Split edge fails when performed on a doloop jump on x86 port

[Jim Wilson]

Mostafa Hagog wrote:
>             (set (pc)
>                 (if_then_else (ne (reg:SI 0 ax)
>                         (const_int 1 [0x1]))
>                     (label_ref 419)
>                     (pc)))
>             (set (mem:SI (plus:SI (reg/f:SI 7 sp)
>                         (const_int 4 [0x4])) [12 S4 A8])
>                 (plus:SI (reg:SI 0 ax)
>                     (const_int -1 [0xffffffff])))
>             (clobber (reg:SI 0 ax))
>             (clobber (reg:CC 17 flags))
> 
> doloop_end_internal pattern (from i386.md ) doesn't match this jump because
> of the (clobber (reg:SI 0 ax)) which is as a result of reload (I suppose).

I don't see a problem with the clobber.  A hard register will match a 
match_scratch.

I believe the problem is that operand 2 is a MEM, but the pattern uses 
register_operand.  There is an inconsistency in the pattern here, the 
cosntraints allow operands that the predicates don't.
     (match_operand:SI 2 "register_operand" "=1,1,*m*r")
The MEM matches the 'm' constraint, but not the register operand 
predicate.  This is not good.  The pattern should use 
nonimmediate_operand here instead of register_operand.

For optimization purposes, we perhaps want to add a condition that 
rejects a MEM before reload.  Maybe something like
   "reload_in_progress || reload_completed
    || register_operand (operands[2], SImode)"
We want to do something like this because the MEM case results in less 
efficient code than the REG case.
-- 
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com