[Bug target/110247] New: suboptimal code about `no_caller_saved_registers` on caller side

[Bug target/110247] New: suboptimal code about `no_caller_saved_registers` on caller side

[lh_mouse at 126 dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110247

            Bug ID: 110247
           Summary: suboptimal code about `no_caller_saved_registers` on
                    caller side
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: lh_mouse at 126 dot com
  Target Milestone: ---

Given:
(https://gcc.godbolt.org/z/xevzx56Y5)

```
int complex(int x, int y)
  __attribute__((__no_caller_saved_registers__));

int test(int x, int y, int z)
  {
    return complex(x, y) + complex(y, z) + complex(z, x);
  }
```


My understanding is that `__no_caller_saved_registers__` says no register will
be clobbered by `complex`, so it is not necessary for GCC to establish a stack
frame and push arguments there. This is gonna help a lot if the function will
be inlined.


Clang generates much better assembly but I wonder whether it is valid to assume
that arguments registers are also preserved, like

```
test:
  push rcx            ; align %rsp
                      ; %edi := x, %esi = y, %edx = z
  call complex        ;
  mov ecx, eax        ; %ecx = complex(x, y)
  xchg edi, esi       ; %edi = y
  xchg esi, edx       ; %esi = z, %edx = x
  call complex        ;
  add ecx, eax        ; %ecx += complex(y, z)
  mov edi, esi        ; %edi = z
  mov esi, edx        ; %esi = x
  call complex        ;
  add eax, ecx        ; %eax = %ecx + complex(z, x)
  pop rcx
  ret
```

[Bug target/110247] suboptimal code about `no_caller_saved_registers` on caller side

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110247

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |documentation

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
The way I read the documentation, it will NOT be used when dealing with calls
to that function but rather the function that has that attribute will save all
registers and restore them at the end.
And it is used only for calls from an interrupt handler (e.g. directly from
assembly). and even more you need to use -mgeneral-regs-only too.

[Bug target/110247] suboptimal code about `no_caller_saved_registers` on caller side

[lh_mouse at 126 dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110247

--- Comment #2 from LIU Hao <lh_mouse at 126 dot com> ---
(In reply to Andrew Pinski from comment #1)
> The way I read the documentation, it will NOT be used when dealing with

If it is known, then why shouldn't it?

One potential usecase where this would be helpful is `__errno_location()` in
GNU libc., and maybe `__emutls_get_address()`.


> And it is used only for calls from an interrupt handler (e.g. directly from
> assembly). and even more you need to use -mgeneral-regs-only too.

That function will be compiled with `-mgeneral-regs-only`, but callers need
not.