Re: [tree-ssa] Copy propagation [patch]

[law at redhat dot com]

In message <1046480389 dot 6056 dot 22 dot camel at frodo>, Diego Novillo writes:
 >On Fri, 2003-02-28 at 19:45, law at redhat dot com wrote:
 >
 >> FWIW, you can do this essentially for free while building the SSA
 >> form, along with value numbering and simple copy propagations.  In
 >> addition to being effectively free, you also get to cut down on the
 >> amount of code all the other SSA opts ever see.
 >> 
 >Right.  I forgot to mention that.  I've still not decided whether I'll
 >do the dominator opts first or the rewriting out of SSA.  But it would
 >probably be a good idea to get in sync first.
Well, given that you've got copy-prop done, we might as well go ahead
and use it to pound the ssa translator.

 >One of the things we could probably do is the propagation of pointer
 >copies.  Once we are in SSA form, propagating pointers is a bit awkward
 >because we need to update their associated INDIRECT_REFs and rewrite the
 >VUSES for the statements that dereference them.
Yup.


 >I also forgot to mention that since the regressions are only at runtime,
 >I did some compile time tests.  For insn-attrtab.c, copy propagation
 >simplifies the program quite a bit, cutting the time for schedule 2 in
 >half (compile time goes from 52 to 45 seconds).  Right now it's largely
 >meaningless, of course.  Many of those copies we are removing may end up
 >being re-added by the unSSA pass.
Yup.


 >> Anyway, I'm damn close to having the double-gimplification problem
 >> nailed down.  There's a couple nits that I'm still tracking down, but
 >> it's real real close.
 >> 
 >Cool.
I think I found it.  The inliner creates INIT_EXPRs :-0  If we don't
do a second simplification pass, they get through to the SSA code
unmolested.  It's easy 'nuff to fix (in fact, prelim tests are 
running now :-)

Oh yea, not only does this speed up gimplification, but the inliner
runs faster too :-)

jeff