[PATCH] MachineBlockPlacement: Strengthen the source order bias when picking an exit block.

[Hal Finkel]

----- Original Message -----
> From: "Benjamin Kramer" <benny.kra at gmail.com>
> To: chandlerc at gmail.com, "benny kra" <benny.kra at gmail.com>
> Cc: llvm-commits at cs.uiuc.edu
> Sent: Sunday, November 17, 2013 8:20:57 AM
> Subject: [PATCH] MachineBlockPlacement: Strengthen the source order bias when	picking an exit block.
> 
> Hi chandlerc,
> 
> We now only allow breaking source order if the exit block frequency
> is
> significantly (20%+) higher than the other exit block. The percentage
> is
> taken from other parts of MBP and doesn't have any scientific basis.
> I
> haven't run the full test suite on it yet, but the benchmarks that
> historically had huge performance swings when changes to MBP were
> committed showed zero difference in the produced asm.
> 
> The test case is the full long_match function from zlib's deflate. I
> was
> reluctant to add it for previous tweaks to branch probabilities
> because
> it's large and potentially fragile, but changed my mind since it's an
> important use case and more likely to break with all the current work
> going into the PGO infrastructure. This change to MBP is worth
> another
> 5% of run time when deflating a random file with zlib.

I've run this through my test-suite benchmarking configuration on the BG/Q; I see no regressions and a 0.5%-1% speedup in MultiSource/Benchmarks/Fhourstones/fhourstones.

 -Hal

> 
> http://llvm-reviews.chandlerc.com/D2202
> 
> Files:
>   lib/CodeGen/MachineBlockPlacement.cpp
>   test/CodeGen/X86/longest_match.ll
> 
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
> 

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory