Re: [tree-ssa] Speeding things up

[law at redhat dot com]

In message <4BDFD7C5-33EC-11D7-8309-003065A77310@apple.com>, Mike Stump writes:
 >On Wednesday, January 29, 2003, at 04:14 PM, law@redhat.com wrote:
 >>   In all we're talking about over 2.6 billion function calls (and this 
 >> is
 >>   *after* the algorithmic improvements).  For reference we really
 >>   need about 100k calls total.
 >
 >Do you have timings for this?  13.98s -> 13.97s might not be worth 
 >making the code less clean, but 13.98s -> 5.32s I think is worthwhile.  
 >In some of my benchmarking, I've discovered that 100% of the code in 
 >entire files is not worth speeding up, as 100% of the time is spent on 
 >one line.
A rough implementation brought it from 13 minutes down to 6.  That
implementation hasn't been tested thoroughly yet as I've been working
to test the stuff that takes us from 30 minutes to 13 :-)


 >I think we have to be willing to make the code slightly less clean for 
 >the right performance enhancement.  Ideally, if people can find ways to 
 >get us the performance, without making the code any less readable, 
 >clean, maintainable, that would be great, but, failing that, we should 
 >entertain changes like this.  It all comes down to, how much worse is 
 >the code, and what is the benefit gained.
Actually, it doesn't IMHO make this code any less clean.

jeff