[llvm-commits] Speeding up RegAllocLinearScan on big test-cases

[Chris Lattner]

On May 19, 2008, at 3:13 AM, Roman Levenstein wrote:

>> If your assumption holds, I should see that STL's standard  
>> allocator would perform much  worse than before.
>
> Chris, you're absolutely right!!!
>
> The standard STL allocators is now much slower after a special prepass
> creates artificially a lot of fragmented memory. When I redo the same
> tests that I did before, the performance figures for STl containers
> using standard allocators are about 2 times worse than for all other
> allocators(whose performance is unaffected by the heap fragmentation).
> I can also nicely see that the performance of containers using the
> standard STL allocator constantly degrades as memory fragmentation
> increases.

Yep, that makes logical sense.  The major reason to do per-container  
allocation is locality :)

-Chris