[cfe-commits] [patch] Qualifiers refactor

[John McCall]

John McCall wrote:
> The idea about adding a non-trivial branch to the old version is a good 
> one;  I'll let you know when I have numbers.
>   

I made a fresh checkout of LLVM/clang 82704 and ran the standard benchmark
twice (n=40), once after applying the following patch:


--- include/clang/AST/Type.h    (revision 82704)
+++ include/clang/AST/Type.h    (working copy)
@@ -109,7 +109,7 @@ public:
 
   unsigned getCVRQualifiers() const { return Value.getInt(); }
   void setCVRQualifiers(unsigned Quals) { Value.setInt(Quals); }
-  Type *getTypePtr() const { return Value.getPointer(); }
+  Type *getTypePtr() const { return Value.getPointer() == (Type *) 0xd0 
? 0 : Value.getPointer(); }
 
   void *getAsOpaquePtr() const { return Value.getOpaqueValue(); }
   static QualType getFromOpaquePtr(void *Ptr) {


This is (very roughly) parallel to the change in getTypePtr() that the 
QualType
refactor introduced;  it's probably slightly more work, but it captures 
the general
idea.

No branch in getTypePtr() (pristine 82704):
name      avg     min     med     max      SD    total
user     0.2120  0.2117  0.2120  0.2126  0.0002  8.4810
 sys     0.0316  0.0306  0.0319  0.0308  0.0016  1.2658
wall     0.2478  0.2464  0.2475  0.2482  0.0008  9.9118

With branch in getTypePtr():
user     0.2228  0.2226  0.2228  0.2231  0.0001  8.9120
 sys     0.0317  0.0311  0.0306  0.0300  0.0017  1.2690
wall     0.2593  0.2593  0.2588  0.2586  0.0021 10.3739

So the branch causes a 5% regression in average userland time.

John.