[fpc-devel] Peephole Optimizer - work in reducing number of passes

[J. Gareth Moreton]

Hi everyone,

Being on a working Intel laptop full-time again, I've gone back to 
looking at the Peephole Optimizer like I always used to do. One thing 
I've been focusing on is an attempt to reduce the number of passes of 
the Peephole Optimizer, since a single pass amounts to stepping through 
every instruction in the pre-compiled assembly language, which can get 
long for complex routines.

https://bugs.freepascal.org/view.php?id=37959

The main change is that the mandatory second run of pass 1 has been 
removed, and the upper limit on the number of iterations performed is 
governed by whether the setting is -O2 (2), -03 (5) or -O4 (8), and 1 
otherwise - details are in the bug report.  As a result -O1 will perform 
slightly worse, but I gathered from past conversations that this is 
acceptable as this is generally only done when debugging - if it's not, 
then this can be easily fixed by changing the upper limit from 1 to 2 
for -O1.  -O2 and above should be exactly the same in regards to the 
code that is produced (if it's not, then there's likely something wrong 
with an optimisation not setting Result to True when it should have 
done).  So far, I've had success with the RTL and Lazraus when it comes 
to this.

Ultimately, there should be a slight increase in compiler performance 
under high optimisation settings... at least that's the theory.  Under 
x86 platforms at least, I've been making sure optimisations do set 
Result to True when they should.

Gareth aka. Kit


-- 
This email has been checked for viruses by Avast antivirus software.
https://www.avast.com/antivirus