Ergonomic format_args!

text/0000-ergonomic-format-args.md

@@ -0,0 +1,146 @@
+- Feature Name: (not applicable)
+- Start Date: 2016-05-17
+- RFC PR: (leave this empty)
+- Rust Issue: (leave this empty)
+
+# Summary
+[summary]: #summary
+
+Removes the one-type-only restriction on `format_args!` arguments.
+Expressions like `format_args!("{0:x} {0:o}", foo)` now work as intended,
+where each argument is still evaluated only once, in order of appearance
+(i.e. left-to-right).
+
+# Motivation
+[motivation]: #motivation
+
+The `format_args!` macro and its friends historically only allowed a single
+type per argument, such that trivial format strings like `"{0:?} == {0:x}"` or
+`"rgb({r}, {g}, {b}) is #{r:02x}{g:02x}{b:02x}"` are illegal. This is
+massively inconvenient and counter-intuitive, especially considering the
+formatting syntax is borrowed from Python where such things are perfectly
+valid.
+
+Upon closer investigation, the restriction is in fact an artificial
+implementation detail. For mapping format placeholders to macro arguments the
+`format_args!` implementation did not bother to record type information for
+all the placeholders sequentially, but rather chose to remember only one type
+per argument. Also the formatting logic has not received significant attention
+since after its conception, but the uses have greatly expanded over the years,
+so the mechanism as a whole certainly needs more love.
+
+# Detailed design
+[design]: #detailed-design
+
+Formatting is done during both compile-time (expansion-time to be pedantic)
+and runtime in Rust. As we are concerned with format string parsing, not
+outputting, this RFC only touches the compile-time side of the existing
+formatting mechanism which is `libsyntax_ext` and `libfmt_macros`.
+
+Before continuing with the details, it is worth noting that the core flow of
+current Rust formatting is *mapping arguments to placeholders to format specs*.
+For clarity, we distinguish among *placeholders*, *macro arguments* and
+*argument objects*. They are all *italicized* to provide some
+visual hint for distinction.
+
+To implement the proposed design, the following changes in behavior are made:
+
+* implicit references are resolved during parse of format string;
+* named *macro arguments* are resolved into positional ones;
+* placeholder types are remembered and de-duplicated for each *macro argument*,
+* the *argument objects* are emitted with information gathered in steps above.
+
+As most of the details is best described in the code itself, we only
+illustrate some of the high-level changes below.
+
+## Implicit reference resolution
+
+Currently two forms of implicit references exist: `ArgumentNext` and
+`CountIsNextParam`. Both take a positional *macro argument* and advance the
+same internal pointer, but format is parsed before position, as shown in
+format strings like `"{foo:.*} {} {:.*}"` which is in every way equivalent to
+`"{foo:.0$} {1} {3:.2$}"`.
+
+As the rule is already known even at compile-time, and does not require the
+whole format string to be known beforehand, the resolution can happen just
+inside the parser after a *placeholder* is successfully parsed. As a natural
+consequence, both forms can be removed from the rest of the compiler,
+simplifying work later.
+
+## Named argument resolution
+
+Not seen elsewhere in Rust, named arguments in format macros are best seen as
+syntactic sugar, and we'd better actually treat them as such. Just after
+successfully parsing the *macro arguments*, we immediately rewrite every name
+to its respective position in the argument list, which again simplifies the
+process.
+
+## Processing and expansion
+
+We only have absolute positional references to *macro arguments* at this point,
+and it's straightforward to remember all unique *placeholders* encountered for
+each. The unique *placeholders* are emitted into *argument objects* in order,
+to preserve evaluation order, but no difference in behavior otherwise.
+
+# Drawbacks
+[drawbacks]: #drawbacks
+
+Due to the added data structures and processing, time and memory costs of
+compilations may slightly increase. However this is mere speculation without
+actual profiling and benchmarks. Also the ergonomical benefits alone justifies
+the additional costs.
+
+# Alternatives
+[alternatives]: #alternatives
+
+## Do nothing
+
+One can always write a little more code to simulate the proposed behavior,
+and this is what people have most likely been doing under today's constraints.
+As in:
+
+```rust
+fn main() {
+	let r = 0x66;
+	let g = 0xcc;
+	let b = 0xff;
+
+	// rgb(102, 204, 255) == #66ccff
+	// println!("rgb({r}, {g}, {b}) == #{r:02x}{g:02x}{b:02x}", r=r, g=g, b=b);
+	println!("rgb({}, {}, {}) == #{:02x}{:02x}{:02x}", r, g, b, r, g, b);
+}
+```
+
+Or slightly more verbose when side effects are in play:
+
+```rust
+fn do_something(i: &mut usize) -> usize {
+	let result = *i;
+	*i += 1;
+	result
+}
+
+fn main() {
+	let mut i = 0x1234usize;
+
+	// 0b1001000110100 0o11064 0x1234
+	// 0x1235
+	// println!("{0:#b} {0:#o} {0:#x}", do_something(&mut i));
+	// println!("{:#x}", i);
+
+	// need to consider side effects, hence a temp var
+	{
+		let r = do_something(&mut i);
+		println!("{:#b} {:#o} {:#x}", r, r, r);
+		println!("{:#x}", i);
+	}
+}
+```
+
+While the effects are the same and nothing requires modification, the
+ergonomics is simply bad and the code becomes unnecessarily convoluted.
+
+# Unresolved questions
+[unresolved]: #unresolved-questions
+
+None.