llvm dev - Jul 2004 - [LLVMdev] ConstantPointerRef Changes

LLVMers,

The long awaited conclusion of bug 122 (VMCore cleanups) is very near. I 
have just committed all the changes to rid LLVM of the ConstantPointerRef 
abomination. This class no longer exists. Just use GlobalValue, 
GlobalVariable or Function directly.

The changes pass 100% of Feature and Regression tests.

Two minor discrepancies against the Programs tests were detected but these 
were attributed to either the test program (changes committed) or the test 
machine.

What hasn't been tested: SparcV9 BE (thanks, Brian! :)

Hopefully you won't get any grief from these changes, but if you do, please 
contact either Chris or I.

I must also give at least 50% credit for these changes to Chris Lattner. 
Without his almost constant expertise at hand (on the IRC channel), I would 
not have been able to complete these changes this past week.

Best Regards,

Reid.

On Sat, 17 Jul 2004, Reid Spencer wrote:
> The long awaited conclusion of bug 122 (VMCore cleanups) is very near. I
> have just committed all the changes to rid LLVM of the ConstantPointerRef
> abomination. This class no longer exists. Just use GlobalValue,
> GlobalVariable or Function directly.
Actually, PR122 is now closed, which is a HUGE deal.  If you're
interested, I suggest taking a look at some of the stuff in the bug trail
to see what Reid went through: http://llvm.cs.uiuc.edu/PR122   :)
> The changes pass 100% of Feature and Regression tests.
>
> Two minor discrepancies against the Programs tests were detected but these
> were attributed to either the test program (changes committed) or the test
> machine.
All tests pass on X86.  By tommorow we should get nightly tester data for
other platforms, but I don't anticipate major trouble.

What Reid didn't do is explain what this change DOES. :)  Here's what it
did and what you need to know if you are maintaining code outside of CVS:

1. The GlobalValue class now derives from the Constant class.  This means
   the two classes that derive from GlobalValue (Function & GlobalVariable)
   are now constants.
2. Because of #1, isa/dyn_cast/etc functions on a global value will return
   true for isa<Constant>.  This can break your code, so you should be
   aware of it.
3. Because of #1, the ConstantPointerRef class is now obsolete and is
   gone.  ConstantPointerRef used to be a bridge between the GlobalValue
   class and the Constant classes (which allowed you to treat a GV as a
   constant).  Because global values ARE constants, you can just use the
   GV directly now.  If you use the ConstantPointerRef class in your code,
   this will break your code, just use the GlobalValue where you used CPR.

For clarity, I will point out what it means for global values to be
constants.  :)  In particular, the "value" of a global value is really
the
address of the mentioned object.  For example, if you have a program like
this:

%G = global int 4
...
   load int* %G
...

The operand to the load *is* the GlobalVariable for %G.

Because the address of a global is a link-time constant, it makes sense
for the GlobalValue class to derive from Constant.

Anyways, if you have any questions or if your code is broken somehow,
please let us know and we'll try to help.  Also, HUGE thanks to Reid for
taking charge and getting this bug fixed.  This is a major cleanup for
LLVM and allows the system make more sense.

-Chris

-- 
http://llvm.cs.uiuc.edu/
http://nondot.org/sabre/