Crash on alignments >= 2**16 (PR26444)

Clang Frontend
1 
Sema::AddAlignment() defines MaxValidAlignment like this:


  // Alignment calculations can wrap around if it's greater than 2**28.
  unsigned MaxValidAlignment =
      Context.getTargetInfo().getTriple().isOSBinFormatCOFF() ? 8192
                                                              : 268435456;

But AggValueSlot stores Alignment as an unsigned short:

class AggValueSlot {
...
  unsigned short Alignment;

The max value for an unsigned short is 2**16 -1, but you are passing 2**16, which AddeAlignment says is okay.  However, it ends up getting stored as 0 in an unsigned short.

The fix is to make these sizes consistent, but I'm not sure which should be changed  (though I'm guessing Alignment should be unsigned instead of unsigned short).
2

From the LLVM perspective, the maximum alignment is specified as:

In Value.h:
00520   /// \brief The maximum alignment for instructions.
00521   ///
00522   /// This is the greatest alignment value supported by load, store, and alloca
00523   /// instructions, and global values.
[00524](http://llvm.org/docs/doxygen/html/classllvm_1_1Value.html#a50e02a54cc6f87a00a671265e1447e61)   static const unsigned [MaxAlignmentExponent](http://llvm.org/docs/doxygen/html/classllvm_1_1Value.html#a50e02a54cc6f87a00a671265e1447e61) = 29;
[00525](http://llvm.org/docs/doxygen/html/classllvm_1_1Value.html#a2885f6c5817ff68bc24bd1e22e221740)   static const unsigned [MaximumAlignment](http://llvm.org/docs/doxygen/html/classllvm_1_1Value.html#a2885f6c5817ff68bc24bd1e22e221740) = 1u << [MaxAlignmentExponent](http://llvm.org/docs/doxygen/html/classllvm_1_1Value.html#a50e02a54cc6f87a00a671265e1447e61);

This is checked in the Verifier via:
Assert(GV.getAlignment() <= Value::MaximumAlignment,
“huge alignment values are unsupported”, &GV);
(and similiar checks for Load, Store, Alloca, but not the various atomic instructions)

3

Hi Philip:

Clang does check to make sure max alignment is <= 29, so that is consistent, however, AggValueSlot stores it in an unsigned short. On my system, OSX, that means values >= 2**16 are stored as 0, which causes and assert/crash.

I submitted a patch over the weekend to change it to an unsigned, http://reviews.llvm.org/D17248, but it hasn’t been reviewed yet.

thanks for taking a look…
don