Re: Small update to reversed_comparison_code: Request Reversal

[Jeffrey Oldham <oldham at codesourcery dot com>]

On Mon, Mar 12, 2001 at 02:41:48PM +0100, Jan Hubicka wrote:
> > 
> > On Tue, 27 Feb 2001, Jan Hubicka wrote:
> > 
> > > This patch makes reversed_comparison_code smarter - it allows reversals
> > > like LT->UNGE, since they are safe - the resulting comparison may not trap.
> > > Also allows the opposite direction of reversals for -ffast-math mode.
> > >
> > > This patch is needed to make ifcvt match some cmoves on i386 w/o -ffast-math.
> > >
> > > Honza
> > >
> > > Tue Feb 27 17:52:14 CET 2001  Jan Hubicka  <jh@suse.cz>
> > > 	* jump.c (reversed_comparison_code_parts): Allow reversal of
> > > 	unordered compares in -ffast-math mode; reverse ordered compares
> > > 	for FP even w/o -ffast-math.
> > 
> > Hi,
> > 
> > This patch broke the arm-linux builds on the mainline.  My last
> > successful build was immediately before this patch was applied.  Reverting
> > this patch on the next day's build allows it to pass the failure listed
> > below.  All builds since March 1st fail with the same failure.
> > 
> > It is true that this change could simply have exposed another latent bug
> > in the compiler.  Let me know if more info is needed, I can send along
> > the .i file.
> Hi,
> 
> I will try to see what happends.  It looks this is the first change making
> gcc regulary use the unordered compares, so some previously untested code
> got triggered.

May we just have the patch reversed until the problems are fixed?  The
patch is an optimization that breaks correctness; bootstrapping on
mips-sig-irix6.5 and arm-linux have been broken since 01March.  Other
patches may also be breaking these bootstraps, but we do not know.

Thanks,
Jeffrey D. Oldham
oldham@codesourcery.com