OLD | NEW |
---|---|
1 /* | 1 /* |
2 * Copyright (C) 2011, Google Inc. All rights reserved. | 2 * Copyright (C) 2011, Google Inc. All rights reserved. |
3 * | 3 * |
4 * Redistribution and use in source and binary forms, with or without | 4 * Redistribution and use in source and binary forms, with or without |
5 * modification, are permitted provided that the following conditions | 5 * modification, are permitted provided that the following conditions |
6 * are met: | 6 * are met: |
7 * 1. Redistributions of source code must retain the above copyright | 7 * 1. Redistributions of source code must retain the above copyright |
8 * notice, this list of conditions and the following disclaimer. | 8 * notice, this list of conditions and the following disclaimer. |
9 * 2. Redistributions in binary form must reproduce the above copyright | 9 * 2. Redistributions in binary form must reproduce the above copyright |
10 * notice, this list of conditions and the following disclaimer in the | 10 * notice, this list of conditions and the following disclaimer in the |
(...skipping 16 matching lines...) Expand all Loading... | |
27 | 27 |
28 #include "wtf/CPU.h" | 28 #include "wtf/CPU.h" |
29 #include "wtf/MathExtras.h" | 29 #include "wtf/MathExtras.h" |
30 #include <float.h> | 30 #include <float.h> |
31 | 31 |
32 namespace WebCore { | 32 namespace WebCore { |
33 | 33 |
34 // Deal with denormals. They can very seriously impact performance on x86. | 34 // Deal with denormals. They can very seriously impact performance on x86. |
35 | 35 |
36 // Define HAVE_DENORMAL if we support flushing denormals to zero. | 36 // Define HAVE_DENORMAL if we support flushing denormals to zero. |
37 #if OS(WIN) && COMPILER(MSVC) | 37 |
38 #if OS(WIN) && COMPILER(MSVC) && (_M_IX86_FP == 2) | |
39 // Windows compiled using MSVC with SSE2 (_M_IX86_FP == 2) can flush denormals. | |
Ken Russell (switch to Gerrit)
2014/07/18 18:46:56
A little searching indicates that _M_IX86_FP isn't
Raymond Toy
2014/07/18 20:08:34
Should just probably get rid of this check, now th
| |
38 #define HAVE_DENORMAL 1 | 40 #define HAVE_DENORMAL 1 |
39 #endif | 41 #endif |
40 | 42 |
41 #if COMPILER(GCC) && (CPU(X86) || CPU(X86_64)) | 43 #if COMPILER(GCC) && (CPU(X86) || CPU(X86_64)) |
44 // X86 chips can flush denormals | |
42 #define HAVE_DENORMAL 1 | 45 #define HAVE_DENORMAL 1 |
43 #endif | 46 #endif |
44 | 47 |
48 #if CPU(ARM) || CPU(ARM64) | |
49 #define HAVE_DENORMAL 1 | |
50 #endif | |
51 | |
45 #if HAVE(DENORMAL) | 52 #if HAVE(DENORMAL) |
46 class DenormalDisabler { | 53 class DenormalDisabler { |
47 public: | 54 public: |
48 DenormalDisabler() | 55 DenormalDisabler() |
49 : m_savedCSR(0) | 56 : m_savedCSR(0) |
50 { | 57 { |
51 #if OS(WIN) && COMPILER(MSVC) | 58 disableDenormals(); |
52 // Save the current state, and set mode to flush denormals. | |
53 // | |
54 // http://stackoverflow.com/questions/637175/possible-bug-in-controlfp-s -may-not-restore-control-word-correctly | |
55 _controlfp_s(&m_savedCSR, 0, 0); | |
56 unsigned unused; | |
57 _controlfp_s(&unused, _DN_FLUSH, _MCW_DN); | |
58 #else | |
59 m_savedCSR = getCSR(); | |
60 setCSR(m_savedCSR | 0x8040); | |
61 #endif | |
62 } | 59 } |
63 | 60 |
64 ~DenormalDisabler() | 61 ~DenormalDisabler() |
65 { | 62 { |
66 #if OS(WIN) && COMPILER(MSVC) | 63 restoreState(); |
67 unsigned unused; | |
68 _controlfp_s(&unused, m_savedCSR, _MCW_DN); | |
69 #else | |
70 setCSR(m_savedCSR); | |
71 #endif | |
72 } | 64 } |
73 | 65 |
74 // This is a nop if we can flush denormals to zero in hardware. | 66 // This is a nop if we can flush denormals to zero in hardware. |
75 static inline float flushDenormalFloatToZero(float f) | 67 static inline float flushDenormalFloatToZero(float f) |
76 { | 68 { |
77 #if OS(WIN) && COMPILER(MSVC) && (!_M_IX86_FP) | |
78 // For systems using x87 instead of sse, there's no hardware support | |
79 // to flush denormals automatically. Hence, we need to flush | |
80 // denormals to zero manually. | |
81 return (fabs(f) < FLT_MIN) ? 0.0f : f; | |
82 #else | |
83 return f; | 69 return f; |
84 #endif | |
85 } | 70 } |
86 private: | 71 private: |
72 unsigned m_savedCSR; | |
73 | |
87 #if COMPILER(GCC) && (CPU(X86) || CPU(X86_64)) | 74 #if COMPILER(GCC) && (CPU(X86) || CPU(X86_64)) |
75 inline void disableDenormals() | |
76 { | |
77 m_savedCSR = getCSR(); | |
78 setCSR(m_savedCSR | 0x8040); | |
79 } | |
80 | |
81 inline void restoreState() | |
82 { | |
83 setCSR(m_savedCSR); | |
84 } | |
85 | |
88 inline int getCSR() | 86 inline int getCSR() |
89 { | 87 { |
90 int result; | 88 int result; |
91 asm volatile("stmxcsr %0" : "=m" (result)); | 89 asm volatile("stmxcsr %0" : "=m" (result)); |
92 return result; | 90 return result; |
93 } | 91 } |
94 | 92 |
95 inline void setCSR(int a) | 93 inline void setCSR(int a) |
96 { | 94 { |
97 int temp = a; | 95 int temp = a; |
98 asm volatile("ldmxcsr %0" : : "m" (temp)); | 96 asm volatile("ldmxcsr %0" : : "m" (temp)); |
99 } | 97 } |
100 | 98 |
99 #elif OS(WIN) && COMPILER(MSVC) | |
Ken Russell (switch to Gerrit)
2014/07/18 18:46:56
The conditions in this #elif don't match those in
| |
100 inline void disableDenormals() | |
101 { | |
102 // Save the current state, and set mode to flush denormals. | |
103 // | |
104 // http://stackoverflow.com/questions/637175/possible-bug-in-controlfp-s -may-not-restore-control-word-correctly | |
105 _controlfp_s(&m_savedCSR, 0, 0); | |
106 unsigned unused; | |
107 _controlfp_s(&unused, _DN_FLUSH, _MCW_DN); | |
108 } | |
109 | |
110 inline void restoreState() | |
111 { | |
112 unsigned unused; | |
113 _controlfp_s(&unused, m_savedCSR, _MCW_DN); | |
114 } | |
115 #elif CPU(ARM) || CPU(ARM64) | |
116 inline void disableDenormals() | |
117 { | |
118 m_savedCSR = getStatusWord(); | |
119 // Bit 24 is the flush-to-zero mode control bit. Setting it to 1 flushes denormals to 0. | |
120 setStatusWord(m_savedCSR | (1 << 24)); | |
121 } | |
122 | |
123 inline void restoreState() | |
124 { | |
125 setStatusWord(m_savedCSR); | |
126 } | |
127 | |
128 inline int getStatusWord() | |
129 { | |
130 int result; | |
131 #if CPU(ARM64) | |
132 asm volatile("mrs %[result], FPCR" : [result] "=r" (result)); | |
133 #else | |
134 asm volatile("vmrs %[result], FPSCR" : [result] "=r" (result)); | |
135 #endif | |
136 return result; | |
137 } | |
138 | |
139 inline void setStatusWord(int a) | |
140 { | |
141 #if CPU(ARM64) | |
142 asm volatile("msr FPCR, %[src]" : : [src] "r" (a)); | |
143 #else | |
144 asm volatile("vmsr FPSCR, %[src]" : : [src] "r" (a)); | |
145 #endif | |
146 } | |
147 | |
101 #endif | 148 #endif |
102 | 149 |
103 unsigned m_savedCSR; | |
104 }; | 150 }; |
105 | 151 |
106 #else | 152 #else |
107 // FIXME: add implementations for other architectures and compilers | 153 // FIXME: add implementations for other architectures and compilers |
108 class DenormalDisabler { | 154 class DenormalDisabler { |
109 public: | 155 public: |
110 DenormalDisabler() { } | 156 DenormalDisabler() { } |
111 | 157 |
112 // Assume the worst case that other architectures and compilers | 158 // Assume the worst case that other architectures and compilers |
113 // need to flush denormals to zero manually. | 159 // need to flush denormals to zero manually. |
114 static inline float flushDenormalFloatToZero(float f) | 160 static inline float flushDenormalFloatToZero(float f) |
115 { | 161 { |
116 return (fabs(f) < FLT_MIN) ? 0.0f : f; | 162 return (fabs(f) < FLT_MIN) ? 0.0f : f; |
117 } | 163 } |
118 }; | 164 }; |
119 | 165 |
120 #endif | 166 #endif |
121 | 167 |
122 } // WebCore | 168 } // WebCore |
123 | 169 |
124 #undef HAVE_DENORMAL | 170 #undef HAVE_DENORMAL |
125 #endif // DenormalDisabler_h | 171 #endif // DenormalDisabler_h |
OLD | NEW |