-(-0.0) raises const_err "attempt to negate with overflow"

[sinkuu]

fn main() {
    let _ = -(-0.0);
}

error: this expression will panic at runtime
 --> test.rs:2:13
  |
2 |     let _ = -(-0.0);
  |             ^^^^^^^ attempt to negate with overflow
  |
  = note: `#[deny(const_err)]` on by default

error: aborting due to previous error

Meta

rustc 1.39.0-nightly (fba38ac27 2019-08-31)
binary: rustc
commit-hash: fba38ac27e2ade309f4c2504a6d6cd3556972a28
commit-date: 2019-08-31
host: x86_64-unknown-linux-gnu
release: 1.39.0-nightly
LLVM version: 9.0

[Centril]

cc @oli-obk

[oli-obk]

Huh...

const X: f32 = -(-0.0);
const Y: f64 = -(-0.0);

compiles fine.

rust/src/librustc_mir/transform/const_prop.rs

Line 414 in d0677b9

if prim.to_bits()? == (1 << (prim.layout.size.bits() - 1)) {

is the cause. It was created just for signed integers, and I believe we don't need this anymore. We are now emitting the warning on the TerminatorKind::Assert that precedes it (actually causing duplicate warnings: https://play.rust-lang.org/?version=stable&mode=debug&edition=2018&gist=432142e19caa14135dfa261f9549a2b7)

So in order to fix this (and deduplicate a bunch of warnings in our test suite), remove

rust/src/librustc_mir/transform/const_prop.rs

Lines 410 to 421 in d0677b9

	match op {
	UnOp::Neg => {
	// Need to do overflow check here: For actual CTFE, MIR
	// generation emits code that does this before calling the op.
	if prim.to_bits()? == (1 << (prim.layout.size.bits() - 1)) {
	throw_panic!(OverflowNeg)
	}
	}
	UnOp::Not => {
	// Cannot overflow
	}
	}

and rerun all ui tests with --bless