Re: [PATCH v2 1/3] compiler_types: Introduce the Clang __preserve_most function attribute

[Peter Zijlstra]

On Mon, Aug 07, 2023 at 01:41:07PM +0200, Florian Weimer wrote:
> * Marco Elver:
> 
> > [1]: "On X86-64 and AArch64 targets, this attribute changes the calling
> > convention of a function. The preserve_most calling convention attempts
> > to make the code in the caller as unintrusive as possible. This
> > convention behaves identically to the C calling convention on how
> > arguments and return values are passed, but it uses a different set of
> > caller/callee-saved registers. This alleviates the burden of saving and
> > recovering a large register set before and after the call in the
> > caller."
> >
> > [1] https://clang.llvm.org/docs/AttributeReference.html#preserve-most
> 
> You dropped the interesting part:
> 
> | If the arguments are passed in callee-saved registers, then they will
> | be preserved by the callee across the call. This doesn’t apply for
> | values returned in callee-saved registers.
> | 
> |  ·  On X86-64 the callee preserves all general purpose registers, except
> |     for R11. R11 can be used as a scratch register. Floating-point
> |     registers (XMMs/YMMs) are not preserved and need to be saved by the
> |     caller.
> |     
> |  ·  On AArch64 the callee preserve all general purpose registers, except
> |     X0-X8 and X16-X18.
> 
> Ideally, this would be documented in the respective psABI supplement.
> I filled in some gaps and filed:
> 
>   Document the ABI for __preserve_most__ function calls
>   <https://gitlab.com/x86-psABIs/x86-64-ABI/-/merge_requests/45>
> 
> Doesn't this change impact the kernel module ABI?
> 
> I would really expect a check here

So in the GCC bugzilla you raise the point of unwinding.

So in arch/x86 we've beeing doing what this attribute proposes (except
without that weird r11 exception) for a long while.

We simply do: asm("call foo_thunk"); instead of a C call, and then
have the 'thunk' thing PUSH/POP all the registers around a regular C
function.

Paravirt does quite a lot of that as well.

In a very few cases we implement a function specifically to avoid all
the PUSH/POP nonsense because it's so small the stack overhead kills it.

For unwinding this means that the 'thunk' becomes invisible when IP is
not inside it. But since the thunk is purely 'uninteresting' PUSH/POP
around a real C call, this is not an issue.

[[ tail calls leaving their sibling invisible is a far more annoying
   issue ]]

If the function is one of those special things, then it will be a leaf
function and we get to see it throught he IP.


Now, the problem with __preserve_most is that it makes it really easy to
deviate from this pattern, you can trivially write a function that is
not a trivial wrapper and then does not show up on unwind. This might
indeed be a problem.

Ofc. the kernel doesn't longjmp much, and we also don't throw many
exxceptions, but the live-patch people might care (although ORC unwinder
should be able to deal with all this perfectly fine).

In userspace, would not .eh_frame still be able to unwind this?