This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [RFC] By default if-convert only basic blocks that will be vectorized (take 4)


On Wed, Oct 23, 2013 at 07:22:20PM +0200, Jakub Jelinek wrote:
> On Tue, Oct 22, 2013 at 08:27:54PM +0400, Sergey Ostanevich wrote:
> > still fails on 403 et al.
> 
> Ok, reproduced, unfortunately the pending stmt sequences already pretty much
> assume that they will end up in a single combined basic block.  I went
> through various alternatives (deferring update_ssa (TODO_update_ssa) call
> until after combine_blocks - doesn't work, because it is unhappy about
> basic blocks being removed, temporarily putting all the stmts into latch
> (doesn't work, because there are no PHIs for it in the loop), so the final
> fix as discussed with Richard on IRC is not to predicate_bbs early before
> versioning (unless -ftree-loop-if-convert-stores it is easily achievable
> by just using a better dominance check for that), or for the stores stuff
> doing it and freeing again (at least for now).
> 
> The predicate_bbs stuff would certainly appreciate more TLC in the future.
> 
> Attaching whole new patchset, the above mentioned fix is mostly in the
> first patch (which also contains a tree-cfg.h include that is needed for
> today's header reshufling), the other two patches are just tweaked to
> apply on top of that.  All 3 patches together have been
> bootstrapped/regtested on x86_64-linux and i686-linux, the first one
> and first+second just compile time tested.

BTW, to test the effect of not disabling if-conversion for non-vectorized
loops unless there were conditional loads/stores, instead of patching
you could just test with additional -ftree-loop-if-convert I guess.

	Jakub


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]