From 1d6e2e1833864238b903325b37d05fef9b794393 Mon Sep 17 00:00:00 2001 From: Mark Dickinson Date: Sat, 24 Oct 2009 13:28:38 +0000 Subject: Issue #7117 (backport py3k float repr) continued: - add double endianness detection to configure script - add configure-time check to see whether we can use inline assembly to get and set x87 control word in configure script - add functions to get and set x87 control word in Python/pymath.c - add pyport.h logic to determine whether it's safe to use the short float repr or not --- Include/pymacconfig.h | 6 + Include/pymath.h | 5 + Include/pyport.h | 74 ++++++++++++ PC/pyconfig.h | 4 + Python/pymath.c | 18 +++ configure | 305 +++++++++++++++++++++++++++++++++++++++++++++++++- configure.in | 112 +++++++++++++++++- pyconfig.h.in | 16 +++ 8 files changed, 529 insertions(+), 11 deletions(-) diff --git a/Include/pymacconfig.h b/Include/pymacconfig.h index b2cc0b7..3a3428c 100644 --- a/Include/pymacconfig.h +++ b/Include/pymacconfig.h @@ -17,6 +17,9 @@ # undef SIZEOF_VOID_P # undef SIZEOF__BOOL # undef WORDS_BIGENDIAN +# undef DOUBLE_IS_ARM_MIXED_ENDIAN_IEEE754 +# undef DOUBLE_IS_BIG_ENDIAN_IEEE754 +# undef DOUBLE_IS_LITTLE_ENDIAN_IEEE754 # undef VA_LIST_IS_ARRAY # if defined(__LP64__) && defined(__x86_64__) @@ -65,6 +68,9 @@ #ifdef __BIG_ENDIAN__ #define WORDS_BIGENDIAN 1 +#define DOUBLE_IS_BIG_ENDIAN_IEEE754 +#else +#define DOUBLE_IS_LITTLE_ENDIAN_IEEE754 #endif /* __BIG_ENDIAN */ /* diff --git a/Include/pymath.h b/Include/pymath.h index 7623efa..dc2c427 100644 --- a/Include/pymath.h +++ b/Include/pymath.h @@ -92,6 +92,11 @@ PyAPI_FUNC(double) _Py_force_double(double); # endif #endif +#ifdef HAVE_GCC_ASM_FOR_X87 +PyAPI_FUNC(unsigned short) _Py_get_387controlword(void); +PyAPI_FUNC(void) _Py_set_387controlword(unsigned short); +#endif + /* Py_IS_NAN(X) * Return 1 if float or double arg is a NaN, else 0. * Caution: diff --git a/Include/pyport.h b/Include/pyport.h index da83196..62d4524 100644 --- a/Include/pyport.h +++ b/Include/pyport.h @@ -488,6 +488,80 @@ extern "C" { errno = 0; \ } while(0) +/* The functions _Py_dg_strtod and _Py_dg_dtoa in Python/dtoa.c (which are + * required to support the short float repr introduced in Python 3.1) require + * that the floating-point unit that's being used for arithmetic operations + * on C doubles is set to use 53-bit precision. It also requires that the + * FPU rounding mode is round-half-to-even, but that's less often an issue. + * + * If your FPU isn't already set to 53-bit precision/round-half-to-even, and + * you want to make use of _Py_dg_strtod and _Py_dg_dtoa, then you should + * + * #define HAVE_PY_SET_53BIT_PRECISION 1 + * + * and also give appropriate definitions for the following three macros: + * + * _PY_SET_53BIT_PRECISION_START : store original FPU settings, and + * set FPU to 53-bit precision/round-half-to-even + * _PY_SET_53BIT_PRECISION_END : restore original FPU settings + * _PY_SET_53BIT_PRECISION_HEADER : any variable declarations needed to + * use the two macros above. + * + * The macros are designed to be used within a single C function: see + * Python/pystrtod.c for an example of their use. + */ + +/* get and set x87 control word for gcc/x86 */ +#ifdef HAVE_GCC_ASM_FOR_X87 +#define HAVE_PY_SET_53BIT_PRECISION 1 +/* _Py_get/set_387controlword functions are defined in Python/pymath.c */ +#define _Py_SET_53BIT_PRECISION_HEADER \ + unsigned short old_387controlword, new_387controlword +#define _Py_SET_53BIT_PRECISION_START \ + do { \ + old_387controlword = _Py_get_387controlword(); \ + new_387controlword = (old_387controlword & ~0x0f00) | 0x0200; \ + if (new_387controlword != old_387controlword) \ + _Py_set_387controlword(new_387controlword); \ + } while (0) +#define _Py_SET_53BIT_PRECISION_END \ + if (new_387controlword != old_387controlword) \ + _Py_set_387controlword(old_387controlword) +#endif + +/* default definitions are empty */ +#ifndef HAVE_PY_SET_53BIT_PRECISION +#define _Py_SET_53BIT_PRECISION_HEADER +#define _Py_SET_53BIT_PRECISION_START +#define _Py_SET_53BIT_PRECISION_END +#endif + +/* If we can't guarantee 53-bit precision, don't use the code + in Python/dtoa.c, but fall back to standard code. This + means that repr of a float will be long (17 sig digits). + + Realistically, there are two things that could go wrong: + + (1) doubles aren't IEEE 754 doubles, or + (2) we're on x86 with the rounding precision set to 64-bits + (extended precision), and we don't know how to change + the rounding precision. + */ + +#if !defined(DOUBLE_IS_LITTLE_ENDIAN_IEEE754) && \ + !defined(DOUBLE_IS_BIG_ENDIAN_IEEE754) && \ + !defined(DOUBLE_IS_ARM_MIXED_ENDIAN_IEEE754) +#define PY_NO_SHORT_FLOAT_REPR +#endif + +/* double rounding is symptomatic of use of extended precision on x86. If + we're seeing double rounding, and we don't have any mechanism available for + changing the FPU rounding precision, then don't use Python/dtoa.c. */ +#if defined(X87_DOUBLE_ROUNDING) && !defined(HAVE_PY_SET_53BIT_PRECISION) +#define PY_NO_SHORT_FLOAT_REPR +#endif + + /* Py_DEPRECATED(version) * Declare a variable, type, or function deprecated. * Usage: diff --git a/PC/pyconfig.h b/PC/pyconfig.h index d4524c8..c9cd2c0 100644 --- a/PC/pyconfig.h +++ b/PC/pyconfig.h @@ -749,4 +749,8 @@ Py_NO_ENABLE_SHARED to find out. Also support MS_NO_COREDLL for b/w compat */ socket handles greater than FD_SETSIZE */ #define Py_SOCKET_FD_CAN_BE_GE_FD_SETSIZE +/* Define if C doubles are 64-bit IEEE 754 binary format, stored with the + least significant byte first */ +#define DOUBLE_IS_LITTLE_ENDIAN_IEEE754 1 + #endif /* !Py_CONFIG_H */ diff --git a/Python/pymath.c b/Python/pymath.c index 6438058..db2920c 100644 --- a/Python/pymath.c +++ b/Python/pymath.c @@ -13,6 +13,24 @@ double _Py_force_double(double x) } #endif +#ifdef HAVE_GCC_ASM_FOR_X87 + +/* inline assembly for getting and setting the 387 FPU control word on + gcc/x86 */ + +unsigned short _Py_get_387controlword(void) { + unsigned short cw; + __asm__ __volatile__ ("fnstcw %0" : "=m" (cw)); + return cw; +} + +void _Py_set_387controlword(unsigned short cw) { + __asm__ __volatile__ ("fldcw %0" : : "m" (cw)); +} + +#endif + + #ifndef HAVE_HYPOT double hypot(double x, double y) { diff --git a/configure b/configure index 82f60be..e774d5e 100755 --- a/configure +++ b/configure @@ -1,5 +1,5 @@ #! /bin/sh -# From configure.in Revision: 74978 . +# From configure.in Revision: 75131 . # Guess values for system-dependent variables and create Makefiles. # Generated by GNU Autoconf 2.61 for python 2.7. # @@ -23233,12 +23233,298 @@ echo "${ECHO_T}default LIBC=\"$LIBC\"" >&6; } fi -# ************************************ -# * Check for mathematical functions * -# ************************************ +# ************************************************** +# * Check for various properties of floating point * +# ************************************************** -LIBS_SAVE=$LIBS -LIBS="$LIBS $LIBM" +{ echo "$as_me:$LINENO: checking whether C doubles are little-endian IEEE 754 binary64" >&5 +echo $ECHO_N "checking whether C doubles are little-endian IEEE 754 binary64... $ECHO_C" >&6; } +if test "${ac_cv_little_endian_double+set}" = set; then + echo $ECHO_N "(cached) $ECHO_C" >&6 +else + +if test "$cross_compiling" = yes; then + ac_cv_little_endian_double=no +else + cat >conftest.$ac_ext <<_ACEOF +/* confdefs.h. */ +_ACEOF +cat confdefs.h >>conftest.$ac_ext +cat >>conftest.$ac_ext <<_ACEOF +/* end confdefs.h. */ + +#include +int main() { + double x = 9006104071832581.0; + if (memcmp(&x, "\x05\x04\x03\x02\x01\xff\x3f\x43", 8) == 0) + return 0; + else + return 1; +} + +_ACEOF +rm -f conftest$ac_exeext +if { (ac_try="$ac_link" +case "(($ac_try" in + *\"* | *\`* | *\\*) ac_try_echo=\$ac_try;; + *) ac_try_echo=$ac_try;; +esac +eval "echo \"\$as_me:$LINENO: $ac_try_echo\"") >&5 + (eval "$ac_link") 2>&5 + ac_status=$? + echo "$as_me:$LINENO: \$? = $ac_status" >&5 + (exit $ac_status); } && { ac_try='./conftest$ac_exeext' + { (case "(($ac_try" in + *\"* | *\`* | *\\*) ac_try_echo=\$ac_try;; + *) ac_try_echo=$ac_try;; +esac +eval "echo \"\$as_me:$LINENO: $ac_try_echo\"") >&5 + (eval "$ac_try") 2>&5 + ac_status=$? + echo "$as_me:$LINENO: \$? = $ac_status" >&5 + (exit $ac_status); }; }; then + ac_cv_little_endian_double=yes +else + echo "$as_me: program exited with status $ac_status" >&5 +echo "$as_me: failed program was:" >&5 +sed 's/^/| /' conftest.$ac_ext >&5 + +( exit $ac_status ) +ac_cv_little_endian_double=no +fi +rm -f core *.core core.conftest.* gmon.out bb.out conftest$ac_exeext conftest.$ac_objext conftest.$ac_ext +fi + + +fi + +{ echo "$as_me:$LINENO: result: $ac_cv_little_endian_double" >&5 +echo "${ECHO_T}$ac_cv_little_endian_double" >&6; } +if test "$ac_cv_little_endian_double" = yes +then + +cat >>confdefs.h <<\_ACEOF +#define DOUBLE_IS_LITTLE_ENDIAN_IEEE754 1 +_ACEOF + +fi + +{ echo "$as_me:$LINENO: checking whether C doubles are big-endian IEEE 754 binary64" >&5 +echo $ECHO_N "checking whether C doubles are big-endian IEEE 754 binary64... $ECHO_C" >&6; } +if test "${ac_cv_big_endian_double+set}" = set; then + echo $ECHO_N "(cached) $ECHO_C" >&6 +else + +if test "$cross_compiling" = yes; then + ac_cv_big_endian_double=no +else + cat >conftest.$ac_ext <<_ACEOF +/* confdefs.h. */ +_ACEOF +cat confdefs.h >>conftest.$ac_ext +cat >>conftest.$ac_ext <<_ACEOF +/* end confdefs.h. */ + +#include +int main() { + double x = 9006104071832581.0; + if (memcmp(&x, "\x43\x3f\xff\x01\x02\x03\x04\x05", 8) == 0) + return 0; + else + return 1; +} + +_ACEOF +rm -f conftest$ac_exeext +if { (ac_try="$ac_link" +case "(($ac_try" in + *\"* | *\`* | *\\*) ac_try_echo=\$ac_try;; + *) ac_try_echo=$ac_try;; +esac +eval "echo \"\$as_me:$LINENO: $ac_try_echo\"") >&5 + (eval "$ac_link") 2>&5 + ac_status=$? + echo "$as_me:$LINENO: \$? = $ac_status" >&5 + (exit $ac_status); } && { ac_try='./conftest$ac_exeext' + { (case "(($ac_try" in + *\"* | *\`* | *\\*) ac_try_echo=\$ac_try;; + *) ac_try_echo=$ac_try;; +esac +eval "echo \"\$as_me:$LINENO: $ac_try_echo\"") >&5 + (eval "$ac_try") 2>&5 + ac_status=$? + echo "$as_me:$LINENO: \$? = $ac_status" >&5 + (exit $ac_status); }; }; then + ac_cv_big_endian_double=yes +else + echo "$as_me: program exited with status $ac_status" >&5 +echo "$as_me: failed program was:" >&5 +sed 's/^/| /' conftest.$ac_ext >&5 + +( exit $ac_status ) +ac_cv_big_endian_double=no +fi +rm -f core *.core core.conftest.* gmon.out bb.out conftest$ac_exeext conftest.$ac_objext conftest.$ac_ext +fi + + +fi + +{ echo "$as_me:$LINENO: result: $ac_cv_big_endian_double" >&5 +echo "${ECHO_T}$ac_cv_big_endian_double" >&6; } +if test "$ac_cv_big_endian_double" = yes +then + +cat >>confdefs.h <<\_ACEOF +#define DOUBLE_IS_BIG_ENDIAN_IEEE754 1 +_ACEOF + +fi + +# Some ARM platforms use a mixed-endian representation for doubles. +# While Python doesn't currently have full support for these platforms +# (see e.g., issue 1762561), we can at least make sure that float <-> string +# conversions work. +{ echo "$as_me:$LINENO: checking whether C doubles are ARM mixed-endian IEEE 754 binary64" >&5 +echo $ECHO_N "checking whether C doubles are ARM mixed-endian IEEE 754 binary64... $ECHO_C" >&6; } +if test "${ac_cv_mixed_endian_double+set}" = set; then + echo $ECHO_N "(cached) $ECHO_C" >&6 +else + +if test "$cross_compiling" = yes; then + ac_cv_mixed_endian_double=no +else + cat >conftest.$ac_ext <<_ACEOF +/* confdefs.h. */ +_ACEOF +cat confdefs.h >>conftest.$ac_ext +cat >>conftest.$ac_ext <<_ACEOF +/* end confdefs.h. */ + +#include +int main() { + double x = 9006104071832581.0; + if (memcmp(&x, "\x01\xff\x3f\x43\x05\x04\x03\x02", 8) == 0) + return 0; + else + return 1; +} + +_ACEOF +rm -f conftest$ac_exeext +if { (ac_try="$ac_link" +case "(($ac_try" in + *\"* | *\`* | *\\*) ac_try_echo=\$ac_try;; + *) ac_try_echo=$ac_try;; +esac +eval "echo \"\$as_me:$LINENO: $ac_try_echo\"") >&5 + (eval "$ac_link") 2>&5 + ac_status=$? + echo "$as_me:$LINENO: \$? = $ac_status" >&5 + (exit $ac_status); } && { ac_try='./conftest$ac_exeext' + { (case "(($ac_try" in + *\"* | *\`* | *\\*) ac_try_echo=\$ac_try;; + *) ac_try_echo=$ac_try;; +esac +eval "echo \"\$as_me:$LINENO: $ac_try_echo\"") >&5 + (eval "$ac_try") 2>&5 + ac_status=$? + echo "$as_me:$LINENO: \$? = $ac_status" >&5 + (exit $ac_status); }; }; then + ac_cv_mixed_endian_double=yes +else + echo "$as_me: program exited with status $ac_status" >&5 +echo "$as_me: failed program was:" >&5 +sed 's/^/| /' conftest.$ac_ext >&5 + +( exit $ac_status ) +ac_cv_mixed_endian_double=no +fi +rm -f core *.core core.conftest.* gmon.out bb.out conftest$ac_exeext conftest.$ac_objext conftest.$ac_ext +fi + + +fi + +{ echo "$as_me:$LINENO: result: $ac_cv_mixed_endian_double" >&5 +echo "${ECHO_T}$ac_cv_mixed_endian_double" >&6; } +if test "$ac_cv_mixed_endian_double" = yes +then + +cat >>confdefs.h <<\_ACEOF +#define DOUBLE_IS_ARM_MIXED_ENDIAN_IEEE754 1 +_ACEOF + +fi + +# The short float repr introduced in Python 3.1 requires the +# correctly-rounded string <-> double conversion functions from +# Python/dtoa.c, which in turn require that the FPU uses 53-bit +# rounding; this is a problem on x86, where the x87 FPU has a default +# rounding precision of 64 bits. For gcc/x86, we try to fix this by +# using inline assembler to get and set the x87 FPU control word. +if test "$GCC" = yes && test -n "`$CC -dM -E - &5 +echo $ECHO_N "checking whether we can use gcc inline assembler to get and set x87 control word... $ECHO_C" >&6; } + cat >conftest.$ac_ext <<_ACEOF +/* confdefs.h. */ +_ACEOF +cat confdefs.h >>conftest.$ac_ext +cat >>conftest.$ac_ext <<_ACEOF +/* end confdefs.h. */ + +int +main () +{ + + unsigned short cw; + __asm__ __volatile__ ("fnstcw %0" : "=m" (cw)); + __asm__ __volatile__ ("fldcw %0" : : "m" (cw)); + + ; + return 0; +} +_ACEOF +rm -f conftest.$ac_objext +if { (ac_try="$ac_compile" +case "(($ac_try" in + *\"* | *\`* | *\\*) ac_try_echo=\$ac_try;; + *) ac_try_echo=$ac_try;; +esac +eval "echo \"\$as_me:$LINENO: $ac_try_echo\"") >&5 + (eval "$ac_compile") 2>conftest.er1 + ac_status=$? + grep -v '^ *+' conftest.er1 >conftest.err + rm -f conftest.er1 + cat conftest.err >&5 + echo "$as_me:$LINENO: \$? = $ac_status" >&5 + (exit $ac_status); } && { + test -z "$ac_c_werror_flag" || + test ! -s conftest.err + } && test -s conftest.$ac_objext; then + have_gcc_asm_for_x87=yes +else + echo "$as_me: failed program was:" >&5 +sed 's/^/| /' conftest.$ac_ext >&5 + + have_gcc_asm_for_x87=no +fi + +rm -f core conftest.err conftest.$ac_objext conftest.$ac_ext + { echo "$as_me:$LINENO: result: $have_gcc_asm_for_x87" >&5 +echo "${ECHO_T}$have_gcc_asm_for_x87" >&6; } + if test "$have_gcc_asm_for_x87" = yes + then + +cat >>confdefs.h <<\_ACEOF +#define HAVE_GCC_ASM_FOR_X87 1 +_ACEOF + + fi +fi # Detect whether system arithmetic is subject to x87-style double # rounding issues. The result of this test has little meaning on non @@ -23327,6 +23613,13 @@ _ACEOF fi +# ************************************ +# * Check for mathematical functions * +# ************************************ + +LIBS_SAVE=$LIBS +LIBS="$LIBS $LIBM" + # Multiprocessing check for broken sem_getvalue { echo "$as_me:$LINENO: checking for broken sem_getvalue" >&5 echo $ECHO_N "checking for broken sem_getvalue... $ECHO_C" >&6; } diff --git a/configure.in b/configure.in index 4713d57..28d0c80 100644 --- a/configure.in +++ b/configure.in @@ -3240,12 +3240,107 @@ else AC_MSG_ERROR([proper usage is --with-libc=STRING]) fi], [AC_MSG_RESULT(default LIBC="$LIBC")]) -# ************************************ -# * Check for mathematical functions * -# ************************************ +# ************************************************** +# * Check for various properties of floating point * +# ************************************************** -LIBS_SAVE=$LIBS -LIBS="$LIBS $LIBM" +AC_MSG_CHECKING(whether C doubles are little-endian IEEE 754 binary64) +AC_CACHE_VAL(ac_cv_little_endian_double, [ +AC_TRY_RUN([ +#include +int main() { + double x = 9006104071832581.0; + if (memcmp(&x, "\x05\x04\x03\x02\x01\xff\x3f\x43", 8) == 0) + return 0; + else + return 1; +} +], +ac_cv_little_endian_double=yes, +ac_cv_little_endian_double=no, +ac_cv_little_endian_double=no)]) +AC_MSG_RESULT($ac_cv_little_endian_double) +if test "$ac_cv_little_endian_double" = yes +then + AC_DEFINE(DOUBLE_IS_LITTLE_ENDIAN_IEEE754, 1, + [Define if C doubles are 64-bit IEEE 754 binary format, stored + with the least significant byte first]) +fi + +AC_MSG_CHECKING(whether C doubles are big-endian IEEE 754 binary64) +AC_CACHE_VAL(ac_cv_big_endian_double, [ +AC_TRY_RUN([ +#include +int main() { + double x = 9006104071832581.0; + if (memcmp(&x, "\x43\x3f\xff\x01\x02\x03\x04\x05", 8) == 0) + return 0; + else + return 1; +} +], +ac_cv_big_endian_double=yes, +ac_cv_big_endian_double=no, +ac_cv_big_endian_double=no)]) +AC_MSG_RESULT($ac_cv_big_endian_double) +if test "$ac_cv_big_endian_double" = yes +then + AC_DEFINE(DOUBLE_IS_BIG_ENDIAN_IEEE754, 1, + [Define if C doubles are 64-bit IEEE 754 binary format, stored + with the most significant byte first]) +fi + +# Some ARM platforms use a mixed-endian representation for doubles. +# While Python doesn't currently have full support for these platforms +# (see e.g., issue 1762561), we can at least make sure that float <-> string +# conversions work. +AC_MSG_CHECKING(whether C doubles are ARM mixed-endian IEEE 754 binary64) +AC_CACHE_VAL(ac_cv_mixed_endian_double, [ +AC_TRY_RUN([ +#include +int main() { + double x = 9006104071832581.0; + if (memcmp(&x, "\x01\xff\x3f\x43\x05\x04\x03\x02", 8) == 0) + return 0; + else + return 1; +} +], +ac_cv_mixed_endian_double=yes, +ac_cv_mixed_endian_double=no, +ac_cv_mixed_endian_double=no)]) +AC_MSG_RESULT($ac_cv_mixed_endian_double) +if test "$ac_cv_mixed_endian_double" = yes +then + AC_DEFINE(DOUBLE_IS_ARM_MIXED_ENDIAN_IEEE754, 1, + [Define if C doubles are 64-bit IEEE 754 binary format, stored + in ARM mixed-endian order (byte order 45670123)]) +fi + +# The short float repr introduced in Python 3.1 requires the +# correctly-rounded string <-> double conversion functions from +# Python/dtoa.c, which in turn require that the FPU uses 53-bit +# rounding; this is a problem on x86, where the x87 FPU has a default +# rounding precision of 64 bits. For gcc/x86, we try to fix this by +# using inline assembler to get and set the x87 FPU control word. +if test "$GCC" = yes && test -n "`$CC -dM -E -