[PATCH][cunroll] Add unroll-known-loop-iterations-only param and use it in aarch64

classic Classic list List threaded Threaded
13 messages Options
Reply | Threaded
Open this post in threaded view
|

[PATCH][cunroll] Add unroll-known-loop-iterations-only param and use it in aarch64

Kyrill  Tkachov-2
Hi all,

In this testcase the codegen for VLA SVE is worse than it could be due to unrolling:

fully_peel_me:
         mov     x1, 5
         ptrue   p1.d, all
         whilelo p0.d, xzr, x1
         ld1d    z0.d, p0/z, [x0]
         fadd    z0.d, z0.d, z0.d
         st1d    z0.d, p0, [x0]
         cntd    x2
         addvl   x3, x0, #1
         whilelo p0.d, x2, x1
         beq     .L1
         ld1d    z0.d, p0/z, [x0, #1, mul vl]
         fadd    z0.d, z0.d, z0.d
         st1d    z0.d, p0, [x3]
         cntw    x2
         incb    x0, all, mul #2
         whilelo p0.d, x2, x1
         beq     .L1
         ld1d    z0.d, p0/z, [x0]
         fadd    z0.d, z0.d, z0.d
         st1d    z0.d, p0, [x0]
.L1:
         ret

In this case, due to the vector-length-agnostic nature of SVE the compiler doesn't know the loop iteration count.
For such loops we don't want to unroll if we don't end up eliminating branches as this just bloats code size
and hurts icache performance.

This patch introduces a new unroll-known-loop-iterations-only param that disables cunroll when the loop iteration
count is unknown (SCEV_NOT_KNOWN). This case occurs much more often for SVE VLA code, but it does help some
Advanced SIMD cases as well where loops with an unknown iteration count are not unrolled when it doesn't eliminate
the branches.

So for the above testcase we generate now:
fully_peel_me:
         mov     x2, 5
         mov     x3, x2
         mov     x1, 0
         whilelo p0.d, xzr, x2
         ptrue   p1.d, all
.L2:
         ld1d    z0.d, p0/z, [x0, x1, lsl 3]
         fadd    z0.d, z0.d, z0.d
         st1d    z0.d, p0, [x0, x1, lsl 3]
         incd    x1
         whilelo p0.d, x1, x3
         bne     .L2
         ret

Not perfect still, but it's preferable to the original code.
The new param is enabled by default on aarch64 but disabled for other targets, leaving their behaviour unchanged
(until other target people experiment with it and set it, if appropriate).

Bootstrapped and tested on aarch64-none-linux-gnu.
Benchmarked on SPEC2017 on a Cortex-A57 and there are no differences in performance.

Ok for trunk?
Thanks,
Kyrill


2018-11-09  Kyrylo Tkachov  <[hidden email]>

        * params.def (PARAM_UNROLL_KNOWN_LOOP_ITERATIONS_ONLY): Define.
        * tree-ssa-loop-ivcanon.c (try_unroll_loop_completely): Use above to
        disable unrolling on unknown iteration count.
        * config/aarch64/aarch64.c (aarch64_override_options_internal): Set
        PARAM_UNROLL_KNOWN_LOOP_ITERATIONS_ONLY to 1.
        * doc/invoke.texi (--param unroll-known-loop-iterations-only):
        Document.

2018-11-09  Kyrylo Tkachov  <[hidden email]>

        * gcc.target/aarch64/sve/unroll-1.c: New test.


unroll.patch (3K) Download Attachment
Reply | Threaded
Open this post in threaded view
|

Re: [PATCH][cunroll] Add unroll-known-loop-iterations-only param and use it in aarch64

Richard Biener-2
On Fri, Nov 9, 2018 at 11:47 AM Kyrill Tkachov
<[hidden email]> wrote:

>
> Hi all,
>
> In this testcase the codegen for VLA SVE is worse than it could be due to unrolling:
>
> fully_peel_me:
>          mov     x1, 5
>          ptrue   p1.d, all
>          whilelo p0.d, xzr, x1
>          ld1d    z0.d, p0/z, [x0]
>          fadd    z0.d, z0.d, z0.d
>          st1d    z0.d, p0, [x0]
>          cntd    x2
>          addvl   x3, x0, #1
>          whilelo p0.d, x2, x1
>          beq     .L1
>          ld1d    z0.d, p0/z, [x0, #1, mul vl]
>          fadd    z0.d, z0.d, z0.d
>          st1d    z0.d, p0, [x3]
>          cntw    x2
>          incb    x0, all, mul #2
>          whilelo p0.d, x2, x1
>          beq     .L1
>          ld1d    z0.d, p0/z, [x0]
>          fadd    z0.d, z0.d, z0.d
>          st1d    z0.d, p0, [x0]
> .L1:
>          ret
>
> In this case, due to the vector-length-agnostic nature of SVE the compiler doesn't know the loop iteration count.
> For such loops we don't want to unroll if we don't end up eliminating branches as this just bloats code size
> and hurts icache performance.
>
> This patch introduces a new unroll-known-loop-iterations-only param that disables cunroll when the loop iteration
> count is unknown (SCEV_NOT_KNOWN). This case occurs much more often for SVE VLA code, but it does help some
> Advanced SIMD cases as well where loops with an unknown iteration count are not unrolled when it doesn't eliminate
> the branches.
>
> So for the above testcase we generate now:
> fully_peel_me:
>          mov     x2, 5
>          mov     x3, x2
>          mov     x1, 0
>          whilelo p0.d, xzr, x2
>          ptrue   p1.d, all
> .L2:
>          ld1d    z0.d, p0/z, [x0, x1, lsl 3]
>          fadd    z0.d, z0.d, z0.d
>          st1d    z0.d, p0, [x0, x1, lsl 3]
>          incd    x1
>          whilelo p0.d, x1, x3
>          bne     .L2
>          ret
>
> Not perfect still, but it's preferable to the original code.
> The new param is enabled by default on aarch64 but disabled for other targets, leaving their behaviour unchanged
> (until other target people experiment with it and set it, if appropriate).
>
> Bootstrapped and tested on aarch64-none-linux-gnu.
> Benchmarked on SPEC2017 on a Cortex-A57 and there are no differences in performance.
>
> Ok for trunk?

Hum.  Why introduce a new --param and not simply key on
flag_peel_loops instead?  That is
enabled by default at -O3 and with FDO but you of course can control
that in your targets
post-option-processing hook.

It might also make sense to have more fine-grained control for this
and allow a target
to say whether it wants to peel a specific loop or not when the
middle-end thinks that
would be profitable.

Richard.

> Thanks,
> Kyrill
>
>
> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
>
>         * params.def (PARAM_UNROLL_KNOWN_LOOP_ITERATIONS_ONLY): Define.
>         * tree-ssa-loop-ivcanon.c (try_unroll_loop_completely): Use above to
>         disable unrolling on unknown iteration count.
>         * config/aarch64/aarch64.c (aarch64_override_options_internal): Set
>         PARAM_UNROLL_KNOWN_LOOP_ITERATIONS_ONLY to 1.
>         * doc/invoke.texi (--param unroll-known-loop-iterations-only):
>         Document.
>
> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
>
>         * gcc.target/aarch64/sve/unroll-1.c: New test.
>
Reply | Threaded
Open this post in threaded view
|

Re: [PATCH][cunroll] Add unroll-known-loop-iterations-only param and use it in aarch64

Kyrill  Tkachov-2
On 09/11/18 12:18, Richard Biener wrote:

> On Fri, Nov 9, 2018 at 11:47 AM Kyrill Tkachov
> <[hidden email]> wrote:
>>
>> Hi all,
>>
>> In this testcase the codegen for VLA SVE is worse than it could be due to unrolling:
>>
>> fully_peel_me:
>>          mov     x1, 5
>>          ptrue   p1.d, all
>>          whilelo p0.d, xzr, x1
>>          ld1d    z0.d, p0/z, [x0]
>>          fadd    z0.d, z0.d, z0.d
>>          st1d    z0.d, p0, [x0]
>>          cntd    x2
>>          addvl   x3, x0, #1
>>          whilelo p0.d, x2, x1
>>          beq     .L1
>>          ld1d    z0.d, p0/z, [x0, #1, mul vl]
>>          fadd    z0.d, z0.d, z0.d
>>          st1d    z0.d, p0, [x3]
>>          cntw    x2
>>          incb    x0, all, mul #2
>>          whilelo p0.d, x2, x1
>>          beq     .L1
>>          ld1d    z0.d, p0/z, [x0]
>>          fadd    z0.d, z0.d, z0.d
>>          st1d    z0.d, p0, [x0]
>> .L1:
>>          ret
>>
>> In this case, due to the vector-length-agnostic nature of SVE the compiler doesn't know the loop iteration count.
>> For such loops we don't want to unroll if we don't end up eliminating branches as this just bloats code size
>> and hurts icache performance.
>>
>> This patch introduces a new unroll-known-loop-iterations-only param that disables cunroll when the loop iteration
>> count is unknown (SCEV_NOT_KNOWN). This case occurs much more often for SVE VLA code, but it does help some
>> Advanced SIMD cases as well where loops with an unknown iteration count are not unrolled when it doesn't eliminate
>> the branches.
>>
>> So for the above testcase we generate now:
>> fully_peel_me:
>>          mov     x2, 5
>>          mov     x3, x2
>>          mov     x1, 0
>>          whilelo p0.d, xzr, x2
>>          ptrue   p1.d, all
>> .L2:
>>          ld1d    z0.d, p0/z, [x0, x1, lsl 3]
>>          fadd    z0.d, z0.d, z0.d
>>          st1d    z0.d, p0, [x0, x1, lsl 3]
>>          incd    x1
>>          whilelo p0.d, x1, x3
>>          bne     .L2
>>          ret
>>
>> Not perfect still, but it's preferable to the original code.
>> The new param is enabled by default on aarch64 but disabled for other targets, leaving their behaviour unchanged
>> (until other target people experiment with it and set it, if appropriate).
>>
>> Bootstrapped and tested on aarch64-none-linux-gnu.
>> Benchmarked on SPEC2017 on a Cortex-A57 and there are no differences in performance.
>>
>> Ok for trunk?
>
> Hum.  Why introduce a new --param and not simply key on
> flag_peel_loops instead?  That is
> enabled by default at -O3 and with FDO but you of course can control
> that in your targets
> post-option-processing hook.
You mean like this?
It's certainly a simpler patch, but I was just a bit hesitant of making this change for all targets :)
But I suppose it's a reasonable change.

>
> It might also make sense to have more fine-grained control for this
> and allow a target
> to say whether it wants to peel a specific loop or not when the
> middle-end thinks that
> would be profitable.

Can be worth looking at as a follow-up. Do you envisage the target analysing
the gimple statements of the loop to figure out its cost?

Thanks,
Kyrill


2018-11-09  Kyrylo Tkachov  <[hidden email]>

        * tree-ssa-loop-ivcanon.c (try_unroll_loop_completely): Do not unroll
        loop when number of iterations is not known and flag_peel_loops is in
        effect.

2018-11-09  Kyrylo Tkachov  <[hidden email]>

        * gcc.target/aarch64/sve/unroll-1.c: New test.


unroll2.patch (1K) Download Attachment
Reply | Threaded
Open this post in threaded view
|

Re: [PATCH][cunroll] Add unroll-known-loop-iterations-only param and use it in aarch64

Richard Biener-2
On Fri, Nov 9, 2018 at 6:57 PM Kyrill Tkachov
<[hidden email]> wrote:

>
> On 09/11/18 12:18, Richard Biener wrote:
> > On Fri, Nov 9, 2018 at 11:47 AM Kyrill Tkachov
> > <[hidden email]> wrote:
> >>
> >> Hi all,
> >>
> >> In this testcase the codegen for VLA SVE is worse than it could be due to unrolling:
> >>
> >> fully_peel_me:
> >>          mov     x1, 5
> >>          ptrue   p1.d, all
> >>          whilelo p0.d, xzr, x1
> >>          ld1d    z0.d, p0/z, [x0]
> >>          fadd    z0.d, z0.d, z0.d
> >>          st1d    z0.d, p0, [x0]
> >>          cntd    x2
> >>          addvl   x3, x0, #1
> >>          whilelo p0.d, x2, x1
> >>          beq     .L1
> >>          ld1d    z0.d, p0/z, [x0, #1, mul vl]
> >>          fadd    z0.d, z0.d, z0.d
> >>          st1d    z0.d, p0, [x3]
> >>          cntw    x2
> >>          incb    x0, all, mul #2
> >>          whilelo p0.d, x2, x1
> >>          beq     .L1
> >>          ld1d    z0.d, p0/z, [x0]
> >>          fadd    z0.d, z0.d, z0.d
> >>          st1d    z0.d, p0, [x0]
> >> .L1:
> >>          ret
> >>
> >> In this case, due to the vector-length-agnostic nature of SVE the compiler doesn't know the loop iteration count.
> >> For such loops we don't want to unroll if we don't end up eliminating branches as this just bloats code size
> >> and hurts icache performance.
> >>
> >> This patch introduces a new unroll-known-loop-iterations-only param that disables cunroll when the loop iteration
> >> count is unknown (SCEV_NOT_KNOWN). This case occurs much more often for SVE VLA code, but it does help some
> >> Advanced SIMD cases as well where loops with an unknown iteration count are not unrolled when it doesn't eliminate
> >> the branches.
> >>
> >> So for the above testcase we generate now:
> >> fully_peel_me:
> >>          mov     x2, 5
> >>          mov     x3, x2
> >>          mov     x1, 0
> >>          whilelo p0.d, xzr, x2
> >>          ptrue   p1.d, all
> >> .L2:
> >>          ld1d    z0.d, p0/z, [x0, x1, lsl 3]
> >>          fadd    z0.d, z0.d, z0.d
> >>          st1d    z0.d, p0, [x0, x1, lsl 3]
> >>          incd    x1
> >>          whilelo p0.d, x1, x3
> >>          bne     .L2
> >>          ret
> >>
> >> Not perfect still, but it's preferable to the original code.
> >> The new param is enabled by default on aarch64 but disabled for other targets, leaving their behaviour unchanged
> >> (until other target people experiment with it and set it, if appropriate).
> >>
> >> Bootstrapped and tested on aarch64-none-linux-gnu.
> >> Benchmarked on SPEC2017 on a Cortex-A57 and there are no differences in performance.
> >>
> >> Ok for trunk?
> >
> > Hum.  Why introduce a new --param and not simply key on
> > flag_peel_loops instead?  That is
> > enabled by default at -O3 and with FDO but you of course can control
> > that in your targets
> > post-option-processing hook.
>
> You mean like this?
> It's certainly a simpler patch, but I was just a bit hesitant of making this change for all targets :)
> But I suppose it's a reasonable change.

No, that change is backward.  What I said is that peeling is already
conditional on
flag_peel_loops and that is enabled by -O3.  So you want to disable
flag_peel_loops for
SVE instead in the target.

> >
> > It might also make sense to have more fine-grained control for this
> > and allow a target
> > to say whether it wants to peel a specific loop or not when the
> > middle-end thinks that
> > would be profitable.
>
> Can be worth looking at as a follow-up. Do you envisage the target analysing
> the gimple statements of the loop to figure out its cost?

Kind-of.  Sth like

  bool targetm.peel_loop (struct loop *);

I have no idea whether you can easily detect a SVE vectorized loop though.
Maybe there's always a special IV or so (the mask?)

Richard.

> Thanks,
> Kyrill
>
>
> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
>
>         * tree-ssa-loop-ivcanon.c (try_unroll_loop_completely): Do not unroll
>         loop when number of iterations is not known and flag_peel_loops is in
>         effect.
>
> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
>
>         * gcc.target/aarch64/sve/unroll-1.c: New test.
>
Reply | Threaded
Open this post in threaded view
|

Re: [PATCH][cunroll] Add unroll-known-loop-iterations-only param and use it in aarch64

Kyrill  Tkachov-2

On 12/11/18 14:10, Richard Biener wrote:

> On Fri, Nov 9, 2018 at 6:57 PM Kyrill Tkachov
> <[hidden email]> wrote:
>> On 09/11/18 12:18, Richard Biener wrote:
>>> On Fri, Nov 9, 2018 at 11:47 AM Kyrill Tkachov
>>> <[hidden email]> wrote:
>>>> Hi all,
>>>>
>>>> In this testcase the codegen for VLA SVE is worse than it could be due to unrolling:
>>>>
>>>> fully_peel_me:
>>>>           mov     x1, 5
>>>>           ptrue   p1.d, all
>>>>           whilelo p0.d, xzr, x1
>>>>           ld1d    z0.d, p0/z, [x0]
>>>>           fadd    z0.d, z0.d, z0.d
>>>>           st1d    z0.d, p0, [x0]
>>>>           cntd    x2
>>>>           addvl   x3, x0, #1
>>>>           whilelo p0.d, x2, x1
>>>>           beq     .L1
>>>>           ld1d    z0.d, p0/z, [x0, #1, mul vl]
>>>>           fadd    z0.d, z0.d, z0.d
>>>>           st1d    z0.d, p0, [x3]
>>>>           cntw    x2
>>>>           incb    x0, all, mul #2
>>>>           whilelo p0.d, x2, x1
>>>>           beq     .L1
>>>>           ld1d    z0.d, p0/z, [x0]
>>>>           fadd    z0.d, z0.d, z0.d
>>>>           st1d    z0.d, p0, [x0]
>>>> .L1:
>>>>           ret
>>>>
>>>> In this case, due to the vector-length-agnostic nature of SVE the compiler doesn't know the loop iteration count.
>>>> For such loops we don't want to unroll if we don't end up eliminating branches as this just bloats code size
>>>> and hurts icache performance.
>>>>
>>>> This patch introduces a new unroll-known-loop-iterations-only param that disables cunroll when the loop iteration
>>>> count is unknown (SCEV_NOT_KNOWN). This case occurs much more often for SVE VLA code, but it does help some
>>>> Advanced SIMD cases as well where loops with an unknown iteration count are not unrolled when it doesn't eliminate
>>>> the branches.
>>>>
>>>> So for the above testcase we generate now:
>>>> fully_peel_me:
>>>>           mov     x2, 5
>>>>           mov     x3, x2
>>>>           mov     x1, 0
>>>>           whilelo p0.d, xzr, x2
>>>>           ptrue   p1.d, all
>>>> .L2:
>>>>           ld1d    z0.d, p0/z, [x0, x1, lsl 3]
>>>>           fadd    z0.d, z0.d, z0.d
>>>>           st1d    z0.d, p0, [x0, x1, lsl 3]
>>>>           incd    x1
>>>>           whilelo p0.d, x1, x3
>>>>           bne     .L2
>>>>           ret
>>>>
>>>> Not perfect still, but it's preferable to the original code.
>>>> The new param is enabled by default on aarch64 but disabled for other targets, leaving their behaviour unchanged
>>>> (until other target people experiment with it and set it, if appropriate).
>>>>
>>>> Bootstrapped and tested on aarch64-none-linux-gnu.
>>>> Benchmarked on SPEC2017 on a Cortex-A57 and there are no differences in performance.
>>>>
>>>> Ok for trunk?
>>> Hum.  Why introduce a new --param and not simply key on
>>> flag_peel_loops instead?  That is
>>> enabled by default at -O3 and with FDO but you of course can control
>>> that in your targets
>>> post-option-processing hook.
>> You mean like this?
>> It's certainly a simpler patch, but I was just a bit hesitant of making this change for all targets :)
>> But I suppose it's a reasonable change.
> No, that change is backward.  What I said is that peeling is already
> conditional on
> flag_peel_loops and that is enabled by -O3.  So you want to disable
> flag_peel_loops for
> SVE instead in the target.

Sorry, I got confused by the similarly named functions.
I'm talking about try_unroll_loop_completely when run as part of canonicalize_induction_variables i.e. the "ivcanon" pass
(sorry about blaming cunroll here). This doesn't get called through the try_unroll_loops_completely path.

try_unroll_loop_completely doesn't get disabled with -fno-peel-loops or -fno-unroll-loops.
Maybe disabling peeling inside try_unroll_loop_completely itself when !flag_peel_loops is viable?

Thanks,
Kyrill

>>> It might also make sense to have more fine-grained control for this
>>> and allow a target
>>> to say whether it wants to peel a specific loop or not when the
>>> middle-end thinks that
>>> would be profitable.
>> Can be worth looking at as a follow-up. Do you envisage the target analysing
>> the gimple statements of the loop to figure out its cost?
> Kind-of.  Sth like
>
>    bool targetm.peel_loop (struct loop *);
>
> I have no idea whether you can easily detect a SVE vectorized loop though.
> Maybe there's always a special IV or so (the mask?)
>
> Richard.
>
>> Thanks,
>> Kyrill
>>
>>
>> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
>>
>>          * tree-ssa-loop-ivcanon.c (try_unroll_loop_completely): Do not unroll
>>          loop when number of iterations is not known and flag_peel_loops is in
>>          effect.
>>
>> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
>>
>>          * gcc.target/aarch64/sve/unroll-1.c: New test.
>>

Reply | Threaded
Open this post in threaded view
|

Re: [PATCH][cunroll] Add unroll-known-loop-iterations-only param and use it in aarch64

Richard Biener-2
On Mon, Nov 12, 2018 at 7:20 PM Kyrill Tkachov
<[hidden email]> wrote:

>
>
> On 12/11/18 14:10, Richard Biener wrote:
> > On Fri, Nov 9, 2018 at 6:57 PM Kyrill Tkachov
> > <[hidden email]> wrote:
> >> On 09/11/18 12:18, Richard Biener wrote:
> >>> On Fri, Nov 9, 2018 at 11:47 AM Kyrill Tkachov
> >>> <[hidden email]> wrote:
> >>>> Hi all,
> >>>>
> >>>> In this testcase the codegen for VLA SVE is worse than it could be due to unrolling:
> >>>>
> >>>> fully_peel_me:
> >>>>           mov     x1, 5
> >>>>           ptrue   p1.d, all
> >>>>           whilelo p0.d, xzr, x1
> >>>>           ld1d    z0.d, p0/z, [x0]
> >>>>           fadd    z0.d, z0.d, z0.d
> >>>>           st1d    z0.d, p0, [x0]
> >>>>           cntd    x2
> >>>>           addvl   x3, x0, #1
> >>>>           whilelo p0.d, x2, x1
> >>>>           beq     .L1
> >>>>           ld1d    z0.d, p0/z, [x0, #1, mul vl]
> >>>>           fadd    z0.d, z0.d, z0.d
> >>>>           st1d    z0.d, p0, [x3]
> >>>>           cntw    x2
> >>>>           incb    x0, all, mul #2
> >>>>           whilelo p0.d, x2, x1
> >>>>           beq     .L1
> >>>>           ld1d    z0.d, p0/z, [x0]
> >>>>           fadd    z0.d, z0.d, z0.d
> >>>>           st1d    z0.d, p0, [x0]
> >>>> .L1:
> >>>>           ret
> >>>>
> >>>> In this case, due to the vector-length-agnostic nature of SVE the compiler doesn't know the loop iteration count.
> >>>> For such loops we don't want to unroll if we don't end up eliminating branches as this just bloats code size
> >>>> and hurts icache performance.
> >>>>
> >>>> This patch introduces a new unroll-known-loop-iterations-only param that disables cunroll when the loop iteration
> >>>> count is unknown (SCEV_NOT_KNOWN). This case occurs much more often for SVE VLA code, but it does help some
> >>>> Advanced SIMD cases as well where loops with an unknown iteration count are not unrolled when it doesn't eliminate
> >>>> the branches.
> >>>>
> >>>> So for the above testcase we generate now:
> >>>> fully_peel_me:
> >>>>           mov     x2, 5
> >>>>           mov     x3, x2
> >>>>           mov     x1, 0
> >>>>           whilelo p0.d, xzr, x2
> >>>>           ptrue   p1.d, all
> >>>> .L2:
> >>>>           ld1d    z0.d, p0/z, [x0, x1, lsl 3]
> >>>>           fadd    z0.d, z0.d, z0.d
> >>>>           st1d    z0.d, p0, [x0, x1, lsl 3]
> >>>>           incd    x1
> >>>>           whilelo p0.d, x1, x3
> >>>>           bne     .L2
> >>>>           ret
> >>>>
> >>>> Not perfect still, but it's preferable to the original code.
> >>>> The new param is enabled by default on aarch64 but disabled for other targets, leaving their behaviour unchanged
> >>>> (until other target people experiment with it and set it, if appropriate).
> >>>>
> >>>> Bootstrapped and tested on aarch64-none-linux-gnu.
> >>>> Benchmarked on SPEC2017 on a Cortex-A57 and there are no differences in performance.
> >>>>
> >>>> Ok for trunk?
> >>> Hum.  Why introduce a new --param and not simply key on
> >>> flag_peel_loops instead?  That is
> >>> enabled by default at -O3 and with FDO but you of course can control
> >>> that in your targets
> >>> post-option-processing hook.
> >> You mean like this?
> >> It's certainly a simpler patch, but I was just a bit hesitant of making this change for all targets :)
> >> But I suppose it's a reasonable change.
> > No, that change is backward.  What I said is that peeling is already
> > conditional on
> > flag_peel_loops and that is enabled by -O3.  So you want to disable
> > flag_peel_loops for
> > SVE instead in the target.
>
> Sorry, I got confused by the similarly named functions.
> I'm talking about try_unroll_loop_completely when run as part of canonicalize_induction_variables i.e. the "ivcanon" pass
> (sorry about blaming cunroll here). This doesn't get called through the try_unroll_loops_completely path.

Well, peeling gets disabled.  From your patch I see you want to
disable "unrolling" when
the number of loop iteration is not constant.  That is called peeling
where we need to
emit the loop exit test N times.

Did you check your testcases with -fno-peel-loops?

> try_unroll_loop_completely doesn't get disabled with -fno-peel-loops or -fno-unroll-loops.
> Maybe disabling peeling inside try_unroll_loop_completely itself when !flag_peel_loops is viable?
>
> Thanks,
> Kyrill
>
> >>> It might also make sense to have more fine-grained control for this
> >>> and allow a target
> >>> to say whether it wants to peel a specific loop or not when the
> >>> middle-end thinks that
> >>> would be profitable.
> >> Can be worth looking at as a follow-up. Do you envisage the target analysing
> >> the gimple statements of the loop to figure out its cost?
> > Kind-of.  Sth like
> >
> >    bool targetm.peel_loop (struct loop *);
> >
> > I have no idea whether you can easily detect a SVE vectorized loop though.
> > Maybe there's always a special IV or so (the mask?)
> >
> > Richard.
> >
> >> Thanks,
> >> Kyrill
> >>
> >>
> >> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
> >>
> >>          * tree-ssa-loop-ivcanon.c (try_unroll_loop_completely): Do not unroll
> >>          loop when number of iterations is not known and flag_peel_loops is in
> >>          effect.
> >>
> >> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
> >>
> >>          * gcc.target/aarch64/sve/unroll-1.c: New test.
> >>
>
Reply | Threaded
Open this post in threaded view
|

Re: [PATCH][cunroll] Add unroll-known-loop-iterations-only param and use it in aarch64

Kyrill  Tkachov-2
Hi Richard,

On 13/11/18 08:24, Richard Biener wrote:

> On Mon, Nov 12, 2018 at 7:20 PM Kyrill Tkachov
> <[hidden email]> wrote:
>>
>> On 12/11/18 14:10, Richard Biener wrote:
>>> On Fri, Nov 9, 2018 at 6:57 PM Kyrill Tkachov
>>> <[hidden email]> wrote:
>>>> On 09/11/18 12:18, Richard Biener wrote:
>>>>> On Fri, Nov 9, 2018 at 11:47 AM Kyrill Tkachov
>>>>> <[hidden email]> wrote:
>>>>>> Hi all,
>>>>>>
>>>>>> In this testcase the codegen for VLA SVE is worse than it could be due to unrolling:
>>>>>>
>>>>>> fully_peel_me:
>>>>>>            mov     x1, 5
>>>>>>            ptrue   p1.d, all
>>>>>>            whilelo p0.d, xzr, x1
>>>>>>            ld1d    z0.d, p0/z, [x0]
>>>>>>            fadd    z0.d, z0.d, z0.d
>>>>>>            st1d    z0.d, p0, [x0]
>>>>>>            cntd    x2
>>>>>>            addvl   x3, x0, #1
>>>>>>            whilelo p0.d, x2, x1
>>>>>>            beq     .L1
>>>>>>            ld1d    z0.d, p0/z, [x0, #1, mul vl]
>>>>>>            fadd    z0.d, z0.d, z0.d
>>>>>>            st1d    z0.d, p0, [x3]
>>>>>>            cntw    x2
>>>>>>            incb    x0, all, mul #2
>>>>>>            whilelo p0.d, x2, x1
>>>>>>            beq     .L1
>>>>>>            ld1d    z0.d, p0/z, [x0]
>>>>>>            fadd    z0.d, z0.d, z0.d
>>>>>>            st1d    z0.d, p0, [x0]
>>>>>> .L1:
>>>>>>            ret
>>>>>>
>>>>>> In this case, due to the vector-length-agnostic nature of SVE the compiler doesn't know the loop iteration count.
>>>>>> For such loops we don't want to unroll if we don't end up eliminating branches as this just bloats code size
>>>>>> and hurts icache performance.
>>>>>>
>>>>>> This patch introduces a new unroll-known-loop-iterations-only param that disables cunroll when the loop iteration
>>>>>> count is unknown (SCEV_NOT_KNOWN). This case occurs much more often for SVE VLA code, but it does help some
>>>>>> Advanced SIMD cases as well where loops with an unknown iteration count are not unrolled when it doesn't eliminate
>>>>>> the branches.
>>>>>>
>>>>>> So for the above testcase we generate now:
>>>>>> fully_peel_me:
>>>>>>            mov     x2, 5
>>>>>>            mov     x3, x2
>>>>>>            mov     x1, 0
>>>>>>            whilelo p0.d, xzr, x2
>>>>>>            ptrue   p1.d, all
>>>>>> .L2:
>>>>>>            ld1d    z0.d, p0/z, [x0, x1, lsl 3]
>>>>>>            fadd    z0.d, z0.d, z0.d
>>>>>>            st1d    z0.d, p0, [x0, x1, lsl 3]
>>>>>>            incd    x1
>>>>>>            whilelo p0.d, x1, x3
>>>>>>            bne     .L2
>>>>>>            ret
>>>>>>
>>>>>> Not perfect still, but it's preferable to the original code.
>>>>>> The new param is enabled by default on aarch64 but disabled for other targets, leaving their behaviour unchanged
>>>>>> (until other target people experiment with it and set it, if appropriate).
>>>>>>
>>>>>> Bootstrapped and tested on aarch64-none-linux-gnu.
>>>>>> Benchmarked on SPEC2017 on a Cortex-A57 and there are no differences in performance.
>>>>>>
>>>>>> Ok for trunk?
>>>>> Hum.  Why introduce a new --param and not simply key on
>>>>> flag_peel_loops instead?  That is
>>>>> enabled by default at -O3 and with FDO but you of course can control
>>>>> that in your targets
>>>>> post-option-processing hook.
>>>> You mean like this?
>>>> It's certainly a simpler patch, but I was just a bit hesitant of making this change for all targets :)
>>>> But I suppose it's a reasonable change.
>>> No, that change is backward.  What I said is that peeling is already
>>> conditional on
>>> flag_peel_loops and that is enabled by -O3.  So you want to disable
>>> flag_peel_loops for
>>> SVE instead in the target.
>> Sorry, I got confused by the similarly named functions.
>> I'm talking about try_unroll_loop_completely when run as part of canonicalize_induction_variables i.e. the "ivcanon" pass
>> (sorry about blaming cunroll here). This doesn't get called through the try_unroll_loops_completely path.
> Well, peeling gets disabled.  From your patch I see you want to
> disable "unrolling" when
> the number of loop iteration is not constant.  That is called peeling
> where we need to
> emit the loop exit test N times.
>
> Did you check your testcases with -fno-peel-loops?

-fno-peel-loops doesn't help in the testcases. The code that does this peeling (try_unroll_loop_completely)
can be called through two paths, only one of which is gated on flag_peel_loops.

Thanks,
Kyrill


>> try_unroll_loop_completely doesn't get disabled with -fno-peel-loops or -fno-unroll-loops.
>> Maybe disabling peeling inside try_unroll_loop_completely itself when !flag_peel_loops is viable?
>>
>> Thanks,
>> Kyrill
>>
>>>>> It might also make sense to have more fine-grained control for this
>>>>> and allow a target
>>>>> to say whether it wants to peel a specific loop or not when the
>>>>> middle-end thinks that
>>>>> would be profitable.
>>>> Can be worth looking at as a follow-up. Do you envisage the target analysing
>>>> the gimple statements of the loop to figure out its cost?
>>> Kind-of.  Sth like
>>>
>>>     bool targetm.peel_loop (struct loop *);
>>>
>>> I have no idea whether you can easily detect a SVE vectorized loop though.
>>> Maybe there's always a special IV or so (the mask?)
>>>
>>> Richard.
>>>
>>>> Thanks,
>>>> Kyrill
>>>>
>>>>
>>>> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
>>>>
>>>>           * tree-ssa-loop-ivcanon.c (try_unroll_loop_completely): Do not unroll
>>>>           loop when number of iterations is not known and flag_peel_loops is in
>>>>           effect.
>>>>
>>>> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
>>>>
>>>>           * gcc.target/aarch64/sve/unroll-1.c: New test.
>>>>

Reply | Threaded
Open this post in threaded view
|

Re: [PATCH][cunroll] Add unroll-known-loop-iterations-only param and use it in aarch64

Richard Biener-2
On Tue, Nov 13, 2018 at 10:15 AM Kyrill Tkachov
<[hidden email]> wrote:

>
> Hi Richard,
>
> On 13/11/18 08:24, Richard Biener wrote:
> > On Mon, Nov 12, 2018 at 7:20 PM Kyrill Tkachov
> > <[hidden email]> wrote:
> >>
> >> On 12/11/18 14:10, Richard Biener wrote:
> >>> On Fri, Nov 9, 2018 at 6:57 PM Kyrill Tkachov
> >>> <[hidden email]> wrote:
> >>>> On 09/11/18 12:18, Richard Biener wrote:
> >>>>> On Fri, Nov 9, 2018 at 11:47 AM Kyrill Tkachov
> >>>>> <[hidden email]> wrote:
> >>>>>> Hi all,
> >>>>>>
> >>>>>> In this testcase the codegen for VLA SVE is worse than it could be due to unrolling:
> >>>>>>
> >>>>>> fully_peel_me:
> >>>>>>            mov     x1, 5
> >>>>>>            ptrue   p1.d, all
> >>>>>>            whilelo p0.d, xzr, x1
> >>>>>>            ld1d    z0.d, p0/z, [x0]
> >>>>>>            fadd    z0.d, z0.d, z0.d
> >>>>>>            st1d    z0.d, p0, [x0]
> >>>>>>            cntd    x2
> >>>>>>            addvl   x3, x0, #1
> >>>>>>            whilelo p0.d, x2, x1
> >>>>>>            beq     .L1
> >>>>>>            ld1d    z0.d, p0/z, [x0, #1, mul vl]
> >>>>>>            fadd    z0.d, z0.d, z0.d
> >>>>>>            st1d    z0.d, p0, [x3]
> >>>>>>            cntw    x2
> >>>>>>            incb    x0, all, mul #2
> >>>>>>            whilelo p0.d, x2, x1
> >>>>>>            beq     .L1
> >>>>>>            ld1d    z0.d, p0/z, [x0]
> >>>>>>            fadd    z0.d, z0.d, z0.d
> >>>>>>            st1d    z0.d, p0, [x0]
> >>>>>> .L1:
> >>>>>>            ret
> >>>>>>
> >>>>>> In this case, due to the vector-length-agnostic nature of SVE the compiler doesn't know the loop iteration count.
> >>>>>> For such loops we don't want to unroll if we don't end up eliminating branches as this just bloats code size
> >>>>>> and hurts icache performance.
> >>>>>>
> >>>>>> This patch introduces a new unroll-known-loop-iterations-only param that disables cunroll when the loop iteration
> >>>>>> count is unknown (SCEV_NOT_KNOWN). This case occurs much more often for SVE VLA code, but it does help some
> >>>>>> Advanced SIMD cases as well where loops with an unknown iteration count are not unrolled when it doesn't eliminate
> >>>>>> the branches.
> >>>>>>
> >>>>>> So for the above testcase we generate now:
> >>>>>> fully_peel_me:
> >>>>>>            mov     x2, 5
> >>>>>>            mov     x3, x2
> >>>>>>            mov     x1, 0
> >>>>>>            whilelo p0.d, xzr, x2
> >>>>>>            ptrue   p1.d, all
> >>>>>> .L2:
> >>>>>>            ld1d    z0.d, p0/z, [x0, x1, lsl 3]
> >>>>>>            fadd    z0.d, z0.d, z0.d
> >>>>>>            st1d    z0.d, p0, [x0, x1, lsl 3]
> >>>>>>            incd    x1
> >>>>>>            whilelo p0.d, x1, x3
> >>>>>>            bne     .L2
> >>>>>>            ret
> >>>>>>
> >>>>>> Not perfect still, but it's preferable to the original code.
> >>>>>> The new param is enabled by default on aarch64 but disabled for other targets, leaving their behaviour unchanged
> >>>>>> (until other target people experiment with it and set it, if appropriate).
> >>>>>>
> >>>>>> Bootstrapped and tested on aarch64-none-linux-gnu.
> >>>>>> Benchmarked on SPEC2017 on a Cortex-A57 and there are no differences in performance.
> >>>>>>
> >>>>>> Ok for trunk?
> >>>>> Hum.  Why introduce a new --param and not simply key on
> >>>>> flag_peel_loops instead?  That is
> >>>>> enabled by default at -O3 and with FDO but you of course can control
> >>>>> that in your targets
> >>>>> post-option-processing hook.
> >>>> You mean like this?
> >>>> It's certainly a simpler patch, but I was just a bit hesitant of making this change for all targets :)
> >>>> But I suppose it's a reasonable change.
> >>> No, that change is backward.  What I said is that peeling is already
> >>> conditional on
> >>> flag_peel_loops and that is enabled by -O3.  So you want to disable
> >>> flag_peel_loops for
> >>> SVE instead in the target.
> >> Sorry, I got confused by the similarly named functions.
> >> I'm talking about try_unroll_loop_completely when run as part of canonicalize_induction_variables i.e. the "ivcanon" pass
> >> (sorry about blaming cunroll here). This doesn't get called through the try_unroll_loops_completely path.
> > Well, peeling gets disabled.  From your patch I see you want to
> > disable "unrolling" when
> > the number of loop iteration is not constant.  That is called peeling
> > where we need to
> > emit the loop exit test N times.
> >
> > Did you check your testcases with -fno-peel-loops?
>
> -fno-peel-loops doesn't help in the testcases. The code that does this peeling (try_unroll_loop_completely)
> can be called through two paths, only one of which is gated on flag_peel_loops.

I don't see the obvious here so I have to either sit down with a
non-SVE specific testcase
showing this, or I am misunderstanding the actual transform that you
want to avoid.
allow_peel is false when called from canonicalize_induction_variables.
There's the slight
chance that UL_NO_GROWTH lets through cases - is your case one of
that?  That is,
does the following help?

Index: gcc/tree-ssa-loop-ivcanon.c
===================================================================
--- gcc/tree-ssa-loop-ivcanon.c (revision 266056)
+++ gcc/tree-ssa-loop-ivcanon.c (working copy)
@@ -724,7 +724,7 @@ try_unroll_loop_completely (struct loop
     exit = NULL;

   /* See if we can improve our estimate by using recorded loop bounds.  */
-  if ((allow_peel || maxiter == 0 || ul == UL_NO_GROWTH)
+  if ((allow_peel || maxiter == 0)
       && maxiter >= 0
       && (!n_unroll_found || (unsigned HOST_WIDE_INT)maxiter < n_unroll))
     {

IIRC I allowed that case when adding allow_peel simply because it avoided some
testsuite regressions.  This means you eventually want to work on the
size estimate
of SVE style loops?

Richard.

> Thanks,
> Kyrill
>
>
> >> try_unroll_loop_completely doesn't get disabled with -fno-peel-loops or -fno-unroll-loops.
> >> Maybe disabling peeling inside try_unroll_loop_completely itself when !flag_peel_loops is viable?
> >>
> >> Thanks,
> >> Kyrill
> >>
> >>>>> It might also make sense to have more fine-grained control for this
> >>>>> and allow a target
> >>>>> to say whether it wants to peel a specific loop or not when the
> >>>>> middle-end thinks that
> >>>>> would be profitable.
> >>>> Can be worth looking at as a follow-up. Do you envisage the target analysing
> >>>> the gimple statements of the loop to figure out its cost?
> >>> Kind-of.  Sth like
> >>>
> >>>     bool targetm.peel_loop (struct loop *);
> >>>
> >>> I have no idea whether you can easily detect a SVE vectorized loop though.
> >>> Maybe there's always a special IV or so (the mask?)
> >>>
> >>> Richard.
> >>>
> >>>> Thanks,
> >>>> Kyrill
> >>>>
> >>>>
> >>>> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
> >>>>
> >>>>           * tree-ssa-loop-ivcanon.c (try_unroll_loop_completely): Do not unroll
> >>>>           loop when number of iterations is not known and flag_peel_loops is in
> >>>>           effect.
> >>>>
> >>>> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
> >>>>
> >>>>           * gcc.target/aarch64/sve/unroll-1.c: New test.
> >>>>
>
Reply | Threaded
Open this post in threaded view
|

Re: [PATCH][cunroll] Add unroll-known-loop-iterations-only param and use it in aarch64

Kyrill  Tkachov-2

On 13/11/18 09:28, Richard Biener wrote:

> On Tue, Nov 13, 2018 at 10:15 AM Kyrill Tkachov
> <[hidden email]> wrote:
>> Hi Richard,
>>
>> On 13/11/18 08:24, Richard Biener wrote:
>>> On Mon, Nov 12, 2018 at 7:20 PM Kyrill Tkachov
>>> <[hidden email]> wrote:
>>>> On 12/11/18 14:10, Richard Biener wrote:
>>>>> On Fri, Nov 9, 2018 at 6:57 PM Kyrill Tkachov
>>>>> <[hidden email]> wrote:
>>>>>> On 09/11/18 12:18, Richard Biener wrote:
>>>>>>> On Fri, Nov 9, 2018 at 11:47 AM Kyrill Tkachov
>>>>>>> <[hidden email]> wrote:
>>>>>>>> Hi all,
>>>>>>>>
>>>>>>>> In this testcase the codegen for VLA SVE is worse than it could be due to unrolling:
>>>>>>>>
>>>>>>>> fully_peel_me:
>>>>>>>>             mov     x1, 5
>>>>>>>>             ptrue   p1.d, all
>>>>>>>>             whilelo p0.d, xzr, x1
>>>>>>>>             ld1d    z0.d, p0/z, [x0]
>>>>>>>>             fadd    z0.d, z0.d, z0.d
>>>>>>>>             st1d    z0.d, p0, [x0]
>>>>>>>>             cntd    x2
>>>>>>>>             addvl   x3, x0, #1
>>>>>>>>             whilelo p0.d, x2, x1
>>>>>>>>             beq     .L1
>>>>>>>>             ld1d    z0.d, p0/z, [x0, #1, mul vl]
>>>>>>>>             fadd    z0.d, z0.d, z0.d
>>>>>>>>             st1d    z0.d, p0, [x3]
>>>>>>>>             cntw    x2
>>>>>>>>             incb    x0, all, mul #2
>>>>>>>>             whilelo p0.d, x2, x1
>>>>>>>>             beq     .L1
>>>>>>>>             ld1d    z0.d, p0/z, [x0]
>>>>>>>>             fadd    z0.d, z0.d, z0.d
>>>>>>>>             st1d    z0.d, p0, [x0]
>>>>>>>> .L1:
>>>>>>>>             ret
>>>>>>>>
>>>>>>>> In this case, due to the vector-length-agnostic nature of SVE the compiler doesn't know the loop iteration count.
>>>>>>>> For such loops we don't want to unroll if we don't end up eliminating branches as this just bloats code size
>>>>>>>> and hurts icache performance.
>>>>>>>>
>>>>>>>> This patch introduces a new unroll-known-loop-iterations-only param that disables cunroll when the loop iteration
>>>>>>>> count is unknown (SCEV_NOT_KNOWN). This case occurs much more often for SVE VLA code, but it does help some
>>>>>>>> Advanced SIMD cases as well where loops with an unknown iteration count are not unrolled when it doesn't eliminate
>>>>>>>> the branches.
>>>>>>>>
>>>>>>>> So for the above testcase we generate now:
>>>>>>>> fully_peel_me:
>>>>>>>>             mov     x2, 5
>>>>>>>>             mov     x3, x2
>>>>>>>>             mov     x1, 0
>>>>>>>>             whilelo p0.d, xzr, x2
>>>>>>>>             ptrue   p1.d, all
>>>>>>>> .L2:
>>>>>>>>             ld1d    z0.d, p0/z, [x0, x1, lsl 3]
>>>>>>>>             fadd    z0.d, z0.d, z0.d
>>>>>>>>             st1d    z0.d, p0, [x0, x1, lsl 3]
>>>>>>>>             incd    x1
>>>>>>>>             whilelo p0.d, x1, x3
>>>>>>>>             bne     .L2
>>>>>>>>             ret
>>>>>>>>
>>>>>>>> Not perfect still, but it's preferable to the original code.
>>>>>>>> The new param is enabled by default on aarch64 but disabled for other targets, leaving their behaviour unchanged
>>>>>>>> (until other target people experiment with it and set it, if appropriate).
>>>>>>>>
>>>>>>>> Bootstrapped and tested on aarch64-none-linux-gnu.
>>>>>>>> Benchmarked on SPEC2017 on a Cortex-A57 and there are no differences in performance.
>>>>>>>>
>>>>>>>> Ok for trunk?
>>>>>>> Hum.  Why introduce a new --param and not simply key on
>>>>>>> flag_peel_loops instead?  That is
>>>>>>> enabled by default at -O3 and with FDO but you of course can control
>>>>>>> that in your targets
>>>>>>> post-option-processing hook.
>>>>>> You mean like this?
>>>>>> It's certainly a simpler patch, but I was just a bit hesitant of making this change for all targets :)
>>>>>> But I suppose it's a reasonable change.
>>>>> No, that change is backward.  What I said is that peeling is already
>>>>> conditional on
>>>>> flag_peel_loops and that is enabled by -O3.  So you want to disable
>>>>> flag_peel_loops for
>>>>> SVE instead in the target.
>>>> Sorry, I got confused by the similarly named functions.
>>>> I'm talking about try_unroll_loop_completely when run as part of canonicalize_induction_variables i.e. the "ivcanon" pass
>>>> (sorry about blaming cunroll here). This doesn't get called through the try_unroll_loops_completely path.
>>> Well, peeling gets disabled.  From your patch I see you want to
>>> disable "unrolling" when
>>> the number of loop iteration is not constant.  That is called peeling
>>> where we need to
>>> emit the loop exit test N times.
>>>
>>> Did you check your testcases with -fno-peel-loops?
>> -fno-peel-loops doesn't help in the testcases. The code that does this peeling (try_unroll_loop_completely)
>> can be called through two paths, only one of which is gated on flag_peel_loops.
> I don't see the obvious here so I have to either sit down with a
> non-SVE specific testcase
> showing this, or I am misunderstanding the actual transform that you
> want to avoid.
> allow_peel is false when called from canonicalize_induction_variables.
> There's the slight
> chance that UL_NO_GROWTH lets through cases - is your case one of
> that?  That is,
> does the following help?
>
> Index: gcc/tree-ssa-loop-ivcanon.c
> ===================================================================
> --- gcc/tree-ssa-loop-ivcanon.c (revision 266056)
> +++ gcc/tree-ssa-loop-ivcanon.c (working copy)
> @@ -724,7 +724,7 @@ try_unroll_loop_completely (struct loop
>       exit = NULL;
>
>     /* See if we can improve our estimate by using recorded loop bounds.  */
> -  if ((allow_peel || maxiter == 0 || ul == UL_NO_GROWTH)
> +  if ((allow_peel || maxiter == 0)
>         && maxiter >= 0
>         && (!n_unroll_found || (unsigned HOST_WIDE_INT)maxiter < n_unroll))
>       {
>
> IIRC I allowed that case when adding allow_peel simply because it avoided some
> testsuite regressions.  This means you eventually want to work on the
> size estimate
> of SVE style loops?

This doesn't help.

Sorry if we're talking over each other here, I'm not very familiar with this area :(
For this loop:
   for (int i = 0; i < 5; i++)
     x[i] = x[i] * 2;

For normal vectorisation (e.g. AArch64 NEON) we know the exact number of executions of the loop latch.
This gets fully unrolled as:
         ldp     q2, q1, [x0]
         ldr     d0, [x0, 32]
         fadd    v2.2d, v2.2d, v2.2d
         fadd    v1.2d, v1.2d, v1.2d
         fadd    d0, d0, d0
         stp     q2, q1, [x0]
         str     d0, [x0, 32]

For vector length-agnostic SVE vectorisation we don't as we don't know the number of elements we process
with each loop iteration. So the NEON unrolling becomes SVE peeling I guess.
Note that the number of iterations in SVE is still "constant", just not known at compile-time.
In this case peeling doesn't eliminate any branches and only serves to bloat code size.

Kyrill

> Richard.
>
>> Thanks,
>> Kyrill
>>
>>
>>>> try_unroll_loop_completely doesn't get disabled with -fno-peel-loops or -fno-unroll-loops.
>>>> Maybe disabling peeling inside try_unroll_loop_completely itself when !flag_peel_loops is viable?
>>>>
>>>> Thanks,
>>>> Kyrill
>>>>
>>>>>>> It might also make sense to have more fine-grained control for this
>>>>>>> and allow a target
>>>>>>> to say whether it wants to peel a specific loop or not when the
>>>>>>> middle-end thinks that
>>>>>>> would be profitable.
>>>>>> Can be worth looking at as a follow-up. Do you envisage the target analysing
>>>>>> the gimple statements of the loop to figure out its cost?
>>>>> Kind-of.  Sth like
>>>>>
>>>>>      bool targetm.peel_loop (struct loop *);
>>>>>
>>>>> I have no idea whether you can easily detect a SVE vectorized loop though.
>>>>> Maybe there's always a special IV or so (the mask?)
>>>>>
>>>>> Richard.
>>>>>
>>>>>> Thanks,
>>>>>> Kyrill
>>>>>>
>>>>>>
>>>>>> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
>>>>>>
>>>>>>            * tree-ssa-loop-ivcanon.c (try_unroll_loop_completely): Do not unroll
>>>>>>            loop when number of iterations is not known and flag_peel_loops is in
>>>>>>            effect.
>>>>>>
>>>>>> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
>>>>>>
>>>>>>            * gcc.target/aarch64/sve/unroll-1.c: New test.
>>>>>>

Reply | Threaded
Open this post in threaded view
|

Re: [PATCH][cunroll] Add unroll-known-loop-iterations-only param and use it in aarch64

Richard Biener-2
On Tue, Nov 13, 2018 at 10:48 AM Kyrill Tkachov
<[hidden email]> wrote:

>
>
> On 13/11/18 09:28, Richard Biener wrote:
> > On Tue, Nov 13, 2018 at 10:15 AM Kyrill Tkachov
> > <[hidden email]> wrote:
> >> Hi Richard,
> >>
> >> On 13/11/18 08:24, Richard Biener wrote:
> >>> On Mon, Nov 12, 2018 at 7:20 PM Kyrill Tkachov
> >>> <[hidden email]> wrote:
> >>>> On 12/11/18 14:10, Richard Biener wrote:
> >>>>> On Fri, Nov 9, 2018 at 6:57 PM Kyrill Tkachov
> >>>>> <[hidden email]> wrote:
> >>>>>> On 09/11/18 12:18, Richard Biener wrote:
> >>>>>>> On Fri, Nov 9, 2018 at 11:47 AM Kyrill Tkachov
> >>>>>>> <[hidden email]> wrote:
> >>>>>>>> Hi all,
> >>>>>>>>
> >>>>>>>> In this testcase the codegen for VLA SVE is worse than it could be due to unrolling:
> >>>>>>>>
> >>>>>>>> fully_peel_me:
> >>>>>>>>             mov     x1, 5
> >>>>>>>>             ptrue   p1.d, all
> >>>>>>>>             whilelo p0.d, xzr, x1
> >>>>>>>>             ld1d    z0.d, p0/z, [x0]
> >>>>>>>>             fadd    z0.d, z0.d, z0.d
> >>>>>>>>             st1d    z0.d, p0, [x0]
> >>>>>>>>             cntd    x2
> >>>>>>>>             addvl   x3, x0, #1
> >>>>>>>>             whilelo p0.d, x2, x1
> >>>>>>>>             beq     .L1
> >>>>>>>>             ld1d    z0.d, p0/z, [x0, #1, mul vl]
> >>>>>>>>             fadd    z0.d, z0.d, z0.d
> >>>>>>>>             st1d    z0.d, p0, [x3]
> >>>>>>>>             cntw    x2
> >>>>>>>>             incb    x0, all, mul #2
> >>>>>>>>             whilelo p0.d, x2, x1
> >>>>>>>>             beq     .L1
> >>>>>>>>             ld1d    z0.d, p0/z, [x0]
> >>>>>>>>             fadd    z0.d, z0.d, z0.d
> >>>>>>>>             st1d    z0.d, p0, [x0]
> >>>>>>>> .L1:
> >>>>>>>>             ret
> >>>>>>>>
> >>>>>>>> In this case, due to the vector-length-agnostic nature of SVE the compiler doesn't know the loop iteration count.
> >>>>>>>> For such loops we don't want to unroll if we don't end up eliminating branches as this just bloats code size
> >>>>>>>> and hurts icache performance.
> >>>>>>>>
> >>>>>>>> This patch introduces a new unroll-known-loop-iterations-only param that disables cunroll when the loop iteration
> >>>>>>>> count is unknown (SCEV_NOT_KNOWN). This case occurs much more often for SVE VLA code, but it does help some
> >>>>>>>> Advanced SIMD cases as well where loops with an unknown iteration count are not unrolled when it doesn't eliminate
> >>>>>>>> the branches.
> >>>>>>>>
> >>>>>>>> So for the above testcase we generate now:
> >>>>>>>> fully_peel_me:
> >>>>>>>>             mov     x2, 5
> >>>>>>>>             mov     x3, x2
> >>>>>>>>             mov     x1, 0
> >>>>>>>>             whilelo p0.d, xzr, x2
> >>>>>>>>             ptrue   p1.d, all
> >>>>>>>> .L2:
> >>>>>>>>             ld1d    z0.d, p0/z, [x0, x1, lsl 3]
> >>>>>>>>             fadd    z0.d, z0.d, z0.d
> >>>>>>>>             st1d    z0.d, p0, [x0, x1, lsl 3]
> >>>>>>>>             incd    x1
> >>>>>>>>             whilelo p0.d, x1, x3
> >>>>>>>>             bne     .L2
> >>>>>>>>             ret
> >>>>>>>>
> >>>>>>>> Not perfect still, but it's preferable to the original code.
> >>>>>>>> The new param is enabled by default on aarch64 but disabled for other targets, leaving their behaviour unchanged
> >>>>>>>> (until other target people experiment with it and set it, if appropriate).
> >>>>>>>>
> >>>>>>>> Bootstrapped and tested on aarch64-none-linux-gnu.
> >>>>>>>> Benchmarked on SPEC2017 on a Cortex-A57 and there are no differences in performance.
> >>>>>>>>
> >>>>>>>> Ok for trunk?
> >>>>>>> Hum.  Why introduce a new --param and not simply key on
> >>>>>>> flag_peel_loops instead?  That is
> >>>>>>> enabled by default at -O3 and with FDO but you of course can control
> >>>>>>> that in your targets
> >>>>>>> post-option-processing hook.
> >>>>>> You mean like this?
> >>>>>> It's certainly a simpler patch, but I was just a bit hesitant of making this change for all targets :)
> >>>>>> But I suppose it's a reasonable change.
> >>>>> No, that change is backward.  What I said is that peeling is already
> >>>>> conditional on
> >>>>> flag_peel_loops and that is enabled by -O3.  So you want to disable
> >>>>> flag_peel_loops for
> >>>>> SVE instead in the target.
> >>>> Sorry, I got confused by the similarly named functions.
> >>>> I'm talking about try_unroll_loop_completely when run as part of canonicalize_induction_variables i.e. the "ivcanon" pass
> >>>> (sorry about blaming cunroll here). This doesn't get called through the try_unroll_loops_completely path.
> >>> Well, peeling gets disabled.  From your patch I see you want to
> >>> disable "unrolling" when
> >>> the number of loop iteration is not constant.  That is called peeling
> >>> where we need to
> >>> emit the loop exit test N times.
> >>>
> >>> Did you check your testcases with -fno-peel-loops?
> >> -fno-peel-loops doesn't help in the testcases. The code that does this peeling (try_unroll_loop_completely)
> >> can be called through two paths, only one of which is gated on flag_peel_loops.
> > I don't see the obvious here so I have to either sit down with a
> > non-SVE specific testcase
> > showing this, or I am misunderstanding the actual transform that you
> > want to avoid.
> > allow_peel is false when called from canonicalize_induction_variables.
> > There's the slight
> > chance that UL_NO_GROWTH lets through cases - is your case one of
> > that?  That is,
> > does the following help?
> >
> > Index: gcc/tree-ssa-loop-ivcanon.c
> > ===================================================================
> > --- gcc/tree-ssa-loop-ivcanon.c (revision 266056)
> > +++ gcc/tree-ssa-loop-ivcanon.c (working copy)
> > @@ -724,7 +724,7 @@ try_unroll_loop_completely (struct loop
> >       exit = NULL;
> >
> >     /* See if we can improve our estimate by using recorded loop bounds.  */
> > -  if ((allow_peel || maxiter == 0 || ul == UL_NO_GROWTH)
> > +  if ((allow_peel || maxiter == 0)
> >         && maxiter >= 0
> >         && (!n_unroll_found || (unsigned HOST_WIDE_INT)maxiter < n_unroll))
> >       {
> >
> > IIRC I allowed that case when adding allow_peel simply because it avoided some
> > testsuite regressions.  This means you eventually want to work on the
> > size estimate
> > of SVE style loops?
>
> This doesn't help.
>
> Sorry if we're talking over each other here, I'm not very familiar with this area :(
> For this loop:
>    for (int i = 0; i < 5; i++)
>      x[i] = x[i] * 2;
>
> For normal vectorisation (e.g. AArch64 NEON) we know the exact number of executions of the loop latch.
> This gets fully unrolled as:
>          ldp     q2, q1, [x0]
>          ldr     d0, [x0, 32]
>          fadd    v2.2d, v2.2d, v2.2d
>          fadd    v1.2d, v1.2d, v1.2d
>          fadd    d0, d0, d0
>          stp     q2, q1, [x0]
>          str     d0, [x0, 32]
>
> For vector length-agnostic SVE vectorisation we don't as we don't know the number of elements we process
> with each loop iteration. So the NEON unrolling becomes SVE peeling I guess.
> Note that the number of iterations in SVE is still "constant", just not known at compile-time.
> In this case peeling doesn't eliminate any branches and only serves to bloat code size.

So I sat down with a cross and indeed for the late unrolling pass we
simply pass in allow_peel == true
given that try_unroll_loop_completely also performs peeling (and not
just try_peel_loops which guards
itself with flag_peel_loops).  That means instead of the above the
below fixes this (with -fno-peel-loops):

Index: gcc/tree-ssa-loop-ivcanon.c
===================================================================
--- gcc/tree-ssa-loop-ivcanon.c (revision 266072)
+++ gcc/tree-ssa-loop-ivcanon.c (working copy)
@@ -724,7 +724,7 @@ try_unroll_loop_completely (struct loop
     exit = NULL;

   /* See if we can improve our estimate by using recorded loop bounds.  */
-  if ((allow_peel || maxiter == 0 || ul == UL_NO_GROWTH)
+  if (((allow_peel && flag_peel_loops) || maxiter == 0 || ul == UL_NO_GROWTH)
       && maxiter >= 0
       && (!n_unroll_found || (unsigned HOST_WIDE_INT)maxiter < n_unroll))
     {

there might be quite some testsuite fallout since flag_peel_loops is
only enabled at -O3+,
but one has to double-check.  As said, a per-loop target control
whether loop peeling is
desirable would be an improvement I guess (apart from generally
disabling peeling for aarch64).
I suppose you can benchmark that together with the above fix.

Richard.

> Kyrill
>
> > Richard.
> >
> >> Thanks,
> >> Kyrill
> >>
> >>
> >>>> try_unroll_loop_completely doesn't get disabled with -fno-peel-loops or -fno-unroll-loops.
> >>>> Maybe disabling peeling inside try_unroll_loop_completely itself when !flag_peel_loops is viable?
> >>>>
> >>>> Thanks,
> >>>> Kyrill
> >>>>
> >>>>>>> It might also make sense to have more fine-grained control for this
> >>>>>>> and allow a target
> >>>>>>> to say whether it wants to peel a specific loop or not when the
> >>>>>>> middle-end thinks that
> >>>>>>> would be profitable.
> >>>>>> Can be worth looking at as a follow-up. Do you envisage the target analysing
> >>>>>> the gimple statements of the loop to figure out its cost?
> >>>>> Kind-of.  Sth like
> >>>>>
> >>>>>      bool targetm.peel_loop (struct loop *);
> >>>>>
> >>>>> I have no idea whether you can easily detect a SVE vectorized loop though.
> >>>>> Maybe there's always a special IV or so (the mask?)
> >>>>>
> >>>>> Richard.
> >>>>>
> >>>>>> Thanks,
> >>>>>> Kyrill
> >>>>>>
> >>>>>>
> >>>>>> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
> >>>>>>
> >>>>>>            * tree-ssa-loop-ivcanon.c (try_unroll_loop_completely): Do not unroll
> >>>>>>            loop when number of iterations is not known and flag_peel_loops is in
> >>>>>>            effect.
> >>>>>>
> >>>>>> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
> >>>>>>
> >>>>>>            * gcc.target/aarch64/sve/unroll-1.c: New test.
> >>>>>>
>
Reply | Threaded
Open this post in threaded view
|

Re: [PATCH][cunroll] Add unroll-known-loop-iterations-only param and use it in aarch64

Richard Biener-2
On Tue, Nov 13, 2018 at 3:33 PM Richard Biener
<[hidden email]> wrote:

>
> On Tue, Nov 13, 2018 at 10:48 AM Kyrill Tkachov
> <[hidden email]> wrote:
> >
> >
> > On 13/11/18 09:28, Richard Biener wrote:
> > > On Tue, Nov 13, 2018 at 10:15 AM Kyrill Tkachov
> > > <[hidden email]> wrote:
> > >> Hi Richard,
> > >>
> > >> On 13/11/18 08:24, Richard Biener wrote:
> > >>> On Mon, Nov 12, 2018 at 7:20 PM Kyrill Tkachov
> > >>> <[hidden email]> wrote:
> > >>>> On 12/11/18 14:10, Richard Biener wrote:
> > >>>>> On Fri, Nov 9, 2018 at 6:57 PM Kyrill Tkachov
> > >>>>> <[hidden email]> wrote:
> > >>>>>> On 09/11/18 12:18, Richard Biener wrote:
> > >>>>>>> On Fri, Nov 9, 2018 at 11:47 AM Kyrill Tkachov
> > >>>>>>> <[hidden email]> wrote:
> > >>>>>>>> Hi all,
> > >>>>>>>>
> > >>>>>>>> In this testcase the codegen for VLA SVE is worse than it could be due to unrolling:
> > >>>>>>>>
> > >>>>>>>> fully_peel_me:
> > >>>>>>>>             mov     x1, 5
> > >>>>>>>>             ptrue   p1.d, all
> > >>>>>>>>             whilelo p0.d, xzr, x1
> > >>>>>>>>             ld1d    z0.d, p0/z, [x0]
> > >>>>>>>>             fadd    z0.d, z0.d, z0.d
> > >>>>>>>>             st1d    z0.d, p0, [x0]
> > >>>>>>>>             cntd    x2
> > >>>>>>>>             addvl   x3, x0, #1
> > >>>>>>>>             whilelo p0.d, x2, x1
> > >>>>>>>>             beq     .L1
> > >>>>>>>>             ld1d    z0.d, p0/z, [x0, #1, mul vl]
> > >>>>>>>>             fadd    z0.d, z0.d, z0.d
> > >>>>>>>>             st1d    z0.d, p0, [x3]
> > >>>>>>>>             cntw    x2
> > >>>>>>>>             incb    x0, all, mul #2
> > >>>>>>>>             whilelo p0.d, x2, x1
> > >>>>>>>>             beq     .L1
> > >>>>>>>>             ld1d    z0.d, p0/z, [x0]
> > >>>>>>>>             fadd    z0.d, z0.d, z0.d
> > >>>>>>>>             st1d    z0.d, p0, [x0]
> > >>>>>>>> .L1:
> > >>>>>>>>             ret
> > >>>>>>>>
> > >>>>>>>> In this case, due to the vector-length-agnostic nature of SVE the compiler doesn't know the loop iteration count.
> > >>>>>>>> For such loops we don't want to unroll if we don't end up eliminating branches as this just bloats code size
> > >>>>>>>> and hurts icache performance.
> > >>>>>>>>
> > >>>>>>>> This patch introduces a new unroll-known-loop-iterations-only param that disables cunroll when the loop iteration
> > >>>>>>>> count is unknown (SCEV_NOT_KNOWN). This case occurs much more often for SVE VLA code, but it does help some
> > >>>>>>>> Advanced SIMD cases as well where loops with an unknown iteration count are not unrolled when it doesn't eliminate
> > >>>>>>>> the branches.
> > >>>>>>>>
> > >>>>>>>> So for the above testcase we generate now:
> > >>>>>>>> fully_peel_me:
> > >>>>>>>>             mov     x2, 5
> > >>>>>>>>             mov     x3, x2
> > >>>>>>>>             mov     x1, 0
> > >>>>>>>>             whilelo p0.d, xzr, x2
> > >>>>>>>>             ptrue   p1.d, all
> > >>>>>>>> .L2:
> > >>>>>>>>             ld1d    z0.d, p0/z, [x0, x1, lsl 3]
> > >>>>>>>>             fadd    z0.d, z0.d, z0.d
> > >>>>>>>>             st1d    z0.d, p0, [x0, x1, lsl 3]
> > >>>>>>>>             incd    x1
> > >>>>>>>>             whilelo p0.d, x1, x3
> > >>>>>>>>             bne     .L2
> > >>>>>>>>             ret
> > >>>>>>>>
> > >>>>>>>> Not perfect still, but it's preferable to the original code.
> > >>>>>>>> The new param is enabled by default on aarch64 but disabled for other targets, leaving their behaviour unchanged
> > >>>>>>>> (until other target people experiment with it and set it, if appropriate).
> > >>>>>>>>
> > >>>>>>>> Bootstrapped and tested on aarch64-none-linux-gnu.
> > >>>>>>>> Benchmarked on SPEC2017 on a Cortex-A57 and there are no differences in performance.
> > >>>>>>>>
> > >>>>>>>> Ok for trunk?
> > >>>>>>> Hum.  Why introduce a new --param and not simply key on
> > >>>>>>> flag_peel_loops instead?  That is
> > >>>>>>> enabled by default at -O3 and with FDO but you of course can control
> > >>>>>>> that in your targets
> > >>>>>>> post-option-processing hook.
> > >>>>>> You mean like this?
> > >>>>>> It's certainly a simpler patch, but I was just a bit hesitant of making this change for all targets :)
> > >>>>>> But I suppose it's a reasonable change.
> > >>>>> No, that change is backward.  What I said is that peeling is already
> > >>>>> conditional on
> > >>>>> flag_peel_loops and that is enabled by -O3.  So you want to disable
> > >>>>> flag_peel_loops for
> > >>>>> SVE instead in the target.
> > >>>> Sorry, I got confused by the similarly named functions.
> > >>>> I'm talking about try_unroll_loop_completely when run as part of canonicalize_induction_variables i.e. the "ivcanon" pass
> > >>>> (sorry about blaming cunroll here). This doesn't get called through the try_unroll_loops_completely path.
> > >>> Well, peeling gets disabled.  From your patch I see you want to
> > >>> disable "unrolling" when
> > >>> the number of loop iteration is not constant.  That is called peeling
> > >>> where we need to
> > >>> emit the loop exit test N times.
> > >>>
> > >>> Did you check your testcases with -fno-peel-loops?
> > >> -fno-peel-loops doesn't help in the testcases. The code that does this peeling (try_unroll_loop_completely)
> > >> can be called through two paths, only one of which is gated on flag_peel_loops.
> > > I don't see the obvious here so I have to either sit down with a
> > > non-SVE specific testcase
> > > showing this, or I am misunderstanding the actual transform that you
> > > want to avoid.
> > > allow_peel is false when called from canonicalize_induction_variables.
> > > There's the slight
> > > chance that UL_NO_GROWTH lets through cases - is your case one of
> > > that?  That is,
> > > does the following help?
> > >
> > > Index: gcc/tree-ssa-loop-ivcanon.c
> > > ===================================================================
> > > --- gcc/tree-ssa-loop-ivcanon.c (revision 266056)
> > > +++ gcc/tree-ssa-loop-ivcanon.c (working copy)
> > > @@ -724,7 +724,7 @@ try_unroll_loop_completely (struct loop
> > >       exit = NULL;
> > >
> > >     /* See if we can improve our estimate by using recorded loop bounds.  */
> > > -  if ((allow_peel || maxiter == 0 || ul == UL_NO_GROWTH)
> > > +  if ((allow_peel || maxiter == 0)
> > >         && maxiter >= 0
> > >         && (!n_unroll_found || (unsigned HOST_WIDE_INT)maxiter < n_unroll))
> > >       {
> > >
> > > IIRC I allowed that case when adding allow_peel simply because it avoided some
> > > testsuite regressions.  This means you eventually want to work on the
> > > size estimate
> > > of SVE style loops?
> >
> > This doesn't help.
> >
> > Sorry if we're talking over each other here, I'm not very familiar with this area :(
> > For this loop:
> >    for (int i = 0; i < 5; i++)
> >      x[i] = x[i] * 2;
> >
> > For normal vectorisation (e.g. AArch64 NEON) we know the exact number of executions of the loop latch.
> > This gets fully unrolled as:
> >          ldp     q2, q1, [x0]
> >          ldr     d0, [x0, 32]
> >          fadd    v2.2d, v2.2d, v2.2d
> >          fadd    v1.2d, v1.2d, v1.2d
> >          fadd    d0, d0, d0
> >          stp     q2, q1, [x0]
> >          str     d0, [x0, 32]
> >
> > For vector length-agnostic SVE vectorisation we don't as we don't know the number of elements we process
> > with each loop iteration. So the NEON unrolling becomes SVE peeling I guess.
> > Note that the number of iterations in SVE is still "constant", just not known at compile-time.
> > In this case peeling doesn't eliminate any branches and only serves to bloat code size.
>
> So I sat down with a cross and indeed for the late unrolling pass we
> simply pass in allow_peel == true
> given that try_unroll_loop_completely also performs peeling (and not
> just try_peel_loops which guards
> itself with flag_peel_loops).  That means instead of the above the
> below fixes this (with -fno-peel-loops):
>
> Index: gcc/tree-ssa-loop-ivcanon.c
> ===================================================================
> --- gcc/tree-ssa-loop-ivcanon.c (revision 266072)
> +++ gcc/tree-ssa-loop-ivcanon.c (working copy)
> @@ -724,7 +724,7 @@ try_unroll_loop_completely (struct loop
>      exit = NULL;
>
>    /* See if we can improve our estimate by using recorded loop bounds.  */
> -  if ((allow_peel || maxiter == 0 || ul == UL_NO_GROWTH)
> +  if (((allow_peel && flag_peel_loops) || maxiter == 0 || ul == UL_NO_GROWTH)
>        && maxiter >= 0
>        && (!n_unroll_found || (unsigned HOST_WIDE_INT)maxiter < n_unroll))
>      {
>
> there might be quite some testsuite fallout since flag_peel_loops is
> only enabled at -O3+,
> but one has to double-check.  As said, a per-loop target control
> whether loop peeling is
> desirable would be an improvement I guess (apart from generally
> disabling peeling for aarch64).
> I suppose you can benchmark that together with the above fix.

Oh - and I completely forgot about loop->unroll which the vectorizer
could set (for SVE loops) to 1 which disables any unrolling.

Richard.

> Richard.
>
> > Kyrill
> >
> > > Richard.
> > >
> > >> Thanks,
> > >> Kyrill
> > >>
> > >>
> > >>>> try_unroll_loop_completely doesn't get disabled with -fno-peel-loops or -fno-unroll-loops.
> > >>>> Maybe disabling peeling inside try_unroll_loop_completely itself when !flag_peel_loops is viable?
> > >>>>
> > >>>> Thanks,
> > >>>> Kyrill
> > >>>>
> > >>>>>>> It might also make sense to have more fine-grained control for this
> > >>>>>>> and allow a target
> > >>>>>>> to say whether it wants to peel a specific loop or not when the
> > >>>>>>> middle-end thinks that
> > >>>>>>> would be profitable.
> > >>>>>> Can be worth looking at as a follow-up. Do you envisage the target analysing
> > >>>>>> the gimple statements of the loop to figure out its cost?
> > >>>>> Kind-of.  Sth like
> > >>>>>
> > >>>>>      bool targetm.peel_loop (struct loop *);
> > >>>>>
> > >>>>> I have no idea whether you can easily detect a SVE vectorized loop though.
> > >>>>> Maybe there's always a special IV or so (the mask?)
> > >>>>>
> > >>>>> Richard.
> > >>>>>
> > >>>>>> Thanks,
> > >>>>>> Kyrill
> > >>>>>>
> > >>>>>>
> > >>>>>> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
> > >>>>>>
> > >>>>>>            * tree-ssa-loop-ivcanon.c (try_unroll_loop_completely): Do not unroll
> > >>>>>>            loop when number of iterations is not known and flag_peel_loops is in
> > >>>>>>            effect.
> > >>>>>>
> > >>>>>> 2018-11-09  Kyrylo Tkachov  <[hidden email]>
> > >>>>>>
> > >>>>>>            * gcc.target/aarch64/sve/unroll-1.c: New test.
> > >>>>>>
> >
Reply | Threaded
Open this post in threaded view
|

[PATCH] Disable unrolling for loops vectorised with non-constant VF (was: [PATCH][cunroll] Add unroll-known-loop-iterations-only param and use it in aarch64)

Kyrill  Tkachov-2
Hi all,

This is an alternative to https://gcc.gnu.org/ml/gcc-patches/2018-11/msg00694.html
As richi suggested, this disables unrolling of loops vectorised with variable-length SVE
in the vectoriser itself through the loop->unroll member.

It took me a few tries to get it right, as it needs to be set to '1' to disable unrolling,
the rationale for that mechanism is described in the comment in cfgloop.h.

Bootstrapped and tested on aarch64-none-linux-gnu.

Is this ok for trunk?

Thanks,
Kyrill

2018-11-15  Kyrylo Tkachov  <[hidden email]>

     * tree-vect-loop.c (vect_transform_loop): Disable further unrolling
     of the loop if vf is non-constant.

2018-11-15  Kyrylo Tkachov  <[hidden email]>

     * gcc.target/aarch64/sve/unroll-1.c: New test.

unroll-vect.patch (1K) Download Attachment
Reply | Threaded
Open this post in threaded view
|

Re: [PATCH] Disable unrolling for loops vectorised with non-constant VF (was: [PATCH][cunroll] Add unroll-known-loop-iterations-only param and use it in aarch64)

Richard Biener-2
On Fri, Nov 16, 2018 at 7:02 PM Kyrill Tkachov
<[hidden email]> wrote:

>
> Hi all,
>
> This is an alternative to https://gcc.gnu.org/ml/gcc-patches/2018-11/msg00694.html
> As richi suggested, this disables unrolling of loops vectorised with variable-length SVE
> in the vectoriser itself through the loop->unroll member.
>
> It took me a few tries to get it right, as it needs to be set to '1' to disable unrolling,
> the rationale for that mechanism is described in the comment in cfgloop.h.
>
> Bootstrapped and tested on aarch64-none-linux-gnu.
>
> Is this ok for trunk?

OK.

Richard.

> Thanks,
> Kyrill
>
> 2018-11-15  Kyrylo Tkachov  <[hidden email]>
>
>      * tree-vect-loop.c (vect_transform_loop): Disable further unrolling
>      of the loop if vf is non-constant.
>
> 2018-11-15  Kyrylo Tkachov  <[hidden email]>
>
>      * gcc.target/aarch64/sve/unroll-1.c: New test.