x i32> %vecinit3.i to <8 x i16>
  %pshuflw = shufflevector <8 x i16> %0, <8 x i16> poison, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 4, i32 5, i32 6, i32 7>
  %1 = bitcast <8 x i16> %pshuflw to <2 x i64>
  ret <2 x i64> %1
}