Re: [PATCH 11/23] target/riscv: Rewrite vext_ldff

On 8/26/26 11:40, Max Chou wrote:
> Sorry, my previous example was unclear. It does not need a 1-byte PMP
> region: RISC-V PMP definition allows the minimum 4-byte PMP region,
> which exactly matches one e32 element.
>
> For example, with vl=3, vstart=0, e32, and v0.mask = [1, 0, 1],
> configure locked PMP entries as follows:
>
>    PMP0: NA4   [base + 4, base + 7], L, ---    # element 1, no read
>    PMP1: NAPOT target page,          L, R      # lower-priority page allow
>
>    bytes:       base       base + 4     base + 8
>                 +-----+    +-----+      +-----+
>    element:       0          1            2
>    mask:          1          0            1
>    access:     active      inactive      active
>    PMP:        PMP1 R      PMP0 ---      PMP1 R
>
> The RISC-V spec defines that masked vector loads access memory and raise
> exceptions only for active elements.
> Therefore this situation has reads only for elements 0 and 2; PMP0 must
> not create an access-fault condition for this instruction.
>
> My concern is that the page probe may be unsafe in some situations
> because PMP permissions can change at 4-byte granularity inside a
> target page, while the probe includes bytes which are not part of
> active vector memory operations.
>
> Maybe we could use the page probe only when it returns flags == 0,
> and otherwise fall back to checking only active elements or active
> runs?

Ok, thanks for the clear example.  flags == 0 does not help with this case.

What we want is to probe the first active element, with 
probe_access_full.  For this particular case, where a PMP rule matches 
something smaller than the page, we will have set lg_page_size < 
TARGET_PAGE_BITS.  So we get the fault (or not) for the first element 
exactly, and can then tell via flags and lg_page_size whether the rest 
of the page can use the fast or slow paths.

> I'm trying to create masked fauly-only-first + PMP test for this.

Thanks, that would be helpful.  Similarly for masked not-first-fault, 
please?


r~