[PATCH] [RFT] Improve estimate_num_insns

[Mark Mitchell]

Richard Guenther wrote:
> This treating of CALL_EXPRs reduces the compile time impact of the
> patch for the tramp3d-v3 testcase from a regression of 135% to
> a regression of 35% while still reducing the performance regression
> from 3.4 to 18% from 350%.  More tuning of inline-unit-growth and
> the other parameters will possibly give even better numbers.

If that kind of improvement holds up in other tests, I'd say this seems 
worthwhile.  I think most people who are compiling with optimization 
would pay the 35% compile-time penalty to get the performance 
improvement you describe.  Ideally, the penalty would be smaller, but 
this is still a win.

-- 
Mark Mitchell
CodeSourcery, LLC
mark@codesourcery.com
(916) 791-8304