Re: computation_cost too blind to hard regs

[James E Wilson <wilson at specifixinc dot com>]

DJ Delorie wrote:
> Am I missing something obvious here?  We're creating a temporary
> sequence only to compute its cost, but we're using HARD registers in
> ways that they might not support (in my case, my backend aborted due
> to a subreg of a register in a mode it didn't support).  In my case,
> using FIRST_PSEUDO_REGISTER intead of 0 as the initial regno works
> better.

There is similar code in init_expmed in expmed.c that uses a register 
number of 10000.  The number is arbitrary, but high enough to avoid 
matching hard registers.  We could use FIRST_PSEUDO_REGISTER here and it 
would probably work just as well.  In this case, though, we avoid 
generating any RTL.  We create just enough fake RTL to make rtx_cost happy.

In the computation_cost case, a pseudo-reg number is likewise better 
than a hard reg number, but this one creates RTL.  So we need to make 
sure we don't accidentally create a pseudo-reg conflict, but since 
gen_raw_REG is used here, it seems safe to use any pseudo register number.
--
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com