------- Comment #13 from law at redhat dot com 2006-03-28 19:13 -------
Subject: Re: [4.1/4.2 Regression] missed jump
threading after unroller
On Wed, 2006-03-22 at 16:06 +0100, Richard Guenther wrote:
> On 3/22/06, Jeffrey A Law <[EMAIL PROTECTED]> wrote:
> > On Wed, 2006-03-22 at 12:14 +0100, Richard Guenther wrote:
> > > On 3/21/06, Jeffrey A Law <[EMAIL PROTECTED]> wrote:
> > > > It turns out this specialized PHI optimization pass is as effective
> > > > as running copy-prop and CCP on PHI nodes after DOM. Better yet, it
> > > > is a teeny tiny slowdown compared to just running the stripped down
> > > > copyprop. ie, for an almost unmeasurable slowdown we can do both
> > > > constant and copy propagation instead of just copy propagation.
> > >
> > > This patch caused a compile-time regression from 139s to 143s, resp.
> > > 192s to 197s (leafify) accounted by increases of operand scan / SSA
> > > incremental
> > > and tree CCP times for compiling tramp3d. Also memory usage during
> > > compiling
> > > went up from 655494 kB to 660626kB (this may be due to the VRP patch,
> > > though).
> > >
> > > Runtime of tramp3d did not improve but regress slightly (but that
> > > might be in the
> > > noise - we'll see).
> > >
> > > For this simple cleanup pass can you try updating SSA form manually
> > > please?
> > I'm more than happy to look at it; however, be aware that if you're
> > seeing increased time in CCP then either you're seeing some truly
> > bizzarre secondary effect or your testing methodology is suspect.
> > The patch did not affect CCP. In fact, the changes only affect
> > passes which run *after* CCP in the optimization pipeline.
>
> struct tree_opt_pass pass_phi_only_cprop =
> {
> "phicprop", /* name */
> gate_dominator, /* gate */
> eliminate_degenerate_phis, /* execute */
> NULL, /* sub */
> NULL, /* next */
> 0, /* static_pass_number */
> TV_TREE_CCP, /* tv_id */
> PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
> 0, /* properties_provided */
> PROP_smt_usage, /* properties_destroyed */
> 0, /* todo_flags_start */
> TODO_cleanup_cfg | TODO_dump_func
> | TODO_ggc_collect | TODO_verify_ssa
> | TODO_verify_stmts | TODO_update_smt_usage
> | TODO_update_ssa, /* todo_flags_finish */
> 0 /* letter */
> };
>
> see tv_id - so I guess increased CCP times are expected.
I've created a separate timevar for the phi-only cprop code so
at least we can see how much time's taking. Note that while we
should be seeing a small amount of time in phi-cprop, we should
be seeing a reduction in TV_TREE_CCP and TV_TREE_COPY_PROP.
Can you send me a recent tramp3d.ii file so that I can poke
around and see if there's anything we can do about the compile
time regression?
Thanks,
jeff
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=21829
