[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: AMD64 64bit vs 32bit speed increase

>>>>> "Scott" == C Scott Ananian <cananian@lesser-magoo.csail.mit.edu> writes:

Scott> OTOH, ogg encoding is heavily floating-point dependent, ...

That is likely to be a large chunk of the perfomance benefit, given
that the amd64 abi uses the xmm registers for floating point rather
than the x87 stack based fp stuff.  Gcc can do a much better job w/
the register-based fp ops, and the args are passed in registers, so
fewer mem ops are required for each function call.

But, I bet were one to test a -m32 -march=k8 compile it would also
be faster than the generic i386 compile in the debian binary tree.

Things will get even better when gcc can parallelize loops itself
in the optimization phase, as icc already does.  Then the full
benefit of the xmm registers and SSE{,2} instruction sets will be
evident even for code that was not manually parallelized.


Reply to: