optimization bug

[Jeffrey A Law]

  In message < Pine.OSF.3.96.970922001305.6099A-100000@hepunix1.physics.yale.edu >you write:
  > The problem shows up on an alpha-dec-osf4.0 with haifa schedule.
OK.  You need to make sure to let us know this kind of information
when you make the initial bug report.  It'll keep everyone from wasting
time.

  > The RTL files t1.c.sched and t1.c.lreg are in the appendix.
  > 
  > Note the INSN in t1.c.sched
  > ======>(insn 88 4 15 (use (reg:DI 88)) -1 (nil)
  > is changed in t1.c.lreg to 
  > ======>(insn 88 4 15 (use (const:DI (plus:DI (symbol_ref:DI ("pure"))
  > which should not have happened.
  > 
  > With this change, pseudo-reg 88 is never set, but it is used later which
  > cause the program crash.
OK.  This is also something you need to tell us when you send patches;
otherwise we have to figure these things out ourselves to determine
if your patch is correct.

Your patch is probably correct, though I want Jim's opinion on
one aspect.

Jim -- the problem is combine creates a use, but leaves reg_n_refs
as 2.

local-alloc handles reg_n_refs == 2 specially if the register is
equivalent to a constant (tries to replace references to the register
with the corresponding constant expression).  In thise case there's
actually two insns that reference the reg as a source operand and
local-alloc substituted the constant expression in the wrong one
(ie the USE insn).

Should register references in USE insns be counted in reg_n_refs?

jeff