[llvm] [SystemZ] Remove getInliningThresholdMultiplier() override (PR #94612)

[Ulrich Weigand via llvm-commits]

uweigand wrote:

This is interesting.  To answer the question if there's anything fundamental about s390x that would require higher inlining thresholds - not really.  However, our ABI historically has had a somewhat higher overhead than some others, which made inlining in some benchmarks more helpful than elsewhere.  Also, in our typical usage, executable size tends to be less of a concern than on some other platforms.  As a result, we've always thought somewhat more aggressive inlining might be a better tradeoff for us.

That said, I agree we certainly do not want inlining thresholds so high that they cause extreme increases in compile time and/or executable size - in particular, if that introduces notable differences between s390x and other server architectures. If that is indeed the case, we likely have to revisit this choice.

Given that, one remaining question would be whether instead of simply generally increasing the threshold, we can find a way to make inlining decisions more intelligent so that we can get back those benchmark regressions.  @JonPsson1 that might be something worth spending some time investigating.

https://github.com/llvm/llvm-project/pull/94612