[clang] [clang][NFC] Refactor `Selector` to use `PointerIntPair` inside (PR #69916)

Aaron Ballman via cfe-commits cfe-commits at lists.llvm.org
Tue Oct 24 05:24:38 PDT 2023


================
@@ -809,43 +927,42 @@ class Selector {
   enum IdentifierInfoFlag {
     // Empty selector = 0. Note that these enumeration values must
     // correspond to the enumeration values of DeclarationName::StoredNameKind
-    ZeroArg  = 0x01,
-    OneArg   = 0x02,
+    ZeroArg = 0x01,
+    OneArg = 0x02,
     MultiArg = 0x07,
-    ArgFlags = 0x07
   };
 
   /// A pointer to the MultiKeywordSelector or IdentifierInfo. We use the low
-  /// three bits of InfoPtr to store an IdentifierInfoFlag. Note that in any
+  /// three bits of InfoPtr to store an IdentifierInfoFlag, but the highest
+  /// of them is also a discriminator for pointer type. Note that in any
   /// case IdentifierInfo and MultiKeywordSelector are already aligned to
   /// 8 bytes even on 32 bits archs because of DeclarationName.
-  uintptr_t InfoPtr = 0;
+  llvm::PointerIntPair<
+      llvm::PointerUnion<IdentifierInfo *, MultiKeywordSelector *>, 2>
+      InfoPtr;
 
   Selector(IdentifierInfo *II, unsigned nArgs) {
-    InfoPtr = reinterpret_cast<uintptr_t>(II);
-    assert((InfoPtr & ArgFlags) == 0 &&"Insufficiently aligned IdentifierInfo");
     assert(nArgs < 2 && "nArgs not equal to 0/1");
-    InfoPtr |= nArgs+1;
+    InfoPtr.setPointerAndInt(II, nArgs + 1);
   }
 
   Selector(MultiKeywordSelector *SI) {
-    InfoPtr = reinterpret_cast<uintptr_t>(SI);
-    assert((InfoPtr & ArgFlags) == 0 &&"Insufficiently aligned IdentifierInfo");
-    InfoPtr |= MultiArg;
+    InfoPtr.setPointerAndInt(SI, MultiArg & 0b11);
----------------
AaronBallman wrote:

It's a bit strange to me that `MultiArg` is 0x7 and not 0x3 to begin with; are we able to reorder `DeclarationName::StoredNameKind` such that `StoredDeclarationNameExtra` is value 3 and then we can shave a bit off from the value of `MultiArg` and not have to do this odd masking here?

https://github.com/llvm/llvm-project/pull/69916


More information about the cfe-commits mailing list