RE: [Patch, fortran] Inline DOT_PRODUCT

["THOMAS Paul Richard 169137" <paul dot richard dot thomas at cea dot fr>]

Janne,

> -----Message d'origine-----
> De : Janne Blomqvist [mailto:jblomqvi@cc.hut.fi]
> Envoye : lundi 27 fevrier 2006 13:13
> A : THOMAS Paul Richard 169137
> Cc : paolo.bonzini@lu.unisi.ch; fortran@gcc.gnu.org;
> gcc-patches@gcc.gnu.org
> Objet : Re: [Patch, fortran] Inline DOT_PRODUCT
> 
> 
> On Mon, Feb 27, 2006 at 12:27:12PM +0100, THOMAS Paul Richard 
> 169137 wrote:
> > Paolo,
> > > 
> > > > Have you tested -ftree-vectorize?
> > > 
> > > No, I haven't and I should.  I will report back tonight.
> > > 
> > 
> > It makes no difference whatsoever - should I be changing 
> the TREEery to benefit from it? 
> > 
> > On a PIV 2.5GHz, under Cygwin_NT
> 
> I'm not that familiar with cygwin, but you might need the -msse2 flag
> to tell the compiler that you actually have vector hw available,

This does not make any difference that I can see.

> otherwise it'll use x87 only. You might want -mfpmath=sse too just to

This option breaks the compiler - the executable does, well.... nothing.

> make sure it doesn't waste time moving stuff between sse and x87 regs
> (though I suspect your benchmark should be safe). And with
> -ftree-vectorizer-verbose=5 you might get some additional info about
> what the vectorizer is doing.

This gives me a "f951: error: unrecognized command line option
"-ftree-vectorize-verbose=5"" for my pains.

> 
> As for the patch itself, now that everything is done inline, can't one
> get rid of gfc_resolve_dot_product entirely, or does it still do
> something useful?

It is combining the expressions into an intrinsic expression and setting
the type of the result.  This could be done elsewhere but the usage is
consistent with other functions.  The writing of the library function
name could be discarded.

Paul