llvm dev - Sep 2016 - sret vs noalias

Hi,

In ArgumentPromotion there is a piece of code that replaces the "sret"
attribute with "noalias" and I'm puzzled about this:

// Replace sret attribute with noalias. This reduces register pressure by
// avoiding a register copy.

And the commit message when this was introduced says

"Because sret implies noalias, we also replace the former with the
latter."

But if I read about "sret" at

  http://llvm.org/docs/LangRef.html#parameter-attributes

I don't see anything saying that "sret" must only be used when it 
doesn't alias with anything else?

Also, there are several backends that look for the "sret" attribute,
how
does that work if "sret" has already been removed by ArgumentPromotion
when running opt?

And finally, if my own backend doesn't look at the "sret"
attribute at
all, is it then ok for my frontend to avoid emitting it, and just use 
noalias (when applicable) instead?

Thanks,
Mikael

On Thu, Sep 8, 2016 at 12:25 AM, Mikael Holmén via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> Hi,
>
> In ArgumentPromotion there is a piece of code that replaces the
"sret"
> attribute with "noalias" and I'm puzzled about this:
>
> // Replace sret attribute with noalias. This reduces register pressure by
> // avoiding a register copy.
>
> And the commit message when this was introduced says
>
> "Because sret implies noalias, we also replace the former with the
latter."
>
> But if I read about "sret" at
>
>  http://llvm.org/docs/LangRef.html#parameter-attributes
>
> I don't see anything saying that "sret" must only be used
when it doesn't
> alias with anything else?
>
Clang appears to set both attributes (sret and noalias) on implicit struct
return pointer parameters. I think we might be better off having sret *not*
imply noalias, and letting frontends set noalias separately. Argument
promotion would then simply drop 'sret' when possible.

> Also, there are several backends that look for the "sret"
attribute, how
> does that work if "sret" has already been removed by
ArgumentPromotion when
> running opt?
>
It works as intended: those backends don't find sret, so they don't do
the
extra register copy that sret normally requires.

> And finally, if my own backend doesn't look at the "sret"
attribute at
> all, is it then ok for my frontend to avoid emitting it, and just use
> noalias (when applicable) instead?
>
Yes.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160908/56169e89/attachment.html>