[Bug c/111558] New: RISCV: shrink-wrapper optimization question

[Bug c/111558] New: RISCV: shrink-wrapper optimization question

[alexey.lapshin at espressif dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111558

            Bug ID: 111558
           Summary: RISCV: shrink-wrapper optimization question
           Product: gcc
           Version: 13.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: alexey.lapshin at espressif dot com
                CC: manolis.tsamis at vrull dot eu
  Target Milestone: ---

Created attachment 55977
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=55977&action=edit
shrink-synthetic-example.c

When I was studying shrink-wrapping behavior, I discovered something strange
code generation.

See the code and assembly at https://godbolt.org/z/P3WfTszYn

Every branch has the same save/restore callee-saved registers.
Therefore, the shrink-wrapping optimization may reduce the speed of the
function when input parameters are not zero...

I understand that this is not a synthetic code snippet... 

But may this code generation be reproduced in real code (without asm inlining)?

[Bug rtl-optimization/111558] RISCV: shrink-wrapper optimization question

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111558

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
>But may this code generation be reproduced in real code (without asm inlining)?


It could in theory but I highly doubt it since most of the time the code blocks
are not written in such a way they would happen this way where they are
independent and such. Maybe initialization code might be done this way but that
code is usually only executed once so the speed will not have a huge impact ...

[Bug rtl-optimization/111558] RISCV: shrink-wrapper optimization question

[mxlol233 at outlook dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111558

Xiao Ma <mxlol233 at outlook dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |mxlol233 at outlook dot com

--- Comment #2 from Xiao Ma <mxlol233 at outlook dot com> ---
For this piece of code, LLVM seems to be a bit concise: during the pass of
Prologue/Epilogue Insertion & Frame Finalization, it only inserts sd/ld blocks
in the entry and exit BB:

https://godbolt.org/z/qxdrMKc46

[Bug rtl-optimization/111558] RISCV: shrink-wrapper optimization question

[mxlol233 at outlook dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111558

--- Comment #3 from Xiao Ma <mxlol233 at outlook dot com> ---
(In reply to Xiao Ma from comment #2)
> For this piece of code, LLVM seems to be a bit concise: during the pass of
> Prologue/Epilogue Insertion & Frame Finalization, it only inserts sd/ld
> blocks in the entry and exit BB:
> 
> https://godbolt.org/z/qxdrMKc46

On the contrary, GCC emits sd/ld blocks for each of BB, in the pro_and_epilogue
pass:

https://godbolt.org/z/vY4M71snz