error output between self-hosted versus cross-compile target differs significantly for some tests

[pnkfelix]

The error output between self-hosted versus cross-compile target can differ significantly for some tests.

In some cases one can get around this via error-pattern, but it does not work universally.

This bug is to mark such cases, and hopefully go back and identify why the error output is so different (in some cases the reason may be self-evident, but that was not universally true).

[pnkfelix]

Ah this case
https://github.com/rust-lang/rust/blob/2a9de1d989d7f95846b711eec2695cbd86794ee3/src/test/compile-fail/huge-array-simple.rs
is one of the ones that seems self-evident:

• on 32-bit hosts, the shift overflows.
• But on 64-bit hosts, the shift succeeds, and then the attempt to create the huge array overflows.

Maybe the easiest thing would be to make 32- and 64-bit variants of that same test. (That or conditionalize the size we feed into the array construction based on pointer size.)

Update: another thing to maybe check is to see if trying to allocate an [u8, (1<<31)] array on the stack causes the same compile-time failure on both targets.

[steveklabnik]

Triage: no changes I'm aware of.

[Mark-Simulacrum]

There are two other tests currently tagged with this issue, const-eval-overflow-3.rs and const-eval-overflow-3b.rs, both in compile-fail. The issue with both appears to be the same as with the previous test mentioned; we either can or cannot create and evaluate the const array depending on platform.

[pnkfelix]

I wonder if we might be able to leverage the revision system in some way here. That, or as I noted earlier: make pointer-size-specific variants of each test. (And then presumably ignore the tests on platforms that cannot support the 64-bit pointer-sized variant.)

[pnkfelix]

Better solution: make target-specific versions of each fn main with a customized const expression for the target of interest, and then use // normalize-stderr-test REGEXP -> OUTPUT to normalize to a single canonical output for the test.