[FPEnv] clang support for constrained FP builtins

[kpn]

Summary

Currently if constrained fp math is requested clang is still emitting calls to the non-constrained versions of math builtins. This patch corrects that for calls that are not target-specific.

I really wanted to keep constrained/non-constrained decisions in the IRBuilder, but that causes complications. This version stays close to that ideal without the complications.

[rjmccall]

clang/lib/CodeGen/CGBuiltin.cpp

359

I don't see any situation where not_intrinsic is passed in here; why all these checks?

2299

What can this not use the standard path?

[kpn]

D62731 is required for testing this patch.

[kpn]

clang/lib/CodeGen/CGBuiltin.cpp

359

I tend to be conservative, and I'm not good at predicting the future. I can remove the checks easily in the next round.

2299

I'll take another look. It wasn't using the standard path before so I made my added code match. If it can use the standard path in both cases then maybe it should?

[rjmccall]

clang/lib/CodeGen/CGBuiltin.cpp

359

I just can't imagine why you would ever call this function without a constrained intrinsic ID when its only purpose is to choose between a constrained and unconstrained intrinsic. It's always better to start with stronger preconditions and then generalize later if necessary.

2299

Yeah. It's possible the type-heterogeneity (the power always being an int) stops that from working for some reason, but if not, it's always nicer to re-use more code.

[kpn]

Update for review comments.

[rjmccall]

Thanks, LGTM.