[llvm] Revert "[TailDuplicator] Do not restrict the computed gotos" (PR #132431)

[Alexander Kornienko via llvm-commits]

alexfh wrote:

> Is there some possible middle ground, where we bump the limit a bit to catch most perf relevant cases, but avoid cases that won’t compile in reasonable time?

If you're talking about `-tail-dup-pred-size`, then setting it to 1000 (before #114990) for my reduced example still results in a reasonable compilation time, while 10000 is already at least 10x slower.

But I suggest to first revert to unbreak Clang, and then start experimenting with the value.

https://github.com/llvm/llvm-project/pull/132431