#3285 Suspected quadratic performance hog when compiling inline functions

[Oleg Endo]

I'm compiling larger, mostly generated, code which consists of a lot of (nested) inline functions for MCS-51. The compile times of those codes are just too long, in the minutes for a single source file.

Is this a known issue? Any suggestions where to look or how to improve it?

[Erik Petrich]

I don't know of anything specific to inlining, but I am aware of a performance problem in GCSE optimization in large functions that have lots of branches. Inlining will create a lot of temporary variables that GCSE will try to clean up and perhaps this is triggering the problem? You might try the --nogcse option when compiling to see if that makes any significant effect to the compile time. I have started on a possible improvement to the GCSE time, but have not had time to complete and test it.

[Oleg Endo]

Yes, it looks like we got a winner.

Recompiling single file in my project without --nogcse:
real 2m26.212s
user 2m24.992s
sys 0m0.641s

and with --nogcse:
real 0m3.638s
user 0m3.356s
sys 0m0.386s

Something's terribly wrong there ;)
Any suggestions where to look in the code? Maybe I can come up with a quick hack for myself. It's really a showstopper for me at the moment. Compile-Run cycles are in the minutes.

[Erik Petrich]

If it's the problem I think it is, then removing these lines from the cseBBlock() function in src/SDCCcse.c should speed it up (these were added in [r9700]):

  if (recomputeDataFlow)
     computeDataFlow (ebbi);

However, that would also reenable bug [#2495].

My current plan is to replace the call to computeDataFlow(), which recomputes the data flow of the entire function, with a call to a new function that only recomputes the data flow of the expressions changed by the GCSE optimization. I have this partially completed and will give finishing it priority when I have time available.

[Oleg Endo]

I've tried to comment out those lines in SDCCcse.c. It is a significant improvement. With those lines commented out and GCSE enabled I get

real 0m11.426s
user 0m11.219s
sys 0m0.306s

It's noticeably slower than no GCSE optimization, but looks like the right direction.