[PATCH] OpenCL: Add new types for OpenCL 2.0
sam.parker at arm.com
Thu Jun 11 02:01:23 PDT 2015
I had noticed that event_t is also handled this way, but it still seems odd to me to declare something as a pointer type when it is not a pointer (sorry if this is what is blinding me). But as OpenCL is gaining popularity, and clang being a popular driver, I wonder if it would be a good idea to encode the size and alignment of OpenCL types in the target datalayout?
From: Pedro Ferreira [mailto:arkangath at gmail.com]
Sent: 11 June 2015 09:01
To: Sam Parker; cfe-commits at cs.uiuc.edu
Subject: Re: [PATCH] OpenCL: Add new types for OpenCL 2.0
I don't understand what you're trying to say.
You quoted the standard which says that the "size of those types is implementation defined"; you then infer that this is the reason why it can /not/ make it a pointer to opaque struct. I would think the other way around: it's because the size is implementation defined that we can use whatever representation we wish.
Further to that, this is how event_t is already defined (I didn't add it).
OpenCL implementations will likely either patch clang to do this or add a typedef in the implicit header which will end up generating the same llvm intermediate anyway.
What these structs map to in final HW code is indeed target defined, so I suppose your comment on "hooks to get the specific types" would be accurate.
In any case, if the representation of those types is not acceptable for Clang, I'll strip them and leave only the images in.
On Thu, 11 Jun 2015 at 08:41 Sam Parker <sam.parker at arm.com> wrote:
I was looking at these types a couple of months back, what stumped me is that ndrange_t is not a pointer type as I had originally tried. You too have made it a pointer-to-a-struct type but I feel this is wrong. The spec says, ‘values returned by applying the sizeof operator to the queue_t, clk_event_t, nrange_t and reserve_id_t types is implementation defined.’ So I understand that it would not be valid to make this types as pointers and that implementations would probably just have to create definitions for them in the distributed header file; unless there were some hooks to the backend to get the specific implementation defined types.
From: cfe-commits-bounces at cs.uiuc.edu [mailto:cfe-commits-bounces at cs.uiuc.edu] On Behalf Of Pedro Ferreira
Sent: 11 June 2015 08:19
To: cfe-commits at cs.uiuc.edu
Subject: [PATCH] OpenCL: Add new types for OpenCL 2.0
This patch adds the new OpenCL types for 2.0 described at https://www.khronos.org/registry/cl/sdk/2.0/docs/man/xhtml/otherDataTypes.html <https://urldefense.proofpoint.com/v2/url?u=https-3A__www.khronos.org_registry_cl_sdk_2.0_docs_man_xhtml_otherDataTypes.html&d=AwMFaQ&c=8hUWFZcy2Z-Za5rBPlktOQ&r=BSqEv9KvKMW_Ob8SyngJ70KdZISM_ASROnREeq0cCxk&m=42YnWExwxwpeU6GPDY2_3RFxCqQakUbj_CXZsMsQ2jU&s=REOBNoaDio7qDyIDCqmXhxFvZYjMOK6vuXAttjOVsNI&e=>
I also opened https://llvm.org/bugs/show_bug.cgi?id=23794 <https://urldefense.proofpoint.com/v2/url?u=https-3A__llvm.org_bugs_show-5Fbug.cgi-3Fid-3D23794&d=AwMFaQ&c=8hUWFZcy2Z-Za5rBPlktOQ&r=BSqEv9KvKMW_Ob8SyngJ70KdZISM_ASROnREeq0cCxk&m=42YnWExwxwpeU6GPDY2_3RFxCqQakUbj_CXZsMsQ2jU&s=TAV4suAMaHgdIPA83Da3pQl7c68On7bAFWtnrUbt_Uk&e=> for this. I keep forgetting you prefer patches sent to this mailing list.
This also adds lldb entries (fixes switch warnings).
The types are:
let me know if something looks wrong,
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the cfe-commits