Avoid ICE when global_asm const operand fails to evaluate

[BeetleFunk]

Fixes #1466

Follows the same approach used in rust-lang/rust#122691. The error is reported when calling const_eval_poly(). codegen_global_asm_item() doesn't have a way to propagate the error but can detect the ErrorHandled::Reported case and avoid the panic.

[BeetleFunk]

With these changes, compiling the test case file:

//@ build-fail
//@ needs-asm-support
#![feature(asm_const)]

use std::arch::global_asm;

fn main() {}

global_asm!("/* {} */", const 1 << 500); //~ ERROR evaluation of constant value failed [E0080]

global_asm!("/* {} */", const 1 / 0); //~ ERROR evaluation of constant value failed [E0080]

Results in STDERR output:

error[E0080]: evaluation of constant value failed
 --> src/main.rs:9:31
  |
9 | global_asm!("/* {} */", const 1 << 500); //~ ERROR evaluation of constant value failed [E0080]
  |                               ^^^^^^^^ attempt to shift left by `500_i32`, which would overflow

error[E0080]: evaluation of constant value failed
  --> src/main.rs:11:31
   |
11 | global_asm!("/* {} */", const 1 / 0); //~ ERROR evaluation of constant value failed [E0080]
   |                               ^^^^^ attempt to divide `1_i32` by zero

For more information about this error, try `rustc --explain E0080`.
error: could not compile `sandbox` (bin "sandbox") due to 2 previous errors

[BeetleFunk]

What's standard practice on automated tests for compiler errors like this?

I noticed that the test runner in scripts/test_rustc_tests.sh seems to be filtering out most of the rustc test files containing "build-fail" or "ERR" patterns.

EDIT: I see this is being tracked in issue #381. Makes sense!

[bjorn3]

Thanks!