comments on loop changes

Jeffrey A Law
Sun Feb 28 18:15:00 GMT 1999

  In message < >you write:
  > On Sun, Feb 14, 1999 at 09:25:40PM +1300, Michael Hayes wrote:
  > > The following hack makes all the difference for me.  With this, all my
  > > test routines generate good code.  Maybe the test should be more
  > > selective but I'm only tinkering and do not understand the heuristics
  > > employed by GIV combination.
  > The problem is that on all machines I have regular contact with,
  > all addressing modes cost the same execution-wise.
Yes.  sparc, pa, mips and most modern risc iron.

  > So to avoid
  > bad code on those machines, some artificial costing would have 
  > to be added just for the benefit of loop.  Which seems too gross
  > for words.
Don't we already fudge things in ADDRESS_COST on the effected targets?
Specifically to slightly discourage reg+reg to avoid keeping extra regs alive.

I've always thought that was a little hokey though.


More information about the Gcc-patches mailing list