Fix attribute test case.

@Anastasia is this test added OK?

Add attribute test.

I found that there is a newly added patch about attribute that will fail the test, and I am working on it.
https://reviews.llvm.org/D30009
https://reviews.llvm.org/D32176

Add duplicate warning test case.

Fix some varibale up case problems.

Refine some comments.

Fix some minor issues from bader.

Sorry for the late reply, I was busy with some backend problem.
LGTM, thanks for the fix!

I was on a vecation. LGTM, thanks!

Make all tests in OpenCL 2.0.

Yes, I found although the khronos have make a clarify with implicit declarations but they sames to be useful with some program link. And I found some test case about link could not pass by this reason.
But how should we handle the new added spec?

LGTM!
Thanks!

Sorry about late reply.
LGTM!

LGTM!
I think we may add optimization on this base later.

LGTM

You can update the SUMMARY of this diff as we are now using generic codegen output.
And there are some inline comments about spec.

The pointer type seems to be not that important, we can just cast it to any type we want.

LGTM, thanks!

In D18369#422367, @yaxunl wrote:

In D18369#421963, @pxli168 wrote:

If we want to save some space, could we use some macro to expand the gentype or some script to expand the gentype into each types when the clang is build?

We need to balance between space and readability. When I absorbed attribute((overloadable)) into __const_func to save space, it did not sacrifice readability. However using macro with gentype will cause the header file more difficult to read.

If we want to save some space, could we use some macro to expand the gentype or some script to expand the gentype into each types when the clang is build?

Could we output a generic function in CodeGen?
This seems to have no big difference to have a lot of declaration in an opencl c header file.

Add test and fix some old test.

I will add a test and send a new version.

LGTM.

I think merge them into one file is a good idea.
And the layout by sepc order is OK, is will makes the review easy.

LGTM!
Thanks!

Hi Anastasia,
https://cvs.khronos.org/bugzilla/show_bug.cgi?id=15603
Now we got clarify from khronos. I will continue on this patch.
BTW, https://cvs.khronos.org/bugzilla/show_bug.cgi?id=15599 is also fixed in the same revision.
Thanks
Xiuli

Add test for correct case.

Add test for spir64.

Refine other check using the changed diag.
Change test as well.

The logic is still to complex and I hope it can be optimized.

LGTM!
It fixed the problem that our project could not built with llvm 3.9.
Hope this fix can merge in soon.
Thanks.

I think this is just a small change to help identify the which standard llvm-ir is using.

https://cvs.khronos.org/bugzilla/show_bug.cgi?id=15603
Bug reported.

foo((event_t)0);

Is above use have been hint by some test cases?

1. Make new indent and leave space for the incoming OpenCL C++.
2. Check for the index to see if they are integers.

Remove test case for access quilifier in test/SemaOpenCL/invalid-kernel-attrs.cl.
Due to the patch http://reviews.llvm.org/D17437.
read_only can only be used in parameters with pipe and image type.

It now gives a warning for both C99 and OpenCL.
But for target SPIR it gives an err only because it treat unprototyped function as varidic function:

/// \brief Checks whether the given calling convention supports variadic
/// calls. Unprototyped calls also use the variadic call rules.
inline bool sthiupportsVariadicCall(CallingConv CC) {...}

and in OpenCL s6.9.e it says

Add array check for types.

It seems this patch is useless.
The spec does not tell about implicit declaration of function, but now clang with -triple spir will output err if there is implicit declaration of function.
I have read about spir and opencl spec but could not find anything talk about that.
If this is a clang bug then this patch is useless.

Then there will be no problem.
I will change my patches as well.
LGTM!
Thanks.

Refine comments and diag.

Block is an OpenCL v2.0 feature, I think all test should be handled only for CL2.0 or newer version.

It seems this patch will check block for CL1.2 or eailer? But the spec reference is for OpenCL v2.0.

LGTM!
Thanks

Refine the pipe parse to solve the problem the attribute for pipe will be handled twice within Declarator

It seems it is related with the pipe type. I am still working on how to fix this problem.

Hi Anastasia,
It seems the access qualifier attribute may need to be handled like the you have mentioned in clang-commit that to be in type.
What is you opinion? I can write a patch for that too.

LGTM!
Thanks

Sorry for the late reply, I was on a break.
LGTM!

Need to refine access qualifier with pipe type.

Update doc for OpenCL access qualifier, still find where to handle access qualifier for pipe type.

Rebase for partition