[cfe-dev] RFC: Representation of OpenCL Memory Spaces

[Peter Collingbourne]

Hi Justin,

Thanks for bringing this up, I think it's important to discuss
these issues here.

On Thu, Oct 13, 2011 at 09:46:28AM -0400, Justin Holewinski wrote:
> It is becoming increasingly clear to me that LLVM address spaces are not the
> general solution to OpenCL/CUDA memory spaces. They are a convenient hack to
> get things working in the short term, but I think a more long-term approach
> should be discussed and decided upon now before the OpenCL and CUDA
> implementations in Clang/LLVM get too mature. To be clear, I am not
> advocating that *targets* change to a different method for representing
> device memory spaces. The current use of address spaces to represent
> different types of device memory is perfectly valid, IMHO. However, this
> knowledge should not be encoded in front-ends and pre-SelectionDAG
> optimization passes.

I disagree.  The targets should expose all the address spaces they
provide, and the frontend should know about the various address spaces
it needs to know about.  It is incumbent on the frontend to deliver
a valid IR for a particular language implementation, and part of
that involves knowing about the ABI requirements for the language
implementation (which may involve using specific address spaces)
and the capabilities of each target (including the capabilities of
the target's address spaces), together with the language semantics.
It is not the job of the optimisers or backend to know the semantics
for a specific language, a specific implementation of that language
or a specific ABI.

> 
> 
> *2. Solutions*
> 
> A couple of solutions to this problem are presented here, with the hope that
> the Clang/LLVM community will offer a constructive discussion on how best to
> proceed with OpenCL/CUDA support in Clang/LLVM. The following list is in no
> way meant to be exhaustive; it merely serves as a starting basis for
> discussion.
> 
> 
> *2A. Extend TBAA*
> 
> In theory, the type-based alias analysis pass could be extended to
> (properly) support aliasing queries for pointers in OpenCL kernels.
>  Currently, it has no way of knowing if two pointers in different address
> spaces can alias, and in fact cannot know if this is the case given the
> definition of LLVM address spaces.  Instead of programming it with
> target-specific knowledge, it can be extended with language-specific
> knowledge.  Instead of considering address spaces, the Clang portion of TBAA
> can be programmed to use OpenCL attributes to extend its pointer metadata.
>  Specifically, pointers to different memory spaces are in essence different
> types and cannot alias.  For the kernel shown above, the resulting LLVM IR
> could be:
> 
> ; ModuleID = 'test1.cl'
> target datalayout = "e-p:32:32-i64:64:64-f64:64:64-n1:8:16:32:64"
> target triple = "ptx32--"
> 
> define ptx_kernel void @foo(float* nocapture %a, float addrspace(4)*
> nocapture %b) nounwind noinline {
> entry:
>   %0 = load float* %a, align 4, !tbaa !1
>   store float %0, float addrspace(4)* %b, align 4, !tbaa *!2*
>   ret void
> }
> 
> !opencl.kernels = !{!0}
> 
> !0 = metadata !{void (float*, float addrspace(4)*)* @foo}
> *!1 = metadata !{metadata !"float$__global", metadata !3}*
> *!2 = metadata !{metadata !"float$__local", metadata !3}*
> !3 = metadata !{metadata !"omnipotent char", metadata !4}
> !4 = metadata !{metadata !"Simple C/C++ TBAA", null}
> 
> Differences are bolded.  Here, the TBAA pass would be able to identify that
> the loads and stores do not alias.  Of course, when compiling in
> non-OpenCL/CUDA mode, TBAA would work just as before.

I have to say that I much prefer the TBAA solution, as it encodes the
language semantics using the existing metadata for language semantics.

> *Pros:*
> 
> Relatively easy to implement
> 
> *Cons:*
> 
> Does not solve the full problem, such as how to represent OpenCL memory
> spaces in other backends, such as X86 which uses LLVM address spaces for
> different purposes.

This presupposes that we need a way of representing OpenCL address
spaces in IR targeting X86 (and targets which lack GPU-like address
spaces).  As far as I can tell, the only real representations of
OpenCL address spaces on such targets that we need are a way of
distinguishing the different address spaces for alias analysis
and a representation for __local variables allocated on the stack.
TBAA metadata would solve the first problem, and we already have
mechanisms in the frontend that could be used to solve the second.

> I see this solution as more of a short-term hack to solve the pointer
> aliasing issue without actually addressing the larger issues.

I remain to be persuaded that there are any "larger issues" to solve.

> *2B. Emit OpenCL/CUDA-specific Metadata or Attributes*
> 
> Instead of using LLVM address spaces to represent OpenCL/CUDA memory spaces,
> language-specific annotations can be provided on types.  This can take the
> form of metadata, or additional LLVM IR attributes on types and parameters,
> such as:
> 
> ; ModuleID = 'test1.cl'
> target datalayout = "e-p:32:32-i64:64:64-f64:64:64-n1:8:16:32:64"
> target triple = "ptx32--"
> 
> define *ocl_kernel* void @foo(float* nocapture *ocl_global* %a, float*
> nocapture *ocl_local* %b) nounwind noinline {
> entry:
>   %0 = load float* %a, align 4
>   store float %0, float* %b, align 4
>   ret void
> }
> 
> Instead of extending the LLVM IR language, this information could also be
> encoded as metadata by either (1) emitting some global metadata that binds
> useful properties to globals and parameters, or (2) extending LLVM IR to
> allow attributes on parameters and globals.
> 
> Optimization passes can make use of these additional attributes to derive
> useful properties, such as %a cannot alias %b. Then, back-ends can use these
> attributes to emit proper code sequences based on the pointer attributes.
> 
> *Pros:*
> *
> *
> If done right, would solve the general problem
> 
> *Cons:*
> *
> *
> Large implementation commitment; could potentially touch many parts of LLVM.

You are being vague about what is required here.  A complete solution
following 2B would involve allowing these attributes on all pointer
types.  It would be very expensive to allow custom attributes or
metadata on pointer types, since they are used frequently in the IR,
and the common case is not to have attributes or metadata.  Also,
depending on how this is implemented, this would encode far too much
language specific information in the IR.

Thanks,
-- 
Peter