[PING] set libfunc entry for sdivmod_optab to NULL in optabs.def

On 15 September 2016 at 17:53, Richard Biener
<richard.guenther@gmail.com> wrote:
> On Thu, Sep 15, 2016 at 1:21 PM, Prathamesh Kulkarni

> <prathamesh.kulkarni@linaro.org> wrote:

>> On 15 September 2016 at 16:31, Richard Sandiford

>> <richard.sandiford@arm.com> wrote:

>>> Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org> writes:

>>>> On 15 September 2016 at 04:21, Richard Sandiford

>>>> <rdsandiford@googlemail.com> wrote:

>>>>> Richard Sandiford <rdsandiford@googlemail.com> writes:

>>>>>> Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org> writes:

>>>>>>> Hi,

>>>>>>> I would like to ping the following patch:

>>>>>>> https://gcc.gnu.org/ml/gcc-patches/2016-08/msg01015.html

>>>>>>>

>>>>>>> While implementing divmod transform:

>>>>>>> https://gcc.gnu.org/ml/gcc-patches/2016-05/msg01757.html

>>>>>>>

>>>>>>> I ran into an  issue with optab_libfunc().

>>>>>>> It appears optab_libfunc (sdivmod_optab, mode) returns

>>>>>>> bogus libfunc for unsupported modes, for instance

>>>>>>> on x86_64, optab_libfunc (sdivmod_optab, DImode) returns

>>>>>>> a libfunc with name "__divmoddi4", even though such a libfunc

>>>>>>> does not exist in libgcc. This happens because in optabs.def

>>>>>>> the libfunc entry for sdivmod_optab has gen_int_libfunc,

>>>>>>> and call to optab_libfunc (sdivmo_optab, DImode) lazily

>>>>>>> creates a bogus libfunc "__divmoddi4" by calling gen_int_libfunc().

>>>>>>>

>>>>>>> To work around this issue I set libfunc entry for sdivmod_optab to NULL

>>>>>>> and verified that optab_libfunc (sdivmod_optab, DImode) returns NULL_RTX

>>>>>>> instead of a bogus libfunc if it's not overriden by the target.

>>>>>>>

>>>>>>> Bootstrapped and tested on ppc64le-linux-gnu, x86_64-linux-gnu.

>>>>>>> Cross tested on arm*-*-*, aarch64*-*-*.

>>>>>>> OK for trunk ?

>>>>>>

>>>>>> I'm not a maintainer for this area, but:

>>>>>

>>>>> ...in https://gcc.gnu.org/ml/gcc-patches/2016-05/msg01757.html

>>>>> you said that c6x follows the return-by-pointer convention.

>>>>> I'm no c6x expert, but from a quick look, I think its divrem

>>>>> function returns a div/mod pair in A4/A5, which matches the

>>>>> ARM convention of returning both results by value.

>>>>>

>>>>> Does anyone know if the optab function registered by the SPU

>>>>> backend is ever called directly?

>>>>>

>>>>> You mention that this is latent as far as expand_twoval_binop_libfunc

>>>>> is concerned.  AIUI expand_twoval_binop_libfunc implements the ARM/c6x

>>>>> convention and expects the two values to be returned as a pair.

>>>>> It then extracts one half of the pair and discards the other.

>>>>> So my worry is that we're leaving the udivmod entry intact even though

>>>>> the standard __udivmoddi4 doesn't do what expand_twoval_binop_libfunc

>>>>> expects.

>>>>>

>>>>> Would it make sense to set both entries to null and treat __udivmoddi4

>>>>> as a non-optab function?  ARM and c6x could then continue to register

>>>>> their current optab functions and a non-null optab function would

>>>>> indicate a return value pair.

>>>> AFAIU, there are only three targets (c6x, spu, arm) that override

>>>> optab_libfunc for udivmod_optab for following modes:

>>>> ./c6x/c6x.c:  set_optab_libfunc (udivmod_optab, SImode, "__c6xabi_divremu");

>>>> ./c6x/c6x.c:  set_optab_libfunc (udivmod_optab, DImode, "__c6xabi_divremull");

>>>> ./arm/arm.c:  set_optab_libfunc (udivmod_optab, DImode, "__aeabi_uldivmod");

>>>> ./arm/arm.c:  set_optab_libfunc (udivmod_optab, SImode, "__aeabi_uidivmod");

>>>> ./spu/spu.c:  set_optab_libfunc (udivmod_optab, DImode, "__udivmoddi4");

>>>> ./spu/spu.c:  set_optab_libfunc (udivmod_optab, TImode, "__udivmodti4");

>>>>

>>>> Out of these only the arm, and c6x have target-specific divmod libfuncs which

>>>> return <div, mod> pair, while spu merely makes it point to the

>>>> standard functions.

>>>>

>>>> So we could set libfunc entry for udivmod_optab to NULL, thus dropping

>>>> support for generic

>>>> divmod functions (__udivmoddi4, __udivmodti4). For targets that

>>>> require standard divmod libfuncs like __udivmoddi4,

>>>> they could explicitly override  optab_libfunc and set it to

>>>> __udivmoddi4, just as spu does.

>>>>

>>>> However this implies that non-null optab function doesn't necessarily

>>>> follow arm/c6x convention.

>>>> (i686-gcc for instance generates call to libgcc routines

>>>> __udivdi3/__umoddi3 for DImode division/mod operations

>>>> and could profit from divmod transform by calling __udivmoddi4).

>>>

>>> What I meant was that we shouldn't treat udivmoddi4 as an optab function

>>> at all, but handle it with some on-the-side mechanism.  That seems like

>>> quite a natural fit if we handle the fused div/mod operation as an

>>> internal function during gimple.

>> Ah right, thanks for pointing out. So if optab function for [us]divmod_optab

>> is defined, then it must follow the arm/c6x convention ?

>>>

>>> I think the current SPU code is wrong, but it looks like a latent bug.

>>> (Like I say, does the udivmodti4 function that it registers ever

>>> actually get used?  It seems unlikely.)

>>>

>>> In that scenario no other targets should do what SPU does.

>> I am testing the following patch which sets libfunc entries for both

>> sdivmod_optab, udivmod_optab to NULL.

>> This won't change the current (broken) behavior for SPU port since it

>> explicitly overrides optab_libfunc for udivmod_optab

>> and sets it to __udivmoddi4.

>

> Just

>

> -OPTAB_NL(sdivmod_optab, "divmod$a4", UNKNOWN, "divmod", '4', gen_int_libfunc)

> -OPTAB_NL(udivmod_optab, "udivmod$a4", UNKNOWN, "udivmod", '4', gen_int_libfunc)

> +OPTAB_NC(sdivmod_optab, "divmod$a4", UNKNOWN)

> +OPTAB_NC(udivmod_optab, "udivmod$a4", UNKNOWN)

>

> If I remember correctly.

Thanks for the pointers, I verified with the patch,
optab_libfunc (udivmod_optab, DImode) returns NULL for targets that
don't override
libfunc for udivmod_optab like x86_64, and returns correct libfunc for
targets that
override libfunc for udivmod_optab like arm.

The patch would make divmod transform more simpler, since now we require that
the divmod libfunc has a consistent interface identical to arm/c6x convention
(take 2 args and return value <div, mod> pair), so we can get rid of
the target hook
in divmod patch.

However I see one potential issue with SPU port for divmod transform:
So far, it appears the only function that could generate call to __udivmoddi4()
was expand_twoval_binop_libfunc() which was called from expand_divmod().
Maybe that code-path never got triggered for SPU and the issue remained latent.

However with divmod patch, we are introducing another code-path to generate
call to __udivmoddi4(), and since __udivmoddi4() and divmod libfunc conventions
don't agree, this will result in wrong code with the divmod transform on SPU.

We could work around this by SPU port not overriding optab_libfunc for
udivmod_optab,
by removing following lines in spu_init_libfuncs():
set_optab_libfunc (udivmod_optab, DImode, "__udivmoddi4");
set_optab_libfunc (udivmod_optab, TImode, "__udivmodti4");
If that's not desirable, we could disable the divmod transform for spu
for now and revisit the issue later.

The patch passes bootstrap+test on x86_64-unknown-linux-gnu,
aarch64-linux-gnu, ppc64le-linux-gnu, and cross tested on arm*-*-*,
aarch64*-*-*.
Would it be OK to commit ?
I am happy to do more testing if required and if something breaks due
to this patch,
I will revert it immediately.

Thanks,
Prathamesh
>

> Richard.

>

>

>> Thanks,

>> Prathamesh

>>>

>>> Thanks,

>>> Richard