[Patch, libgfortran] PR21468 Vectorizing matmul, other perf improvements

[Tim Prince]

Thomas Koenig wrote:
> On Sun, Nov 13, 2005 at 05:52:37PM +0200, Janne Blomqvist wrote:
> 

> 
>>Another thing that it does is enabling -funroll-loops for matmul_* and
>>matmull_*. This is actually a bigger performance booster than
>>vectorizing, almost doubling performance.
> 
> 
> I fear that this may be platform-dependant.  Loop unrolling may
> be a big win for architectures with many registers, while on
> architectures with few registers, such as i686, it may cause
> performance degradation.
> 
I doubt it would be worth while to try to discriminate between ancient 
CPUs which lack hardware register renaming and those of the last 6 years 
which have that feature.  CPUs like P-III and later do very well in 
pipelining code which assigns independent data to the same named register.
For CPUs with many registers but no hardware renaming, like Itanium, it 
seems unpredictable whether gcc -frename-registers will accomplish the 
desired optimization.