Re: [PATCH 2/4] ARM: Fix register constraints in run_in_exception_handler()

[Andrew Cooper <andrew.cooper3@citrix.com>]

On 10/02/2025 9:29 pm, Julien Grall wrote:
> Hi,
>
> On 08/02/2025 00:02, Andrew Cooper wrote:
>> Right now, run_in_exception_handler() takes an input in an arbitrary
>> register,
>> and clobbers BUG_FN_REG.  This causes the compiler to calculate fn in
>> the
>> wrong regsiter.
>
> Just to confirm, you mean, the compiler is not clever enough to notice
> that the value should be in the register BUG_FN_REG and therefore, two
> registers will be clobbered. Is that correct?

Not quite.

The clobbered register set is always disjoint from inputs and outputs,
so the combination of one clobbered + one input always means two
different registers.

For "here's an input but it gets modified", you need to express that as
an output into a variable which isn't subsequently used.

For ARM, that is best spelt "+r" (foo) so it can also be used with
register asm() to tie to a single register.  On x86, you can use "=a"
(tmp) : "a" (input).  In principle you can do it with named parameters,
so [fn] "=r" (tmp) : "[fn]" (input) I believe works too.

Here is a contrived example https://godbolt.org/z/WjqTKjWWb showing how
the output (discard only) is forced into r0, causing the compiler to
copy a into r3 around the asm block.  Notice that GCC and Clang pick the
input operand differently, as both r0 and r3 are valid candidates in
this case.


However, for run_in_exception_handler(), "fn" isn't even modified
(AFAICT), so it's correct to describe it as an input only.

~Andrew