finalize printf format calling syntax

[StefanKarpinski]

See #29 (comment):

We should write a wrapper for printf that looks at the format string to check/convert arguments, then calls printf on each format specifier individually.

[JeffBezanson]

Shall we 2.0 this?

[StefanKarpinski]

No, I'm still aiming to get to this. Working on the last sections of the manual currently. After that I'm going to tackle all the printing and formatting stuff. It's hard to get any real work done without this.

[StefanKarpinski]

Concluded at JuliaCon 1.0day1: should probably implement plain old functions that implement the various printf formatting features, then play around with different ways of invoking them.

[JeffBezanson]

This is looking pretty good. What's left to do here?

[StefanKarpinski]

The parsing all works, but actual printing doesn't always work. E.g. Zero padding is horribly broken, most conversion specifiers don't actually work, etc. I'll take another pass at it soon and whip it all into shape.

[StefanKarpinski]

We still need to finalize the API to printf. That is a breaking change.

[JeffBezanson]

Next step is to benchmark the function call version, printf("%d", n), with a cache mapping format strings to generated code. If performance is ok, we should go with that. Otherwise it should be printf(f"%d", n), or maybe print(f"%d", n) to cut down on f's, requiring the first argument to be an f"" formatter.

[StefanKarpinski]

I kind of like print(f"%d", n). I've thought of that before. As a corollary, I feel like we'd also have to rename print_to_string as sprint.

[nolta]

printf"%d"(n) ?

[JeffBezanson]

I think we pretty much concluded it had to look like a normal function call or it would be too weird.

The devious thing about print(f"%d",...) is that if you forget the f it's obviously not a printf, just a normal call to print where the first argument is a string. I like it.

The only remaining case is building a format string at run time, which people sometimes do. The best thing I can think of is an alternate formatting API that doesn't require parsing format strings. Hopefully it could call bits of printf's functionality. Then we can say to use that instead of building format strings.

[JeffBezanson]

Other thought: maybe f"" should give something of a specific type other than Function.

[pao]

FormatString, maybe, containing an anonymous formatter function in a field that gets accessed by printf and sprintf?

[StefanKarpinski]

Ok, I like this plan. A FormatString object can also print nicely (like Regex objects do). The only issue this leaves is making sure that code using the print(f"", ...) and constructing an object containing a function at compile and then applying that at run time can be made fast enough. If that's not the case, I may as well rewrite the printf functionality to just parse at runtime like the C version does. I still think the code gen approach can with work on I/O be made faster than C, so I'd rather keep the code gen approach.

[StefanKarpinski]

Closed by 04f542d.