[v10,0/2] riscv: Add runtime constant support

Message

Charlie Jenkins March 19, 2025, 6:35 p.m. UTC

Ard brought this to my attention in this patch [1].

I benchmarked this patch on the Nezha D1 (which does not contain Zba or
Zbkb so it uses the default algorithm) by navigating through a large
directory structure. I created a 1000-deep directory structure and then
cd and ls through it. With this patch there was a 0.57% performance
improvement.

[1] https://lore.kernel.org/lkml/CAMj1kXE4DJnwFejNWQu784GvyJO=[e-mail]/

Signed-off-by: Charlie Jenkins <[e-mail]>
---
Changes in v10:
- Use _AC() instead of just adding U to the end of a constant
- Link to v9: https://lore.kernel.org/r/[e-mail]

Changes in v9:
- Fix bug where stale register data may be used when an lui is replaced
  with a nop. In the following addiw, add to register x0 instead of the
  stale register to resolve.
- Add locks for text_mutex before using patch_insn_write()
- Link to v8: https://lore.kernel.org/r/[e-mail]

Changes in v8:
- Rebase to linux v6.14-rc5
- Link to v7: https://lore.kernel.org/r/[e-mail]

Changes in v7:
- Added benchmarking info
- Added CONFIG_RISCV_ISA_ZBA and CONFIG_RISCV_ISA_ZBKB to check that the
  compiler supports the extensions.
- Link to v6: https://lore.kernel.org/r/[e-mail]

Changes in v6:
- .option arch only became officially supported by clang in version 17.
  Add a config to check that and guard the alternatives uses .option
  arch.
- Link to v5: https://lore.kernel.org/r/[e-mail]

Changes in v5:
- Split instructions into 16-bit parcels to avoid alignment (Emil)
- Link to v4: https://lore.kernel.org/r/[e-mail]

Changes in v4:
- Add newlines after riscv32 assembler directives
- Align instructions along 32-bit boundary (Emil)
- Link to v3: https://lore.kernel.org/r/[e-mail]

Changes in v3:
- Leverage "pack" instruction for runtime_const_ptr() to reduce hot path
  by 3 instructions if Zbkb is supported. Suggested by Pasha Bouzarjomehri ([e-mail])
- Link to v2: https://lore.kernel.org/r/[e-mail]

Changes in v2:
- Treat instructions as __le32 and do proper conversions (Ben)
- Link to v1: https://lore.kernel.org/r/[e-mail]

---
Charlie Jenkins (2):
      riscv: Move nop definition to insn-def.h
      riscv: Add runtime constant support

 arch/riscv/Kconfig                     |  22 +++
 arch/riscv/include/asm/asm.h           |   1 +
 arch/riscv/include/asm/ftrace.h        |   1 -
 arch/riscv/include/asm/insn-def.h      |   3 +
 arch/riscv/include/asm/runtime-const.h | 265 +++++++++++++++++++++++++++++++++
 arch/riscv/kernel/ftrace.c             |   6 +-
 arch/riscv/kernel/jump_label.c         |   4 +-
 arch/riscv/kernel/vmlinux.lds.S        |   3 +
 8 files changed, 299 insertions(+), 6 deletions(-)
---
base-commit: 2014c95afecee3e76ca4a56956a936e23283f05b
change-id: 20250123-runtime_const_riscv-6cd854ee2817

Comments

[e-mail] March 27, 2025, 3:24 a.m. UTC | #1

Hello:

This series was applied to riscv/linux.git (for-next)
by Alexandre Ghiti <[e-mail]>:

On Wed, 19 Mar 2025 11:35:18 -0700 you wrote:
> Ard brought this to my attention in this patch [1].
> 
> I benchmarked this patch on the Nezha D1 (which does not contain Zba or
> Zbkb so it uses the default algorithm) by navigating through a large
> directory structure. I created a 1000-deep directory structure and then
> cd and ls through it. With this patch there was a 0.57% performance
> improvement.
> 
> [...]

Here is the summary with links:
  - [v10,1/2] riscv: Move nop definition to insn-def.h
    https://git.kernel.org/riscv/c/afa8a93932aa
  - [v10,2/2] riscv: Add runtime constant support
    https://git.kernel.org/riscv/c/a44fb5722199

You are awesome, thank you!