C2x semantics for <tgmath.h>

<tgmath.h> implements semantics for integer generic arguments that
handle cases involving _FloatN / _FloatNx types as specified in TS
18661-3 plus some defect fixes.

C2x has further changes to the semantics for <tgmath.h> macros with
such types, which should also be considered defect fixes (although
handled through the integration of TS 18661-3 in C2x rather than
through an issue tracking process).  Specifically, the rules were
changed because of problems raised with using the macros with the
evaluation format types such as float_t and _Float32_t: the older
version of the rules didn't allow passing _FloatN / _FloatNx types to
the narrowing macros returning float or double, or passing float /
double / long double to the narrowing macros returning _FloatN /
_FloatNx, which was a problem with the evaluation format types which
could be either kind of type depending on the value of
FLT_EVAL_METHOD.

Thus the new rules allow cases of mixing types which were not allowed
before, and, as part of the changes, the handling of integer arguments
was also changed: if there is any _FloatNx generic argument, integer
generic arguments are treated as _Float32x (not double), while the
rule about treating integer arguments to narrowing macros returning
_FloatN or _FloatNx as _Float64 not double was removed (no longer
needed now double is a valid argument to such macros).

I've implemented the changes in GCC's __builtin_tgmath, which thus
requires updates to glibc's test expectations so that the tests
continue to build with GCC 13 (the test is also updated to test the
argument types that weren't allowed before but are now valid under C2x
rules).

Given those test changes, it's then also necessary to fix the
implementations in <tgmath.h> to have appropriate semantics with older
GCC so that the tests pass with GCC versions before GCC 13 as well.
For some cases (non-narrowing macros with two or three generic
arguments; narrowing macros returning _Float32x), the older version of
__builtin_tgmath doesn't correspond sufficiently well to C2x
semantics, so in those cases <tgmath.h> is adjusted to use the older
macro implementation instead of __builtin_tgmath.  The older macro
implementation is itself adjusted to give the desired semantics, with
GCC 7 and later.  (It's not possible to get the right semantics in all
cases for the narrowing macros with GCC 6 and before when the _FloatN
/ _FloatNx names are typedefs rather than distinct types.)

Tested as follows: with the full glibc testsuite for x86_64, GCC 6, 7,
11, 13; with execution of the math/tests for aarch64, arm, powerpc and
powerpc64le, GCC 6, 7, 12 and 13 (powerpc64le only with GCC 12 and
13); with build-many-glibcs.py with GCC 6, 7, 12 and 13.

1 files changed, 166 insertions, 80 deletions

diff --git a/math/tgmath.h b/math/tgmath.h
index a50c218a52..7a1e77f393 100644
--- a/math/tgmath.h
+++ b/math/tgmath.h
@@ -37,9 +37,17 @@
    for older GCC, using other compiler extensions but with macros
    expanding their arguments many times (so resulting in exponential
    blowup of the size of expansions when calls to such macros are
-   nested inside arguments to such macros).  */
+   nested inside arguments to such macros).  Because of a long series
+   of defect fixes made after the initial release of TS 18661-1, GCC
+   versions before GCC 13 have __builtin_tgmath semantics that, when
+   integer arguments are passed to narrowing macros returning
+   _Float32x, or non-narrowing macros with at least two generic
+   arguments, do not always correspond to the C2X semantics, so more
+   complicated macro definitions are also used in some cases for
+   versions from GCC 8 to GCC 12.  */
 
 #define __HAVE_BUILTIN_TGMATH __GNUC_PREREQ (8, 0)
+#define __HAVE_BUILTIN_TGMATH_C2X __GNUC_PREREQ (13, 0)
 
 #if __GNUC_PREREQ (2, 7)
 
@@ -163,7 +171,7 @@
   __builtin_tgmath (__TGMATH_NARROW_FUNCS_F64 (F) (X), (Y))
 #  define __TGMATH_3_NARROW_F64(F, X, Y, Z)				\
   __builtin_tgmath (__TGMATH_NARROW_FUNCS_F64 (F) (X), (Y), (Z))
-#  if __HAVE_FLOAT128
+#  if __HAVE_FLOAT128 && __HAVE_BUILTIN_TGMATH_C2X
 #   define __TGMATH_1_NARROW_F32X(F, X)				\
   __builtin_tgmath (__TGMATH_NARROW_FUNCS_F32X (F) (X))
 #   define __TGMATH_2_NARROW_F32X(F, X, Y)			\
@@ -172,8 +180,9 @@
   __builtin_tgmath (__TGMATH_NARROW_FUNCS_F32X (F) (X), (Y), (Z))
 #  endif
 
-# else /* !__HAVE_BUILTIN_TGMATH.  */
+# endif
 
+# if !__HAVE_BUILTIN_TGMATH_C2X
 #  ifdef __NO_LONG_DOUBLE_MATH
 #   define __tgml(fct) fct
 #  else
@@ -213,13 +222,17 @@
 /* Whether an expression (of arithmetic type) has a real type.  */
 #  define __expr_is_real(E) (__builtin_classify_type (E) != 9)
 
+/* Type T1 if E is 1, type T2 is E is 0.  */
+#  define __tgmath_type_if(T1, T2, E)					\
+  __typeof__ (*(0 ? (__typeof__ (0 ? (T2 *) 0 : (void *) (E))) 0	\
+		: (__typeof__ (0 ? (T1 *) 0 : (void *) (!(E)))) 0))
+
 /* The tgmath real type for T, where E is 0 if T is an integer type
    and 1 for a floating type.  If T has a complex type, it is
    unspecified whether the return type is real or complex (but it has
    the correct corresponding real type).  */
 #  define __tgmath_real_type_sub(T, E) \
-  __typeof__ (*(0 ? (__typeof__ (0 ? (double *) 0 : (void *) (E))) 0	      \
-		  : (__typeof__ (0 ? (T *) 0 : (void *) (!(E)))) 0))
+  __tgmath_type_if (T, double, E)
 
 /* The tgmath real type of EXPR.  */
 #  define __tgmath_real_type(expr) \
@@ -247,6 +260,56 @@
 			     __real_integer_type (__typeof__ (+(expr))), \
 			     __complex_integer_type (__typeof__ (+(expr))))
 
+/* The tgmath real type of EXPR1 combined with EXPR2, without handling
+   the C2X rule of interpreting integer arguments as _Float32x if any
+   argument is _FloatNx.  */
+#  define __tgmath_real_type2_base(expr1, expr2)			\
+  __typeof ((__tgmath_real_type (expr1)) 0 + (__tgmath_real_type (expr2)) 0)
+
+/* The tgmath complex type of EXPR1 combined with EXPR2, without
+   handling the C2X rule of interpreting integer arguments as
+   _Float32x if any argument is _FloatNx.  */
+#  define __tgmath_complex_type2_base(expr1, expr2)	\
+  __typeof ((__tgmath_complex_type (expr1)) 0		\
+	    + (__tgmath_complex_type (expr2)) 0)
+
+/* The tgmath real type of EXPR1 combined with EXPR2 and EXPR3,
+   without handling the C2X rule of interpreting integer arguments as
+   _Float32x if any argument is _FloatNx.  */
+#  define __tgmath_real_type3_base(expr1, expr2, expr3)	\
+  __typeof ((__tgmath_real_type (expr1)) 0		\
+	    + (__tgmath_real_type (expr2)) 0		\
+	    + (__tgmath_real_type (expr3)) 0)
+
+/* The tgmath real or complex type of EXPR1 combined with EXPR2 (and
+   EXPR3 if applicable).  */
+#  if __HAVE_FLOATN_NOT_TYPEDEF
+#   define __tgmath_real_type2(expr1, expr2)				\
+  __tgmath_type_if (_Float32x, __tgmath_real_type2_base (expr1, expr2), \
+		    _Generic ((expr1) + (expr2), _Float32x: 1, default: 0))
+#   define __tgmath_complex_type2(expr1, expr2)				\
+  __tgmath_type_if (_Float32x,						\
+		    __tgmath_type_if (_Complex _Float32x,		\
+				      __tgmath_complex_type2_base (expr1, \
+								   expr2), \
+				      _Generic ((expr1) + (expr2),	\
+						_Complex _Float32x: 1,	\
+						default: 0)),		\
+		    _Generic ((expr1) + (expr2), _Float32x: 1, default: 0))
+#   define __tgmath_real_type3(expr1, expr2, expr3)			\
+  __tgmath_type_if (_Float32x,						\
+		    __tgmath_real_type3_base (expr1, expr2, expr3),	\
+		    _Generic ((expr1) + (expr2) + (expr3),		\
+			      _Float32x: 1, default: 0))
+#  else
+#   define __tgmath_real_type2(expr1, expr2)	\
+  __tgmath_real_type2_base (expr1, expr2)
+#   define __tgmath_complex_type2(expr1, expr2)	\
+  __tgmath_complex_type2_base (expr1, expr2)
+#   define __tgmath_real_type3(expr1, expr2, expr3)	\
+  __tgmath_real_type3_base (expr1, expr2, expr3)
+#  endif
+
 #  if (__HAVE_DISTINCT_FLOAT16			\
       || __HAVE_DISTINCT_FLOAT32		\
       || __HAVE_DISTINCT_FLOAT64		\
@@ -258,7 +321,10 @@
 
 /* Expand to text that checks if ARG_COMB has type _Float128, and if
    so calls the appropriately suffixed FCT (which may include a cast),
-   or FCT and CFCT for complex functions, with arguments ARG_CALL.  */
+   or FCT and CFCT for complex functions, with arguments ARG_CALL.
+   __TGMATH_F128LD (only used in the __HAVE_FLOAT64X_LONG_DOUBLE case,
+   for narrowing macros) handles long double the same as
+   _Float128.  */
 #  if __HAVE_DISTINCT_FLOAT128 && __GLIBC_USE (IEC_60559_TYPES_EXT)
 #   if (!__HAVE_FLOAT64X			\
        || __HAVE_FLOAT64X_LONG_DOUBLE		\
@@ -266,6 +332,10 @@
 #    define __TGMATH_F128(arg_comb, fct, arg_call)			\
   __builtin_types_compatible_p (__typeof (+(arg_comb)), _Float128)	\
   ? fct ## f128 arg_call :
+#    define __TGMATH_F128LD(arg_comb, fct, arg_call)			\
+  (__builtin_types_compatible_p (__typeof (+(arg_comb)), _Float128)	\
+   || __builtin_types_compatible_p (__typeof (+(arg_comb)), long double)) \
+  ? fct ## f128 arg_call :
 #    define __TGMATH_CF128(arg_comb, fct, cfct, arg_call)		\
   __builtin_types_compatible_p (__typeof (+__real__ (arg_comb)), _Float128) \
   ? (__expr_is_real (arg_comb)						\
@@ -291,7 +361,7 @@
 #   define __TGMATH_CF128(arg_comb, fct, cfct, arg_call) /* Nothing.  */
 #  endif
 
-# endif /* !__HAVE_BUILTIN_TGMATH.  */
+# endif /* !__HAVE_BUILTIN_TGMATH_C2X.  */
 
 /* We have two kinds of generic macros: to support functions which are
    only defined on real valued parameters and those which are defined
@@ -304,14 +374,18 @@
   __TGMATH_2 (Fct, (Val1), (Val2))
 #  define __TGMATH_BINARY_FIRST_REAL_STD_ONLY(Val1, Val2, Fct)	\
   __TGMATH_2STD (Fct, (Val1), (Val2))
-#  define __TGMATH_BINARY_REAL_ONLY(Val1, Val2, Fct)	\
+#  if __HAVE_BUILTIN_TGMATH_C2X
+#   define __TGMATH_BINARY_REAL_ONLY(Val1, Val2, Fct)	\
   __TGMATH_2 (Fct, (Val1), (Val2))
+#  endif
 #  define __TGMATH_BINARY_REAL_STD_ONLY(Val1, Val2, Fct)	\
   __TGMATH_2STD (Fct, (Val1), (Val2))
-#  define __TGMATH_TERNARY_FIRST_SECOND_REAL_ONLY(Val1, Val2, Val3, Fct) \
+#  if __HAVE_BUILTIN_TGMATH_C2X
+#   define __TGMATH_TERNARY_FIRST_SECOND_REAL_ONLY(Val1, Val2, Val3, Fct) \
   __TGMATH_3 (Fct, (Val1), (Val2), (Val3))
-#  define __TGMATH_TERNARY_REAL_ONLY(Val1, Val2, Val3, Fct)	\
+#   define __TGMATH_TERNARY_REAL_ONLY(Val1, Val2, Val3, Fct)	\
   __TGMATH_3 (Fct, (Val1), (Val2), (Val3))
+#  endif
 #  define __TGMATH_TERNARY_FIRST_REAL_RET_ONLY(Val1, Val2, Val3, Fct)	\
   __TGMATH_3 (Fct, (Val1), (Val2), (Val3))
 #  define __TGMATH_UNARY_REAL_IMAG(Val, Fct, Cfct)	\
@@ -321,11 +395,14 @@
   __TGMATH_1C (Fct, Cfct, (Val))
 #  define __TGMATH_UNARY_REAL_IMAG_RET_REAL_SAME(Val, Cfct)	\
   __TGMATH_1 (Cfct, (Val))
-#  define __TGMATH_BINARY_REAL_IMAG(Val1, Val2, Fct, Cfct)	\
+#  if __HAVE_BUILTIN_TGMATH_C2X
+#   define __TGMATH_BINARY_REAL_IMAG(Val1, Val2, Fct, Cfct)	\
   __TGMATH_2C (Fct, Cfct, (Val1), (Val2))
+#  endif
 
-# else /* !__HAVE_BUILTIN_TGMATH.  */
+# endif
 
+# if !__HAVE_BUILTIN_TGMATH
 #  define __TGMATH_UNARY_REAL_ONLY(Val, Fct)				\
   (__extension__ ((sizeof (+(Val)) == sizeof (double)			      \
 		      || __builtin_classify_type (Val) != 8)		      \
@@ -362,29 +439,28 @@
 		     : (sizeof (+(Val1)) == sizeof (float))		      \
 		     ? (__tgmath_real_type (Val1)) Fct##f (Val1, Val2)	      \
 		     : (__tgmath_real_type (Val1)) __tgml(Fct) (Val1, Val2)))
+# endif
 
+# if !__HAVE_BUILTIN_TGMATH_C2X
 #  define __TGMATH_BINARY_REAL_ONLY(Val1, Val2, Fct) \
      (__extension__ ((sizeof ((Val1) + (Val2)) > sizeof (double)	      \
 		      && __builtin_classify_type ((Val1) + (Val2)) == 8)      \
 		     ? __TGMATH_F128 ((Val1) + (Val2),			      \
-				      (__typeof				      \
-				       ((__tgmath_real_type (Val1)) 0	      \
-					+ (__tgmath_real_type (Val2)) 0)) Fct, \
+				      (__tgmath_real_type2 (Val1, Val2)) Fct, \
 				      (Val1, Val2))			      \
-		     (__typeof ((__tgmath_real_type (Val1)) 0		      \
-				+ (__tgmath_real_type (Val2)) 0))	      \
+		     (__tgmath_real_type2 (Val1, Val2))			      \
 		     __tgml(Fct) (Val1, Val2)				      \
 		     : (sizeof (+(Val1)) == sizeof (double)		      \
 			|| sizeof (+(Val2)) == sizeof (double)		      \
 			|| __builtin_classify_type (Val1) != 8		      \
 			|| __builtin_classify_type (Val2) != 8)		      \
-		     ? (__typeof ((__tgmath_real_type (Val1)) 0		      \
-				   + (__tgmath_real_type (Val2)) 0))	      \
+		     ? (__tgmath_real_type2 (Val1, Val2))		      \
 		       Fct (Val1, Val2)					      \
-		     : (__typeof ((__tgmath_real_type (Val1)) 0		      \
-				   + (__tgmath_real_type (Val2)) 0))	      \
+		     : (__tgmath_real_type2 (Val1, Val2))		      \
 		       Fct##f (Val1, Val2)))
+# endif
 
+# if !__HAVE_BUILTIN_TGMATH
 #  define __TGMATH_BINARY_REAL_STD_ONLY(Val1, Val2, Fct) \
      (__extension__ ((sizeof ((Val1) + (Val2)) > sizeof (double)	      \
 		      && __builtin_classify_type ((Val1) + (Val2)) == 8)      \
@@ -401,27 +477,24 @@
 		     : (__typeof ((__tgmath_real_type (Val1)) 0		      \
 				   + (__tgmath_real_type (Val2)) 0))	      \
 		       Fct##f (Val1, Val2)))
+# endif
 
+# if !__HAVE_BUILTIN_TGMATH_C2X
 #  define __TGMATH_TERNARY_FIRST_SECOND_REAL_ONLY(Val1, Val2, Val3, Fct) \
      (__extension__ ((sizeof ((Val1) + (Val2)) > sizeof (double)	      \
 		      && __builtin_classify_type ((Val1) + (Val2)) == 8)      \
 		     ? __TGMATH_F128 ((Val1) + (Val2),			      \
-				      (__typeof				      \
-				       ((__tgmath_real_type (Val1)) 0	      \
-					+ (__tgmath_real_type (Val2)) 0)) Fct, \
+				      (__tgmath_real_type2 (Val1, Val2)) Fct, \
 				      (Val1, Val2, Val3))		      \
-		     (__typeof ((__tgmath_real_type (Val1)) 0		      \
-				+ (__tgmath_real_type (Val2)) 0))	      \
+		     (__tgmath_real_type2 (Val1, Val2))			      \
 		     __tgml(Fct) (Val1, Val2, Val3)			      \
 		     : (sizeof (+(Val1)) == sizeof (double)		      \
 			|| sizeof (+(Val2)) == sizeof (double)		      \
 			|| __builtin_classify_type (Val1) != 8		      \
 			|| __builtin_classify_type (Val2) != 8)		      \
-		     ? (__typeof ((__tgmath_real_type (Val1)) 0		      \
-				   + (__tgmath_real_type (Val2)) 0))	      \
+		     ? (__tgmath_real_type2 (Val1, Val2))		      \
 		       Fct (Val1, Val2, Val3)				      \
-		     : (__typeof ((__tgmath_real_type (Val1)) 0		      \
-				   + (__tgmath_real_type (Val2)) 0))	      \
+		     : (__tgmath_real_type2 (Val1, Val2))		      \
 		       Fct##f (Val1, Val2, Val3)))
 
 #  define __TGMATH_TERNARY_REAL_ONLY(Val1, Val2, Val3, Fct) \
@@ -429,14 +502,10 @@
 		      && __builtin_classify_type ((Val1) + (Val2) + (Val3))   \
 			 == 8)						      \
 		     ? __TGMATH_F128 ((Val1) + (Val2) + (Val3),		      \
-				      (__typeof				      \
-				       ((__tgmath_real_type (Val1)) 0	      \
-					+ (__tgmath_real_type (Val2)) 0	      \
-					+ (__tgmath_real_type (Val3)) 0)) Fct, \
+				      (__tgmath_real_type3 (Val1, Val2,	      \
+							    Val3)) Fct,	      \
 				      (Val1, Val2, Val3))		      \
-		     (__typeof ((__tgmath_real_type (Val1)) 0		      \
-				+ (__tgmath_real_type (Val2)) 0		      \
-				+ (__tgmath_real_type (Val3)) 0))	      \
+		     (__tgmath_real_type3 (Val1, Val2, Val3))		      \
 		       __tgml(Fct) (Val1, Val2, Val3)			      \
 		     : (sizeof (+(Val1)) == sizeof (double)		      \
 			|| sizeof (+(Val2)) == sizeof (double)		      \
@@ -444,15 +513,13 @@
 			|| __builtin_classify_type (Val1) != 8		      \
 			|| __builtin_classify_type (Val2) != 8		      \
 			|| __builtin_classify_type (Val3) != 8)		      \
-		     ? (__typeof ((__tgmath_real_type (Val1)) 0		      \
-				   + (__tgmath_real_type (Val2)) 0	      \
-				   + (__tgmath_real_type (Val3)) 0))	      \
+		     ? (__tgmath_real_type3 (Val1, Val2, Val3))		      \
 		       Fct (Val1, Val2, Val3)				      \
-		     : (__typeof ((__tgmath_real_type (Val1)) 0		      \
-				   + (__tgmath_real_type (Val2)) 0	      \
-				   + (__tgmath_real_type (Val3)) 0))	      \
+		     : (__tgmath_real_type3 (Val1, Val2, Val3))		      \
 		       Fct##f (Val1, Val2, Val3)))
+# endif
 
+# if !__HAVE_BUILTIN_TGMATH
 #  define __TGMATH_TERNARY_FIRST_REAL_RET_ONLY(Val1, Val2, Val3, Fct) \
      (__extension__ ((sizeof (+(Val1)) == sizeof (double)		\
 		      || __builtin_classify_type (Val1) != 8)		\
@@ -528,7 +595,9 @@
 		      __tgml(Cfct) (Val))))
 #  define __TGMATH_UNARY_REAL_IMAG_RET_REAL_SAME(Val, Cfct)	\
   __TGMATH_UNARY_REAL_IMAG_RET_REAL ((Val), Cfct, Cfct)
+# endif
 
+# if !__HAVE_BUILTIN_TGMATH_C2X
 /* XXX This definition has to be changed as soon as the compiler understands
    the imaginary keyword.  */
 #  define __TGMATH_BINARY_REAL_IMAG(Val1, Val2, Fct, Cfct) \
@@ -537,41 +606,33 @@
 		      && __builtin_classify_type (__real__ (Val1)	      \
 						  + __real__ (Val2)) == 8)    \
 		     ? __TGMATH_CF128 ((Val1) + (Val2),			      \
-				       (__typeof			      \
-					((__tgmath_complex_type (Val1)) 0     \
-					 + (__tgmath_complex_type (Val2)) 0)) \
+				       (__tgmath_complex_type2 (Val1, Val2))  \
 				       Fct,				      \
-				       (__typeof			      \
-					((__tgmath_complex_type (Val1)) 0     \
-					 + (__tgmath_complex_type (Val2)) 0)) \
+				       (__tgmath_complex_type2 (Val1, Val2))  \
 				       Cfct,				      \
 				       (Val1, Val2))			      \
 		     (__expr_is_real ((Val1) + (Val2))			      \
-		      ? (__typeof ((__tgmath_complex_type (Val1)) 0	      \
-				   + (__tgmath_complex_type (Val2)) 0))	      \
+		      ? (__tgmath_complex_type2 (Val1, Val2))		      \
 		      __tgml(Fct) (Val1, Val2)				      \
-		      : (__typeof ((__tgmath_complex_type (Val1)) 0	      \
-				   + (__tgmath_complex_type (Val2)) 0))	      \
+		      : (__tgmath_complex_type2 (Val1, Val2))		      \
 		      __tgml(Cfct) (Val1, Val2))			      \
 		     : (sizeof (+__real__ (Val1)) == sizeof (double)	      \
 			|| sizeof (+__real__ (Val2)) == sizeof (double)	      \
 			|| __builtin_classify_type (__real__ (Val1)) != 8     \
 			|| __builtin_classify_type (__real__ (Val2)) != 8)    \
 		     ? (__expr_is_real ((Val1) + (Val2))		      \
-			? (__typeof ((__tgmath_complex_type (Val1)) 0	      \
-				   + (__tgmath_complex_type (Val2)) 0))	      \
+			? (__tgmath_complex_type2 (Val1, Val2))		      \
 			  Fct (Val1, Val2)				      \
-			: (__typeof ((__tgmath_complex_type (Val1)) 0	      \
-				   + (__tgmath_complex_type (Val2)) 0))	      \
+			: (__tgmath_complex_type2 (Val1, Val2))		      \
 			  Cfct (Val1, Val2))				      \
 		     : (__expr_is_real ((Val1) + (Val2))		      \
-			? (__typeof ((__tgmath_complex_type (Val1)) 0	      \
-				   + (__tgmath_complex_type (Val2)) 0))	      \
+			? (__tgmath_complex_type2 (Val1, Val2))		      \
 			  Fct##f (Val1, Val2)				      \
-			: (__typeof ((__tgmath_complex_type (Val1)) 0	      \
-				   + (__tgmath_complex_type (Val2)) 0))	      \
+			: (__tgmath_complex_type2 (Val1, Val2))		      \
 			  Cfct##f (Val1, Val2))))
+# endif
 
+# if !__HAVE_BUILTIN_TGMATH
 #  define __TGMATH_1_NARROW_F(F, X)					\
   (__extension__ (sizeof ((__tgmath_real_type (X)) 0) > sizeof (double) \
 		  ? F ## l (X)						\
@@ -587,6 +648,7 @@
 			  + (__tgmath_real_type (Z)) 0) > sizeof (double) \
 		  ? F ## l (X, Y, Z)					\
 		  : F (X, Y, Z)))
+# endif
 /* In most cases, these narrowing macro definitions based on sizeof
    ensure that the function called has the right argument format, as
    for other <tgmath.h> macros for compilers before GCC 8, but may not
@@ -595,60 +657,74 @@
 
    In the case of macros for _Float32x return type, when _Float64x
    exists, _Float64 arguments should result in the *f64 function being
-   called while _Float32x arguments should result in the *f64x
-   function being called.  These cases cannot be distinguished using
-   sizeof (or at all if the types are typedefs rather than different
-   types).  However, for these functions it is OK (does not affect the
-   final result) to call a function with any argument format at least
-   as wide as all the floating-point arguments, unless that affects
-   rounding of integer arguments.  Integer arguments are considered to
-   have type _Float64, so the *f64 functions are preferred for f32x*
-   macros when no argument has a wider floating-point type.  */
-#  if __HAVE_FLOAT64X_LONG_DOUBLE && __HAVE_DISTINCT_FLOAT128
+   called while _Float32x, float and double arguments should result in
+   the *f64x function being called (and integer arguments are
+   considered to have type _Float32x if any argument has type
+   _FloatNx, or double otherwise).  These cases cannot be
+   distinguished using sizeof (or at all if the types are typedefs
+   rather than different types, in which case we err on the side of
+   using the wider type if unsure).  */
+# if !__HAVE_BUILTIN_TGMATH_C2X
+#  if __HAVE_FLOATN_NOT_TYPEDEF
+#   define __TGMATH_NARROW_F32X_USE_F64X(X)			\
+  !__builtin_types_compatible_p (__typeof (+(X)), _Float64)
+#  else
+#   define __TGMATH_NARROW_F32X_USE_F64X(X)			\
+  (__builtin_types_compatible_p (__typeof (+(X)), double)	\
+   || __builtin_types_compatible_p (__typeof (+(X)), float)	\
+   || !__floating_type (__typeof (+(X))))
+#  endif
+# endif
+# if __HAVE_FLOAT64X_LONG_DOUBLE && __HAVE_DISTINCT_FLOAT128
+#  if !__HAVE_BUILTIN_TGMATH
 #   define __TGMATH_1_NARROW_F32(F, X)					\
   (__extension__ (sizeof ((__tgmath_real_type (X)) 0) > sizeof (_Float64) \
-		  ? __TGMATH_F128 ((X), F, (X))				\
+		  ? __TGMATH_F128LD ((X), F, (X))			\
 		  F ## f64x (X)						\
 		  : F ## f64 (X)))
 #   define __TGMATH_2_NARROW_F32(F, X, Y)				\
   (__extension__ (sizeof ((__tgmath_real_type (X)) 0			\
 			  + (__tgmath_real_type (Y)) 0) > sizeof (_Float64) \
-		  ? __TGMATH_F128 ((X) + (Y), F, (X, Y))		\
+		  ? __TGMATH_F128LD ((X) + (Y), F, (X, Y))		\
 		  F ## f64x (X, Y)					\
 		  : F ## f64 (X, Y)))
 #   define __TGMATH_3_NARROW_F32(F, X, Y, Z)				\
   (__extension__ (sizeof ((__tgmath_real_type (X)) 0			\
 			  + (__tgmath_real_type (Y)) 0			\
 			  + (__tgmath_real_type (Z)) 0) > sizeof (_Float64) \
-		  ? __TGMATH_F128 ((X) + (Y) + (Z), F, (X, Y, Z))	\
+		  ? __TGMATH_F128LD ((X) + (Y) + (Z), F, (X, Y, Z))	\
 		  F ## f64x (X, Y, Z)					\
 		  : F ## f64 (X, Y, Z)))
 #   define __TGMATH_1_NARROW_F64(F, X)					\
   (__extension__ (sizeof ((__tgmath_real_type (X)) 0) > sizeof (_Float64) \
-		  ? __TGMATH_F128 ((X), F, (X))				\
+		  ? __TGMATH_F128LD ((X), F, (X))			\
 		  F ## f64x (X)						\
 		  : F ## f128 (X)))
 #   define __TGMATH_2_NARROW_F64(F, X, Y)				\
   (__extension__ (sizeof ((__tgmath_real_type (X)) 0			\
 			  + (__tgmath_real_type (Y)) 0) > sizeof (_Float64) \
-		  ? __TGMATH_F128 ((X) + (Y), F, (X, Y))		\
+		  ? __TGMATH_F128LD ((X) + (Y), F, (X, Y))		\
 		  F ## f64x (X, Y)					\
 		  : F ## f128 (X, Y)))
 #   define __TGMATH_3_NARROW_F64(F, X, Y, Z)				\
   (__extension__ (sizeof ((__tgmath_real_type (X)) 0			\
 			  + (__tgmath_real_type (Y)) 0			\
 			  + (__tgmath_real_type (Z)) 0) > sizeof (_Float64) \
-		  ? __TGMATH_F128 ((X) + (Y) + (Z), F, (X, Y, Z))	\
+		  ? __TGMATH_F128LD ((X) + (Y) + (Z), F, (X, Y, Z))	\
 		  F ## f64x (X, Y, Z)					\
 		  : F ## f128 (X, Y, Z)))
+#  endif
+#  if !__HAVE_BUILTIN_TGMATH_C2X
 #   define __TGMATH_1_NARROW_F32X(F, X)					\
   (__extension__ (sizeof ((__tgmath_real_type (X)) 0) > sizeof (_Float64) \
+		  || __TGMATH_NARROW_F32X_USE_F64X (X)			\
 		  ? __TGMATH_F128 ((X), F, (X))				\
 		  F ## f64x (X)						\
 		  : F ## f64 (X)))
 #   define __TGMATH_2_NARROW_F32X(F, X, Y)				\
   (__extension__ (sizeof ((__tgmath_real_type (X)) 0			\
 			  + (__tgmath_real_type (Y)) 0) > sizeof (_Float64) \
+		  || __TGMATH_NARROW_F32X_USE_F64X ((X) + (Y))		\
 		  ? __TGMATH_F128 ((X) + (Y), F, (X, Y))		\
 		  F ## f64x (X, Y)					\
 		  : F ## f64 (X, Y)))
@@ -656,10 +732,13 @@
   (__extension__ (sizeof ((__tgmath_real_type (X)) 0			\
 			  + (__tgmath_real_type (Y)) 0			\
 			  + (__tgmath_real_type (Z)) 0) > sizeof (_Float64) \
+		  || __TGMATH_NARROW_F32X_USE_F64X ((X) + (Y) + (Z))	\
 		  ? __TGMATH_F128 ((X) + (Y) + (Z), F, (X, Y, Z))	\
 		  F ## f64x (X, Y, Z)					\
 		  : F ## f64 (X, Y, Z)))
-#  elif __HAVE_FLOAT128
+#  endif
+# elif __HAVE_FLOAT128
+#  if !__HAVE_BUILTIN_TGMATH
 #   define __TGMATH_1_NARROW_F32(F, X)					\
   (__extension__ (sizeof ((__tgmath_real_type (X)) 0) > sizeof (_Float64) \
 		  ? F ## f128 (X)					\
@@ -681,22 +760,29 @@
   (F ## f128 (X, Y))
 #   define __TGMATH_3_NARROW_F64(F, X, Y, Z)	\
   (F ## f128 (X, Y, Z))
+#  endif
+#  if !__HAVE_BUILTIN_TGMATH_C2X
 #   define __TGMATH_1_NARROW_F32X(F, X)					\
   (__extension__ (sizeof ((__tgmath_real_type (X)) 0) > sizeof (_Float32x) \
+		  || __TGMATH_NARROW_F32X_USE_F64X (X)			\
 		  ? F ## f64x (X)					\
 		  : F ## f64 (X)))
 #   define __TGMATH_2_NARROW_F32X(F, X, Y)				\
   (__extension__ (sizeof ((__tgmath_real_type (X)) 0			\
 			  + (__tgmath_real_type (Y)) 0) > sizeof (_Float32x) \
+		  || __TGMATH_NARROW_F32X_USE_F64X ((X) + (Y))		\
 		  ? F ## f64x (X, Y)					\
 		  : F ## f64 (X, Y)))
 #   define __TGMATH_3_NARROW_F32X(F, X, Y, Z)				\
   (__extension__ (sizeof ((__tgmath_real_type (X)) 0			\
 			  + (__tgmath_real_type (Y)) 0			\
 			  + (__tgmath_real_type (Z)) 0) > sizeof (_Float32x) \
+		  || __TGMATH_NARROW_F32X_USE_F64X ((X) + (Y) + (Z))	\
 		  ? F ## f64x (X, Y, Z)					\
 		  : F ## f64 (X, Y, Z)))
-#  else
+#  endif
+# else
+#  if !__HAVE_BUILTIN_TGMATH
 #   define __TGMATH_1_NARROW_F32(F, X)		\
   (F ## f64 (X))
 #   define __TGMATH_2_NARROW_F32(F, X, Y)	\
@@ -704,7 +790,7 @@
 #   define __TGMATH_3_NARROW_F32(F, X, Y, Z)	\
   (F ## f64 (X, Y, Z))
 #  endif
-# endif /* !__HAVE_BUILTIN_TGMATH.  */
+# endif
 #else
 # error "Unsupported compiler; you cannot use <tgmath.h>"
 #endif