On Wednesday, 14 July 2021 14:42:01 CEST H.J. Lu wrote:
> On Wed, Jul 14, 2021 at 12:32 AM Matthias Kretz <m.kr...@gsi.de> wrote:
> > OK?
> >
> > On Wednesday, 30 June 2021 10:59:28 CEST Matthias Kretz wrote:
> > > Library code, especially in headers, sometimes needs to know how the
> > > compiler interprets / optimizes floating-point types and operations.
> > > This information can be used for additional optimizations or for
> > > ensuring correctness. This change makes -freciprocal-math,
> > > -fno-signed-zeros, -fno-trapping-math, -fassociative-math, and
> > > -frounding-math report their state via corresponding pre-defined macros.
> > >
> > > Signed-off-by: Matthias Kretz <m.kr...@gsi.de>
> > >
> > > gcc/testsuite/ChangeLog:
> > > * gcc.dg/associative-math-1.c: New test.
> > > * gcc.dg/associative-math-2.c: New test.
> > > * gcc.dg/no-signed-zeros-1.c: New test.
> > > * gcc.dg/no-signed-zeros-2.c: New test.
> > > * gcc.dg/no-trapping-math-1.c: New test.
> > > * gcc.dg/no-trapping-math-2.c: New test.
> > > * gcc.dg/reciprocal-math-1.c: New test.
> > > * gcc.dg/reciprocal-math-2.c: New test.
> > > * gcc.dg/rounding-math-1.c: New test.
> > > * gcc.dg/rounding-math-2.c: New test.
> > >
> > > gcc/c-family/ChangeLog:
> > > * c-cppbuiltin.c (c_cpp_builtins_optimize_pragma): Define or
> > > undefine __RECIPROCAL_MATH__, __NO_SIGNED_ZEROS__,
> > > __NO_TRAPPING_MATH__, __ASSOCIATIVE_MATH__, and
> > > __ROUNDING_MATH__ according to the new optimization flags.
> > >
> > > gcc/ChangeLog:
> > > * cppbuiltin.c (define_builtin_macros_for_compilation_flags):
> > > Define __RECIPROCAL_MATH__, __NO_SIGNED_ZEROS__,
> > > __NO_TRAPPING_MATH__, __ASSOCIATIVE_MATH__, and
> > > __ROUNDING_MATH__ according to their corresponding flags.
> > > * doc/cpp.texi: Document __RECIPROCAL_MATH__,
> > > __NO_SIGNED_ZEROS__, __NO_TRAPPING_MATH__, __ASSOCIATIVE_MATH__,
> > > and __ROUNDING_MATH__.
> > >
>
> Hi Hongtao,
>
> Can this be used to address
>
> https://gcc.gnu.org/pipermail/gcc/2021-July/236778.html
It should help to determine when a workaround is necessary. I use inline asm
to implement the workaround. Relevant libstdc++ code (not upstream yet and not
making use of __ASSOCIATIVE_MATH__ yet):
/*
* Ensure the expressions leading up to the @p __x argument are evaluated at
least once.
*
* Example: __force_evaluation(x + y) - y will not optimize to x with -
fassociative-math.
* _TV is expected to be __vector_type_t<floating-point type, N>.
*/
template <typename _TV>
[[__gnu__::__flatten__, __gnu__::__const__]]
_GLIBCXX_SIMD_INTRINSIC constexpr
_TV
__force_evaluation(_TV __x) noexcept
{
if (__builtin_is_constant_evaluated())
return __x;
else
return [&] {
if constexpr(__have_sse)
{
if constexpr (sizeof(__x) >= 16)
{
asm("" :: "x"(__x));
asm("" : "+x"(__x));
}
else if constexpr (is_same_v<__vector_type_t<float, 2>, _TV>)
{
asm("" :: "x"(__x[0]), "x"(__x[1]));
asm("" : "+x"(__x[0]), "+x"(__x[1]));
}
else
__assert_unreachable<_TV>();
}
else if constexpr(__have_neon)
{
asm("" :: "w"(__x));
asm("" : "+w"(__x));
}
else if constexpr (__have_power_vmx)
{
if constexpr (is_same_v<__vector_type_t<float, 2>, _TV>)
{
asm("" :: "fgr"(__x[0]), "fgr"(__x[1]));
asm("" : "+fgr"(__x[0]), "+fgr"(__x[1]));
}
else
{
asm("" :: "v"(__x));
asm("" : "+v"(__x));
}
}
else
{
asm("" :: "g"(__x));
asm("" : "+g"(__x));
}
return __x;
}();
}
// Returns __x + __y - __y without -fassociative-math optimizing to __x.
// - _TV must be __vector_type_t<floating-point type, N>.
// - _UV must be _TV or floating-point type.
template <typename _TV, typename _UV>
[[__gnu__::__const__]]
_GLIBCXX_SIMD_INTRINSIC constexpr
_TV
__plus_minus(_TV __x, _UV __y) noexcept
{
#if defined __clang__ || __GCC_IEC_559 > 0
return (__x + __y) - __y;
#else
if (__builtin_is_constant_evaluated()
|| (__builtin_constant_p(__x) && __builtin_constant_p(__y)))
return (__x + __y) - __y;
#if defined __i386__ && !defined __SSE_MATH__
else if constexpr (sizeof(__x) == 8)
{ // operations on __x would use the FPU
static_assert(is_same_v<_TV, __vector_type_t<float, 2>>);
const auto __x4 = __vector_bitcast<float, 4>(__x);
if constexpr (is_same_v<_TV, _UV>)
return __vector_bitcast<float, 2>(
__plus_minus(__x4, __vector_bitcast<float, 4>(__y)));
else
return __vector_bitcast<float, 2>(__plus_minus(__x4, __y));
}
#endif
else
return __force_evaluation(__x + __y) - __y;
#endif
}
--
──────────────────────────────────────────────────────────────────────────
Dr. Matthias Kretz https://mattkretz.github.io
GSI Helmholtz Centre for Heavy Ion Research https://gsi.de
std::experimental::simd https://github.com/VcDevel/std-simd
──────────────────────────────────────────────────────────────────────────
