clang inserts stack protector in function with inline assembly "m" constraint

[nickdesaulniers]

Bugzilla Link	37583
Resolution	DUPLICATE
Resolved on	Jun 01, 2018 07:33
Version	trunk
OS	Linux
CC	@tstellar

Extended Description

https://godbolt.org/g/TmLd9z

The two examples show that clang will insert a stack guard in a function with inline assembly and the "m" constraint compiled with -fstack-protector-strong. Is there a good reason to do so, or this a bug in clang?

I know the function attribute no_stack_protector was recently added to clang. That seems to help.

It just seems weird the just "=r" generates no stack protector, but "=rm" does.

From discussion:
https://lkml.org/lkml/2018/5/24/1389

[tstellar]

It's generating a stack protector, because the function takes the address of a local variable.

I think the real question here is why does clang decide pass the input as an address rather than a register.

[llvmbot]

Inline assembly arguments are conservatively chosen to be stored memory if there's a choice to avoid running out of registers later in allocation.

*** This bug has been marked as a duplicate of bug #20571 ***