[clang] [llvm] [ARM] enable FENV_ACCESS pragma support for hard-float targets (PR #137101)

[David Green via cfe-commits]

davemgreen wrote:

> Will it be enough to add strict fp tests similar to those made for AArch64?

Yeah I think so for the most part. There is the added complication of soft-fp but it should mostly be a case of testing all the strict-fp operations under a couple of configs to make sure everything works OK.
Note that there is talk of changing how the constrained-fp intrinsics are represented in the IR https://discourse.llvm.org/t/rfc-change-of-strict-fp-operation-representation-in-ir/85021, but my understanding is that they will work the same for codegen.

> Regarding the changes in some tests (e.g. Thumb2/mve-fmas.ll): they arise after updating patterns in ARMInstrVFP.td, which in turn reorders some instructions. For VFP tests, as far as I understand, such strict fp behavior can be expected, but I’m unsure to what extent the MVE tests should be affected.

I think we need to teach it that the fpscr bits of nofpexcept instructions and `fpscr_nzcv` do not interact and can be  treated separately for scheduling. In general we need to be careful not to make perf worse for all the people who do not use strict-fp. Maybe in this case it is caused by the `vmrs    APSR_nzcv, fpscr` instructions?

https://github.com/llvm/llvm-project/pull/137101