[clang] [compiler-rt] [llvm] [InstrFDO][TypeProf] Implement binary instrumentation and profile read/write (PR #66825)

[Snehasish Kumar via llvm-commits]

================
@@ -429,20 +439,36 @@ uint64_t ComputeHash(StringRef K);
 class InstrProfSymtab {
 public:
   using AddrHashMap = std::vector<std::pair<uint64_t, uint64_t>>;
+  using RangeHashMap =
+      std::vector<std::pair<std::pair<uint64_t, uint64_t>, uint64_t>>;
 
 private:
   StringRef Data;
   uint64_t Address = 0;
-  // Unique name strings.
+  // Unique name strings. Used to ensure entries in MD5NameMap (a vector that's
+  // going to be sorted) has unique MD5 keys in the first place.
   StringSet<> NameTab;
+  // Records the unique virtual table names. This is used by InstrProfWriter to
+  // write out an on-disk chained hash table of virtual table names.
+  // InstrProfWriter stores per function profile data (keyed by function names)
+  // so it doesn't use a StringSet for function names.
+  StringSet<> VTableNames;
   // A map from MD5 keys to function name strings.
   std::vector<std::pair<uint64_t, StringRef>> MD5NameMap;
+  // A map from MD5 keys to virtual table definitions. Only populated when
+  // building the Symtab from a module.
+  std::vector<std::pair<uint64_t, GlobalVariable *>> MD5VTableMap;
----------------
snehasish wrote:

Having `Map` in the name is confusing when the data structure is a vector. In this case I think using a DenseMap is better since the keys will be random and spread apart. So it should have better performance than log(N) operations necessary for lookup in a sorted vector.

https://github.com/llvm/llvm-project/pull/66825