http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29776
--- Comment #4 from Uros Bizjak <ubizjak at gmail dot com> 2012-07-30 11:36:23
UTC ---
> Perhaps REE can be taught about ctz giving a non-negative result.
Maybe we need some VRP pass to remove these extensions. Please note an example
from (duplicate) PR54115, where we generate:
#include <stdint.h>
uint64_t foo(long x){
return __builtin_ctzl(x);
}
foo:
bsfq %rdi, %rax
cltq
this is DImode -> DImode operation, followed by sign-extend from SImode partial
reg.
Your examples in comment #3:
bsfl %esi, %esi
movslq %esi, %rsi
can be "fixed" by slapping any_extend:DI to CTZ pattern, to consume either
ZERO_EXTEND or SIGN_EXTEND of the value (it doesn't matter for ranges [0..(some
small number)]).
The DImode example above can be "fixed" by adding SUBREG to all patterns. But,
I think that there is more optimal implementation than exploding the number of
bit manipulating patterns.
