Re: atomic memory operation builtins, ia64

[Geoff Keating <geoffk at geoffk dot org>]

On 18/04/2005, at 8:42 AM, Richard Henderson wrote:

> On Sun, Apr 17, 2005 at 08:45:54PM -0700, Geoffrey Keating wrote:
> > Is this really supposed to be generating
> >
> > (set (mem:BLK (mem:BLK (scratch))) ...)
>
> Um... no?

I was wondering if you were doing something tricky.  You have

  [(set (mem:BLK (match_dup 0))

but then you write

  operands[0] = gen_rtx_MEM (BLKmode, gen_rtx_SCRATCH (DImode));

which means that the (match_dup 0) gets substituted with (mem:BLK 
(scratch)) giving you two MEMs.

For PPC, I'm writing

(define_insn "memory_barrier"
  [(set (mem:BLK (match_scratch 0 "X"))
        (unspec:BLK [(mem:BLK (match_scratch 1 "X"))] UNSPEC_SYNC))]
  "TARGET_POWERPC"
  "sync")

which seems to be pretty effective, and saves a define_expand.

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