stdcxx-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Travis Vitek (JIRA)" <j...@apache.org>
Subject [jira] Issue Comment Edited: (STDCXX-722) [gcc] use math __builtins
Date Wed, 13 Feb 2008 19:25:10 GMT

    [ https://issues.apache.org/jira/browse/STDCXX-722?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12568684#action_12568684
] 

vitek edited comment on STDCXX-722 at 2/13/08 11:25 AM:
---------------------------------------------------------------

Martin, a few comments on your patch.

{noformat}
+#if 4 <= __GNUG__
+
+// define helpers to call gcc 4.x math builtins
+
+_RWSTD_NAMESPACE (__rw) {
+
+inline float __rw_atan2 (float __x, float __y)
+{
+    return __builtin_atan2f (__x, __y);
+}
+
+inline float __rw_cos (float __x)
+{
+    return __builtin_cosf (__x);
+}

+// use gcc 4 builtins for efficiency and namespace cleanliness
+#  define _RWSTD_ATAN2   _RW::__rw_atan2
+#  define _RWSTD_COS     _RW::__rw_cos
+#  define _RWSTD_COSH    _RW::__rw_cosh
+#  define _RWSTD_EXP     _RW::__rw_exp
+#  define _RWSTD_LOG     _RW::__rw_log
+#  define _RWSTD_POW     _RW::__rw_pow
+#  define _RWSTD_SIN     _RW::__rw_sin
+#  define _RWSTD_SINH    _RW::__rw_sinh
+#  define _RWSTD_SQRT    _RW::__rw_sqrt
+
+#else   // gcc < 4.0
+#  include _RWSTD_CMATH
+
+#  define _RWSTD_ATAN2   _RWSTD_C::atan2
+#  define _RWSTD_COS     _RWSTD_C::cos
+#  define _RWSTD_COSH    _RWSTD_C::cosh
+#  define _RWSTD_EXP     _RWSTD_C::exp
+#  define _RWSTD_LOG     _RWSTD_C::log
+#  define _RWSTD_POW     _RWSTD_C::pow
+#  define _RWSTD_SIN     _RWSTD_C::sin
+#  define _RWSTD_SINH    _RWSTD_C::sinh
+#  define _RWSTD_SQRT    _RWSTD_C::sqrt
+
+#endif   // gcc 4.0
{noformat}

It would be nice if these definitions and macros could go into a common header. That way the
valarray transcendentals could take advantage of them.

{noformat}
+#if 4 <= __GNUG__
+
+    // use gcc 4.x builtins
+    complex (const float __complex__ &__rhs)
+        : _C_re (__builtin_crealf (__rhs)),
+          _C_im (__builtin_cimagf (__rhs)) { }
+
+#endif   // gcc >= 4.0
{noformat}

Should these new constructors be guarded by a \_RWSTD_NO_EXT_* macro? They are extensions,
so it seems that we should provide a way to disable them in strict mode. Also, C99 adds _Complex
and functions for manipulating them. Should we add an enhancement to support these since you've
opened the door here?

{noformat}
+#if 4 <= __GNUG__
+
+// use gcc 4.x C99 complex builtins
+
+#  define _RWSTD_COMPLEX_CAST(T, arg)                   \
+    _RWSTD_REINTERPRET_CAST (const T __complex__&, arg)
+
+
+_RWSTD_SPECIALIZED_FUNCTION inline float
+abs (const complex<float> &__x)
+{
+    return __builtin_cabsf (_RWSTD_COMPLEX_CAST (float, __x));
+}
+
{noformat}

I think \_RWSTD_COMPLEX_CAST is dangerous at best. I don't know for sure how a GCC \_\_complex__
is laid out in memory. If I were to guess, I'd expect a `float \_\_complex\_\_' to be two
contiguous floats like `float \_C_complex [2]', which may be completely different from `float
\_C_real, \_C_imag' if any padding is inserted. Fortunately the [documentation|http://gcc.gnu.org/onlinedocs/gcc-4.2.3/gcc/Complex.html#Complex]
says some interesting things about the layout of a \_\_complex__. 

It also appears that the IBM compiler supports a similar set of [functions|http://publib.boulder.ibm.com/infocenter/pseries/v5r3/index.jsp?topic=/com.ibm.xlcpp8a.doc/compiler/ref/bif_fp.htm].
It might be nice to add support for them while we're in there.

      was (Author: vitek):
    Martin, a few comments on your patch.

{noformat}
+#if 4 <= __GNUG__
+
+// define helpers to call gcc 4.x math builtins
+
+_RWSTD_NAMESPACE (__rw) {
+
+inline float __rw_atan2 (float __x, float __y)
+{
+    return __builtin_atan2f (__x, __y);
+}
+
+inline float __rw_cos (float __x)
+{
+    return __builtin_cosf (__x);
+}

+// use gcc 4 builtins for efficiency and namespace cleanliness
+#  define _RWSTD_ATAN2   _RW::__rw_atan2
+#  define _RWSTD_COS     _RW::__rw_cos
+#  define _RWSTD_COSH    _RW::__rw_cosh
+#  define _RWSTD_EXP     _RW::__rw_exp
+#  define _RWSTD_LOG     _RW::__rw_log
+#  define _RWSTD_POW     _RW::__rw_pow
+#  define _RWSTD_SIN     _RW::__rw_sin
+#  define _RWSTD_SINH    _RW::__rw_sinh
+#  define _RWSTD_SQRT    _RW::__rw_sqrt
+
+#else   // gcc < 4.0
+#  include _RWSTD_CMATH
+
+#  define _RWSTD_ATAN2   _RWSTD_C::atan2
+#  define _RWSTD_COS     _RWSTD_C::cos
+#  define _RWSTD_COSH    _RWSTD_C::cosh
+#  define _RWSTD_EXP     _RWSTD_C::exp
+#  define _RWSTD_LOG     _RWSTD_C::log
+#  define _RWSTD_POW     _RWSTD_C::pow
+#  define _RWSTD_SIN     _RWSTD_C::sin
+#  define _RWSTD_SINH    _RWSTD_C::sinh
+#  define _RWSTD_SQRT    _RWSTD_C::sqrt
+
+#endif   // gcc 4.0
{noformat}

It would be nice if these definitions and macros could go into a common header. That way the
valarray transcendentals could take advantage of them.

{noformat}
+#if 4 <= __GNUG__
+
+    // use gcc 4.x builtins
+    complex (const float __complex__ &__rhs)
+        : _C_re (__builtin_crealf (__rhs)),
+          _C_im (__builtin_cimagf (__rhs)) { }
+
+#endif   // gcc >= 4.0
{noformat}

Should these new constructors be guarded by a _RWSTD_NO_EXT_* macro? They are extensions,
so it seems that we should provide a way to disable them in strict mode. Also, C99 adds _Complex
and functions for manipulating them. Should we add an enhancement to support these since you've
opened the door here?

{noformat}
+#if 4 <= __GNUG__
+
+// use gcc 4.x C99 complex builtins
+
+#  define _RWSTD_COMPLEX_CAST(T, arg)                   \
+    _RWSTD_REINTERPRET_CAST (const T __complex__&, arg)
+
+
+_RWSTD_SPECIALIZED_FUNCTION inline float
+abs (const complex<float> &__x)
+{
+    return __builtin_cabsf (_RWSTD_COMPLEX_CAST (float, __x));
+}
+
{noformat}

I think _RWSTD_COMPLEX_CAST() is dangerous at best. I don't know for sure how a __complex__
is laid out in memory. If I were to guess, I'd expect a 'float __complex__' to be two contiguous
floats like 'float _C_complex [2]', which may be completely different from 'float _C_real,
_C_imag' if any padding is inserted. Fortunately the [documentation|http://gcc.gnu.org/onlinedocs/gcc-4.2.3/gcc/Complex.html#Complex]
says some interesting things about the layout of a __complex__. 

It also appears that the IBM compiler supports a similar set of [functions|http://publib.boulder.ibm.com/infocenter/pseries/v5r3/index.jsp?topic=/com.ibm.xlcpp8a.doc/compiler/ref/bif_fp.htm].
It might be nice to add support for them while we're in there.
  
> [gcc] use math __builtins
> -------------------------
>
>                 Key: STDCXX-722
>                 URL: https://issues.apache.org/jira/browse/STDCXX-722
>             Project: C++ Standard Library
>          Issue Type: Sub-task
>          Components: 26. Numerics
>    Affects Versions: 4.2.0
>         Environment: gcc 4
>            Reporter: Martin Sebor
>            Assignee: Martin Sebor
>            Priority: Minor
>             Fix For: 4.2.1
>
>         Attachments: stdcxx-722.diff, stdcxx-722.log
>
>   Original Estimate: 2h
>          Time Spent: 4h
>  Remaining Estimate: 0h
>
> For better efficiency and to reduce namespace pollution we can replace all the math functions
used in [<complex>|http://svn.apache.org/repos/asf/stdcxx/trunk/include/complex] with
gcc's built-in  equivalents.
> Quoting from section [5.46 Other built-in functions provided by GCC|http://gcc.gnu.org/onlinedocs/gcc-4.1.2/gcc/Other-Builtins.html#Other-Builtins]
of the gcc online manual:
> {quote}
> The ISO C99 functions {{_Exit}}, {{acoshf}}, {{acoshl}}, {{acosh}}, {{asinhf}}, {{asinhl}},
{{asinh}}, {{atanhf}}, {{atanhl}}, {{atanh}}, {{cabsf}}, {{cabsl}}, {{cabs}}, {{cacosf}},
{{cacoshf}}, {{cacoshl}}, {{cacosh}}, {{cacosl}}, {{cacos}}, {{cargf}}, {{cargl}}, {{carg}},
{{casinf}}, {{casinhf}}, {{casinhl}}, {{casinh}}, {{casinl}}, {{casin}}, {{catanf}}, {{catanhf}},
{{catanhl}}, {{catanh}}, {{catanl}}, {{catan}}, {{cbrtf}}, {{cbrtl}}, {{cbrt}}, {{ccosf}},
{{ccoshf}}, {{ccoshl}}, {{ccosh}}, {{ccosl}}, {{ccos}}, {{cexpf}}, {{cexpl}}, {{cexp}}, {{cimagf}},
{{cimagl}}, {{cimag}}, {{clogf}}, {{clogl}}, {{clog}}, {{conjf}}, {{conjl}}, {{conj}}, {{copysignf}},
{{copysignl}}, {{copysign}}, {{cpowf}}, {{cpowl}}, {{cpow}}, {{cprojf}}, {{cprojl}}, {{cproj}},
{{crealf}}, {{creall}}, {{creal}}, {{csinf}}, {{csinhf}}, {{csinhl}}, {{csinh}}, {{csinl}},
{{csin}}, {{csqrtf}}, {{csqrtl}}, {{csqrt}}, {{ctanf}}, {{ctanhf}}, {{ctanhl}}, {{ctanh}},
{{ctanl}}, {{ctan}}, {{erfcf}}, {{erfcl}}, {{erfc}}, {{erff}}, {{erfl}}, {{erf}}, {{exp2f}},
{{exp2l}}, {{exp2}}, {{expm1f}}, {{expm1l}}, {{expm1}}, {{fdimf}}, {{fdiml}}, {{fdim}}, {{fmaf}},
{{fmal}}, {{fmaxf}}, {{fmaxl}}, {{fmax}}, {{fma}}, {{fminf}}, {{fminl}}, {{fmin}}, {{hypotf}},
{{hypotl}}, {{hypot}}, {{ilogbf}}, {{ilogbl}}, {{ilogb}}, {{imaxabs}}, {{isblank}}, {{iswblank}},
{{lgammaf}}, {{lgammal}}, {{lgamma}}, {{llabs}}, {{llrintf}}, {{llrintl}}, {{llrint}}, {{llroundf}},
{{llroundl}}, {{llround}}, {{log1pf}}, {{log1pl}}, {{log1p}}, {{log2f}}, {{log2l}}, {{log2}},
{{logbf}}, {{logbl}}, {{logb}}, {{lrintf}}, {{lrintl}}, {{lrint}}, {{lroundf}}, {{lroundl}},
{{lround}}, {{nearbyintf}}, {{nearbyintl}}, {{nearbyint}}, {{nextafterf}}, {{nextafterl}},
{{nextafter}}, {{nexttowardf}}, {{nexttowardl}}, {{nexttoward}}, {{remainderf}}, {{remainderl}},
{{remainder}}, {{remquof}}, {{remquol}}, {{remquo}}, {{rintf}}, {{rintl}}, {{rint}}, {{roundf}},
{{roundl}}, {{round}}, {{scalblnf}}, {{scalblnl}}, {{scalbln}}, {{scalbnf}}, {{scalbnl}},
{{scalbn}}, {{snprintf}}, {{tgammaf}}, {{tgammal}}, {{tgamma}}, {{truncf}}, {{truncl}}, {{trunc}},
{{vfscanf}}, {{vscanf}}, {{vsnprintf}} and {{vsscanf}} are handled as built-in functions except
in strict ISO C90 mode ({{-ansi}} or {{-std=c89}}).
> There are also built-in versions of the ISO C99 functions {{acosf}}, {{acosl}}, {{asinf}},
{{asinl}}, {{atan2f}}, {{atan2l}}, {{atanf}}, {{atanl}}, {{ceilf}}, {{ceill}}, {{cosf}}, {{coshf}},
{{coshl}}, {{cosl}}, {{expf}}, {{expl}}, {{fabsf}}, {{fabsl}}, {{floorf}}, {{floorl}}, {{fmodf}},
{{fmodl}}, {{frexpf}}, {{frexpl}}, {{ldexpf}}, {{ldexpl}}, {{log10f}}, {{log10l}}, {{logf}},
{{logl}}, {{modfl}}, {{modf}}, {{powf}}, {{powl}}, {{sinf}}, {{sinhf}}, {{sinhl}}, {{sinl}},
{{sqrtf}}, {{sqrtl}}, {{tanf}}, {{tanhf}}, {{tanhl}} and {{tanl}} that are recognized in any
mode since ISO C90 reserves these names for the purpose to which ISO C99 puts them. All these
functions have corresponding versions prefixed with {{__builtin_}}.
> {quote}

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


Mime
View raw message