[PATCH v2, rtl-optimization]: Fix PR54457, [x32] Fail to combine 64bit index + constant

[Uros Bizjak <ubizjak at gmail dot com>]

On Thu, Sep 27, 2012 at 4:25 PM, Richard Sandiford
<rdsandiford@googlemail.com> wrote:

>>>> I agree (subreg:M (op:N A C) 0) to (op:M (subreg:N (A 0)) C) is
>>>> a good transformation, but why do we need to handle as special
>>>> the case where the subreg is itself the operand of a plus or minus?
>>>> I think it should happen regardless of where the subreg occurs.
>>>
>>> Don't we need to restrict this to the low part though?
>>
>> I have tried this approach with attached patch.  Unfortunately,
>> although it survived bootstrap without libjava on x86_64, it failed
>> building libjava with:
>>
>> /home/uros/gcc-svn/trunk/libjava/classpath/javax/swing/plaf/basic/BasicSliderUI.java:1299:0:
>> error: insn does not satisfy its constraints:
>>    }
>>  ^
>> (insn 237 398 399 7 (set (reg:SI 1 dx [125])
>>         (plus:SI (subreg:SI (mult:DI (reg:DI 1 dx [orig:72 D.78627 ] [72])
>>                     (const_int 2 [0x2])) 0)
>>             (reg:SI 5 di)))
>> /home/uros/gcc-svn/trunk/libjava/classpath/javax/swing/plaf/basic/BasicSliderUI.java:1271
>> 240 {*leasi}
>>      (expr_list:REG_DEAD (reg:DI 5 di)
>>         (nil)))
>>
>> Original RTX was (subreg:SI (plus:DI (mult:DI (...) reg:DI))), which
>> is valid RTX pattern for lea insn, the above is not.
>>
>> Due to these problems, I think the safer approach is to limit the
>> transformation to (plus:SI (subreg:SI (plus:DI (...) 0)) RTXes, as was
>> the case with original patch. This approach would fix a specific
>> problem where simplify_plus_minus is not able to simplify the combined
>> RTX at combine time. Please note, that combined RTXes are always
>> checked for correctness at combine pass.
>
> I think instead the (subreg (plus ...)) handling should be applied
> to (subreg (mult ...)) too.  IMO the correct form of the above address
> ought to be:
>
>     (set (reg:SI 1 dx [125])
>          (plus:SI (mult:SI (reg:SI 1 dx [orig:72 D.78627 ] [72])
>                            (const_int 2 [0x2]))
>                   (reg:SI 5 di))

Great, this works as expected!

After some off-line discussion with Richard, attached is v2 of the patch.

2012-09-27  Uros Bizjak  <ubizjak@gmail.com>

        PR rtl-optimization/54457
        * simplify-rtx.c (simplify_subreg):
	Simplify (subreg:SI (op:DI ((x:DI) (y:DI)), 0)
     	to (op:SI (subreg:SI (x:DI) 0) (subreg:SI (x:DI) 0)).

testsuite/ChangeLog:

2012-09-27  Uros Bizjak  <ubizjak@gmail.com>

        PR rtl-optimization/54457
        * gcc.target/i386/pr54457.c: New test.

Patch was bootstrapped and regression tested on x86_64-pc-linux-gnu {,-m32}.

BTW: I propose that we start with limited selection of opcodes, so x32
autotester will pick and test the patch with SImode addresses.

OK for mainline?

Uros.

Attachment: p.diff.txt
Description: Text document