[v9] Generate strided vector loads/stores with tcg nodes

[PATCH v9 0/2] Generate strided vector loads/stores with tcg nodes

Posted by Chao Liu 4 months, 3 weeks ago

Hi all,

Thanks Richard for the review. In patch v9:

- Simplify the implementation of gen_check_vext_elem_mask():
  remove the `mask` argument, compute the mask directly inside the function,
  and eliminate redundant code.

- Limit the bit width to 8 bits when loading the mask from memory.

- Remove the `vreg` argument in gen_ldst_vreg().

History of changes:

patch v8:
- Use the right TCGv type for each variable — for example, make mask_elem
  type TCGv_i64.
- Use tcg_gen_trunc_i64_ptr() to change TCGv types — don't use C-style
  casting.
- Use TCG_COND_TSTNE, not TCG_COND_NE in tcg_gen_brcond_i64() to represent:
  if (vext_elem_mask(v0, i) != 0)
  https://lore.kernel.org/qemu-devel/cover.1757690407.git.chao.liu@zevorn.cn/

patch v7:
- Standardize the subject line of patch 1 and remove the trailing period.
- Split into sub-functions to improve the patch's code readability and
  facilitate review.
- Use more faster TCG ops, use tcg_gen_andi_tl() instead of tcg_gen_rem_tl().
- Add a tested-by signature for patch 2, as Eric has already tested it.
  https://lore.kernel.org/qemu-devel/cover.1756975571.git.chao.liu@zevorn.cn/

patch v6:
- If a strided vector memory access instruction has non-zero vstart, 
  check it through vlse/vsse helpers function.
- Adjust the tcg test Makefile.
  https://lore.kernel.org/qemu-devel/cover.1756906528.git.chao.liu@zevorn.cn/

Patch v5:
- Removed the redundant call to mark_vs_dirty(s) within the
  gen_ldst_stride_main_loop() function.
  https://lore.kernel.org/qemu-riscv/cover.1755609029.git.chao.liu@zevorn.cn/

Patch v4:
- Use ctz32() replace to for-loop
  https://lore.kernel.org/qemu-devel/cover.1755333616.git.chao.liu@yeah.net/

Patch v3:
- Fix the get_log2() function:
  https://lore.kernel.org/qemu-riscv/cover.1755287531.git.chao.liu@yeah.net/T/#t
- Add test for vlsseg8e32 instruction.
- Rebase on top of the latest master.

Patch v2:
- Split the TCG node emulation of the complex strided load/store operation into
  two separate functions to simplify the implementation:
  https://lore.kernel.org/qemu-riscv/20250312155547.289642-1-paolo.savini@embecosm.com/

Patch v1:
- Paolo submitted the initial version of the patch.
  https://lore.kernel.org/qemu-devel/20250211182056.412867-1-paolo.savini@embecosm.com/


Thanks,
Chao

Chao Liu (2):
  target/riscv: Use tcg nodes for strided vector ld/st generation
  tests/tcg/riscv64: Add test for vlsseg8e32 instruction

 target/riscv/insn_trans/trans_rvv.c.inc   | 354 ++++++++++++++++++++--
 tests/tcg/riscv64/Makefile.softmmu-target |   7 +-
 tests/tcg/riscv64/test-vlsseg8e32.S       | 107 +++++++
 3 files changed, 450 insertions(+), 18 deletions(-)
 create mode 100644 tests/tcg/riscv64/test-vlsseg8e32.S

-- 
2.51.0

Re: [PATCH v9 0/2] Generate strided vector loads/stores with tcg nodes

Posted by Chao Liu 2 months ago

On Tue, 16 Sep 2025 17:21:36 +0800, Chao Liu wrote:
>Hi all,
>
>Thanks Richard for the review. In patch v9:
>
>- Simplify the implementation of gen_check_vext_elem_mask():
>  remove the `mask` argument, compute the mask directly inside the function,
>  and eliminate redundant code.
>
>- Limit the bit width to 8 bits when loading the mask from memory.
>
>- Remove the `vreg` argument in gen_ldst_vreg().
>
>History of changes:
>
>patch v8:
>- Use the right TCGv type for each variable — for example, make mask_elem
>  type TCGv_i64.
>- Use tcg_gen_trunc_i64_ptr() to change TCGv types — don't use C-style
>  casting.
>- Use TCG_COND_TSTNE, not TCG_COND_NE in tcg_gen_brcond_i64() to represent:
>  if (vext_elem_mask(v0, i) != 0)
>  https://lore.kernel.org/qemu-devel/cover.1757690407.git.chao.liu@zevorn.cn/
>
>patch v7:
>- Standardize the subject line of patch 1 and remove the trailing period.
>- Split into sub-functions to improve the patch's code readability and
>  facilitate review.
>- Use more faster TCG ops, use tcg_gen_andi_tl() instead of tcg_gen_rem_tl().
>- Add a tested-by signature for patch 2, as Eric has already tested it.
>  https://lore.kernel.org/qemu-devel/cover.1756975571.git.chao.liu@zevorn.cn/
>
>patch v6:
>- If a strided vector memory access instruction has non-zero vstart, 
>  check it through vlse/vsse helpers function.
>- Adjust the tcg test Makefile.
>  https://lore.kernel.org/qemu-devel/cover.1756906528.git.chao.liu@zevorn.cn/
>
>Patch v5:
>- Removed the redundant call to mark_vs_dirty(s) within the
>  gen_ldst_stride_main_loop() function.
>  https://lore.kernel.org/qemu-riscv/cover.1755609029.git.chao.liu@zevorn.cn/
>
>Patch v4:
>- Use ctz32() replace to for-loop
>  https://lore.kernel.org/qemu-devel/cover.1755333616.git.chao.liu@yeah.net/
>
>Patch v3:
>- Fix the get_log2() function:
>  https://lore.kernel.org/qemu-riscv/cover.1755287531.git.chao.liu@yeah.net/T/#t
>- Add test for vlsseg8e32 instruction.
>- Rebase on top of the latest master.
>
>Patch v2:
>- Split the TCG node emulation of the complex strided load/store operation into
>  two separate functions to simplify the implementation:
>  https://lore.kernel.org/qemu-riscv/20250312155547.289642-1-paolo.savini@embecosm.com/
>
>Patch v1:
>- Paolo submitted the initial version of the patch.
>  https://lore.kernel.org/qemu-devel/20250211182056.412867-1-paolo.savini@embecosm.com/
>

Ping.

Does anyone have any further comments or revisions on this set of patches?
I have been using them in the production environment for a long time, and so far,
they have been working reliably.

Thanks,
Chao