[PATCH] D24666: [OpenCL] Allow half type kernel argument when cl_khr_fp16 is enabled

Alexey Bader via cfe-commits cfe-commits at lists.llvm.org
Mon Sep 19 08:26:01 PDT 2016


bader added inline comments.

================
Comment at: lib/Sema/SemaDecl.cpp:7599-7602
@@ -7595,3 +7598,6 @@
     // of event_t type.
-    S.Diag(Param->getLocation(), diag::err_bad_kernel_param_type) << PT;
+    // Do not diagnose half type since it is diagnosed as invalid argument
+    // type for any function eleswhere.
+    if (!PT->isHalfType())
+      S.Diag(Param->getLocation(), diag::err_bad_kernel_param_type) << PT;
     D.setInvalidType();
----------------
yaxunl wrote:
> bader wrote:
> > It looks strange to me. First we check if parameter type is half - we set InvalidKernelParam status, later we check again and do not report an error.
> > I think it would be simpler just return ValidKernelParam for half data type from getOpenCLKernelParameterType,
> > 
> > I think the whole patch should two deleted lines from getOpenCLKernelParameterType function + test case.
> getOpenCLKernelParameterType should report half type as InvalidKernelParam since it really is an invalid kernel argument type. We do not emit diagnostic msg because the msg is redundant, not because half type is a valid kernel argument type. 
> 
> getOpenCLKernelParameterType may be used for other purpose. Reporting half type as a valid kernel argument violates the semantics of getOpenCLKernelParameterType and can cause confusion and potential error.
Maybe we should the other way.
Leave half parameter check here only and remove duplicate check that reports "declaring function parameter of type 'half' is not allowed; did you forget * ?" message.


https://reviews.llvm.org/D24666





More information about the cfe-commits mailing list