Re: 100% speed improvement 3.4 -> 3.5 for ia64!! (?)

[Jim Wilson <wilson at specifixinc dot com>]

Richard Guenther wrote:
> Using -minline-float-divide-min-latency -minline-int-divide-min-latency
> doesn't change performance at all.  Either all these options are not
> working as expected, or there are more differences between 3.4 and 3.5 in
> this area.

These options will have a small effect on most programs.  The latency 
difference between the latency-optimized and throughput-optimized 
sequences are small compared to the entire latency of the divide 
operation.  So you probably won't notice any difference unless you have 
divides on the critical path inside an inner loop.

Meanwhile, keep in mind that the latency-optimized sequences use more 
hardware resources (i.e. more instructions) than the 
throughput-optimized sequences.  So if you have other operations that 
can be performed in parallel with the divides, then use of the 
latency-optimized sequences can actually make the program run slower.

If you aren't seeing any performance difference, then probably the gains 
you get from shorter divide latencies is canceled out by the losses you 
get from the longer instruction sequences.

It is trivial to see whether the options work by compiling a one line 
function that contains a divide, and see if you get different results.
--
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com