llvm dev - Mar 2021 - Controlling parameter alignment

On Wed, Mar 10, 2021 at 9:51 PM Reid Kleckner <rnk at google.com>
wrote:
>
> On Wed, Mar 10, 2021 at 6:35 AM Momchil Velikov <momchil.velikov at
gmail.com>
wrote:
>>
>> Just to be clear, the suggestion is to introduce `alignstack` to
>> affect argument alignment, while retaining
>> the current semantics for `align`?
>>
>> Thus, a pointer argument having both `align(A)` and `alignstack(B)`
>> would itself be allocated at B boundary (if it happens to be passed in
>> memory),
>> while it would contain an A-aligned address?
>
>
> Yes, that's the proposal as I understand it.
Something is not quite right here.

We have up to three relevant alignment properties for a parameter:
 * the alignment of the parameter itself (if it happenes to be passed in
memory)
 * if it's a pointer, the actual alignment of the pointed to memory (as an
optimisation aid)
 * if it's a `byval` or a `byref` argument, the minimum alignment of the
storage, allocated
   for the original argument value (ABI affecting).

For non-pointer arguments `alignstack(N)` gives stack slot alignment.
For pointer arguments, we retain that use of `alignstack(N)` and also have
`align(M)` to give
the actual alignment of the contained pointer.

Now when we add `byval` or `byref` to the above, there is no attribute left
to give the alignment
of the allocated memory. We thought of using `alignstack(N)`, but that
would leave us without a way
to specify the pointer alignment itself.

~chill

-- 
Compiler scrub, Arm
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210324/f8b50a43/attachment.html>

On Wed, Mar 24, 2021 at 11:28 AM Momchil Velikov <momchil.velikov at
gmail.com>
wrote:
>
> On Wed, Mar 10, 2021 at 9:51 PM Reid Kleckner <rnk at google.com>
wrote:
> >
> > On Wed, Mar 10, 2021 at 6:35 AM Momchil Velikov <
momchil.velikov at gmail.com> wrote:
> >>
> >> Just to be clear, the suggestion is to introduce `alignstack` to
> >> affect argument alignment, while retaining
> >> the current semantics for `align`?
> >>
> >> Thus, a pointer argument having both `align(A)` and
`alignstack(B)`
> >> would itself be allocated at B boundary (if it happens to be
passed in
> >> memory),
> >> while it would contain an A-aligned address?
> >
> >
> > Yes, that's the proposal as I understand it.
>
> Something is not quite right here.
>
> We have up to three relevant alignment properties for a parameter:
>  * the alignment of the parameter itself (if it happenes to be passed in
memory)
>  * if it's a pointer, the actual alignment of the pointed to memory (as
an optimisation aid)
>  * if it's a `byval` or a `byref` argument, the minimum alignment of
the
storage, allocated
>    for the original argument value (ABI affecting).
>
> For non-pointer arguments `alignstack(N)` gives stack slot alignment.
> For pointer arguments, we retain that use of `alignstack(N)` and also
have `align(M)` to give
> the actual alignment of the contained pointer.
>
> Now when we add `byval` or `byref` to the above, there is no attribute
left to give the alignment
> of the allocated memory. We thought of using `alignstack(N)`, but that
would leave us without a way
> to specify the pointer alignment itself.

I hope I'm not missing something obvious, and if I don't here's an
idea:

Extend the `byval(Ty)` attribute to `byval(Ty [, Align])` (same for
`byref`).

(Most of the attributes take zero or one parameters, but there's a
precedent with `allocsize(<EltSizeParam>[, <NumEltsParam>])`)

So we end up with :
* `align(N)` for pointer content
* `stackalign(N)` for the minimum alignment for of the actual argument, if
it ends up in memory
* `byval(Ty[, N])` and `byref(Ty[, N])` for the original argument value

Thus something like `call %f(%struct.S * alignstack(8) align(32)
byval(%struct.S, 16) p);`
would mean:
* the caller has allocated a slot for `struct S`, that slot is at least 16
bytes aligned
* the caller is passing a pointer to that slot, which pointer itself should
be 8 bytes aligned, if it ends up in memory
* that pointer happens to have the lower five bits clear

~chill

--
Compiler scrub, Arm
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210324/63cb3d69/attachment.html>