Re: DFA for PPro, P2, P3

[Richard Henderson <rth at redhat dot com>]

On Thu, May 02, 2002 at 04:55:46PM -0600, law@redhat.com wrote:
>   3. There is a performance problem in building some of the tables in
>      genautomata/genattrtab that is preventing me from modeling 
>      fdiv/fsqrt like the old description.  Vlad is working on fixing
>      the performance issue with an algorithmic change in how the
>      particular table is built.

Incidentally, I ran into this same problem on Alpha.

Originally I put all the units (for a given cpu) in the same automata.
This was of the unacceptable go-out-for-dinner-come-back-and-kill-it
sort of build time.

Then I followed advice seen elsewhere about putting all of the
long-latency units in a separate automata.  This helped, bringing
the build time down to ~3 minutes.

Today I put each pair of long-latency units in multiple separate
automata.  For instance, given 

(define_insn_reservation "ev4_imulsi" 21
  (and (eq_attr "cpu" "ev4")
       (and (eq_attr "type" "imul")
            (eq_attr "opsize" "si")))
  "ev4_ib0+ev4_imul,ev4_imul*18,ev4_ebox")

(define_insn_reservation "ev4_fdivsf" 34
  (and (eq_attr "cpu" "ev4")
       (and (eq_attr "type" "fdiv")
            (eq_attr "opsize" "si")))
  "ev4_ib1+ev4_fdiv,ev4_fdiv*28,ev4_fdiv+ev4_fbox,ev4_fbox")

then ev4_imul and ev4_ebox are related long-latency units, but are
placed into a different automata than ev4_fdiv and ev4_fbox.  I.e.

(define_cpu_unit "ev4_ebox,ev4_imul" "ev4_1")
(define_cpu_unit "ev4_fbox,ev4_fdiv" "ev4_2")

This reduced build time to 3 seconds, and insn-attrtab.c is 300k
rather than 1.8MB.


r~