Add lint for large const items

[petrochenkov]

const-ants are defined to be inline rvalues and this definition sometimes leads to quite non-intuitive behavior.
For example, in this code the array ARR is (correctly) copied to the stack 3 times (see the unoptimized IR/asm http://is.gd/UMQz1B):

const ARR: [u8; 100] = [33; 100];

fn main() {
    let _a = ARR[4] + ARR[7] + ARR[47];
}

The lint prevents such mistakes by checking the size of a const-ant and warning if it's larger than MAX_CONST_SIZE.

In this PR the value of MAX_CONST_SIZE is selected to be 64 bytes, so nothing in Rust itself breaks (except for one place in libstd/thread_local), but it can (should?) be smaller.
Obviously, the value should be larger than size of fat reference, because the code like

const MY_STR: &'static str = "Hello!";

shouldn't warn.
Any ideas on what should the value of MAX_CONST_SIZE ideally be?

[rust-highfive]

r? @pcwalton

(rust_highfive has picked a reviewer for you, use r? to override)

[petrochenkov]

r? @eddyb
Will the warning still be relevant when #21744 lands?

[eddyb]

I made non-constant evaluation descend into expression that have [T; N] in their interior, at that is the only expression that has a constant evaluation cost that doesn't scale linearly with source size, and it's always more efficient to initialize in-place than copy it from somewhere.
This means const ARR: [u8; 100] = [33; 100]; will behave as a non-const, when ARR is copied somewhere. Only if it's borrowed it will be placed in static memory. We could also optimize that to generate zeroinitializer in the common case of [all_zero_bytes; N] instead of using O(N) memory and time.

[bors]

☔ The latest upstream changes (presumably #22801) made this pull request unmergeable. Please resolve the merge conflicts.

[petrochenkov]

I'll gladly close this PR myself if large const items are never duplicated or copied when used
by immutable or by mutable reference, including inter-crate usage.
cc rust-lang/rfcs#885

[steveklabnik]

So, what's the status on this?

[petrochenkov]

@steveklabnik
Needs a decision from someone familiar with the implementation of constants.
rust-lang/rfcs#885 needs a decision too, because it proposes a backward-incompatible change.