http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45903
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |ASSIGNED
AssignedTo|unassigned at gcc dot |jakub at gcc dot gnu.org
|gnu.org |
--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> 2010-10-06
13:04:23 UTC ---
Created attachment 21971
--> http://gcc.gnu.org/bugzilla/attachment.cgi?id=21971
gcc46-pr45903.patch
Untested patch that deals with this in the combiner. Of course it can't handle
f64or2 (and doesn't handle even cases when combiner rejects it for other
reasons, e.g.
unsigned long long a;
unsigned char c;
void foo (void)
{
c = a >> 8;
}
is not optimized, because (set (mem:QI (symbol_ref:P "c")) (mem:QI (const:P
(plus:P (symbol_ref:P "a") (const_int 1))))) is not recognized as insn.
At the tree level optionally it would be possible to perform such arithmetics
in narrower type and thus making this optimization to hit more often.
