Re: fix 20650

[Geoffrey Keating <geoffk at apple dot com>]

On 02/04/2005, at 11:04 AM, David Edelsohn wrote:

> -  [(parallel [(set (match_operand:SI 0 "reg_or_mem_operand" "")
> +  [(parallel [(set (match_operand:SI 0 "gpc_reg_operand" "")
>
> 	I created a special e500 reg_or_mem_operand predicate for
> expandsfdf2.  You could do something similar for fix_truncdfsi2 with
> GFXOPT and try rth's SUBREG suggestion, if you want.

The problem I have is that if I make the expander only take memory 
operands (forgetting about the non-GFXOPT case for now), then optabs 
fails, because it assumes that every instruction can take a register; 
but if I make the expander take registers, then either I have to make 
temporary memory locations in the expander for all pseudos, which can't 
be removed by any later pass, or reload fails, apparently because it 
doesn't know how to put a pseudo into memory.

There must be other ports that have instructions that can operate only 
on memory; does anyone know an example?

Attachment: smime.p7s
Description: S/MIME cryptographic signature