1 //===-- x86_64 floating point env manipulation functions --------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8
9 #ifndef LLVM_LIBC_SRC___SUPPORT_FPUTIL_X86_64_FENVIMPL_H
10 #define LLVM_LIBC_SRC___SUPPORT_FPUTIL_X86_64_FENVIMPL_H
11
12 #include "src/__support/macros/attributes.h" // LIBC_INLINE
13 #include "src/__support/macros/config.h"
14 #include "src/__support/macros/properties/architectures.h"
15
16 #if !defined(LIBC_TARGET_ARCH_IS_X86)
17 #error "Invalid include"
18 #endif
19
20 #include <stdint.h>
21
22 #include "hdr/types/fenv_t.h"
23 #include "src/__support/macros/sanitizer.h"
24
25 namespace LIBC_NAMESPACE_DECL {
26 namespace fputil {
27
28 namespace internal {
29
30 // Normally, one should be able to define FE_* macros to the exact rounding mode
31 // encodings. However, since we want LLVM libc to be compiled against headers
32 // from other libcs, we cannot assume that FE_* macros are always defined in
33 // such a manner. So, we will define enums corresponding to the x86_64 bit
34 // encodings. The implementations can map from FE_* to the corresponding enum
35 // values.
36
37 // The rounding control values in the x87 control register and the MXCSR
38 // register have the same 2-bit enoding but have different bit positions.
39 // See below for the bit positions.
40 struct RoundingControlValue {
41 static constexpr uint16_t TO_NEAREST = 0x0;
42 static constexpr uint16_t DOWNWARD = 0x1;
43 static constexpr uint16_t UPWARD = 0x2;
44 static constexpr uint16_t TOWARD_ZERO = 0x3;
45 };
46
47 static constexpr uint16_t X87_ROUNDING_CONTROL_BIT_POSITION = 10;
48 static constexpr uint16_t MXCSR_ROUNDING_CONTROL_BIT_POSITION = 13;
49
50 // The exception flags in the x87 status register and the MXCSR have the same
51 // encoding as well as the same bit positions.
52 struct ExceptionFlags {
53 static constexpr uint16_t INVALID_F = 0x1;
54 // Some libcs define __FE_DENORM corresponding to the denormal input
55 // exception and include it in FE_ALL_EXCEPTS. We define and use it to
56 // support compiling against headers provided by such libcs.
57 static constexpr uint16_t DENORMAL_F = 0x2;
58 static constexpr uint16_t DIV_BY_ZERO_F = 0x4;
59 static constexpr uint16_t OVERFLOW_F = 0x8;
60 static constexpr uint16_t UNDERFLOW_F = 0x10;
61 static constexpr uint16_t INEXACT_F = 0x20;
62 };
63
64 // The exception control bits occupy six bits, one bit for each exception.
65 // In the x87 control word, they occupy the first 6 bits. In the MXCSR
66 // register, they occupy bits 7 to 12.
67 static constexpr uint16_t X87_EXCEPTION_CONTROL_BIT_POSITION = 0;
68 static constexpr uint16_t X87_EXCEPTION_CONTROL_BIT_POSITION_HIGH = 24;
69 static constexpr uint16_t MXCSR_EXCEPTION_CONTOL_BIT_POISTION = 7;
70
71 // Exception flags are individual bits in the corresponding registers.
72 // So, we just OR the bit values to get the full set of exceptions.
get_status_value_for_except(int excepts)73 LIBC_INLINE uint16_t get_status_value_for_except(int excepts) {
74 // We will make use of the fact that exception control bits are single
75 // bit flags in the control registers.
76 return ((excepts & FE_INVALID) ? ExceptionFlags::INVALID_F : 0) |
77 #ifdef __FE_DENORM
78 ((excepts & __FE_DENORM) ? ExceptionFlags::DENORMAL_F : 0) |
79 #endif // __FE_DENORM
80 ((excepts & FE_DIVBYZERO) ? ExceptionFlags::DIV_BY_ZERO_F : 0) |
81 ((excepts & FE_OVERFLOW) ? ExceptionFlags::OVERFLOW_F : 0) |
82 ((excepts & FE_UNDERFLOW) ? ExceptionFlags::UNDERFLOW_F : 0) |
83 ((excepts & FE_INEXACT) ? ExceptionFlags::INEXACT_F : 0);
84 }
85
exception_status_to_macro(uint16_t status)86 LIBC_INLINE int exception_status_to_macro(uint16_t status) {
87 return ((status & ExceptionFlags::INVALID_F) ? FE_INVALID : 0) |
88 #ifdef __FE_DENORM
89 ((status & ExceptionFlags::DENORMAL_F) ? __FE_DENORM : 0) |
90 #endif // __FE_DENORM
91 ((status & ExceptionFlags::DIV_BY_ZERO_F) ? FE_DIVBYZERO : 0) |
92 ((status & ExceptionFlags::OVERFLOW_F) ? FE_OVERFLOW : 0) |
93 ((status & ExceptionFlags::UNDERFLOW_F) ? FE_UNDERFLOW : 0) |
94 ((status & ExceptionFlags::INEXACT_F) ? FE_INEXACT : 0);
95 }
96
97 struct X87StateDescriptor {
98 uint16_t control_word;
99 uint16_t unused1;
100 uint16_t status_word;
101 uint16_t unused2;
102 // TODO: Elaborate the remaining 20 bytes as required.
103 uint32_t _[5];
104 };
105
get_x87_control_word()106 LIBC_INLINE uint16_t get_x87_control_word() {
107 uint16_t w;
108 __asm__ __volatile__("fnstcw %0" : "=m"(w)::);
109 MSAN_UNPOISON(&w, sizeof(w));
110 return w;
111 }
112
write_x87_control_word(uint16_t w)113 LIBC_INLINE void write_x87_control_word(uint16_t w) {
114 __asm__ __volatile__("fldcw %0" : : "m"(w) :);
115 }
116
get_x87_status_word()117 LIBC_INLINE uint16_t get_x87_status_word() {
118 uint16_t w;
119 __asm__ __volatile__("fnstsw %0" : "=m"(w)::);
120 MSAN_UNPOISON(&w, sizeof(w));
121 return w;
122 }
123
clear_x87_exceptions()124 LIBC_INLINE void clear_x87_exceptions() {
125 __asm__ __volatile__("fnclex" : : :);
126 }
127
get_mxcsr()128 LIBC_INLINE uint32_t get_mxcsr() {
129 uint32_t w;
130 __asm__ __volatile__("stmxcsr %0" : "=m"(w)::);
131 MSAN_UNPOISON(&w, sizeof(w));
132 return w;
133 }
134
write_mxcsr(uint32_t w)135 LIBC_INLINE void write_mxcsr(uint32_t w) {
136 __asm__ __volatile__("ldmxcsr %0" : : "m"(w) :);
137 }
138
get_x87_state_descriptor(X87StateDescriptor & s)139 LIBC_INLINE void get_x87_state_descriptor(X87StateDescriptor &s) {
140 __asm__ __volatile__("fnstenv %0" : "=m"(s));
141 MSAN_UNPOISON(&s, sizeof(s));
142 }
143
write_x87_state_descriptor(const X87StateDescriptor & s)144 LIBC_INLINE void write_x87_state_descriptor(const X87StateDescriptor &s) {
145 __asm__ __volatile__("fldenv %0" : : "m"(s) :);
146 }
147
fwait()148 LIBC_INLINE void fwait() { __asm__ __volatile__("fwait"); }
149
150 } // namespace internal
151
enable_except(int excepts)152 LIBC_INLINE int enable_except(int excepts) {
153 // In the x87 control word and in MXCSR, an exception is blocked
154 // if the corresponding bit is set. That is the reason for all the
155 // bit-flip operations below as we need to turn the bits to zero
156 // to enable them.
157
158 uint16_t bit_mask = internal::get_status_value_for_except(excepts);
159
160 uint16_t x87_cw = internal::get_x87_control_word();
161 uint16_t old_excepts = ~x87_cw & 0x3F; // Save previously enabled exceptions.
162 x87_cw &= ~bit_mask;
163 internal::write_x87_control_word(x87_cw);
164
165 // Enabling SSE exceptions via MXCSR is a nice thing to do but
166 // might not be of much use practically as SSE exceptions and the x87
167 // exceptions are independent of each other.
168 uint32_t mxcsr = internal::get_mxcsr();
169 mxcsr &= ~(bit_mask << internal::MXCSR_EXCEPTION_CONTOL_BIT_POISTION);
170 internal::write_mxcsr(mxcsr);
171
172 // Since the x87 exceptions and SSE exceptions are independent of each,
173 // it doesn't make much sence to report both in the return value. Most
174 // often, the standard floating point functions deal with FPU operations
175 // so we will retrun only the old x87 exceptions.
176 return internal::exception_status_to_macro(old_excepts);
177 }
178
disable_except(int excepts)179 LIBC_INLINE int disable_except(int excepts) {
180 // In the x87 control word and in MXCSR, an exception is blocked
181 // if the corresponding bit is set.
182
183 uint16_t bit_mask = internal::get_status_value_for_except(excepts);
184
185 uint16_t x87_cw = internal::get_x87_control_word();
186 uint16_t old_excepts = ~x87_cw & 0x3F; // Save previously enabled exceptions.
187 x87_cw |= bit_mask;
188 internal::write_x87_control_word(x87_cw);
189
190 // Just like in enable_except, it is not clear if disabling SSE exceptions
191 // is required. But, we will still do it only as a "nice thing to do".
192 uint32_t mxcsr = internal::get_mxcsr();
193 mxcsr |= (bit_mask << internal::MXCSR_EXCEPTION_CONTOL_BIT_POISTION);
194 internal::write_mxcsr(mxcsr);
195
196 return internal::exception_status_to_macro(old_excepts);
197 }
198
get_except()199 LIBC_INLINE int get_except() {
200 uint16_t mxcsr = static_cast<uint16_t>(internal::get_mxcsr());
201 uint16_t enabled_excepts = ~(mxcsr >> 7) & 0x3F;
202 return internal::exception_status_to_macro(enabled_excepts);
203 }
204
clear_except(int excepts)205 LIBC_INLINE int clear_except(int excepts) {
206 internal::X87StateDescriptor state;
207 internal::get_x87_state_descriptor(state);
208 state.status_word &=
209 static_cast<uint16_t>(~internal::get_status_value_for_except(excepts));
210 internal::write_x87_state_descriptor(state);
211
212 uint32_t mxcsr = internal::get_mxcsr();
213 mxcsr &= ~internal::get_status_value_for_except(excepts);
214 internal::write_mxcsr(mxcsr);
215 return 0;
216 }
217
test_except(int excepts)218 LIBC_INLINE int test_except(int excepts) {
219 uint16_t status_word = internal::get_x87_status_word();
220 uint32_t mxcsr = internal::get_mxcsr();
221 // Check both x87 status word and MXCSR.
222 uint16_t status_value = internal::get_status_value_for_except(excepts);
223 return internal::exception_status_to_macro(
224 static_cast<uint16_t>(status_value & (status_word | mxcsr)));
225 }
226
227 // Sets the exception flags but does not trigger the exception handler.
set_except(int excepts)228 LIBC_INLINE int set_except(int excepts) {
229 uint16_t status_value = internal::get_status_value_for_except(excepts);
230 internal::X87StateDescriptor state;
231 internal::get_x87_state_descriptor(state);
232 state.status_word |= status_value;
233 internal::write_x87_state_descriptor(state);
234
235 uint32_t mxcsr = internal::get_mxcsr();
236 mxcsr |= status_value;
237 internal::write_mxcsr(mxcsr);
238
239 return 0;
240 }
241
raise_except(int excepts)242 LIBC_INLINE int raise_except(int excepts) {
243 uint16_t status_value = internal::get_status_value_for_except(excepts);
244
245 // We set the status flag for exception one at a time and call the
246 // fwait instruction to actually get the processor to raise the
247 // exception by calling the exception handler. This scheme is per
248 // the description in "8.6 X87 FPU EXCEPTION SYNCHRONIZATION"
249 // of the "Intel 64 and IA-32 Architectures Software Developer's
250 // Manual, Vol 1".
251
252 // FPU status word is read for each exception separately as the
253 // exception handler can potentially write to it (typically to clear
254 // the corresponding exception flag). By reading it separately, we
255 // ensure that the writes by the exception handler are maintained
256 // when raising the next exception.
257
258 auto raise_helper = [](uint16_t singleExceptFlag) {
259 internal::X87StateDescriptor state;
260 uint32_t mxcsr = 0;
261 internal::get_x87_state_descriptor(state);
262 mxcsr = internal::get_mxcsr();
263 state.status_word |= singleExceptFlag;
264 mxcsr |= singleExceptFlag;
265 internal::write_x87_state_descriptor(state);
266 internal::write_mxcsr(mxcsr);
267 internal::fwait();
268 };
269
270 if (status_value & internal::ExceptionFlags::INVALID_F)
271 raise_helper(internal::ExceptionFlags::INVALID_F);
272 if (status_value & internal::ExceptionFlags::DIV_BY_ZERO_F)
273 raise_helper(internal::ExceptionFlags::DIV_BY_ZERO_F);
274 if (status_value & internal::ExceptionFlags::OVERFLOW_F)
275 raise_helper(internal::ExceptionFlags::OVERFLOW_F);
276 if (status_value & internal::ExceptionFlags::UNDERFLOW_F)
277 raise_helper(internal::ExceptionFlags::UNDERFLOW_F);
278 if (status_value & internal::ExceptionFlags::INEXACT_F)
279 raise_helper(internal::ExceptionFlags::INEXACT_F);
280 #ifdef __FE_DENORM
281 if (status_value & internal::ExceptionFlags::DENORMAL_F) {
282 raise_helper(internal::ExceptionFlags::DENORMAL_F);
283 }
284 #endif // __FE_DENORM
285
286 // There is no special synchronization scheme available to
287 // raise SEE exceptions. So, we will ignore that for now.
288 // Just plain writing to the MXCSR register does not guarantee
289 // the exception handler will be called.
290
291 return 0;
292 }
293
get_round()294 LIBC_INLINE int get_round() {
295 uint16_t bit_value =
296 (internal::get_mxcsr() >> internal::MXCSR_ROUNDING_CONTROL_BIT_POSITION) &
297 0x3;
298 switch (bit_value) {
299 case internal::RoundingControlValue::TO_NEAREST:
300 return FE_TONEAREST;
301 case internal::RoundingControlValue::DOWNWARD:
302 return FE_DOWNWARD;
303 case internal::RoundingControlValue::UPWARD:
304 return FE_UPWARD;
305 case internal::RoundingControlValue::TOWARD_ZERO:
306 return FE_TOWARDZERO;
307 default:
308 return -1; // Error value.
309 }
310 }
311
set_round(int mode)312 LIBC_INLINE int set_round(int mode) {
313 uint16_t bit_value;
314 switch (mode) {
315 case FE_TONEAREST:
316 bit_value = internal::RoundingControlValue::TO_NEAREST;
317 break;
318 case FE_DOWNWARD:
319 bit_value = internal::RoundingControlValue::DOWNWARD;
320 break;
321 case FE_UPWARD:
322 bit_value = internal::RoundingControlValue::UPWARD;
323 break;
324 case FE_TOWARDZERO:
325 bit_value = internal::RoundingControlValue::TOWARD_ZERO;
326 break;
327 default:
328 return 1; // To indicate failure
329 }
330
331 uint16_t x87_value = static_cast<uint16_t>(
332 bit_value << internal::X87_ROUNDING_CONTROL_BIT_POSITION);
333 uint16_t x87_control = internal::get_x87_control_word();
334 x87_control = static_cast<uint16_t>(
335 (x87_control &
336 ~(uint16_t(0x3) << internal::X87_ROUNDING_CONTROL_BIT_POSITION)) |
337 x87_value);
338 internal::write_x87_control_word(x87_control);
339
340 uint32_t mxcsr_value = bit_value
341 << internal::MXCSR_ROUNDING_CONTROL_BIT_POSITION;
342 uint32_t mxcsr_control = internal::get_mxcsr();
343 mxcsr_control = (mxcsr_control &
344 ~(0x3 << internal::MXCSR_ROUNDING_CONTROL_BIT_POSITION)) |
345 mxcsr_value;
346 internal::write_mxcsr(mxcsr_control);
347
348 return 0;
349 }
350
351 namespace internal {
352
353 #if defined(_WIN32)
354 // MSVC fenv.h defines a very simple representation of the floating point state
355 // which just consists of control and status words of the x87 unit.
356 struct FPState {
357 uint32_t control_word;
358 uint32_t status_word;
359 };
360 #elif defined(__APPLE__)
361 struct FPState {
362 uint16_t control_word;
363 uint16_t status_word;
364 uint32_t mxcsr;
365 uint8_t reserved[8];
366 };
367 #else
368 struct FPState {
369 X87StateDescriptor x87_status;
370 uint32_t mxcsr;
371 };
372 #endif // _WIN32
373
374 } // namespace internal
375
376 static_assert(
377 sizeof(fenv_t) == sizeof(internal::FPState),
378 "Internal floating point state does not match the public fenv_t type.");
379
380 #ifdef _WIN32
381
382 // The exception flags in the Windows FEnv struct and the MXCSR have almost
383 // reversed bit positions.
384 struct WinExceptionFlags {
385 static constexpr uint32_t INEXACT_WIN = 0x01;
386 static constexpr uint32_t UNDERFLOW_WIN = 0x02;
387 static constexpr uint32_t OVERFLOW_WIN = 0x04;
388 static constexpr uint32_t DIV_BY_ZERO_WIN = 0x08;
389 static constexpr uint32_t INVALID_WIN = 0x10;
390 static constexpr uint32_t DENORMAL_WIN = 0x20;
391
392 // The Windows FEnv struct has a second copy of all of these bits in the high
393 // byte of the 32 bit control word. These are used as the source of truth when
394 // calling fesetenv.
395 static constexpr uint32_t HIGH_OFFSET = 24;
396
397 static constexpr uint32_t HIGH_INEXACT = INEXACT_WIN << HIGH_OFFSET;
398 static constexpr uint32_t HIGH_UNDERFLOW = UNDERFLOW_WIN << HIGH_OFFSET;
399 static constexpr uint32_t HIGH_OVERFLOW = OVERFLOW_WIN << HIGH_OFFSET;
400 static constexpr uint32_t HIGH_DIV_BY_ZERO = DIV_BY_ZERO_WIN << HIGH_OFFSET;
401 static constexpr uint32_t HIGH_INVALID = INVALID_WIN << HIGH_OFFSET;
402 static constexpr uint32_t HIGH_DENORMAL = DENORMAL_WIN << HIGH_OFFSET;
403 };
404
405 /*
406 fenv_t control word format:
407
408 Windows (at least for x64) uses a 4 byte control fenv control word stored in
409 a 32 bit integer. The first byte contains just the rounding mode and the
410 exception masks, while the last two bytes contain that same information as
411 well as the flush-to-zero and denormals-are-zero flags. The flags are
412 represented with a truth table:
413
414 00 - No flags set
415 01 - Flush-to-zero and Denormals-are-zero set
416 11 - Flush-to-zero set
417 10 - Denormals-are-zero set
418
419 U represents unused.
420
421 +-----Rounding Mode-----+
422 | |
423 ++ ++
424 || ||
425 RRMMMMMM UUUUUUUU UUUUFFRR UUMMMMMM
426 | | || | |
427 +----+ flags---++ +----+
428 | |
429 +------Exception Masks-----+
430
431
432 fenv_t status word format:
433
434 The status word is a lot simpler for this conversion, since only the
435 exception flags are used in the MXCSR.
436
437 +----+---Exception Flags---+----+
438 | | | |
439 UUEEEEEE UUUUUUUU UUUUUUUU UUEEEEEE
440
441
442
443 MXCSR Format:
444
445 The MXCSR format is the same information, just organized differently. Since
446 the fenv_t struct for windows doesn't include the mxcsr bits, they must be
447 generated from the control word bits.
448
449 Exception Masks---+ +---Exception Flags
450 | |
451 Flush-to-zero---+ +----+ +----+
452 | | | | |
453 FRRMMMMMMDEEEEEE
454 || |
455 ++ +---Denormals-are-zero
456 |
457 +---Rounding Mode
458
459
460 The mask and flag order is as follows:
461
462 fenv_t mxcsr
463
464 denormal inexact
465 invalid underflow
466 div by 0 overflow
467 overflow div by 0
468 underflow denormal
469 inexact invalid
470
471 This is almost reverse, except for denormal and invalid which are in the
472 same order in both.
473 */
474
get_env(fenv_t * envp)475 LIBC_INLINE int get_env(fenv_t *envp) {
476 internal::FPState *state = reinterpret_cast<internal::FPState *>(envp);
477
478 uint32_t status_word = 0;
479 uint32_t control_word = 0;
480
481 uint32_t mxcsr = internal::get_mxcsr();
482
483 // Set exception flags in the status word
484 status_word |= (mxcsr & (internal::ExceptionFlags::INVALID_F |
485 internal::ExceptionFlags::DENORMAL_F))
486 << 4;
487 status_word |= (mxcsr & internal::ExceptionFlags::DIV_BY_ZERO_F) << 1;
488 status_word |= (mxcsr & internal::ExceptionFlags::OVERFLOW_F) >> 1;
489 status_word |= (mxcsr & internal::ExceptionFlags::UNDERFLOW_F) >> 3;
490 status_word |= (mxcsr & internal::ExceptionFlags::INEXACT_F) >> 5;
491 status_word |= status_word << WinExceptionFlags::HIGH_OFFSET;
492
493 // Set exception masks in bits 0-5 and 24-29
494 control_word |= (mxcsr & ((internal::ExceptionFlags::INVALID_F |
495 internal::ExceptionFlags::DENORMAL_F)
496 << 7)) >>
497 3;
498 control_word |= (mxcsr & (internal::ExceptionFlags::DIV_BY_ZERO_F << 7)) >> 6;
499 control_word |= (mxcsr & (internal::ExceptionFlags::OVERFLOW_F << 7)) >> 8;
500 control_word |= (mxcsr & (internal::ExceptionFlags::UNDERFLOW_F << 7)) >> 10;
501 control_word |= (mxcsr & (internal::ExceptionFlags::INEXACT_F << 7)) >> 12;
502 control_word |= control_word << WinExceptionFlags::HIGH_OFFSET;
503
504 // Set rounding in bits 8-9 and 30-31
505 control_word |= (mxcsr & 0x6000) >> 5;
506 control_word |= (mxcsr & 0x6000) << 17;
507
508 // Set flush-to-zero in bit 10
509 control_word |= (mxcsr & 0x8000) >> 5;
510
511 // Set denormals-are-zero xor flush-to-zero in bit 11
512 control_word |= (((mxcsr & 0x8000) >> 9) ^ (mxcsr & 0x0040)) << 5;
513
514 state->control_word = control_word;
515 state->status_word = status_word;
516 return 0;
517 }
518
set_env(const fenv_t * envp)519 LIBC_INLINE int set_env(const fenv_t *envp) {
520 const internal::FPState *state =
521 reinterpret_cast<const internal::FPState *>(envp);
522
523 uint32_t mxcsr = 0;
524
525 // Set exception flags from the status word
526 mxcsr |= static_cast<uint16_t>(
527 (state->status_word &
528 (WinExceptionFlags::HIGH_DENORMAL | WinExceptionFlags::HIGH_INVALID)) >>
529 28);
530 mxcsr |= static_cast<uint16_t>(
531 (state->status_word & WinExceptionFlags::HIGH_DIV_BY_ZERO) >> 25);
532 mxcsr |= static_cast<uint16_t>(
533 (state->status_word & WinExceptionFlags::HIGH_OVERFLOW) >> 23);
534 mxcsr |= static_cast<uint16_t>(
535 (state->status_word & WinExceptionFlags::HIGH_UNDERFLOW) >> 21);
536 mxcsr |= static_cast<uint16_t>(
537 (state->status_word & WinExceptionFlags::HIGH_INEXACT) >> 19);
538
539 // Set denormals-are-zero from bit 10 xor bit 11
540 mxcsr |= static_cast<uint16_t>(
541 (((state->control_word & 0x800) >> 1) ^ (state->control_word & 0x400)) >>
542 4);
543
544 // Set exception masks from bits 24-29
545 mxcsr |= static_cast<uint16_t>(
546 (state->control_word &
547 (WinExceptionFlags::HIGH_DENORMAL | WinExceptionFlags::HIGH_INVALID)) >>
548 21);
549 mxcsr |= static_cast<uint16_t>(
550 (state->control_word & WinExceptionFlags::HIGH_DIV_BY_ZERO) >> 18);
551 mxcsr |= static_cast<uint16_t>(
552 (state->control_word & WinExceptionFlags::HIGH_OVERFLOW) >> 16);
553 mxcsr |= static_cast<uint16_t>(
554 (state->control_word & WinExceptionFlags::HIGH_UNDERFLOW) >> 14);
555 mxcsr |= static_cast<uint16_t>(
556 (state->control_word & WinExceptionFlags::HIGH_INEXACT) >> 12);
557
558 // Set rounding from bits 30-31
559 mxcsr |= static_cast<uint16_t>((state->control_word & 0xc0000000) >> 17);
560
561 // Set flush-to-zero from bit 10
562 mxcsr |= static_cast<uint16_t>((state->control_word & 0x400) << 5);
563
564 internal::write_mxcsr(mxcsr);
565 return 0;
566 }
567 #else
get_env(fenv_t * envp)568 LIBC_INLINE int get_env(fenv_t *envp) {
569 internal::FPState *state = reinterpret_cast<internal::FPState *>(envp);
570 #ifdef __APPLE__
571 internal::X87StateDescriptor x87_status;
572 internal::get_x87_state_descriptor(x87_status);
573 state->control_word = x87_status.control_word;
574 state->status_word = x87_status.status_word;
575 #else
576 internal::get_x87_state_descriptor(state->x87_status);
577 #endif // __APPLE__
578 state->mxcsr = internal::get_mxcsr();
579 return 0;
580 }
581
set_env(const fenv_t * envp)582 LIBC_INLINE int set_env(const fenv_t *envp) {
583 // envp contains everything including pieces like the current
584 // top of FPU stack. We cannot arbitrarily change them. So, we first
585 // read the current status and update only those pieces which are
586 // not disruptive.
587 internal::X87StateDescriptor x87_status;
588 internal::get_x87_state_descriptor(x87_status);
589
590 if (envp == FE_DFL_ENV) {
591 // Reset the exception flags in the status word.
592 x87_status.status_word &= ~uint16_t(0x3F);
593 // Reset other non-sensitive parts of the status word.
594 for (int i = 0; i < 5; i++)
595 x87_status._[i] = 0;
596 // In the control word, we do the following:
597 // 1. Mask all exceptions
598 // 2. Set rounding mode to round-to-nearest
599 // 3. Set the internal precision to double extended precision.
600 x87_status.control_word |= uint16_t(0x3F); // Mask all exceptions.
601 x87_status.control_word &= ~(uint16_t(0x3) << 10); // Round to nearest.
602 x87_status.control_word |= (uint16_t(0x3) << 8); // Extended precision.
603 internal::write_x87_state_descriptor(x87_status);
604
605 // We take the exact same approach MXCSR register as well.
606 // MXCSR has two additional fields, "flush-to-zero" and
607 // "denormals-are-zero". We reset those bits. Also, MXCSR does not
608 // have a field which controls the precision of internal operations.
609 uint32_t mxcsr = internal::get_mxcsr();
610 mxcsr &= ~uint16_t(0x3F); // Clear exception flags.
611 mxcsr &= ~(uint16_t(0x1) << 6); // Reset denormals-are-zero
612 mxcsr |= (uint16_t(0x3F) << 7); // Mask exceptions
613 mxcsr &= ~(uint16_t(0x3) << 13); // Round to nearest.
614 mxcsr &= ~(uint16_t(0x1) << 15); // Reset flush-to-zero
615 internal::write_mxcsr(mxcsr);
616
617 return 0;
618 }
619
620 const internal::FPState *fpstate =
621 reinterpret_cast<const internal::FPState *>(envp);
622
623 // Copy the exception status flags from envp.
624 x87_status.status_word &= ~uint16_t(0x3F);
625 #ifdef __APPLE__
626 x87_status.status_word |= (fpstate->status_word & 0x3F);
627 // We can set the x87 control word as is as there no sensitive bits.
628 x87_status.control_word = fpstate->control_word;
629 #else
630 x87_status.status_word |= (fpstate->x87_status.status_word & 0x3F);
631 // Copy other non-sensitive parts of the status word.
632 for (int i = 0; i < 5; i++)
633 x87_status._[i] = fpstate->x87_status._[i];
634 // We can set the x87 control word as is as there no sensitive bits.
635 x87_status.control_word = fpstate->x87_status.control_word;
636 #endif // __APPLE__
637 internal::write_x87_state_descriptor(x87_status);
638
639 // We can write the MXCSR state as is as there are no sensitive bits.
640 internal::write_mxcsr(fpstate->mxcsr);
641 return 0;
642 }
643 #endif
644
645 } // namespace fputil
646 } // namespace LIBC_NAMESPACE_DECL
647
648 #endif // LLVM_LIBC_SRC___SUPPORT_FPUTIL_X86_64_FENVIMPL_H
649