This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: ifcvt conditional execution and rtx_cost


On Wed, 7 Jul 2004, Uros Bizjak wrote:
> The problem here is, that complex FP instruction execution times depend
> on their parameter value.
> Reusing fsqrt cost is indeed a bad idea, because we hide the fact that
> we actually want an arbitrary high number to prevent ceratin
> "optimizations". Perhaps value of 1024 would be a good number?

For my work on rs6000_rtx_costs I considered adding something like

#define COST_UNKNOWN   1024

but decided to cautiously preserve the current behaviour whenever
I didn't know what the true instruction latency should be.  This
is the reason why I didn't provide rs6000 values for "fsqrt" as
not all CPUs implement it, and which value to list for those CPUs
that don't is less clear.


However, for the x86 trigonmetric and complex floating point
operations I'd propose that we use the median value between
the upper and lower bounds.  Without any extra information on
the distribution of cycle times, (upper+lower)/2 seems like
as good an approximation as any.

Roger
--


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]