Re: ia64 branch register opts

[David Mosberger <davidm at napali dot hpl dot hp dot com>]

>>>>> On Fri, 25 Jul 2003 08:46:59 -0700, Richard Henderson <rth@twiddle.net> said:

  Rich> As discussed at OLS, wrt optimizing indirect function calls.
  Rich> Will you try this and see if it provides enough improvement
  Rich> to warrent cleaning up properly?

Well, the initial results are encouraging!  It's generating exactly
the code I was hoping for!  Here is an example:

 $ cat t1.c
 int r;

 int foo (int count, void (*fp0) (void), void (*fp1) (void)) {
   int i;

   for (i = 0; i < count; ++i) {
      (*fp0)();
      (*fp1)();
   }
   return r;
 }
 $ cc1 -O2 t1.c
 $ cat t1.s
 foo:
        alloc r36 = ar.pfs, 3, 6, 0, 0
        mov r38 = ar.lc
        cmp4.ge p6, p7 = 0, r32
        mov r37 = r1
        sxt4 r14 = r32
        mov r39 = b1 ;;
        adds r14 = -1, r14
        mov r40 = b2
        mov r35 = b0
   (p6) br.cond.dpnt .L7 ;;
        mov ar.lc = r14
        ld8 r15 = [r33], 8 ;;
        mov b2 = r15
        ld8 r15 = [r34], 8
        ld8 r33 = [r33] ;;
        mov b1 = r15
        ld8 r34 = [r34]
 .L8:   mov r1 = r33
        br.call.sptk.many b0 = b2 ;;
        mov r1 = r34
        br.call.sptk.many b0 = b1 ;;
        br.cloop.sptk.few .L8
        mov r1 = r37 ;;
 .L7:   addl r14 = @ltoff(r#), r1
        mov b0 = r35 ;;
        ld8 r14 = [r14]
        mov b1 = r39
        mov b2 = r40 ;;
        ld4 r8 = [r14]
        mov ar.pfs = r36
        mov ar.lc = r38
        br.ret.sptk.many b0

When passing empty (null) functions to fp0 and fp1, the new code
is 58% faster on McKinley.  Even on Merced (which has a more
complicated indirect branch predictor) the new code is 16% faster.

I'll try it now on the Linux kernel.

Thanks,

	--david