[Bug c/112537] New: Is there a way to disable cpymem pass for rvv

[Bug c/112537] New: Is there a way to disable cpymem pass for rvv

[fanghuaqi at vip dot qq.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112537

            Bug ID: 112537
           Summary: Is there a way to disable cpymem pass for rvv
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: fanghuaqi at vip dot qq.com
  Target Milestone: ---

See
https://gcc.gnu.org/git/?p=gcc.git;a=blob;f=gcc/testsuite/gcc.target/riscv/rvv/base/cpymem-2.c;h=7b706b6ef52544cabdce2007256c9e67b72ccd79;hb=9464e72bcc9123b619215af8cfef491772a3ebd9
this case, if I pass -march=rv64imafdcv to compile code like this, it will
generate vector related code, which I think it should only be enabled when
--param=riscv-autovec-preference=scalable is enabled, but actually it is auto
enable d even I pass  --param=riscv-autovec-preference=none, see
https://godbolt.org/z/TazWKEjzM

Is there a way to disable it, since if vector load and store for smaller
structure assign is costable and harmfull to cpu pipeline.

Thanks

[Bug target/112537] Is there a way to disable cpymem pass for rvv

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112537

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
>since if vector load and store for smaller structure assign is costable and harmfull to cpu pipeline.

This depends on the Pipeline. Seems like there is a missing cost model for your
cpu somewhere.

[Bug target/112537] Is there a way to disable cpymem pass for rvv

[juzhe.zhong at rivai dot ai]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112537

--- Comment #2 from JuzheZhong <juzhe.zhong at rivai dot ai> ---
Currently, we don't have a compile option to disable cpymem by RVV.

I can tell how to disable it:

in riscv.md:

(define_expand "cpymem<mode>"
  [(parallel [(set (match_operand:BLK 0 "general_operand")
                   (match_operand:BLK 1 "general_operand"))
              (use (match_operand:P 2 ""))
              (use (match_operand:SI 3 "const_int_operand"))])]
  ""
{
  if (riscv_vector::expand_block_move (operands[0], operands[1], operands[2]))
    DONE;
  else if (riscv_expand_block_move (operands[0], operands[1], operands[2]))
    DONE;
  else
    FAIL;
})

remove 
if (riscv_vector::expand_block_move (operands[0], operands[1], operands[2]))
    DONE;

Then it will be disabled.

[Bug target/112537] Is there a way to disable cpymem pass for rvv

[fanghuaqi at vip dot qq.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112537

--- Comment #3 from Huaqi <fanghuaqi at vip dot qq.com> ---
Hi Juzhe, thanks for the help, could it be controlled by
--param=riscv-autovec-preference= option, so if I want to enable cpymem
optimization, I can enable it by auto vectorzation.

Thanks

[Bug target/112537] Is there a way to disable cpymem pass for rvv

[juzhe.zhong at rivai dot ai]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112537

--- Comment #4 from JuzheZhong <juzhe.zhong at rivai dot ai> ---
You mean you want to disable it when disabling auto-vectorization ?

[Bug target/112537] Is there a way to disable cpymem pass for rvv

[fanghuaqi at vip dot qq.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112537

--- Comment #5 from Huaqi <fanghuaqi at vip dot qq.com> ---
(In reply to JuzheZhong from comment #4)
> You mean you want to disable it when disabling auto-vectorization ?

Yes, I think this could be better, so we can control whether vector instruction
should be generated, and if auto generated rvv instruction is not good, we can
enable it via gcc pragma control for selected code block.

Thanks

[Bug target/112537] Is there a way to disable cpymem pass for rvv

[juzhe.zhong at rivai dot ai]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112537

--- Comment #6 from JuzheZhong <juzhe.zhong at rivai dot ai> ---
We can add 

  if (!flag_tree_vectorize)
    return false;

in riscv_vector::expand_block_move

But I am not sure whether other RISC-V folks is happy with that.

CCing other folks to see whether they allow it.

[Bug target/112537] Is there a way to disable cpymem pass for rvv

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112537

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |mumuxi_ll at outlook dot com

--- Comment #7 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
*** Bug 112538 has been marked as a duplicate of this bug. ***

[Bug target/112537] Is there a way to disable cpymem pass for rvv

[kito at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112537

--- Comment #8 from Kito Cheng <kito at gcc dot gnu.org> ---
That remind me we may need one option like something -mgeneral-regs-only in
aarch64 and also for target attribute.

BTW, clang has an generic option called -mno-implicit-float can did similar
thing

[Bug target/112537] Is there a way to disable cpymem pass for rvv

[juzhe.zhong at rivai dot ai]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112537

--- Comment #9 from JuzheZhong <juzhe.zhong at rivai dot ai> ---
I think we can use -fno-tree-vectorize or --param=riscv-autovec-preference=none
to
disable it.

[Bug target/112537] Is there a way to disable cpymem pass for rvv

[juzhe.zhong at rivai dot ai]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112537

--- Comment #10 from JuzheZhong <juzhe.zhong at rivai dot ai> ---
This following patch:

diff --git a/gcc/config/riscv/riscv-string.cc
b/gcc/config/riscv/riscv-string.cc
index 57e8ad698d7..c135d4efdb2 100644
--- a/gcc/config/riscv/riscv-string.cc
+++ b/gcc/config/riscv/riscv-string.cc
@@ -773,7 +773,7 @@ expand_block_move (rtx dst_in, rtx src_in, rtx length_in)
        bnez a2, loop                   # Any more?
        ret                             # Return
   */
-  if (!TARGET_VECTOR)
+  if (!TARGET_VECTOR || riscv_autovec_preference == NO_AUTOVEC ||
!flag_tree_vectorize)
     return false;
   HOST_WIDE_INT potential_ew
     = (MIN (MIN (MEM_ALIGN (src_in), MEM_ALIGN (dst_in)), BITS_PER_WORD)


Can fix this issue. Is it OK ?

[Bug target/112537] Is there a way to disable cpymem pass for rvv

[kito at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112537

--- Comment #11 from Kito Cheng <kito at gcc dot gnu.org> ---
It's not scope of auto vectorization, so I would suggest add something like
`-mstringop-strategy=*` or `-mmemcpy-strategy=*` (from x86) or
`-param=riscv-mops-memcpy-size-threshold=` (from aarch64).

Personally I prefer x86 approach.

[Bug target/112537] Is there a way to disable cpymem pass for rvv

[amylaar at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112537

Jorn Wolfgang Rennecke <amylaar at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |amylaar at gcc dot gnu.org

--- Comment #12 from Jorn Wolfgang Rennecke <amylaar at gcc dot gnu.org> ---
(In reply to JuzheZhong from comment #2)
> Currently, we don't have a compile option to disable cpymem by RVV.

If you don't want any vector instructions to be emitted, why do you tell the
compiler to enable the 'v' extsnsion of the architecture?

[Bug target/112537] Is there a way to disable cpymem pass for rvv

[amylaar at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112537

--- Comment #13 from Jorn Wolfgang Rennecke <amylaar at gcc dot gnu.org> ---
Before we can consider any costs, we first have to know what they are.  Is
there any manual for a hardware implementation that specifies costs?

[Bug target/112537] Is there a way to disable cpymem pass for rvv

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112537

--- Comment #14 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Li Xu <xuli@gcc.gnu.org>:

https://gcc.gnu.org/g:e6269bb69c0734a5af716bfbded3621de6ca351d

commit r14-5602-ge6269bb69c0734a5af716bfbded3621de6ca351d
Author: xuli <xuli1@eswincomputing.com>
Date:   Fri Nov 17 04:48:47 2023 +0000

    RISC-V: Implement -mmemcpy-strategy= options[PR112537]

    https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112537

    -mmemcpy-strategy=[auto|libcall|scalar|vector]

    auto: Current status, use scalar or vector instructions.
    libcall: Always use a library call.
    scalar: Only use scalar instructions.
    vector: Only use vector instructions.

            PR target/112537

    gcc/ChangeLog:

            * config/riscv/riscv-opts.h (enum riscv_stringop_strategy_enum):
Strategy enum.
            * config/riscv/riscv-string.cc (riscv_expand_block_move): Disabled
based on options.
            (expand_block_move): Ditto.
            * config/riscv/riscv.opt: Add -mmemcpy-strategy=.

    gcc/testsuite/ChangeLog:

            * gcc.target/riscv/rvv/base/cpymem-strategy-1.c: New test.
            * gcc.target/riscv/rvv/base/cpymem-strategy-2.c: New test.
            * gcc.target/riscv/rvv/base/cpymem-strategy-3.c: New test.
            * gcc.target/riscv/rvv/base/cpymem-strategy-4.c: New test.
            * gcc.target/riscv/rvv/base/cpymem-strategy-5.c: New test.
            * gcc.target/riscv/rvv/base/cpymem-strategy.h: New test.