[cfe-dev] OpenCL support - using metadata

[Anton Lokhmotov]

Hi Guy,

We like the idea of a single node for each kernel.  (We tried to avoid
providing default information but I don't think we should try and save space
on this.)

> required work group size (or (0,0,0) if not defined), work group size
> hint (or (0,0,0) if not defined), vec type hint (as a string),
I think the default for vec_type_hint is "int" (6.7.2).

> kernel's signature (as a string too),
We are not sure whether this is required.  It seems this information can be
obtained directly from the kernel's value?

> and the name of a named metadata
> node, that holds a list of the kernel's locals' names.
We need to discuss a bit more __local and __constant variables.

> BTW, local variables in our solution are treated much like "static" C
> variables; they are promoted to the program scope with prefixes (as you
> might see below).
We use no metadata for this at all and Clang's conversion for static
variables.  So for:

  kernel void foo(...) { local int i; ... }

instead of:

> !opencl_foo_locals_anchor = !{!4}
> !4 = metadata !{metadata !"opencl_foo_local_i"}

we generate something like:

  @foo.i = internal addrspace(2) global i32 0, align 4

(I assume you use metadata in addition to this; otherwise, how would you be
able represent the type of individual __local variables?)

I see that with your representation one immediately gets hold of all names
for __local variables for a kernel.  Do you think it is important?

Is there any benefit in using underscores over dots? (In certain cases,
underscores could lead to a name clash.)

> According to my understanding, function scope
> constant variables are prohibited, the OpenCL 1.1 spec says at 6.5.3:
> "Variables allocated in the __constant address space can only be
> defined as program scope variables and are required to be initialized."
This statement has been removed since revision 37.  This statement is deemed
correct: "Variables inside a function or in program scope can also be
declared with the __constant address qualifier."  (Note that it doesn't say
that a function must be a kernel function.)

Best wishes,
Anton.