Re: [PATCH 00/22] arm: New framework for MVE intrinsics

[Christophe Lyon <christophe.lyon@arm.com>]

On 5/2/23 17:04, Christophe Lyon via Gcc-patches wrote:
> 
> 
> On 5/2/23 11:18, Kyrylo Tkachov wrote:
>> Hi Christophe,
>>
>>> -----Original Message-----
>>> From: Christophe Lyon <christophe.lyon@arm.com>
>>> Sent: Tuesday, April 18, 2023 2:46 PM
>>> To: gcc-patches@gcc.gnu.org; Kyrylo Tkachov <Kyrylo.Tkachov@arm.com>;
>>> Richard Earnshaw <Richard.Earnshaw@arm.com>; Richard Sandiford
>>> <Richard.Sandiford@arm.com>
>>> Cc: Christophe Lyon <Christophe.Lyon@arm.com>
>>> Subject: [PATCH 00/22] arm: New framework for MVE intrinsics
>>>
>>> Hi,
>>>
>>> This is the beginning of a long patch series to change the way Arm MVE
>>> intrinsics are implemented. The goal is to get rid of arm_mve.h, which
>>> takes a long time to parse and compile.
>>>
>>
>> Thanks for doing this. It is a significant improvement to the MVE 
>> intrinsics and should address some of the biggest maintainability and 
>> scalability issues we have in that area.
>> I'll be going through the patches one-by-one (I've looked at these 
>> offline already before), but the approach looks good to me at a high 
>> level.
>>
>> My hope is that we'll move all the intrinsics, including the Neon ones 
>> to use this framework in the future, but getting the framework in 
>> place first is a good major first step in that direction.
>>
> 
> Indeed. Ideally we'd probably want to make this framework more generic 
> so that it supports aarch64 SVE, arm MVE and Neon, but that can be done 
> later. I tried to highlight the differences I noticed compared to SVE, 
> so that it helps us think what needs to be specialized for different 
> targets, as opposed to what is already generic enough.
> 
> Thanks,
> 
> Christophe
> 

Thank you Kyrill for the quick review.
I've pushed the series with the minor changes you requested.

I am going to prepare the next batch :-)

Thanks,

Christophe

>> Thanks,
>> Kyrill
>>
>>> Roughly speaking, it's about using a framework very similar to what is
>>> implemented for AArch64/SVE intrinsics. I haven't converted all the
>>> intrinsics yet, but I think it would be good to start the conversion
>>> when stage-1 reopens.
>>>
>>> * Factorizing names
>>> One of the main implementation differences I noticed between SVE and
>>> MVE is that mve.md provides only full builtin names at the moment, and
>>> makes almost no use of "parameterized names"
>>> (https://gcc.gnu.org/onlinedocs/gccint/Parameterized-
>>> Names.html#Parameterized-Names).
>>>
>>> Without this, we'd need the builtin expander to use a large
>>> switch/case of the form:
>>>
>>> switch (code)
>>> case VADDQ_S: insn_code = code_for_mve_vaddq_s (...)
>>> case VADDQ_U: insn_code = code_for_mve_vaddq_u (...)
>>> case VSUBQ_S: insn_code = code_for_mve_vsubq_s (...)
>>> case VSUBQ_U: insn_code = code_for_mve_vsubq_u (...)
>>> ....
>>>
>>> so part of the work (which I called "factorize" in the commit
>>> messages) is about replacing
>>>
>>> (define_insn "mve_vaddq_n_<supf><mode>"
>>> with
>>> (define_insn "@mve_<mve_insn>q_n_<supf><mode>"
>>> with the help of a new iterator (mve_insn).
>>>
>>> Doing so makes it more obvious that some patterns are identical,
>>> except for the instruction name. I took this opportunity to merge
>>> them, so for instance I have a patch which merges add, sub and mul
>>> patterns.  Although not strictly necessary for the MVE intrinsics
>>> restructuring work, this is a good opportunity to reduce such code
>>> duplication (I did notice a few bugs during that process, which led me
>>> to post a few small patches in the past months).  Note that identical
>>> patterns will probably remain after the series, they can be merged
>>> later if we want.
>>>
>>> This factorization also implies the introduction of new iterators, but
>>> also means that several existing ones become useless. These patches do
>>> not remove them because it's a bit painful to reorder patches which
>>> remove lines at some "random" places, leading to merge conflicts. It's
>>> much simpler to write a big cleanup patch at the end of the serie to
>>> remove all such useless iterators at once.
>>>
>>> * Intrinsic re-implementation
>>> After intrinsic names have been factorized, the actual
>>> re-implementation patch is small:
>>> - add 1 line in each of arm-mve-builtins-base.{cc,def,h} describing
>>>    the intrinsic shape/signature, types and predicates involved,
>>>    RTX/unspec codes
>>> - remove the intrinsic definitions from arm_mve.h
>>>
>>> The full series of ~140 patches is organized like this:
>>> - patches 1 and 2 introduce the new framework
>>> - new implementation of vreinterpretq
>>> - new implementation of vuninitialized
>>> - patch groups of varying size, consisting in:
>>>    - add a new "shape" if needed (e.g. unary, binary, ternary, ....)
>>>    - add framework support functions if needed
>>>    - factorize a set of intrinsics (at minimum, just make use of
>>>      parameterized-names)
>>>    - actual re-implementation of the intrinsics
>>>
>>> I kept patches small so the incremental progress is easy to follow and
>>> check.  I'll submit the patches in small groups, this first one will
>>> make sure we agree on the implementation.
>>>
>>> Tested on arm-eabi with -mthumb/-mfloat-abi=hard/-march=armv8.1-
>>> m.main+mve.
>>>
>>> To help reviewers, I suggest to compare arm-mve-builtins.cc with
>>> aarch64-sve-builtins.cc.
>>>
>>> Christophe Lyon (22):
>>>    arm: move builtin function codes into general numberspace
>>>    arm: [MVE intrinsics] Add new framework
>>>    arm: [MVE intrinsics] Rework vreinterpretq
>>>    arm: [MVE intrinsics] Rework vuninitialized
>>>    arm: [MVE intrinsics] add binary_opt_n shape
>>>    arm: [MVE intrinsics] add unspec_based_mve_function_exact_insn
>>>    arm: [MVE intrinsics] factorize vadd vsubq vmulq
>>>    arm: [MVE intrinsics] rework vaddq vmulq vsubq
>>>    arm: [MVE intrinsics] add binary shape
>>>    arm: [MVE intrinsics] factorize vandq veorq vorrq vbicq
>>>    arm: [MVE intrinsics] rework vandq veorq
>>>    arm: [MVE intrinsics] add binary_orrq shape
>>>    arm: [MVE intrinsics] rework vorrq
>>>    arm: [MVE intrinsics] add unspec_mve_function_exact_insn
>>>    arm: [MVE intrinsics] add create shape
>>>    arm: [MVE intrinsics] factorize vcreateq
>>>    arm: [MVE intrinsics] rework vcreateq
>>>    arm: [MVE intrinsics] factorize several binary_m operations
>>>    arm: [MVE intrinsics] factorize several binary _n operations
>>>    arm: [MVE intrinsics] factorize several binary _m_n operations
>>>    arm: [MVE intrinsics] factorize several binary operations
>>>    arm: [MVE intrinsics] rework vhaddq vhsubq vmulhq vqaddq vqsubq
>>>      vqdmulhq vrhaddq vrmulhq
>>>
>>>   gcc/config.gcc                                |    2 +-
>>>   gcc/config/arm/arm-builtins.cc                |  237 +-
>>>   gcc/config/arm/arm-builtins.h                 |    1 +
>>>   gcc/config/arm/arm-c.cc                       |   42 +-
>>>   gcc/config/arm/arm-mve-builtins-base.cc       |  163 +
>>>   gcc/config/arm/arm-mve-builtins-base.def      |   50 +
>>>   gcc/config/arm/arm-mve-builtins-base.h        |   47 +
>>>   gcc/config/arm/arm-mve-builtins-functions.h   |  387 +
>>>   gcc/config/arm/arm-mve-builtins-shapes.cc     |  529 ++
>>>   gcc/config/arm/arm-mve-builtins-shapes.h      |   47 +
>>>   gcc/config/arm/arm-mve-builtins.cc            | 2013 ++++-
>>>   gcc/config/arm/arm-mve-builtins.def           |   40 +-
>>>   gcc/config/arm/arm-mve-builtins.h             |  672 +-
>>>   gcc/config/arm/arm-protos.h                   |   24 +
>>>   gcc/config/arm/arm.cc                         |   27 +
>>>   gcc/config/arm/arm_mve.h                      | 7581 +----------------
>>>   gcc/config/arm/arm_mve_builtins.def           |    6 -
>>>   gcc/config/arm/arm_mve_types.h                | 1430 ----
>>>   gcc/config/arm/iterators.md                   |  240 +-
>>>   gcc/config/arm/mve.md                         | 1747 +---
>>>   gcc/config/arm/predicates.md                  |    4 +
>>>   gcc/config/arm/t-arm                          |   32 +-
>>>   gcc/config/arm/unspecs.md                     |    1 +
>>>   gcc/config/arm/vec-common.md                  |    8 +-
>>>   gcc/testsuite/g++.target/arm/mve.exp          |    8 +-
>>>   .../arm/mve/general-c++/nomve_fp_1.c          |   15 +
>>>   .../arm/mve/general-c++/vreinterpretq_1.C     |   25 +
>>>   .../gcc.target/arm/mve/general-c/nomve_fp_1.c |   15 +
>>>   .../arm/mve/general-c/vreinterpretq_1.c       |   25 +
>>>   29 files changed, 4926 insertions(+), 10492 deletions(-)
>>>   create mode 100644 gcc/config/arm/arm-mve-builtins-base.cc
>>>   create mode 100644 gcc/config/arm/arm-mve-builtins-base.def
>>>   create mode 100644 gcc/config/arm/arm-mve-builtins-base.h
>>>   create mode 100644 gcc/config/arm/arm-mve-builtins-functions.h
>>>   create mode 100644 gcc/config/arm/arm-mve-builtins-shapes.cc
>>>   create mode 100644 gcc/config/arm/arm-mve-builtins-shapes.h
>>>   create mode 100644 gcc/testsuite/g++.target/arm/mve/general-
>>> c++/nomve_fp_1.c
>>>   create mode 100644 gcc/testsuite/g++.target/arm/mve/general-
>>> c++/vreinterpretq_1.C
>>>   create mode 100644 gcc/testsuite/gcc.target/arm/mve/general-
>>> c/nomve_fp_1.c
>>>   create mode 100644 gcc/testsuite/gcc.target/arm/mve/general-
>>> c/vreinterpretq_1.c
>>>
>>> -- 
>>> 2.34.1
>>