https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85720
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Severity|normal |enhancement
--- Comment #8 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Mathias Stearn from comment #2)
> Hmm. Taking the example from the -ftree-loop-distribute-patterns
> documentation, it still seems to generate poor code, this time at both -O2
> and -O3: https://godbolt.org/g/EsQDj8
>
> Why isn't that transformed to memset(A, 0, N); memset(B, 1, N); ? This feels
> similar to https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85721. Should I make
> a new ticket with this example?
For this case (with the mentioned B[i] = A[i] + 1 fix), in GCC 10+ we started
to produce memset for one of the alias conditions.
-ftree-loop-distribute-patterns is turned on at -O2 for starting in GCC 10
also.
I should note clang, ICC nor MSVC is able to do this second loop to convert it
to two memset even with an alias check; others just vectorize the loop with an
aliasing check. clang and MSVC are able to detect the one in comment #0 and
convert it to memset though so I know they have this kind of optimization for
sure.
