[llvm] [SDAG] Fix deferring constrained function calls (PR #153029)

[Serge Pavlov via llvm-commits]

spavloff wrote:

> Rounding mode can be changed by an arbitrary function called from inside a basic block. Thus, we can't model it perfectly and need to take it into account when reordering.

What's wrong with the current way FP operations are modeled? Each operation that may depend on dynamic rounding has some memory effect, which indicates that this operation reads FP control register. Functions that changes the control register must have memory effect, that this function writes FP control register. Any instruction movement must honor this relations, just as for real memory access.

https://github.com/llvm/llvm-project/pull/153029