[llvm-dev] Opaque Pointers Help Wanted

Tue Jun 22 09:56:25 PDT 2021

> Maybe the byref attribute is suitable for this? That is @llvm.some.intrinsic(ptr byref(%ElemTy) %p).

I’m always hesitant to repurpose existing attributes, but byref does seem very close. The line in the LangRef saying explicitly that it is intended for ABI constraints and shouldn’t be used for optimization worries me.

Also, byref assumes the pointer is pointing to a singe element. For a GEP-like intrinsic, that’s not what I’d want. It would take some time to craft a precise definition, but I was thinking something like “Within the scope of this function call, the pointer may be assumed to point to zero or more elements of the specified type. This attribute does not guarantee the amount of dereferenceable memory, but whatever memory is dereferenceable may be assumed to be in increments of the size of the element type.” It may need to say more than that.

From: Nikita Popov <nikita.ppv at gmail.com>
Sent: Tuesday, June 22, 2021 12:23 PM
To: Kaylor, Andrew <andrew.kaylor at intel.com>
Cc: David Blaikie <dblaikie at gmail.com>; craig.topper at gmail.com; llvm-dev at lists.llvm.org
Subject: Re: [llvm-dev] Opaque Pointers Help Wanted

On Tue, Jun 22, 2021 at 6:16 PM Kaylor, Andrew via llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>> wrote:
Honestly, I didn’t (and still don’t) understand enough about the intrinsics Craig mentioned to even be sure they were analogous to the case I brought up. It seemed like they might be more interested in debug types than IR types and may or may not be going down the same paths of reasoning, but I just don’t understand what the intrinsics are doing.

We’re scrambling to get our out-of-tree code caught up with the opaque pointer transition. I think the scatter/gather intrinsics are likely to be our first useful overlap with the in-tree work to be done, though now that I’ve thought about it a bit more, I think the type in that can just be inferred from non-pointer arguments. In any case, I guess talking about it and agreeing on an approach is a good first step.

I started trying to mock up some IR for how this might work, but I wasn’t entirely happy with it. Specifically, how do you describe an IR type in metadata. We could use an undef value in the metadata, similar to what Arthur proposed as an argument, but that only moves the ugliness. This is kind of what I was thinking

define void @f() {
  ptr %p = <however we got %p>
  call void @llvm.some.intrinsic(ptr %p, metadata !1)
  <…>
  return void
}

define void @llvm.some.intrinsic(ptr, metadata)

%MyElementTy = type { <whatever> }

!1 = !{ %MyElementTy undef }

I guess that would work, but I don’t really like it. I looked at the way TBAA and debug info describe types, but that’s not describing the IR type and so it seems simultaneously fragile and too verbose. Is there a better way to do this? What about introducing a parameter attribute that would be considered required for the intrinsic? Something like this:

define void @f() {
  ptr %p = <however we got %p>
  call void @llvm.some.intrinsic(ptr elementtype(%MyElementTy) %p)
  <…>
  return void
}

; Requires ‘elementype’ parameter attribute
define void @llvm.some.intrinsic(ptr)

%MyElementTy = type { <whatever> }

Does that feel like a step backward to you? I don’t know how hard it would be for front end’s to generate something like that, but it hasn’t the benefit that the elementtype would have intentional and well-defined semantic meaning.

Maybe the byref attribute is suitable for this? That is @llvm.some.intrinsic(ptr byref(%ElemTy) %p).

I believe byref() doesn't carry any particular semantics apart from providing the element type. It's currently used to carry ABI information in the presence of opaque pointers (for AMDGPU kernels), but I think it would be suitable for use with intrinsics as well.

Regards,
Nikita

From: David Blaikie <dblaikie at gmail.com<mailto:dblaikie at gmail.com>>
Sent: Monday, June 21, 2021 6:00 PM
To: Kaylor, Andrew <andrew.kaylor at intel.com<mailto:andrew.kaylor at intel.com>>; craig.topper at gmail.com<mailto:craig.topper at gmail.com>
Cc: Arthur Eubanks <aeubanks at google.com<mailto:aeubanks at google.com>>; llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>
Subject: Re: [llvm-dev] Opaque Pointers Help Wanted

It looks like Craig's pointed out some cases that come up in LLVM. Perhaps you could try fixing one of the in-tree examples to help explore/set the direction for your out of tree work?

Metadata arguments sound like a direction worth exploring. (& would be consistent with the way parameter attributes have been approached - still using a type argument even though the length is all that matters in those cases)

On Mon, Jun 21, 2021 at 2:28 PM Kaylor, Andrew via llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>> wrote:
This has probably been discussed somewhere, but I missed it. Can you elaborate a bit on this?

  *   Allow bitcode auto-upgrade of legacy pointer type to the new opaque pointer type (not to be turned on until ready)

     *   To support legacy bitcode, such as legacy stores/loads, we need to track pointee types for all values since legacy instructions may infer the types from a pointer operand's pointee type
I‘m specifically trying to understand what will happen when typed pointer support is removed. How will IR with typed pointers be auto-upgraded to pure opaque pointer IR? Will the bitcode reader keep some level of typed pointer support indefinitely?

Also, do you have a plan for replacing intrinsics that currently rely on pointee types? For example, the load instruction was updated to take an explicit type operand but I don’t think we can do the same thing for an intrinsic like llvm.masked.load since there is Value for Type. This is an easy problem to work around for something like masked.load, but more complicated if anyone has a downstream GEP-like intrinsic that needs more than the size of an element (spoiler alert: I do have such an intrinsic). Would you use a metadata argument?

Thanks,
Andy

From: llvm-dev <llvm-dev-bounces at lists.llvm.org<mailto:llvm-dev-bounces at lists.llvm.org>> On Behalf Of Arthur Eubanks via llvm-dev
Sent: Monday, June 21, 2021 12:03 PM
To: llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>>
Subject: [llvm-dev] Opaque Pointers Help Wanted

For the opaque pointers project, https://llvm.org/docs/OpaquePointers.html#transition-plan contains high level steps for what to do before we can enable opaque pointers. (Looks like the page hasn't been rebuilt in a while, https://github.com/llvm/llvm-project/blob/main/llvm/docs/OpaquePointers.rst#transition-plan contains some more concrete steps)

Essentially almost all of the work boils down to figuring out how to remove calls to `PointerType::getElementType()` and `Type::getPointerElementType()`. Everything else derives from that.

Any help with this is welcome and appreciated!
_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20210622/f03afea6/attachment-0001.html>