Windows2003-3790/windows/advcore/gdiplus/engine/runtime/real.hpp

748 lines
24 KiB
C++
Raw Permalink Normal View History

2001-01-01 00:00:00 +01:00
/**************************************************************************\
*
* Copyright (c) 1998 Microsoft Corporation
*
* Abstract:
*
* Contains macros for helping with floating point arithmetic.
*
* History:
*
* 07/08/1999 agodfrey
* Remove MSVCRT dependency.
* 12/06/1998 andrewgo
* Created it.
*
\**************************************************************************/
#ifndef _REAL_HPP_
#define _REAL_HPP_
// The following stuff is taken from Office code
/*****************************************************************************
Intrinsic functions - it is essential that these be used because there is
no library implementation. Note that the intrinsic forms often have
restricted behavior, e.g. the argument to a trignometric function must be
less than 2^63 radians.
******************************************************************* JohnBo **/
#pragma intrinsic(sin, cos, tan)
#pragma intrinsic(atan, atan2)
#pragma intrinsic(sqrt)
#pragma intrinsic(log, log10, exp)
#pragma intrinsic(fabs)
#pragma intrinsic(fmod)
namespace GpRuntime
{
/*
[JohnBo]
These in-line functions are required to force direct use of the
instructions - without this the CI versions are used unless g optimization
is switched on, with g optimization on the in-line function calls will be
removed completely.
*/
#pragma optimize("g", on)
inline double InlineSin(double x) { return sin(x); }
inline double InlineCos(double x) { return cos(x); }
inline double InlineTan(double x) { return tan(x); }
inline double InlineATan(double x) { return atan(x); }
inline double InlineATan2(double y, double x) { return atan2(y, x); }
inline double InlineSqrt(double x) { return sqrt(x); }
inline double InlineLog(double x) { return log(x); }
inline double InlineLog10(double x) { return log10(x); }
inline double InlineExp(double x) { return exp(x); }
/* Restore default optimization. */
#pragma optimize("", on)
// Out-of-line math functions
// pow: We implemented it ourselves
// exp: Because the inline version is so long, the compiler won't
// inline it unless the original caller has 'generate fast code'
// set. Instead, we use an out-of-line version.
double Pow(double, double);
double Exp(double);
/*// Trying something:
inline double FPX86InlineFmod(double x, double y) { return fmod(x,y); }
#define fmod(x,y) FPX86InlineFmod(x,y)
*/
};
/* Force use of the in-line functions. */
#define sin(x) GpRuntime::InlineSin(x)
#define cos(x) GpRuntime::InlineCos(x)
#define tan(x) GpRuntime::InlineTan(x)
#define atan(x) GpRuntime::InlineATan(x)
#define atan2(y,x) GpRuntime::InlineATan2(y,x)
#define sqrt(x) GpRuntime::InlineSqrt(x)
#define log(x) GpRuntime::InlineLog(x)
#define log10(x) GpRuntime::InlineLog10(x)
#define exp(x) GpRuntime::Exp(x)
#define pow(x,y) GpRuntime::Pow(x,y)
/* Integer interfaces */
#pragma intrinsic(abs, labs)
// End of Office code
// Our pixel positioning uses 28.4 fixed point arithmetic and therefore
// anything below the threshold of 1/32 should be irrelevant.
// Our choice of PIXEL_EPSILON is 1/64 which should give us correct pixel
// comparisons even in the event of accumulated floating point error.
#define PIXEL_EPSILON 0.015625f // 1/64
#ifndef REAL_EPSILON
#define REAL_EPSILON FLT_EPSILON
#endif
// This is for computing the complexity of matrices. When you compose matrices
// or scale them up by large factors, it's really easy to hit the REAL_EPSILON
// limits without actually affecting the transform in any noticable way.
// e.g. a matrix with a rotation of 1e-5 degrees is, for all practical purposes,
// not a rotation.
#define CPLX_EPSILON (REAL_EPSILON*5000.0f)
// This is a tolerance error for the difference of each real coordinates.
// This leaves some acurracy to calculate tangent or normal for the two points.
#define POINTF_EPSILON (REAL_EPSILON*5000.0f)
#define REALFMOD fmodf
#define REALSQRT sqrtf
#ifndef REALABS
#define REALABS fabsf
#endif
#define REALSIN sinf
#define REALCOS cosf
#define REALATAN2 atan2f
// #define REAL_EPSILON DBL_EPSILON
// #define REALFMOD fmod
// #define REALSQRT sqrt
// #define REALABS fabs
// #define REALSIN sin
// #define REALCOS cos
// #define REALATAN2 atan2
// convert from unknown FLOAT type => REAL
#define TOREAL(x) (static_cast<REAL>(x))
// defined separately for possibly future optimization LONG to REAL
#define LTOF(x) (static_cast<REAL>(x))
// Return the positive integer remainder of a/b. b should not be zero.
// note that a % b will return a negative number
// for a<0 xor b<0 which is not suitable for texture mapping
// or brush tiling.
// This macro computes the remainder of a/b
// correctly adjusted for tiling negative coordinates.
#define RemainderI(a, b)\
((a) >= 0 ? (a) % (b) : \
(b) - 1 - ((-(a) - 1) % (b)))
// This definition assumes y > 0
inline REAL GpModF(REAL x, REAL y)
{
if(x > 0)
{
return static_cast<REAL> (x - ((INT) (x/y))*y);
}
else if(x < 0)
{
REAL z;
x = - x;
z = static_cast<REAL> (x - ((INT) (x/y))*y);
if(z > 0)
z = y - z;
return z;
}
else // x == 0
return 0.0;
/*
// This assumes fmod(x, y) = - fmod(-x, y) when x < 0.
REAL z = REALFMOD(x, y);
if(z >= 0)
return z;
else
return y + z;
*/
}
#if defined(_X86_) && !defined(_COMPLUS_GDI)
#define _USE_X86_ASSEMBLY
#else
#undef _USE_X86_ASSEMBLY
#endif
#if defined(_USE_X86_ASSEMBLY)
#define FPU_STATE() \
UINT32 cwSave; \
UINT32 cwTemp;
#define FPU_GET_STATE() \
UINT32 cwSave = this->cwSave; \
UINT32 cwTemp = this->cwTemp;
#define FPU_SAVE_STATE() \
this->cwSave = cwSave; \
this->cwTemp = cwTemp;
#define FPU_SAVE_MODE() \
UINT32 cwSave; \
UINT32 cwTemp; \
\
__asm { \
_asm fnstcw WORD PTR cwSave \
_asm mov eax, cwSave \
_asm mov cwTemp, eax \
} \
this->cwSave =
#define FPU_RESTORE_MODE() \
__asm { \
_asm fldcw WORD PTR cwSave \
}
#define FPU_RESTORE_MODE_NO_EXCEPTIONS()\
__asm { \
_asm fnclex \
_asm fldcw WORD PTR cwSave \
}
#define FPU_CHOP_ON() \
__asm { \
_asm mov eax, cwTemp \
_asm or eax, 0x0c00 \
_asm mov cwTemp, eax \
_asm fldcw WORD PTR cwTemp \
}
#define FPU_ROUND_ON() \
__asm { \
_asm mov eax, cwTemp \
_asm and eax,0xf3ff \
_asm mov cwTemp, eax \
_asm fldcw WORD PTR cwTemp \
}
#define FPU_ROUND_ON_PREC_HI() \
__asm { \
_asm mov eax, cwTemp \
_asm and eax,0xf0ff \
_asm or eax,0x0200 \
_asm mov cwTemp, eax \
_asm fldcw WORD PTR cwTemp \
}
#define FPU_PREC_LOW() \
__asm { \
_asm mov eax, cwTemp \
_asm and eax, 0xfcff \
_asm mov cwTemp, eax \
_asm fldcw WORD PTR cwTemp \
}
#define FPU_PREC_LOW_MASK_EXCEPTIONS() \
__asm { \
_asm mov eax, cwTemp \
_asm and eax, 0xfcff \
_asm or eax, 0x3f \
_asm mov cwTemp, eax \
_asm fldcw WORD PTR cwTemp \
}
#define FPU_CHOP_ON_PREC_LOW() \
__asm { \
_asm mov eax, cwTemp \
_asm or eax, 0x0c00 \
_asm and eax, 0xfcff \
_asm mov cwTemp, eax \
_asm fldcw WORD PTR cwTemp \
}
#define FPU_CHOP_OFF_PREC_HI() \
__asm { \
_asm mov eax, cwTemp \
_asm mov ah, 2 \
_asm mov cwTemp, eax \
_asm fldcw WORD PTR cwTemp \
}
#define CHOP_ROUND_ON()
#define CHOP_ROUND_OFF()
#if DBG
#define ASSERT_CHOP_ROUND() \
{ \
WORD cw; \
__asm { \
__asm fnstcw cw \
} \
ASSERT((cw & 0xc00) == 0xc00, "Chop round must be on"); \
}
#else
#define ASSERT_CHOP_ROUND()
#endif
#else // _USE_X86_ASSEMBLY
#define FLT_TO_FIX_SCALE(value_in, scale) \
((INT)((REAL)(value_in) * scale))
#define FLT_TO_UCHAR_SCALE(value_in, scale) \
((UCHAR)((INT)((REAL)(value_in) * scale)))
#define UNSAFE_FLT_TO_FIX(value_in) \
FLT_TO_FIX(value_in)
#define FPU_SAVE_MODE()
#define FPU_RESTORE_MODE()
#define FPU_RESTORE_MODE_NO_EXCEPTIONS()
#define FPU_CHOP_ON()
#define FPU_ROUND_ON()
#define FPU_ROUND_ON_PREC_HI()
#define FPU_PREC_LOW()
#define FPU_PREC_LOW_MASK_EXCEPTIONS()
#define FPU_CHOP_ON_PREC_LOW()
#define FPU_CHOP_OFF_PREC_HI()
#define CHOP_ROUND_ON()
#define CHOP_ROUND_OFF()
#define ASSERT_CHOP_ROUND()
#endif //_USE_X86_ASSEMBLY
#if defined(_USE_X86_ASSEMBLY)
__inline INT __fastcall FLOOR(
REAL a)
{
INT i;
_asm {
fld a
fistp i
}
return i;
}
// Can cause overflow exceptions
__inline INT __fastcall UNSAFE_FLOOR(
REAL a)
{
INT l;
_asm {
fld a
fistp l
}
return l;
}
#else
#define FLOOR(a) ((INT) floor(a))
#define UNSAFE_FLOOR(a) ((INT) floor(a))
#endif
//--------------------------------------------------------------------------
// Class for invoking 'floor' mode, and restoring it afterwards
//--------------------------------------------------------------------------
// These define the bits in the FPU control word that we care about.
// The high byte 0x0c is the rounding control and the low byte
// 0x3F is the exception mask flags.
#define FP_CTRL_MASK 0x0c3F
// Set the rounding control and mask PE. DE is no longer masked because
// taking the exception can be helpful in finding bugs. However, we should
// probably turn on DE masking when we ship.
// Round Down
#define FP_CTRL_ROUNDDOWN 0x0400
// Mask PE only.
// Use this define to track down nasty FP bugs.
// #define FP_CTRL_MASKEDEXCEPTIONS 0x0020
// Mask all FP exceptions.
#define FP_CTRL_MASKEDEXCEPTIONS 0x003F
#define FP_CTRL_STATE (FP_CTRL_ROUNDDOWN | FP_CTRL_MASKEDEXCEPTIONS)
class FPUStateSaver
{
private:
UINT32 SavedState;
#if DBG
// For AssertMode, we use SaveLevel to keep track
// of the nesting level of FPUState. This is not thread safe.
// At worst, that could cause spurious asserts; more likely it could
// fail to catch some errors. To prevent the spurious asserts,
// used interlocked instructions to modify it. To fix it properly,
// we'd need to use per-thread storage.
static LONG SaveLevel;
#endif
public:
FPUStateSaver()
{
// Here we set our thread's round mode to 'round down', and
// save the old mode. 'Round to nearest' is actually the
// default state for a process, but applications can set it
// to whatever they like, and we should also respect and save
// their preferred mode.
#if defined(_USE_X86_ASSEMBLY)
UINT32 tempState;
UINT32 savedState;
// clear the current exception state.
// fnclex is a non-wait version of fclex - which means it clears
// the exceptions without taking any pending unmasked exceptions.
// We issue a fclex so that any unmasked exceptions are
// triggered immediately. This indicates a bug in the caller
// of the API.
_asm fclex
// Save control word:
_asm fnstcw WORD PTR savedState
this->SavedState = savedState;
// Floor mode on & set up our prefered exception masks
_asm mov eax, savedState
_asm and eax, ~FP_CTRL_MASK
_asm or eax, FP_CTRL_STATE
_asm mov tempState, eax
_asm fldcw WORD PTR tempState
#endif
#if DBG
InterlockedIncrement(&SaveLevel);
#endif
}
~FPUStateSaver()
{
AssertMode();
#if DBG
InterlockedDecrement(&SaveLevel);
#endif
#if defined(_USE_X86_ASSEMBLY)
UINT32 savedState = this->SavedState;
// Clear the exception state.
// Note: We issue the fwait and then an fnclex - which is equivalent
// to the fclex instruction (9B DB E2) which causes us to
// immediately take any unmasked pending exceptions.
// Because we clear the exception state on the way in, hitting
// an exception on this line means we generated an exception
// in our code (or a call out to other code between the
// FPUStateSaver constructor and destructor nesting) which was
// pending and not handled.
_asm fclex
// Restore control word (rounding mode and exception masks):
_asm fldcw WORD PTR savedState
#endif
}
// AssertMode.
//
// AssertMode does nothing in Free builds, unless the FREE_BUILD_FP_BARRIER
// define is set to 1. Debug builds always have FP barriers.
// If exceptions are unmasked and you're getting delayed FP exceptions,
// turn on the FP barriers and add FPUStateSaver::AssertMode() calls
// bracketing all your calls. This will allow you to isolate the FP
// exception generator.
#define FREE_BUILD_FP_BARRIER 0
#if DBG
static VOID AssertMode();
#else
static VOID AssertMode()
{
#if defined(_USE_X86_ASSEMBLY)
#if FREE_BUILD_FP_BARRIER
_asm fwait
#endif
#endif
}
#endif
//--------------------------------------------------------------------------
// The following conversion routines are MUCH faster than the standard
// C routines, because they assume that the FPU floating point state is
// properly set.
//
// However, because they assume that the FPU floating point state has been
// properly set, they can only be used if an instance of the
// 'FPUStateSaver' class is in scope.
//--------------------------------------------------------------------------
static INT Floor(REAL x)
{
AssertMode();
return((INT) FLOOR(x));
}
static INT Trunc(REAL x)
{
AssertMode();
return (x>=0) ? FLOOR(x) : -FLOOR(-x);
}
static INT Ceiling(REAL x)
{
AssertMode();
return((INT) -FLOOR(-x));
}
static INT Round(REAL x)
{
AssertMode();
return((INT) FLOOR((x) + TOREAL(0.5)));
}
// Saturation versions of the above conversion routines. Don't test for
// equality to INT_MAX because, when converted to floating-point for the
// comparison, the value is (INT_MAX + 1):
#define SATURATE(op) \
static INT op##Sat(REAL x) \
{ \
return (x >= INT_MIN) ? ((x < INT_MAX) ? op(x) \
: INT_MAX) \
: INT_MIN; \
}
SATURATE(Floor);
SATURATE(Trunc);
SATURATE(Ceiling);
SATURATE(Round);
#undef SATURATE
};
// FPUStateSandbox
//
// This object is designed to sandbox FPU unsafe code.
// For example, many badly written printer drivers on win9x codebases
// manipulate the FPU state without restoring it on exit. In order to
// prevent code like that from hosing us, we wrap calls to potentially
// unsafe code (like driver escapes) in an FPUStateSandbox.
//
// This will guarantee that after calling the unsafe code, the FPU state
// (rounding mode and exceptions) are reset to our preferred state.
// Because we assume that we're restoring to our preferred state, we
// ASSERT on our preferred state being set on entry. This means that
// the sandbox must be declared inside some top level FPUStateSaver block.
// This condition is not strictly necessary and if there is a requirement
// for an FPUStateSandbox not contained inside an FPUStateSaver, this
// ASSERT can be removed. The sandbox saves the current state and restores
// it on exit, so it can operate outside of our preferred state if required.
//
// So far we've found a number of printer drivers on win9x codebase that
// require sandboxing - e.g. HP4500c pcl.
//
// Couple of caveats: This code is designed to wrap simple calls out of
// GDI+ to the printer driver, such as Escape. It's not intended to be
// nested or for use with GDI+ code. However, nesting will work. In
// particular you should not call FPUStateSaver functions inside of
// an FPUStateSandbox unless you've acquired another nested FPUStateSaver.
// The only anticipated need for this is for sandboxing a callback function
// that calls into our API again. In this case it's ok, because all the
// GDI+ calls will be wrapped by a nested FPUStateSaver acquired at the
// API.
//
// NOTE: The ASSERTs in GpRound will not catch the scenario when they're
// called inside of a sandbox and not properly surrounded by a FPUStateSaver.
// GpRound may work incorrectly inside of a sandbox because the unsafe code
// could change the rounding mode. It could also generate exceptions.
class FPUStateSandbox
{
private:
UINT32 SavedState;
public:
FPUStateSandbox()
{
// it is assumed that this call is issued inside of an
// FPUStateSaver block, so that the CTRL word is set to
// our preferred state.
// Lets not do this ASSERT - turns out that it gets called on the device
// destructor durning InternalGdiplusShutdown and we don't want to wrap
// that with an FPUStateSaver.
// FPUStateSaver::AssertMode();
#if defined(_USE_X86_ASSEMBLY)
UINT32 savedState;
// We must protect the sandboxed code from clearing the exception
// masks and taking an exception generated by GDI+.
// We do this by issuing fclex - which takes any unmasked exceptions
// and clears all of the exceptions after that (masked and unmasked)
// giving the sandboxed code a clean slate.
_asm fclex
// Save control word:
_asm fnstcw WORD PTR savedState
this->SavedState = savedState;
#endif
}
~FPUStateSandbox()
{
#if defined(_USE_X86_ASSEMBLY)
UINT32 savedState = this->SavedState;
// clear the current exception state.
// fnclex is a non-wait version of fclex - which means it clears
// the exceptions without taking any pending unmasked exceptions.
// We issue an fnclex so that any unmasked exceptions are ignored.
// This is designed to prevent the sandboxed code from blowing
// up in code outside of the sandbox.
_asm fnclex
// Restore control word (rounding mode and exception masks):
_asm fldcw WORD PTR savedState
#endif
}
};
//--------------------------------------------------------------------------
// The following are simply handy versions that require less typing to
// use (i.e., 'GpFloor(x)' instead of 'FPUStateSaver::Floor(x)').
//
// These functions require that a version of 'FPUStateSaver' has been
// instantiated already for the current thread.
//--------------------------------------------------------------------------
inline INT GpFloor(REAL x) { return(FPUStateSaver::Floor(x)); }
inline INT GpTrunc(REAL x) { return(FPUStateSaver::Trunc(x)); }
inline INT GpCeiling(REAL x) { return(FPUStateSaver::Ceiling(x)); }
inline INT GpRound(REAL x) { return(FPUStateSaver::Round(x)); }
inline INT GpFloorSat(REAL x) { return(FPUStateSaver::FloorSat(x)); }
inline INT GpTruncSat(REAL x) { return(FPUStateSaver::TruncSat(x)); }
inline INT GpCeilingSat(REAL x) { return(FPUStateSaver::CeilingSat(x)); }
inline INT GpRoundSat(REAL x) { return(FPUStateSaver::RoundSat(x)); }
/**************************************************************************\
*
* Function Description:
*
* Return TRUE if two points are close. Close is defined as near enough
* that the rounding to 32bit float precision could have resulted in the
* difference. We define an arbitrary number of allowed rounding errors (10).
* We divide by b to normalize the difference. It doesn't matter which point
* we divide by - if they're significantly different, we'll return true, and
* if they're really close, then a==b (almost).
*
* Arguments:
*
* a, b - input numbers to compare.
*
* Return Value:
*
* TRUE if the numbers are close enough.
*
* Created:
*
* 12/11/2000 asecchia
*
\**************************************************************************/
inline BOOL IsCloseReal(const REAL a, const REAL b)
{
// if b == 0.0f we don't want to divide by zero. If this happens
// it's sufficient to use 1.0 as the divisor because REAL_EPSILON
// should be good enough to test if a number is close enough to zero.
// NOTE: if b << a, this could cause an FP overflow. Currently we mask
// these exceptions, but if we unmask them, we should probably check
// the divide.
// We assume we can generate an overflow exception without taking down
// the system. We will still get the right results based on the FPU
// default handling of the overflow.
#if !(FP_CTRL_STATE & 0x8)
// Ensure that anyone clearing the overflow mask comes and revisits this
// assumption.
#error #O exception cleared. Go check FP_CTRL_MASKEDEXCEPTIONS.
#endif
FPUStateSaver::AssertMode();
return( REALABS( (a-b) / ((b==0.0f)?1.0f:b) ) < 10.0f*REAL_EPSILON );
}
inline BOOL IsClosePointF(const PointF &pt1, const PointF &pt2)
{
return (
IsCloseReal(pt1.X, pt2.X) &&
IsCloseReal(pt1.Y, pt2.Y)
);
}
#endif // !_REAL_HPP_