dust3d/thirdparty/cgal/CGAL-4.13/include/CGAL/FPU.h

517 lines
16 KiB
C
Raw Normal View History

// Copyright (c) 1998-2008
// Utrecht University (The Netherlands),
// ETH Zurich (Switzerland),
// INRIA Sophia-Antipolis (France),
// Max-Planck-Institute Saarbruecken (Germany),
// and Tel-Aviv University (Israel). All rights reserved.
//
// This file is part of CGAL (www.cgal.org); you can redistribute it and/or
// modify it under the terms of the GNU Lesser General Public License as
// published by the Free Software Foundation; either version 3 of the License,
// or (at your option) any later version.
//
// Licensees holding a valid commercial license may use this file in
// accordance with the commercial license agreement provided with the software.
//
// This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
// WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.
//
// $URL$
// $Id$
// SPDX-License-Identifier: LGPL-3.0+
//
//
// Author(s) : Sylvain Pion
#ifndef CGAL_FPU_H
#define CGAL_FPU_H
#include <CGAL/assertions.h>
#include <CGAL/use.h>
#ifndef __INTEL_COMPILER
#include <cmath> // for HUGE_VAL
#endif
// This file specifies some platform dependant functions, regarding the FPU
// directed rounding modes. There is only support for double precision.
//
// It also contains the definition of the Protect_FPU_rounding<> class,
// which helps to protect blocks of code needing a particular rounding mode.
#if defined __alpha__ && defined __linux__
extern "C" {
# include <fenv.h>
}
#elif defined __SUNPRO_CC && defined __sun
# include <ieeefp.h>
#elif defined __osf || defined __osf__
# ifdef __GNUG__
// GCC seems to remove (fixincludes) read_rnd/write_rnd...
# include "/usr/include/float.h"
# else
# include <cfloat>
# endif
#elif defined _MSC_VER || defined __sparc__ || \
(defined __i386__ && !defined __PGI && !defined __SUNPRO_CC \
&& !defined __SSE2__)
// Nothing to include.
#else
// By default we use the ISO C99 version.
# include <fenv.h>
#endif
// Some useful constants
#if defined CGAL_CFG_NO_LIMITS
# if defined CGAL_CFG_DENORMALS_COMPILE_BUG
// For compilers crashing when dealing with denormalized values.
// So we have to generate it at run time instead.
# define CGAL_IA_MIN_DOUBLE (CGAL::internal::get_static_minimin())
# else
# define CGAL_IA_MIN_DOUBLE (5e-324)
# endif
# define CGAL_IA_MAX_DOUBLE (1.7976931348623157081e+308)
#else
# include <limits>
# define CGAL_IA_MIN_DOUBLE std::numeric_limits<double>::denorm_min()
# define CGAL_IA_MAX_DOUBLE (std::numeric_limits<double>::max)()
#endif
// Pure and safe SSE2 mode (g++ -mfpmath=sse && (-msse2 || -march=pentium4))
// can be detected by :
// TODO : see what Intel and VC++ have to say about this.
#if defined __FLT_EVAL_METHOD__ && defined __SSE2_MATH__ && \
(__FLT_EVAL_METHOD__ == 0 || __FLT_EVAL_METHOD__ == 1)
# define CGAL_SAFE_SSE2
# include <xmmintrin.h>
#endif
// The CGAL_FPU_HAS_EXCESS_PRECISION macro is defined if some computations with
// double can use more than the 53bits of precision of IEEE754, and/or if the
// exponent has a wider range. This can produce double rounding effects and
// other bad things that we need to protect against.
// The typical offender is the traditional FPU of x86 (SSE2-only mode is not affected).
// Are there others, besides itanium and m68k?
#if !defined CGAL_IA_NO_X86_OVER_UNDER_FLOW_PROTECT && \
(((defined __i386__ || defined __x86_64__) && !defined CGAL_SAFE_SSE2) \
|| defined __ia64__ \
|| defined _M_IX86 || defined _M_X64 || defined _M_IA64 \
|| (defined FLT_EVAL_METHOD && FLT_EVAL_METHOD != 0 && FLT_EVAL_METHOD != 1))
# define CGAL_FPU_HAS_EXCESS_PRECISION
#endif
// Presence of SSE2 (for explicit use)
#if defined(__SSE2__) \
|| (defined(_M_IX86_FP) && _M_IX86_FP >= 2) \
|| defined(_M_X64)
# include <emmintrin.h>
# define CGAL_HAS_SSE2 1
#endif
// Only define CGAL_USE_SSE2 for 64 bits where malloc has a suitable
// alignment, 32 bits is too dangerous.
#if defined(CGAL_HAS_SSE2) && (defined(__x86_64__) || defined(_M_X64))
# define CGAL_USE_SSE2 1
#endif
#ifdef CGAL_CFG_DENORMALS_COMPILE_BUG
double& get_static_minimin(); // Defined in Interval_arithmetic_impl.h
#endif
namespace CGAL {
#ifdef CGAL_HEADER_ONLY
// Defined in test_FPU_rounding_mode_impl.h
struct Check_FPU_rounding_mode_is_restored;
inline const Check_FPU_rounding_mode_is_restored&
get_static_check_fpu_rounding_mode_is_restored();
#endif
// Inline function to stop compiler optimizations that shouldn't happen with
// pragma fenv on.
// - constant propagation
// - migration of fesetround across floating point operations
// - (-a)-b -> -(a+b)
// - (-a)*b -> -(a*b)
// etc
inline double IA_opacify(double x)
{
#ifdef __llvm__
// LLVM's support for inline asm is completely messed up:
// http://llvm.org/bugs/show_bug.cgi?id=17958
// http://llvm.org/bugs/show_bug.cgi?id=17959
// etc.
// This seems to produce code that is ok (not optimal but better than
// volatile). In case of trouble, use volatile instead.
# ifdef CGAL_HAS_SSE2
asm volatile ("" : "+x"(x) );
# elif (defined __VFP_FP__ && !defined __SOFTFP__) || defined __aarch64__
// ARM
asm volatile ("" : "+w"(x) );
# else
asm volatile ("" : "+m"(x) );
# endif
return x;
#elif defined __xlC__
// PowerPC - XL C++ (the z/OS version supposedly does not define this macro)
// If we give it an alternative "+fm", it gets confused and generates worse code.
asm volatile ("" : "+f"(x) );
return x;
#elif defined __GNUG__
// Intel used not to emulate this perfectly, we'll see.
// If we create a version of IA_opacify for vectors, note that gcc < 4.8
// fails with "+g" and we need to use "+mx" instead.
// "+X" ICEs ( http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59155 ) and
// may not be safe?
// The constraint 'g' doesn't include floating point registers ???
// Intel has a bug where -mno-sse still defines __SSE__ and __SSE2__
// (-mno-sse2 works though), no work-around for now.
# if defined __SSE2_MATH__ || (defined __INTEL_COMPILER && defined __SSE2__)
# if __GNUC__ * 100 + __GNUC_MINOR__ >= 409
// ICEs in reload/LRA with older versions.
asm volatile ("" : "+gx"(x) );
# else
asm volatile ("" : "+mx"(x) );
# endif
# elif (defined __i386__ || defined __x86_64__)
// "+f" doesn't compile on x86(_64)
// ( http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59157 )
// Don't mix "t" with "g": http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59180
// We can't put "t" with "x" either, prefer "x" for -mfpmath=sse,387.
// ( http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59181 )
asm volatile ("" : "+mt"(x) );
# elif (defined __VFP_FP__ && !defined __SOFTFP__) || defined __aarch64__
// ARM
asm volatile ("" : "+gw"(x) );
# elif defined __powerpc__ || defined __POWERPC__
// PowerPC
asm volatile ("" : "+gd"(x) );
# elif defined __sparc
// Sparc
asm volatile ("" : "+ge"(x) );
# elif defined __ia64
// Itanium
asm volatile ("" : "+gf"(x) );
# else
asm volatile ("" : "+g"(x) );
# endif
return x;
#else
volatile double e = x;
return e;
#endif
}
// Inline function to drop excess precision before we forget the rounding mode,
// and stop compiler optimizations at the same time.
inline double IA_force_to_double(double x)
{
#ifndef CGAL_FPU_HAS_EXCESS_PRECISION
return IA_opacify (x);
#else
#if defined __GNUG__
# ifdef CGAL_HAS_SSE2
// For an explanation of volatile:
// http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56027
asm volatile ("" : "+mx"(x) );
# else
// Similar to writing to a volatile and reading back, except that calling
// it k times in a row only goes through memory once.
asm volatile ("" : "+m"(x) );
# endif
return x;
#else
volatile double e = x;
return e;
#endif
#endif
}
// Interval arithmetic needs to protect against double-rounding effects
// caused by excess FPU precision, even if it forces the 53bit mantissa
// precision, because there is no way to fix the problem for the exponent
// which has the same problem. This affects underflow and overflow cases.
// In case one does not care about such "extreme" situations, one can
// set CGAL_IA_NO_X86_OVER_UNDER_FLOW_PROTECT to pretend there is no excess
// precision.
#if defined CGAL_FPU_HAS_EXCESS_PRECISION
# define CGAL_IA_FORCE_TO_DOUBLE(x) CGAL::IA_force_to_double(x)
#elif 1
// LLVM doesn't have -frounding-math so needs extra protection.
// GCC also migrates fesetround calls over FP instructions, so protect
// everyone.
# define CGAL_IA_FORCE_TO_DOUBLE(x) CGAL::IA_opacify(x)
#else
// Unused, reserved to compilers without excess precision and pragma fenv on.
// ??? Should we trust Visual Studio not to optimize too much and let it use
// this when CGAL_IA_NO_X86_OVER_UNDER_FLOW_PROTECT?
# define CGAL_IA_FORCE_TO_DOUBLE(x) (x)
#endif
// We sometimes need to stop constant propagation,
// because operations are done with a wrong rounding mode at compile time.
#ifndef CGAL_IA_DONT_STOP_CONSTANT_PROPAGATION
# define CGAL_IA_STOP_CPROP(x) CGAL::IA_opacify(x)
#else
# define CGAL_IA_STOP_CPROP(x) (x)
#endif
// std::sqrt(double) on VC++ and CygWin is buggy when not optimizing.
#if defined ( _MSC_VER ) && ! defined ( _WIN64 )
inline double IA_bug_sqrt(double d)
{
_asm
{
fld d
fsqrt
fstp d
}
return d;
}
# define CGAL_BUG_SQRT(d) IA_bug_sqrt(d)
#elif defined __SSE2_MATH__
// For SSE2, we need to call __builtin_sqrt() instead of libc's sqrt().
# define CGAL_BUG_SQRT(d) __builtin_sqrt(d)
#elif defined __CYGWIN__
inline double IA_bug_sqrt(double d)
{
double r;
asm volatile ("fsqrt" : "=t"(r) : "0"(d));
return r;
}
# define CGAL_BUG_SQRT(d) IA_bug_sqrt(d)
#else
# define CGAL_BUG_SQRT(d) std::sqrt(d)
#endif
// Here are the operator macros that make use of the above.
// With GCC, we can do slightly better : test with __builtin_constant_p()
// that both arguments are constant before stopping one of them.
// Use inline functions instead ?
#define CGAL_IA_ADD(a,b) CGAL_IA_FORCE_TO_DOUBLE((a)+CGAL_IA_STOP_CPROP(b))
#define CGAL_IA_SUB(a,b) CGAL_IA_FORCE_TO_DOUBLE(CGAL_IA_STOP_CPROP(a)-(b))
#define CGAL_IA_MUL(a,b) CGAL_IA_FORCE_TO_DOUBLE((a)*CGAL_IA_STOP_CPROP(b))
#define CGAL_IA_DIV(a,b) CGAL_IA_FORCE_TO_DOUBLE((a)/CGAL_IA_STOP_CPROP(b))
#define CGAL_IA_SQUARE(a) CGAL_IA_MUL(a,a)
#define CGAL_IA_SQRT(a) \
CGAL_IA_FORCE_TO_DOUBLE(CGAL_BUG_SQRT(CGAL_IA_STOP_CPROP(a)))
#if defined CGAL_SAFE_SSE2
#define CGAL_IA_SETFPCW(CW) _MM_SET_ROUNDING_MODE(CW)
#define CGAL_IA_GETFPCW(CW) CW = _MM_GET_ROUNDING_MODE()
typedef unsigned int FPU_CW_t;
#define CGAL_FE_TONEAREST _MM_ROUND_NEAREST
#define CGAL_FE_TOWARDZERO _MM_ROUND_TOWARD_ZERO
#define CGAL_FE_UPWARD _MM_ROUND_UP
#define CGAL_FE_DOWNWARD _MM_ROUND_DOWN
#elif defined __i386__ && !defined __PGI && !defined __SUNPRO_CC \
&& !defined CGAL_HAS_SSE2
// If we use both 387 and sse2, be safe and drop to fe[gs]etround.
// Can we test CGAL_USE_SSE2 instead?
// The GNU libc version (cf powerpc) is nicer, but doesn't work on libc 5 :(
// This one also works with CygWin.
// Note that the ISO C99 version may not be enough because of the extended
// mantissa issue on x86 (may be required by some kinds of computation, but
// as far as CGAL::Interval_nt<> is concerned, the double-rounding issues
// are taking care of there).
#define CGAL_IA_SETFPCW(CW) asm volatile ("fldcw %0" : :"m" (CW))
#define CGAL_IA_GETFPCW(CW) asm volatile ("fnstcw %0" : "=m" (CW))
typedef unsigned short FPU_CW_t;
#define CGAL_FE_TONEAREST (0x000 | 0x127f)
#define CGAL_FE_TOWARDZERO (0xc00 | 0x127f)
#define CGAL_FE_UPWARD (0x800 | 0x127f)
#define CGAL_FE_DOWNWARD (0x400 | 0x127f)
#elif defined __SUNPRO_CC && defined __sun
#define CGAL_IA_SETFPCW(CW) fpsetround(fp_rnd(CW))
#define CGAL_IA_GETFPCW(CW) CW = fpgetround()
typedef unsigned int FPU_CW_t;
#define CGAL_FE_TONEAREST FP_RN
#define CGAL_FE_TOWARDZERO FP_RZ
#define CGAL_FE_UPWARD FP_RP
#define CGAL_FE_DOWNWARD FP_RM
#elif defined __sparc__
#define CGAL_IA_SETFPCW(CW) asm volatile ("ld %0,%%fsr" : :"m" (CW))
#define CGAL_IA_GETFPCW(CW) asm volatile ("st %%fsr,%0" : "=m" (CW))
typedef unsigned int FPU_CW_t;
#define CGAL_FE_TONEAREST (0x0 | 0x20000000 | 0x1f)
#define CGAL_FE_TOWARDZERO (0x40000000 | 0x20000000 | 0x1f)
#define CGAL_FE_UPWARD (0x80000000 | 0x20000000 | 0x1f)
#define CGAL_FE_DOWNWARD (0xc0000000 | 0x20000000 | 0x1f)
#elif defined __mips__
#define CGAL_IA_SETFPCW(CW) asm volatile ("ctc1 %0,$31" : :"r" (CW))
#define CGAL_IA_GETFPCW(CW) asm volatile ("cfc1 %0,$31" : "=r" (CW))
typedef unsigned int FPU_CW_t;
#define CGAL_FE_TONEAREST (0x0)
#define CGAL_FE_TOWARDZERO (0x1)
#define CGAL_FE_UPWARD (0x2)
#define CGAL_FE_DOWNWARD (0x3)
#elif defined __osf || defined __osf__ // Not yet supported.
#define CGAL_IA_SETFPCW(CW) write_rnd(CW)
#define CGAL_IA_GETFPCW(CW) CW = read_rnd()
typedef unsigned int FPU_CW_t;
#define CGAL_FE_TONEAREST FP_RND_RN
#define CGAL_FE_TOWARDZERO FP_RND_RZ
#define CGAL_FE_UPWARD FP_RND_RP
#define CGAL_FE_DOWNWARD FP_RND_RM
#elif defined ( _MSC_VER )
#if ( _MSC_VER < 1400)
#define CGAL_IA_SETFPCW(CW) _controlfp (CW, _MCW_RC )
#define CGAL_IA_GETFPCW(CW) CW = _controlfp (0, 0 ) & _MCW_RC
typedef unsigned short FPU_CW_t;
#else
#define CGAL_IA_SETFPCW(CW) unsigned int dummy; _controlfp_s (&dummy, CW, _MCW_RC )
#define CGAL_IA_GETFPCW(CW)_controlfp_s (&CW, 0, 0 ); CW &= _MCW_RC
typedef unsigned int FPU_CW_t;
#endif
#define CGAL_FE_TONEAREST _RC_NEAR
#define CGAL_FE_TOWARDZERO _RC_CHOP
#define CGAL_FE_UPWARD _RC_UP
#define CGAL_FE_DOWNWARD _RC_DOWN
#else
// This is a version following the ISO C99 standard, which aims at portability.
// The drawbacks are speed on one hand, and also, on x86, it doesn't fix the
// extended mantissa issue (this is not a problem for IA, but it is one for
// some future modular computations).
#define CGAL_IA_SETFPCW(CW) fesetround(CW)
#define CGAL_IA_GETFPCW(CW) CW = fegetround()
typedef int FPU_CW_t;
#define CGAL_FE_TONEAREST FE_TONEAREST
#define CGAL_FE_TOWARDZERO FE_TOWARDZERO
#define CGAL_FE_UPWARD FE_UPWARD
#define CGAL_FE_DOWNWARD FE_DOWNWARD
#endif
// User interface:
inline
FPU_CW_t
FPU_get_cw (void)
{
FPU_CW_t cw;
CGAL_IA_GETFPCW(cw);
return cw;
}
// User interface (cont):
inline
void
FPU_set_cw (FPU_CW_t cw)
{
CGAL_IA_SETFPCW(cw);
}
inline
FPU_CW_t
FPU_get_and_set_cw (FPU_CW_t cw)
{
FPU_CW_t old = FPU_get_cw();
FPU_set_cw(cw);
return old;
}
// A class whose constructor sets the FPU mode to +inf, saves a backup of it,
// and whose destructor resets it back to the saved state.
template <bool Protected = true> struct Protect_FPU_rounding;
template <>
struct Protect_FPU_rounding<true>
{
Protect_FPU_rounding(FPU_CW_t r = CGAL_FE_UPWARD)
: backup( FPU_get_and_set_cw(r) ) {}
~Protect_FPU_rounding()
{
FPU_set_cw(backup);
}
private:
FPU_CW_t backup;
};
template <>
struct Protect_FPU_rounding<false>
{
Protect_FPU_rounding() {}
Protect_FPU_rounding(FPU_CW_t /*= CGAL_FE_UPWARD*/) {}
};
// A wrapper on top of the Protect_FPU_rounding to add "expensive" checks
// of the rounding mode. It is used internally, to benefit from the
// protector declarations to add checks in non-protected mode.
template <bool Protected = true>
struct Checked_protect_FPU_rounding
: Protect_FPU_rounding<Protected>
{
Checked_protect_FPU_rounding()
{
CGAL_expensive_assertion(FPU_get_cw() == CGAL_FE_UPWARD);
}
Checked_protect_FPU_rounding(FPU_CW_t r)
: Protect_FPU_rounding<Protected>(r)
{
CGAL_expensive_assertion(FPU_get_cw() == CGAL_FE_UPWARD);
}
};
// The class Set_ieee_double_precision forces the double precision (53bit mantissa),
// to protect from double rounding effects on x86 FPU.
// ( Note that it also sets the rounding mode to nearest. )
// Its destructor restores the FPU state as it was previously.
// Note that this affects "long double" as well, and other potential side effects.
// And note that it does not (cannot) "fix" the same problem for the exponent.
struct Set_ieee_double_precision
#ifdef CGAL_FPU_HAS_EXCESS_PRECISION
: public Protect_FPU_rounding<>
{
Set_ieee_double_precision()
: Protect_FPU_rounding<>(CGAL_FE_TONEAREST) {}
};
#else
{
Set_ieee_double_precision() {} // only to kill compiler warnings.
};
#endif
// The following function serves the same goal as Set_ieee_double_precision but
// does the change globally (no destructor resets the previous behavior).
inline void force_ieee_double_precision()
{
#ifdef CGAL_FPU_HAS_EXCESS_PRECISION
FPU_set_cw(CGAL_FE_TONEAREST);
#endif
}
} //namespace CGAL
#ifdef CGAL_HEADER_ONLY
#include <CGAL/test_FPU_rounding_mode_impl.h>
#endif // CGAL_HEADER_ONLY
#endif // CGAL_FPU_H