Extreme PRE compile times

[Daniel Berlin]

On Thu, 13 Nov 2003 law@redhat.com wrote:

> In message <Pine.LNX.4.56.0311131122300.29117@dberlin.org>, Daniel Berlin write
> s:
>  >> >
>  >> > Clearly PRE is still doing something that's not terribly intelligent give
>  >n
>  >> > that PRE + the increase the GC time accounts for 90% of the total
>  >> > compilation time for this file.
>  >
>  >Looking at the dumps, it looks (it's not finished yet) like all the
>  >replacements it's performing are ones that could  have been caught
>  >by the dominator optimizers:
>  >
>  >In BB 1, insert save of this.2_4528 + 2B to pretmp.12002_44554 before
>  >statement T.4_4566 = pretmp.12002_44554
>  > on line 417
>  >In BB 15988, insert reload of this.2_4528 + 2B from pretmp.12002_44554 in
>  >statement T.4_44534 = this.2_4528 + 2B;
>  >
>  > on line 417
>  >
>  >
>  >
>  >In BB 3, insert save of this.2_4528 + 4B to pretmp.12004_44563 before
>  >statement T.6_4572 = pretmp.12004_44563
>  > on line 417
>  >In BB 15982, insert reload of this.2_4528 + 4B from pretmp.12004_44563 in
>  >statement T.6_44520 = this.2_4528 + 4B;
>  >
>  >
>  >
>  >None of these expressions are causing new phis to be inserted, which means
>  >they aren't partially redundant.
> Which probably argues that you should have some kind of filter which
> eliminates certain expressions from your candidates for PRE.  Trying to
> PRE something which isn't partially redundant is kindof silly.

Um, I can't.  There's no quick test i'm aware of for partial redundancy,
we can't determine it until well into the SSAPRE algorithm.

These expressions have EPHI's, they just turn out to be pointless.

It also looks like it's the splitting of critical edges that exposes them,
at first glance.


>
> Similar filtering on variables before PHI insertion on this testcase was
> a huge win.
>
> Also, as I've stated before, if we're leaving in redundancies like this after
> DOM, I'm more than happy to look at it if you can give me a reasonable
> testcase.

Uh, this is a reasonable testcase.

What exactly are you looking for?

You want me to remove all but one of those insertions so that it's a bit
smaller?