[llvm] [llvm-mca] Abort on parse error without -skip-unsupported-instructions (PR #90474)

[Michael Maitland via llvm-commits]

michaelmaitland wrote:

> I can tell you that it was not intentional. In case of parse errors, it leads to wrong/misleading perf reports. I may be wrong, but I don't think that this is what most users want. In general, I don't think that people should rely on this.

+1

> The analysis outputs may be compromised to some degree for bad inputs, but so long as the user is being explicit about allowing it and being warned that the analysis will be less accurate, I don't see the harm in supporting this case.

+1

I am still concerned about the case where the program contains an instruction that is absolute garbage:

> like passing not enough registers, or incorrect types of registers, or even mnemonics that don't exist.

I really dislike that we are going to allow this. Can we exclude these kinds of errors from any `skip-parse-errors` option? Maybe we need two levels: (1) ignore all parse errors and (2) ignore parse errors because CPU did not support those instructions?

Other than that, I would support having an explicit option to skip over parse errors.

https://github.com/llvm/llvm-project/pull/90474