Re: PATCH: [gcc3.5 improvement branch] Very Simple constant propagation

[Caroline Tice <ctice at apple dot com>]

In an attempt to figure out why my numbers were so different from 
everybody else's I re-downloaded
both gcc3,4 and gcc3.5, made two copies of each, and applied my and 
Roger's patches to both 3.4
and 3.5 (so I ended up with four compilers, a 3.4 & a 3.5 version of 
each of our patches).  On
re-running all my timing measurements I discovered the following:

1).  I was no longer able to get my low numbers.  My best guess is that 
I originally had some out-of-date
version of the compiler, and when I updated everything I lost whatever 
was giving me the good numbers.

2).  In comparing my patch against Roger's patch, the times were always 
nearly identical.  In light of
this, it seems best for Roger's patch to be committed (subject to 
approval of course :-)), and mine to
be withdrawn.

3).  Just a point of interest: in every case the compile times for 3.5 
were about 10% slower than
3.4.  Sigh.


So, conditional on someone approving Roger's patch, consider mine 
withdrawn.  I apologize for
all the confusion.

-- Caroline Tice
ctice@apple.com

On Feb 4, 2004, at 12:15 AM, Andrew Pinski wrote:

> On Jan 22, 2004, at 17:46, Andrew Pinski wrote:
>
> > On Jan 22, 2004, at 15:06, Caroline Tice wrote:
> >
> > > Hi Roger,
> > >
> > > I've been playing around with your patch and comparing it to mine, 
> > > with the following
> > > results.
> >
> > > Benchmark        My Patch          Your Patch
> > >
> > > gzip                6.65            12.74
> > > vpr                20.83            36.98
> > > gcc               244.13           399.41
> > > mcf                 1.75             3.28
> > > crafty             31.95            61.89
> > > parser             15.61            31.29
>
> I was asked by Caroline to do some timings for both of these patches 
> as a semi-third party.
> Here are my results.
>
> Here are the results from my machine (a PPC G4 800MHz laptop running 
> Mac OS X 10.3.2 aka
> darwin7.2.0), I can also do an x86 box running Linux also if need be.  
> All sources were
> from "Tue Feb  3 19:34:23 UTC 2004" except for 3.3.3 was from 20040201.
>
> I only did one IMO because the SPEC benchmarks are not really C 
> complaint and I do not
> have a patch to ignore the difference in function prototypes.
>
> Benchmark       Without   Caroline    Roger     3.3.3 (for reference)
> crafty          61.59     62.310      62.250
> mcf (with IMO)  3.260     3.290       3.340     2.570 (not with IMO)
> fold-const.i    94.55     96.070      95.300    67.590
>   targeting PPC64 from June
>
> What I can conclude from this is that their patches are almost equally 
> in compile time but
> Caroline's can cause memory overhead to go up which we consider bad as 
> there was just a
> reduction of memory usage on the mainline done by Jan Hubicka.  Also 
> the compile time has
> regressed from 3.3.3 but which is already a PR 13987; this is mostly 
> because of a patch
> by me to fully support -mpowerpc64 on Darwin as HOST_WIDE_INT needs to 
> be 64bit now.  This
> might also explain what Caroline did wrong in her testings as it looks 
> like she did not use
> the same source tree for both patches but used one where HOST_WIDE_INT 
> was 32bit for hers
> while 64bit for Roger's which causes this compile time regression.
>
>
> Thanks,
> Andrew Pinski