[Bug c/121235] New: Local register variable constraints aren't respected in extended asm

[pixel@nobis-crew.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121235

            Bug ID: 121235
           Summary: Local register variable constraints aren't respected
                    in extended asm
           Product: gcc
           Version: 15.1.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: pixel@nobis-crew.org
  Target Milestone: ---

I was following this documentation page here:

https://gcc.gnu.org/onlinedocs/gcc/Local-Register-Variables.html

And I tried to naively copy/past the given example into compilable code: 

https://godbolt.org/z/bxj1TToTr

In this code, I'm creating a function that's using the example from the
documentation, having one input and one output, both with specific constraints
on register names. The generated code with the comments shows precisely which
register has been assigned, and it's not always respecting those constraints.

Basically, when the inner function is compiled as a stand alone function, the
constraints are properly respected, which could be by happenstance, but when
the inner function is used inline by the outer function, suddenly the compiler
is ignoring the constraints for the input values, for virtually every CPU
architecture I can think of. The selected output register seems to be properly
respected, but this may be happenstance, although the generated code seems to
try to respect it nonetheless in some of the architecture, by swapping
registers around.

Am I reading the documentation wrong? It seems to me the provided example
literally doesn't work, for any architecture, nor any version of gcc I can try
on godbolt, ever since 4.x. It could simply be that the documentation isn't
clear enough, but then this bug might be changed to a documentation bug to
clarify what exactly is the only supported use for this feature. Currently, the
only workaround I have is to force the function to be noinline.