[Atomics] Rename and change prototype for atomic memcpy intrinsic

[dneilson]

Summary

Background: http://lists.llvm.org/pipermail/llvm-dev/2017-May/112779.html

This change is to alter the prototype for the atomic memcpy intrinsic. The prototype itself is being changed to more closely resemble the semantics and parameters of the llvm.memcpy intrinsic -- to ease later combination of the llvm.memcpy and atomic memcpy intrinsics. Furthermore, the name of the atomic memcpy intrinsic is being changed to make it clear that it is not a generic atomic memcpy, but specifically a memcpy is unordered atomic.

[dneilson]

Note that this is the first of a series of patches that are being developed for the unordered atomic memcpy. Minimally, the plan is to push the following changes one at a time to minimize risk and impact on others:
i. Change intrinsic name, prototype (to match memcpy closely), & documentation.
ii. Add code to loop idiom to recognize the element unordered atomic memcpy.
iii. Add code to instcombine & selection dag builder to lower the intrinsic.
iv. Add an isunordered() to the MemIntrinsic introspection class (returning false for all existing intrinsics), and add calls to it to all passes it's relevant.
v. Add intrinsic into the introspection hierarchy & complete support for new intrinsic in passes.

[reames]

Reviewing only the LangRef changes for the moment. Let's iterate on those until we're happy and then I can go looking for code issues.

docs/LangRef.rst

13582

Your revised text is missing key aspects of the old text. You need to preserve the "as a sequence of intrinsic. It differs in that the `dest and src are treated as arrays with elements that are element_size` bytes wide and aligned at an element size boundary. " wording from the original, because this is semantically important.

13592

Er, huh? What do these major values mean? And why do we need anything other than an i1 boolean?

13595

This sentence is important and shouldn't be dropped.

13607

Ah, the answer to my question above. I think it would be cleaner to have two i1 params instead of encoding the bitmask. Does that complicate anything for you?

13614

This is slightly wrong. You don't need the writes to be unordered atomic if the src/dest doesn't need it, but you do still want to allow concurrent reads and writes. I think you want something along the lines of: "It is well defined to have concurrent reads and writes to both source and destination provided those reads and writes are unordered atomic when specified.

[efriedma]

Do we get any practical benefit from separately specifying whether the source and destination require unordered operations?

Why are you adding an alignment parameter? The alignment is already specified with attributes. (There was a plan at one point to change memcpy to specify alignment like this; IIRC it got committed, then reverted? I don't recall what happened after that.)

Why do we want to specify the length in bytes, as opposed to the number of elements to copy? Any implementation is inevitably just going to divide a length in bytes by the element size.

I'm not really sure why you're messing with the signature of the intrinsic in the first place; we went through most of this design space when it was initially proposed.

[dneilson]

In D33240#756419, @efriedma wrote:

Do we get any practical benefit from separately specifying whether the source and destination require unordered operations?

I don't know enough about the possible source languages to know with 100% certainty that it's not possible to mix, say, unordered loads with ordered stores. For example, Java only requires the unordered ops for shared data (i.e. stuff on the heap). It's conceivable that a memcpy is desired to copy, say, from the stack to the heap; only the heap stores would need to be unordered in this case -- playing devil's advocate, it would not be wrong (just unnecessary) to use unordered loads of the stack data here as well. Erring on the side of flexible/generic here.

Why are you adding an alignment parameter? The alignment is already specified with attributes. (There was a plan at one point to change memcpy to specify alignment like this; IIRC it got committed, then reverted? I don't recall what happened after that.)

Compatibility with the existing memcpy intrinsic. I have no problem removing the alignment parameter if that's the long-term goal/vision. My only concern is that I don't know whether that difference will lead me into trouble when I get to the point of adding the unordered atomic intrinsic to the MemTransferInst introspection class hierarchy.

Why do we want to specify the length in bytes, as opposed to the number of elements to copy? Any implementation is inevitably just going to divide a length in bytes by the element size.

Compatibility with memcpy semantics. The 'length' parameter of a MemTransferInst intrinsic is semantically understood by transforms/analysis to be the size of the transfer in bytes. Having an intrinsic in that hierarchy that specifies a non-bytes value for that parameter strikes me as a recipe for bugs.

I'm not really sure why you're messing with the signature of the intrinsic in the first place; we went through most of this design space when it was initially proposed.

It's being revisited with a long-term eye towards merging the unordered atomic semantics into the existing memcpy intrinsic. By making the proposed/new intrinsic's definition much closer to that of the existing llvm.memcpy it should be much easier to do that eventual merging of the two without having to revisit the semantic understanding of the unordered intrinsic in every analysis/transform.

[dneilson]

I'll make the suggested changes to the LangRef.

docs/LangRef.rst

13607

Easy enough to do if it's desired. Only reason that I didn't do that originally is that I'm already adding two parameters on top of memcpy for unordered memcpy; seemed like a good way to prevent that from becoming three additional parameters.

[dneilson]

• Addressed suggested changes to the LangRef doc for the intrinsic.
• Split out the is unordered parameter into two separate parameters -- dest_unordered & src_unordered.

[reames]

I'm going to let Daniel and Eli debate the general direction before reviewing further. I want to make sure Eli is on board with the general direction before we invest lots of effort in cleaning up the code.

[skatkov]

include/llvm/IR/IntrinsicInst.h

199

why not enum?

200

ARG_SOURCE? For consistency with other names.

205

ARG_SOURCE_UNORDERED? For consistency with other names.

343

bool isDestUnordered?

348

bool isSrcUnordered?

385

Don't you want to check (or assert) the constraints for the alignment here?

391

setSourceUnordered? For consistency with other names.

393

Don't you want to check (or assert) the constraints for the element size here?

include/llvm/IR/Intrinsics.td

810

you have two unordered arguments.

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4869–4870

You introduced the special getters, can you use them here?

lib/IR/Verifier.cpp

3997

is zero alignment allowed? At least it is strange that the 0 is a power of 2 :) if it is allowed, please update text "or zero"

[efriedma]

I don't know enough about the possible source languages to know with 100% certainty that it's not possible to mix, say, unordered loads with ordered stores.

It's almost certainly possible to mix them... even if the source language doesn't allow mixing them, we could transform unordered operations to non-atomic operations if we can prove the memory isn't accessed from another thread. (I don't think we actually do this transform at the moment, but LICM has a similar sort of check.)

The question is whether there's actually any optimization that would actually check the unordered bit. Maybe there is? (See my comments on __llvm_memcpy_element_unordered_atomic_*.)

Compatibility with the existing memcpy intrinsic. I have no problem removing the alignment parameter if that's the long-term goal/vision. My only concern is that I don't know whether that difference will lead me into trouble when I get to the point of adding the unordered atomic intrinsic to the MemTransferInst introspection class hierarchy.

Yes, this is the direction we want to go in. You should be able to hide the difference in the implementation of MemIntrinsic, I think. If it does cause problems, we can revisit.

Compatibility with memcpy semantics. The 'length' parameter of a MemTransferInst intrinsic is semantically understood by transforms/analysis to be the size of the transfer in bytes. Having an intrinsic in that hierarchy that specifies a non-bytes value for that parameter strikes me as a recipe for bugs.

Okay. It's a little ugly, but I guess it isn't that terrible.

docs/LangRef.rst

13600

"must" is kind of confusing in this context. Probably need to explicitly say "if len is not a multiple of element_size, the behavior is undefined", or something like that, since we can't actually tell until runtime.

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4887–4888

I'm not sure changing the signature of __llvm_memcpy_element_unordered_atomic_* like this makes sense.

I guess changing from the number of elements to the length in bytes is fine.

I'm not sure why you want to pass the alignment to the function; the implementation can easily compute the alignment itself based on the pointers passed in.

Not sure what the implementation is going to do with the DestUnordered and SrcUnordered parameters. Maybe if the source/dest is non-atomic, it could use unaligned load/store operations?

If we do need DestUnordered and SrcUnordered, it probably makes sense to merge them to save an instruction in the caller.

[dneilson]

docs/LangRef.rst

13582

I'm not seeing the difference here. "Sequence" doesn't imply any sort of ordering. So, to me the old and new are semantically equivalent -- there is a copy happening, and it's being done with unordered atomic load/stores.

13600

Fair. I'll make that change.

include/llvm/IR/IntrinsicInst.h

199

No particularly good reason. Just playing around.

200

Fair.

385

I figured that's handled by the verifier.

393

Handled by verifier.

include/llvm/IR/Intrinsics.td

810

Good catch! I updated the prototype, but neglected the comment.

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4887–4888

Fair. I have no strong preference with the library function taking size in bytes vs. number of elements. The only benefit that I can see to passing size in bytes instead of number of elements here is the possibility for runtime checks being implemented in a debug version of the library -- one that verifies that the length is a multiple of element size.

Good point about alignment. Not the smartest move on my part.

I *think* that you're right about passing dest_unordered & src_unordered to the library -- it might be unnecessary. The lib function could just be implemented assuming that both source & dest require unordered atomic ops; there shouldn't be any harm in it, since unordered just means that we can't break up an element into partial loads/stores, and we wouldn't want to do that in a high performance library anyways.

I'll change the lib prototype to match memcpy exactly.
_llvm_memcpy_unordered_atomic(i8* noalias dest, i8* noalias src, uint64 length)

lib/IR/Verifier.cpp

3997

I would think not, but this is the same check as exists for memcpy. So, for compatibility I think it should be allowed unless we can definitively say that it's not allowed for memcpy.

[dneilson]

Addressing suggestions.

[anna]

docs/LangRef.rst

13604–13605

Nit: if and only if stores to the destination buffer *are*

[skatkov]

include/llvm/IR/IntrinsicInst.h

385

ok, but to me it is one of a primary goal for setter to check incoming args.

lib/IR/Verifier.cpp

3991

Use getters?

[dneilson]

Addressing some comments -- use of getters, adding assertions to setters, and some minor wording changes to LangRef.

[skatkov]

lib/IR/Verifier.cpp

3992

Extra semicolon

[dneilson]

• Another iteration on the intrinsic prototype. I've removed the align, and dest/src_unordered arguments.
  • Having align both as arg attributes and as an arg could cause challenges if we need to resolve a difference, and it is the desired future direction for intrinsics.
  • Upon further thought, and digging into where passes would have to be made aware of this intrinsic -- I'm no longer convinced about the value of the separate dest_unordered/src_unordered args.
    • It seems sufficient to have the semantics of the intrinsic being that all loads/stores are unordered atomic; we can still "promote" idioms that mix, say, unordered loads with simple stores.
    • Any library implementation will just use unordered atomic loads & stores throughout, anyways.
    • There will be a side-effect of promoting simple ops to unordered-atomic if we recognize a loop idiom, and then later lower it into loads/stores. The tradeoff is that it should be easier to work with the intrinsic in passes.
    • The only value that I can see in having the separate dest_unordered/src_unordered args is that in lowering passes that change the intrinsic into explicit loads/stores we wouldn't "promote" a simple op into an unordered-atomic op.

• Added in updating the InstCombine lowering of the intrinsic so that the change doesn't lose functionality; even temporarily.

[efriedma]

docs/LangRef.rst

13573

Hmm... I didn't mention isvolatile earlier? See https://reviews.llvm.org/D27133?id=79305#629884 for original discussion of it. At the very least, we need a better description of what it means.

[skatkov]

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4887–4888

MI.getLength() == Length

lib/Transforms/InstCombine/InstCombineCalls.cpp

110

assert LengthInBytes % ElementSizeInBytes == 0 and LengthInBytes > 0?

[reames]

Did a scan through the code, didn't spot anything major. Once we settle the last few design/specification questions, this looks basically ready to go in.

docs/LangRef.rst

13573

I think we can just remove this. The original motivation was essentially future proofing, and I don't think it's worth keeping the complexity for now. We can change our minds later if it turns out we actually need this.

lib/Transforms/InstCombine/InstCombineCalls.cpp

110

Agreed. Also, length might actually be zero. We should remove such calls.

1900

Hm, I might sink this into the helper function. Optional, and can be submitted separately without further review.

[dneilson]

docs/LangRef.rst

13573

Agreed. I'll remove it.

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4887–4888

Not quite. Doesn't seem to be a straightforward way to go from an SDValue to a Type*, so I don't think this sort of replacement can be made.

lib/Transforms/InstCombine/InstCombineCalls.cpp

110

Re: The assert. I think that it would be better to check this in the verifier. In the LangRef we've said that it's undefined behaviour if length isn't a multiple of element size, so I think it's okay to blindly do this divide here and add a check to the verifier.

Re: Zero length; see the in-line comment below.

1900

This is following the same pattern/code-flow as the normal memcpy/memmove/memset handlers just above this. i.e. Check for a null length -- if there is one, then remove the call, else call the simplify method for the intrinsic. I'm inclined to stick to this pattern to make the later merging of the introspection classes cleaner.

[dneilson]

• Remove volatile arg from intrinsic.
• Add check to verifier to ensure that constant length is a multiple of element size & add corresponding test.

[reames]

LGTM w/one comment addressed before submission.

lib/Transforms/InstCombine/InstCombineCalls.cpp

116

Where did this check come from and why is it needed? It looks like an attempt to handle a length which isn't an even interval of element size, but the verifier should reject that?

[dneilson]

lib/Transforms/InstCombine/InstCombineCalls.cpp

116

Just me being extra cautious. The verifier checks for the case where constant length is not a multiple of element size, and the zero length case is handled elsewhere. However, I'm not sure that the verifier runs after every single pass. So, I figure there's no harm in handling the corner case.

[dneilson]

• Loop idiom patch was dropped, so update loop idiom recognition as well.

[dneilson]

include/llvm/IR/IntrinsicInst.h

197

I'm inclined to change this name to 'EUAMemcpyInst' to cut down on the length of its name. Any objections?

[dneilson]

• rebase

docs/LangRef.rst

 are described in :doc:`StackMaps`.
 
 Element Wise Atomic Memory Intrinsics
 -------------------------------------
 
 These intrinsics are similar to the standard library memory intrinsics except
 that they perform memory transfer as a sequence of atomic memory accesses.
 
-.. _int_memcpy_element_atomic:
+.. _int_memcpy_element_unordered_atomic:
 
-'``llvm.memcpy.element.atomic``' Intrinsic
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+'``llvm.memcpy.element.unordered.atomic``' Intrinsic
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
-This is an overloaded intrinsic. You can use ``llvm.memcpy.element.atomic`` on
+This is an overloaded intrinsic. You can use ``llvm.memcpy.element.unordered.atomic`` on
 any integer bit width and for different address spaces. Not all targets
 support all bit widths however.
 
 ::
 
-      declare void @llvm.memcpy.element.atomic.p0i8.p0i8(i8* <dest>, i8* <src>,
-                                              i64 <num_elements>, i32 <element_size>)
+      declare void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* <dest>,
+                                                                       i8* <src>,
+                                                                       i32 <len>,
+                                                                       i32 <align>,
+                                                                       i1 <isvolatile>,
+                                                                       i1 <dest_unordered>,
+                                                                       i1 <src_unordered>,
+                                                                       i8 <element_size>)
+      declare void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i64(i8* <dest>,
+                                                                       i8* <src>,
+                                                                       i64 <len>,
+                                                                       i32 <align>,
+                                                                       i1 <isvolatile>,
+                                                                       i1 <dest_unordered>,
+                                                                       i1 <src_unordered>,
+                                                                       i8 <element_size>)
 
 Overview:
 """""""""
 
-The '``llvm.memcpy.element.atomic.*``' intrinsic performs copy of a block of 
-memory from the source location to the destination location as a sequence of
-unordered atomic memory accesses where each access is a multiple of
-``element_size`` bytes wide and aligned at an element size boundary. For example
-each element is accessed atomically in source and destination buffers.
+The '``llvm.memcpy.element.unordered.atomic.*``' intrinsic is a specialization of the
+'``llvm.memcpy.*``' intrinsic. It differs in that the ``dest`` and ``src`` are treated
+as arrays with elements that are exactly ``element_size`` bytes, and the copy between
+buffers uses a sequence of :ref:`unordered atomic <ordering>` load/store operations
+that are a positive integer multiple of the ``element_size`` in size.
 
 Arguments:
 """"""""""
 
-The first argument is a pointer to the destination, the second is a
-pointer to the source. The third argument is an integer argument
-specifying the number of elements to copy, the fourth argument is size of
-the single element in bytes.
+The first five arguments are the same as they are in the :ref:`@llvm.memcpy <int_memcpy>`
+intrinsic, with the added constraint that ``len`` is required to be a positive integer
+multiple of the ``element_size``. If ``len`` is not a positive integer multiple of
+``element_size``, then the behaviour of the intrinsic is undefined.
 
-``element_size`` should be a power of two, greater than zero and less than
-a target-specific atomic access size limit.
+``dest_unordered`` is ``true`` if and only if stores to the destination buffer are
+unordered atomic stores.
 
-For each of the input pointers ``align`` parameter attribute must be specified.
-It must be a power of two and greater than or equal to the ``element_size``.
-Caller guarantees that both the source and destination pointers are aligned to
-that boundary.
+``src_unordered`` is ``true`` if and only if loads from the source buffer are
+unordered atomic loads.
+
+``element_size`` must be a compile-time constant positive power of two no greater than
+target-specific atomic access size limit.
+
+For each of the input pointers ``align`` parameter attribute must be specified. It
+must be a power of two and greater than or equal to the ``element_size``. Caller
+guarantees that both the source and destination pointers are aligned to that boundary.
 
 Semantics:
 """"""""""
 
-The '``llvm.memcpy.element.atomic.*``' intrinsic copies
-'``num_elements`` * ``element_size``' bytes of memory from the source location to
-the destination location. These locations are not allowed to overlap. Memory copy
-is performed as a sequence of unordered atomic memory accesses where each access
-is guaranteed to be a multiple of ``element_size`` bytes wide and aligned at an
-element size boundary.
+The '``llvm.memcpy.element.unordered.atomic.*``' intrinsic copies ``len`` bytes of
+memory from the source location to the destination location. These locations are not
+allowed to overlap. The memory copy is performed as a sequence of load/store operations
+where each access is guaranteed to be a multiple of ``element_size`` bytes wide and
+aligned at an ``element_size`` boundary. 
 
 The order of the copy is unspecified. The same value may be read from the source
 buffer many times, but only one write is issued to the destination buffer per
-element. It is well defined to have concurrent reads and writes to both source
-and destination provided those reads and writes are at least unordered atomic.
+element. It is well defined to have concurrent reads and writes to both source and
+destination provided those reads and writes are unordered atomic when specified.
 
 This intrinsic does not provide any additional ordering guarantees over those
 provided by a set of unordered loads from the source location and stores to the
 destination.
 
 Lowering:
 """""""""
 
-In the most general case call to the '``llvm.memcpy.element.atomic.*``' is lowered
-to a call to the symbol ``__llvm_memcpy_element_atomic_*``. Where '*' is replaced
-with an actual element size.
+In the most general case call to the '``llvm.memcpy.element.unordered.atomic.*``' is
+lowered to a call to the symbol ``__llvm_memcpy_element_unordered_atomic_*``. Where '*'
+is replaced with an actual element size.
 
-Optimizer is allowed to inline memory copy when it's profitable to do so.
+The optimizer is allowed to inline the memory copy when it's profitable to do so.

include/llvm/CodeGen/RuntimeLibcalls.h

enum Libcall {
     O_F128,
     O_PPCF128,
 
     // MEMORY
     MEMCPY,
     MEMSET,
     MEMMOVE,
 
-    // ELEMENT-WISE ATOMIC MEMORY
-    MEMCPY_ELEMENT_ATOMIC_1,
-    MEMCPY_ELEMENT_ATOMIC_2,
-    MEMCPY_ELEMENT_ATOMIC_4,
-    MEMCPY_ELEMENT_ATOMIC_8,
-    MEMCPY_ELEMENT_ATOMIC_16,
+    // ELEMENT-WISE UNORDERED-ATOMIC MEMORY of different element sizes
+    MEMCPY_ELEMENT_UNORDERED_ATOMIC_1,
+    MEMCPY_ELEMENT_UNORDERED_ATOMIC_2,
+    MEMCPY_ELEMENT_UNORDERED_ATOMIC_4,
+    MEMCPY_ELEMENT_UNORDERED_ATOMIC_8,
+    MEMCPY_ELEMENT_UNORDERED_ATOMIC_16,
 
     // EXCEPTION HANDLING
     UNWIND_RESUME,
 
     // Note: there's two sets of atomics libcalls; see
     // <http://llvm.org/docs/Atomics.html> for more info on the
     // difference between them.
 
namespace RTLIB {
   /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or
   /// UNKNOWN_LIBCALL if there is none.
   Libcall getUINTTOFP(EVT OpVT, EVT RetVT);
 
   /// Return the SYNC_FETCH_AND_* value for the given opcode and type, or
   /// UNKNOWN_LIBCALL if there is none.
   Libcall getSYNC(unsigned Opc, MVT VT);
 
-  /// getMEMCPY_ELEMENT_ATOMIC - Return MEMCPY_ELEMENT_ATOMIC_* value for the
-  /// given element size or UNKNOW_LIBCALL if there is none.
-  Libcall getMEMCPY_ELEMENT_ATOMIC(uint64_t ElementSize);
+  /// getMEMCPY_ELEMENT_UNORDERED_ATOMIC - Return
+  /// MEMCPY_ELEMENT_UNORDERED_ATOMIC_* value for the given element size or
+  /// UNKNOW_LIBCALL if there is none.
+  Libcall getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize);
 }
 }
 
 #endif

include/llvm/IR/IntrinsicInst.h

static inline bool classof(const IntrinsicInst *I) {
       }
     }
     static inline bool classof(const Value *V) {
       return isa<IntrinsicInst>(V) && classof(cast<IntrinsicInst>(V));
     }
   };
 
   /// This class represents atomic memcpy intrinsic
-  /// TODO: Integrate this class into MemIntrinsic hierarchy.
-  class ElementAtomicMemCpyInst : public IntrinsicInst {
+  /// TODO: Integrate this class into MemIntrinsic hierarchy; for now this is
+  /// C&P of all methods from that hierarchy
+  class ElementUnorderedAtomicMemCpyInst : public IntrinsicInst {
+  private:
+    enum {
+      ARG_DEST = 0,
+      ARG_SOURCE = 1,
+      ARG_LENGTH = 2,
+      ARG_ALIGN = 3,
+      ARG_VOLATILE = 4,
+      ARG_DEST_UNORDERED = 5,
+      ARG_SOURCE_UNORDERED = 6,
+      ARG_ELEMENTSIZE = 7
+    };
+
   public:
-    Value *getRawDest() const { return getArgOperand(0); }
-    Value *getRawSource() const { return getArgOperand(1); }
+    Value *getRawDest() const {
+      return const_cast<Value *>(getArgOperand(ARG_DEST));
+    }
+    const Use &getRawDestUse() const { return getArgOperandUse(ARG_DEST); }
+    Use &getRawDestUse() { return getArgOperandUse(ARG_DEST); }
+
+    /// Return the arguments to the instruction.
+    Value *getRawSource() const {
+      return const_cast<Value *>(getArgOperand(ARG_SOURCE));
+    }
+    const Use &getRawSourceUse() const { return getArgOperandUse(ARG_SOURCE); }
+    Use &getRawSourceUse() { return getArgOperandUse(ARG_SOURCE); }
+
+    Value *getLength() const {
+      return const_cast<Value *>(getArgOperand(ARG_LENGTH));
+    }
+    const Use &getLengthUse() const { return getArgOperandUse(ARG_LENGTH); }
+    Use &getLengthUse() { return getArgOperandUse(ARG_LENGTH); }
+
+    Value *getRawAlignment() const {
+      return const_cast<Value *>(getArgOperand(ARG_ALIGN));
+    }
+
+    ConstantInt *getAlignmentCst() const {
+      return cast<ConstantInt>(getRawAlignment());
+    }
+
+    unsigned getAlignment() const { return getAlignmentCst()->getZExtValue(); }
+
+    Type *getAlignmentType() const { return getRawAlignment()->getType(); }
+
+    ConstantInt *getVolatileCst() const {
+      return cast<ConstantInt>(
+          const_cast<Value *>(getArgOperand(ARG_VOLATILE)));
+    }
+
+    bool isVolatile() const { return !getVolatileCst()->isZero(); }
+
+    Value *getRawDestUnordered() const {
+      return const_cast<Value *>(getArgOperand(ARG_DEST_UNORDERED));
+    }
+
+    ConstantInt *getDestUnorderedCst() const {
+      return cast<ConstantInt>(getRawDestUnordered());
+    }
+
+    bool isDestUnordered() const { return !getDestUnorderedCst()->isZero(); }
+
+    Value *getRawSourceUnordered() const {
+      return const_cast<Value *>(getArgOperand(ARG_SOURCE_UNORDERED));
+    }
+
+    ConstantInt *getSourceUnorderedCst() const {
+      return cast<ConstantInt>(getRawSourceUnordered());
+    }
+
+    bool isSrcUnordered() const { return !getSourceUnorderedCst()->isZero(); }
+
+    Value *getRawElementSizeInBytes() const {
+      return const_cast<Value *>(getArgOperand(ARG_ELEMENTSIZE));
+    }
+
+    ConstantInt *getElementSizeInBytesCst() const {
+      return cast<ConstantInt>(getRawElementSizeInBytes());
+    }
+
+    uint8_t getElementSizeInBytes() const {
+      return getElementSizeInBytesCst()->getZExtValue();
+    }
+
+    /// This is just like getRawDest, but it strips off any cast
+    /// instructions that feed it, giving the original input.  The returned
+    /// value is guaranteed to be a pointer.
+    Value *getDest() const { return getRawDest()->stripPointerCasts(); }
+
+    /// This is just like getRawSource, but it strips off any cast
+    /// instructions that feed it, giving the original input.  The returned
+    /// value is guaranteed to be a pointer.
+    Value *getSource() const { return getRawSource()->stripPointerCasts(); }
+
+    unsigned getDestAddressSpace() const {
+      return cast<PointerType>(getRawDest()->getType())->getAddressSpace();
+    }
+
+    unsigned getSourceAddressSpace() const {
+      return cast<PointerType>(getRawSource()->getType())->getAddressSpace();
+    }
 
-    Value *getNumElements() const { return getArgOperand(2); }
-    void setNumElements(Value *V) { setArgOperand(2, V); }
+    /// Set the specified arguments of the instruction.
+    void setDest(Value *Ptr) {
+      assert(getRawDest()->getType() == Ptr->getType() &&
+             "setDest called with pointer of wrong type!");
+      setArgOperand(ARG_DEST, Ptr);
+    }
 
-    uint64_t getSrcAlignment() const { return getParamAlignment(0); }
-    uint64_t getDstAlignment() const { return getParamAlignment(1); }
+    void setSource(Value *Ptr) {
+      assert(getRawSource()->getType() == Ptr->getType() &&
+             "setSource called with pointer of wrong type!");
+      setArgOperand(ARG_SOURCE, Ptr);
+    }
+
+    void setLength(Value *L) {
+      assert(getLength()->getType() == L->getType() &&
+             "setLength called with value of wrong type!");
+      setArgOperand(ARG_LENGTH, L);
+    }
+
+    void setAlignment(Constant *A) {
+      assert(A->getType() == Type::getInt32Ty(getContext()) &&
+             "setAlignment called with value of wrong type!");
+      setArgOperand(ARG_ALIGN, A);
+    }
+
+    void setVolatile(Constant *V) {
+      assert(V->getType() == Type::getInt1Ty(getContext()) &&
+             "setVolatile called with value of wrong type!");
+      setArgOperand(ARG_VOLATILE, V);
+    }
+
+    void setDestUnordered(Constant *V) {
+      assert(V->getType() == Type::getInt1Ty(getContext()) &&
+             "setDestUnordered called with value of wrong type!");
+      setArgOperand(ARG_DEST_UNORDERED, V);
+    }
+
+    void setSourceUnordered(Constant *V) {
+      assert(V->getType() == Type::getInt1Ty(getContext()) &&
+             "setSourceUnordered called with value of wrong type!");
+      setArgOperand(ARG_SOURCE_UNORDERED, V);
+    }
 
-    uint64_t getElementSizeInBytes() const {
-      Value *Arg = getArgOperand(3);
-      return cast<ConstantInt>(Arg)->getZExtValue();
+    void setElementSizeInBytes(Constant *V) {
+      assert(V->getType() == Type::getInt8Ty(getContext()) &&
+             "setElementSizeInBytes called with value of wrong type!");
+      setArgOperand(ARG_ELEMENTSIZE, V);
     }
 
     static inline bool classof(const IntrinsicInst *I) {
-      return I->getIntrinsicID() == Intrinsic::memcpy_element_atomic;
+      return I->getIntrinsicID() == Intrinsic::memcpy_element_unordered_atomic;
     }
     static inline bool classof(const Value *V) {
       return isa<IntrinsicInst>(V) && classof(cast<IntrinsicInst>(V));
     }
   };
 
   /// This is the common base class for memset/memcpy/memmove.
   class MemIntrinsic : public IntrinsicInst {
public:
     }
 
     bool isVolatile() const {
       return !getVolatileCst()->isZero();
     }
 
     unsigned getDestAddressSpace() const {
       return cast<PointerType>(getRawDest()->getType())->getAddressSpace();
     }
 
     /// This is just like getRawDest, but it strips off any cast
     /// instructions that feed it, giving the original input.  The returned
     /// value is guaranteed to be a pointer.
     Value *getDest() const { return getRawDest()->stripPointerCasts(); }
 
     /// Set the specified arguments of the instruction.
     void setDest(Value *Ptr) {
       assert(getRawDest()->getType() == Ptr->getType() &&
              "setDest called with pointer of wrong type!");
       setArgOperand(0, Ptr);
     }
 
     void setLength(Value *L) {
       assert(getLength()->getType() == L->getType() &&
              "setLength called with value of wrong type!");

include/llvm/IR/Intrinsics.td

 // Takes a pointer to a string and the length of the string.
 def int_xray_customevent : Intrinsic<[], [llvm_ptr_ty, llvm_i32_ty],
                                      [NoCapture<0>, ReadOnly<0>, IntrWriteMem]>;
 //===----------------------------------------------------------------------===//
 
 //===------ Memory intrinsics with element-wise atomicity guarantees ------===//
 //
 
-def int_memcpy_element_atomic  : Intrinsic<[],
-                                           [llvm_anyptr_ty, llvm_anyptr_ty,
-                                            llvm_i64_ty, llvm_i32_ty],
-                                 [IntrArgMemOnly, NoCapture<0>, NoCapture<1>,
-                                  WriteOnly<0>, ReadOnly<1>]>;
+// llvm.memcpy.element.unordered.atomic(dest, src, length, alignment, volatile,
+// dest_unordered, source_unordered, elementsize)
+def int_memcpy_element_unordered_atomic
+    : Intrinsic<[],
+                [
+                  llvm_anyptr_ty, llvm_anyptr_ty, llvm_anyint_ty, llvm_i32_ty,
+                  llvm_i1_ty, llvm_i1_ty, llvm_i1_ty, llvm_i8_ty
+                ],
+                [
+                  IntrArgMemOnly, NoCapture<0>, NoCapture<1>, WriteOnly<0>,
+                  ReadOnly<1>
+                ]>;
 
 //===------------------------ Reduction Intrinsics ------------------------===//
 //
 def int_experimental_vector_reduce_fadd : Intrinsic<[llvm_anyfloat_ty],
                                                     [llvm_anyfloat_ty,
                                                      llvm_anyvector_ty],
                                                     [IntrNoMem]>;
 def int_experimental_vector_reduce_fmul : Intrinsic<[llvm_anyfloat_ty],

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

if (!Align)
       Align = 1; // @llvm.memmove defines 0 and 1 to both mean no alignment.
     bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
     bool isTC = I.isTailCall() && isInTailCallPosition(&I, DAG.getTarget());
     SDValue MM = DAG.getMemmove(getRoot(), sdl, Op1, Op2, Op3, Align, isVol,
                                 isTC, MachinePointerInfo(I.getArgOperand(0)),
                                 MachinePointerInfo(I.getArgOperand(1)));
     updateDAGForMaybeTailCall(MM);
     return nullptr;
   }
-  case Intrinsic::memcpy_element_atomic: {
-    SDValue Dst = getValue(I.getArgOperand(0));
-    SDValue Src = getValue(I.getArgOperand(1));
-    SDValue NumElements = getValue(I.getArgOperand(2));
-    SDValue ElementSize = getValue(I.getArgOperand(3));
+  case Intrinsic::memcpy_element_unordered_atomic: {
+    const ElementUnorderedAtomicMemCpyInst &MI =
+        cast<ElementUnorderedAtomicMemCpyInst>(I);
+    SDValue Dst = getValue(MI.getRawDest());
+    SDValue Src = getValue(MI.getRawSource());
+    SDValue Length = getValue(MI.getLength());
 
     // Emit a library call.
     TargetLowering::ArgListTy Args;
     TargetLowering::ArgListEntry Entry;
     Entry.Ty = DAG.getDataLayout().getIntPtrType(*DAG.getContext());
     Entry.Node = Dst;
     Args.push_back(Entry);
 
     Entry.Node = Src;
     Args.push_back(Entry);
 
-    Entry.Ty = I.getArgOperand(2)->getType();
-    Entry.Node = NumElements;
-    Args.push_back(Entry);
-
-    Entry.Ty = Type::getInt32Ty(*DAG.getContext());
-    Entry.Node = ElementSize;
+    Entry.Ty = MI.getLength()->getType();
+    Entry.Node = Length;
     Args.push_back(Entry);
 
-    uint64_t ElementSizeConstant =
-        cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
+    uint64_t ElementSizeConstant = MI.getElementSizeInBytes();
     RTLIB::Libcall LibraryCall =
-        RTLIB::getMEMCPY_ELEMENT_ATOMIC(ElementSizeConstant);
+        RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(ElementSizeConstant);
     if (LibraryCall == RTLIB::UNKNOWN_LIBCALL)
       report_fatal_error("Unsupported element size");
 
     TargetLowering::CallLoweringInfo CLI(DAG);
     CLI.setDebugLoc(sdl).setChain(getRoot()).setLibCallee(
         TLI.getLibcallCallingConv(LibraryCall),
         Type::getVoidTy(*DAG.getContext()),
         DAG.getExternalSymbol(TLI.getLibcallName(LibraryCall),

lib/CodeGen/TargetLoweringBase.cpp

static void InitLibcallNames(const char **Names, const Triple &TT) {
   Names[RTLIB::UO_PPCF128] = "__gcc_qunord";
   Names[RTLIB::O_F32] = "__unordsf2";
   Names[RTLIB::O_F64] = "__unorddf2";
   Names[RTLIB::O_F128] = "__unordtf2";
   Names[RTLIB::O_PPCF128] = "__gcc_qunord";
   Names[RTLIB::MEMCPY] = "memcpy";
   Names[RTLIB::MEMMOVE] = "memmove";
   Names[RTLIB::MEMSET] = "memset";
-  Names[RTLIB::MEMCPY_ELEMENT_ATOMIC_1] = "__llvm_memcpy_element_atomic_1";
-  Names[RTLIB::MEMCPY_ELEMENT_ATOMIC_2] = "__llvm_memcpy_element_atomic_2";
-  Names[RTLIB::MEMCPY_ELEMENT_ATOMIC_4] = "__llvm_memcpy_element_atomic_4";
-  Names[RTLIB::MEMCPY_ELEMENT_ATOMIC_8] = "__llvm_memcpy_element_atomic_8";
-  Names[RTLIB::MEMCPY_ELEMENT_ATOMIC_16] = "__llvm_memcpy_element_atomic_16";
+  Names[RTLIB::MEMCPY_ELEMENT_UNORDERED_ATOMIC_1] =
+      "__llvm_memcpy_element_unordered_atomic_1";
+  Names[RTLIB::MEMCPY_ELEMENT_UNORDERED_ATOMIC_2] =
+      "__llvm_memcpy_element_unordered_atomic_2";
+  Names[RTLIB::MEMCPY_ELEMENT_UNORDERED_ATOMIC_4] =
+      "__llvm_memcpy_element_unordered_atomic_4";
+  Names[RTLIB::MEMCPY_ELEMENT_UNORDERED_ATOMIC_8] =
+      "__llvm_memcpy_element_unordered_atomic_8";
+  Names[RTLIB::MEMCPY_ELEMENT_UNORDERED_ATOMIC_16] =
+      "__llvm_memcpy_element_unordered_atomic_16";
   Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume";
   Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1] = "__sync_val_compare_and_swap_1";
   Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2] = "__sync_val_compare_and_swap_2";
   Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4] = "__sync_val_compare_and_swap_4";
   Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8] = "__sync_val_compare_and_swap_8";
   Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_16] = "__sync_val_compare_and_swap_16";
   Names[RTLIB::SYNC_LOCK_TEST_AND_SET_1] = "__sync_lock_test_and_set_1";
   Names[RTLIB::SYNC_LOCK_TEST_AND_SET_2] = "__sync_lock_test_and_set_2";
switch (Opc) {
     OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN)
   }
 
 #undef OP_TO_LIBCALL
 
   return UNKNOWN_LIBCALL;
 }
 
-RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_ATOMIC(uint64_t ElementSize) {
+RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) {
   switch (ElementSize) {
   case 1:
-    return MEMCPY_ELEMENT_ATOMIC_1;
+    return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1;
   case 2:
-    return MEMCPY_ELEMENT_ATOMIC_2;
+    return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2;
   case 4:
-    return MEMCPY_ELEMENT_ATOMIC_4;
+    return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4;
   case 8:
-    return MEMCPY_ELEMENT_ATOMIC_8;
+    return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8;
   case 16:
-    return MEMCPY_ELEMENT_ATOMIC_16;
+    return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16;
   default:
     return UNKNOWN_LIBCALL;
   }
-
 }
 
 /// InitCmpLibcallCCs - Set default comparison libcall CC.
 ///
 static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
   memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL);
   CCs[RTLIB::OEQ_F32] = ISD::SETEQ;
   CCs[RTLIB::OEQ_F64] = ISD::SETEQ;

lib/IR/Verifier.cpp

case Intrinsic::memset: {
     const APInt &AlignVal = AlignCI->getValue();
     Assert(AlignCI->isZero() || AlignVal.isPowerOf2(),
            "alignment argument of memory intrinsics must be a power of 2", CS);
     Assert(isa<ConstantInt>(CS.getArgOperand(4)),
            "isvolatile argument of memory intrinsics must be a constant int",
            CS);
     break;
   }
-  case Intrinsic::memcpy_element_atomic: {
-    ConstantInt *ElementSizeCI = dyn_cast<ConstantInt>(CS.getArgOperand(3));
-    Assert(ElementSizeCI, "element size of the element-wise atomic memory "
-                          "intrinsic must be a constant int",
+  case Intrinsic::memcpy_element_unordered_atomic: {
+    const ElementUnorderedAtomicMemCpyInst *MI =
+      cast<ElementUnorderedAtomicMemCpyInst>(CS.getInstruction());;
+    ConstantInt *AlignCI = dyn_cast<ConstantInt>(MI->getRawAlignment());
+    Assert(AlignCI,
+           "alignment argument of element-wise unordered atomic memory "
+           "intrinsics must be a constant int",
+           CS);
+    const APInt &AlignVal = AlignCI->getValue();
+    Assert(AlignCI->isZero() || AlignVal.isPowerOf2(),
+           "alignment argument of element-wise unordered atomic memory "
+           "intrinsics must be a power of 2",
+           CS);
+
+    ConstantInt *DestUnorderedCI = dyn_cast<ConstantInt>(MI->getRawDestUnordered());
+    Assert(DestUnorderedCI,
+           "dest_unordered of the element-wise unordered atomic memory "
+           "intrinsic must be a constant int",
+           CS);
+
+    ConstantInt *SrcUnorderedCI = dyn_cast<ConstantInt>(MI->getRawSourceUnordered());
+    Assert(SrcUnorderedCI,
+           "src_unordered of the element-wise unordered atomic memory "
+           "intrinsic must be a constant int",
+           CS);
+
+    // Cannot have both unordered flags being false.
+    Assert(!(DestUnorderedCI->isZero() && SrcUnorderedCI->isZero()),
+           "dest_unordered and src_unordered cannot both be zero on the "
+           "element-wise unordered atomic memory intrinsic",
+           CS);
+
+    ConstantInt *ElementSizeCI = dyn_cast<ConstantInt>(MI->getRawElementSizeInBytes());
+    Assert(ElementSizeCI,
+           "element size of the element-wise unordered atomic memory "
            CS);
     const APInt &ElementSizeVal = ElementSizeCI->getValue();
     Assert(ElementSizeVal.isPowerOf2(),
            "element size of the element-wise atomic memory intrinsic "
            "must be a power of 2",
            CS);
 
     auto IsValidAlignment = [&](uint64_t Alignment) {
       return isPowerOf2_64(Alignment) && ElementSizeVal.ule(Alignment);
     };
-
     uint64_t DstAlignment = CS.getParamAlignment(0),
              SrcAlignment = CS.getParamAlignment(1);
-
     Assert(IsValidAlignment(DstAlignment),
-           "incorrect alignment of the destination argument",
-           CS);
+           "incorrect alignment of the destination argument", CS);
     Assert(IsValidAlignment(SrcAlignment),
-           "incorrect alignment of the source argument",
-           CS);
+           "incorrect alignment of the source argument", CS);
     break;
   }
   case Intrinsic::gcroot:
   case Intrinsic::gcwrite:
   case Intrinsic::gcread:
     if (ID == Intrinsic::gcroot) {
       AllocaInst *AI =
         dyn_cast<AllocaInst>(CS.getArgOperand(0)->stripPointerCasts());

lib/Transforms/InstCombine/InstCombineCalls.cpp

for (unsigned I = 0, E = V->getNumElements(); I != E; ++I) {
     bool Sign = V->getElementType()->isIntegerTy()
                     ? cast<ConstantInt>(Elt)->isNegative()
                     : cast<ConstantFP>(Elt)->isNegative();
     BoolVec.push_back(ConstantInt::get(BoolTy, Sign));
   }
   return ConstantVector::get(BoolVec);
 }
 
+/* -- temp removal to aid staging
 Instruction *
 InstCombiner::SimplifyElementAtomicMemCpy(ElementAtomicMemCpyInst *AMI) {
   // Try to unfold this intrinsic into sequence of explicit atomic loads and
   // stores.
   // First check that number of elements is compile time constant.
   auto *NumElementsCI = dyn_cast<ConstantInt>(AMI->getNumElements());
   if (!NumElementsCI)
     return nullptr;
 
   // Check that there are not too many elements.
   uint64_t NumElements = NumElementsCI->getZExtValue();
   if (NumElements >= UnfoldElementAtomicMemcpyMaxElements)
     return nullptr;
 
   // Don't unfold into illegal integers
   uint64_t ElementSizeInBytes = AMI->getElementSizeInBytes() * 8;
   if (!getDataLayout().isLegalInteger(ElementSizeInBytes))
     return nullptr;
 
   // Cast source and destination to the correct type. Intrinsic input arguments
   // are usually represented as i8*.
   // Often operands will be explicitly casted to i8* and we can just strip
   // those casts instead of inserting new ones. However it's easier to rely on
   // other InstCombine rules which will cover trivial cases anyway.
   Value *Src = AMI->getRawSource();
   Value *Dst = AMI->getRawDest();
   Type *ElementPointerType = Type::getIntNPtrTy(
for (uint64_t i = 0; i < NumElements; ++i) {
     Store->setDebugLoc(AMI->getDebugLoc());
   }
 
   // Set the number of elements of the copy to 0, it will be deleted on the
   // next iteration.
   AMI->setNumElements(Constant::getNullValue(NumElementsCI->getType()));
   return AMI;
 }
+*/
 
 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
   unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), DL, MI, &AC, &DT);
   unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), DL, MI, &AC, &DT);
   unsigned MinAlign = std::min(DstAlign, SrcAlign);
   unsigned CopyAlign = MI->getAlignment();
 
   if (CopyAlign < MinAlign) {
if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
     } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
       if (Instruction *I = SimplifyMemSet(MSI))
         return I;
     }
 
     if (Changed) return II;
   }
 
+  /* -- temp removal to simplify staging
   if (auto *AMI = dyn_cast<ElementAtomicMemCpyInst>(II)) {
     if (Constant *C = dyn_cast<Constant>(AMI->getNumElements()))
       if (C->isNullValue())
         return eraseInstFromFunction(*AMI);
 
     if (Instruction *I = SimplifyElementAtomicMemCpy(AMI))
       return I;
   }
+  */
   
   if (Instruction *I = SimplifyNVVMIntrinsic(II, *this))
     return I;
 
   auto SimplifyDemandedVectorEltsLow = [this](Value *Op, unsigned Width,
                                               unsigned DemandedWidth) {
     APInt UndefElts(Width, 0);
     APInt DemandedElts = APInt::getLowBitsSet(Width, DemandedWidth);
     return SimplifyDemandedVectorElts(Op, DemandedElts, UndefElts);

lib/Transforms/InstCombine/InstCombineInternal.h

Instruction *OptAndOp(BinaryOperator *Op, ConstantInt *OpRHS,
                         ConstantInt *AndRHS, BinaryOperator &TheAnd);
 
   Value *insertRangeTest(Value *V, const APInt &Lo, const APInt &Hi,
                          bool isSigned, bool Inside);
   Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocaInst &AI);
   Instruction *MatchBSwap(BinaryOperator &I);
   bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
 
-  Instruction *SimplifyElementAtomicMemCpy(ElementAtomicMemCpyInst *AMI);
+  // Instruction *SimplifyElementAtomicMemCpy(ElementAtomicMemCpyInst *AMI); --
+  // temp removal to aid staging
   Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
   Instruction *SimplifyMemSet(MemSetInst *MI);
 
   Value *EvaluateInDifferentType(Value *V, Type *Ty, bool isSigned);
 
   /// \brief Returns a value X such that Val = X * Scale, or null if none.
   ///
   /// If the multiplication is known not to overflow then NoSignedWrap is set.
   Value *Descale(Value *Val, APInt Scale, bool &NoSignedWrap);
 };
 
 } // end namespace llvm.
 
 #undef DEBUG_TYPE
 
 #endif

test/CodeGen/X86/element-wise-atomic-memory-intrinsics.ll

 ; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu | FileCheck %s
 
 define i8* @test_memcpy1(i8* %P, i8* %Q) {
   ; CHECK: test_memcpy
-  call void @llvm.memcpy.element.atomic.p0i8.p0i8(i8* align 4 %P, i8* align 4 %Q, i64 1, i32 1)
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 %P, i8* align 4 %Q, i32 1, i32 4, i1 0, i1 1, i1 1, i8 1)
   ret i8* %P
+  ; 3rd arg (%edx) -- length
   ; CHECK-DAG: movl $1, %edx
-  ; CHECK-DAG: movl $1, %ecx
-  ; CHECK: __llvm_memcpy_element_atomic_1
+  ; CHECK: __llvm_memcpy_element_unordered_atomic_1
 }
 
 define i8* @test_memcpy2(i8* %P, i8* %Q) {
   ; CHECK: test_memcpy2
-  call void @llvm.memcpy.element.atomic.p0i8.p0i8(i8* align 4 %P, i8* align 4 %Q, i64 2, i32 2)
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 %P, i8* align 4 %Q, i32 2, i32 4, i1 0, i1 1, i1 1, i8 2)
   ret i8* %P
+  ; 3rd arg (%edx) -- length
   ; CHECK-DAG: movl $2, %edx
-  ; CHECK-DAG: movl $2, %ecx
-  ; CHECK: __llvm_memcpy_element_atomic_2
+  ; CHECK: __llvm_memcpy_element_unordered_atomic_2
 }
 
 define i8* @test_memcpy4(i8* %P, i8* %Q) {
   ; CHECK: test_memcpy4
-  call void @llvm.memcpy.element.atomic.p0i8.p0i8(i8* align 4 %P, i8* align 4 %Q, i64 4, i32 4)
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 %P, i8* align 4 %Q, i32 4, i32 4, i1 0, i1 1, i1 1, i8 4)
   ret i8* %P
+  ; 3rd arg (%edx) -- length
   ; CHECK-DAG: movl $4, %edx
-  ; CHECK-DAG: movl $4, %ecx
-  ; CHECK: __llvm_memcpy_element_atomic_4
+  ; CHECK: __llvm_memcpy_element_unordered_atomic_4
 }
 
 define i8* @test_memcpy8(i8* %P, i8* %Q) {
   ; CHECK: test_memcpy8
-  call void @llvm.memcpy.element.atomic.p0i8.p0i8(i8* align 8 %P, i8* align 8 %Q, i64 8, i32 8)
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 8 %P, i8* align 8 %Q, i32 8, i32 8, i1 0, i1 1, i1 1, i8 8)
   ret i8* %P
+  ; 3rd arg (%edx) -- length
   ; CHECK-DAG: movl $8, %edx
-  ; CHECK-DAG: movl $8, %ecx
-  ; CHECK: __llvm_memcpy_element_atomic_8
+  ; CHECK: __llvm_memcpy_element_unordered_atomic_8
 }
 
 define i8* @test_memcpy16(i8* %P, i8* %Q) {
   ; CHECK: test_memcpy16
-  call void @llvm.memcpy.element.atomic.p0i8.p0i8(i8* align 16 %P, i8* align 16 %Q, i64 16, i32 16)
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 16 %P, i8* align 16 %Q, i32 16, i32 16, i1 0, i1 1, i1 1, i8 16)
   ret i8* %P
+  ; 3rd arg (%edx) -- length
   ; CHECK-DAG: movl $16, %edx
-  ; CHECK-DAG: movl $16, %ecx
-  ; CHECK: __llvm_memcpy_element_atomic_16
+  ; CHECK: __llvm_memcpy_element_unordered_atomic_16
 }
 
 define void @test_memcpy_args(i8** %Storage) {
   ; CHECK: test_memcpy_args
   %Dst = load i8*, i8** %Storage
   %Src.addr = getelementptr i8*, i8** %Storage, i64 1
   %Src = load i8*, i8** %Src.addr
 
-  ; First argument
+  ; 1st arg (%rdi)
   ; CHECK-DAG: movq (%rdi), [[REG1:%r.+]]
   ; CHECK-DAG: movq [[REG1]], %rdi
-  ; Second argument
+  ; 2nd arg (%rsi)
   ; CHECK-DAG: movq 8(%rdi), %rsi
-  ; Third argument
+  ; 3rd arg (%edx) -- length
   ; CHECK-DAG: movl $4, %edx
-  ; Fourth argument
-  ; CHECK-DAG: movl $4, %ecx
-  ; CHECK: __llvm_memcpy_element_atomic_4
-  call void @llvm.memcpy.element.atomic.p0i8.p0i8(i8* align 4 %Dst, i8* align 4 %Src, i64 4, i32 4)
-  ret void
+  ; CHECK: __llvm_memcpy_element_unordered_atomic_4
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 %Dst, i8* align 4 %Src, i32 4, i32 4, i1 0, i1 1, i1 1, i8 4)  ret void
 }
 
-declare void @llvm.memcpy.element.atomic.p0i8.p0i8(i8* nocapture, i8* nocapture, i64, i32) nounwind
+declare void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* nocapture, i8* nocapture, i32, i32, i1, i1, i1, i8) nounwind

test/Transforms/InstCombine/element-atomic-memcpy-to-loads.ll

 ; RUN: opt -instcombine -unfold-element-atomic-memcpy-max-elements=8 -S < %s | FileCheck %s
+; Temporarily an expected failure until inst combine is updated in the next patch
+; XFAIL: *
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 
 ; Test basic unfolding
 define void @test1(i8* %Src, i8* %Dst) {
 ; CHECK-LABEL: test1
 ; CHECK-NOT: llvm.memcpy.element.atomic
 
 ; CHECK-DAG: %memcpy_unfold.src_casted = bitcast i8* %Src to i32*

test/Verifier/element-wise-atomic-memory-intrinsics.ll

 ; RUN: not opt -verify < %s 2>&1 | FileCheck %s
 
-define void @test_memcpy(i8* %P, i8* %Q) {
+define void @test_memcpy(i8* %P, i8* %Q, i32 %A, i8 %E, i1 %V) {
+  ; CHECK: alignment argument of element-wise unordered atomic memory intrinsics must be a constant int
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 %P, i8* align 4 %Q, i32 1, i32 %A, i1 0, i1 1, i1 1, i8 1)
+
+  ; CHECK: alignment argument of element-wise unordered atomic memory intrinsics must be a power of 2
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 %P, i8* align 4 %Q, i32 1, i32 5, i1 0, i1 1, i1 1, i8 1)
+
+  ; CHECK: dest_unordered of the element-wise unordered atomic memory intrinsic must be a constant int
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 %P, i8* align 4 %Q, i32 1, i32 4, i1 0, i1 %V, i1 1, i8 1)
+
+  ; CHECK: src_unordered of the element-wise unordered atomic memory intrinsic must be a constant int
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 %P, i8* align 4 %Q, i32 1, i32 4, i1 0, i1 1, i1 %V, i8 1)
+
+  ; CHECK: dest_unordered and src_unordered cannot both be zero on the element-wise unordered atomic memory intrinsic
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 %P, i8* align 4 %Q, i32 1, i32 4, i1 0, i1 0, i1 0, i8 1)
+
+  ; CHECK: element size of the element-wise unordered atomic memory intrinsic must be a constant int
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 %P, i8* align 4 %Q, i32 1, i32 4, i1 0, i1 1, i1 1, i8 %E)
   ; CHECK: element size of the element-wise atomic memory intrinsic must be a power of 2
-  call void @llvm.memcpy.element.atomic.p0i8.p0i8(i8* align 2 %P, i8* align 2 %Q, i64 4, i32 3)
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 %P, i8* align 4 %Q, i32 1, i32 4, i1 0, i1 1, i1 1, i8 3)
 
   ; CHECK: incorrect alignment of the destination argument
-  call void @llvm.memcpy.element.atomic.p0i8.p0i8(i8* align 2 %P, i8* align 4 %Q, i64 4, i32 4)
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* %P, i8* align 4 %Q, i32 1, i32 4, i1 0, i1 1, i1 1, i8 1)
+  ; CHECK: incorrect alignment of the destination argument
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 1 %P, i8* align 4 %Q, i32 4, i32 4, i1 0, i1 1, i1 1, i8 4)
 
   ; CHECK: incorrect alignment of the source argument
-  call void @llvm.memcpy.element.atomic.p0i8.p0i8(i8* align 4 %P, i8* align 2 %Q, i64 4, i32 4)
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 %P, i8* %Q, i32 1, i32 4, i1 0, i1 1, i1 1, i8 1)
+  ; CHECK: incorrect alignment of the source argument
+  call void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* align 4 %P, i8* align 1 %Q, i32 4, i32 4, i1 0, i1 1, i1 1, i8 4)
+
 
   ret void
 }
-declare void @llvm.memcpy.element.atomic.p0i8.p0i8(i8* nocapture, i8* nocapture, i64, i32) nounwind
-
+declare void @llvm.memcpy.element.unordered.atomic.p0i8.p0i8.i32(i8* nocapture, i8* nocapture, i32, i32, i1, i1, i1, i8) nounwind
 ; CHECK: input module is broken!