[Bug optimization/7871] [3.3/3.4/3.5 regression] ICE on legal code, global register variables problems

["wilson at gcc dot gnu dot org" <gcc-bugzilla at gcc dot gnu dot org>]

------- Additional Comments From wilson at gcc dot gnu dot org  2004-02-18 08:49 -------
The patch that Eric Botcazou is proposing is identical to one of the early
patches I tried.  See comment #9.  Unfortunately, I neglected to include the
patch in this bug report, which obscured it.  The patch did not work for me when
I tried it.  Indeed, if I apply Eric's patch to gcc-3.3.x, it does not work.

It works in current sources apparently by accident.  See the Oct 10 combine.c
distribute_links patch from Richard Kenner.
    http://gcc.gnu.org/ml/gcc-patches/2003-10/msg00820.html
Note that he added a reg_set_p call, and note that reg_set_p assumes that a call
clobbers all hard registers, which is technically not correct.  This wrong
assumption is documented with a ??? comment.

I do think Eric's patch is a good one.  However, if we go this route, then we
have a latent problem with LOG_LINKS that may reappear in the future.  We should
be concerned about this.  I discuss this issue in comment #9 a little bit.  We
need to make sure that LOG_LINKS involving global_regs are treated specially. 
Currently, they are treated correctly by accident.

Adding some comments to distribute_notes and reg_set_p to document this would be
a start.  Perhaps that is all we can do for now.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=7871