type_name depends on -Zverbose flag

[tmiasko]

#![feature(adt_const_params)]
#![feature(const_type_name)]
#![allow(incomplete_features)]

pub trait True {}

pub struct Equal<const A: &'static str, const B: &'static str>;

impl<const A: &'static str> True for Equal<{A}, {A}> {}

pub fn assert<T: True>(_: T) {}

fn main() {
    assert(Equal::<{std::any::type_name::<[u32; 0]>()}, "[u32; 0]">);
}

$ rustc t.rs
$ rustc t.rs -Zverbose
error[E0277]: the trait bound `Equal<Const(Value(Slice { data: Allocation { bytes: [91, 117, 51, 50, 59, 32, 67, 111, 110, 115, 116, 32, 123, 32, 116, 121, 58, 32, 117, 115, 105, 122, 101, 44, 32, 118, 97, 108, 58, 32, 86, 97, 108, 117, 101, 40, 83, 99, 97, 108, 97, 114, 40, 48, 120, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 41, 41, 32, 125, 93], relocations: Relocations(SortedMap { data: [] }), init_mask: InitMask { blocks: [18446744073709551615, 3], len: Size { raw: 66 } }, align: Align { pow2: 0 }, mutability: Not, extra: () }, start: 0, end: 66 }): &ReStatic str), Const(Value(Slice { data: Allocation { bytes: [91, 117, 51, 50, 59, 32, 48, 93], relocations: Relocations(SortedMap { data: [] }), init_mask: InitMask { blocks: [255], len: Size { raw: 8 } }, align: Align { pow2: 0 }, mutability: Not, extra: () }, start: 0, end: 8 }): &ReStatic str)>: True` is not satisfied
  --> t.rs:14:12
   |
14 |     assert(Equal::<{std::any::type_name::<[u32; 0]>()}, "[u32; 0]">);
   |     ------ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ the trait `True` is not implemented for `Equal<Const(Value(Slice { data: Allocation { bytes: [91, 117, 51, 50, 59, 32, 67, 111, 110, 115, 116, 32, 123, 32, 116, 121, 58, 32, 117, 115, 105, 122, 101, 44, 32, 118, 97, 108, 58, 32, 86, 97, 108, 117, 101, 40, 83, 99, 97, 108, 97, 114, 40, 48, 120, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 41, 41, 32, 125, 93], relocations: Relocations(SortedMap { data: [] }), init_mask: InitMask { blocks: [18446744073709551615, 3], len: Size { raw: 66 } }, align: Align { pow2: 0 }, mutability: Not, extra: () }, start: 0, end: 66 }): &ReStatic str), Const(Value(Slice { data: Allocation { bytes: [91, 117, 51, 50, 59, 32, 48, 93], relocations: Relocations(SortedMap { data: [] }), init_mask: InitMask { blocks: [255], len: Size { raw: 8 } }, align: Align { pow2: 0 }, mutability: Not, extra: () }, start: 0, end: 8 }): &ReStatic str)>`
   |     |
   |     required by a bound introduced by this call
   |
   = help: the following implementations were found:
             <Equal<Const(Param(A/#0): &ReStatic str), Const(Param(A/#0): &ReStatic str)> as True>
note: required by a bound in `assert`
  --> t.rs:11:18
   |
11 | pub fn assert<T: True>(_: T) {}
   |                  ^^^^ required by this bound in `assert`

error: aborting due to previous error

cc #63084

[RalfJung]

Cc @rust-lang/wg-const-eval strange type-level equality quirk?

[tmiasko]

pretty_print_const used by type_name implementation depends on -Zverbose:

rust/compiler/rustc_middle/src/ty/print/pretty.rs

Lines 1152 to 1162 in c1aa854

	fn pretty_print_const(
	mut self,
	ct: ty::Const<'tcx>,
	print_ty: bool,
	) -> Result<Self::Const, Self::Error> {
	define_scoped_cx!(self);
	if self.tcx().sess.verbose() {
	p!(write("Const({:?}: {:?})", ct.val(), ct.ty()));
	return Ok(self);
	}

[oli-obk]

That's not really a const eval issue. This likely also happens for any lifetimes or impl trait types or anything else affected by the verbose flag.

Not sure we want to fix this, relying on a specific name is not something that should be done afaik. You should only use type names for printing to users

[tmiasko]

-Zverbose is an untracked option:

• Producing different query output based on its value is incompatible with incremental compilation.
• Its value could vary between crates which makes type system inconsistent.

[lcnr]

jup, should fix that 😅 type_name should always not be verbose i think

[oli-obk]

As a first step we should probably just remove the verbose flag check from rendering constants. That looks just useless even for compiler debugging.

But to fix this properly we'll likely need to create a way to tell the pretty printer whether to use verbose mode or not, and entirely avoid using verbosity checks within it.