insn-attrtab.c on diet

[Jan Hubicka]

> On Mon, Aug 28, 2000 at 01:41:02AM +0200, Jan Hubicka wrote:
> > This patch reduces code of insn-attrtab from 1300Kb to roughly 400Kb (170Kb
> > when compiled) by avoiding some inlining.  The attr_rtx_cost is introduced
> > to limit cost of the inlined exression.  While this introduces somoe
> > ineffectivity to the attribtutes, with my previous patch I've measured no
> > slowdowns (in fact small speedups) in the cc1 ...
> 
> Hmm.  Perhaps the cost needs tuning, as it expands the code size on
> Alpha by about 3%.  There are changes like
> 
> I would suggest to consider this as a factoring problem.  Especially for
> result_ready_cost, in which we have lots of expressions
> like
Hmm tunning cost seems to be trickier than I've expected.  Problem is that
real cost comes from the fact, that inlining will make expression different
for more insns.  I can't compute real cost of that, since it is not known yet
for which insns the expressions match at the moment.

Perhaps we should process by simplifying all expression first, finding the
matches and then try to inline, but I would rather try something easier.
Disabling which_alternative cost should be enought for alpha testcase.
> 
> If we began this function with 
> 
>   switch (ix86_cpu)
>     {
>     case CPU_K6:
>       ...
>     }
> 
> I think that the code size would be much reduced.
Result_ready_cost don't seems to bring major code bloats tought..
In the other functions we don't surfer much similar problems.
I will try that.

Honza
> 
> 
> 
> r~