Re: ICE with current mainline gcc

[Richard Henderson <rth at cygnus dot com>]

On Mon, Sep 25, 2000 at 06:08:18PM -0700, Ulrich Drepper wrote:
> With the current mainline gcc as of 2000-09-26T18:00-0700 I'll get an
> ICE for the file at
> 
> 	http://www.cygnus.com/~drepper/dcigettext.i.gz

Here's a "reduced" test case.  It appears to be the expansion of

  stpcpy (mempcpy (stpcpy (stpcpy (xdomainname, categoryname), "/"),
                  domainname, domainname_len),
          ".mo");

Beware.  This expands to over 180k of text, all one big
monsterous expression.  It does compress well though.  ;-)

It appears to be a bad interaction between the new functions
as trees switch table code and the alternate insn chains of
a call_placeholder.  I.e. the nesting around a call isn't correct,
and a label wound up inside one of the chains that was discarded.

I think it is clear that fold-const.c should be taught to recognize
when __builtin_constant_p is true so that we can prune some of these
expressions immediately.  Recognizing when it will be false is harder,
since global constant propogation should be done first.


r~

zz.c.gz