[PATCH] Fix -fsanitize=alignment reports on DenseMap empty/tombstone keys.

[Alexey Samsonov]

In http://reviews.llvm.org/D6768#103892, @chandlerc wrote:

> I'm curious why we need to support forward declared types. It seems really
>  brittle. Have you looked into this much? If not, I can.


If your class has a member `llvm::DenseMap<llvm::AllocaInst *, Foo>`, it's a common practice to forward-declare llvm::AllocaInst.

In http://reviews.llvm.org/D6768#103896, @majnemer wrote:

> I had a similar patch here: http://reviews.llvm.org/D5428
>
> I went with a different approach to handle global objects which were massively over-aligned (unlikely, but possible).


Right, your patch looks slightly better. I'd be happy to see either of these commited :)


================
Comment at: include/llvm/ADT/DenseMapInfo.h:38-47
@@ -34,8 +37,12 @@
     uintptr_t Val = static_cast<uintptr_t>(-1);
-    Val <<= PointerLikeTypeTraits<T*>::NumLowBitsAvailable;
+    Val <<= PointerLikeTypeTraits<T *>::NumLowBitsAvailable < MaxAlignmentLog
+                ? MaxAlignmentLog
+                : PointerLikeTypeTraits<T *>::NumLowBitsAvailable;
     return reinterpret_cast<T*>(Val);
   }
   static inline T* getTombstoneKey() {
     uintptr_t Val = static_cast<uintptr_t>(-2);
-    Val <<= PointerLikeTypeTraits<T*>::NumLowBitsAvailable;
+    Val <<= PointerLikeTypeTraits<T *>::NumLowBitsAvailable < MaxAlignmentLog
+                ? MaxAlignmentLog
+                : PointerLikeTypeTraits<T *>::NumLowBitsAvailable;
     return reinterpret_cast<T*>(Val);
----------------
majnemer wrote:
> Would using `std::max` make this more readable?
This header is included by almost every LLVM source. I didn't want to #include <algorithm> here.

http://reviews.llvm.org/D6768

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/