Introduce llvm.loop.parallel_accesses and llvm.access.group metadata.

[Meinersbur]

Summary

The current llvm.mem.parallel_loop_access metadata has a problem in that it uses LoopIDs. LoopID unfortunately is not loop identifier. It is neither unique (there's even a regression test assigning the some LoopID to multiple loops; can otherwise happen if passes such as LoopVersioning make copies of entire loops) nor persistent (every time a property is removed/added from a LoopID's MDNode, it will also receive a new LoopID; this happens e.g. when calling Loop::setLoopAlreadyUnrolled()).
Since most loop transformation passes change the loop attributes (even if it just to mark that a loop should not be processed again as llvm.loop.isvectorized does, for the versioned and unversioned loop), the parallel access information is lost for any subsequent pass.

This patch unlinks LoopIDs and parallel accesses. llvm.mem.parallel_loop_access metadata on instruction is replaced by llvm.access.group metadata. llvm.access.group points to a distinct MDNode with no operands (avoiding the problem to ever need to add/remove operands), called "access group". Alternatively, it can point to a list of access groups. The LoopID then has an attribute llvm.loop.parallel_accesses with all the access groups that are parallel (no dependencies carries by this loop).

This intentionally avoid any kind of "ID". Loops that are clones/have their attributes modifies retain the llvm.loop.parallel_accesses attribute. Access instructions that a cloned point to the same access group. It is not necessary for each access to have it's own "ID" MDNode, but those memory access instructions with the same behavior can be grouped together.

The behavior of llvm.mem.parallel_loop_access is not changed by this patch, but should be considered deprecated.

Possible extensions/follow-up patches:

• AutoUpgrade llvm.mem.parallel_loop_access to llvm.access.group such that we can remove its handling in the passes.

[pekka.jaaskelainen]

Thanks for making this MD more robust! It is essential for the vectorization performance of pocl. Sorry for my slow response time.

Related to OpenCL, there's usually a 3D work-item loop all of which levels are parallel. I didn't spot a test that shows multiple loop levels and thus multiple parallel access groups attached to an instruction. Moreover, the inliner MD transfer would indeed be much improved in case it then replicated the parallel access info from all loop hierarchy levels downwards. The comment which mentions that you focus on inner loop is true, but if there is (or will be) a loop interchange optimization pass that utilizes the parallel loop info, then any loop level might be potentially transferred to be the inner loop.

This is again very useful for OpenCL work-item loops as it allows optimizing memory access patterns via loop interchange, thus essentially performing outer loop vectorization when it's the best way to get performance.

[Meinersbur]

Thank you for your feedback. If you don'r mind, before I commit this, I will prepare a patch for clang generating this new kind of metadata (since atm nothing is generating it) and give others some time for feedback.

Dropping one of the llvm.access.group annotation instead of merging is a trade-off I had to make. It only affects inlining situation where the a loop containing a function call to a function call containing a loop, both loop being marked as parallel. I can make patch suggestion later.

[Meinersbur]

I actually already created a review for the Clang part which is D52117, but uploaded the wrong diff. Corrected. I'll wait for both being accepted before committing.

[Meinersbur]

• Rebase

[Meinersbur]

@hfinkel ping

[hfinkel]

In D52116#1283910, @Meinersbur wrote:

@hfinkel ping

I'm basically happy with this, but we shouldn't add things to the LangRef without an RFC. I don't recall seeing one.

lib/Transforms/Scalar/LoopVersioningLICM.cpp

636

I don't understand what this FIXME is saying. What needs to be fixed?

lib/Transforms/Utils/InlineFunction.cpp

809

Is the problem with "updating all uses of one of the access groups" that we don't have a way to efficiently enumerate them? Would we need to scan the functions for branches and collect all of the loop-id metadata that's relevant first? It would be nice not to lose this information.

[Meinersbur]

I am going to to prepare an RFC.

lib/Transforms/Scalar/LoopVersioningLICM.cpp

636

The line below adds llvm.mem.parallel_loop_access to a LoopID, but is expected as annotations of instructions that access memory (with a the loop it is parallel to as parameter).

There is no code that looks for llvm.mem.parallel_loop_access in LoopID metadata. Hence, adding the property has no effect.

lib/Transforms/Utils/InlineFunction.cpp

809

Either search (and update) all LoopIDs that reference the access group or create a new 'meta-access-group' as outlined in the FIXME.

Of course it would be nice to not lose information, but as for any analyses there is a trade-off between accuracy and computational complexity. E.g. it would be nice if alias-analysis would be control-flow-sensitive. At there moment the only pass making use of Loop::isAnnotatedParallel() are LoopVectorize, LoopVersioningLICM, LoopDistribute and LoopLoadElimination. All of which only process innermost loops. That is, there would be no effect of keeping more information. There's also the FIXME still in the code such that the issue is not forgotten.

Do you still want me to implement one of the solutions?

[hfinkel]

lib/Transforms/Scalar/LoopVersioningLICM.cpp

636

OIC, okay. Thanks.

lib/Transforms/Utils/InlineFunction.cpp

809

Of course it would be nice to not lose information, but as for any analyses there is a trade-off between accuracy and computational complexity.

Clearly I know this ;)

At there moment the only pass making use of Loop::isAnnotatedParallel() are LoopVectorize, LoopVersioningLICM, LoopDistribute and LoopLoadElimination. All of which only process innermost loops. That is, there would be no effect of keeping more information.

Two things: First, this might be true today, but very soon won't be true (vectorization will soon handle outer loops, and we are developing other loop-nest transformations) and if we do the right thing now, we won't later need to go back and fix this later. Second, while it's true that full loop unrolling generally happens before inlining, we could have a loop that, at the point of inlining is an inner loop, but is later fully unrolled such that before vectorization (etc.) the loop here might become, once again, the inner loop.

Do you still want me to implement one of the solutions?

Yes.

[Meinersbur]

• Allow multiple access groups per instructions, i.e. an instruction can be in multiple access groups. This allows a simple 'union' operation that occurs when inlining into another function. A memory access is considered parallel when at least one access group is listed in llvm.loop.parallel_accesses. This is prioritized over the 'intersect' case for combining instructions which would be dual. We only do best-effort here.

[Meinersbur]

• Rebase to trunk

[Meinersbur]

• Reinsert parts of LangRef.rst that went missing after merge.

lib/Transforms/Utils/LoopUtils.cpp

190–219

This has moved to LoopInfo.cpp. LoopUtils.cpp belongs to libTransform, but not all users have a dependency to libTransform (such as Loop::isAnnotatedParallel()).

[hfinkel]

docs/LangRef.rst

5469

of -> if

5472

update -> updating

include/llvm/Analysis/VectorUtils.h

120

access group lists -> access-group lists

126

access group list -> access-group list

lib/Analysis/LoopInfo.cpp

328

Repeating this set of asserts seems unfortunate. Also, they have no comment. Maybe make a function?

assert(AccGroup->isDistinct() && "Access group metadata nodes must be distinct");
assert(AccGroup->getNumOperands() == 0 && "Access group metadata nodes must have zero operands");

you also repeat these asserts in lib/Analysis/VectorUtils.cpp below.

lib/Analysis/VectorUtils.cpp

474

Indentation here is odd.

477

Indentation here looks odd too.

[Meinersbur]

• Address @hfinkel's review
• clang-format

[hfinkel]

Aside from the requested renaming (see below), this LGTM.

include/llvm/Analysis/LoopInfo.h

1016

I'd prefer to name this isValidAsAccessGroup() (because the function does not actually determine whether the MD node *is* an access group, but rather, whether it might be valid to use it as one).

[Meinersbur]

• Rename: isAccessGroup -> isValidAsAccessGroup

docs/LangRef.rst

 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Metadata prefixed with ``llvm.loop.vectorize`` or ``llvm.loop.interleave`` are
 used to control per-loop vectorization and interleaving parameters such as
 vectorization width and interleave count. These metadata should be used in
 conjunction with ``llvm.loop`` loop identification metadata. The
 ``llvm.loop.vectorize`` and ``llvm.loop.interleave`` metadata are only
 optimization hints and the optimizer will only interleave and vectorize loops if
-it believes it is safe to do so. The ``llvm.mem.parallel_loop_access`` metadata
+it believes it is safe to do so. The ``llvm.loop.parallel_accesses`` metadata
 which contains information about loop-carried memory dependencies can be helpful
 in determining the safety of these transformations.
 
 '``llvm.loop.interleave.count``' Metadata
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 This metadata suggests an interleave count to the loop interleaver.
 The first operand is the string ``llvm.loop.interleave.count`` and the

 
 '``llvm.loop.distribute.followup_all``' Metadata
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Thes attributes in this metdata is added to all followup loops of the
 loop distribution pass. See
 :ref:`Transformation Metadata <transformation-metadata>` for details.
 
-'``llvm.mem``'
-^^^^^^^^^^^^^^^
+'``llvm.access.group``' Metadata
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-Metadata types used to annotate memory accesses with information helpful
-for optimizations are prefixed with ``llvm.mem``.
+``llvm.access.group`` metadata can be attached to any instruction that
+potentially accesses memory. It can point to a single distinct metadata
+node, which we call access group. This node represents all memory access
+instructions referring to it via ``llvm.access.group``. When an
+instruction belongs to multiple access groups, it can also point to a
+list of accesses groups, illustrated by the following example.
 
-'``llvm.mem.parallel_loop_access``' Metadata
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+.. code-block:: llvm
 
-The ``llvm.mem.parallel_loop_access`` metadata refers to a loop identifier,
-or metadata containing a list of loop identifiers for nested loops.
-The metadata is attached to memory accessing instructions and denotes that
-no loop carried memory dependence exist between it and other instructions denoted
-with the same loop identifier. The metadata on memory reads also implies that
-if conversion (i.e. speculative execution within a loop iteration) is safe.
-
-Precisely, given two instructions ``m1`` and ``m2`` that both have the
-``llvm.mem.parallel_loop_access`` metadata, with ``L1`` and ``L2`` being the
-set of loops associated with that metadata, respectively, then there is no loop
-carried dependence between ``m1`` and ``m2`` for loops in both ``L1`` and
-``L2``.
-
-As a special case, if all memory accessing instructions in a loop have
-``llvm.mem.parallel_loop_access`` metadata that refers to that loop, then the
-loop has no loop carried memory dependences and is considered to be a parallel
-loop.
+   %val = load i32, i32* %arrayidx, !llvm.access.group !0
+   ...
+   !0 = !{!1, !2}
+   !1 = distinct !{}
+   !2 = distinct !{}
+
+It is illegal for the list node to be empty since it might be confused
+with an access group.
+
+The access group metadata node must be 'distinct' to avoid collapsing
+multiple access groups by content. A access group metadata node must
+always be empty which can be used to distinguish an access group
+metadata node from a list of access groups. Being empty avoids the
+situation that the content must be updated which, because metadata is
+immutable by design, would required finding and updating all references
+to the access group node.
+
+The access group can be used to refer to a memory access instruction
+without pointing to it directly (which is not possible in global
+metadata). Currently, the only metadata making use of it is
+``llvm.loop.parallel_accesses``.
+
+'``llvm.loop.parallel_accesses``' Metadata
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-Note that if not all memory access instructions have such metadata referring to
-the loop, then the loop is considered not being trivially parallel. Additional
+The ``llvm.loop.parallel_accesses`` metadata refers to one or more
+access group metadata nodes (see ``llvm.access.group``). It denotes that
+no loop-carried memory dependence exist between it and other instructions
+in the loop with this metadata.
+
+Let ``m1`` and ``m2`` be two instructions that both have the
+``llvm.access.group`` metadata to the access group ``g1``, respectively
+``g2`` (which might be identical). If a loop contains both access groups
+in its ``llvm.loop.parallel_accesses`` metadata, then the compiler can
+assume that there is no dependency between ``m1`` and ``m2`` carried by
+this loop. Instructions that belong to multiple access groups are
+considered having this property if at least one of the access groups
+matches the ``llvm.loop.parallel_accesses`` list.
+
+If all memory-accessing instructions in a loop have
+``llvm.loop.parallel_accesses`` metadata that refers to that loop, then the
+loop has no loop carried memory dependences and is considered to be a
+parallel loop.
+
+Note that if not all memory access instructions belong to an access
+group referred to by ``llvm.loop.parallel_accesses``, then the loop must
+not be considered trivially parallel. Additional
 memory dependence analysis is required to make that determination. As a fail
 safe mechanism, this causes loops that were originally parallel to be considered
 sequential (if optimization passes that are unaware of the parallel semantics
 insert new memory instructions into the loop body).
 
 Example of a loop that is considered parallel due to its correct use of
-both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
-metadata types that refer to the same loop identifier metadata.
+both ``llvm.access.group`` and ``llvm.loop.parallel_accesses``
+metadata types.
 
 .. code-block:: llvm
 
    for.body:
      ...
-     %val0 = load i32, i32* %arrayidx, !llvm.mem.parallel_loop_access !0
+     %val0 = load i32, i32* %arrayidx, !llvm.access.group !1
      ...
-     store i32 %val0, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
+     store i32 %val0, i32* %arrayidx1, !llvm.access.group !1
      ...
      br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
 
    for.end:
    ...
-   !0 = !{!0}
+   !0 = distinct !{!0, !{!"llvm.loop.parallel_accesses", !1}}
+   !1 = distinct !{}
 
-It is also possible to have nested parallel loops. In that case the
-memory accesses refer to a list of loop identifier metadata nodes instead of
-the loop identifier metadata node directly:
+It is also possible to have nested parallel loops:
 
 .. code-block:: llvm
 
    outer.for.body:
      ...
-     %val1 = load i32, i32* %arrayidx3, !llvm.mem.parallel_loop_access !2
+     %val1 = load i32, i32* %arrayidx3, !llvm.access.group !4
      ...
      br label %inner.for.body
 
    inner.for.body:
      ...
-     %val0 = load i32, i32* %arrayidx1, !llvm.mem.parallel_loop_access !0
+     %val0 = load i32, i32* %arrayidx1, !llvm.access.group !3
      ...
-     store i32 %val0, i32* %arrayidx2, !llvm.mem.parallel_loop_access !0
+     store i32 %val0, i32* %arrayidx2, !llvm.access.group !3
      ...
      br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
 
    inner.for.end:
      ...
-     store i32 %val1, i32* %arrayidx4, !llvm.mem.parallel_loop_access !2
+     store i32 %val1, i32* %arrayidx4, !llvm.access.group !4
      ...
      br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
 
    outer.for.end:                                          ; preds = %for.body
    ...
-   !0 = !{!1, !2} ; a list of loop identifiers
-   !1 = !{!1} ; an identifier for the inner loop
-   !2 = !{!2} ; an identifier for the outer loop
+   !1 = distinct !{!1, !{!"llvm.loop.parallel_accesses", !3}}     ; metadata for the inner loop
+   !2 = distinct !{!2, !{!"llvm.loop.parallel_accesses", !3, !4}} ; metadata for the outer loop
+   !3 = distinct !{} ; access group for instructions in the inner loop (which are implicitly contained in outer loop as well)
+   !4 = distinct !{} ; access group for instructions in the outer, but not the inner loop
 
 '``irr_loop``' Metadata
 ^^^^^^^^^^^^^^^^^^^^^^^
 
 ``irr_loop`` metadata may be attached to the terminator instruction of a basic
 block that's an irreducible loop header (note that an irreducible loop has more
 than once header basic blocks.) If ``irr_loop`` metadata is attached to the
 terminator instruction of a basic block that is not really an irreducible loop

 
 The '``fneg``' instruction returns the negation of its operand.
 
 Arguments:
 """"""""""
 
 The argument to the '``fneg``' instruction must be a
 :ref:`floating-point <t_floating>` or :ref:`vector <t_vector>` of
-floating-point values. 
+floating-point values.
 
 Semantics:
 """"""""""
 
 The value produced is a copy of the operand with its sign bit flipped.
 This instruction can also take any number of :ref:`fast-math
 flags <fastmath>`, which are optimization hints to enable otherwise
 unsafe floating-point optimizations:
::
       declare <type>
       @llvm.experimental.constrained.maxnum(<type> <op1>, <type> <op2>
                                             metadata <rounding mode>,
                                             metadata <exception behavior>)
 
 Overview:
 """""""""
 
-The '``llvm.experimental.constrained.maxnum``' intrinsic returns the maximum 
+The '``llvm.experimental.constrained.maxnum``' intrinsic returns the maximum
 of the two arguments.
 
 Arguments:
 """"""""""
 
-The first two arguments and the return value are floating-point numbers 
+The first two arguments and the return value are floating-point numbers
 of the same type.
 
 The third and forth arguments specify the rounding mode and exception
 behavior as described above.
 
 Semantics:
 """"""""""
 
::
       declare <type>
       @llvm.experimental.constrained.ceil(<type> <op1>,
                                           metadata <rounding mode>,
                                           metadata <exception behavior>)
 
 Overview:
 """""""""
 
-The '``llvm.experimental.constrained.ceil``' intrinsic returns the ceiling of the 
+The '``llvm.experimental.constrained.ceil``' intrinsic returns the ceiling of the
 first operand.
 
 Arguments:
 """"""""""
 
 The first argument and the return value are floating-point numbers of the same
 type.
 
::
       declare <type>
       @llvm.experimental.constrained.floor(<type> <op1>,
                                            metadata <rounding mode>,
                                            metadata <exception behavior>)
 
 Overview:
 """""""""
 
-The '``llvm.experimental.constrained.floor``' intrinsic returns the floor of the 
+The '``llvm.experimental.constrained.floor``' intrinsic returns the floor of the
 first operand.
 
 Arguments:
 """"""""""
 
 The first argument and the return value are floating-point numbers of the same
 type.
 
 The second and third arguments specify the rounding mode and exception
 behavior as described above. The rounding mode is currently unused for this
 intrinsic.
 
 Semantics:
 """"""""""
 
 This function returns the same values as the libm ``floor`` functions
-would and handles error conditions in the same way. 
+would and handles error conditions in the same way.
 
 
 '``llvm.experimental.constrained.round``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax:
 """""""
 
 ::
 
       declare <type>
       @llvm.experimental.constrained.round(<type> <op1>,
                                            metadata <rounding mode>,
                                            metadata <exception behavior>)
 
 Overview:
 """""""""
 
-The '``llvm.experimental.constrained.round``' intrinsic returns the first 
+The '``llvm.experimental.constrained.round``' intrinsic returns the first
 operand rounded to the nearest integer.
 
 Arguments:
 """"""""""
 
 The first argument and the return value are floating-point numbers of the same
 type.
 
::
       declare <type>
       @llvm.experimental.constrained.trunc(<type> <op1>,
                                            metadata <truncing mode>,
                                            metadata <exception behavior>)
 
 Overview:
 """""""""
 
-The '``llvm.experimental.constrained.trunc``' intrinsic returns the first 
-operand rounded to the nearest integer not larger in magnitude than the 
+The '``llvm.experimental.constrained.trunc``' intrinsic returns the first
+operand rounded to the nearest integer not larger in magnitude than the
 operand.
 
 Arguments:
 """"""""""
 
 The first argument and the return value are floating-point numbers of the same
 type.
 

include/llvm/Analysis/LoopInfo.h

#endif
   }
 
   /// Verify loop structure
   void verifyLoop() const;
 
   /// Verify loop structure of this loop and all nested loops.
   void verifyLoopNest(DenseSet<const LoopT *> *Loops) const;
 
+  /// Returns true if the loop is annotated parallel.
+  ///
+  /// Derived classes can override this method using static template
+  /// polymorphism.
+  bool isAnnotatedParallel() const { return false; }
+
   /// Print loop with all the BBs inside it.
   void print(raw_ostream &OS, unsigned Depth = 0, bool Verbose = false) const;
 
 protected:
   friend class LoopInfoBase<BlockT, LoopT>;
 
   /// This creates an empty loop.
   LoopBase() : ParentLoop(nullptr) {}
public:
   void print(raw_ostream &O, const Module *M = nullptr) const override;
 
   void getAnalysisUsage(AnalysisUsage &AU) const override;
 };
 
 /// Function to print a loop's contents as LLVM's text IR assembly.
 void printLoop(Loop &L, raw_ostream &OS, const std::string &Banner = "");
 
+/// Find and return the loop attribute node for the attribute @p Name in
+/// @p LoopID. Return nullptr if there is no such attribute.
+MDNode *findOptionMDForLoopID(MDNode *LoopID, StringRef Name);
+
+/// Find string metadata for a loop.
+///
+/// Returns the MDNode where the first operand is the metadata's name. The
+/// following operands are the metadata's values. If no metadata with @p Name is
+/// found, return nullptr.
+MDNode *findOptionMDForLoop(const Loop *TheLoop, StringRef Name);
+
+/// Return whether an MDNode might represent an access group.
+///
+/// Access group metadata nodes have to be distinct and empty. Being
+/// always-empty ensures that it never needs to be changed (which -- because
+/// MDNodes are designed immutable -- would require creating a new MDNode). Note
+/// that this is not a sufficient condition: not every distinct and empty NDNode
+/// is representing an access group.
+bool isValidAsAccessGroup(MDNode *AccGroup);
+
 } // End llvm namespace
 
 #endif

include/llvm/Analysis/LoopInfoImpl.h

void LoopBase<BlockT, LoopT>::verifyLoopNest(
   // Verify the subloops.
   for (iterator I = begin(), E = end(); I != E; ++I)
     (*I)->verifyLoopNest(Loops);
 }
 
 template <class BlockT, class LoopT>
 void LoopBase<BlockT, LoopT>::print(raw_ostream &OS, unsigned Depth,
                                     bool Verbose) const {
-  OS.indent(Depth * 2) << "Loop at depth " << getLoopDepth() << " containing: ";
+  OS.indent(Depth * 2);
+  if (static_cast<const LoopT *>(this)->isAnnotatedParallel())
+    OS << "Parallel ";
+  OS << "Loop at depth " << getLoopDepth() << " containing: ";
 
   BlockT *H = getHeader();
   for (unsigned i = 0; i < getBlocks().size(); ++i) {
     BlockT *BB = getBlocks()[i];
     if (!Verbose) {
       if (i)
         OS << ",";
       BB->printAsOperand(OS, false);

include/llvm/Analysis/VectorUtils.h

 ///
 /// If the optional TargetTransformInfo is provided, this function tries harder
 /// to do less work by only looking at illegal types.
 MapVector<Instruction*, uint64_t>
 computeMinimumValueSizes(ArrayRef<BasicBlock*> Blocks,
                          DemandedBits &DB,
                          const TargetTransformInfo *TTI=nullptr);
 
+/// Compute the union of two access-group lists.
+///
+/// If the list contains just one access group, it is returned directly. If the
+/// list is empty, returns nullptr.
+MDNode *uniteAccessGroups(MDNode *AccGroups1, MDNode *AccGroups2);
+
+/// Compute the access-group list of access groups that @p Inst1 and @p Inst2
+/// are both in. If either instruction does not access memory at all, it is
+/// considered to be in every list.
+///
+/// If the list contains just one access group, it is returned directly. If the
+/// list is empty, returns nullptr.
+MDNode *intersectAccessGroups(const Instruction *Inst1,
+                              const Instruction *Inst2);
+
 /// Specifically, let Kinds = [MD_tbaa, MD_alias_scope, MD_noalias, MD_fpmath,
-/// MD_nontemporal].  For K in Kinds, we get the MDNode for K from each of the
+/// MD_nontemporal, MD_access_group].
+/// For K in Kinds, we get the MDNode for K from each of the
 /// elements of VL, compute their "intersection" (i.e., the most generic
 /// metadata value that covers all of the individual values), and set I's
 /// metadata for M equal to the intersection value.
 ///
 /// This function always sets a (possibly null) value for each K in Kinds.
 Instruction *propagateMetadata(Instruction *I, ArrayRef<Value *> VL);
 
 /// Create a mask that filters the members of an interleave group where there
 /// are gaps.
 ///
 /// For example, the mask for \p Group with interleave-factor 3
 /// and \p VF 4, that has only its first member present is:
 ///
 ///   <1,0,0,1,0,0,1,0,0,1,0,0>
 ///
 /// Note: The result is a mask of 0's and 1's, as opposed to the other
 /// create[*]Mask() utilities which create a shuffle mask (mask that
 /// consists of indices).
 Constant *createBitMaskForGaps(IRBuilder<> &Builder, unsigned VF,
                                const InterleaveGroup<Instruction> &Group);
 
 /// Create a mask with replicated elements.
 ///
-/// This function creates a shuffle mask for replicating each of the \p VF 
+/// This function creates a shuffle mask for replicating each of the \p VF
 /// elements in a vector \p ReplicationFactor times. It can be used to
 /// transform a mask of \p VF elements into a mask of
 /// \p VF * \p ReplicationFactor elements used by a predicated
 /// interleaved-group of loads/stores whose Interleaved-factor ==
 /// \p ReplicationFactor.
 ///
 /// For example, the mask for \p ReplicationFactor=3 and \p VF=4 is:
 ///

include/llvm/IR/LLVMContext.h

enum : unsigned {
     MD_align = 17,                    // "align"
     MD_loop = 18,                     // "llvm.loop"
     MD_type = 19,                     // "type"
     MD_section_prefix = 20,           // "section_prefix"
     MD_absolute_symbol = 21,          // "absolute_symbol"
     MD_associated = 22,               // "associated"
     MD_callees = 23,                  // "callees"
     MD_irr_loop = 24,                 // "irr_loop"
+    MD_access_group = 25,             // "llvm.access.group"
   };
 
   /// Known operand bundle tag IDs, which always have the same value.  All
   /// operand bundle tags that LLVM has special knowledge of are listed here.
   /// Additionally, this scheme allows LLVM to efficiently check for specific
   /// operand bundle tags without comparing strings.
   enum : unsigned {
     OB_deopt = 0,         // "deopt"

include/llvm/Transforms/Utils/LoopUtils.h

 /// Returns the instructions that use values defined in the loop.
 SmallVector<Instruction *, 8> findDefsUsedOutsideOfLoop(Loop *L);
 
 /// Find string metadata for loop
 ///
 /// If it has a value (e.g. {"llvm.distribute", 1} return the value as an
 /// operand or null otherwise.  If the string metadata is not found return
 /// Optional's not-a-value.
-Optional<const MDOperand *> findStringMetadataForLoop(Loop *TheLoop,
+Optional<const MDOperand *> findStringMetadataForLoop(const Loop *TheLoop,
                                                       StringRef Name);
 
 /// Find named metadata for a loop with an integer value.
 llvm::Optional<int> getOptionalIntLoopAttribute(Loop *TheLoop, StringRef Name);
 
 /// Create a new loop identifier for a loop created from a loop transformation.
 ///
 /// @param OrigLoopID The loop ID of the loop before the transformation.

lib/Analysis/LoopInfo.cpp

 }
 
 bool Loop::isAnnotatedParallel() const {
   MDNode *DesiredLoopIdMetadata = getLoopID();
 
   if (!DesiredLoopIdMetadata)
     return false;
 
+  MDNode *ParallelAccesses =
+      findOptionMDForLoop(this, "llvm.loop.parallel_accesses");
+  SmallPtrSet<MDNode *, 4>
+      ParallelAccessGroups; // For scalable 'contains' check.
+  if (ParallelAccesses) {
+    for (const MDOperand &MD : drop_begin(ParallelAccesses->operands(), 1)) {
+      MDNode *AccGroup = cast<MDNode>(MD.get());
+      assert(isValidAsAccessGroup(AccGroup) &&
+             "List item must be an access group");
+      ParallelAccessGroups.insert(AccGroup);
+    }
+  }
+
   // The loop branch contains the parallel loop metadata. In order to ensure
   // that any parallel-loop-unaware optimization pass hasn't added loop-carried
   // dependencies (thus converted the loop back to a sequential loop), check
-  // that all the memory instructions in the loop contain parallelism metadata
-  // that point to the same unique "loop id metadata" the loop branch does.
+  // that all the memory instructions in the loop belong to an access group that
+  // is parallel to this loop.
   for (BasicBlock *BB : this->blocks()) {
     for (Instruction &I : *BB) {
       if (!I.mayReadOrWriteMemory())
         continue;
 
+      if (MDNode *AccessGroup = I.getMetadata(LLVMContext::MD_access_group)) {
+        auto ContainsAccessGroup = [&ParallelAccessGroups](MDNode *AG) -> bool {
+          if (AG->getNumOperands() == 0) {
+            assert(isValidAsAccessGroup(AG) && "Item must be an access group");
+            return ParallelAccessGroups.count(AG);
+          }
+
+          for (const MDOperand &AccessListItem : AG->operands()) {
+            MDNode *AccGroup = cast<MDNode>(AccessListItem.get());
+            assert(isValidAsAccessGroup(AccGroup) &&
+                   "List item must be an access group");
+            if (ParallelAccessGroups.count(AccGroup))
+              return true;
+          }
+          return false;
+        };
+
+        if (ContainsAccessGroup(AccessGroup))
+          continue;
+      }
+
       // The memory instruction can refer to the loop identifier metadata
       // directly or indirectly through another list metadata (in case of
       // nested parallel loops). The loop identifier metadata refers to
       // itself so we can check both cases with the same routine.
       MDNode *LoopIdMD =
           I.getMetadata(LLVMContext::MD_mem_parallel_loop_access);
 
       if (!LoopIdMD)
if (!ExitBlocks.empty()) {
     for (auto *Block : ExitBlocks)
       if (Block)
         Block->print(OS);
       else
         OS << "Printing <null> block";
   }
 }
 
+MDNode *llvm::findOptionMDForLoopID(MDNode *LoopID, StringRef Name) {
+  // No loop metadata node, no loop properties.
+  if (!LoopID)
+    return nullptr;
+
+  // First operand should refer to the metadata node itself, for legacy reasons.
+  assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
+  assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
+
+  // Iterate over the metdata node operands and look for MDString metadata.
+  for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
+    MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
+    if (!MD || MD->getNumOperands() < 1)
+      continue;
+    MDString *S = dyn_cast<MDString>(MD->getOperand(0));
+    if (!S)
+      continue;
+    // Return the operand node if MDString holds expected metadata.
+    if (Name.equals(S->getString()))
+      return MD;
+  }
+
+  // Loop property not found.
+  return nullptr;
+}
+
+MDNode *llvm::findOptionMDForLoop(const Loop *TheLoop, StringRef Name) {
+  return findOptionMDForLoopID(TheLoop->getLoopID(), Name);
+}
+
+bool llvm::isValidAsAccessGroup(MDNode *Node) {
+  return Node->getNumOperands() == 0 && Node->isDistinct();
+}
+
 //===----------------------------------------------------------------------===//
 // LoopInfo implementation
 //
 
 char LoopInfoWrapperPass::ID = 0;
 INITIALIZE_PASS_BEGIN(LoopInfoWrapperPass, "loops", "Natural Loop Information",
                       true, true)
 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)

lib/Analysis/VectorUtils.cpp

for (auto MI = ECs.member_begin(I), ME = ECs.member_end(); MI != ME; ++MI) {
       if (MinBW < Ty->getScalarSizeInBits())
         MinBWs[cast<Instruction>(*MI)] = MinBW;
     }
   }
 
   return MinBWs;
 }
 
+/// Add all access groups in @p AccGroups to @p List.
+template <typename ListT>
+static void addToAccessGroupList(ListT &List, MDNode *AccGroups) {
+  // Interpret an access group as a list containing itself.
+  if (AccGroups->getNumOperands() == 0) {
+    assert(isValidAsAccessGroup(AccGroups) && "Node must be an access group");
+    List.insert(AccGroups);
+    return;
+  }
+
+  for (auto &AccGroupListOp : AccGroups->operands()) {
+    auto *Item = cast<MDNode>(AccGroupListOp.get());
+    assert(isValidAsAccessGroup(Item) && "List item must be an access group");
+    List.insert(Item);
+  }
+};
+
+MDNode *llvm::uniteAccessGroups(MDNode *AccGroups1, MDNode *AccGroups2) {
+  if (!AccGroups1)
+    return AccGroups2;
+  if (!AccGroups2)
+    return AccGroups1;
+  if (AccGroups1 == AccGroups2)
+    return AccGroups1;
+
+  SmallSetVector<Metadata *, 4> Union;
+  addToAccessGroupList(Union, AccGroups1);
+  addToAccessGroupList(Union, AccGroups2);
+
+  if (Union.size() == 0)
+    return nullptr;
+  if (Union.size() == 1)
+    return cast<MDNode>(Union.front());
+
+  LLVMContext &Ctx = AccGroups1->getContext();
+  return MDNode::get(Ctx, Union.getArrayRef());
+}
+
+MDNode *llvm::intersectAccessGroups(const Instruction *Inst1,
+                                    const Instruction *Inst2) {
+  bool MayAccessMem1 = Inst1->mayReadOrWriteMemory();
+  bool MayAccessMem2 = Inst2->mayReadOrWriteMemory();
+
+  if (!MayAccessMem1 && !MayAccessMem2)
+    return nullptr;
+  if (!MayAccessMem1)
+    return Inst2->getMetadata(LLVMContext::MD_access_group);
+  if (!MayAccessMem2)
+    return Inst1->getMetadata(LLVMContext::MD_access_group);
+
+  MDNode *MD1 = Inst1->getMetadata(LLVMContext::MD_access_group);
+  MDNode *MD2 = Inst2->getMetadata(LLVMContext::MD_access_group);
+  if (!MD1 || !MD2)
+    return nullptr;
+  if (MD1 == MD2)
+    return MD1;
+
+  // Use set for scalable 'contains' check.
+  SmallPtrSet<Metadata *, 4> AccGroupSet2;
+  addToAccessGroupList(AccGroupSet2, MD2);
+
+  SmallVector<Metadata *, 4> Intersection;
+  if (MD1->getNumOperands() == 0) {
+    assert(isValidAsAccessGroup(MD1) && "Node must be an access group");
+    if (AccGroupSet2.count(MD1))
+      Intersection.push_back(MD1);
+  } else {
+    for (const MDOperand &Node : MD1->operands()) {
+      auto *Item = cast<MDNode>(Node.get());
+      assert(isValidAsAccessGroup(Item) && "List item must be an access group");
+      if (AccGroupSet2.count(Item))
+        Intersection.push_back(Item);
+    }
+  }
+
+  if (Intersection.size() == 0)
+    return nullptr;
+  if (Intersection.size() == 1)
+    return cast<MDNode>(Intersection.front());
+
+  LLVMContext &Ctx = Inst1->getContext();
+  return MDNode::get(Ctx, Intersection);
+}
+
 /// \returns \p I after propagating metadata from \p VL.
 Instruction *llvm::propagateMetadata(Instruction *Inst, ArrayRef<Value *> VL) {
   Instruction *I0 = cast<Instruction>(VL[0]);
   SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
   I0->getAllMetadataOtherThanDebugLoc(Metadata);
 
-  for (auto Kind :
-       {LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
+  for (auto Kind : {LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
                     LLVMContext::MD_noalias, LLVMContext::MD_fpmath,
-        LLVMContext::MD_nontemporal, LLVMContext::MD_invariant_load}) {
+                    LLVMContext::MD_nontemporal, LLVMContext::MD_invariant_load,
+                    LLVMContext::MD_access_group}) {
     MDNode *MD = I0->getMetadata(Kind);
 
     for (int J = 1, E = VL.size(); MD && J != E; ++J) {
       const Instruction *IJ = cast<Instruction>(VL[J]);
       MDNode *IMD = IJ->getMetadata(Kind);
       switch (Kind) {
       case LLVMContext::MD_tbaa:
         MD = MDNode::getMostGenericTBAA(MD, IMD);
         break;
       case LLVMContext::MD_alias_scope:
         MD = MDNode::getMostGenericAliasScope(MD, IMD);
         break;
       case LLVMContext::MD_fpmath:
         MD = MDNode::getMostGenericFPMath(MD, IMD);
         break;
       case LLVMContext::MD_noalias:
       case LLVMContext::MD_nontemporal:
       case LLVMContext::MD_invariant_load:
         MD = MDNode::intersect(MD, IMD);
         break;
+      case LLVMContext::MD_access_group:
+        MD = intersectAccessGroups(Inst, IJ);
+        break;
       default:
         llvm_unreachable("unhandled metadata");
       }
     }
 
     Inst->setMetadata(Kind, MD);
   }
 

lib/IR/LLVMContext.cpp

 #include <utility>
 
 using namespace llvm;
 
 LLVMContext::LLVMContext() : pImpl(new LLVMContextImpl(*this)) {
   // Create the fixed metadata kinds. This is done in the same order as the
   // MD_* enum values so that they correspond.
   std::pair<unsigned, StringRef> MDKinds[] = {
       {MD_dbg, "dbg"},
       {MD_tbaa, "tbaa"},
       {MD_prof, "prof"},
       {MD_fpmath, "fpmath"},
       {MD_range, "range"},
       {MD_tbaa_struct, "tbaa.struct"},
       {MD_invariant_load, "invariant.load"},
       {MD_alias_scope, "alias.scope"},
       {MD_noalias, "noalias"},
       {MD_nontemporal, "nontemporal"},
       {MD_mem_parallel_loop_access, "llvm.mem.parallel_loop_access"},
       {MD_nonnull, "nonnull"},
       {MD_dereferenceable, "dereferenceable"},
       {MD_dereferenceable_or_null, "dereferenceable_or_null"},
       {MD_make_implicit, "make.implicit"},
       {MD_unpredictable, "unpredictable"},
       {MD_invariant_group, "invariant.group"},
       {MD_align, "align"},
       {MD_loop, "llvm.loop"},
       {MD_type, "type"},
       {MD_section_prefix, "section_prefix"},
       {MD_absolute_symbol, "absolute_symbol"},
       {MD_associated, "associated"},
       {MD_callees, "callees"},
       {MD_irr_loop, "irr_loop"},
+      {MD_access_group, "llvm.access.group"},
   };
 
   for (auto &MDKind : MDKinds) {
     unsigned ID = getMDKindID(MDKind.second);
     assert(ID == MDKind.first && "metadata kind id drifted");
     (void)ID;
   }
 

lib/Transforms/InstCombine/InstCombineCalls.cpp

Instruction *InstCombiner::SimplifyAnyMemTransfer(AnyMemTransferInst *MI) {
   // Alignment from the mem intrinsic will be better, so use it.
   L->setAlignment(CopySrcAlign);
   if (CopyMD)
     L->setMetadata(LLVMContext::MD_tbaa, CopyMD);
   MDNode *LoopMemParallelMD =
     MI->getMetadata(LLVMContext::MD_mem_parallel_loop_access);
   if (LoopMemParallelMD)
     L->setMetadata(LLVMContext::MD_mem_parallel_loop_access, LoopMemParallelMD);
+  MDNode *AccessGroupMD = MI->getMetadata(LLVMContext::MD_access_group);
+  if (AccessGroupMD)
+    L->setMetadata(LLVMContext::MD_access_group, AccessGroupMD);
 
   StoreInst *S = Builder.CreateStore(L, Dest);
   // Alignment from the mem intrinsic will be better, so use it.
   S->setAlignment(CopyDstAlign);
   if (CopyMD)
     S->setMetadata(LLVMContext::MD_tbaa, CopyMD);
   if (LoopMemParallelMD)
     S->setMetadata(LLVMContext::MD_mem_parallel_loop_access, LoopMemParallelMD);
+  if (AccessGroupMD)
+    S->setMetadata(LLVMContext::MD_access_group, AccessGroupMD);
 
   if (auto *MT = dyn_cast<MemTransferInst>(MI)) {
     // non-atomics can be volatile
     L->setVolatile(MT->isVolatile());
     S->setVolatile(MT->isVolatile());
   }
   if (isa<AtomicMemTransferInst>(MI)) {
     // atomics have to be unordered

lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp

for (const auto &MDPair : MD) {
     case LLVMContext::MD_prof:
     case LLVMContext::MD_fpmath:
     case LLVMContext::MD_tbaa_struct:
     case LLVMContext::MD_invariant_load:
     case LLVMContext::MD_alias_scope:
     case LLVMContext::MD_noalias:
     case LLVMContext::MD_nontemporal:
     case LLVMContext::MD_mem_parallel_loop_access:
+    case LLVMContext::MD_access_group:
       // All of these directly apply.
       NewLoad->setMetadata(ID, N);
       break;
 
     case LLVMContext::MD_nonnull:
       copyNonnullMetadata(LI, N, *NewLoad);
       break;
     case LLVMContext::MD_align:

lib/Transforms/InstCombine/InstCombinePHI.cpp

unsigned KnownIDs[] = {
     LLVMContext::MD_range,
     LLVMContext::MD_invariant_load,
     LLVMContext::MD_alias_scope,
     LLVMContext::MD_noalias,
     LLVMContext::MD_nonnull,
     LLVMContext::MD_align,
     LLVMContext::MD_dereferenceable,
     LLVMContext::MD_dereferenceable_or_null,
+    LLVMContext::MD_access_group,
   };
 
   for (unsigned ID : KnownIDs)
     NewLI->setMetadata(ID, FirstLI->getMetadata(ID));
 
   // Add all operands to the new PHI and combine TBAA metadata.
   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
     LoadInst *LI = cast<LoadInst>(PN.getIncomingValue(i));

lib/Transforms/Scalar/GVNHoist.cpp

public:
   const VNtoInsns &getStoreVNTable() const { return VNtoCallsStores; }
 };
 
 static void combineKnownMetadata(Instruction *ReplInst, Instruction *I) {
   static const unsigned KnownIDs[] = {
       LLVMContext::MD_tbaa,           LLVMContext::MD_alias_scope,
       LLVMContext::MD_noalias,        LLVMContext::MD_range,
       LLVMContext::MD_fpmath,         LLVMContext::MD_invariant_load,
-      LLVMContext::MD_invariant_group};
+      LLVMContext::MD_invariant_group, LLVMContext::MD_access_group};
   combineMetadata(ReplInst, I, KnownIDs, true);
 }
 
 // This pass hoists common computations across branches sharing common
 // dominator. The primary goal is to reduce the code size, and in some
 // cases reduce critical path (by exposing more ILP).
 class GVNHoist {
 public:

lib/Transforms/Scalar/LoopVersioningLICM.cpp

if (isLegalForVersioning()) {
     DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
     LoopVersioning LVer(*LAI, CurLoop, LI, DT, SE, true);
     LVer.versionLoop();
     // Set Loop Versioning metaData for original loop.
     addStringMetadataToLoop(LVer.getNonVersionedLoop(), LICMVersioningMetaData);
     // Set Loop Versioning metaData for version loop.
     addStringMetadataToLoop(LVer.getVersionedLoop(), LICMVersioningMetaData);
     // Set "llvm.mem.parallel_loop_access" metaData to versioned loop.
+    // FIXME: "llvm.mem.parallel_loop_access" annotates memory access
+    // instructions, not loops.
     addStringMetadataToLoop(LVer.getVersionedLoop(),
                             "llvm.mem.parallel_loop_access");
     // Update version loop with aggressive aliasing assumption.
     setNoAliasToLoop(LVer.getVersionedLoop());
     Changed = true;
   }
   return Changed;
 }

lib/Transforms/Scalar/MemCpyOptimizer.cpp

bool MemCpyOptPass::performCallSlotOptzn(Instruction *cpy, Value *cpyDest,
   // its dependence information by changing its parameter.
   MD->removeInstruction(C);
 
   // Update AA metadata
   // FIXME: MD_tbaa_struct and MD_mem_parallel_loop_access should also be
   // handled here, but combineMetadata doesn't support them yet
   unsigned KnownIDs[] = {LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
                          LLVMContext::MD_noalias,
-                         LLVMContext::MD_invariant_group};
+                         LLVMContext::MD_invariant_group,
+                         LLVMContext::MD_access_group};
   combineMetadata(C, cpy, KnownIDs, true);
 
   // Remove the memcpy.
   MD->removeInstruction(cpy);
   ++NumMemCpyInstr;
 
   return true;
 }

lib/Transforms/Scalar/SROA.cpp

if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
           IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
       Old = convertValue(DL, IRB, Old, IntTy);
       assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
       uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
       V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
     }
     V = convertValue(DL, IRB, V, NewAllocaTy);
     StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
-    Store->copyMetadata(SI, LLVMContext::MD_mem_parallel_loop_access);
+    Store->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
+                             LLVMContext::MD_access_group});
     if (AATags)
       Store->setAAMetadata(AATags);
     Pass.DeadInsts.insert(&SI);
     LLVM_DEBUG(dbgs() << "          to: " << *Store << "\n");
     return true;
   }
 
   bool visitStoreInst(StoreInst &SI) {
if (NewBeginOffset == NewAllocaBeginOffset &&
       NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
                                      SI.isVolatile());
     } else {
       unsigned AS = SI.getPointerAddressSpace();
       Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo(AS));
       NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
                                      SI.isVolatile());
     }
-    NewSI->copyMetadata(SI, LLVMContext::MD_mem_parallel_loop_access);
+    NewSI->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
+                             LLVMContext::MD_access_group});
     if (AATags)
       NewSI->setAAMetadata(AATags);
     if (SI.isVolatile())
       NewSI->setAtomic(SI.getOrdering(), SI.getSyncScopeID());
     Pass.DeadInsts.insert(&SI);
     deleteIfTriviallyDead(OldOp);
 
     LLVM_DEBUG(dbgs() << "          to: " << *NewSI << "\n");
for (;;) {
       auto AS = LI->getPointerAddressSpace();
       auto *PartPtrTy = PartTy->getPointerTo(AS);
       LoadInst *PLoad = IRB.CreateAlignedLoad(
           getAdjustedPtr(IRB, DL, BasePtr,
                          APInt(DL.getIndexSizeInBits(AS), PartOffset),
                          PartPtrTy, BasePtr->getName() + "."),
           getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
           LI->getName());
-      PLoad->copyMetadata(*LI, LLVMContext::MD_mem_parallel_loop_access);
+      PLoad->copyMetadata(*LI, {LLVMContext::MD_mem_parallel_loop_access,
+                                LLVMContext::MD_access_group});
 
       // Append this load onto the list of split loads so we can find it later
       // to rewrite the stores.
       SplitLoads.push_back(PLoad);
 
       // Now build a new slice for the alloca.
       NewSlices.push_back(
           Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
for (User *LU : LI->users()) {
 
         auto AS = SI->getPointerAddressSpace();
         StoreInst *PStore = IRB.CreateAlignedStore(
             PLoad,
             getAdjustedPtr(IRB, DL, StoreBasePtr,
                            APInt(DL.getIndexSizeInBits(AS), PartOffset),
                            PartPtrTy, StoreBasePtr->getName() + "."),
             getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
-        PStore->copyMetadata(*LI, LLVMContext::MD_mem_parallel_loop_access);
+        PStore->copyMetadata(*LI, {LLVMContext::MD_mem_parallel_loop_access,
+                                   LLVMContext::MD_access_group});
         LLVM_DEBUG(dbgs() << "      +" << PartOffset << ":" << *PStore << "\n");
       }
 
       // We want to immediately iterate on any allocas impacted by splitting
       // this store, and we have to track any promotable alloca (indicated by
       // a direct store) as needing to be resplit because it is no longer
       // promotable.
       if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {

lib/Transforms/Scalar/Scalarizer.cpp

 // vector to scalar instructions.
 bool ScalarizerVisitor::canTransferMetadata(unsigned Tag) {
   return (Tag == LLVMContext::MD_tbaa
           || Tag == LLVMContext::MD_fpmath
           || Tag == LLVMContext::MD_tbaa_struct
           || Tag == LLVMContext::MD_invariant_load
           || Tag == LLVMContext::MD_alias_scope
           || Tag == LLVMContext::MD_noalias
-          || Tag == ParallelLoopAccessMDKind);
+          || Tag == ParallelLoopAccessMDKind
+          || Tag == LLVMContext::MD_access_group);
 }
 
 // Transfer metadata from Op to the instructions in CV if it is known
 // to be safe to do so.
 void ScalarizerVisitor::transferMetadata(Instruction *Op, const ValueVector &CV) {
   SmallVector<std::pair<unsigned, MDNode *>, 4> MDs;
   Op->getAllMetadataOtherThanDebugLoc(MDs);
   for (unsigned I = 0, E = CV.size(); I != E; ++I) {

lib/Transforms/Utils/InlineFunction.cpp

 #include "llvm/Analysis/BlockFrequencyInfo.h"
 #include "llvm/Analysis/CallGraph.h"
 #include "llvm/Analysis/CaptureTracking.h"
 #include "llvm/Analysis/EHPersonalities.h"
 #include "llvm/Analysis/InstructionSimplify.h"
 #include "llvm/Analysis/ProfileSummaryInfo.h"
 #include "llvm/Transforms/Utils/Local.h"
 #include "llvm/Analysis/ValueTracking.h"
+#include "llvm/Analysis/VectorUtils.h"
 #include "llvm/IR/Argument.h"
 #include "llvm/IR/BasicBlock.h"
 #include "llvm/IR/CFG.h"
 #include "llvm/IR/CallSite.h"
 #include "llvm/IR/Constant.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/DIBuilder.h"
 #include "llvm/IR/DataLayout.h"
static void HandleInlinedEHPad(InvokeInst *II, BasicBlock *FirstNewBlock,
 
   // Now that everything is happy, we have one final detail.  The PHI nodes in
   // the exception destination block still have entries due to the original
   // invoke instruction. Eliminate these entries (which might even delete the
   // PHI node) now.
   UnwindDest->removePredecessor(InvokeBB);
 }
 
-/// When inlining a call site that has !llvm.mem.parallel_loop_access metadata,
-/// that metadata should be propagated to all memory-accessing cloned
-/// instructions.
+/// When inlining a call site that has !llvm.mem.parallel_loop_access or
+/// llvm.access.group metadata, that metadata should be propagated to all
+/// memory-accessing cloned instructions.
 static void PropagateParallelLoopAccessMetadata(CallSite CS,
                                                 ValueToValueMapTy &VMap) {
   MDNode *M =
     CS.getInstruction()->getMetadata(LLVMContext::MD_mem_parallel_loop_access);
-  if (!M)
+  MDNode *CallAccessGroup =
+      CS.getInstruction()->getMetadata(LLVMContext::MD_access_group);
+  if (!M && !CallAccessGroup)
     return;
 
   for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end();
        VMI != VMIE; ++VMI) {
     if (!VMI->second)
       continue;
 
     Instruction *NI = dyn_cast<Instruction>(VMI->second);
     if (!NI)
       continue;
 
-    if (MDNode *PM = NI->getMetadata(LLVMContext::MD_mem_parallel_loop_access)) {
+    if (M) {
+      if (MDNode *PM =
+              NI->getMetadata(LLVMContext::MD_mem_parallel_loop_access)) {
         M = MDNode::concatenate(PM, M);
       NI->setMetadata(LLVMContext::MD_mem_parallel_loop_access, M);
       } else if (NI->mayReadOrWriteMemory()) {
         NI->setMetadata(LLVMContext::MD_mem_parallel_loop_access, M);
       }
     }
+
+    if (NI->mayReadOrWriteMemory()) {
+      MDNode *UnitedAccGroups = uniteAccessGroups(
+          NI->getMetadata(LLVMContext::MD_access_group), CallAccessGroup);
+      NI->setMetadata(LLVMContext::MD_access_group, UnitedAccGroups);
+    }
+  }
 }
 
 /// When inlining a function that contains noalias scope metadata,
 /// this metadata needs to be cloned so that the inlined blocks
 /// have different "unique scopes" at every call site. Were this not done, then
 /// aliasing scopes from a function inlined into a caller multiple times could
 /// not be differentiated (and this would lead to miscompiles because the
 /// non-aliasing property communicated by the metadata could have

lib/Transforms/Utils/Local.cpp

 #include "llvm/Analysis/ConstantFolding.h"
 #include "llvm/Analysis/EHPersonalities.h"
 #include "llvm/Analysis/InstructionSimplify.h"
 #include "llvm/Analysis/LazyValueInfo.h"
 #include "llvm/Analysis/MemoryBuiltins.h"
 #include "llvm/Analysis/MemorySSAUpdater.h"
 #include "llvm/Analysis/TargetLibraryInfo.h"
 #include "llvm/Analysis/ValueTracking.h"
+#include "llvm/Analysis/VectorUtils.h"
 #include "llvm/BinaryFormat/Dwarf.h"
 #include "llvm/IR/Argument.h"
 #include "llvm/IR/Attributes.h"
 #include "llvm/IR/BasicBlock.h"
 #include "llvm/IR/CFG.h"
 #include "llvm/IR/CallSite.h"
 #include "llvm/IR/Constant.h"
 #include "llvm/IR/ConstantRange.h"
switch (Kind) {
         break;
       case LLVMContext::MD_alias_scope:
         K->setMetadata(Kind, MDNode::getMostGenericAliasScope(JMD, KMD));
         break;
       case LLVMContext::MD_noalias:
       case LLVMContext::MD_mem_parallel_loop_access:
         K->setMetadata(Kind, MDNode::intersect(JMD, KMD));
         break;
+      case LLVMContext::MD_access_group:
+        K->setMetadata(LLVMContext::MD_access_group,
+                       intersectAccessGroups(K, J));
+        break;
       case LLVMContext::MD_range:
 
         // If K does move, use most generic range. Otherwise keep the range of
         // K.
         if (DoesKMove)
           // FIXME: If K does move, we should drop the range info and nonnull.
           //        Currently this function is used with DoesKMove in passes
           //        doing hoisting/sinking and the current behavior of using the

 void llvm::combineMetadataForCSE(Instruction *K, const Instruction *J,
                                  bool KDominatesJ) {
   unsigned KnownIDs[] = {
       LLVMContext::MD_tbaa,            LLVMContext::MD_alias_scope,
       LLVMContext::MD_noalias,         LLVMContext::MD_range,
       LLVMContext::MD_invariant_load,  LLVMContext::MD_nonnull,
       LLVMContext::MD_invariant_group, LLVMContext::MD_align,
       LLVMContext::MD_dereferenceable,
-      LLVMContext::MD_dereferenceable_or_null};
+      LLVMContext::MD_dereferenceable_or_null,
+      LLVMContext::MD_access_group};
   combineMetadata(K, J, KnownIDs, KDominatesJ);
 }
 
 void llvm::patchReplacementInstruction(Instruction *I, Value *Repl) {
   auto *ReplInst = dyn_cast<Instruction>(Repl);
   if (!ReplInst)
     return;
 
void llvm::patchReplacementInstruction(Instruction *I, Value *Repl) {
 
   // In general, GVN unifies expressions over different control-flow
   // regions, and so we need a conservative combination of the noalias
   // scopes.
   static const unsigned KnownIDs[] = {
       LLVMContext::MD_tbaa,            LLVMContext::MD_alias_scope,
       LLVMContext::MD_noalias,         LLVMContext::MD_range,
       LLVMContext::MD_fpmath,          LLVMContext::MD_invariant_load,
-      LLVMContext::MD_invariant_group, LLVMContext::MD_nonnull};
+      LLVMContext::MD_invariant_group, LLVMContext::MD_nonnull,
+      LLVMContext::MD_access_group};
   combineMetadata(ReplInst, I, KnownIDs, false);
 }
 
 template <typename RootType, typename DominatesFn>
 static unsigned replaceDominatedUsesWith(Value *From, Value *To,
                                          const RootType &Root,
                                          const DominatesFn &Dominates) {
   assert(From->getType() == To->getType());

lib/Transforms/Utils/LoopUtils.cpp

void llvm::initializeLoopPassPass(PassRegistry &Registry) {
   INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)
   INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
   INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)
   INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
   INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass)
   INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
 }
 
-static Optional<MDNode *> findOptionMDForLoopID(MDNode *LoopID,
-                                                StringRef Name) {
-  // Return none if LoopID is false.
-  if (!LoopID)
-    return None;
-
-  // First operand should refer to the loop id itself.
-  assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
-  assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
-
-  // Iterate over LoopID operands and look for MDString Metadata
-  for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
-    MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
-    if (!MD)
-      continue;
-    MDString *S = dyn_cast<MDString>(MD->getOperand(0));
-    if (!S)
-      continue;
-    // Return true if MDString holds expected MetaData.
-    if (Name.equals(S->getString()))
-      return MD;
-  }
-  return None;
-}
-
-static Optional<MDNode *> findOptionMDForLoop(const Loop *TheLoop,
-                                              StringRef Name) {
-  return findOptionMDForLoopID(TheLoop->getLoopID(), Name);
-}
-
 /// Find string metadata for loop
 ///
 /// If it has a value (e.g. {"llvm.distribute", 1} return the value as an
 /// operand or null otherwise.  If the string metadata is not found return
 /// Optional's not-a-value.
-Optional<const MDOperand *> llvm::findStringMetadataForLoop(Loop *TheLoop,
+Optional<const MDOperand *> llvm::findStringMetadataForLoop(const Loop *TheLoop,
                                                             StringRef Name) {
-  auto MD = findOptionMDForLoop(TheLoop, Name).getValueOr(nullptr);
+  MDNode *MD = findOptionMDForLoop(TheLoop, Name);
   if (!MD)
     return None;
   switch (MD->getNumOperands()) {
   case 1:
     return nullptr;
   case 2:
     return &MD->getOperand(1);
   default:
     llvm_unreachable("loop metadata has 0 or 1 operand");
   }
 }
 
 static Optional<bool> getOptionalBoolLoopAttribute(const Loop *TheLoop,
                                                    StringRef Name) {
-  Optional<MDNode *> MD = findOptionMDForLoop(TheLoop, Name);
-  if (!MD.hasValue())
-    return None;
-  MDNode *OptionNode = MD.getValue();
-  if (OptionNode == nullptr)
+  MDNode *MD = findOptionMDForLoop(TheLoop, Name);
+  if (!MD)
     return None;
-  switch (OptionNode->getNumOperands()) {
+  switch (MD->getNumOperands()) {
   case 1:
     // When the value is absent it is interpreted as 'attribute set'.
     return true;
   case 2:
-    return mdconst::extract_or_null<ConstantInt>(
-        OptionNode->getOperand(1).get());
+    return mdconst::extract_or_null<ConstantInt>(MD->getOperand(1).get());
   }
   llvm_unreachable("unexpected number of options");
 }
 
 static bool getBooleanLoopAttribute(const Loop *TheLoop, StringRef Name) {
   return getOptionalBoolLoopAttribute(TheLoop, Name).getValueOr(false);
 }
 
if (InheritAllAttrs || InheritSomeAttrs) {
     }
   } else {
     // Modified if we dropped at least one attribute.
     Changed = OrigLoopID->getNumOperands() > 1;
   }
 
   bool HasAnyFollowup = false;
   for (StringRef OptionName : FollowupOptions) {
-    MDNode *FollowupNode =
-        findOptionMDForLoopID(OrigLoopID, OptionName).getValueOr(nullptr);
+    MDNode *FollowupNode = findOptionMDForLoopID(OrigLoopID, OptionName);
     if (!FollowupNode)
       continue;
 
     HasAnyFollowup = true;
     for (const MDOperand &Option : drop_begin(FollowupNode->operands(), 1)) {
       MDs.push_back(Option.get());
       Changed = true;
     }

lib/Transforms/Utils/SimplifyCFG.cpp

if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
                              LLVMContext::MD_range,
                              LLVMContext::MD_fpmath,
                              LLVMContext::MD_invariant_load,
                              LLVMContext::MD_nonnull,
                              LLVMContext::MD_invariant_group,
                              LLVMContext::MD_align,
                              LLVMContext::MD_dereferenceable,
                              LLVMContext::MD_dereferenceable_or_null,
-                             LLVMContext::MD_mem_parallel_loop_access};
+                             LLVMContext::MD_mem_parallel_loop_access,
+                             LLVMContext::MD_access_group};
       combineMetadata(I1, I2, KnownIDs, true);
 
       // I1 and I2 are being combined into a single instruction.  Its debug
       // location is the merged locations of the original instructions.
       I1->applyMergedLocation(I1->getDebugLoc(), I2->getDebugLoc());
 
       I2->eraseFromParent();
       Changed = true;

test/Analysis/LoopInfo/annotated-parallel-complex.ll

+; RUN: opt -loops -analyze < %s | FileCheck %s
+;
+; void func(long n, double A[static const restrict 4*n], double B[static const restrict 4*n]) {
+;   for (long i = 0; i < n; i += 1)
+;     for (long j = 0; j < n; j += 1)
+;       for (long k = 0; k < n; k += 1)
+;         for (long l = 0; l < n; l += 1) {
+;           A[i + j + k + l] = 21;
+;           B[i + j + k + l] = 42;
+;         }
+; }
+;
+; Check that isAnnotatedParallel is working as expected.
+;
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+define void @func(i64 %n, double* noalias nonnull %A, double* noalias nonnull %B) {
+entry:
+  br label %for.cond
+
+for.cond:
+  %i.0 = phi i64 [ 0, %entry ], [ %add28, %for.inc27 ]
+  %cmp = icmp slt i64 %i.0, %n
+  br i1 %cmp, label %for.cond2, label %for.end29
+
+for.cond2:
+  %j.0 = phi i64 [ %add25, %for.inc24 ], [ 0, %for.cond ]
+  %cmp3 = icmp slt i64 %j.0, %n
+  br i1 %cmp3, label %for.cond6, label %for.inc27
+
+for.cond6:
+  %k.0 = phi i64 [ %add22, %for.inc21 ], [ 0, %for.cond2 ]
+  %cmp7 = icmp slt i64 %k.0, %n
+  br i1 %cmp7, label %for.cond10, label %for.inc24
+
+for.cond10:
+  %l.0 = phi i64 [ %add20, %for.body13 ], [ 0, %for.cond6 ]
+  %cmp11 = icmp slt i64 %l.0, %n
+  br i1 %cmp11, label %for.body13, label %for.inc21
+
+for.body13:
+  %add = add nuw nsw i64 %i.0, %j.0
+  %add14 = add nuw nsw i64 %add, %k.0
+  %add15 = add nuw nsw i64 %add14, %l.0
+  %arrayidx = getelementptr inbounds double, double* %A, i64 %add15
+  store double 2.100000e+01, double* %arrayidx, align 8, !llvm.access.group !5
+  %add16 = add nuw nsw i64 %i.0, %j.0
+  %add17 = add nuw nsw i64 %add16, %k.0
+  %add18 = add nuw nsw i64 %add17, %l.0
+  %arrayidx19 = getelementptr inbounds double, double* %B, i64 %add18
+  store double 4.200000e+01, double* %arrayidx19, align 8, !llvm.access.group !6
+  %add20 = add nuw nsw i64 %l.0, 1
+  br label %for.cond10, !llvm.loop !11
+
+for.inc21:
+  %add22 = add nuw nsw i64 %k.0, 1
+  br label %for.cond6, !llvm.loop !14
+
+for.inc24:
+  %add25 = add nuw nsw i64 %j.0, 1
+  br label %for.cond2, !llvm.loop !16
+
+for.inc27:
+  %add28 = add nuw nsw i64 %i.0, 1
+  br label %for.cond, !llvm.loop !18
+
+for.end29:
+  ret void
+}
+
+; access groups
+!7 = distinct !{}
+!8 = distinct !{}
+!10 = distinct !{}
+
+; access group lists
+!5 = !{!7, !10}
+!6 = !{!7, !8, !10}
+
+; LoopIDs
+!11 = distinct !{!11, !{!"llvm.loop.parallel_accesses", !10}}
+!14 = distinct !{!14, !{!"llvm.loop.parallel_accesses", !8, !10}}
+!16 = distinct !{!16, !{!"llvm.loop.parallel_accesses", !8}}
+!18 = distinct !{!18, !{!"llvm.loop.parallel_accesses", !7}}
+
+
+; CHECK: Parallel Loop at depth 1
+; CHECK-NOT: Parallel
+; CHECK:     Loop at depth 2
+; CHECK:         Parallel Loop
+; CHECK:             Parallel Loop

test/Analysis/LoopInfo/annotated-parallel-simple.ll

+; RUN: opt -loops -analyze < %s | FileCheck %s
+;
+; void func(long n, double A[static const restrict n]) {
+;   for (long i = 0; i < n; i += 1)
+;     A[i] = 21;
+; }
+;
+; Check that isAnnotatedParallel is working as expected.
+;
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+define void @func(i64 %n, double* noalias nonnull %A) {
+entry:
+  br label %for.cond
+
+for.cond:
+  %i.0 = phi i64 [ 0, %entry ], [ %add, %for.body ]
+  %cmp = icmp slt i64 %i.0, %n
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:
+  %arrayidx = getelementptr inbounds double, double* %A, i64 %i.0
+  store double 2.100000e+01, double* %arrayidx, align 8, !llvm.access.group !6
+  %add = add nuw nsw i64 %i.0, 1
+  br label %for.cond, !llvm.loop !7
+
+for.end:
+  ret void
+}
+
+!6 = distinct !{} ; access group
+
+!7 = distinct !{!7, !9} ; LoopID
+!9 = !{!"llvm.loop.parallel_accesses", !6}
+
+
+; CHECK: Parallel Loop

test/ThinLTO/X86/lazyload_metadata.ll

 ; Do setup work for all below tests: generate bitcode and combined index
 ; RUN: opt -module-summary %s -o %t.bc -bitcode-mdindex-threshold=0
 ; RUN: opt -module-summary %p/Inputs/lazyload_metadata.ll -o %t2.bc -bitcode-mdindex-threshold=0
 ; RUN: llvm-lto -thinlto-action=thinlink -o %t3.bc %t.bc %t2.bc
 ; REQUIRES: asserts
 
 ; Check that importing @globalfunc1 does not trigger loading all the global
 ; metadata for @globalfunc2 and @globalfunc3
 
 ; RUN: llvm-lto -thinlto-action=import %t2.bc -thinlto-index=%t3.bc \
 ; RUN:          -o /dev/null -stats \
 ; RUN:  2>&1 | FileCheck %s -check-prefix=LAZY
-; LAZY: 55 bitcode-reader  - Number of Metadata records loaded
+; LAZY: 57 bitcode-reader  - Number of Metadata records loaded
 ; LAZY: 2 bitcode-reader  - Number of MDStrings loaded
 
 ; RUN: llvm-lto -thinlto-action=import %t2.bc -thinlto-index=%t3.bc \
 ; RUN:          -o /dev/null -disable-ondemand-mds-loading -stats \
 ; RUN:  2>&1 | FileCheck %s -check-prefix=NOTLAZY
-; NOTLAZY: 64 bitcode-reader  - Number of Metadata records loaded
+; NOTLAZY: 66 bitcode-reader  - Number of Metadata records loaded
 ; NOTLAZY: 7 bitcode-reader  - Number of MDStrings loaded
 
 
 target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-apple-macosx10.11.0"
 
 define void @globalfunc1(i32 %arg) {
   %x = call i1 @llvm.type.test(i8* undef, metadata !"typeid1")

test/Transforms/Inline/parallel-loop-md-callee.ll

+; RUN: opt -S -inline < %s | FileCheck %s
+;
+; Check that the !llvm.access.group is still present after inlining.
+;
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+define void @Body(i32* nocapture %res, i32* nocapture readnone %c, i32* nocapture readonly %d, i32* nocapture readonly %p, i32 %i) {
+entry:
+  %idxprom = sext i32 %i to i64
+  %arrayidx = getelementptr inbounds i32, i32* %p, i64 %idxprom
+  %0 = load i32, i32* %arrayidx, align 4, !llvm.access.group !0
+  %cmp = icmp eq i32 %0, 0
+  %arrayidx2 = getelementptr inbounds i32, i32* %res, i64 %idxprom
+  %1 = load i32, i32* %arrayidx2, align 4, !llvm.access.group !0
+  br i1 %cmp, label %cond.end, label %cond.false
+
+cond.false:
+  %arrayidx6 = getelementptr inbounds i32, i32* %d, i64 %idxprom
+  %2 = load i32, i32* %arrayidx6, align 4, !llvm.access.group !0
+  %add = add nsw i32 %2, %1
+  br label %cond.end
+
+cond.end:
+  %cond = phi i32 [ %add, %cond.false ], [ %1, %entry ]
+  store i32 %cond, i32* %arrayidx2, align 4
+  ret void
+}
+
+define void @Test(i32* %res, i32* %c, i32* %d, i32* %p, i32 %n) {
+entry:
+  br label %for.cond
+
+for.cond:
+  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  %cmp = icmp slt i32 %i.0, 1600
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:
+  call void @Body(i32* %res, i32* undef, i32* %d, i32* %p, i32 %i.0), !llvm.access.group !0
+  %inc = add nsw i32 %i.0, 1
+  br label %for.cond, !llvm.loop !1
+
+for.end:
+  ret void
+}
+
+!0 = distinct !{}                                          ; access group
+!1 = distinct !{!1, !{!"llvm.loop.parallel_accesses", !0}} ; LoopID
+
+
+; CHECK-LABEL: @Test
+; CHECK: load i32,{{.*}}, !llvm.access.group !0
+; CHECK: load i32,{{.*}}, !llvm.access.group !0
+; CHECK: load i32,{{.*}}, !llvm.access.group !0
+; CHECK: store i32 {{.*}}, !llvm.access.group !0
+; CHECK: br label %for.cond, !llvm.loop !1

test/Transforms/Inline/parallel-loop-md-merge.ll

+; RUN: opt -always-inline -globalopt -S < %s | FileCheck %s
+;
+; static void __attribute__((always_inline)) callee(long n, double A[static const restrict n], long i) {
+;   for (long j = 0; j < n; j += 1)
+;     A[i * n + j] = 42;
+; }
+;
+; void caller(long n, double A[static const restrict n]) {
+;   for (long i = 0; i < n; i += 1)
+;     callee(n, A, i);
+; }
+;
+; Check that the access groups (llvm.access.group) are correctly merged.
+;
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+define internal void @callee(i64 %n, double* noalias nonnull %A, i64 %i) #0 {
+entry:
+  br label %for.cond
+
+for.cond:
+  %j.0 = phi i64 [ 0, %entry ], [ %add1, %for.body ]
+  %cmp = icmp slt i64 %j.0, %n
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:
+  %mul = mul nsw i64 %i, %n
+  %add = add nsw i64 %mul, %j.0
+  %arrayidx = getelementptr inbounds double, double* %A, i64 %add
+  store double 4.200000e+01, double* %arrayidx, align 8, !llvm.access.group !6
+  %add1 = add nuw nsw i64 %j.0, 1
+  br label %for.cond, !llvm.loop !7
+
+for.end:
+  ret void
+}
+
+attributes #0 = { alwaysinline }
+
+!6 = distinct !{}       ; access group
+!7 = distinct !{!7, !9} ; LoopID
+!9 = !{!"llvm.loop.parallel_accesses", !6}
+
+
+define void @caller(i64 %n, double* noalias nonnull %A) {
+entry:
+  br label %for.cond
+
+for.cond:
+  %i.0 = phi i64 [ 0, %entry ], [ %add, %for.body ]
+  %cmp = icmp slt i64 %i.0, %n
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:
+  call void @callee(i64 %n, double* %A, i64 %i.0), !llvm.access.group !10
+  %add = add nuw nsw i64 %i.0, 1
+  br label %for.cond, !llvm.loop !11
+
+for.end:
+  ret void
+}
+
+!10 = distinct !{}         ; access group
+!11 = distinct !{!11, !12} ; LoopID
+!12 = !{!"llvm.loop.parallel_accesses", !10}
+
+
+; CHECK: store double 4.200000e+01, {{.*}} !llvm.access.group ![[ACCESS_GROUP_LIST_3:[0-9]+]]
+; CHECK: br label %for.cond.i, !llvm.loop ![[LOOP_INNER:[0-9]+]]
+; CHECK: br label %for.cond, !llvm.loop ![[LOOP_OUTER:[0-9]+]]
+
+; CHECK: ![[ACCESS_GROUP_LIST_3]] = !{![[ACCESS_GROUP_INNER:[0-9]+]], ![[ACCESS_GROUP_OUTER:[0-9]+]]}
+; CHECK: ![[ACCESS_GROUP_INNER]] = distinct !{}
+; CHECK: ![[ACCESS_GROUP_OUTER]] = distinct !{}
+; CHECK: ![[LOOP_INNER]] = distinct !{![[LOOP_INNER]], ![[ACCESSES_INNER:[0-9]+]]}
+; CHECK: ![[ACCESSES_INNER]] = !{!"llvm.loop.parallel_accesses", ![[ACCESS_GROUP_INNER]]}
+; CHECK: ![[LOOP_OUTER]] = distinct !{![[LOOP_OUTER]], ![[ACCESSES_OUTER:[0-9]+]]}
+; CHECK: ![[ACCESSES_OUTER]] = !{!"llvm.loop.parallel_accesses", ![[ACCESS_GROUP_OUTER]]}

test/Transforms/Inline/parallel-loop-md.ll

entry:
   br label %for.cond
 
 for.cond:                                         ; preds = %for.body, %entry
   %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
   %cmp = icmp slt i32 %i.0, 1600
   br i1 %cmp, label %for.body, label %for.end
 
 for.body:                                         ; preds = %for.cond
-  call void @Body(i32* %res, i32* undef, i32* %d, i32* %p, i32 %i.0), !llvm.mem.parallel_loop_access !0
+  call void @Body(i32* %res, i32* undef, i32* %d, i32* %p, i32 %i.0), !llvm.access.group !0
   %inc = add nsw i32 %i.0, 1
-  br label %for.cond, !llvm.loop !0
+  br label %for.cond, !llvm.loop !1
 
 for.end:                                          ; preds = %for.cond
   ret void
 }
 
 ; CHECK-LABEL: @Test
-; CHECK: load i32,{{.*}}, !llvm.mem.parallel_loop_access !0
-; CHECK: load i32,{{.*}}, !llvm.mem.parallel_loop_access !0
-; CHECK: load i32,{{.*}}, !llvm.mem.parallel_loop_access !0
-; CHECK: store i32{{.*}}, !llvm.mem.parallel_loop_access !0
-; CHECK: br label %for.cond, !llvm.loop !0
+; CHECK: load i32,{{.*}}, !llvm.access.group !0
+; CHECK: load i32,{{.*}}, !llvm.access.group !0
+; CHECK: load i32,{{.*}}, !llvm.access.group !0
+; CHECK: store i32{{.*}}, !llvm.access.group !0
+; CHECK: br label %for.cond, !llvm.loop !1
 
 attributes #0 = { norecurse nounwind uwtable }
 
-!0 = distinct !{!0}
+!0 = distinct !{}
+!1 = distinct !{!0, !{!"llvm.loop.parallel_accesses", !0}}
+
 

test/Transforms/InstCombine/intersect-accessgroup.ll

+; RUN: opt -instcombine -S < %s | FileCheck %s
+; 
+; void func(long n, double A[static const restrict n]) {
+;   for (int i = 0; i <  n; i+=1)
+;     for (int j = 0; j <  n;j+=1)
+;       for (int k = 0; k < n; k += 1) 
+;         for (int l = 0; l < n; l += 1) {
+;           double *p = &A[i + j + k + l];
+;           double x = *p;
+;           double y = *p;
+;           arg(x + y);
+;         }
+; }
+;
+; Check for correctly merging access group metadata for instcombine
+; (only common loops are parallel == intersection)
+; Note that combined load would be parallel to loop !16 since both
+; origin loads are parallel to it, but it references two access groups
+; (!8 and !9), neither of which contain both loads. As such, the
+; information that the combined load is parallel to !16 is lost.
+;
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+declare void @arg(double)
+
+define void @func(i64 %n, double* noalias nonnull %A) {
+entry:
+  br label %for.cond
+
+for.cond:
+  %i.0 = phi i32 [ 0, %entry ], [ %add31, %for.inc30 ]
+  %conv = sext i32 %i.0 to i64
+  %cmp = icmp slt i64 %conv, %n
+  br i1 %cmp, label %for.cond2, label %for.end32
+
+for.cond2:
+  %j.0 = phi i32 [ %add28, %for.inc27 ], [ 0, %for.cond ]
+  %conv3 = sext i32 %j.0 to i64
+  %cmp4 = icmp slt i64 %conv3, %n
+  br i1 %cmp4, label %for.cond8, label %for.inc30
+
+for.cond8:
+  %k.0 = phi i32 [ %add25, %for.inc24 ], [ 0, %for.cond2 ]
+  %conv9 = sext i32 %k.0 to i64
+  %cmp10 = icmp slt i64 %conv9, %n
+  br i1 %cmp10, label %for.cond14, label %for.inc27
+
+for.cond14:
+  %l.0 = phi i32 [ %add23, %for.body19 ], [ 0, %for.cond8 ]
+  %conv15 = sext i32 %l.0 to i64
+  %cmp16 = icmp slt i64 %conv15, %n
+  br i1 %cmp16, label %for.body19, label %for.inc24
+
+for.body19:
+  %add = add nsw i32 %i.0, %j.0
+  %add20 = add nsw i32 %add, %k.0
+  %add21 = add nsw i32 %add20, %l.0
+  %idxprom = sext i32 %add21 to i64
+  %arrayidx = getelementptr inbounds double, double* %A, i64 %idxprom
+  %0 = load double, double* %arrayidx, align 8, !llvm.access.group !1
+  %1 = load double, double* %arrayidx, align 8, !llvm.access.group !2
+  %add22 = fadd double %0, %1
+  call void @arg(double %add22), !llvm.access.group !3
+  %add23 = add nsw i32 %l.0, 1
+  br label %for.cond14, !llvm.loop !11
+
+for.inc24:
+  %add25 = add nsw i32 %k.0, 1
+  br label %for.cond8, !llvm.loop !14
+
+for.inc27:
+  %add28 = add nsw i32 %j.0, 1
+  br label %for.cond2, !llvm.loop !16
+
+for.inc30:
+  %add31 = add nsw i32 %i.0, 1
+  br label %for.cond, !llvm.loop !18
+
+for.end32:
+  ret void
+}
+
+
+; access groups
+!7 = distinct !{}
+!8 = distinct !{}
+!9 = distinct !{}
+
+; access group lists
+!1 = !{!7, !9}
+!2 = !{!7, !8}
+!3 = !{!7, !8, !9}
+
+!11 = distinct !{!11, !13}
+!13 = !{!"llvm.loop.parallel_accesses", !7}
+
+!14 = distinct !{!14, !15}
+!15 = !{!"llvm.loop.parallel_accesses", !8}
+
+!16 = distinct !{!16, !17}
+!17 = !{!"llvm.loop.parallel_accesses", !8, !9}
+
+!18 = distinct !{!18, !19}
+!19 = !{!"llvm.loop.parallel_accesses", !9}
+
+
+; CHECK: load double, {{.*}} !llvm.access.group ![[ACCESSGROUP_0:[0-9]+]]
+; CHECK: br label %for.cond14, !llvm.loop ![[LOOP_4:[0-9]+]]
+
+; CHECK: ![[ACCESSGROUP_0]] = distinct !{}
+
+; CHECK: ![[LOOP_4]] = distinct !{![[LOOP_4]], ![[PARALLEL_ACCESSES_5:[0-9]+]]}
+; CHECK: ![[PARALLEL_ACCESSES_5]] = !{!"llvm.loop.parallel_accesses", ![[ACCESSGROUP_0]]}

test/Transforms/InstCombine/loadstore-metadata.ll

entry:
   %l = load i32, i32* %ptr, !range !5
   %c = bitcast i32 %l to float
   ret float %c
 }
 
 define i32 @test_load_cast_combine_invariant(float* %ptr) {
 ; Ensure (cast (load (...))) -> (load (cast (...))) preserves invariant metadata.
 ; CHECK-LABEL: @test_load_cast_combine_invariant(
-; CHECK: load i32, i32* %{{.*}}, !invariant.load !5
+; CHECK: load i32, i32* %{{.*}}, !invariant.load !7
 entry:
   %l = load float, float* %ptr, !invariant.load !6
   %c = bitcast float %l to i32
   ret i32 %c
 }
 
 define i32 @test_load_cast_combine_nontemporal(float* %ptr) {
 ; Ensure (cast (load (...))) -> (load (cast (...))) preserves nontemporal
 ; metadata.
 ; CHECK-LABEL: @test_load_cast_combine_nontemporal(
-; CHECK: load i32, i32* %{{.*}}, !nontemporal !6
+; CHECK: load i32, i32* %{{.*}}, !nontemporal !8
 entry:
   %l = load float, float* %ptr, !nontemporal !7
   %c = bitcast float %l to i32
   ret i32 %c
 }
 
 define i8* @test_load_cast_combine_align(i32** %ptr) {
 ; Ensure (cast (load (...))) -> (load (cast (...))) preserves align
 ; metadata.
 ; CHECK-LABEL: @test_load_cast_combine_align(
-; CHECK: load i8*, i8** %{{.*}}, !align !7
+; CHECK: load i8*, i8** %{{.*}}, !align !9
 entry:
   %l = load i32*, i32** %ptr, !align !8
   %c = bitcast i32* %l to i8*
   ret i8* %c
 }
 
 define i8* @test_load_cast_combine_deref(i32** %ptr) {
 ; Ensure (cast (load (...))) -> (load (cast (...))) preserves dereferenceable
 ; metadata.
 ; CHECK-LABEL: @test_load_cast_combine_deref(
-; CHECK: load i8*, i8** %{{.*}}, !dereferenceable !7
+; CHECK: load i8*, i8** %{{.*}}, !dereferenceable !9
 entry:
   %l = load i32*, i32** %ptr, !dereferenceable !8
   %c = bitcast i32* %l to i8*
   ret i8* %c
 }
 
 define i8* @test_load_cast_combine_deref_or_null(i32** %ptr) {
 ; Ensure (cast (load (...))) -> (load (cast (...))) preserves
 ; dereferenceable_or_null metadata.
 ; CHECK-LABEL: @test_load_cast_combine_deref_or_null(
-; CHECK: load i8*, i8** %{{.*}}, !dereferenceable_or_null !7
+; CHECK: load i8*, i8** %{{.*}}, !dereferenceable_or_null !9
 entry:
   %l = load i32*, i32** %ptr, !dereferenceable_or_null !8
   %c = bitcast i32* %l to i8*
   ret i8* %c
 }
 
 define void @test_load_cast_combine_loop(float* %src, i32* %dst, i32 %n) {
 ; Ensure (cast (load (...))) -> (load (cast (...))) preserves loop access
 ; metadata.
 ; CHECK-LABEL: @test_load_cast_combine_loop(
-; CHECK: load i32, i32* %{{.*}}, !llvm.mem.parallel_loop_access !4
+; CHECK: load i32, i32* %{{.*}}, !llvm.access.group !6
 entry:
   br label %loop
 
 loop:
   %i = phi i32 [ 0, %entry ], [ %i.next, %loop ]
   %src.gep = getelementptr inbounds float, float* %src, i32 %i
   %dst.gep = getelementptr inbounds i32, i32* %dst, i32 %i
-  %l = load float, float* %src.gep, !llvm.mem.parallel_loop_access !4
+  %l = load float, float* %src.gep, !llvm.access.group !9
   %c = bitcast float %l to i32
   store i32 %c, i32* %dst.gep
   %i.next = add i32 %i, 1
   %cmp = icmp slt i32 %i.next, %n
   br i1 %cmp, label %loop, label %exit, !llvm.loop !1
 
 exit:
   ret void

 }
 
 ; This is the metadata tuple that we reference above:
 ; CHECK: ![[MD]] = !{i64 1, i64 0}
 !0 = !{!1, !1, i64 0}
 !1 = !{!"scalar type", !2}
 !2 = !{!"root"}
 !3 = distinct !{!3, !4}
-!4 = distinct !{!4}
+!4 = distinct !{!4, !{!"llvm.loop.parallel_accesses", !9}}
 !5 = !{i32 0, i32 42}
 !6 = !{}
 !7 = !{i32 1}
 !8 = !{i64 8}
+!9 = distinct !{}

test/Transforms/InstCombine/mem-par-metadata-memcpy.ll

 ; RUN: opt < %s -instcombine -S | FileCheck %s
 ;
-; Make sure the llvm.mem.parallel_loop_access meta-data is preserved
+; Make sure the llvm.access.group meta-data is preserved
 ; when a memcpy is replaced with a load+store by instcombine
 ;
 ; #include <string.h>
 ; void test(char* out, long size)
 ; {
 ;     #pragma clang loop vectorize(assume_safety)
 ;     for (long i = 0; i < size; i+=2) {
 ;         memcpy(&(out[i]), &(out[i+size]), 2);
 ;     }
 ; }
 
 ; CHECK: for.body:
-; CHECK:  %{{.*}} = load i16, i16* %{{.*}}, align 1, !llvm.mem.parallel_loop_access !1
-; CHECK:  store i16 %{{.*}}, i16* %{{.*}}, align 1, !llvm.mem.parallel_loop_access !1
+; CHECK:  %{{.*}} = load i16, i16* %{{.*}}, align 1, !llvm.access.group !1
+; CHECK:  store i16 %{{.*}}, i16* %{{.*}}, align 1, !llvm.access.group !1
 
 
 ; ModuleID = '<stdin>'
 source_filename = "memcpy.pragma.cpp"
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 ; Function Attrs: nounwind uwtable
 define void @_Z4testPcl(i8* %out, i64 %size) #0 {
 entry:
   br label %for.cond
 
 for.cond:                                         ; preds = %for.inc, %entry
   %i.0 = phi i64 [ 0, %entry ], [ %add2, %for.inc ]
   %cmp = icmp slt i64 %i.0, %size
   br i1 %cmp, label %for.body, label %for.end
 
 for.body:                                         ; preds = %for.cond
   %arrayidx = getelementptr inbounds i8, i8* %out, i64 %i.0
   %add = add nsw i64 %i.0, %size
   %arrayidx1 = getelementptr inbounds i8, i8* %out, i64 %add
-  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %arrayidx, i8* %arrayidx1, i64 2, i1 false), !llvm.mem.parallel_loop_access !1
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %arrayidx, i8* %arrayidx1, i64 2, i1 false), !llvm.access.group !4
   br label %for.inc
 
 for.inc:                                          ; preds = %for.body
   %add2 = add nsw i64 %i.0, 2
   br label %for.cond, !llvm.loop !2
 
 for.end:                                          ; preds = %for.cond
   ret void
 }
 
 ; Function Attrs: argmemonly nounwind
 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture writeonly, i8* nocapture readonly, i64, i1) #1
 
 attributes #0 = { nounwind uwtable "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
 attributes #1 = { argmemonly nounwind }
 
 !llvm.ident = !{!0}
 
 !0 = !{!"clang version 4.0.0 (cfe/trunk 277751)"}
-!1 = distinct !{!1, !2, !3}
+!1 = distinct !{!1, !2, !3, !{!"llvm.loop.parallel_accesses", !4}}
 !2 = distinct !{!2, !3}
 !3 = !{!"llvm.loop.vectorize.enable", i1 true}
+!4 = distinct !{} ; access group

test/Transforms/LoopVectorize/X86/force-ifcvt.ll

 ; RUN: opt -loop-vectorize -S < %s | FileCheck %s
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 ; Function Attrs: norecurse nounwind uwtable
 define void @Test(i32* nocapture %res, i32* nocapture readnone %c, i32* nocapture readonly %d, i32* nocapture readonly %p) #0 {
 entry:
   br label %for.body
 
 ; CHECK-LABEL: @Test
 ; CHECK: <4 x i32>
 
 for.body:                                         ; preds = %cond.end, %entry
   %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %cond.end ]
   %arrayidx = getelementptr inbounds i32, i32* %p, i64 %indvars.iv
-  %0 = load i32, i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
+  %0 = load i32, i32* %arrayidx, align 4, !llvm.access.group !1
   %cmp1 = icmp eq i32 %0, 0
   %arrayidx3 = getelementptr inbounds i32, i32* %res, i64 %indvars.iv
-  %1 = load i32, i32* %arrayidx3, align 4, !llvm.mem.parallel_loop_access !0
+  %1 = load i32, i32* %arrayidx3, align 4, !llvm.access.group !1
   br i1 %cmp1, label %cond.end, label %cond.false
 
 cond.false:                                       ; preds = %for.body
   %arrayidx7 = getelementptr inbounds i32, i32* %d, i64 %indvars.iv
-  %2 = load i32, i32* %arrayidx7, align 4, !llvm.mem.parallel_loop_access !0
+  %2 = load i32, i32* %arrayidx7, align 4, !llvm.access.group !1
   %add = add nsw i32 %2, %1
   br label %cond.end
 
 cond.end:                                         ; preds = %for.body, %cond.false
   %cond = phi i32 [ %add, %cond.false ], [ %1, %for.body ]
-  store i32 %cond, i32* %arrayidx3, align 4, !llvm.mem.parallel_loop_access !0
+  store i32 %cond, i32* %arrayidx3, align 4, !llvm.access.group !1
   %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
   %exitcond = icmp eq i64 %indvars.iv.next, 16
   br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
 
 for.end:                                          ; preds = %cond.end
   ret void
 }
 
 attributes #0 = { norecurse nounwind uwtable "target-cpu"="x86-64" "target-features"="+fxsr,+mmx,+sse,+sse2,+x87" }
 
-!0 = distinct !{!0}
+!0 = distinct !{!0, !{!"llvm.loop.parallel_accesses", !1}}
+!1 = distinct !{}

test/Transforms/LoopVectorize/X86/parallel-loops-after-reg2mem.ll

entry:
   %indvars.iv.reg2mem = alloca i64
   %"reg2mem alloca point" = bitcast i32 0 to i32
   store i64 0, i64* %indvars.iv.reg2mem
   br label %for.body
 
 for.body:                                         ; preds = %for.body.for.body_crit_edge, %entry
   %indvars.iv.reload = load i64, i64* %indvars.iv.reg2mem
   %arrayidx = getelementptr inbounds i32, i32* %b, i64 %indvars.iv.reload
-  %0 = load i32, i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !3
+  %0 = load i32, i32* %arrayidx, align 4, !llvm.access.group !4
   %arrayidx2 = getelementptr inbounds i32, i32* %a, i64 %indvars.iv.reload
-  %1 = load i32, i32* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !3
+  %1 = load i32, i32* %arrayidx2, align 4, !llvm.access.group !4
   %idxprom3 = sext i32 %1 to i64
   %arrayidx4 = getelementptr inbounds i32, i32* %a, i64 %idxprom3
-  store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !3
+  store i32 %0, i32* %arrayidx4, align 4, !llvm.access.group !4
   %indvars.iv.next = add i64 %indvars.iv.reload, 1
   ; A new store without the parallel metadata here:
   store i64 %indvars.iv.next, i64* %indvars.iv.next.reg2mem
   %indvars.iv.next.reload1 = load i64, i64* %indvars.iv.next.reg2mem
   %arrayidx6 = getelementptr inbounds i32, i32* %b, i64 %indvars.iv.next.reload1
-  %2 = load i32, i32* %arrayidx6, align 4, !llvm.mem.parallel_loop_access !3
-  store i32 %2, i32* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !3
+  %2 = load i32, i32* %arrayidx6, align 4, !llvm.access.group !4
+  store i32 %2, i32* %arrayidx2, align 4, !llvm.access.group !4
   %indvars.iv.next.reload = load i64, i64* %indvars.iv.next.reg2mem
   %lftr.wideiv = trunc i64 %indvars.iv.next.reload to i32
   %exitcond = icmp eq i32 %lftr.wideiv, 512
   br i1 %exitcond, label %for.end, label %for.body.for.body_crit_edge, !llvm.loop !3
 
 for.body.for.body_crit_edge:                      ; preds = %for.body
   %indvars.iv.next.reload2 = load i64, i64* %indvars.iv.next.reg2mem
   store i64 %indvars.iv.next.reload2, i64* %indvars.iv.reg2mem
   br label %for.body
 
 for.end:                                          ; preds = %for.body
   ret void
 }
 
-!3 = !{!3}
+!3 = !{!3, !{!"llvm.loop.parallel_accesses", !4}}
+!4 = distinct !{}

test/Transforms/LoopVectorize/X86/parallel-loops.ll

 ;CHECK: <4 x i32>
 define void @parallel_loop(i32* nocapture %a, i32* nocapture %b) nounwind uwtable {
 entry:
   br label %for.body
 
 for.body:                                         ; preds = %for.body, %entry
   %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
   %arrayidx = getelementptr inbounds i32, i32* %b, i64 %indvars.iv
-  %0 = load i32, i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !3
+  %0 = load i32, i32* %arrayidx, align 4, !llvm.access.group !13
   %arrayidx2 = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
-  %1 = load i32, i32* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !3
+  %1 = load i32, i32* %arrayidx2, align 4, !llvm.access.group !13
   %idxprom3 = sext i32 %1 to i64
   %arrayidx4 = getelementptr inbounds i32, i32* %a, i64 %idxprom3
   ; This store might have originated from inlining a function with a parallel
   ; loop. Refers to a list with the "original loop reference" (!4) also included.
-  store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !5
+  store i32 %0, i32* %arrayidx4, align 4, !llvm.access.group !15
   %indvars.iv.next = add i64 %indvars.iv, 1
   %arrayidx6 = getelementptr inbounds i32, i32* %b, i64 %indvars.iv.next
-  %2 = load i32, i32* %arrayidx6, align 4, !llvm.mem.parallel_loop_access !3
-  store i32 %2, i32* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !3
+  %2 = load i32, i32* %arrayidx6, align 4, !llvm.access.group !13
+  store i32 %2, i32* %arrayidx2, align 4, !llvm.access.group !13
   %lftr.wideiv = trunc i64 %indvars.iv.next to i32
   %exitcond = icmp eq i32 %lftr.wideiv, 512
   br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !3
 
 for.end:                                          ; preds = %for.body
   ret void
 }
 
 ; The same loop with an illegal parallel loop metadata: the memory
 ; accesses refer to a different loop's identifier.
 
 ;CHECK-LABEL: @mixed_metadata(
 ;CHECK-NOT: <4 x i32>
 
 define void @mixed_metadata(i32* nocapture %a, i32* nocapture %b) nounwind uwtable {
 entry:
   br label %for.body
 
 for.body:                                         ; preds = %for.body, %entry
   %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
   %arrayidx = getelementptr inbounds i32, i32* %b, i64 %indvars.iv
-  %0 = load i32, i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !6
+  %0 = load i32, i32* %arrayidx, align 4, !llvm.access.group !16
   %arrayidx2 = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
-  %1 = load i32, i32* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !6
+  %1 = load i32, i32* %arrayidx2, align 4, !llvm.access.group !16
   %idxprom3 = sext i32 %1 to i64
   %arrayidx4 = getelementptr inbounds i32, i32* %a, i64 %idxprom3
   ; This refers to the loop marked with !7 which we are not in at the moment.
   ; It should prevent detecting as a parallel loop.
-  store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !7
+  store i32 %0, i32* %arrayidx4, align 4, !llvm.access.group !17
   %indvars.iv.next = add i64 %indvars.iv, 1
   %arrayidx6 = getelementptr inbounds i32, i32* %b, i64 %indvars.iv.next
-  %2 = load i32, i32* %arrayidx6, align 4, !llvm.mem.parallel_loop_access !6
-  store i32 %2, i32* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !6
+  %2 = load i32, i32* %arrayidx6, align 4, !llvm.access.group !16
+  store i32 %2, i32* %arrayidx2, align 4, !llvm.access.group !16
   %lftr.wideiv = trunc i64 %indvars.iv.next to i32
   %exitcond = icmp eq i32 %lftr.wideiv, 512
   br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !6
 
 for.end:                                          ; preds = %for.body
   ret void
 }
 
-!3 = !{!3}
-!4 = !{!4}
-!5 = !{!3, !4}
-!6 = !{!6}
-!7 = !{!7}
+!3 = !{!3, !{!"llvm.loop.parallel_accesses", !13, !15}}
+!4 = !{!4, !{!"llvm.loop.parallel_accesses", !14, !15}}
+!6 = !{!6, !{!"llvm.loop.parallel_accesses", !16}}
+!7 = !{!7, !{!"llvm.loop.parallel_accesses", !17}}
+!13 = distinct !{}
+!14 = distinct !{}
+!15 = distinct !{}
+!16 = distinct !{}
+!17 = distinct !{}

test/Transforms/LoopVectorize/X86/pr34438.ll

 ; CHECK:    load <8 x float>, <8 x float>*
 ; CHECK:    fadd fast <8 x float>
 entry:
   br label %for.body
 
 for.body:
   %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
   %arrayidx = getelementptr inbounds float, float* %B, i64 %indvars.iv
-  %0 = load float, float* %arrayidx, align 4, !llvm.mem.parallel_loop_access !3
+  %0 = load float, float* %arrayidx, align 4, !llvm.access.group !5
   %arrayidx2 = getelementptr inbounds float, float* %A, i64 %indvars.iv
-  %1 = load float, float* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !3
+  %1 = load float, float* %arrayidx2, align 4, !llvm.access.group !5
   %add = fadd fast float %0, %1
-  store float %add, float* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !3
+  store float %add, float* %arrayidx2, align 4, !llvm.access.group !5
   %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
   %exitcond = icmp eq i64 %indvars.iv.next, 8
   br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !4
 
 for.end:
   ret void
 }
 
-!3 = !{!3}
+!3 = !{!3, !{!"llvm.loop.parallel_accesses", !5}}
 !4 = !{!4}
+!5 = distinct !{}

test/Transforms/LoopVectorize/X86/vect.omp.force.ll

 
 define void @vectorized(float* noalias nocapture %A, float* noalias nocapture %B) {
 entry:
   br label %for.body
 
 for.body:
   %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
   %arrayidx = getelementptr inbounds float, float* %B, i64 %indvars.iv
-  %0 = load float, float* %arrayidx, align 4, !llvm.mem.parallel_loop_access !1
+  %0 = load float, float* %arrayidx, align 4, !llvm.access.group !11
   %call = tail call float @llvm.sin.f32(float %0)
   %arrayidx2 = getelementptr inbounds float, float* %A, i64 %indvars.iv
-  store float %call, float* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !1
+  store float %call, float* %arrayidx2, align 4, !llvm.access.group !11
   %indvars.iv.next = add nuw nsw i64 %indvars.iv, 2
   %lftr.wideiv = trunc i64 %indvars.iv.next to i32
   %exitcond = icmp eq i32 %lftr.wideiv, 1000
   br i1 %exitcond, label %for.end.loopexit, label %for.body, !llvm.loop !1
 
 for.end.loopexit:
   br label %for.end
 
 for.end:
   ret void
 }
 
-!1 = !{!1, !2}
+!1 = !{!1, !2, !{!"llvm.loop.parallel_accesses", !11}}
 !2 = !{!"llvm.loop.vectorize.enable", i1 true}
+!11 = distinct !{}
 
 ;
 ; This method will not be vectorized, as scalar cost is lower than any of vector costs.
 ;
 
 define void @not_vectorized(float* noalias nocapture %A, float* noalias nocapture %B) {
 entry:
   br label %for.body
 
 for.body:
   %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
   %arrayidx = getelementptr inbounds float, float* %B, i64 %indvars.iv
-  %0 = load float, float* %arrayidx, align 4, !llvm.mem.parallel_loop_access !3
+  %0 = load float, float* %arrayidx, align 4, !llvm.access.group !13
   %call = tail call float @llvm.sin.f32(float %0)
   %arrayidx2 = getelementptr inbounds float, float* %A, i64 %indvars.iv
-  store float %call, float* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !3
+  store float %call, float* %arrayidx2, align 4, !llvm.access.group !13
   %indvars.iv.next = add nuw nsw i64 %indvars.iv, 2
   %lftr.wideiv = trunc i64 %indvars.iv.next to i32
   %exitcond = icmp eq i32 %lftr.wideiv, 1000
   br i1 %exitcond, label %for.end.loopexit, label %for.body, !llvm.loop !3
 
 for.end.loopexit:
   br label %for.end
 
 for.end:
   ret void
 }
 
 declare float @llvm.sin.f32(float) nounwind readnone
 
 ; Dummy metadata
-!3 = !{!3}
+!3 = !{!3, !{!"llvm.loop.parallel_accesses", !13}}
+!13 = distinct !{}
 

test/Transforms/LoopVectorize/X86/vect.omp.force.small-tc.ll

 ; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds float, float* [[TMP4]], i32 0
 ; CHECK-NEXT:    [[TMP6:%.*]] = bitcast float* [[TMP5]] to <8 x float>*
 ; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <8 x float>, <8 x float>* [[TMP6]], align 4
 ; CHECK-NEXT:    [[TMP7:%.*]] = fadd fast <8 x float> [[WIDE_LOAD]], [[WIDE_LOAD1]]
 ; CHECK-NEXT:    [[TMP8:%.*]] = bitcast float* [[TMP5]] to <8 x float>*
 ; CHECK-NEXT:    store <8 x float> [[TMP7]], <8 x float>* [[TMP8]], align 4
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 8
 ; CHECK-NEXT:    [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], 16
-; CHECK-NEXT:    br i1 [[TMP9]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !0
+; CHECK-NEXT:    br i1 [[TMP9]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !1
 ; CHECK:       middle.block:
 ; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 20, 16
 ; CHECK-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 16, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
 ; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds float, float* [[B]], i64 [[INDVARS_IV]]
-; CHECK-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX]], align 4, !llvm.mem.parallel_loop_access !3
+; CHECK-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX]], align 4, !llvm.access.group !0
 ; CHECK-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds float, float* [[A]], i64 [[INDVARS_IV]]
-; CHECK-NEXT:    [[TMP11:%.*]] = load float, float* [[ARRAYIDX2]], align 4, !llvm.mem.parallel_loop_access !3
+; CHECK-NEXT:    [[TMP11:%.*]] = load float, float* [[ARRAYIDX2]], align 4, !llvm.access.group !0
 ; CHECK-NEXT:    [[ADD:%.*]] = fadd fast float [[TMP10]], [[TMP11]]
-; CHECK-NEXT:    store float [[ADD]], float* [[ARRAYIDX2]], align 4, !llvm.mem.parallel_loop_access !3
+; CHECK-NEXT:    store float [[ADD]], float* [[ARRAYIDX2]], align 4, !llvm.access.group !0
 ; CHECK-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
 ; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], 20
-; CHECK-NEXT:    br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop !4
+; CHECK-NEXT:    br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop !5
 ; CHECK:       for.end:
 ; CHECK-NEXT:    ret void
 ;
 entry:
   br label %for.body
 
 for.body:
   %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
   %arrayidx = getelementptr inbounds float, float* %B, i64 %indvars.iv
-  %0 = load float, float* %arrayidx, align 4, !llvm.mem.parallel_loop_access !1
+  %0 = load float, float* %arrayidx, align 4, !llvm.access.group !11
   %arrayidx2 = getelementptr inbounds float, float* %A, i64 %indvars.iv
-  %1 = load float, float* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !1
+  %1 = load float, float* %arrayidx2, align 4, !llvm.access.group !11
   %add = fadd fast float %0, %1
-  store float %add, float* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !1
+  store float %add, float* %arrayidx2, align 4, !llvm.access.group !11
   %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
   %exitcond = icmp eq i64 %indvars.iv.next, 20
   br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !1
 
 for.end:
   ret void
 }
 
-!1 = !{!1, !2}
+!1 = !{!1, !2, !{!"llvm.loop.parallel_accesses", !11}}
 !2 = !{!"llvm.loop.vectorize.enable", i1 true}
+!11 = distinct !{}
 
 ;
 ; This loop will be vectorized as the trip count is below the threshold but no
 ; scalar iterations are needed thanks to folding its tail.
 ;
 define void @vectorized1(float* noalias nocapture %A, float* noalias nocapture readonly %B) {
 ; CHECK-LABEL: @vectorized1(
 ; CHECK-NEXT:  entry:

 ; CHECK-NEXT:    [[TMP6:%.*]] = bitcast float* [[TMP5]] to <8 x float>*
 ; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <8 x float>, <8 x float>* [[TMP6]], align 4
 ; CHECK-NEXT:    [[TMP7:%.*]] = fadd fast <8 x float> [[WIDE_LOAD]], [[WIDE_LOAD1]]
 ; CHECK-NEXT:    [[TMP8:%.*]] = icmp ule <8 x i64> [[INDUCTION]], <i64 19, i64 19, i64 19, i64 19, i64 19, i64 19, i64 19, i64 19>
 ; CHECK-NEXT:    [[TMP9:%.*]] = bitcast float* [[TMP5]] to <8 x float>*
 ; CHECK-NEXT:    call void @llvm.masked.store.v8f32.p0v8f32(<8 x float> [[TMP7]], <8 x float>* [[TMP9]], i32 4, <8 x i1> [[TMP8]])
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 8
 ; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], 24
-; CHECK-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !6
+; CHECK-NEXT:    br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !8
 ; CHECK:       middle.block:
 ; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK:       for.end:
 ; CHECK-NEXT:    ret void
 ;
 entry:
   br label %for.body
 
 for.body:
   %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
   %arrayidx = getelementptr inbounds float, float* %B, i64 %indvars.iv
-  %0 = load float, float* %arrayidx, align 4, !llvm.mem.parallel_loop_access !3
+  %0 = load float, float* %arrayidx, align 4, !llvm.access.group !13
   %arrayidx2 = getelementptr inbounds float, float* %A, i64 %indvars.iv
-  %1 = load float, float* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !3
+  %1 = load float, float* %arrayidx2, align 4, !llvm.access.group !13
   %add = fadd fast float %0, %1
-  store float %add, float* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !3
+  store float %add, float* %arrayidx2, align 4, !llvm.access.group !13
   %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
   %exitcond = icmp eq i64 %indvars.iv.next, 20
   br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !3
 
 for.end:
   ret void
 }
 
-!3 = !{!3}
+!3 = !{!3, !{!"llvm.loop.parallel_accesses", !13}}
+!13 = distinct !{}
 
 ;
 ; This loop will be vectorized as the trip count is below the threshold but no
 ; scalar iterations are needed.
 ;
 define void @vectorized2(float* noalias nocapture %A, float* noalias nocapture readonly %B) {
 ; CHECK-LABEL: @vectorized2(
 ; CHECK-NEXT:  entry:

 ; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds float, float* [[TMP4]], i32 0
 ; CHECK-NEXT:    [[TMP6:%.*]] = bitcast float* [[TMP5]] to <8 x float>*
 ; CHECK-NEXT:    [[WIDE_LOAD1:%.*]] = load <8 x float>, <8 x float>* [[TMP6]], align 4
 ; CHECK-NEXT:    [[TMP7:%.*]] = fadd fast <8 x float> [[WIDE_LOAD]], [[WIDE_LOAD1]]
 ; CHECK-NEXT:    [[TMP8:%.*]] = bitcast float* [[TMP5]] to <8 x float>*
 ; CHECK-NEXT:    store <8 x float> [[TMP7]], <8 x float>* [[TMP8]], align 4
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 8
 ; CHECK-NEXT:    [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], 16
-; CHECK-NEXT:    br i1 [[TMP9]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !9
+; CHECK-NEXT:    br i1 [[TMP9]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !11
 ; CHECK:       middle.block:
 ; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 16, 16
 ; CHECK-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 16, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
 ; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds float, float* [[B]], i64 [[INDVARS_IV]]
-; CHECK-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX]], align 4, !llvm.mem.parallel_loop_access !7
+; CHECK-NEXT:    [[TMP10:%.*]] = load float, float* [[ARRAYIDX]], align 4, !llvm.access.group !7
 ; CHECK-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds float, float* [[A]], i64 [[INDVARS_IV]]
-; CHECK-NEXT:    [[TMP11:%.*]] = load float, float* [[ARRAYIDX2]], align 4, !llvm.mem.parallel_loop_access !7
+; CHECK-NEXT:    [[TMP11:%.*]] = load float, float* [[ARRAYIDX2]], align 4, !llvm.access.group !7
 ; CHECK-NEXT:    [[ADD:%.*]] = fadd fast float [[TMP10]], [[TMP11]]
-; CHECK-NEXT:    store float [[ADD]], float* [[ARRAYIDX2]], align 4, !llvm.mem.parallel_loop_access !7
+; CHECK-NEXT:    store float [[ADD]], float* [[ARRAYIDX2]], align 4, !llvm.access.group !7
 ; CHECK-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
 ; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], 16
-; CHECK-NEXT:    br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop !10
+; CHECK-NEXT:    br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop !12
 ; CHECK:       for.end:
 ; CHECK-NEXT:    ret void
 ;
 entry:
   br label %for.body
 
 for.body:
   %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
   %arrayidx = getelementptr inbounds float, float* %B, i64 %indvars.iv
-  %0 = load float, float* %arrayidx, align 4, !llvm.mem.parallel_loop_access !3
+  %0 = load float, float* %arrayidx, align 4, !llvm.access.group !13
   %arrayidx2 = getelementptr inbounds float, float* %A, i64 %indvars.iv
-  %1 = load float, float* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !3
+  %1 = load float, float* %arrayidx2, align 4, !llvm.access.group !13
   %add = fadd fast float %0, %1
-  store float %add, float* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !3
+  store float %add, float* %arrayidx2, align 4, !llvm.access.group !13
   %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
   %exitcond = icmp eq i64 %indvars.iv.next, 16
   br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !4
 
 for.end:
   ret void
 }
 
 !4 = !{!4}
 

test/Transforms/LoopVectorize/X86/vector_max_bandwidth.ll

 ; CHECK-LABEL: not_too_small_tc
 ; CHECK-AVX2: LV: Selecting VF: 16.
 entry:
   br label %for.body
 
 for.body:
   %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
   %arrayidx = getelementptr inbounds i8, i8* %B, i64 %indvars.iv
-  %l1 = load i8, i8* %arrayidx, align 4, !llvm.mem.parallel_loop_access !3
+  %l1 = load i8, i8* %arrayidx, align 4, !llvm.access.group !13
   %arrayidx2 = getelementptr inbounds i8, i8* %A, i64 %indvars.iv
-  %l2 = load i8, i8* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !3
+  %l2 = load i8, i8* %arrayidx2, align 4, !llvm.access.group !13
   %add = add i8 %l1, %l2
-  store i8 %add, i8* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !3
+  store i8 %add, i8* %arrayidx2, align 4, !llvm.access.group !13
   %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
   %exitcond = icmp eq i64 %indvars.iv.next, 16
   br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !4
 
 for.end:
   ret void
 }
-!3 = !{!3}
+!3 = !{!3, !{!"llvm.loop.parallel_accesses", !13}}
 !4 = !{!4}
+!13 = distinct !{}

test/Transforms/SROA/mem-par-metadata-sroa.ll

 ; RUN: opt < %s -sroa -S | FileCheck %s
 ;
-; Make sure the llvm.mem.parallel_loop_access meta-data is preserved
+; Make sure the llvm.access.group meta-data is preserved
 ; when a load/store is replaced with another load/store by sroa
 ;
 ; class Complex {
 ; private:
 ;  float real_;
 ;  float imaginary_;
 ;
 ; public:

 ;     for (long offset = 0; offset < size; ++offset) {
 ;       Complex t0 = out[offset];
 ;       out[offset] = t0 + t0;
 ;     }
 ; }
 
 ; CHECK: for.body:
 ; CHECK-NOT:  store i32 %{{.*}}, i32* %{{.*}}, align 4
-; CHECK: store i32 %{{.*}}, i32* %{{.*}}, align 4, !llvm.mem.parallel_loop_access !1
+; CHECK: store i32 %{{.*}}, i32* %{{.*}}, align 4, !llvm.access.group !1
 ; CHECK-NOT:  store i32 %{{.*}}, i32* %{{.*}}, align 4
-; CHECK: store i32 %{{.*}}, i32* %{{.*}}, align 4, !llvm.mem.parallel_loop_access !1
+; CHECK: store i32 %{{.*}}, i32* %{{.*}}, align 4, !llvm.access.group !1
 ; CHECK-NOT:  store i32 %{{.*}}, i32* %{{.*}}, align 4
 ; CHECK: br label
 
 ; ModuleID = '<stdin>'
 source_filename = "mem-par-metadata-sroa1.cpp"
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
for.cond:                                         ; preds = %for.body, %entry
   %offset.0 = phi i64 [ 0, %entry ], [ %inc, %for.body ]
   %cmp = icmp slt i64 %offset.0, %size
   br i1 %cmp, label %for.body, label %for.end
 
 for.body:                                         ; preds = %for.cond
   %arrayidx = getelementptr inbounds %class.Complex, %class.Complex* %out, i64 %offset.0
   %real_.i = getelementptr inbounds %class.Complex, %class.Complex* %t0, i64 0, i32 0
   %real_.i.i = getelementptr inbounds %class.Complex, %class.Complex* %arrayidx, i64 0, i32 0
-  %0 = load float, float* %real_.i.i, align 4, !llvm.mem.parallel_loop_access !1
-  store float %0, float* %real_.i, align 4, !llvm.mem.parallel_loop_access !1
+  %0 = load float, float* %real_.i.i, align 4, !llvm.access.group !11
+  store float %0, float* %real_.i, align 4, !llvm.access.group !11
   %imaginary_.i = getelementptr inbounds %class.Complex, %class.Complex* %t0, i64 0, i32 1
   %imaginary_.i.i = getelementptr inbounds %class.Complex, %class.Complex* %arrayidx, i64 0, i32 1
-  %1 = load float, float* %imaginary_.i.i, align 4, !llvm.mem.parallel_loop_access !1
-  store float %1, float* %imaginary_.i, align 4, !llvm.mem.parallel_loop_access !1
+  %1 = load float, float* %imaginary_.i.i, align 4, !llvm.access.group !11
+  store float %1, float* %imaginary_.i, align 4, !llvm.access.group !11
   %arrayidx1 = getelementptr inbounds %class.Complex, %class.Complex* %out, i64 %offset.0
   %real_.i1 = getelementptr inbounds %class.Complex, %class.Complex* %t0, i64 0, i32 0
-  %2 = load float, float* %real_.i1, align 4, !noalias !3, !llvm.mem.parallel_loop_access !1
+  %2 = load float, float* %real_.i1, align 4, !noalias !3, !llvm.access.group !11
   %real_2.i = getelementptr inbounds %class.Complex, %class.Complex* %t0, i64 0, i32 0
-  %3 = load float, float* %real_2.i, align 4, !noalias !3, !llvm.mem.parallel_loop_access !1
+  %3 = load float, float* %real_2.i, align 4, !noalias !3, !llvm.access.group !11
   %add.i = fadd float %2, %3
   %imaginary_.i2 = getelementptr inbounds %class.Complex, %class.Complex* %t0, i64 0, i32 1
-  %4 = load float, float* %imaginary_.i2, align 4, !noalias !3, !llvm.mem.parallel_loop_access !1
+  %4 = load float, float* %imaginary_.i2, align 4, !noalias !3, !llvm.access.group !11
   %imaginary_3.i = getelementptr inbounds %class.Complex, %class.Complex* %t0, i64 0, i32 1
-  %5 = load float, float* %imaginary_3.i, align 4, !noalias !3, !llvm.mem.parallel_loop_access !1
+  %5 = load float, float* %imaginary_3.i, align 4, !noalias !3, !llvm.access.group !11
   %add4.i = fadd float %4, %5
   %real_.i.i3 = getelementptr inbounds %class.Complex, %class.Complex* %tmpcast, i64 0, i32 0
-  store float %add.i, float* %real_.i.i3, align 4, !alias.scope !3, !llvm.mem.parallel_loop_access !1
+  store float %add.i, float* %real_.i.i3, align 4, !alias.scope !3, !llvm.access.group !11
   %imaginary_.i.i4 = getelementptr inbounds %class.Complex, %class.Complex* %tmpcast, i64 0, i32 1
-  store float %add4.i, float* %imaginary_.i.i4, align 4, !alias.scope !3, !llvm.mem.parallel_loop_access !1
+  store float %add4.i, float* %imaginary_.i.i4, align 4, !alias.scope !3, !llvm.access.group !11
   %6 = bitcast %class.Complex* %arrayidx1 to i64*
-  %7 = load i64, i64* %ref.tmp, align 8, !llvm.mem.parallel_loop_access !1
-  store i64 %7, i64* %6, align 4, !llvm.mem.parallel_loop_access !1
+  %7 = load i64, i64* %ref.tmp, align 8, !llvm.access.group !11
+  store i64 %7, i64* %6, align 4, !llvm.access.group !11
   %inc = add nsw i64 %offset.0, 1
   br label %for.cond, !llvm.loop !1
 
 for.end:                                          ; preds = %for.cond
   ret void
 }
 
 ; Function Attrs: argmemonly nounwind
 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture writeonly, i8* nocapture readonly, i64, i1) #1
 
 attributes #0 = { norecurse nounwind uwtable "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
 attributes #1 = { argmemonly nounwind }
 
 !llvm.ident = !{!0}
 
 !0 = !{!"clang version 4.0.0 (cfe/trunk 277751)"}
-!1 = distinct !{!1, !2}
+!1 = distinct !{!1, !2, !{!"llvm.loop.parallel_accesses", !11}}
 !2 = !{!"llvm.loop.vectorize.enable", i1 true}
 !3 = !{!4}
 !4 = distinct !{!4, !5, !"_ZNK7ComplexplERKS_: %agg.result"}
 !5 = distinct !{!5, !"_ZNK7ComplexplERKS_"}
+!11 = distinct !{}

test/Transforms/Scalarizer/basic.ll

 ; CHECK: store i32 %add.i3, i32* %dst.i3, align 4, !tbaa.struct ![[TAG]]
 ; CHECK: ret void
   %val = load <4 x i32> , <4 x i32> *%src, !tbaa.struct !5
   %add = add <4 x i32> %val, %val
   store <4 x i32> %add, <4 x i32> *%dst, !tbaa.struct !5
   ret void
 }
 
-; Check that llvm.mem.parallel_loop_access information is preserved.
+; Check that llvm.access.group information is preserved.
 define void @f5(i32 %count, <4 x i32> *%src, <4 x i32> *%dst) {
 ; CHECK-LABEL: @f5(
-; CHECK: %val.i0 = load i32, i32* %this_src.i0, align 16, !llvm.mem.parallel_loop_access ![[TAG:[0-9]*]]
-; CHECK: %val.i1 = load i32, i32* %this_src.i1, align 4, !llvm.mem.parallel_loop_access ![[TAG]]
-; CHECK: %val.i2 = load i32, i32* %this_src.i2, align 8, !llvm.mem.parallel_loop_access ![[TAG]]
-; CHECK: %val.i3 = load i32, i32* %this_src.i3, align 4, !llvm.mem.parallel_loop_access ![[TAG]]
-; CHECK: store i32 %add.i0, i32* %this_dst.i0, align 16, !llvm.mem.parallel_loop_access ![[TAG]]
-; CHECK: store i32 %add.i1, i32* %this_dst.i1, align 4, !llvm.mem.parallel_loop_access ![[TAG]]
-; CHECK: store i32 %add.i2, i32* %this_dst.i2, align 8, !llvm.mem.parallel_loop_access ![[TAG]]
-; CHECK: store i32 %add.i3, i32* %this_dst.i3, align 4, !llvm.mem.parallel_loop_access ![[TAG]]
+; CHECK: %val.i0 = load i32, i32* %this_src.i0, align 16, !llvm.access.group ![[TAG:[0-9]*]]
+; CHECK: %val.i1 = load i32, i32* %this_src.i1, align 4, !llvm.access.group ![[TAG]]
+; CHECK: %val.i2 = load i32, i32* %this_src.i2, align 8, !llvm.access.group ![[TAG]]
+; CHECK: %val.i3 = load i32, i32* %this_src.i3, align 4, !llvm.access.group ![[TAG]]
+; CHECK: store i32 %add.i0, i32* %this_dst.i0, align 16, !llvm.access.group ![[TAG]]
+; CHECK: store i32 %add.i1, i32* %this_dst.i1, align 4, !llvm.access.group ![[TAG]]
+; CHECK: store i32 %add.i2, i32* %this_dst.i2, align 8, !llvm.access.group ![[TAG]]
+; CHECK: store i32 %add.i3, i32* %this_dst.i3, align 4, !llvm.access.group ![[TAG]]
 ; CHECK: ret void
 entry:
   br label %loop
 
 loop:
   %index = phi i32 [ 0, %entry ], [ %next_index, %loop ]
   %this_src = getelementptr <4 x i32>, <4 x i32> *%src, i32 %index
   %this_dst = getelementptr <4 x i32>, <4 x i32> *%dst, i32 %index
-  %val = load <4 x i32> , <4 x i32> *%this_src, !llvm.mem.parallel_loop_access !3
+  %val = load <4 x i32> , <4 x i32> *%this_src, !llvm.access.group !13
   %add = add <4 x i32> %val, %val
-  store <4 x i32> %add, <4 x i32> *%this_dst, !llvm.mem.parallel_loop_access !3
+  store <4 x i32> %add, <4 x i32> *%this_dst, !llvm.access.group !13
   %next_index = add i32 %index, -1
   %continue = icmp ne i32 %next_index, %count
   br i1 %continue, label %loop, label %end, !llvm.loop !3
 
 end:
   ret void
 }
 

 
 exit:
   ret <4 x float> %next_acc
 }
 
 !0 = !{ !"root" }
 !1 = !{ !"set1", !0 }
 !2 = !{ !"set2", !0 }
-!3 = !{ !3 }
+!3 = !{ !3, !{!"llvm.loop.parallel_accesses", !13} }
 !4 = !{ float 4.0 }
 !5 = !{ i64 0, i64 8, null }
+!13 = distinct !{}

test/Transforms/SimplifyCFG/combine-parallel-mem-md.ll

 ; RUN: opt -simplifycfg -S < %s | FileCheck %s
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 ; Function Attrs: norecurse nounwind uwtable
 define void @Test(i32* nocapture %res, i32* nocapture readnone %c, i32* nocapture readonly %d, i32* nocapture readonly %p) #0 {
 entry:
   br label %for.body
 
 ; CHECK-LABEL: @Test
-; CHECK: load i32, i32* {{.*}}, align 4, !llvm.mem.parallel_loop_access !0
-; CHECK: load i32, i32* {{.*}}, align 4, !llvm.mem.parallel_loop_access !0
-; CHECK: store i32 {{.*}}, align 4, !llvm.mem.parallel_loop_access !0
+; CHECK: load i32, i32* {{.*}}, align 4, !llvm.access.group !0
+; CHECK: load i32, i32* {{.*}}, align 4, !llvm.access.group !0
+; CHECK: store i32 {{.*}}, align 4, !llvm.access.group !0
 ; CHECK-NOT: load
 ; CHECK-NOT: store
 
 for.body:                                         ; preds = %cond.end, %entry
   %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %cond.end ]
   %arrayidx = getelementptr inbounds i32, i32* %p, i64 %indvars.iv
-  %0 = load i32, i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
+  %0 = load i32, i32* %arrayidx, align 4, !llvm.access.group !0
   %cmp1 = icmp eq i32 %0, 0
   br i1 %cmp1, label %cond.true, label %cond.false
 
 cond.false:                                       ; preds = %for.body
   %arrayidx3 = getelementptr inbounds i32, i32* %res, i64 %indvars.iv
-  %v = load i32, i32* %arrayidx3, align 4, !llvm.mem.parallel_loop_access !0
+  %v = load i32, i32* %arrayidx3, align 4, !llvm.access.group !0
   %arrayidx7 = getelementptr inbounds i32, i32* %d, i64 %indvars.iv
-  %1 = load i32, i32* %arrayidx7, align 4, !llvm.mem.parallel_loop_access !0
+  %1 = load i32, i32* %arrayidx7, align 4, !llvm.access.group !0
   %add = add nsw i32 %1, %v
   br label %cond.end
 
 cond.true:                                       ; preds = %for.body
   %arrayidx4 = getelementptr inbounds i32, i32* %res, i64 %indvars.iv
-  %w = load i32, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
+  %w = load i32, i32* %arrayidx4, align 4, !llvm.access.group !0
   %arrayidx8 = getelementptr inbounds i32, i32* %d, i64 %indvars.iv
-  %2 = load i32, i32* %arrayidx8, align 4, !llvm.mem.parallel_loop_access !0
+  %2 = load i32, i32* %arrayidx8, align 4, !llvm.access.group !0
   %add2 = add nsw i32 %2, %w
   br label %cond.end
 
 cond.end:                                         ; preds = %for.body, %cond.false
   %cond = phi i32 [ %add, %cond.false ], [ %add2, %cond.true ]
   %arrayidx9 = getelementptr inbounds i32, i32* %res, i64 %indvars.iv
-  store i32 %cond, i32* %arrayidx9, align 4, !llvm.mem.parallel_loop_access !0
+  store i32 %cond, i32* %arrayidx9, align 4, !llvm.access.group !0
   %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
   %exitcond = icmp eq i64 %indvars.iv.next, 16
   br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
 
 for.end:                                          ; preds = %cond.end
   ret void
 }
 
 attributes #0 = { norecurse nounwind uwtable }
 
-!0 = distinct !{!0, !1}
+!0 = distinct !{!0, !1, !{!"llvm.loop.parallel_accesses", !10}}
 !1 = !{!"llvm.loop.vectorize.enable", i1 true}
+!10 = distinct !{}