[patch] Improvement to implicit scalar-vector conversions

[Stephen Canon]

While it does move closer to the OpenCL behavior in some regards, this patch does not strictly hew to the OpenCL conversion rules of section 6.2.6.  Specifically, it does not adhere to:

> An error shall occur if any scalar operand has greater rank than the type of the vector element. 

This is not an accident; it is the experience of every vector programmer I surveyed that this rule was ill-considered (and , as an early contributor to the OpenCL spec, I’m as responsible for this mistake as anyone; I wish we had detected it back then).  I will work on getting this changed in OpenCL, but that will be a lengthy process, and we should relax it for the vector extensions in C/ObjC/C++, rather than blindly follow after OpenCL’s error.

The best example of why the OpenCL rule is a bad idea is that it makes working with vectors integer types smaller than int extremely painful.  Here’s a simple example:
	
	ushort8 foo(ushort8 x, ushort8 y) { return (x + y + 1) >> 1; }

This seems perfectly reasonable, and naively one would like it to work, but it doesn’t:

	error: can't convert between vector values of different size ('short8' and 'int')

Instead one must write:

	ushort8 foo(ushort8 x, ushort8 y) { return (x + y + (unsigned short)1) >> (unsigned short)1; }

These extraneous casts do not help readability (the intended meaning is unambiguous in the original), and do not help maintain correctness.  They are noise and they are frustrating to the programmer.

With this patch, the example above works as written, but informative warnings are also provided when the conversion would change the value, to help protect new vector programmers who are only familiar with the C scalar promotion rules:

	uchar16 bar(uchar16 x) { return x*257 >> 3; }
	warning: implicit conversion from 'int' to 'uchar16' changes value from 257 to 1 [-Wconstant-conversion]

	float4 bar(float4 x) { return x/M_PI; }
	warning: implicit conversion loses floating-point precision: 'double' to 'float4' [-Wconversion]

I realize that the merit of this change is perhaps non-obvious to those who don’t have a lot of experience writing vector code, but it relieves a major pain point that essentially everyone who has ever tried to seriously use ext_vector_type has experienced and complained about.  I am recommending that we choose to be slightly more permissive than OpenCL is to relieve this pain, while providing clear warnings to mitigate the risk of that permissiveness.

– Steve

On Apr 2, 2014, at 4:58 AM, Pekka Jääskeläinen <pekka.jaaskelainen at tut.fi> wrote:

> Hi,
> 
> Should we try to comply to the OpenCL C specified behavior here or is
> there some other relevant standard to look at in this case?
> 
> This piece from the OpenCL 1.2 specs is relevant here:
> 
> "6.2.6 Usual Arithmetic Conversions <snip>
> 
> If I understood it correctly, your patch goes towards the OpenCL C specified
> behavior. It LGTM.
> 
> BR,
> Pekka
> 
> On 03/27/2014 02:34 AM, Stephen Canon wrote:
>> ExtVectors currently support basic operations with scalar data (which is
>> interpreted as an implicit splat).  However, this support has some serious
>> issues.  Most critically, at present the type of the result depends on
>> operand order:
>> 
>> typedef float __attribute__((ext_vector_type(2))) float2;
>> 
>> float2 x; double y = 2.0 + x; // reinterprets y as double, scalar
>> double-precision add. float2 z = x + 2.0; // reinterprets x as float2, does
>> packed single-precision add.
>> 
>> Both behaviors are pretty busted; the odds are overwhelming that the
>> programmer's intention was to add two to both lanes of x.  What’s worse, +,
>> which is a commutative operator for any reasonable FP type, doesn’t even
>> return the same type when the operand order is flipped.
>> 
>> This patch makes it so that “real scalar OP vector” is interpreted as
>> “convert the scalar to vector element type and splat, then perform OP”,
>> regardless of operand order or conversion rank of the scalar and vector
>> type (i.e. the type of the vector elements always “wins”, even if the rank
>> of the scalar type is greater).  This is somewhat different from the
>> arithmetic promotions for scalar types, but it is by far the most sensible
>> behavior; it is what most vector programmers want to get.
>> 
>> This also improves the state of affairs for integer scalars in ExtVector
>> expressions.  When operating on vectors with elements smaller than int, it
>> has until now been necessary to sprinkle in lots of casts:
>> 
>> typedef unsigned char __attribute__((__ext_vector_type__(16))) uchar16;
>> uchar16 baz(uchar16 x) { return x + (unsigned char)2; }
>> 
>> The extra cast adds little to nothing, and makes simple expressions overly
>> verbose.  With this patch, the following works just fine:
>> 
>> uchar16 baz(uchar16 x) { return x + 2; }
>> 
>> I also improved the state of warnings for implicit scalar->vector casts to
>> make it easier to identify suspicious conversions:
>> 
>> short4 bar( ) { return 65536; }
>> 
>> previously this produced no error or warning.  Now, with -Wconversion we
>> get:
>> 
>> foo.c:9:12: warning: implicit conversion from 'int' to 'short4' changes
>> value from 65536 to 0 [-Wconstant-conversion]
>> 
>> Thanks in advance for your feedback! – Steve
>> 
>> 
>> 
>> _______________________________________________ cfe-commits mailing list
>> cfe-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits