[Serialization] The hash seed for DeclCommentTableInfo can't change

Unlike compiled modules, swiftdoc files are considered a stable
format, so we can't change how information is stored in them. If we
add any more string-hashed tables to swiftdoc files, we should
consider using a new hash seed for those.

Author: jrose-apple

lib/Serialization/DocFormat.h

@@ -69,6 +69,15 @@ const uint16_t SWIFTDOC_VERSION_MAJOR = 1;
 /// adhering to the 80-column limit for this line.
 const uint16_t SWIFTDOC_VERSION_MINOR = 1; // Last change: skipping 0 for testing purposes
 
+/// The hash seed used for the string hashes in a Swift 5.1 swiftdoc file.
+///
+/// 0 is not a good seed for llvm::djbHash, but swiftdoc files use a stable
+/// format, so we can't change the hash seed without a version bump. Any new
+/// hashed strings should use a new stable hash seed constant. (No such constant
+/// has been defined at the time this doc comment was last updated because there
+/// are no other strings to hash.)
+const uint32_t SWIFTDOC_HASH_SEED_5_1 = 0;
+
 /// The record types within the comment block.
 ///
 /// Be very careful when changing this block; it must remain

lib/Serialization/ModuleFile.cpp

@@ -972,8 +972,7 @@ class ModuleFile::DeclCommentTableInfo {
 
   hash_value_type ComputeHash(internal_key_type key) {
     assert(!key.empty());
-    // FIXME: DJB seed=0, audit whether the default seed could be used.
-    return llvm::djbHash(key, 0);
+    return llvm::djbHash(key, SWIFTDOC_HASH_SEED_5_1);
   }
 
   static bool EqualKey(internal_key_type lhs, internal_key_type rhs) {

lib/Serialization/SerializeDoc.cpp

@@ -197,8 +197,7 @@ class DeclCommentTableInfo {
 
   hash_value_type ComputeHash(key_type_ref key) {
     assert(!key.empty());
-    // FIXME: DJB seed=0, audit whether the default seed could be used.
-    return llvm::djbHash(key, 0);
+    return llvm::djbHash(key, SWIFTDOC_HASH_SEED_5_1);
   }
 
   std::pair<unsigned, unsigned>