Re: CSE bug (gcc.c-torture/execute/980605-1.c)

["David S. Miller" <davem at dm dot cobaltmicro dot com>]

   From: Carlo Wood <carlo@runaway.xs4all.nl>
   Date: Mon, 6 Jul 1998 16:49:04 +0200 (CEST)

   I have investigating the same bug for a few days, and I think I
   have evidence that this patch can't be correct.

   This is a preliminary post, I am not done with my research at all.
   But I thought it would be better to tell you that I am working on
   this too - and that I'd like to do a little more research before
   Davids patch is added :).

Don't fear, Jeff wrote a more complete fix and checked it in last
night.  But I believe you are looking at a different bug.

However I want to mention that I think fundamentally the problem
exposed with this CSE bug still exists generically, even with Jeff's
fix.  I do not believe that everyone in the compiler who rewrites
insns within' a libcall block properly update the REG_EQUAL notes at
the end.  I could argue that every call to validate_change() should
check for this situation, but even then all possible cases won't be
handled.

Jeff, when I showed my version of the patch to a few people I was
cautioned that I might have broken the SH target because it does very
strange things with libcalls.  Did you check to see if your version of
the changes have this effect?  Essentially I believe they make a
multiply or some other insn which clobbers a lot of hard registers
look like a libcall so that CSE can still occur in their presence.

Later,
David S. Miller
davem@dm.cobaltmicro.com