[PATCH] Use properly aligned pointers for empty/tombstone DenseMap keys.

[David Blaikie]

On Wed, Aug 27, 2014 at 4:46 PM, Alexey Samsonov <vonosmas at gmail.com> wrote:

>
> On Wed, Aug 27, 2014 at 4:15 PM, David Blaikie <dblaikie at gmail.com> wrote:
>
>>
>>
>>
>> On Wed, Aug 27, 2014 at 4:09 PM, Alexey Samsonov <vonosmas at gmail.com>
>> wrote:
>>
>>>
>>> On Tue, Aug 19, 2014 at 4:20 PM, David Blaikie <dblaikie at gmail.com>
>>> wrote:
>>>
>>>> On Tue, Aug 19, 2014 at 2:59 PM, Alexey Samsonov <vonosmas at gmail.com>
>>>> wrote:
>>>> > PointerLikeTypeTraits::NumLowBitsAvailable have different values for
>>>> > different kinds of pointers (e.g. certain pointers can be cache-line
>>>> > aligned). Do you suggest to bump the default value of
>>>> NumLowBitsAvailable
>>>> > for generic T* to return the natural alignment of T*?
>>>>
>>>> I /think/ that'd be a reasonable thing to do. If anyone cares about
>>>> the number of low bits available, presumably they care even if those
>>>> bits are just provided by the natural alignment.
>>>>
>>>
>>> OK, I was confused =/ PointerLikeTypeTraits<T*>::NumLowBitsAvailable
>>> provides information
>>> about the contents of T*, so what actually matters is alignOf<T>, not
>>> alignOf<T*>. However,
>>> alignOf<T> requires that T is defined, which is probably a restriction
>>> we don't want to impose.
>>>
>>
>> Do we ever provide PointerLikeTypeTraits<T*> for T somewhere other than
>> where T is defined? That seems subtle/easily surprising.
>>
>
> Well, yes, you can declare a class member of type DenseMap<T*, Foo> if T
> is forward-declared (which is convenient).
>

(wait, did you mean DenseMap<T, Foo>? Because DenseMap<T*, Foo> needs T* to
be defined, and T* is defined even if T is only declared)

Sure - but I mean we provide specializations of PointerLikeTypeTraits<T*>
separate from the definition of T? For the default PointerLikeTypeTraits
(presumably with a NumLowBitsAvailable of zero) would allow the above. But
it seems a tricksy to have unrelated bits of code declaring that pointers
to certain objects have a certain non-zero number of low bits available.

For example PointerLikeTypeTraits<Value*> is in Value.h, Use** is in Use.h,
etc. So it wouldn't be bad for those to also use alignof<T>.

Hmm, maybe I've lost the plot somewhere along here...



>
>> I suppose there's the case where PointerLikeTypeTraits<T*> is the default
>> implementation - and in that case we don't want to impose the limitation
>> that T is defined? Is there any way we can do that conditionally?
>>
>> - David
>>
>>
>>>
>>>
>>>>
>>>> - David
>>>>
>>>> >
>>>> >
>>>> > On Tue, Aug 19, 2014 at 2:09 PM, David Blaikie <dblaikie at gmail.com>
>>>> wrote:
>>>> >>
>>>> >> On Tue, Aug 19, 2014 at 1:30 PM, Alexey Samsonov <vonosmas at gmail.com
>>>> >
>>>> >> wrote:
>>>> >> > Hi rsmith, chandlerc,
>>>> >> >
>>>> >> > We should respect alignOf<T*>() when creating empty/tombstone keys
>>>> >> > for DenseMap<T*, foo>. Otherwise this can lead to casts on
>>>> misaligned
>>>> >> > addresses, which is undefined behavior, as reported by UBSan.
>>>> >> >
>>>> >> > http://reviews.llvm.org/D4976
>>>> >> >
>>>> >> > Files:
>>>> >> >   include/llvm/ADT/DenseMapInfo.h
>>>> >> >
>>>> >> > Index: include/llvm/ADT/DenseMapInfo.h
>>>> >> > ===================================================================
>>>> >> > --- include/llvm/ADT/DenseMapInfo.h
>>>> >> > +++ include/llvm/ADT/DenseMapInfo.h
>>>> >> > @@ -14,6 +14,7 @@
>>>> >> >  #ifndef LLVM_ADT_DENSEMAPINFO_H
>>>> >> >  #define LLVM_ADT_DENSEMAPINFO_H
>>>> >> >
>>>> >> > +#include "llvm/Support/AlignOf.h"
>>>> >> >  #include "llvm/Support/PointerLikeTypeTraits.h"
>>>> >> >  #include "llvm/Support/type_traits.h"
>>>> >> >
>>>> >> > @@ -32,12 +33,20 @@
>>>> >> >  struct DenseMapInfo<T*> {
>>>> >> >    static inline T* getEmptyKey() {
>>>> >> >      uintptr_t Val = static_cast<uintptr_t>(-1);
>>>> >> > -    Val <<= PointerLikeTypeTraits<T*>::NumLowBitsAvailable;
>>>> >> > +    const unsigned Alignment =
>>>> >> > +        PointerLikeTypeTraits<T*>::NumLowBitsAvailable >
>>>> alignOf<T*>()
>>>> >> > +            ?
>>>> (unsigned)PointerLikeTypeTraits<T*>::NumLowBitsAvailable
>>>> >> > +            : alignOf<T*>();
>>>> >>
>>>> >> Should this be built into PointerLikeTypeTraits instead?
>>>> >>
>>>> >> > +    Val <<= Alignment;
>>>> >> >      return reinterpret_cast<T*>(Val);
>>>> >> >    }
>>>> >> >    static inline T* getTombstoneKey() {
>>>> >> >      uintptr_t Val = static_cast<uintptr_t>(-2);
>>>> >> > -    Val <<= PointerLikeTypeTraits<T*>::NumLowBitsAvailable;
>>>> >> > +    const unsigned Alignment =
>>>> >> > +        PointerLikeTypeTraits<T*>::NumLowBitsAvailable >
>>>> alignOf<T*>()
>>>> >> > +            ?
>>>> (unsigned)PointerLikeTypeTraits<T*>::NumLowBitsAvailable
>>>> >> > +            : alignOf<T*>();
>>>> >> > +    Val <<= Alignment;
>>>> >> >      return reinterpret_cast<T*>(Val);
>>>> >> >    }
>>>> >> >    static unsigned getHashValue(const T *PtrVal) {
>>>> >> >
>>>> >> > _______________________________________________
>>>> >> > llvm-commits mailing list
>>>> >> > llvm-commits at cs.uiuc.edu
>>>> >> > http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>>> >> >
>>>> >
>>>> >
>>>> >
>>>> >
>>>> > --
>>>> > Alexey Samsonov
>>>> > vonosmas at gmail.com
>>>>
>>>
>>>
>>>
>>> --
>>> Alexey Samsonov
>>> vonosmas at gmail.com
>>>
>>
>>
>
>
> --
> Alexey Samsonov
> vonosmas at gmail.com
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140827/2c1881da/attachment.html>