[LLVMdev] Casting between address spaces and address space semantics
Mon P Wang
wangmp at apple.com
Mon Aug 11 19:35:01 PDT 2008
On Aug 11, 2008, at 4:09 AM, Matthijs Kooijman wrote:
>> I don't have a problem having another class, TargetAddrSpace, to
>> store this
>> information. However, I don't think it make sense being a
>> standalone pass.
>> Address spaces seems to part of the TargetData and it seems more
>> to ask the TargetData to return the TargetAddrSpace object
>> (similar to
>> struct layout) to describe the relationships between address spaces.
> This is pretty much what I did in my first patch, but Chris didn't
> like it.
> Currently, the TargetData class can be completely described using
> it's string
> representation (which is also stored inside a module). Adding
> address space
> information to target data breaks this, unless we also make a string
> representation of the address spaces that we can put in a module.
> I think that having a seperate class is somewhat cleaner, since it
> also makes
> sure that this info is only made available to the passes that
> actually use it.
If I remember correctly, I was also not fond of passing another
TargetAddrSpace reference to the TargetData object. I was hoping that
we could encode the information as a target description string like we
do for ABI information. I just don't want to end up with too many
objects that describe the machine. One can argue that we shouldn't
pollute the TargetData since it describes the ABI alignment, size,
and object layoutbut I feel that this data fits naturally there. If
you and other people feel it is cleaner with a separate pass, I'm fine
I want to treat my next point with some delicacy as I don't want to
start a religious war. I just want to get clarification from the
community on the use of multiple inheritance for the case of Phases
like AllDisjointAddrspaces. From what I can gather, the use of
multiple inheritance is to separate the interface (TargetAddrSpace) to
access data from the interface of the phase (ImmutablePhase). In this
case, will we ever create a concrete class from TargetAddrSpace that
doesn't also derive from ImmutablePass? If not, I don't think is worth
using multiple inheritance in this case.
>> [Deleted text]
>>> The last part of this patch is an addition to InstCombine to make
>>> use of
>>> this information: It removes any bitcasts from a subset to a
>>> address space. It gets at the address space information by
>>> requiring the
>>> TargetAddrspaces analysis, which will give it the default
>>> implementation in
>>> all current tools.
>> For the case of a GetElementPointer, we are replacing a bitcast to a
>> pointer, getelem with a getelem bitcast. The assumption is the
>> bitcast will hopefully go away when we iterate through those uses.
> Uh? Is this a comment about what the current code or my patch does,
> or what it
> should do? I don't understand what you mean here.
My comment was more on what I thought the patch did and I wanted to
confirm that it will cleanup newly generated bit cast that are created.
>>> [Deleted Text]
>> What I'm suggesting is that Alias Analysis can be a client to where
>> store address space information. In the example you gave, alias
>> will examine two memory locations, ask the TargetAddressSpace what
>> relationship is and if it is disjoint, it will return no alias. If
>> address spaces are in subset relationship, the alias analysis
>> returns maybe
>> unless it has more information. If a client doesn't tell the
>> compiler the
>> correct address space information, the client shouldn't expect
>> answers from coming out of the compiler.
> True, anyone actually using address space should make sure that this
> info is
> correct anyway. So, no need for an unknown default?
That is my feeling.
>>> Lastly, I'm still not so sure if InstCombine is the right place
>>> for this
>>> simplification. This needs some more thought, but currently it is
>>> a problem
>>> that instcombine does not process BitCastConstantExprs. I might
>>> end up
>>> writing a seperate pass for just this.
>> I'm not sure either. At some level, what we want is to propagate
>> the most
>> precise address space (or restrict) information to its use.
>> This means that ideally we would want to be able to handle copies
>> of the
>> value stored in some temporary and track it all the way through to
>> it use.
>> InstCombine will not handle this case, e.g, address space 1 is a
>> subset of 2
>> int<1>* ptr = ...
>> int<2>* ptr2 = ptr1+4
>> *ptr2 = ...
> Won't this code produce a bitcast in the IR, which can be
> propagated? My
> current patch doesn't do this, but it should be easy to extend it to
> propagate a bitcast past pointer arithmetic.
> Ie, it should change
> %tmp = bitcast i32 addrspace(1)* %ptr1 to i32 addrspace(2)*
> %ptr2 = add i32 addrspace(2)* %tmp, 4
> store i32 0 i32 addrspace(2)* %ptr2
> %tmp = add i32 addrspace(1)* %ptr, 4
> %ptr2 = bitcast %tmp to i32 addrspace(2)
> store i32 0 i32 addrspace(2)* %ptr2
> and then to
> %ptr2 = add i32 addrspace(1)* %ptr, 4
> store i32 0 i32 addrspace(1)* %ptr2
> Which shouldn't be too hard?
> Coming to think of it, the above won't even use the add instruction,
> but will
> probably use a GEP to do the +4. The current code already propagates
> past GEP instructions. Also, I suspect that our C frontends will
> produce the second version of the IR for the C code you give.
> Or am I totally missing the point you are making here?
No, you got my point even though my example is not a good one. If the
address calculation was using a variable, I don't think we can fold it
into the GEP and we might lose this information. The point I was
trying to make is that the information needs to be propagated through
any address calculation when possible.
-- Mon Ping
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-dev