Questions about the instruction selection for sse2 intrinsic with -O optimization

Questions about the instruction selection for sse2 intrinsic with -O optimization

[haibo]

Hi:
	I am trying to compile sse2 program with gcc-3.4-02-25(prerelease version).
I found that I will get the suprise assembly code which is  different from
the syntax of the intrinsic function and the final result is wrong.
For example,for intrinsic :
 mark0=_mm_set1_epi16(0xffff);
 mark0 = _mm_insert_epi16(mark0,0,0);
I will get :
		movl    $65535, %eax
        movd    %eax, %xmm0
        punpcklwd       %xmm0, %xmm0
        pshufd  $0, %xmm0, %xmm0
        movdqa  %xmm0, -888(%ebp)
        movw    $15, %ax
        movd    %eax, %xmm0
        punpcklbw       %xmm0, %xmm0
        punpcklbw       %xmm0, %xmm0
        pshufd  $0, %xmm0, %xmm0
        movdqa  %xmm0, -568(%ebp) 
and no "pinsrw" found.
I wonder when and where gcc do these things in source code,
I have tried to step into the source code,and it do reach the code that
processes the _mm_insert_epi16 intrinsics and emit the rtl for it,but 
when gcc do insn match,It doesn't math it with sse2_pinsrw.
thanks.

 				

　　　　　　　　haibo
　　　　　　　　chbchb1130@sina.com
　　　　　　　　　　2004-03-09

Re: Questions about the instruction selection for sse2 intrinsic with -O optimization

[Giovanni Bajo]

haibo wrote:

> Hi:
> I am trying to compile sse2 program with gcc-3.4-02-25(prerelease
> version). I found that I will get the suprise assembly code which is
> different from the syntax of the intrinsic function and the final
> result is wrong.

Can you please file a bugreport on Bugzilla about this, together with a minimal
testcase to reproduce the bug? Thank you!

Giovanni Bajo

Re: Questions about the instruction selection for sse2 intrinsic with -O optimization

[Jim Wilson]

haibo wrote:
> I wonder when and where gcc do these things in source code,
> I have tried to step into the source code,and it do reach the code that
> processes the _mm_insert_epi16 intrinsics and emit the rtl for it,but 
> when gcc do insn match,It doesn't math it with sse2_pinsrw.

This is a bit too vague to easily answer.

You didn't provide a proper testcase that I can compile.  You didn't
provide example RTL dumps, so I am not sure exactly what you are
complaining about.

I am guessing that the pattern that the sse2_pinsrw pattern emitted is
also matched by other patterns, and hence you got other instructions.
This might be an optimization.  This might be an accidental bug.  It
might be a false impression on your part that the mm_insert_epi16
intrinsic should always map to a pinsrw instruction.  There isn't enough
information here to tell.
-- 
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com