[llvm] [ARM][ConstantIslands] Correct MinNoSplitDisp calculation (PR #114590)

[via llvm-commits]

pzhengqc wrote:

> It has been a while since I looked into the constant island pass, but this sounds OK. As you mention this should only kick in when we are already 15 iterations in, so is aiming to converge quickly at the expense of optimality. Did you run any checks (the llvm-test-suite or a boot strap), that can check we don't iterate for a long time now? Thanks.

Yes, @davemgreen, I've run some checks (the llvm-test-suite, a boot strap, spec2000 and spec2006), but none of these tests seems to take more than 15 iterations for the constant island pass to converge. I guess, in reality, it's very rare for the check `if (CloserWater && MinNoSplitDisp > U.getMaxDisp() / 2)` to be true.

https://github.com/llvm/llvm-project/pull/114590