LLVM vs stock GCC

[Mike Swingler]

The core functionality that Henri's test are evaluating are effectively complete. I would speculate that the minor drop in performance is likely the combination of forcing stack pointer alignment to 16-byte boundaries, and using -Os instead of -O3, as opposed to any systemic issue with LLVM.

It would be interesting to find the combination of options that cause xalan, lusearch, and eclipse to fall behind the other OpenJDK variants. If there is an issue with using the LLVM backend, a bug should be filed against it ASAP (of course comparing -Os to -Os or -O3 to -O3, and ensuring all other factors are equal).

Regards,
Mike Swingler
Java Engineering
Apple Inc.

On Apr 18, 2011, at 1:21 PM, Gary Meyer wrote:

> There's nothing wrong with Henri's tests.  making any conclusions about OpenJDK-macosx-port isn't really fair since it isn't anywhere near finished (or even functional, really).  We have to get it to work before we will make it fast.
> Gary.
> =======================
> Gary Meyer
> Java Software Build Engineer
> gary.meyer at apple.com
> 
> On Apr 18, 2011, at 1:16 PM, Julien Ponge wrote:
> 
>> What's wrong in Henri's tests?
>> 
>> -- 
>> Julien Ponge
>> 
>> On lundi 18 avril 2011 at 22:03, Gary Meyer wrote:
>> 
>>> no, llvm is actually faster … to a certain extent you are comparing apple to oranges
>>> =======================
>>> Gary Meyer
>>> Java Software Build Engineer
>>> gary.meyer at apple.com
>>> 
>>> On Apr 18, 2011, at 12:46 PM, Henri Gomez wrote:
>>> 
>>>> Hi to all,
>>>> 
>>>> I do some benchmarks with various JVMs available on OS/X
>>>> (http://blog.hgomez.net/2011/04/18/apple-jdk’s-openjdks-–-core2duo-vs-core-i7/)
>>>> 
>>>> Apple VM (1.6.0-22/1.6.0-24), OpenJDK 6, OpenJDK 7 from bsd-port and
>>>> OpenJDK 7 from macosx-port.
>>>> 
>>>> I notice a small performance loss in macosx-port version, which is
>>>> similar to bsd-port code base, except it's using LLVM.
>>>> 
>>>> Did there is a penalty using LLVM instead of stock GCC ?
>>>> 
>>>> Thanks
>> 
>