llvm dev - Oct 2013 - [LLVMdev] [CLang] Comparing vector types - invalid error and proposed fix

I was investigating an error diagnostic in the following test example:

	typedef signed char char16 __attribute__((ext_vector_type(16)));

	void test( char16 srcA, char16 srcB, char16 *dst) {
	  *dst = ( srcA == srcB );
	}

which produces the message:

	mismatch.c:5:10: error: assigning to 'char16' from incompatible type
'char __attribute__((ext_vector_type(16)))'
	    *dst = ( srcA == srcB );
	         ^ ~~~~~~~~~~~~~~~~
	1 error generated.

There appear to be two things going on here.  The first is the notion of
compatible types which are not supported by GCC/CLang for vector extensions,
even in the trivial case where 'signed char' == 'char' (in C).

The second thing is that the result of a comparison between two vectors is
supposed to result in the a 'signed' integer vector result.

'char' is always troublesome, though in C++ it is a discrete type.  I
made
the following change in my copy (clang-3.3) that I think is a correct and
valid change:

	File:		tools/clang/lib/Sema/SemaChecking.cpp
	Function:	Sema::GetSignedVectorType()

Replace the fragment:

	if (TypeSize == Context.getTypeSize(Context.CharTy))
	  return Context.getExtVectorType(Context.CharTy,
VTy->getNumElements());
	                                          ^^^^^^
with:
	if (TypeSize == Context.getTypeSize(Context.CharTy))
	  return Context.getExtVectorType(Context.SignedCharTy,
VTy->getNumElements());
	                                          ^^^^^^^^^^^^

I would like to propose this as a fix for a future release of CLang.

Thanks,

	Martin O'Riordan - Movidius Ltd.

Hi Martin,

On Oct 2, 2013, at 6:25 AM, Martin O'Riordan <Martin.ORiordan at
movidius.com> wrote:
> I was investigating an error diagnostic in the following test example:
> 
> 	typedef signed char char16 __attribute__((ext_vector_type(16)));
> 
> 	void test( char16 srcA, char16 srcB, char16 *dst) {
> 	  *dst = ( srcA == srcB );
> 	}
> 
> which produces the message:
> 
> 	mismatch.c:5:10: error: assigning to 'char16' from incompatible
type
> 'char __attribute__((ext_vector_type(16)))'
> 	    *dst = ( srcA == srcB );
> 	         ^ ~~~~~~~~~~~~~~~~
> 	1 error generated.
> 
> There appear to be two things going on here.  The first is the notion of
> compatible types which are not supported by GCC/CLang for vector
extensions,
> even in the trivial case where 'signed char' == 'char' (in
C).
The right place to discuss clang issues is cfe-dev at cs.uiuc.edu.

Having said that, in C/C++, char, signed char, and unsigned char are three
distinct types, so I don’t think your proposed fix below is correct.
> 
> The second thing is that the result of a comparison between two vectors is
> supposed to result in the a 'signed' integer vector result.
> 
> 'char' is always troublesome, though in C++ it is a discrete type. 
I made
> the following change in my copy (clang-3.3) that I think is a correct and
> valid change:
> 
> 	File:		tools/clang/lib/Sema/SemaChecking.cpp
> 	Function:	Sema::GetSignedVectorType()
> 
> Replace the fragment:
> 
> 	if (TypeSize == Context.getTypeSize(Context.CharTy))
> 	  return Context.getExtVectorType(Context.CharTy,
> VTy->getNumElements());
> 	                                          ^^^^^^
> with:
> 	if (TypeSize == Context.getTypeSize(Context.CharTy))
> 	  return Context.getExtVectorType(Context.SignedCharTy,
> VTy->getNumElements());
> 	                                          ^^^^^^^^^^^^
> 
> I would like to propose this as a fix for a future release of CLang.
> 
> Thanks,
> 
> 	Martin O'Riordan - Movidius Ltd.
> 
> 
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

Thanks Mark, hadn't spotted the CLang distribution list.

	MartinO

-----Original Message-----
From: Mark Lacey [mailto:mark.lacey at apple.com] 
Sent: 02 October 2013 22:50
To: Martin.ORiordan at movidius.com
Cc: LLVM Developers Mailing List
Subject: Re: [LLVMdev] [CLang] Comparing vector types - invalid error and
proposed fix

Hi Martin,

On Oct 2, 2013, at 6:25 AM, Martin O'Riordan <Martin.ORiordan at
movidius.com>
wrote:
> I was investigating an error diagnostic in the following test example:
> 
> 	typedef signed char char16 __attribute__((ext_vector_type(16)));
> 
> 	void test( char16 srcA, char16 srcB, char16 *dst) {
> 	  *dst = ( srcA == srcB );
> 	}
> 
> which produces the message:
> 
> 	mismatch.c:5:10: error: assigning to 'char16' from incompatible
type
> 'char __attribute__((ext_vector_type(16)))'
> 	    *dst = ( srcA == srcB );
> 	         ^ ~~~~~~~~~~~~~~~~
> 	1 error generated.
> 
> There appear to be two things going on here.  The first is the notion 
> of compatible types which are not supported by GCC/CLang for vector 
> extensions, even in the trivial case where 'signed char' ==
'char' (in C).
The right place to discuss clang issues is cfe-dev at cs.uiuc.edu.

Having said that, in C/C++, char, signed char, and unsigned char are three
distinct types, so I don't think your proposed fix below is correct.
> 
> The second thing is that the result of a comparison between two 
> vectors is supposed to result in the a 'signed' integer vector
result.
> 
> 'char' is always troublesome, though in C++ it is a discrete type. 
I
> made the following change in my copy (clang-3.3) that I think is a 
> correct and valid change:
> 
> 	File:		tools/clang/lib/Sema/SemaChecking.cpp
> 	Function:	Sema::GetSignedVectorType()
> 
> Replace the fragment:
> 
> 	if (TypeSize == Context.getTypeSize(Context.CharTy))
> 	  return Context.getExtVectorType(Context.CharTy,
> VTy->getNumElements());
> 	                                          ^^^^^^
> with:
> 	if (TypeSize == Context.getTypeSize(Context.CharTy))
> 	  return Context.getExtVectorType(Context.SignedCharTy,
> VTy->getNumElements());
> 	                                          ^^^^^^^^^^^^
> 
> I would like to propose this as a fix for a future release of CLang.
> 
> Thanks,
> 
> 	Martin O'Riordan - Movidius Ltd.
> 
> 
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev