Re: [ARM] PR98435: Missed optimization in expanding vector constructor

[Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org>]

On Thu, 5 Aug 2021 at 18:05, Christophe Lyon
<christophe.lyon.oss@gmail.com> wrote:
>
>
>
> On Thu, Aug 5, 2021 at 2:28 PM Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org> wrote:
>>
>> On Tue, 3 Aug 2021 at 20:52, Christophe Lyon
>> <christophe.lyon.oss@gmail.com> wrote:
>> >
>> >
>> >
>> > On Tue, Aug 3, 2021 at 12:57 PM Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org> wrote:
>> >>
>> >> On Tue, 3 Aug 2021 at 14:59, Christophe Lyon
>> >> <christophe.lyon.oss@gmail.com> wrote:
>> >> >
>> >> >
>> >> >
>> >> > On Tue, Jul 6, 2021 at 11:26 AM Prathamesh Kulkarni via Gcc-patches <gcc-patches@gcc.gnu.org> wrote:
>> >> >>
>> >> >> On Tue, 6 Jul 2021 at 13:33, Kyrylo Tkachov <Kyrylo.Tkachov@arm.com> wrote:
>> >> >> >
>> >> >> >
>> >> >> >
>> >> >> > > -----Original Message-----
>> >> >> > > From: Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org>
>> >> >> > > Sent: 06 July 2021 08:06
>> >> >> > > To: Christophe LYON <christophe.lyon@foss.st.com>
>> >> >> > > Cc: Kyrylo Tkachov <Kyrylo.Tkachov@arm.com>; gcc Patches <gcc-
>> >> >> > > patches@gcc.gnu.org>
>> >> >> > > Subject: Re: [ARM] PR98435: Missed optimization in expanding vector
>> >> >> > > constructor
>> >> >> > >
>> >> >> > > On Thu, 1 Jul 2021 at 16:26, Prathamesh Kulkarni
>> >> >> > > <prathamesh.kulkarni@linaro.org> wrote:
>> >> >> > > >
>> >> >> > > > On Wed, 30 Jun 2021 at 20:51, Christophe LYON
>> >> >> > > > <christophe.lyon@foss.st.com> wrote:
>> >> >> > > > >
>> >> >> > > > >
>> >> >> > > > > On 29/06/2021 12:46, Prathamesh Kulkarni wrote:
>> >> >> > > > > > On Mon, 28 Jun 2021 at 14:48, Christophe LYON
>> >> >> > > > > > <christophe.lyon@foss.st.com> wrote:
>> >> >> > > > > >>
>> >> >> > > > > >> On 28/06/2021 10:40, Kyrylo Tkachov via Gcc-patches wrote:
>> >> >> > > > > >>>> -----Original Message-----
>> >> >> > > > > >>>> From: Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org>
>> >> >> > > > > >>>> Sent: 28 June 2021 09:38
>> >> >> > > > > >>>> To: Kyrylo Tkachov <Kyrylo.Tkachov@arm.com>
>> >> >> > > > > >>>> Cc: Christophe Lyon <christophe.lyon@linaro.org>; gcc Patches
>> >> >> > > <gcc-
>> >> >> > > > > >>>> patches@gcc.gnu.org>
>> >> >> > > > > >>>> Subject: Re: [ARM] PR98435: Missed optimization in expanding
>> >> >> > > vector
>> >> >> > > > > >>>> constructor
>> >> >> > > > > >>>>
>> >> >> > > > > >>>> On Thu, 24 Jun 2021 at 22:01, Kyrylo Tkachov
>> >> >> > > <Kyrylo.Tkachov@arm.com>
>> >> >> > > > > >>>> wrote:
>> >> >> > > > > >>>>>
>> >> >> > > > > >>>>>> -----Original Message-----
>> >> >> > > > > >>>>>> From: Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org>
>> >> >> > > > > >>>>>> Sent: 14 June 2021 09:02
>> >> >> > > > > >>>>>> To: Christophe Lyon <christophe.lyon@linaro.org>
>> >> >> > > > > >>>>>> Cc: gcc Patches <gcc-patches@gcc.gnu.org>; Kyrylo Tkachov
>> >> >> > > > > >>>>>> <Kyrylo.Tkachov@arm.com>
>> >> >> > > > > >>>>>> Subject: Re: [ARM] PR98435: Missed optimization in expanding
>> >> >> > > vector
>> >> >> > > > > >>>>>> constructor
>> >> >> > > > > >>>>>>
>> >> >> > > > > >>>>>> On Wed, 9 Jun 2021 at 15:58, Prathamesh Kulkarni
>> >> >> > > > > >>>>>> <prathamesh.kulkarni@linaro.org> wrote:
>> >> >> > > > > >>>>>>> On Fri, 4 Jun 2021 at 13:15, Christophe Lyon
>> >> >> > > > > >>>> <christophe.lyon@linaro.org>
>> >> >> > > > > >>>>>> wrote:
>> >> >> > > > > >>>>>>>> On Fri, 4 Jun 2021 at 09:27, Prathamesh Kulkarni via Gcc-
>> >> >> > > patches
>> >> >> > > > > >>>>>>>> <gcc-patches@gcc.gnu.org> wrote:
>> >> >> > > > > >>>>>>>>> Hi,
>> >> >> > > > > >>>>>>>>> As mentioned in PR, for the following test-case:
>> >> >> > > > > >>>>>>>>>
>> >> >> > > > > >>>>>>>>> #include <arm_neon.h>
>> >> >> > > > > >>>>>>>>>
>> >> >> > > > > >>>>>>>>> bfloat16x4_t f1 (bfloat16_t a)
>> >> >> > > > > >>>>>>>>> {
>> >> >> > > > > >>>>>>>>>     return vdup_n_bf16 (a);
>> >> >> > > > > >>>>>>>>> }
>> >> >> > > > > >>>>>>>>>
>> >> >> > > > > >>>>>>>>> bfloat16x4_t f2 (bfloat16_t a)
>> >> >> > > > > >>>>>>>>> {
>> >> >> > > > > >>>>>>>>>     return (bfloat16x4_t) {a, a, a, a};
>> >> >> > > > > >>>>>>>>> }
>> >> >> > > > > >>>>>>>>>
>> >> >> > > > > >>>>>>>>> Compiling with arm-linux-gnueabi -O3 -mfpu=neon -mfloat-
>> >> >> > > > > >>>> abi=softfp
>> >> >> > > > > >>>>>>>>> -march=armv8.2-a+bf16+fp16 results in f2 not being
>> >> >> > > vectorized:
>> >> >> > > > > >>>>>>>>>
>> >> >> > > > > >>>>>>>>> f1:
>> >> >> > > > > >>>>>>>>>           vdup.16 d16, r0
>> >> >> > > > > >>>>>>>>>           vmov    r0, r1, d16  @ v4bf
>> >> >> > > > > >>>>>>>>>           bx      lr
>> >> >> > > > > >>>>>>>>>
>> >> >> > > > > >>>>>>>>> f2:
>> >> >> > > > > >>>>>>>>>           mov     r3, r0  @ __bf16
>> >> >> > > > > >>>>>>>>>           adr     r1, .L4
>> >> >> > > > > >>>>>>>>>           ldrd    r0, [r1]
>> >> >> > > > > >>>>>>>>>           mov     r2, r3  @ __bf16
>> >> >> > > > > >>>>>>>>>           mov     ip, r3  @ __bf16
>> >> >> > > > > >>>>>>>>>           bfi     r1, r2, #0, #16
>> >> >> > > > > >>>>>>>>>           bfi     r0, ip, #0, #16
>> >> >> > > > > >>>>>>>>>           bfi     r1, r3, #16, #16
>> >> >> > > > > >>>>>>>>>           bfi     r0, r2, #16, #16
>> >> >> > > > > >>>>>>>>>           bx      lr
>> >> >> > > > > >>>>>>>>>
>> >> >> > > > > >>>>>>>>> This seems to happen because vec_init pattern in neon.md
>> >> >> > > has VDQ
>> >> >> > > > > >>>>>> mode
>> >> >> > > > > >>>>>>>>> iterator, which doesn't include V4BF. In attached patch, I
>> >> >> > > changed
>> >> >> > > > > >>>>>>>>> mode
>> >> >> > > > > >>>>>>>>> to VDQX which seems to work for the test-case, and the
>> >> >> > > compiler
>> >> >> > > > > >>>> now
>> >> >> > > > > >>>>>> generates:
>> >> >> > > > > >>>>>>>>> f2:
>> >> >> > > > > >>>>>>>>>           vdup.16 d16, r0
>> >> >> > > > > >>>>>>>>>           vmov    r0, r1, d16  @ v4bf
>> >> >> > > > > >>>>>>>>>           bx      lr
>> >> >> > > > > >>>>>>>>>
>> >> >> > > > > >>>>>>>>> However, the pattern is also gated on TARGET_HAVE_MVE
>> >> >> > > and I am
>> >> >> > > > > >>>>>> not
>> >> >> > > > > >>>>>>>>> sure if either VDQ or VDQX are correct modes for MVE since
>> >> >> > > MVE
>> >> >> > > > > >>>> has
>> >> >> > > > > >>>>>>>>> only 128-bit vectors ?
>> >> >> > > > > >>>>>>>>>
>> >> >> > > > > >>>>>>>> I think patterns common to both Neon and MVE should be
>> >> >> > > moved to
>> >> >> > > > > >>>>>>>> vec-common.md, I don't know why such patterns were left in
>> >> >> > > > > >>>> neon.md.
>> >> >> > > > > >>>>>>> Since we end up calling neon_expand_vector_init for both
>> >> >> > > NEON and
>> >> >> > > > > >>>> MVE,
>> >> >> > > > > >>>>>>> I am not sure if we should separate the pattern ?
>> >> >> > > > > >>>>>>> Would it make sense to FAIL if the mode size isn't 16 bytes for
>> >> >> > > MVE as
>> >> >> > > > > >>>>>>> in attached patch so
>> >> >> > > > > >>>>>>> it will call neon_expand_vector_init only for 128-bit vectors ?
>> >> >> > > > > >>>>>>> Altho hard-coding 16 in the pattern doesn't seem a good idea to
>> >> >> > > me
>> >> >> > > > > >>>> either.
>> >> >> > > > > >>>>>> ping https://gcc.gnu.org/pipermail/gcc-patches/2021-
>> >> >> > > June/572342.html
>> >> >> > > > > >>>>>> (attaching patch as text).
>> >> >> > > > > >>>>>>
>> >> >> > > > > >>>>> --- a/gcc/config/arm/neon.md
>> >> >> > > > > >>>>> +++ b/gcc/config/arm/neon.md
>> >> >> > > > > >>>>> @@ -459,10 +459,12 @@
>> >> >> > > > > >>>>>    )
>> >> >> > > > > >>>>>
>> >> >> > > > > >>>>>    (define_expand "vec_init<mode><V_elem_l>"
>> >> >> > > > > >>>>> -  [(match_operand:VDQ 0 "s_register_operand")
>> >> >> > > > > >>>>> +  [(match_operand:VDQX 0 "s_register_operand")
>> >> >> > > > > >>>>>       (match_operand 1 "" "")]
>> >> >> > > > > >>>>>      "TARGET_NEON || TARGET_HAVE_MVE"
>> >> >> > > > > >>>>>    {
>> >> >> > > > > >>>>> +  if (TARGET_HAVE_MVE && GET_MODE_SIZE (GET_MODE
>> >> >> > > > > >>>> (operands[0])) != 16)
>> >> >> > > > > >>>>> +    FAIL;
>> >> >> > > > > >>>>>      neon_expand_vector_init (operands[0], operands[1]);
>> >> >> > > > > >>>>>      DONE;
>> >> >> > > > > >>>>>    })
>> >> >> > > > > >>>>>
>> >> >> > > > > >>>>> I think we should move this to vec-common.md like Christophe
>> >> >> > > said.
>> >> >> > > > > >>>>> Perhaps rather than making it FAIL for non-16 MVE sizes we just
>> >> >> > > disable it in
>> >> >> > > > > >>>> the expander condition?
>> >> >> > > > > >>>>> "TARGET_NEON || (TARGET_HAVE_MVE && GET_MODE_SIZE (<
>> >> >> > > > > >>>> VDQ>mode) != 16)"
>> >> >> > > > > >>>> Is it OK to use <MODE>mode ? Because using <VDQ>mode resulted
>> >> >> > > in lot
>> >> >> > > > > >>>> of build errors.
>> >> >> > > > > >>>> Also, I think the comparison should be inverted, ie, GET_MODE_SIZE
>> >> >> > > > > >>>> (<MODE>mode) == 16 since
>> >> >> > > > > >>>> we want to make the pattern pass if target is MVE and vector size is
>> >> >> > > 16 bytes ?
>> >> >> > > > > >>>> Do these changes in attached patch look OK ?
>> >> >> > > > > >>> Yes, you're right.
>> >> >> > > > > >>
>> >> >> > > > > >> Can't this be ARM_HAVE_<MODE>_ARITH like in most expanders in
>> >> >> > > vec-common.md?
>> >> >> > > > > >>
>> >> >> > > > > >> (maybe with a && !TARGET_REALLY_IWMMXT if needed)
>> >> >> > > > > > I wonder if this should be ARM_HAVE_<MODE>_LDST instead since
>> >> >> > > we're
>> >> >> > > > > > initializing the vector ?
>> >> >> > > > >
>> >> >> > > > >
>> >> >> > > > > Well, it really depends on which modes you want to enable.
>> >> >> > > > >
>> >> >> > > > >
>> >> >> > > > > Looks like your move VDQ -> VDQ adds V4BF, V8BF and DI.
>> >> >> > > > >
>> >> >> > > > > Are they all OK for Neon?
>> >> >> > > > >
>> >> >> > > > > They are not OK for MVE.
>> >> >> > > > >
>> >> >> > > > > Ideally you could add testcases to cover to the supported and
>> >> >> > > > > unsupported modes for both Neon and MVE.\
>> >> >> > > > >
>> >> >> > > > > Before your patch, the expander is enabled for MVE for 64 bit modes
>> >> >> > > > > (V8QI, V4HI, V2SI): what happens in this case? Does the compiler crash
>> >> >> > > > > or is there something else preventing the match?
>> >> >> > > > Hi,
>> >> >> > > > Apparently there is VALID_MVE_MODE macro, so is it better to use:
>> >> >> > > > TARGET_NEON || (TARGET_HAVE_MVE &&
>> >> >> > > VALID_MVE_MODE(<MODE>mode))
>> >> >> > > > as in the attached patch ?
>> >> >> >
>> >> >> > The change is ok. I would like to see some testcases like Christophe suggested, but this patch just moves the expander around rather than introducing new functionality.
>> >> >> Hi Kyrill,
>> >> >> As mentioned in the first email, the patch improves code-gen for
>> >> >> following test-case:
>> >> >>
>> >> >> bfloat16x4_t f (bfloat16_t a)
>> >> >> {
>> >> >>   return (bfloat16x4_t) {a, a, a, a};
>> >> >> }
>> >> >>
>> >> >> Before patch:
>> >> >> f:
>> >> >>         mov     r3, r0  @ __bf16
>> >> >>         adr     r1, .L4
>> >> >>         ldrd    r0, [r1]
>> >> >>         mov     r2, r3  @ __bf16
>> >> >>         mov     ip, r3  @ __bf16
>> >> >>         bfi     r1, r2, #0, #16
>> >> >>         bfi     r0, ip, #0, #16
>> >> >>         bfi     r1, r3, #16, #16
>> >> >>         bfi     r0, r2, #16, #16
>> >> >>         bx      lr
>> >> >>
>> >> >> After patch:
>> >> >> f:
>> >> >>         vdup.16 d16, r0
>> >> >>         vmov    r0, r1, d16  @ v4bf
>> >> >>         bx      lr
>> >> >>
>> >> >> because the patch changes mode from VDQ to VDQX to accommodate bf modes.
>> >> >> I have included the test in the attached patch.
>> >> >> I think Christophe's concerns were mainly about the right modes
>> >> >> getting enabled for MVE.
>> >> >> Unfortunately, I am not sure how to test for that because the FE
>> >> >> catches invalid modes, and we don't
>> >> >> end up hitting the pattern.
>> >> >>
>> >> >
>> >> > Hi Prathamesh,
>> >> >
>> >> > The new testcase fails on arm-linux-gnueabihf:
>> >> >  FAIL: gcc.target/arm/simd/pr98435.c (test for excess errors)
>> >> > Excess errors:
>> >> > /aci-gcc-fsf/builds/gcc-fsf-gccsrc/sysroot-arm-none-linux-gnueabihf/usr/include/gnu/stubs.h:7:11: fatal error: gnu/stubs-soft.h: No such file or directory
>> >> > compilation terminated.
>> >> >
>> >> > Because you don't check whether  -mfloat-abi=softfp is actually supported.
>> >> >
>> >> > Can you fix that?
>> >> Oops, sorry about that.
>> >> The attached patch fixes the test by requiring arm_softfloat and makes
>> >> it UNSUPPORTED on arm-linux-gnueabihf.
>> >> Does it look OK ?
>> >>
>> >
>> > I don't think that's right: it would make the test unsupported if softfp is not the default even if the toolchain has the needed multilibs.
>> > Did you check eg. with arm-eabi and multilibs enabled?
>> Ah OK, thanks for pointing it out!
>> Does the attached patch look correct ?
>>
>
> I don't think: this would skip the test even if the toolchain has multilibs enabled.
> Did you check eg. with arm-eabi and multilibs enabled and the usual option overrides?
It showed 3 PASS with second patch:
/* { dg-skip-if "skip test for hard float" { *-*-* } {
"-mfloat-abi=hard" } { "" } } */

I ran it using make check-gcc RUNTESTFLAGS="simd.exp=pr98435.c"
and built the toolchain using:
abe.sh --target arm-eabi --build all --set multilib=aprofile gcc=gcc.git~master.
I suppose that's correct ?

gcc -v output:
Configured with:
'/home/prathamesh.kulkarni/abe-toolchain-2/snapshots/gcc.git~master/configure'
SHELL=/bin/bash
--with-mpc=/home/prathamesh.kulkarni/abe-toolchain-2/builds/destdir/x86_64-pc-linux-gnu
--with-mpfr=/home/prathamesh.kulkarni/abe-toolchain-2/builds/destdir/x86_64-pc-linux-gnu
--with-gmp=/home/prathamesh.kulkarni/abe-toolchain-2/builds/destdir/x86_64-pc-linux-gnu
--with-gnu-as --with-gnu-ld --disable-libmudflap --enable-lto
--enable-shared --without-included-gettext --enable-nls
--with-system-zlib --disable-sjlj-exceptions
--enable-gnu-unique-object --enable-linker-build-id
--disable-libstdcxx-pch --enable-c99 --enable-clocale=gnu
--enable-libstdcxx-debug --enable-long-long --with-cloog=no
--with-ppl=no --with-isl=no --enable-multilib
--with-multilib-list=aprofile --enable-threads=no --disable-multiarch
--with-sysroot=/home/prathamesh.kulkarni/abe-toolchain-2/builds/destdir/x86_64-pc-linux-gnu/arm-eabi
--with-newlib --enable-checking=yes --disable-bootstrap
--enable-languages=c,c++,lto
--prefix=/home/prathamesh.kulkarni/abe-toolchain-2/builds/destdir/x86_64-pc-linux-gnu
--build=x86_64-pc-linux-gnu --host=x86_64-pc-linux-gnu
--target=arm-eabi

Thanks,
Prathamesh
>
>
> Christophe
>
>> Thanks,
>> Prathamesh
>> >
>> > Christophe
>> >
>> >>
>> >> Thanks,
>> >> Prathamesh
>> >> >
>> >> > Thanks
>> >> >
>> >> > Christophe
>> >> >
>> >> >
>> >> >>
>> >> >> Thanks,
>> >> >> Prathamesh
>> >> >> > Thanks,
>> >> >> > Kyrill
>> >> >> >
>> >> >> > > ping https://gcc.gnu.org/pipermail/gcc-patches/2021-July/574206.html
>> >> >> > >
>> >> >> > > Thanks,
>> >> >> > > Prathamesh
>> >> >> > > >
>> >> >> > > > Thanks,
>> >> >> > > > Prathamesh
>> >> >> > > > >
>> >> >> > > > >
>> >> >> > > > > Thanks,
>> >> >> > > > >
>> >> >> > > > >
>> >> >> > > > > Christophe
>> >> >> > > > >
>> >> >> > > > >
>> >> >> > > > > > Thanks,
>> >> >> > > > > > Prathamesh
>> >> >> > > > > >>
>> >> >> > > > > >> Christophe
>> >> >> > > > > >>
>> >> >> > > > > >>
>> >> >> > > > > >>> Ok.
>> >> >> > > > > >>> Thanks,
>> >> >> > > > > >>> Kyrill
>> >> >> > > > > >>>
>> >> >> > > > > >>>
>> >> >> > > > > >>>> Thanks,
>> >> >> > > > > >>>> Prathamesh
>> >> >> > > > > >>>>> Thanks,
>> >> >> > > > > >>>>> Kyrill
>> >> >> > > > > >>>>>
>> >> >> > > > > >>>>>> Thanks,
>> >> >> > > > > >>>>>> Prathamesh
>> >> >> > > > > >>>>>>> Thanks,
>> >> >> > > > > >>>>>>> Prathamesh
>> >> >> > > > > >>>>>>>> That being said, I suggest you look at other similar patterns in
>> >> >> > > > > >>>>>>>> vec-common.md, most of which are gated on
>> >> >> > > > > >>>>>>>> ARM_HAVE_<MODE>_ARITH
>> >> >> > > > > >>>>>>>> and possibly beware of issues with iwmmxt :-)
>> >> >> > > > > >>>>>>>>
>> >> >> > > > > >>>>>>>> Christophe
>> >> >> > > > > >>>>>>>>
>> >> >> > > > > >>>>>>>>> Thanks,
>> >> >> > > > > >>>>>>>>> Prathamesh