Add fmin/fmax functions to math.h
#303
Conversation
mos-platform/common/c/math.cc
Outdated
| @@ -0,0 +1,22 @@ | |||
| #include <math.h> | |||
|
|
|||
| template <typename T> static inline bool _is_nan(const T x) { return x != x; } | |||
There was a problem hiding this comment.
May as well actually implement the macro in <math.h>:
#define isnan(arg) __builtin_isfpclass(arg, __FPCLASS_SNAN | __FPCLASS_QNAN)
There was a problem hiding this comment.
Or use the existing __builtin_isnan over __builtin_isfpclass? I just noticed that std::numeric_limits<float>::has_quiet_NaN == false and NAN is not defined. Is nan at all supported?
Update: also std::numeric_limits<float>::has_infinity == false.
There was a problem hiding this comment.
Also, I now notice that __builtin_fmin and __builtin_fmax along with several other math functions are already defined and seem to be correct. Perhaps better to use these? Are these covered by existing tests in the separate test suite?
There was a problem hiding this comment.
Yeah, good point. __builtin_isnan lowers to the same thing and is clearer.
By contrast, __builtin_fmin doesn't help us; it lowers to a compiler intrinsic that most naturally lowers to a call to fmin. By comparison, __builtin_isnan lowers to a call to __unordsf2 in compiler_rt, plus a bit of additional logic.
This adds min and max functions for
floatanddouble.Closes #299.