[PATCH] Avoid the use of signed to unsigned integer conversion implementation-defined behavior

[avieira]

Summary

Hello,

This patch is a one-line change to avoid using implementation-defined behavior to check whether the MSB bit of Alignment is set in MachineConstantPool.

Can someone commit this please?

Cheers,
Andre

[javed.absar]

Looks fine to me but probably someone else should also need to have a look. Could you add a small test, if possible.

[avieira]

Euhm, what sort of test would you propose?

[rengolin]

I'm not a language lawyer. Adding some more people that know the standard much better than I do.

[rnk]

What problem does this solve? I'm pretty sure LLVM does not run on ones' complement machines. The comparison against zero is "more canonical" in LLVM IR anyway.

[avieira]

OK, I just thought having it rely on implementation defined behavior wasn't very nice. Though I do agree it is not a big deal in this case...

I'll drop this. (Hoping thats what "abandon revision" means.

include/llvm/CodeGen/MachineConstantPool.h

MachineConstantPoolEntry(MachineConstantPoolValue *V, unsigned A)
     Val.MachineCPVal = V;
     Alignment |= 1U << (sizeof(unsigned) * CHAR_BIT - 1);
   }
 
   /// isMachineConstantPoolEntry - Return true if the MachineConstantPoolEntry
   /// is indeed a target specific constantpool entry, not a wrapper over a
   /// Constant.
   bool isMachineConstantPoolEntry() const {
-    return (int)Alignment < 0;
+    return Alignment & ((unsigned) INT_MAX + 1);
   }
 
   int getAlignment() const {
     return Alignment & ~(1 << (sizeof(unsigned) * CHAR_BIT - 1));
   }
 
   Type *getType() const;