https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71636
Marc Glisse <glisse at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |missed-optimization
--- Comment #2 from Marc Glisse <glisse at gcc dot gnu.org> ---
For -march=core2, we get for g
movl $-1, %edx
movl %esi, %ecx
movl %edx, %eax
sall %cl, %eax
notl %eax
andl %edi, %eax
The use of %edx seems strange, llvm puts -1 directly in %eax. RA issue?
For the original issue, I guess doing the transformation very early with
something like the following might do
(simplify
(minus (lshift:s integer_each_onep@1 @0) @1)
(bit_not (lshift { build_all_ones_cst (type); } @0)))
(maybe with some extra code to handle conversions in the middle)
Or we could restrict the transformation to the case with bit_and, but I have
the vague impression that it would complicate things, especially if (1U<<b)-1
is used several times (although that case is probably less interesting, since
there is less to gain). I don't know if delaying the transformation to RTL
would be better / worse.
