Make std.math.fmax() and .fmin() variadic. Add explicit NaN handling.

[tsbockman]

No description provided.

[wilzbach]

shouldn't this be the first comparison?

[tsbockman]

No.

[9il]

fmax and fmin should have 2 args and be simple math functions like in C. All additional logic can be located in max and min

[9il]

my suggestion was to use fmax and fmin in max and min.

[9il]

fmin and fmax should be compiler intrinsics.

[9il]

See also https://llvm.org/bugs/show_bug.cgi?id=24314

[tsbockman]

fmin and fmax should be compiler intrinsics.

Shouldn't they be in core.math then?

[9il]

Shouldn't they be in core.math then?

I don't know.

See std.math in LDC Phobos
https://github.com/ldc-developers/phobos/blob/release-0.17.1/std/math.d#L653

The best option for math functions is to be templates (so they can be inlined) and support all FP types.

[tsbockman]

The description of core.math says "Builtin mathematical intrinsics":
https://dlang.org/phobos/core_math.html

[9il]

The description of core.math says "Builtin mathematical intrinsics":
https://dlang.org/phobos/core_math.html

But this is just declarations.

[tsbockman]

But this is just declarations.

It is analogous to core.checkedint and core.bitop, which also contain software fallback implementations.

[9il]

But this is just declarations.
It is analogous to core.checkedint and core.bitop, which also contain software fallback implementations.

This is not really important.

Functions in core.math and std.math are deprecated:

1. They are real centric: this slowdown execution on modern x86_64 CPUs
2. They are not templates, so they are not inlined: this slowdown execution on modern x86_64 CPUs too

The good std.math update should allow to use all math functions from LLVM list http://llvm.org/docs/LangRef.html#standard-c-library-intrinsics.

[tsbockman]

They are real centric: this slowdown execution on modern x86_64 CPUs

This can be fixed. There's nothing stopping us from adding double and float overloads to core.math.

They are not templates, so they are not inlined: this slowdown execution on modern x86_64 CPUs too

core.math is for intrinsics which are not normal functions, but are effectively always inlined (even across modules). They are implemented in the compiler backend, I believe.
https://en.wikipedia.org/wiki/Intrinsic_function