Re: PR2912

[Mark Mitchell <mark at codesourcery dot com>]

>>>>> "Richard" == Richard Henderson <rth@redhat.com> writes:

    Richard> On Thu, May 24, 2001 at 09:22:58PM -0700, Mark Mitchell
    Richard> wrote:
    >> The problem is that these peepholes fall afoul of what looks to
    >> be an undocumented restriction on peepholes: they must not
    >> change the set of live registers.

    Richard> Not so much "undocumented" as "if you do this you've made
    Richard> a mistake".

I guess the point is that I looked at the `define_peephole2' and then
I looked at the manual, and I didn't see why the definition was wrong,
so that means either I was dumb or the manual wasn't good enough.  I'm
willing to believe either alternative.

    >> The new peepholes get rid of instructions we don't need that
    >> mess with `al' and `dl' -- and thereby make those registers no
    >> longer killed.

    Richard> Um.. this makes no difference.

What happenned is that we propogated the fact that LIVE_AT_END
contained %eax (because it was live on entry to the successor blocks).
Then, we noticed that `%al' was set in the block, so me made `%eax'
not live at block start.  Then, we removed the set of `%al' in the
peephole.  Then, we verified our information and found that now `%eax'
was still live at block start, which was different than what we
expected, so we aborted.  
 
OK, I see what you're saying -- you're saying that it's bogus to
remove the set of `%al' because we don't know how else `%al' might be
used after this point.  That makes sense.

    Richard>   peep2_reg_dead_p (3, operands[7]) && peep2_reg_dead_p
    Richard> (3, operands[8])

Would you mind removing my patch (which I guess perpetuates the bug in
the original peephole, but makes it not crash us) with yours?  I'm in
the air very shortly, so I won't get a chance to run another test
cycle.

Thanks,

--
Mark Mitchell                   mark@codesourcery.com
CodeSourcery, LLC               http://www.codesourcery.com