On Mon, Apr 23, 2012 at 5:14 PM, Ulrich Weigand <uweig...@de.ibm.com> wrote:

>> > 2011-11-17  Andreas Krebbel  <andreas.kreb...@de.ibm.com>
>> >
>> >     * reload.c (find_reloads): Change the loop nesting when trying an
>> >     alternative with swapped operands.
>>
>> I would just like to point out that constran_operands will have
>> problems to re-recognize alternative with swapped commutative
>> operands. The "%" operand modifier is ignored there, so i.e. x86 add
>> pattern with swapped commutative operands, like:
>>
>> (set (reg: ax) (plus (reg: bx)(reg: ax)))
>>
>> will be rejected.
>
> Well, yes, that's why reload explicitly tries both versions, and
> keeps the one that matches ...
>
> The only thing Andreas' patch changes is the priority in which
> different options are tried if there are multiple alternatives.
>
> For example, if you have a pattern that has commutative operands,
> and one of them supports both register and memory, reload can
> try:
>
>  1.) use operands as they are
>  2.) spill one operand to memory
>  3.) swap operands
>  4.) swap operands and spill one operand to memory
>
> It used to be that reload tried these in order 1, 2, 3, 4;
> with Andreas' patch they are now tried in order 1, 3, 2, 4.
> (Always assuming there are no priorities like ! or ? specified
> in the .md file.  Those would override in either case.)
>
>
> Could you explain in more detail the problem you see with that?

I don't have any problems with proposed order, but in my failed
attempt to convert x86 to post-reload compare elimination,
constrain_operands failed to recognize generated combined add+compare
pattern, when commutative matched operands were swapped (e.g. the
perfectly valid x86 add pattern above). constrain_operands was called
from the call to validate_change (please see compare-elim.c) when
flags clobber was substituted with compare.

Since the order of how commutative operands are handled will be
changed, I thought that I should mention this issue with matched
commutative operands on the mailing list.

Uros.

Reply via email to