Re: Add __builtin_tgmath for better tgmath.h implementation (bug 81156)

[Joseph Myers]

On Wed, 15 Nov 2017, Richard Biener wrote:

> Thanks - I suppose we can't avoid the repeated expansion by sth like
> 
> #define exp(Val) ({ __typeof__ Val tem = Val; __TGMATH_UNARY_REAL_IMAG
> (tem, exp, cexp); })

Well, that still expands its argument twice.  You'd need to use 
__auto_type to avoid the double expansion.  And then you'd still have 
extremely complicated expansions (that are correspondingly unfriendly if a 
user makes a mistake with a call, e.g. an argument of unsupported type), 
and complications around getting the right semantics when decimal floating 
point is involved.  And use of ({ }) doesn't work in sizeof outside 
functions.  And that wouldn't help with cases such as 
__STDC_TGMATH_OPERATOR_EVALUATION__, whereas it would actually be easy to 
add __builtin_tgmath_operator that's handled the same as __builtin_tgmath 
but ends up calling a function based on evaluation formats and producing 
an EXCESS_PRECISION_EXPR.

(Clang overloadable functions in C don't avoid the multiple expansion 
either, or at least Clang's tgmath.h doesn't.)

-- 
Joseph S. Myers
jos...@codesourcery.com

Re: Add __builtin_tgmath for better tgmath.h implementation (bug 81156)

[Richard Biener]

On Wed, Nov 15, 2017 at 2:54 AM, Joseph Myers  wrote:
> Various implementations of C99/C11  have the property that
> their macro expansions contain many copies of the macro arguments, so
> resulting in exponential blowup of the size of macro expansions where
> a call to such a macro contains other such calls in the macro
> arguments.
>
> This patch adds a (C-only) language feature __builtin_tgmath designed
> to avoid this problem by implementing the  function
> selection rules directly in the compiler.  The effect is that
> type-generic macros can be defined simply as
>
> #define pow(a, b) __builtin_tgmath (powf, pow, powl, \
> cpowf, cpow, cpowl, a, b)
>
> as in the example added to the manual, with each macro argument
> expanded exactly once.  The details of __builtin_tgmath are as
> described in the manual.  This is C-only since C++ uses function
> overloading and just defines  to include  and
> .
>
> __builtin_tgmath handles C99/C11 type-generic macros, and _FloatN,
> _FloatNx and decimal floating-point types (following the proposed
> resolution to the floating-point TS DR#9 that makes the rules for
> finding a common type from arguments to a type-generic macro follow
> the usual arithmetic conversions after adjustment of integer arguments
> to _Decimal64 or double - or to _Complex double in the case of GNU
> complex integer arguments).
>
> Type-generic macros for functions from TS 18661 that round their
> results to a narrower type are handled, but there are still some
> unresolved questions regarding such macros so further changes in that
> regard may be needed in future.  The current implementation follows an
> older version of the DR#13 resolution (allowing a function for a
> wide-enough argument type to be selected if no exactly-matching
> function is available), but with appropriate calls to __builtin_tgmath
> is still fully compatible with the latest version of the resolution
> (not yet in the DR log), and allowing such not-exactly-matching
> argument types to be chosen in that case avoids needing another
> special case to treat integers as _Float64 instead of double in
> certain cases.
>
> Regarding other possible language/library features, not currently
> implemented in GCC:
>
> * Imaginary types could be naturally supported by allowing cases where
>   the type-generic type is an imaginary type T and arguments or return
>   types may be T (as at present), or the corresponding real type to T
>   (as at present), or (new) the corresponding real type if T is real
>   or imaginary but T if T is complex.  (tgmath.h would need a series
>   of functions such as
>
>   static inline _Imaginary double
>   __sin_imag (_Imaginary double __x)
>   {
> return _Imaginary_I * sinh (__imag__ __x);
>   }
>
>   to be used in __builtin_tgmath calls.)
>
> * __builtin_tgmath would use the constant rounding direction in the
>   presence of support for the FENV_ROUND / FENV_DEC_ROUND pragmas.
>   Support for those would also require a new __builtin_ to
>   cause a non-type-generic call to use the constant rounding
>   direction (it seems cleaner to add a new __builtin_ when
>   required than to make __builtin_tgmath handle a non-type-generic
>   case with only one function argument).
>
> * TS 18661-5 __STDC_TGMATH_OPERATOR_EVALUATION__ would require new
>   __builtin_ that evaluates with excess range and precision
>   like arithmetic operators do.
>
> * The proposed C bindings for IEEE 754-2018 augmented arithmetic
>   operations involve struct return types.  As currently implemented
>   __builtin_tgmath does not handle those, but support could be added.
>
> There are many error cases that the implementation diagnoses.  I've
> tried to ensure reasonable error messages for erroneous uses of
> __builtin_tgmath, but the errors for erroneous uses of the resulting
> type-generic macros (that is, when the non-function arguments have
> inappropriate types) are more important as they are more likely to be
> seen by users.
>
> GCC's own tgmath.h, as used for some targets, is updated in this
> patch.  I've tested those changes minimally, via adjusting
> gcc.dg/c99-tgmath-* locally to use that tgmath.h version.  I've also
> run the glibc testsuite (which has much more thorough tests of
> correctness of tgmath.h function selection) with a glibc patch to use
> __builtin_tgmath in glibc's tgmath.h.
>
> Bootstrapped with no regressions on x86_64-pc-linux-gnu.  Applied to
> mainline.

Thanks - I suppose we can't avoid the repeated expansion by sth like

#define exp(Val) ({ __typeof__ Val tem = Val; __TGMATH_UNARY_REAL_IMAG
(tem, exp, cexp); })

?

Richard.

> gcc:
> 2017-11-15  Joseph Myers  
>
> PR c/81156
> * doc/extend.texi (Other Builtins): Document __builtin_tgmath.
> * ginclude/tgmath.h (__tg_cplx, __tg_ldbl, __tg_dbl, __tg_choose)
> (__tg_choose_2, __tg_choose_3, __TGMATH_REAL_1_2)
> 

Add __builtin_tgmath for better tgmath.h implementation (bug 81156)

[Joseph Myers]

Various implementations of C99/C11  have the property that
their macro expansions contain many copies of the macro arguments, so
resulting in exponential blowup of the size of macro expansions where
a call to such a macro contains other such calls in the macro
arguments.

This patch adds a (C-only) language feature __builtin_tgmath designed
to avoid this problem by implementing the  function
selection rules directly in the compiler.  The effect is that
type-generic macros can be defined simply as

#define pow(a, b) __builtin_tgmath (powf, pow, powl, \
cpowf, cpow, cpowl, a, b)

as in the example added to the manual, with each macro argument
expanded exactly once.  The details of __builtin_tgmath are as
described in the manual.  This is C-only since C++ uses function
overloading and just defines  to include  and
.

__builtin_tgmath handles C99/C11 type-generic macros, and _FloatN,
_FloatNx and decimal floating-point types (following the proposed
resolution to the floating-point TS DR#9 that makes the rules for
finding a common type from arguments to a type-generic macro follow
the usual arithmetic conversions after adjustment of integer arguments
to _Decimal64 or double - or to _Complex double in the case of GNU
complex integer arguments).

Type-generic macros for functions from TS 18661 that round their
results to a narrower type are handled, but there are still some
unresolved questions regarding such macros so further changes in that
regard may be needed in future.  The current implementation follows an
older version of the DR#13 resolution (allowing a function for a
wide-enough argument type to be selected if no exactly-matching
function is available), but with appropriate calls to __builtin_tgmath
is still fully compatible with the latest version of the resolution
(not yet in the DR log), and allowing such not-exactly-matching
argument types to be chosen in that case avoids needing another
special case to treat integers as _Float64 instead of double in
certain cases.

Regarding other possible language/library features, not currently
implemented in GCC:

* Imaginary types could be naturally supported by allowing cases where
  the type-generic type is an imaginary type T and arguments or return
  types may be T (as at present), or the corresponding real type to T
  (as at present), or (new) the corresponding real type if T is real
  or imaginary but T if T is complex.  (tgmath.h would need a series
  of functions such as

  static inline _Imaginary double
  __sin_imag (_Imaginary double __x)
  {
return _Imaginary_I * sinh (__imag__ __x);
  }

  to be used in __builtin_tgmath calls.)

* __builtin_tgmath would use the constant rounding direction in the
  presence of support for the FENV_ROUND / FENV_DEC_ROUND pragmas.
  Support for those would also require a new __builtin_ to
  cause a non-type-generic call to use the constant rounding
  direction (it seems cleaner to add a new __builtin_ when
  required than to make __builtin_tgmath handle a non-type-generic
  case with only one function argument).

* TS 18661-5 __STDC_TGMATH_OPERATOR_EVALUATION__ would require new
  __builtin_ that evaluates with excess range and precision
  like arithmetic operators do.

* The proposed C bindings for IEEE 754-2018 augmented arithmetic
  operations involve struct return types.  As currently implemented
  __builtin_tgmath does not handle those, but support could be added.

There are many error cases that the implementation diagnoses.  I've
tried to ensure reasonable error messages for erroneous uses of
__builtin_tgmath, but the errors for erroneous uses of the resulting
type-generic macros (that is, when the non-function arguments have
inappropriate types) are more important as they are more likely to be
seen by users.

GCC's own tgmath.h, as used for some targets, is updated in this
patch.  I've tested those changes minimally, via adjusting
gcc.dg/c99-tgmath-* locally to use that tgmath.h version.  I've also
run the glibc testsuite (which has much more thorough tests of
correctness of tgmath.h function selection) with a glibc patch to use
__builtin_tgmath in glibc's tgmath.h.

Bootstrapped with no regressions on x86_64-pc-linux-gnu.  Applied to 
mainline.

gcc:
2017-11-15  Joseph Myers  

PR c/81156
* doc/extend.texi (Other Builtins): Document __builtin_tgmath.
* ginclude/tgmath.h (__tg_cplx, __tg_ldbl, __tg_dbl, __tg_choose)
(__tg_choose_2, __tg_choose_3, __TGMATH_REAL_1_2)
(__TGMATH_REAL_2_3): Remove macros.
(__TGMATH_CPLX, __TGMATH_CPLX_2, __TGMATH_REAL, __TGMATH_REAL_2)
(__TGMATH_REAL_3, __TGMATH_CPLX_ONLY): Define using
__builtin_tgmath.
(frexp, ldexp, nexttoward, scalbn, scalbln): Define using
__TGMATH_REAL_2.
(remquo): Define using __TGMATH_REAL_3.

gcc/c:
2017-11-15  Joseph Myers  

PR c/81156
* c-parser.c