Re: [PATCH 0/2][IRA,LRA] Fix PR86939, IRA incorrectly creates an interference between a pseudo register and a hard register

["H.J. Lu" <hjl dot tools at gmail dot com>]

On Sun, Sep 30, 2018 at 1:21 PM Peter Bergner <bergner@linux.ibm.com> wrote:
>
> On 9/28/18 4:34 PM, Vladimir Makarov wrote:
> > I remember I experimented with value numbering and using it in building
> > conflicts (that time it was global.c) but it did not improve the code as
> > I expected and made RA much slower.
>
> Yes, value numbering is not cheap to compute for catching special cases
> like this that probably don't happen too often.  Handling copies is fairly
> cheap and easy though.
>
>
> > We had some PRs with old RA code generation in case of using undefined values.
> > I don't remember what problems were exactly but I remember Ken Zadeck worked
> > on this too.  That is why I probably chose the current scheme of conflict
> > building.
> >
> > May be I was wrong because you testing shows that it is not a problem anymore.
>
> Maybe it was due to code relying on undefined behavior?  Anyway, it'll be good
> to let PATCH 1 bake on trunk for a little while before we commit PATCH 2 to
> catch any (if they exist) problems related to that change.
>
>
> > The 1st patch is ok for me.  You can commit it to the trunk.
>
> Ok, PATCH 1 is now committed, thanks!
>
>
> > I'll review the second patch on the next week.
>
> Sounds great as it gives any problems caused or exposed by PATCH 1 time to
> be reported.  Thanks again!

This caused:

FAIL: gcc.target/i386/pr63527.c scan-assembler-not movl[ \t]%[^,]+, %ebx
FAIL: gcc.target/i386/pr63534.c scan-assembler-not movl[ \t]%[^,]+, %ebx
FAIL: gcc.target/i386/pr64317.c scan-assembler addl[
\\t]+[$]_GLOBAL_OFFSET_TABLE_, %ebx
FAIL: gcc.target/i386/pr64317.c scan-assembler movl[ \\t]+c@GOTOFF[(]%ebx[)]

for Linux/i686.


-- 
H.J.