Re: [PATCH 3/4] diff.*.xfuncname which uses "extended" regex's for hunk header selection

Junio C Hamano wrote:

I personally think it was a mistake that the loop returns failure when
there is any pattern that does not match (it is a useful feature that a
match with a negated one forces an early return, though).

We may want to fix the semantics to be something like:

	for (i = 0; i < regs->nr; i++) {
        	struct ff_reg *reg = &regs->array[i];
                if (!regexec(&reg->re, line_buffer, 2, pmatch, 0)) {
			if (reg->negate) {
	                	free(line_buffer);
	                	return -1;
			}
                        break;
		}
	}
        if (regs->nr <= i) {
        	free(line_buffer);
		return -1;
	}
	... use pmatch() ...

I.e. (1) negative match forces an early return (useful for catching
language keywords), (2) first positive match is used, and (3) no match is
a failure.

Of course, by definition the above "fix" changes the semantics, and will
break people's existing setup if somebody has an existing custom pattern
string that does use more than one positive regexp anded together with
"\n", but I somehow suspect nobody sane depends on the current broken
semantics.

It would help making JdS's ObjC alternates easier to write.  You can say:


	/* A or B */
	"...(A|B)$"
        "\n" /* or C or D */
        "...(C|D)$"

and both captures around A|B and C|D would be saved in $1.

Just to let you know another pair of eyes has given the above a
once over, I agree with your proposal. "or" semantics makes more sense
than "and".

-brandon