huwcampbell.com
Reactos
[CRT] Fix amd64 floating point control functions
+243
-32
+4
sdk/include/crt/float.h
+4
sdk/include/crt/float.h
···
59
59
#define _PC_24 0x00020000
60
60
#define _PC_53 0x00010000
61
61
#define _PC_64 0x00000000
62
+
#define _DN_SAVE 0x00000000
63
+
#define _DN_FLUSH 0x01000000
64
+
#define _DN_FLUSH_OPERANDS_SAVE_RESULTS 0x02000000
65
+
#define _DN_SAVE_OPERANDS_FLUSH_RESULTS 0x03000000
62
66
63
67
/* These are also defined in Mingw math.h, needed to work around
64
68
GCC build issues. */
+23
sdk/lib/crt/float/amd64/_clearfp.c
+23
sdk/lib/crt/float/amd64/_clearfp.c
···
1
+
/*
2
+
* PROJECT: ReactOS CRT
3
+
* LICENSE: MIT (https://spdx.org/licenses/MIT)
4
+
* PURPOSE: x64 implementation of _clearfp
5
+
* COPYRIGHT: Copyright 2022 Timo Kreuzer <timo.kreuzer@reactos.org>
6
+
*/
7
+
8
+
#include <float.h>
9
+
#include <xmmintrin.h>
10
+
11
+
unsigned int __cdecl _clearfp(void)
12
+
{
13
+
unsigned int retval;
14
+
15
+
/* Get current status value */
16
+
retval = _statusfp();
17
+
18
+
/* Clear the exception mask */
19
+
_mm_setcsr(_mm_getcsr() & ~_MM_EXCEPT_MASK);
20
+
21
+
/* Return the previous state */
22
+
return retval;
23
+
}
+39
sdk/lib/crt/float/amd64/_control87.c
+39
sdk/lib/crt/float/amd64/_control87.c
···
1
+
/*
2
+
* PROJECT: ReactOS CRT
3
+
* LICENSE: MIT (https://spdx.org/licenses/MIT)
4
+
* PURPOSE: Implementation of _control87
5
+
* COPYRIGHT: Copyright 2022 Timo Kreuzer <timo.kreuzer@reactos.org>
6
+
*/
7
+
8
+
#include <xmmintrin.h>
9
+
#include <float.h>
10
+
11
+
unsigned int _get_native_fpcw(void);
12
+
void _set_native_fpcw(unsigned int value);
13
+
unsigned int _fpcw_native_to_abstract(unsigned int native);
14
+
unsigned int _fpcw_abstract_to_native(unsigned int abstract);
15
+
16
+
unsigned int __cdecl _control87(unsigned int newval, unsigned int mask)
17
+
{
18
+
unsigned int native, oldval, updated;
19
+
20
+
/* Sanatize the mask */
21
+
mask &= _MCW_DN | _MCW_EM | _MCW_RC;
22
+
23
+
/* Get native control word */
24
+
native = _get_native_fpcw();
25
+
26
+
/* Convert to abstract */
27
+
oldval = _fpcw_native_to_abstract(native);
28
+
29
+
/* Update it according to the given parameters */
30
+
updated = (oldval & ~mask) | (newval & mask);
31
+
32
+
/* Convert back to native */
33
+
native = _fpcw_abstract_to_native(updated);
34
+
35
+
/* Set the native value */
36
+
_set_native_fpcw(native);
37
+
38
+
return updated;
39
+
}
+13
sdk/lib/crt/float/amd64/_controlfp.c
+13
sdk/lib/crt/float/amd64/_controlfp.c
···
1
+
/*
2
+
* PROJECT: ReactOS CRT
3
+
* LICENSE: MIT (https://spdx.org/licenses/MIT)
4
+
* PURPOSE: x64 implementation of _controlfp
5
+
* COPYRIGHT: Copyright 2022 Timo Kreuzer <timo.kreuzer@reactos.org>
6
+
*/
7
+
8
+
#include <float.h>
9
+
10
+
unsigned int __cdecl _controlfp(unsigned int newval, unsigned int mask)
11
+
{
12
+
return _control87(newval, mask & ~_EM_DENORMAL);
13
+
}
+14
sdk/lib/crt/float/amd64/_fpreset.c
+14
sdk/lib/crt/float/amd64/_fpreset.c
···
1
+
/*
2
+
* PROJECT: ReactOS CRT
3
+
* LICENSE: MIT (https://spdx.org/licenses/MIT)
4
+
* PURPOSE: x64 implementation of _fpreset
5
+
* COPYRIGHT: Copyright 2022 Timo Kreuzer <timo.kreuzer@reactos.org>
6
+
*/
7
+
8
+
#include <xmmintrin.h>
9
+
10
+
void __cdecl _fpreset(void)
11
+
{
12
+
/* Mask everything */
13
+
_mm_setcsr(_MM_MASK_MASK);
14
+
}
+33
sdk/lib/crt/float/amd64/_statusfp.c
+33
sdk/lib/crt/float/amd64/_statusfp.c
···
1
+
/*
2
+
* PROJECT: ReactOS CRT
3
+
* LICENSE: MIT (https://spdx.org/licenses/MIT)
4
+
* PURPOSE: x64 implementation of _statusfp
5
+
* COPYRIGHT: Copyright 2022 Timo Kreuzer <timo.kreuzer@reactos.org>
6
+
*/
7
+
8
+
#include <float.h>
9
+
#include <xmmintrin.h>
10
+
11
+
unsigned int __cdecl _statusfp(void)
12
+
{
13
+
unsigned int mxcsr, status = 0;
14
+
15
+
/* Get MXCSR */
16
+
mxcsr = _mm_getcsr();
17
+
18
+
/* Convert to abstract status flags */
19
+
if (mxcsr & _MM_EXCEPT_INVALID)
20
+
status |= _SW_INVALID;
21
+
if (mxcsr & _MM_EXCEPT_DENORM)
22
+
status |= _SW_DENORMAL;
23
+
if (mxcsr & _MM_EXCEPT_DIV_ZERO)
24
+
status |= _SW_ZERODIVIDE;
25
+
if (mxcsr & _MM_EXCEPT_OVERFLOW)
26
+
status |= _SW_OVERFLOW;
27
+
if (mxcsr & _MM_EXCEPT_UNDERFLOW)
28
+
status |= _SW_UNDERFLOW;
29
+
if (mxcsr & _MM_EXCEPT_INEXACT)
30
+
status |= _SW_INEXACT;
31
+
32
+
return status;
33
+
}
-15
sdk/lib/crt/float/amd64/clearfp.S
-15
sdk/lib/crt/float/amd64/clearfp.S
-13
sdk/lib/crt/float/amd64/fpreset.S
-13
sdk/lib/crt/float/amd64/fpreset.S
+111
sdk/lib/crt/float/amd64/machfpcw.c
+111
sdk/lib/crt/float/amd64/machfpcw.c
···
1
+
/*
2
+
* PROJECT: ReactOS CRT
3
+
* LICENSE: MIT (https://spdx.org/licenses/MIT)
4
+
* PURPOSE: Implementation of x64 floating point control word helper functions
5
+
* COPYRIGHT: Copyright 2022 Timo Kreuzer <timo.kreuzer@reactos.org>
6
+
*/
7
+
8
+
#include <float.h>
9
+
#include <xmmintrin.h>
10
+
11
+
#define _MM_DENORMALS_ARE_ZERO 0x0040
12
+
13
+
unsigned int _get_native_fpcw(void)
14
+
{
15
+
return _mm_getcsr();
16
+
}
17
+
18
+
void _set_native_fpcw(unsigned int value)
19
+
{
20
+
_mm_setcsr(value);
21
+
}
22
+
23
+
unsigned int _fpcw_native_to_abstract(unsigned int native)
24
+
{
25
+
unsigned int rounding_mask, abstract = 0;
26
+
27
+
/* Handle exception mask */
28
+
if (native & _MM_MASK_INVALID)
29
+
abstract |= _EM_INVALID;
30
+
if (native & _MM_MASK_DENORM)
31
+
abstract |= _EM_DENORMAL;
32
+
if (native & _MM_MASK_DIV_ZERO)
33
+
abstract |= _EM_ZERODIVIDE;
34
+
if (native & _MM_MASK_OVERFLOW)
35
+
abstract |= _EM_OVERFLOW;
36
+
if (native & _MM_MASK_UNDERFLOW)
37
+
abstract |= _EM_UNDERFLOW;
38
+
if (native & _MM_MASK_INEXACT)
39
+
abstract |= _EM_INEXACT;
40
+
41
+
/* Handle rounding mode */
42
+
rounding_mask = (native & _MM_ROUND_MASK);
43
+
if (rounding_mask == _MM_ROUND_DOWN)
44
+
abstract |= _RC_DOWN;
45
+
else if(rounding_mask == _MM_ROUND_UP)
46
+
abstract |= _RC_UP;
47
+
else if (rounding_mask == _MM_ROUND_TOWARD_ZERO)
48
+
abstract |= _RC_CHOP;
49
+
50
+
/* Handle denormal control */
51
+
if (native & _MM_DENORMALS_ARE_ZERO)
52
+
{
53
+
if (native & _MM_FLUSH_ZERO_MASK)
54
+
abstract |= _DN_FLUSH;
55
+
else
56
+
abstract |= _DN_FLUSH_OPERANDS_SAVE_RESULTS;
57
+
}
58
+
else
59
+
{
60
+
if (native & _MM_FLUSH_ZERO_MASK)
61
+
abstract |= _DN_SAVE_OPERANDS_FLUSH_RESULTS;
62
+
else
63
+
abstract |= _DN_SAVE;
64
+
}
65
+
66
+
return abstract;
67
+
}
68
+
69
+
unsigned int _fpcw_abstract_to_native(unsigned int abstract)
70
+
{
71
+
unsigned int rounding_mask, native = 0;
72
+
73
+
/* Handle exception mask */
74
+
if (abstract & _EM_INVALID)
75
+
native |= _MM_MASK_INVALID;
76
+
if (abstract & _EM_DENORMAL)
77
+
native |= _MM_MASK_DENORM;
78
+
if (abstract & _EM_ZERODIVIDE)
79
+
native |= _MM_MASK_DIV_ZERO;
80
+
if (abstract & _EM_OVERFLOW)
81
+
native |= _MM_MASK_OVERFLOW;
82
+
if (abstract & _EM_UNDERFLOW)
83
+
native |= _MM_MASK_UNDERFLOW;
84
+
if (abstract & _EM_INEXACT)
85
+
native |= _MM_MASK_INEXACT;
86
+
87
+
/* Handle rounding mode */
88
+
rounding_mask = (abstract & _MCW_RC);
89
+
if (rounding_mask == _RC_DOWN)
90
+
native |= _MM_ROUND_DOWN;
91
+
else if (rounding_mask == _RC_UP)
92
+
native |= _MM_ROUND_UP;
93
+
else if (rounding_mask == _RC_CHOP)
94
+
native |= _MM_ROUND_TOWARD_ZERO;
95
+
96
+
/* Handle Denormal Control */
97
+
if ((abstract & _MCW_DN) == _DN_FLUSH)
98
+
{
99
+
native |= _MM_DENORMALS_ARE_ZERO | _MM_FLUSH_ZERO_MASK;
100
+
}
101
+
else if ((abstract & _MCW_DN) == _DN_FLUSH_OPERANDS_SAVE_RESULTS)
102
+
{
103
+
native |= _MM_DENORMALS_ARE_ZERO;
104
+
}
105
+
else if ((abstract & _MCW_DN) == _DN_SAVE_OPERANDS_FLUSH_RESULTS)
106
+
{
107
+
native |= _MM_FLUSH_ZERO_MASK;
108
+
}
109
+
110
+
return native;
111
+
}
+6
-4
sdk/lib/crt/float/float.cmake
+6
-4
sdk/lib/crt/float/float.cmake
···
23
23
)
24
24
elseif(ARCH STREQUAL "amd64")
25
25
list(APPEND CRT_FLOAT_SOURCE
26
-
float/i386/cntrlfp.c
27
-
float/i386/statfp.c
26
+
float/amd64/_clearfp.c
27
+
float/amd64/_control87.c
28
+
float/amd64/_controlfp.c
29
+
float/amd64/_fpreset.c
30
+
float/amd64/_statusfp.c
31
+
float/amd64/machfpcw.c
28
32
)
29
33
list(APPEND CRT_FLOAT_ASM_SOURCE
30
-
float/amd64/clearfp.S
31
34
float/amd64/getsetfpcw.S
32
-
float/amd64/fpreset.S
33
35
float/amd64/logb.S
34
36
)
35
37
elseif(ARCH STREQUAL "arm")