This is the mail archive of the gcc@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: [tree-ssa] -fgcse and friends useless



On Feb 6, 2004, at 1:28 PM, Zdenek Dvorak wrote:


Hello,

It seems that the tree-optimizers do all the work that rtl level gcse and
friends (gcse-lm, gcse-sm, gcse-las) do.

I was not aware we do load & store motion on tree-ssa (or does PRE handle this?)

SSAPRE does load motion on loads that occur > 1 time (be they direct or indirect loads).


Doing it on loads that occur only once is too expensive right now.
This is mainly because
1. process_left_occs_and_kills needs to be sped up through algorithmic changes so that it doesn't have to do O(n^3) work in the worst case.
2. we need to be able to do rename_2 without generating tons of new trees. Though I guess i can now ggc_free this stuff explicitly, which may help.


I never implemented SSUPRE (which is gcse-sm), though i could.

But for simplicity sake, we'd probably be better off implementing regular FSM (forward store motion) using dominator based sinking/removal, at least for now.
--Dan



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