Re: x86/MMX/SSE shift by immediate

[Stuart Hastings <stuart at apple dot com>]

> Stuart Hastings <stuart@apple.com> writes:
>
> > #define NOINLINE __attribute__ ((noinline))
> >
> > void nowhere (const char *) NOINLINE;
> >
> > void
> > nowhere (const char *x)
> > {
> > }
>
> You need NOINLINE below here, too, or the attribute will be trampled
> over, and unit-at-a-time will inline "nowhere". (I think there's a PR
> about this, but it is how things are currenly.)
>
> --
> 	Falk

Inline nowhere, eh?  That's exactly what I wanted.

Ahem. ;-)

O.K., serious now.  I didn't know about this issue with unit-at-a-time; 
thank-you for pointing this out.  Fixed in the updated patch attached 
below.

> On Jan 7, 2004, at 2:13 AM, Richard Henderson wrote:
>
> > On Tue, Jan 06, 2004 at 05:51:16PM -0800, Stuart Hastings wrote:
> > >   /* Gratuitous use of a static forces save/restore of PIC register.
> > >      It should be O.K. to clobber the PIC register in this routine.  
> > > */
> > >   nowhere ("fubar");
> >
> > This only works for -fpic/-fPIC,

True, and it's probably unnecessary for the non-PIC case.

> > >   __asm__ ("cpuid" : "=a" (fl1) : "0" (0) : /* "ebx", "ecx", "edx", 
> > > */ "cc");
> >
> > Otherwise this really does kill ebx, and it needn't be restored.

I'm not sure if you're suggesting a change to the code, or just to the 
comments (?).  Maybe I would understand better if I were reading this 
at 2AM.  ;-)

Yes, both of the CPUID instructions here kill %eax, %ebx, %ecx, and 
%edx.  However, if I admit as much in the __asm__, and I compile with 
-fPIC (default on Darwin/x86), GCC stops here with the diagnostic

	/Volumes/sandbox/stuart/gcc.fsf/gcc/testsuite/gcc.dg/i386-cpuid.h:39: 
error: PIC register `ebx' clobbered in `asm'

Ergo, I chose to lie to GCC about the CPUID instruction.  :-(  I really 
dislike doing this; sleazy tricks like this usually cause maintenance 
headaches later on (e.g. if GCC gets smarter about the PIC register, we 
could and should describe CPUID correctly here).  Do you know a better 
way?

> > r~

To reiterate, I've run this testcase with DejaGnu on Darwin/x86, where 
-fPIC is the default, and by hand on Debian GNU/Linux, where it's not.  
I haven't tried it on a P3, but not on any hardware lacking MMX.  I've 
also revised some of the comments, and removed some dead local 
variables.  You've already approved the original patch to i386.md; now 
I'm just trying to get the testcase right, so I may commit everything 
simultaneously.

O.K. to commit ?

stuart hastings
Apple Computer

Attachment: gcc.fsf.diffs.txt
Description: Text document

Attachment: i386-sse-5.c
Description: Text document

Attachment: i386-cpuid.h
Description: Text document