[cfe-commits] [PATCH] Even More Abbreviations for PCH Serialization

[Douglas Gregor]

On Jun 3, 2011, at 11:59 AM, Jonathan Turner wrote:

> This patch expands to even more PCH abbreviations for serialization.  The major change here is now DeclaratorDecl children serialize the TypeSourceInfo last, which allows us to capture it as an array, as it's the last element serialized.
> 
> This gives some major improvements to ParmVarDecl serialization, where it was once only abbreviated a single-digit percentage of the time, now it's abbreviated more than 90% of the time for ObjC headers tested, and more than 50% of the time for C++ headers test.
> 
> With this I also added support for abbreviating EnumDecl, ObjCIvarDecl, TypedefDecl, VarDecl and FieldDecl.
> 
> For ObjC headers like Cocoa and WebKit, this brings our total size reduction to ~6.5%.  For C++ headers like string and iostream, the reduction is in the 5.5 to 6.5% range.

That's great!

> I removed support for EnumConstantDecl from the previous patch, as it was pushing the number of abbreviations into the next bit size, which had a net loss.

Okay.

> I also cleaned up the naming a bit to be more uniform.


Index: lib/Serialization/ASTWriterDecl.cpp
===================================================================
--- lib/Serialization/ASTWriterDecl.cpp	(revision 132563)
+++ lib/Serialization/ASTWriterDecl.cpp	(working copy)
@@ -124,6 +124,13 @@
 void ASTDeclWriter::Visit(Decl *D) {
   DeclVisitor<ASTDeclWriter>::Visit(D);
 
+  // Source locations require array (variable-length) abbreviations.  The
+  // abbreviation infrastructure requires that arrays are encoded last, so
+  // we handle it here in the case of those clases derived from DeclaratorDecl

Typo "clases"

@@ -169,6 +176,20 @@
 void ASTDeclWriter::VisitTypedefDecl(TypedefDecl *D) {
   VisitTypeDecl(D);
   Writer.AddTypeSourceInfo(D->getTypeSourceInfo(), Record);
+
+  if (!D->hasAttrs() &&
+      !D->isImplicit() &&
+      !D->isUsed(false) &&
+      D->getPCHLevel() == 0 &&
+      D->RedeclLink.getNext() == D &&
+      !D->isInvalidDecl() &&
+      !D->isReferenced() &&
+      D->getAccess() == AS_none &&
+      D->getDeclName().getNameKind() == DeclarationName::Identifier &&
+      D->getTypeSourceInfo())
+
+    AbbrevToUse = Writer.getDeclTypedefAbbrev();

Do we actually need the D->getTypeSourceInfo() check here? It seems unnecessary.

Also: that extra blank line before the AbbrevToUse assignment looks weird.

@@ -205,6 +226,22 @@
   Record.push_back(D->isScopedUsingClassTag());
   Record.push_back(D->isFixed());
   Writer.AddDeclRef(D->getInstantiatedFromMemberEnum(), Record);
+
+  if (!D->hasAttrs() &&
+      !D->isImplicit() &&
+      !D->isUsed(false) &&
+      D->getPCHLevel() == 0 &&
+      !D->hasExtInfo() &&
+      D->RedeclLink.getNext() == D &&
+      !D->isInvalidDecl() &&
+      !D->isReferenced() &&
+      D->getAccess() == AS_none &&
+      !CXXRecordDecl::classofKind(D->getKind()) &&
+      !D->getIntegerTypeSourceInfo() &&
+      D->getDeclName().getNameKind() == DeclarationName::Identifier)
+
+    AbbrevToUse = Writer.getDeclEnumAbbrev();

Same comment about the extra line (there are a few other cases of this).

With those little cleanups, this patch looks good! If you haven't done so already, please request commit access as described here:

	http://llvm.org/docs/DeveloperPolicy.html#commitaccess

	- Doug