Bug in Apple's vecLib (OS X < 10.10)

bad.c

/* proof-of-concept incorrect */

I just tried this on El Capitan (10.11.1) and the bug is still there.

Your prototype for SDOT is incorrect. The Fortran parts of Apple vecLib use the g77 calling convention, see e.g. -ff2c at GCC 9.3.0 Code-Gen-Options

So under this ABI the return type for SDOT is double.

/* bad.c */
#include <stdio.h>

extern double SDOT(
		  const int *, const float *, const int *, const float *, const int *);

int main()
{
  int len = 2;
  int stride = 1;
  float A[] = {1.0, 1.0};
  double res;

  res = SDOT(&len, A, &stride, A, &stride);
  printf("%f\n", (float)res);
  return 0;
}

> sw_vers
ProductName:	Mac OS X
ProductVersion:	10.11.6
BuildVersion:	15G31
> gcc -framework Accelerate bad.c -o bad && ./bad
2.000000

@matcross Thanks for the clarification, you're right. Out of curiosity, what led you to this gist? In any case, I will edit the gist to indicate that the code is incorrect.

Out of curiosity, what led you to this gist?

Hi, it was through Google. I was looking into some problems at NAG building our numerical libraries against vecLib. I think the search might have been 'Apple vecLib sdot' - I'd forgotten that the g77 convention did this to float-valued functions...

I've checked my notes from that time, and I was trying to debug a similar issue (with NumPy indirectly calling into vecLib).