[Bug rtl-optimization/39543] [4.4 Regression] Reload failure on mplayer from SVN

[bonzini at gnu dot org]

------- Comment #7 from bonzini at gnu dot org  2009-03-25 16:49 -------
> one is the rtx_cost (SET_SRC (set), SET, speed) > old_cost check
> (I wonder whether it is needed for asms at all and if yes, how to actually do
> it)

For addresses it is already done in should_replace_address.  But for everything
else, there could be problems if one uses

  x = 0x123456789abcdef;

and fwprop propagates x into the asm.  If they used a "r" constraint
(correctly), reload probably will fix everything, but it may cause
pessimizations or I don't know what.  Sorry for the handwaving---the reason I
am a bit weary about propagating into asms is that we are quite careful about
it on the tree level.

pr39543-3.c crashes on f1 with optimization, but crashes on f2 without
optimization.  Which means that for f2, cse_not_expected does not work because
gimplification effectively has already done the same as force_reg.  In my
opinion that's the root cause: we worry about making asms complex, but we're
cavalier in making them simpler (hoping that something later restores the
complexity).

For 4.5, maybe SSA expansion fixes this?  (pr39543-3.c in Jakub's patch).  If
so, modifying fwprop is probably not the best thing to do.


-- 

bonzini at gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |matz at gcc dot gnu dot org
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2009-03-25 16:49:20
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39543