[llvm-commits] [llvm] r151111 - /llvm/trunk/lib/CodeGen/RegAllocFast.cpp

[Jakob Stoklund Olesen]

On Feb 21, 2012, at 5:44 PM, Chad Rosier <mcrosier at apple.com> wrote:

> 
> On Feb 21, 2012, at 5:38 PM, Jakob Stoklund Olesen wrote:
> 
>> 
>> On Feb 21, 2012, at 5:18 PM, Chad Rosier <mcrosier at apple.com> wrote:
>> 
>>> 
>>> On Feb 21, 2012, at 5:02 PM, Jakob Stoklund Olesen wrote:
>>> 
>>>> Author: stoklund
>>>> Date: Tue Feb 21 19:02:37 2012
>>>> New Revision: 151111
>>>> 
>>>> URL: http://llvm.org/viewvc/llvm-project?rev=151111&view=rev
>>>> Log:
>>>> Use SparseSet for the RAFast live virtual register map.
>>>> 
>>>> This makes RAFast 4% faster, and it gets rid of the dodgy DenseMap
>>>> iteration.
>>> 
>>> Very nice!!  What exactly were you testing to get the 4% estimate?
>> 
>> I have a -O0 bitcode file for 464.h264ref / armv7. I ran llc -time-passes and checked the user time for the fast allocator.
> 
> Ah, I was just wondering if this was a single test case or something larger like an aggregate number for the test-suite.

I was lazy because I expect the improvement to be the same for all test cases.

>> Note that RAFast is only 7% of codegen time at -O0, so the overall impact on clang -O0 is probably in the noise.
> 
> So you're telling me RAFast is fast.  :)  Every little bit counts!

If you make RAFast infinity faster, clang -O0 would be 2% faster. Diminishing returns… ;)

/jakob