[llvm] [Support] report_fatal_error: Do not generate crash backtrace by default (PR #128495)

[David Blaikie via llvm-commits]

dwblaikie wrote:

> For cases which are never supposed to be user facing, but are compiler developer facing, report_fatal_error is the correct tool. In particular, the invalid MIR input case.

This I don't agree with - "never supposed to be user facing, but are compiler developer facing" - where, by my definition, a user is anyone using the binaries we release in any way we document (&, ideally, even in ways we don't document). If the state is intended to be reachable via one of these tools, it should have a diagnostic (ideally not a report_fatal, because we are a library and so we would like a different tool (eg: GUI app) that works like the tools we release, but has a different mechanism for diagnostics, should be possible).

If it's only reachable via an invalid state - assert.

report_fatal_error, in practice, gets used for things that should be library-safe diagnostics, but aren't "worth it"/we haven't done the work to get it up to the platonic ideal.

> I hand write MIR all day and expect to see a clean error.

This I don't follow - MIR is user-facing (as a user of an LLVM release, I can write MIR into a file, then feed it into a tool provided by the LLVM release) - its errors are part of the product's interface we produce and should be usable in a library, etc. (as above: someone should be able to write a GUI app that works like our command line apps without having to use process isolation to prevent LLVM from taking down the process when it tries to parse something invalid, etc)

https://github.com/llvm/llvm-project/pull/128495