[libc++] Support Newlib as libc++'s C library (on ARM) [cmath part, part 1]

[jroelofs]

Summary

Add shims for nexttoward/nexttowardf/nexttowardl to patch over newlib missing them.

[jfb]

include/support/newlib/math.h

18

Why __arm__? It seems better to key off long double being the same as double. I'm not sure if clang has a macro for this, but it should.

[jfb]

include/support/newlib/math.h

18

Discussed on IRC:

#if __SIZEOF_DOUBLE__ == __SIZEOF_LONG_DOUBLE__

[rengolin]

src/support/newlib/math.cpp

16

you wont need this assert if you use the same macro as above (double == long double), right? Possibly not even the comment...

[jroelofs]

src/support/newlib/math.cpp

16

right :) I think though that the comment makes it clear: these are not an approximation; they precisely implement the required behavior.

[rengolin]

src/support/newlib/math.cpp

16

good point.

[jroelofs]

Use __SIZEOF_DOUBLE__ == __SIZEOF_LONG_DOUBLE__ for the check, and guards to make sure these new support files are only used for newlib.

[jfb]

Everything looks good but the presence of math.cpp.

src/support/newlib/math.cpp

32

Wouldn't it be better to inline all of these into math.h?

[jroelofs]

Abandoning in favor of plunking these directly in Newlib: https://sourceware.org/ml/newlib/2014/msg00469.html

include/cmath

 #include <__config>
 #include <math.h>
 #include <type_traits>
 
 #ifdef _LIBCPP_MSVCRT
 #include "support/win32/math_win32.h"
 #endif
 
+#if defined(_NEWLIB_VERSION)
+#include "support/newlib/math.h"
+#endif
+
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #pragma GCC system_header
 #endif
 
 // signbit
 
 #ifdef signbit
 

include/support/newlib/math.h

+//===----------------------------------------------------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is dual licensed under the MIT and the University of Illinois Open
+// Source Licenses. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _LIBCPP_SUPPORT_NEWLIB_MATH_H
+#define _LIBCPP_SUPPORT_NEWLIB_MATH_H
+
+#if !defined(_NEWLIB_VERSION)
+#error This should only be used when Newlib is the underlying libc
+#endif
+
+#if __cplusplus
+extern "C" {
+#endif
+
+#if __SIZEOF_DOUBLE__ == __SIZEOF_LONG_DOUBLE__
+double nexttoward(double d, long double td);
+float nexttowardf(float f, long double td);
+long double nexttowardl(long double ld, long double td);
+
+long double logbl(long double x);
+#endif
+
+#if __cplusplus
+}
+#endif
+
+#endif
+

src/support/newlib/math.cpp

+//===----------------------------------------------------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is dual licensed under the MIT and the University of Illinois Open
+// Source Licenses. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+#include <math.h>
+#include "support/newlib/math.h"
+
+#if defined(_NEWLIB_VERSION)
+
+#if __SIZEOF_DOUBLE__ == __SIZEOF_LONG_DOUBLE__
+// nexttoward* functions are precisely the same as their nextafter* siblings
+// when long double is the same as double.
+double nexttoward(double d, long double td) {
+  return nextafter(d, (double)td);
+}
+
+float nexttowardf(float f, long double td) {
+  return nextafterf(f, (float)td);
+}
+
+long double nexttowardl(long double ld, long double td) {
+  return nextafter((double)ld, (double)td);
+}
+
+long double logbl(long double x) {
+  return logb((double)x);
+}
+#endif
+
+#endif