Steven Bosscher wrote:
On 4/12/07, Steven Bosscher <[EMAIL PROTECTED]> wrote:
On 4/12/07, Vladimir Makarov <[EMAIL PROTECTED]> wrote:
> An interesting observation is that the more hard registers the
processor
> has, the bigger slowdown is. Although it might be a coincidence.
Yes, I noticed this too. I don't believe this is a coincidence. It's
the first thing I was planning to look into, in fact.
The problem seems to have to do with how the RTL loop optimizers work.
They analyze a subcfg with df, and this is incredibly slow right now.
To give you some idea, we spend more than 35% of the total compile
time in "loop analysis" (TV_LOOP) for a fairly typical Fortran 90
program with lots of small loops, and almost all that extra time is
spent in df_reorganize_refs_by_reg.
So, Vlad, thanks for helping us identify a serious bottleneck! Kenny
said he has some ideas about how to fix this.
No problem. New DF infrastructure is inevitable (we differ only in
question when to merge). I see a real progress in compilation speed
improvement. I am going to do more benchmarking (on merge points) to
check the progress.
As you probably saw from my email, there is SPECFP2000 score degradation
on ppc64 (all other ports looks good). The biggest degradation is
wupwise (about 7%). Wupwise is relatively small program. It would be
nice if you fix it.
I also see bigger code size on all platforms. As I wrote if you fix
that, you will be closer to 5% target or even there. And ppc is the
best platform to do it.