illumos gate

While looking at issues with 7092, I discovered that there was a weird difference in behavior between ftell and ftello64(). Notably, ftell() has an additional case in the non-buffered case that looks at the buffer and subtracts it from what's there. This makes sense because when you call ungetc(), even when you have a non-buffered stream, a buffer is still used. However, ftello64() is a mostly complete copy of ftell(), but lack this.

Writing a test program for this is a little tricky. A stream can be in a reading or writing mode. In this case we need to be in writing mode for this to matter. The following program (which could be reduced further) demonstrates this.

#include <stdio.h>

int
main(void)
{
        int c;

        FILE *f = fopen("./foobar", "r+");
        setvbuf(f, NULL, _IONBF, 0);
        c = fgetc(f);
        fflush(f);
        fputc('e', f);
        printf("got c %d\n", c);
        printf("%d\n", ftell(f));
        printf("%d\n", ftello64(f));
        printf("%d\n", ungetc(c, f));
        printf("%d\n", ftell(f));
        printf("%d\n", ftello64(f));
        return (0);
}
rm@turin:/ws/rm$ gcc tmp.c 
rm@turin:/ws/rm$ echo 'hello' > foobar
rm@turin:/ws/rm$ ./a.out 
got c 104
2
2
104
1
2

Note how ftell() and ftello64() disagree. This should be the same. As part of this, I'm going to look at consolidating the implementation into a single one so we have less versions of this sitting around and it's easier to modify the behavior required for 7092.