Re: Another problem with PIC port, too many reloads for insn

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

Robert Baruch wrote:
> find_reloads for the insn computes five reloads:
> ...

Try "print debug_reload()".  This is a much easier way to print this info.

> Later on, in find_reload_regs for this instruction, the reload order
> is determined to be 0, 4, 1, 2, 3. Reload 0 snags the W register, and
> of course reload 4 fails and causes GCC to abort, complaining that it
> cannot find a spill register because the W register required for
> reload 4 is already used.

You need more registers to make gcc work.  My recollection is that you 
need at least 3 registers in order to make reload work, but it has been 
a while since I've worked with a register limited target.

A common trick for microcontrollers is to use part of memory as a 
register set.  For instance, targets that can easily access zero page 
memory can define a set of 8 registers in zero page memory.  And once 
you have 8 registers, reload will work fine.  Since you listed constant 
as a valid address, this should work for you.

The code you get will be terrible, as it will be full of moves to/from 
zero page memory "registers", but at least it will work.

You said that you deliberately left out reg+offset addressing modes. 
You will get fewer reloads if you allow that, and that may be enough to 
get reload working.  You might need to allow it even if it gives poor 
code, as poor code is better than no code.
--
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com