Compiletest no longer respects `-Cpanic=abort` in `TARGET_RUSTCFLAGS` after #108905

[djkoloski]

Prior to #108905, compiletest would take TARGET_RUSTCFLAGS into account when determining whether a test target can unwind. It did so by calling rustc --print=cfg --target=$TARGET $TARGET_RUSTCFLAGS, which changes the reported panic strategy depending on the flags:

$ rustc --print=cfg --target=x86_64-unknown-fuchsia
...
panic="unwind"
...

$ rustc --print=cfg --target=x86_64-unknown-fuchsia -Cpanic=abort
...
panic="abort"
...

After the PR, compiletest calls rustc --print=all-target-specs-json -Zunstable-options which always reports the default panic strategy for the target (even if TARGET_RUSTCFLAGS are passed):

$ rustc --print=all-target-specs-json -Zunstable-options
...
  "x86_64-unknown-fuchsia": {
    ... (note that no "panic-strategy" is specified, it defaults to "unwind")
  }
...

$ rustc --print=all-target-specs-json -Zunstable-options -Cpanic=abort
...
  "x86_64-unknown-fuchsia": {
    ... ("panic-strategy" is still not specified)
  }
...

This causes a problem for us because we run the compiler test suite against a toolchain built with -Cpanic=abort. Compiletest will run tests marked with // needs-unwind even though they panic (and thus fail).

I see a few options for fixing this issue:

• Have --print=all-target-specs-json respect any additional flags passed.
• Parse target_rustcflags specifically for -Cpanic=abort and modify the return value of Config::can_unwind() to take it into account.
• Add some new configuration flags to compiletest to control panic strategy for the selected target.

cc @pietroalbini who authored the original PR

[jyn514]

Have --print=all-target-specs-json respect any additional flags passed.

This seems like the best solution to me.

[apiraino]

WG-prioritization assigning priority (Zulip discussion).

@rustbot label -I-prioritize +P-high

[emilyalbini]

Agree with jyn here on the solution. Unfortunately I won't have much time myself in the coming weeks to address this :(