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]

Re: SSA vs cross jumping/tail merging


  In message <20010815105626.F19872@atrey.karlin.mff.cuni.cz>you write:
  > Hmm, I see you are running to similar issue, as reg-stack does.  In that
  > case we also emit code to the edges, that is (partially) equivalent for
  > other edges incomming to the same block.
OK.  Also note we can get this kind of stuff from gcse too, though I suspect
it's less likely than SSA to generate sequences that can be optimized via
cross jumping/tail merging.

  > We do use crossjumping sucesfully to cleanup.  In SSA case this should
  > work well, but you need to consider drawbacks:
  > 1) Crossjumping is rather expensive pass
OK.

  > 2) As implemented, it breaks natural loops
OK.  This is potentially bad since we'd want to zap the unwanted stuff
just after conversion back to normal form (ie before the loop optimizer
runs).

  > 3) Currently it is not able to get around case, where different pseudos
  >    are used as temporaries in different copies of code.
Not an issue.

  > 4) Overall it increases number of basic blocks, that does bad to our local
  >    optimizers. 
For the cases I'm looking at it will decrease the number of basic blocks.

  > All problems are avoidable of course, but overall, maybe practical
  > sollution can be to teach commit edge insertion to discover duplicates
  > and merge them while modifying the CFG,
Possibly.  And the more I hear about the state of the cross jumping code
the more I'm inclined to go this direction.  Particularly since it would
potentially help insertions done by gcse and reg-stack.

jeff



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