Loop schedule found despite various problems?

Clemens Eisserer

Hello,

I've written a hand optimized linear assembly function containing a loop, and the aseembly optimizer does afabulous jop optimizing it, generating a highly optimized loop kernel.

However, there seem to be several problems caused by register pressure as well as long live registers:

;*      Searching for software pipeline schedule at ...
;*         ii = 10 Register is live too long
;*         ii = 10 Cannot allocate machine registers
;*                   Regs Live Always   :  3/3  (A/B-side)
;*                   Max Regs Live      : 27/27
;*                   Max Cond Regs Live :  0/0 
;*         ii = 10 Schedule found with 3 iterations in parallel

What I don't understand is why/how the optimizer can generate a pipelined loop in 10 iterations, although it is reporting those problems with 10 iterations. Was the optimizer able to optimize those problems "away"?

Thanks you in advance, Clemens

over 13 years ago

0 Archaeologist over 13 years ago

TI__Guru* 84285 points

The compiler retries the same ii (initiation interval) with a different heurisitic. Different heuristics approach the problem in a different way, and often one heuristic will succeed where another failed. You are seeing the results of two failed heuristics and one that succeeded; only the last result shown is kept.

0 Clemens Eisserer over 13 years ago in reply to Archaeologist

Expert 2430 points

Thank you for the detailed explanation - I didn't know the optimizer tries several passes.

I have to admit I am very impressed by the code, the optimizer generates for loop kernels. Good job =)

Code Composer Studio™︎

Code Composer Studio forum

Loop schedule found despite various problems?