[IR] Introduce !captures metadata

llvm/docs/LangRef.rst

@@ -1489,6 +1489,8 @@ Currently, only the following parameter attributes are defined:
     function, returning a pointer to allocated storage disjoint from the
     storage for any other object accessible to the caller.
 
+.. _captures_attr:
+
 ``captures(...)``
     This attribute restricts the ways in which the callee may capture the
     pointer. This is not a valid attribute for return values. This attribute
@@ -7543,6 +7545,33 @@ The number of bytes known to be dereferenceable is specified by the integer
 value in the metadata node. This is analogous to the ''dereferenceable_or_null''
 attribute on parameters and return values.
 
+'``captures``' Metadata
+^^^^^^^^^^^^^^^^^^^^^^^
+
+The ``!captures`` metadata can only be applied to ``store`` instructions with
+a pointer-typed value operand. It restricts the capturing behavior of the store
+value operand in the same way the ``captures(...)`` attribute would do on a
+call. See the :ref:`pointer capture section <pointercapture>` for a detailed
+discussion of capture semantics.
+
+The ``!captures`` metadata accepts a non-empty list of strings from the same
+set as the :ref:`captures attribute <captures_attr>`:
+``!"address"``, ``!"address_is_null"``, ``!"provenance"`` and
+``!"read_provenance"``. ``!"none"`` is not supported.
+
+For example ``store ptr %x, ptr %y, !captures !{!"address"}`` indicates that
+the copy of pointer ``%x`` stored to location ``%y`` will only be used to
+inspect its integral address value, and not dereferenced. Dereferencing the
+pointer would result in undefined behavior.
+
+Similarly ``store ptr %x, ptr %y, !captures !{!"address", !"read_provenance"}``
+indicates that while reads through the stored pointer are allowed, writes would
+result in undefined behavior.
+
+The ``!captures`` attribute makes no statement about other uses of ``%x``, or
+uses of the stored-to memory location after it has been overwritten with a
+different value.
+
 .. _llvm.loop:
 
 '``llvm.loop``'

llvm/include/llvm/IR/FixedMetadataKinds.def

@@ -55,3 +55,4 @@ LLVM_FIXED_MD_KIND(MD_mmra, "mmra", 40)
 LLVM_FIXED_MD_KIND(MD_noalias_addrspace, "noalias.addrspace", 41)
 LLVM_FIXED_MD_KIND(MD_callee_type, "callee_type", 42)
 LLVM_FIXED_MD_KIND(MD_nofree, "nofree", 43)
+LLVM_FIXED_MD_KIND(MD_captures, "captures", 44)

llvm/include/llvm/IR/Metadata.h

@@ -41,6 +41,7 @@
 
 namespace llvm {
 
+enum class CaptureComponents : uint8_t;
 class Module;
 class ModuleSlotTracker;
 class raw_ostream;
@@ -1480,6 +1481,13 @@ class MDNode : public Metadata {
   LLVM_ABI static MDNode *getMergedCallsiteMetadata(MDNode *A, MDNode *B);
   LLVM_ABI static MDNode *getMergedCalleeTypeMetadata(const MDNode *A,
                                                       const MDNode *B);
+
+  /// Convert !captures metadata to CaptureComponents. MD may be nullptr.
+  LLVM_ABI static CaptureComponents toCaptureComponents(const MDNode *MD);
+  /// Convert CaptureComponents to !captures metadata. The return value may be
+  /// nullptr.
+  LLVM_ABI static MDNode *fromCaptureComponents(LLVMContext &Ctx,
+                                                CaptureComponents CC);
 };
 
 /// Tuple of metadata.

llvm/lib/Analysis/CaptureTracking.cpp

@@ -320,8 +320,12 @@ UseCaptureInfo llvm::DetermineUseCaptureKind(const Use &U, const Value *Base) {
     return CaptureComponents::None;
   case Instruction::Store:
     // Stored the pointer - conservatively assume it may be captured.
+    if (U.getOperandNo() == 0)
+      return MDNode::toCaptureComponents(
+          I->getMetadata(LLVMContext::MD_captures));
+
     // Volatile stores make the address observable.
-    if (U.getOperandNo() == 0 || cast<StoreInst>(I)->isVolatile())
+    if (cast<StoreInst>(I)->isVolatile())
       return CaptureComponents::All;
     return CaptureComponents::None;
   case Instruction::AtomicRMW: {

llvm/lib/IR/Metadata.cpp

@@ -48,6 +48,7 @@
 #include "llvm/Support/Casting.h"
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Support/MathExtras.h"
+#include "llvm/Support/ModRef.h"
 #include <cassert>
 #include <cstddef>
 #include <cstdint>
@@ -1435,6 +1436,40 @@ MDNode *MDNode::getMostGenericAlignmentOrDereferenceable(MDNode *A, MDNode *B) {
   return B;
 }
 
+CaptureComponents MDNode::toCaptureComponents(const MDNode *MD) {
+  if (!MD)
+    return CaptureComponents::All;
+
+  CaptureComponents CC = CaptureComponents::None;
+  for (Metadata *Op : MD->operands()) {
+    CaptureComponents Component =
+        StringSwitch<CaptureComponents>(cast<MDString>(Op)->getString())
+            .Case("address", CaptureComponents::Address)
+            .Case("address_is_null", CaptureComponents::AddressIsNull)
+            .Case("provenance", CaptureComponents::Provenance)
+            .Case("read_provenance", CaptureComponents::ReadProvenance);
+    CC |= Component;
+  }
+  return CC;
+}
+
+MDNode *MDNode::fromCaptureComponents(LLVMContext &Ctx, CaptureComponents CC) {
+  assert(!capturesNothing(CC) && "Can't encode captures(none)");
+  if (capturesAll(CC))
+    return nullptr;
+
+  SmallVector<Metadata *> Components;
+  if (capturesAddressIsNullOnly(CC))
+    Components.push_back(MDString::get(Ctx, "address_is_null"));
+  else if (capturesAddress(CC))
+    Components.push_back(MDString::get(Ctx, "address"));
+  if (capturesReadProvenanceOnly(CC))
+    Components.push_back(MDString::get(Ctx, "read_provenance"));
+  else if (capturesFullProvenance(CC))
+    Components.push_back(MDString::get(Ctx, "provenance"));
+  return MDNode::get(Ctx, Components);
+}
+
 //===----------------------------------------------------------------------===//
 // NamedMDNode implementation.
 //

llvm/lib/IR/Verifier.cpp

@@ -542,6 +542,7 @@ class Verifier : public InstVisitor<Verifier>, VerifierSupport {
   void visitAliasScopeMetadata(const MDNode *MD);
   void visitAliasScopeListMetadata(const MDNode *MD);
   void visitAccessGroupMetadata(const MDNode *MD);
+  void visitCapturesMetadata(Instruction &I, const MDNode *Captures);
 
   template <class Ty> bool isValidMetadataArray(const MDTuple &N);
 #define HANDLE_SPECIALIZED_MDNODE_LEAF(CLASS) void visit##CLASS(const CLASS &N);
@@ -5373,6 +5374,27 @@ void Verifier::visitAccessGroupMetadata(const MDNode *MD) {
   }
 }
 
+void Verifier::visitCapturesMetadata(Instruction &I, const MDNode *Captures) {
+  static const char *ValidArgs[] = {"address_is_null", "address",
+                                    "read_provenance", "provenance"};
+
+  auto *SI = dyn_cast<StoreInst>(&I);
+  Check(SI, "!captures metadata can only be applied to store instructions", &I);
+  Check(SI->getValueOperand()->getType()->isPointerTy(),
+        "!captures metadata can only be applied to store with value operand of "
+        "pointer type",
+        &I);
+  Check(Captures->getNumOperands() != 0, "!captures metadata cannot be empty",
+        &I);
+
+  for (Metadata *Op : Captures->operands()) {
+    auto *Str = dyn_cast<MDString>(Op);
+    Check(Str, "!captures metadata must be a list of strings", &I);
+    Check(is_contained(ValidArgs, Str->getString()),
+          "invalid entry in !captures metadata", &I, Str);
+  }
+}
+
 /// verifyInstruction - Verify that an instruction is well formed.
 ///
 void Verifier::visitInstruction(Instruction &I) {
@@ -5600,6 +5622,9 @@ void Verifier::visitInstruction(Instruction &I) {
   if (MDNode *Annotation = I.getMetadata(LLVMContext::MD_annotation))
     visitAnnotationMetadata(Annotation);
 
+  if (MDNode *Captures = I.getMetadata(LLVMContext::MD_captures))
+    visitCapturesMetadata(I, Captures);
+
   if (MDNode *N = I.getDebugLoc().getAsMDNode()) {
     CheckDI(isa<DILocation>(N), "invalid !dbg metadata attachment", &I, N);
     visitMDNode(*N, AreDebugLocsAllowed::Yes);

llvm/lib/Transforms/Utils/Local.cpp

@@ -3025,6 +3025,12 @@ static void combineMetadata(Instruction *K, const Instruction *J,
         // Preserve !nosanitize if both K and J have it.
         K->setMetadata(Kind, JMD);
         break;
+      case LLVMContext::MD_captures:
+        K->setMetadata(
+            Kind, MDNode::fromCaptureComponents(
+                      K->getContext(), MDNode::toCaptureComponents(JMD) |
+                                           MDNode::toCaptureComponents(KMD)));
+        break;
       }
   }
   // Set !invariant.group from J if J has it. If both instructions have it

llvm/test/Transforms/FunctionAttrs/nocapture.ll

@@ -1398,5 +1398,73 @@ define void @assume_nonnull(ptr %p) {
   ret void
 }
 
+define void @captures_metadata_address_is_null(ptr %x, ptr %y) {
+; FNATTRS: Function Attrs: mustprogress nofree norecurse nosync nounwind willreturn memory(argmem: write)
+; FNATTRS-LABEL: define void @captures_metadata_address_is_null
+; FNATTRS-SAME: (ptr captures(address_is_null) [[X:%.*]], ptr writeonly captures(none) initializes((0, 8)) [[Y:%.*]]) #[[ATTR17]] {
+; FNATTRS-NEXT:    store ptr [[X]], ptr [[Y]], align 8, !captures [[META0:![0-9]+]]
+; FNATTRS-NEXT:    ret void
+;
+; ATTRIBUTOR: Function Attrs: mustprogress nofree norecurse nosync nounwind willreturn memory(argmem: write)
+; ATTRIBUTOR-LABEL: define void @captures_metadata_address_is_null
+; ATTRIBUTOR-SAME: (ptr nofree writeonly [[X:%.*]], ptr nofree nonnull writeonly captures(none) [[Y:%.*]]) #[[ATTR13]] {
+; ATTRIBUTOR-NEXT:    store ptr [[X]], ptr [[Y]], align 8, !captures [[META0:![0-9]+]]
+; ATTRIBUTOR-NEXT:    ret void
+;
+  store ptr %x, ptr %y, !captures !{!"address_is_null"}
+  ret void
+}
+
+define void @captures_metadata_address(ptr %x, ptr %y) {
+; FNATTRS: Function Attrs: mustprogress nofree norecurse nosync nounwind willreturn memory(argmem: write)
+; FNATTRS-LABEL: define void @captures_metadata_address
+; FNATTRS-SAME: (ptr captures(address) [[X:%.*]], ptr writeonly captures(none) initializes((0, 8)) [[Y:%.*]]) #[[ATTR17]] {
+; FNATTRS-NEXT:    store ptr [[X]], ptr [[Y]], align 8, !captures [[META1:![0-9]+]]
+; FNATTRS-NEXT:    ret void
+;
+; ATTRIBUTOR: Function Attrs: mustprogress nofree norecurse nosync nounwind willreturn memory(argmem: write)
+; ATTRIBUTOR-LABEL: define void @captures_metadata_address
+; ATTRIBUTOR-SAME: (ptr nofree writeonly [[X:%.*]], ptr nofree nonnull writeonly captures(none) [[Y:%.*]]) #[[ATTR13]] {
+; ATTRIBUTOR-NEXT:    store ptr [[X]], ptr [[Y]], align 8, !captures [[META1:![0-9]+]]
+; ATTRIBUTOR-NEXT:    ret void
+;
+  store ptr %x, ptr %y, !captures !{!"address"}
+  ret void
+}
+
+define void @captures_metadata_address_read_provenance(ptr %x, ptr %y) {
+; FNATTRS: Function Attrs: mustprogress nofree norecurse nosync nounwind willreturn memory(argmem: write)
+; FNATTRS-LABEL: define void @captures_metadata_address_read_provenance
+; FNATTRS-SAME: (ptr captures(address, read_provenance) [[X:%.*]], ptr writeonly captures(none) initializes((0, 8)) [[Y:%.*]]) #[[ATTR17]] {
+; FNATTRS-NEXT:    store ptr [[X]], ptr [[Y]], align 8, !captures [[META2:![0-9]+]]
+; FNATTRS-NEXT:    ret void
+;
+; ATTRIBUTOR: Function Attrs: mustprogress nofree norecurse nosync nounwind willreturn memory(argmem: write)
+; ATTRIBUTOR-LABEL: define void @captures_metadata_address_read_provenance
+; ATTRIBUTOR-SAME: (ptr nofree writeonly [[X:%.*]], ptr nofree nonnull writeonly captures(none) [[Y:%.*]]) #[[ATTR13]] {
+; ATTRIBUTOR-NEXT:    store ptr [[X]], ptr [[Y]], align 8, !captures [[META2:![0-9]+]]
+; ATTRIBUTOR-NEXT:    ret void
+;
+  store ptr %x, ptr %y, !captures !{!"address", !"read_provenance"}
+  ret void
+}
+
+define void @captures_metadata_provenance(ptr %x, ptr %y) {
+; FNATTRS: Function Attrs: mustprogress nofree norecurse nosync nounwind willreturn memory(argmem: write)
+; FNATTRS-LABEL: define void @captures_metadata_provenance
+; FNATTRS-SAME: (ptr captures(provenance) [[X:%.*]], ptr writeonly captures(none) initializes((0, 8)) [[Y:%.*]]) #[[ATTR17]] {
+; FNATTRS-NEXT:    store ptr [[X]], ptr [[Y]], align 8, !captures [[META3:![0-9]+]]
+; FNATTRS-NEXT:    ret void
+;
+; ATTRIBUTOR: Function Attrs: mustprogress nofree norecurse nosync nounwind willreturn memory(argmem: write)
+; ATTRIBUTOR-LABEL: define void @captures_metadata_provenance
+; ATTRIBUTOR-SAME: (ptr nofree writeonly [[X:%.*]], ptr nofree nonnull writeonly captures(none) [[Y:%.*]]) #[[ATTR13]] {
+; ATTRIBUTOR-NEXT:    store ptr [[X]], ptr [[Y]], align 8, !captures [[META3:![0-9]+]]
+; ATTRIBUTOR-NEXT:    ret void
+;
+  store ptr %x, ptr %y, !captures !{!"provenance"}
+  ret void
+}
+
 declare ptr @llvm.launder.invariant.group.p0(ptr)
 declare ptr @llvm.strip.invariant.group.p0(ptr)