bli_?setijv accepts void* instead of floating point type

[ivan-pi]

The routines ?setijv/?getijv and ?setijm/?getijm are not strongly typed.

E.g. if you run gcc -P -C -E path/to/blis.h > blis.h.pp the preprocess header will contain the following prototypes:

__attribute__ ((visibility ("default"))) void bli_ssetijv ( double ar, double ai, dim_t i, void* x, inc_t incx ); 
__attribute__ ((visibility ("default"))) void bli_dsetijv ( double ar, double ai, dim_t i, void* x, inc_t incx ); 
__attribute__ ((visibility ("default"))) void bli_csetijv ( double ar, double ai, dim_t i, void* x, inc_t incx ); 
__attribute__ ((visibility ("default"))) void bli_zsetijv ( double ar, double ai, dim_t i, void* x, inc_t incx );

This originates from

blis/frame/base/bli_setgetijv.h

Lines 44 to 54 in 827c50b

	#define GENTPROT( ctype, ch, opname ) \
	\
	BLIS_EXPORT_BLIS void PASTEMAC(ch,opname) \
	( \
	double ar, \
	double ai, \
	dim_t i, \
	void* x, inc_t incx \
	);
	INSERT_GENTPROT_BASIC( setijv )

Could these routines be strongly typed using float/double/scomplex/dcomplex? The reason I'm asking is because the missing type breaks some scripts for generating wrappers.

[fgvanzee]

Hi @ivan-pi. To what extent do you need strongly typed function prototypes? My first reaction to this was that this is just the tip of the tip of the tip of the iceberg for use of void* pointers in lieu of actual typed pointers in BLIS. Or maybe I'm missing something about your particular application setting? (cc @devinamatthews)

[ivan-pi]

The application was creating Fortran interfaces from the C prototypes. The current script iterates over the parameter list and creates Fortran interface blocks such as:

interface
   subroutine bli_?setijv(ar,ai,i,x,incx)
      real(c_double), value :: ar, ai
      integer(dim_kind), value :: i
      real(?) :: x(*)  ! void*
      integer(inc_kind), value :: incx
   end subroutine
end interface

Due to the void* x, the interface-generation script I'm writing fails here. I understand it's a hassle and just adds more work to the existing macro machinery.

Upon further thought I think it'll be better if I just introduce a Fortran version so that it can use 1-based indexing and check bounds automatically.

[devinamatthews]

Hi @ivan-pi I'm going to leave this open since you're right, it is an inconsistency. This could also affect C++ users.

[devinamatthews]

@fgvanzee void* is great for kernels since a) they're internal and b) they're typically being called through a type-erased interface anyways. For user-level typed functions, having strongly typed arguments is definitely helpful.

[ivan-pi]

In that case I'd just note that void* also appears in the prototypes for ??copysc, ??cast[m|v], ??castnzm, ?fprint[m|v] and ?print[m|v].

[devinamatthews]

Thx