SSE2 intrinsic optimization problems

[Khouri Giordano]

I have two problems. Both of them only happen with /O2.

First I have some code that uses the _mm_cvtsi32_si128 intrinsic. There are
two of them close together. I can see that one translates into:

B8 01 00 00 00 mov eax,1
66 0F 6E C0 movd xmm0, eax
66 0F 7F 84 24 E0 01 00 00 movdqa xmmword ptr [esp+1E0h],xmm0

But the other translates into:

BA 02 00 00 00 mov edx,2
0F 6E 94 24 10 02 00 00 movd mm2,dword ptr [esp+210h]

The location [esp+210h] is where that value is supposed to be stored.
The second problem happends when I am loading a bunch of values using
_mm_loadu_si128. Looking at the disassembly, these all translate into movdqu
and after a few it starts putting the values on the stack with movdqa.
Depending on the compiler state, it will switch these two opcodes and try to
load an unaligned value with movdqa and store it on the stack with movdqu.
I've been able to get the proper opcodes to be generated if I intersperse the
_mm_loadu_si128 with the code that uses the values.

I've tried to replicate both of these issues in a small console application,
but haven't been able to.