fpucw.h 4.7 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108
  1. /* Manipulating the FPU control word. -*- coding: utf-8 -*-
  2. Copyright (C) 2007-2020 Free Software Foundation, Inc.
  3. Written by Bruno Haible <bruno@clisp.org>, 2007.
  4. This program is free software: you can redistribute it and/or modify
  5. it under the terms of the GNU General Public License as published by
  6. the Free Software Foundation; either version 3 of the License, or
  7. (at your option) any later version.
  8. This program is distributed in the hope that it will be useful,
  9. but WITHOUT ANY WARRANTY; without even the implied warranty of
  10. MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
  11. GNU General Public License for more details.
  12. You should have received a copy of the GNU General Public License
  13. along with this program. If not, see <https://www.gnu.org/licenses/>. */
  14. #ifndef _FPUCW_H
  15. #define _FPUCW_H
  16. /* The i386 floating point hardware (the 387 compatible FPU, not the modern
  17. SSE/SSE2 hardware) has a controllable rounding precision. It is specified
  18. through the 'PC' bits in the FPU control word ('fctrl' register). (See
  19. the GNU libc i386 <fpu_control.h> header for details.)
  20. On some platforms, such as Linux or Solaris, the default precision setting
  21. is set to "extended precision". This means that 'long double' instructions
  22. operate correctly, but 'double' computations often produce slightly
  23. different results as on strictly IEEE 754 conforming systems.
  24. On some platforms, such as NetBSD, the default precision is set to
  25. "double precision". This means that 'long double' instructions will operate
  26. only as 'double', i.e. lead to wrong results. Similarly on FreeBSD 6.4, at
  27. least for the division of 'long double' numbers.
  28. The FPU control word is under control of the application, i.e. it is
  29. not required to be set either way by the ABI. (In fact, the i386 ABI
  30. https://www.linux-mips.org/pub/linux/mips/doc/ABI/abi386-4.pdf page 3-12 = page 38
  31. is not clear about it. But in any case, gcc treats the control word
  32. like a "preserved" register: it emits code that assumes that the control
  33. word is preserved across calls, and it restores the control word at the
  34. end of functions that modify it.)
  35. See Vincent Lefèvre's page https://www.vinc17.net/research/extended.en.html
  36. for a good explanation.
  37. See https://web.archive.org/web/20060905133417/http://www.uwsg.iu.edu/hypermail/linux/kernel/0103.0/0453.html
  38. some argumentation which setting should be the default. */
  39. /* This header file provides the following facilities:
  40. fpucw_t integral type holding the value of 'fctrl'
  41. FPU_PC_MASK bit mask denoting the precision control
  42. FPU_PC_DOUBLE precision control for 53 bits mantissa
  43. FPU_PC_EXTENDED precision control for 64 bits mantissa
  44. GET_FPUCW () yields the current FPU control word
  45. SET_FPUCW (word) sets the FPU control word
  46. DECL_LONG_DOUBLE_ROUNDING variable declaration for
  47. BEGIN/END_LONG_DOUBLE_ROUNDING
  48. BEGIN_LONG_DOUBLE_ROUNDING () starts a sequence of instructions with
  49. 'long double' safe operation precision
  50. END_LONG_DOUBLE_ROUNDING () ends a sequence of instructions with
  51. 'long double' safe operation precision
  52. */
  53. /* Inline assembler like this works only with GNU C and clang. */
  54. #if (defined __i386__ || defined __x86_64__) && (defined __GNUC__ || defined __clang__)
  55. typedef unsigned short fpucw_t; /* glibc calls this fpu_control_t */
  56. # define FPU_PC_MASK 0x0300
  57. # define FPU_PC_DOUBLE 0x200 /* glibc calls this _FPU_DOUBLE */
  58. # define FPU_PC_EXTENDED 0x300 /* glibc calls this _FPU_EXTENDED */
  59. # define GET_FPUCW() __extension__ \
  60. ({ fpucw_t _cw; \
  61. __asm__ __volatile__ ("fnstcw %0" : "=m" (*&_cw)); \
  62. _cw; \
  63. })
  64. # define SET_FPUCW(word) __extension__ \
  65. (void)({ fpucw_t _ncw = (word); \
  66. __asm__ __volatile__ ("fldcw %0" : : "m" (*&_ncw)); \
  67. })
  68. # define DECL_LONG_DOUBLE_ROUNDING \
  69. fpucw_t oldcw;
  70. # define BEGIN_LONG_DOUBLE_ROUNDING() \
  71. (void)(oldcw = GET_FPUCW (), \
  72. SET_FPUCW ((oldcw & ~FPU_PC_MASK) | FPU_PC_EXTENDED))
  73. # define END_LONG_DOUBLE_ROUNDING() \
  74. SET_FPUCW (oldcw)
  75. #else
  76. typedef unsigned int fpucw_t;
  77. # define FPU_PC_MASK 0
  78. # define FPU_PC_DOUBLE 0
  79. # define FPU_PC_EXTENDED 0
  80. # define GET_FPUCW() 0
  81. # define SET_FPUCW(word) (void)(word)
  82. # define DECL_LONG_DOUBLE_ROUNDING
  83. # define BEGIN_LONG_DOUBLE_ROUNDING()
  84. # define END_LONG_DOUBLE_ROUNDING()
  85. #endif
  86. #endif /* _FPUCW_H */