[llvm-commits] [llvm] r117070 - /llvm/trunk/lib/Analysis/TypeBasedAliasAnalysis.cpp

[Duncan Sands]

Hi John,

>>>>>   // If the third field is present, it's an integer which if equal to 1
>>>>> -// indicates that the type is "constant" (meaning
>>>>> -// pointsToConstantMemory should return true; see
>>>>> +// indicates that the type is "constant" (meaning pointsToConstantMemory
>>>>> +// should return true; see
>>>>>   // http://llvm.org/docs/AliasAnalysis.html#OtherItfs).
>>>>
>>>> is this "constant" field actually useful?  How are you imagining that it will
>>>> be used?  Even in Ada, which has a pretty strong type system and notions of
>>>> constantness, the fact that you have a view of an object via a constant type
>>>> doesn't mean that the object can't be altered, it just means it can't be
>>>> altered via this view.  In C, this is something like saying: suppose the
>>>> address of an int is held in a "const int*", then you can't alter the int via
>>>> this pointer; but you can always alter it by putting the address of the int in
>>>> a "int *" and using that instead.  So the fact that you have a "const int*"
>>>> doesn't mean you that it is valid for pointsToConstantMemory to return true.
>>>
>>> Both C and C++ say that modifying an object defined with a const qualifier
>>> has undefined results, so while you can't use this for pointer-to-const, you
>>> could use it for const variables after they've been initialized.
>>
>> in the case of a cont variable, is the constness part of the type?
>
> Yes;  I was talking about things like 'const A a = foo();' where we can't just constant-propagate values from 'a' (because they're not actually constants) and we can't necessarily mem2reg/SROA them (because the address might be taken) but we can say that they're unchanging for all loads (of non-'mutable' fields) after initialization.  But Dan's argument — that TBAA is going to use this bit to feed an AA result saying that the pointer is globally constant — indicates that this won't work.

thanks for the explanation.  The llvm.invariant intrinsics seem better adapted
to the case you describe.  Since the TBAA "constant" field seems to be fairly
useless, maybe it should just be dropped - though perhaps it's too early to say.

Ciao,

Duncan.