Re: [PATCH 0/2] tests/tcg/riscv64: Add vector load and store tests

On Thu, 2026-08-20 at 22:32 +0930, Joel Stanley wrote:
> Hi Richard, I was testing your riscv vector load/store series and
> noticed there were some regressions with page crossing loads and
> stores.
> We don't have any tests for that in the tree, so I wrote some.
> 
> The tests trigger assertions in linux-user for me:
> 
>   $ make -C build/tests/tcg/riscv64-linux-user run-test-rvv-ldst
>   make: Entering directory 'build/tests/tcg/riscv64-linux-user'
>     TEST    test-rvv-ldst on riscv64
>   **
>   ERROR:../accel/tcg/user-exec.c:810:probe_access: assertion failed:
> (-(addr | TARGET_PAGE_MASK) >= size)
>   timeout: the monitored command dumped core
>   Aborted
>   make: *** [Makefile:223: run-test-rvv-ldst] Error 134
>   make: Leaving directory 'build/tests/tcg/riscv64-linux-user'
> 
> I wrote the assembly one first before realising a C version would be
> easy enough to do too. I'm not sure about using memcmp to check (in
> case
> it's vectorised) so perhaps that's where the asm wins. I can tidy
> them
> up and send the preferred version if we decide to put it in the tree.

Why not just use the C one, but replace memcmp() with a simple for
loop? That way we avoid the assembly and know memcmp() isn't vectorised
and won't be in the future.

Alistair

> 
> Joel Stanley (2):
>   tests/tcg/riscv64: Vector load test
>   tests/tcg/riscv64: Add intrinsics version of test-rvv-ldst
> 
>  tests/tcg/riscv64/test-rvv-ldst-intrinsics.c |  66 ++++++++++
>  tests/tcg/riscv64/Makefile.target            |  19 +++
>  tests/tcg/riscv64/test-rvv-ldst.S            | 131
> +++++++++++++++++++
>  3 files changed, 216 insertions(+)
>  create mode 100644 tests/tcg/riscv64/test-rvv-ldst-intrinsics.c
>  create mode 100644 tests/tcg/riscv64/test-rvv-ldst.S