[llvm-commits] [llvm] r126445 - in /llvm/trunk: lib/Transforms/Utils/SimplifyCFG.cpp test/Transforms/SimplifyCFG/select-gep.ll

[Evan Cheng]

On Feb 25, 2011, at 9:48 AM, Benjamin Kramer wrote:

> 
> On 25.02.2011, at 18:25, Evan Cheng wrote:
> 
>> 
>> On Feb 25, 2011, at 7:26 AM, Benjamin Kramer wrote:
>> 
>>> 
>>> On 25.02.2011, at 08:14, Evan Cheng wrote:
>>> 
>>>> Speculation is nice, but I would feel a lot better if there are hard numbers to back this up. Are we sure these changes are improving performance?
>>> 
>>> My unscientific test case (gcc compiled with clang -O3 parsing itself) showed a 0.25% speedup with the original patch (all-constant GEPs only, median of 5 runs, i386, penryn). I just retested with the change Frits suggested and it slowed down ~0.6% compared to the original run so I reverted it.
>> 
>> Thanks. What tests are you using to evaluate this?
> 
> I routinely use the nice C files from http://people.csail.mit.edu/smcc/projects/single-file-programs/
> which are easy to handle and large enough to get meaningful results for both compile time and run
> time. They're extracted from programs that are also part of SPEC CINT, so I'm confident the
> improvements will also show up there :)

These are nice, but not quite enough. In this future, please at least run the llvm testsuite for anything that might affect performance.

Thanks,

Evan