[PATCH] libstdc++: Fix for non-constexpr math_errhandling

[PATCH] libstdc++: Fix for non-constexpr math_errhandling

[Matthias Kretz]

[-- Attachment #1: Type: text/plain, Size: 1196 bytes --]

Follow-up to my last patch. This one is a more thorough fix. Tested on x86_64-
linux. OK for trunk?

---- 8< ----

Use SFINAE magic to support: "It is unspecified whether math_errhandling
is a macro or an identifier with external linkage." [C Standard]

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>

libstdc++-v3/ChangeLog:

	* include/experimental/bits/simd.h (__floating_point_flags): Do
	not rely on math_errhandling to expand to a constant expression.
---
 libstdc++-v3/include/experimental/bits/simd.h | 26 ++++++++++++++-----
 1 file changed, 20 insertions(+), 6 deletions(-)


--
──────────────────────────────────────────────────────────────────────────
 Dr. Matthias Kretz                           https://mattkretz.github.io
 GSI Helmholtz Centre for Heavy Ion Research               https://gsi.de
 stdₓ::simd
──────────────────────────────────────────────────────────────────────────

[-- Attachment #2: 0001-libstdc-Fix-for-non-constexpr-math_errhandling.patch --]
[-- Type: text/x-patch, Size: 1593 bytes --]

diff --git a/libstdc++-v3/include/experimental/bits/simd.h b/libstdc++-v3/include/experimental/bits/simd.h
index c991e3f223e..82e9841195e 100644
--- a/libstdc++-v3/include/experimental/bits/simd.h
+++ b/libstdc++-v3/include/experimental/bits/simd.h
@@ -283,20 +283,34 @@ constexpr inline bool __have_power_vmx = __have_power_vsx;
 
 namespace __detail
 {
-  constexpr bool __handle_fpexcept =
 #ifdef math_errhandling
-    math_errhandling & MATH_ERREXCEPT;
-#elif defined __FAST_MATH__
-    false;
+  // Determines _S_handle_fpexcept from math_errhandling if it is defined and expands to a constant
+  // expression. math_errhandling may expand to an extern symbol, in which case a constexpr value
+  // must be guessed.
+  template <int = math_errhandling>
+    constexpr bool __handle_fpexcept_impl(int)
+    { return math_errhandling & MATH_ERREXCEPT; }
+#endif
+
+  // Fallback if math_errhandling doesn't work: with fast-math assume floating-point exceptions are
+  // ignored, otherwise implement correct exception behavior.
+  constexpr bool __handle_fpexcept_impl(float)
+  {
+#if defined __FAST_MATH__
+    return false;
 #else
-    true;
+    return true;
 #endif
+  }
+
+  /// True if math functions must raise floating-point exceptions as specified by C17.
+  static constexpr bool _S_handle_fpexcept = __handle_fpexcept_impl(0);
 
   constexpr std::uint_least64_t
   __floating_point_flags()
   {
     std::uint_least64_t __flags = 0;
-    if constexpr (__handle_fpexcept)
+    if constexpr (_S_handle_fpexcept)
       __flags |= 1;
 #ifdef __FAST_MATH__
     __flags |= 1 << 1;

Re: [PATCH] libstdc++: Fix for non-constexpr math_errhandling

[Jonathan Wakely]

On Wed, 19 Jan 2022 at 08:10, Matthias Kretz wrote:

> Follow-up to my last patch. This one is a more thorough fix. Tested on
> x86_64-
> linux. OK for trunk?
>
> ---- 8< ----
>
> Use SFINAE magic to support: "It is unspecified whether math_errhandling
> is a macro or an identifier with external linkage." [C Standard]
>

The patch is OK for trunk, but I don't understand what the C standard means
here.

"the macro math_errhandling expands to [...]. It is unspecified whether
math_errhandling is a macro or an identifier with external linkage."

So is it a macro or not?

Re: [PATCH] libstdc++: Fix for non-constexpr math_errhandling

[Matthias Kretz]

On Wednesday, 19 January 2022 13:07:26 CET Jonathan Wakely wrote:
> On Wed, 19 Jan 2022 at 08:10, Matthias Kretz wrote:
> > Follow-up to my last patch. This one is a more thorough fix. Tested on
> > x86_64-
> > linux. OK for trunk?
> > 
> > ---- 8< ----
> > 
> > Use SFINAE magic to support: "It is unspecified whether math_errhandling
> > is a macro or an identifier with external linkage." [C Standard]
> 
> The patch is OK for trunk, but I don't understand what the C standard means
> here.
> 
> "the macro math_errhandling expands to [...]. It is unspecified whether
> math_errhandling is a macro or an identifier with external linkage."
> 
> So is it a macro or not?

I agree the quote I used is unclear. The complete paragraph:

The macros

MATH_ERRNO
MATH_ERREXCEPT

expand to the integer constants 1 and 2, respectively; the macro

math_errhandling

expands to an expression that has type int and the value MATH_ERRNO, 
MATH_ERREXCEPT, or the bitwise OR of both. The value of math_errhandling is 
constant for the duration of the program. It is unspecified whether 
math_errhandling is a macro or an identifier with external linkage. If a macro
definition is suppressed or a program defines an identifier with the name 
math_errhandling, the behavior is undefined. If the expression 
math_errhandling & MATH_ERREXCEPT can be nonzero, the implementation shall 
define the macros FE_DIVBYZERO, FE_INVALID, and FE_OVERFLOW in
<fenv.h>.

-- 
──────────────────────────────────────────────────────────────────────────
 Dr. Matthias Kretz                           https://mattkretz.github.io
 GSI Helmholtz Centre for Heavy Ion Research               https://gsi.de
 stdₓ::simd
──────────────────────────────────────────────────────────────────────────

Re: [PATCH] libstdc++: Fix for non-constexpr math_errhandling

[Jonathan Wakely]

On Wed, 19 Jan 2022, 12:45 Matthias Kretz, <m.kretz@gsi.de> wrote:

> On Wednesday, 19 January 2022 13:07:26 CET Jonathan Wakely wrote:
> > On Wed, 19 Jan 2022 at 08:10, Matthias Kretz wrote:
> > > Follow-up to my last patch. This one is a more thorough fix. Tested on
> > > x86_64-
> > > linux. OK for trunk?
> > >
> > > ---- 8< ----
> > >
> > > Use SFINAE magic to support: "It is unspecified whether
> math_errhandling
> > > is a macro or an identifier with external linkage." [C Standard]
> >
> > The patch is OK for trunk, but I don't understand what the C standard
> means
> > here.
> >
> > "the macro math_errhandling expands to [...]. It is unspecified whether
> > math_errhandling is a macro or an identifier with external linkage."
> >
> > So is it a macro or not?
>
> I agree the quote I used is unclear. The complete paragraph:
>
> The macros
>
> MATH_ERRNO
> MATH_ERREXCEPT
>
> expand to the integer constants 1 and 2, respectively; the macro
>
> math_errhandling
>
> expands to an expression that has type int and the value MATH_ERRNO,
> MATH_ERREXCEPT, or the bitwise OR of both. The value of math_errhandling
> is
> constant for the duration of the program. It is unspecified whether
> math_errhandling is a macro or an identifier with external linkage. If a
> macro
> definition is suppressed or a program defines an identifier with the name
> math_errhandling, the behavior is undefined. If the expression
> math_errhandling & MATH_ERREXCEPT can be nonzero, the implementation shall
> define the macros FE_DIVBYZERO, FE_INVALID, and FE_OVERFLOW in
> <fenv.h>.
>


But that still says "the macro math_errhandling" and then says it might not
be a macro.

I'll ask some WG14 people for clarity, but it doesn't affect your patch.

Re: [PATCH] libstdc++: Fix for non-constexpr math_errhandling

[Matthias Kretz]

On Wednesday, 19 January 2022 16:21:15 CET Jonathan Wakely wrote:
> On Wed, 19 Jan 2022, 12:45 Matthias Kretz, <m.kretz@gsi.de> wrote:
> > On Wednesday, 19 January 2022 13:07:26 CET Jonathan Wakely wrote:
> > > So is it a macro or not?
> > 
> > I agree the quote I used is unclear. The complete paragraph:
> > 
> > The macros
> > 
> > MATH_ERRNO
> > MATH_ERREXCEPT
> > 
> > expand to the integer constants 1 and 2, respectively; the macro
> > 
> > math_errhandling
> > 
> > expands to an expression that has type int and the value MATH_ERRNO,
> > MATH_ERREXCEPT, or the bitwise OR of both. The value of math_errhandling
> > is
> > constant for the duration of the program. It is unspecified whether
> > math_errhandling is a macro or an identifier with external linkage. If a
> > macro
> > definition is suppressed or a program defines an identifier with the name
> > math_errhandling, the behavior is undefined. If the expression
> > math_errhandling & MATH_ERREXCEPT can be nonzero, the implementation shall
> > define the macros FE_DIVBYZERO, FE_INVALID, and FE_OVERFLOW in
> > <fenv.h>.
> 
> But that still says "the macro math_errhandling" and then says it might not
> be a macro.

There's also [cmath.syn] https://eel.is/c++draft/cmath.syn which says:

#define math_errhandling see below

So, FWIW, libstdc++ is required to define math_errhandling as a macro in 
<cmath>. Thus, the original error (that math_errhandling wasn't defined even 
after <cmath> was included) really needs a fix in <cmath>. :-P

-- 
──────────────────────────────────────────────────────────────────────────
 Dr. Matthias Kretz                           https://mattkretz.github.io
 GSI Helmholtz Centre for Heavy Ion Research               https://gsi.de
 stdₓ::simd
──────────────────────────────────────────────────────────────────────────

Re: [PATCH] libstdc++: Fix for non-constexpr math_errhandling

[Jonathan Wakely]

On Wed, 19 Jan 2022 at 15:46, Matthias Kretz <m.kretz@gsi.de> wrote:
>
> On Wednesday, 19 January 2022 16:21:15 CET Jonathan Wakely wrote:
> > On Wed, 19 Jan 2022, 12:45 Matthias Kretz, <m.kretz@gsi.de> wrote:
> > > On Wednesday, 19 January 2022 13:07:26 CET Jonathan Wakely wrote:
> > > > So is it a macro or not?
> > >
> > > I agree the quote I used is unclear. The complete paragraph:
> > >
> > > The macros
> > >
> > > MATH_ERRNO
> > > MATH_ERREXCEPT
> > >
> > > expand to the integer constants 1 and 2, respectively; the macro
> > >
> > > math_errhandling
> > >
> > > expands to an expression that has type int and the value MATH_ERRNO,
> > > MATH_ERREXCEPT, or the bitwise OR of both. The value of math_errhandling
> > > is
> > > constant for the duration of the program. It is unspecified whether
> > > math_errhandling is a macro or an identifier with external linkage. If a
> > > macro
> > > definition is suppressed or a program defines an identifier with the name
> > > math_errhandling, the behavior is undefined. If the expression
> > > math_errhandling & MATH_ERREXCEPT can be nonzero, the implementation shall
> > > define the macros FE_DIVBYZERO, FE_INVALID, and FE_OVERFLOW in
> > > <fenv.h>.
> >
> > But that still says "the macro math_errhandling" and then says it might not
> > be a macro.
>
> There's also [cmath.syn] https://eel.is/c++draft/cmath.syn which says:
>
> #define math_errhandling see below
>
> So, FWIW, libstdc++ is required to define math_errhandling as a macro in
> <cmath>. Thus, the original error (that math_errhandling wasn't defined even
> after <cmath> was included) really needs a fix in <cmath>. :-P

No, because we get it from libc:

#include_next <math.h>

Re: [PATCH] libstdc++: Fix for non-constexpr math_errhandling

[Jonathan Wakely]

On Wed, 19 Jan 2022 at 16:45, Jonathan Wakely <jwakely.gcc@gmail.com> wrote:
>
> On Wed, 19 Jan 2022 at 15:46, Matthias Kretz <m.kretz@gsi.de> wrote:
> >
> > On Wednesday, 19 January 2022 16:21:15 CET Jonathan Wakely wrote:
> > > On Wed, 19 Jan 2022, 12:45 Matthias Kretz, <m.kretz@gsi.de> wrote:
> > > > On Wednesday, 19 January 2022 13:07:26 CET Jonathan Wakely wrote:
> > > > > So is it a macro or not?
> > > >
> > > > I agree the quote I used is unclear. The complete paragraph:
> > > >
> > > > The macros
> > > >
> > > > MATH_ERRNO
> > > > MATH_ERREXCEPT
> > > >
> > > > expand to the integer constants 1 and 2, respectively; the macro
> > > >
> > > > math_errhandling
> > > >
> > > > expands to an expression that has type int and the value MATH_ERRNO,
> > > > MATH_ERREXCEPT, or the bitwise OR of both. The value of math_errhandling
> > > > is
> > > > constant for the duration of the program. It is unspecified whether
> > > > math_errhandling is a macro or an identifier with external linkage. If a
> > > > macro
> > > > definition is suppressed or a program defines an identifier with the name
> > > > math_errhandling, the behavior is undefined. If the expression
> > > > math_errhandling & MATH_ERREXCEPT can be nonzero, the implementation shall
> > > > define the macros FE_DIVBYZERO, FE_INVALID, and FE_OVERFLOW in
> > > > <fenv.h>.
> > >
> > > But that still says "the macro math_errhandling" and then says it might not
> > > be a macro.
> >
> > There's also [cmath.syn] https://eel.is/c++draft/cmath.syn which says:
> >
> > #define math_errhandling see below
> >
> > So, FWIW, libstdc++ is required to define math_errhandling as a macro in
> > <cmath>. Thus, the original error (that math_errhandling wasn't defined even
> > after <cmath> was included) really needs a fix in <cmath>. :-P
>
> No, because we get it from libc:
>
> #include_next <math.h>

So if you aren't seeing it after <cmath> is included, your libc is broken.