[clang] Add support for option -ffp-eval-method and extend #pragma float_control similarly

[mibintc]

Summary

The Intel compiler supports the option "-fp-model=(source|double|extended)" which causes the compiler to use a wider type for intermediate floating point calculations. Also supported is a way to embed this effect in the source program with #pragma float_control(source|double|extended). This patch proposes to extend pragma float_control syntax, and also to support a new floating point option via "-ffp-eval-method=(source|double|extended)"

Note that the C floating point working group has proposed nearly identical pragma semantics in http://www.open-std.org/jtc1/sc22/wg14/www/docs/n1974.pdf like this
#pragma STDC FENV_FLT_EVAL_METHOD width
I would like to add support for this also, but it's not in this patch. In the WG14 document, the standard widths are 0,1,2 representing source, double and extended precision. Negative values of width are reserved for the implementation.

The ICL option with description of semantics is at this url, https://software.intel.com/content/www/us/en/develop/documentation/cpp-compiler-developer-guide-and-reference/top/compiler-reference/compiler-options/compiler-option-details/floating-point-options/fp-model-fp.html
Quoting that here for easy reference "source
Rounds intermediate results to source-defined precision.
double
Rounds intermediate results to 53-bit (double) precision.
extended
Rounds intermediate results to 64-bit (extended) precision."

[mibintc]

clang/lib/Sema/SemaAttr.cpp

478

@rjmccall I would like to push a new value for FLT_EVAL_METHOD at the start of the scope of this pragma, then pop that value of the macro to restore the previous setting when the scope is exited, should I do that in ParsePragma? Can I do that by injecting "pragma push_macro(...)" into the token stream, do you have other suggestion or is there something similar in clang that I can study how to do this?

[mibintc]

@rjmccall Hoping you can take a look at this patch

[mibintc]

I found a way to get the macro FLT_EVAL_METHOD set properly, so this patch is ready for your review, thanks!

[aaron.ballman]

One question that came up in off-line discussions with @mibintc is whether the evaluation method should have an impact on constant folding or not. My intuition is that it should impact constant folding because it would be pretty strange for the evaluation at runtime to produce different results, but I don't have anything concrete to back my intuition up with.

clang/docs/UsersManual.rst

1464

clang/include/clang/Basic/LangOptions.h

234

FYI: it may be somewhat confusing that we have an enumerator with default in the name but that enumerator isn't the default.

clang/lib/Frontend/CompilerInvocation.cpp

3080

May want to update the comment above for why we're turning the option off here.

clang/lib/Lex/PPMacroExpansion.cpp

1714

It looks like some unrelated formatting changes snuck in.

[mibintc]

clang/include/clang/Basic/LangOptions.h

234

That's true. I had a heck of a time getting the initialization of the lang option setting correct: the setting should come from the TargetInfo if command line options don't override. That seems simple enough but there's somethng really odd about initialization. Adding the "default target" enumeral allowed me to get it initialized the way I wanted.

[mibintc]

clang/docs/LanguageExtensions.rst

3363

I'd like to add Intel ICL compatible syntax, but 'off' doesn't make too much sense to me. In the Intel compiler using'off' requires both ffp-exception-behavior to be ignore, and FENV_ACCESS to be off. If that's true then setting float_control(source|double|extended, off) is allowable and the semantics of float evaluation method set to a 4th level "float eval fast" which is loosely described in the commentary as "don't care". Ideally "float eval fast" would deliver the same results, in this context, as Microsoft's fp:fast command line option. I could make "off" an illegal setting, or warn that it's ignored. @andrew.w.kaylor do you have any input here?