[fpc-devel] Slow if...then...else report
lagprogramming at aim.com
Sat Sep 13 21:11:27 CEST 2014
I'm looking for a 64bit processor that has similar timings with the Mobile Sempron presented inside the slow if then else report.
A very basic(incomplete) project is attached to the bug report that will help you identify if the CPU you are using is what I'm looking for.
If the program gives abnormal timings for "then" part, which can be identified with "(0,0)" lines compared to "else" part,identified with (0,1), please present the results.
I'm looking for:
1)Do we use many CPU models with such results or the presented one is an exception?
2)If the presented one is not unique, are there common things that these CPU share? For example, the same functions present the anomalies.
3)Is there a workaround that we can find for these CPUs?
There are already some hints regarding use of "jne" on a certain CPU, hints that I would like to compare with other CPUs anomalies.
4)Should the documentation regarding optimizing code be modified in order to inform the programmer of such anomalies?
For example extras from programmer's guide, "11.4 Tips to get faster code"
"Write your if/then/else statements so that the code in the "then"-part gets executed most of the time (improves the rate of successful jump prediction)."
Yes, but also at the moment I'm writing this text it looks like gambling with lots of resources at stake when using at least one particular CPU.
Even if your CPU doesn't have "spikes" like the one presented, how reliable is the above statement? You might consider these "micro-benchmarks" useless but looking a little bit in the sources will make you identify these "micro-benchmarks" all over fpc and lazarus source codes. There are so many small functions and lots of small loops...
Maybe "11.2 Processor specific" needs some update.
5)There is also an influence on loops due to these conditional jumps. Simply replacing a "while" with a "for" loop might have a huge impact on speed due to the initial assignment(not the pure precomputed value) of the "for" loop. I wrote this because I don't want to forget it(see the link found in the bug report, where original reporter probably didn't acknowledged this impact).
The bug report can be found at http://bugs.freepascal.org/view.php?id=26089
Most likely the direct link won't work, reason why I identify the report with: issue number is "26089", and (not a good)presentation is "Slow code generated for if...then...else".
The project attached to the bug report is a zip file "ifthenelse.zip". I tried to make it easy to use. A 800MHz CPU will take less than 6 minutes so it's not like an eternity.
If any of you is "lucky" enough to have a CPU with abnormal timings like the one presented in the "AMDSempronMobile3500plusx8664.txt" file, would I ask too much if you can use Delphi to double-check the results!? I've tried my best to keep the code ugly but as easy as possible to be modified for other pascal compilers. Maybe we'll have a surprise...who knows!?
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the fpc-devel