[FPEnv] Make the strictfp attribute rules stricter.

[kpn]

Summary

This is mostly a revert of D148138 but with a twist. I discovered when working on tests so I can land Verifier changes from D146845 that the rules as they are after D148138 don't work.

Specifically, if a non-strictfp function calls a strictfp function, and we look through to the function being called for the strictfp attribute, then D146845 will reject the compile because it thinks we are mixing strictfp and non-strictfp in the same function. By reverting to the rule that the strictfp attribute is required on every single call in a strictfp function we avoid the issue.

However, having duplicate information in every call to a constrained intrinsic doesn't seem worthwhile. So, I added the twist that calls to intrinsics with the strictfp attribute attached to the intrinsic don't need it again on the call instruction. This avoids duplicate information while solving the problem at the boundary between non-strictfp and strictfp code.

[arsenm]

llvm/docs/LangRef.rst

23761–23764

I don't understand why you're specifically mentioning call sites Attribute behavior is always union of call site and declaration, so this should apply to all calls

[arsenm]

Weirdly phrased in a special case sort of way. Attributes are a union of callsite and declaration

[kpn]

With the revised D146845 now approved this change is no longer needed.

llvm/docs/LangRef.rst

 The number and order of floating-point exceptions is NOT guaranteed.  For
 example, a series of FP operations that each may raise exceptions may be
 vectorized into a single instruction that raises each unique exception a single
 time.
 
 Proper :ref:`function attributes <fnattrs>` usage is required for the
 constrained intrinsics to function correctly.
 
 All function *calls* done in a function that uses constrained floating
-point intrinsics must have the ``strictfp`` attribute either on the
-calling instruction or on the declaration or definition of the function
-being called.
+point intrinsics must have the ``strictfp`` attribute. Calls to intrinsics
+that are defined to have the ``strictfp`` are allowed to omit the attribute
+from the call site.
 
 All function *definitions* that use constrained floating point intrinsics
 must have the ``strictfp`` attribute.
 
 '``llvm.experimental.constrained.fadd``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 Syntax: