hi!

I try to convert a normal vector from a DWORD to float3, using the following code:

__m128i n_i;

n_i.m128i_i32[0] = n&0xff;

n_i.m128i_i32[1] = (n >> 8)&0xff;

n_i.m128i_i32[2] = (n >> 16)&0xff;

n_i.m128i_i32[3] = 0;

__m128 n_f = _mm_cvtepi32_ps(n_i);

...

here are the assembly:

...

mov dword ptr [esp], ecx

mov dword ptr [esp+0x4], edx

mov dword ptr [esp+0x8], eax

mov dword ptr [esp+0xc], 0

movdqa xmm0, xmmword ptr [esp]

cvtdq2ps xmm0, xmm0

...

And "cvtdq2ps xmm0, xmm0" has a high CPI rate(1.65).

According to https://fgiesen.wordpress.com/2013/03/04/speculatively-speaking/ , CPU can not forward multiple store to one big load. I wonder whether this is a load hit store or not.

I use _mm_loadu_si128 and process 4 normal at a time, and the stall is gone, thank you!