[llvm] CodeGenPrepare: Remove threadlocal_address intrinsic when cheap to recompute. (PR #87844)

[Matthias Braun via llvm-commits]

MatzeB wrote:

Reworked this change in terms of `CodeGenPrepare` address-mode sinking logic and enabled the duplication on `llvm.threadlocal.address` for X86 `TLSModel::LocalExec` and `TLSModel::InitialExec` where it improves code quality.

https://github.com/llvm/llvm-project/pull/87844