Re: [PATCH] Optimize manual byte swap implementations v3

[Andreas Krebbel <krebbel at linux dot vnet dot ibm dot com>]

Mark Mitchell wrote:
> Andreas Krebbel wrote:
>
> > The timings look much better now:
> > 8.05s			8.097s		0.58%	
>
> Certainly better, but still significant.

These are only the costs for walking over the ssa statements looking for 
OR operations plus a bit more work if an OR has been found. With the 
tweaks I've added after the first reviews I don't think this can be done 
significantly faster. Please also consider that this might also be the 
base for further optimizations on vectors as Richard pointed out. For 
the vector optimizations the stmt walk is needed as well so in the end 
the added compile time overhead will just be the check if it is an OR 
operation.

> What is the performance improvement provided by this pass on what
> benchmarks on what CPUs?  You may have indicated this elsewhere, in
> which case I apologize for asking a redundant question; please just
> point me at the URL.

Unfortunately I'm not able to come up with performance numbers. I agree 
that this is a special purpose optimization but byte swaps occur much 
more often than one might think at first glance. A quick scan reveals 
that the Linux Kernel uses over 3000 bswaps (sha*, wpa (wlan), ...) with 
the default config on x86_64, libc over 200 and openssh about 10. Ok, 
the optimization doesn't help x86 for these packages since they already 
 make heavy use of inline assemblies to implement bswaps.

But how would you recommend an application developer should implement a 
byte swap which is neither bound to GCC (even to a specific version) by 
using the bswap builtins nor bound to GCC and a specific CPU by using 
inline assemblies? I think the only way is to use a bunch of ORs, ANDs 
and SHIFTs. With the bswap pass you can use such a compiler- and 
platform-independent implementation without losing performance on Linux 
with GCC and an architecture which is able to do it better.

Another point is other programming languages.  The bswap optimizer is 
pretty well able to optimize bswaps implemented in Fortran, Java and 
others.  The Integer.reverseBytes method in the Java library is already 
detected and with the bswap builtins enabled for Java the library can 
benefit from hw instructions.

I know this is no replacement for tough performance numbers so I'll keep 
looking for an appropriate benchmark or application.

Bye,

-Andreas-