[VP] Strided loads/stores

[loralb]

Summary

This patch introduces two new experimental IR intrinsics and SDAG nodes to represent vector strided loads and stores.

[craig.topper]

llvm/include/llvm/IR/VPIntrinsics.def

253

Is this an existing mistake in the documentation? Can you just push this as an NFC commit? No need to bury it in this patch.

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

8039

Since this accesses multiple disjoint memory locations, the Size is not MemVT.getStoreSize(). It probably needs to always be unknown. I guess it would be if the stride were statically known to be 0, but not sure it is worth special casing that.

8172

Same as load, the Size can't be based on the store type.

8228

Same

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

7488

This Size doesn't make sense. I think it needs to be unknown.

7509

Same

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.h

574

Can these be SmallVectorImpl<SDValue>?

[rogfer01]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

7476

Does anyone know if we should be passing AAInfo here?

(Non-stride) VPLoad doesn't do that, so we can fix both in a later change if necessary.

[loralb]

ChangeLog:

• add documentation for the new intrinsics
• use MemoryLocation::UnknownSize
• use SmallVectorImpl<SDValue> in functions signatures

[loralb]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.h

574

I'll fix visitVPLoadGather and visitVPStoreScatter above in a later change.

[loralb]

Changelog:

• ensure default alignment behaviour reflects changes in D114423
• align MemoryLocation retrieval for VPStridedLoad to VPLoad (see D115036)

[simoll]

llvm/lib/IR/IntrinsicInst.cpp

510

This will require an explicit return type because of opaque pointers, as in 6213f1dd03e27fd2483f8ac775a346de8e873573

[loralb]

Changelog:

• use explicit return type in Intrinsic::getDeclaration() for experimental_vp_strided_load
• get first argument type from first operand in Intrinsic::getDeclaration() for experimental_vp_strided_store

[frasercrmck]

llvm/docs/LangRef.rst

19964

I think mentioning that the align parameter can be used (and correctly handling that further down in SelectionDAG) would be good.

[loralb]

Changelog:

• Mention that the align parameter attribute can be provided for the base pointer operands

[loralb]

llvm/docs/LangRef.rst

19952

Reflecting on how the strided versions work, I was thinking if, semantically speaking, it would be better to define the pointer operand as a pointer to the scalar element type of the data/return operand instead of a pointer to a vector (this is only relevant while opaque pointers are not the default)

[simoll]

llvm/docs/LangRef.rst

19952

Agreed. Typed pointers will go away but we shouldn't be sloppy here.

[loralb]

Changelog:

• Changed the pointer operand type to match the scalar elements type of the data/return operand

[simoll]

llvm/docs/LangRef.rst

19980

The stride can have a different number of bits than the ptr.
To make the definition unambiguous, we should mention that the stride is always interpreted as a signed integer and all arithmetic occurs in the pointer type.
This implies that an over-sized stride is truncated and a less-bits-than-the-address stride is always sign extended to the address type.

[loralb]

Changelog:

• Update documentation
• Align with recent changes to vp_load and vp_store

[simoll]

LGTM

[loralb]

I do not have commit access, could anyone commit this for me please?

[loralb]

Thanks!