Matching Constraints and Predicate

Matching Constraints and Predicate

[Paulo J. Matos]

Hello,

I am quite confused with the following from the internals:
; Operand predicates can allow operands that are not actually
; acceptable to the hardware, as long as the constraints give
; reload the ability to fix them up (see Constraints). However,
; GCC will usually generate better code if the predicates specify
; the requirements of the machine instructions as closely as
; possible. Reload cannot fix up operands that must be
; constants (immediate operands); you must use a predicate that
; allows only constants, or else enforce the requirement in the
; extra condition.

This seems to imply that if the predicates specify the same
requirements of constraints, then resulting code if better.

However, if I have register constraints that define constraint x to
match a certain register class X and I define predicate:
(define_predicate "x_operand"
    (and (match_operand 0 "register_operand")
             (match_test "REGNO_REG_CLASS(REGNO(op)) == X")))

and then I have a match_operand on an insn (match_operand 0
"x_operand" "x") this is quite disastrous as gcc promptly stops since
any pseudo-register during will fail the predicate and no rule will
not be matched.
It feels like I should say in the predicate something like: it's
either a pseudo, or a register from X.

What am I misinterpreting?

Cheers,

-- 
PMatos

Re: Matching Constraints and Predicate

[Ian Lance Taylor]

"Paulo J. Matos" <pocmatos@gmail.com> writes:

> However, if I have register constraints that define constraint x to
> match a certain register class X and I define predicate:
> (define_predicate "x_operand"
>     (and (match_operand 0 "register_operand")
>              (match_test "REGNO_REG_CLASS(REGNO(op)) == X")))

As you observe, this is wrong.  A pseudo-register has no register class.
You can only test the register class if you first test that you have a
hard register.

This also makes no sense.  The operand predicate should test whether the
value is acceptable.  The constraints should tell the register allocator
where the value should go.  The paragraph you quoted is saying that the
operand predicate should restrict the value as much as possible to
ensure that the constraints will match.  The paragraph is not saying
that you should test the register class of a register in the operand
predicate.

Ian

Re: Matching Constraints and Predicate

[Paulo J. Matos]

On Mon, Aug 23, 2010 at 6:43 PM, Ian Lance Taylor <iant@google.com> wrote:
>
> This also makes no sense.  The operand predicate should test whether the
> value is acceptable.  The constraints should tell the register allocator
> where the value should go.  The paragraph you quoted is saying that the
> operand predicate should restrict the value as much as possible to
> ensure that the constraints will match.  The paragraph is not saying
> that you should test the register class of a register in the operand
> predicate.
>

Thanks for clearing this up.

-- 
PMatos