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, 131 insertions, 82 deletions

diff --git a/math/gen-tgmath-tests.py b/math/gen-tgmath-tests.py
index c841db6000..be10c47877 100755
--- a/math/gen-tgmath-tests.py
+++ b/math/gen-tgmath-tests.py
@@ -19,14 +19,13 @@
 
 # As glibc does not support decimal floating point, the types to
 # consider for generic parameters are standard and binary
-# floating-point types, and integer types which are treated as double.
-# The corresponding complex types may also be used (including complex
-# integer types, which are a GNU extension, but are currently disabled
-# here because they do not work properly with tgmath.h).
-
-# The proposed resolution to TS 18661-1 DR#9
-# <http://www.open-std.org/jtc1/sc22/wg14/www/docs/n2149.htm#dr_9>
-# makes the <tgmath.h> rules for selecting a function to call
+# floating-point types, and integer types which are treated as
+# _Float32x if any argument has a _FloatNx type and otherwise as
+# double.  The corresponding complex types may also be used (including
+# complex integer types, which are a GNU extension, but are currently
+# disabled here because they do not work properly with tgmath.h).
+
+# C2x makes the <tgmath.h> rules for selecting a function to call
 # correspond to the usual arithmetic conversions (applied successively
 # to the arguments for generic parameters in order), which choose the
 # type whose set of values contains that of the other type (undefined
@@ -69,10 +68,6 @@ class Type(object):
     # Real argument types that correspond to a standard floating type
     # (float, double or long double; not _FloatN or _FloatNx).
     standard_real_argument_types_list = []
-    # Real argument types other than float, double and long double
-    # (i.e., those that are valid as arguments to narrowing macros
-    # returning _FloatN or _FloatNx).
-    non_standard_real_argument_types_list = []
     # The real floating types by their order properties (which are
     # tuples giving the positions in both the possible orders above).
     real_types_order = {}
@@ -86,13 +81,16 @@ class Type(object):
     float64_type = None
     # The type _Complex _Float64.
     complex_float64_type = None
+    # The type _Float32x.
+    float32x_type = None
+    # The type _Complex _Float32x.
+    complex_float32x_type = None
     # The type _Float64x.
     float64x_type = None
-    # The type _Float64x if available, otherwise _Float64.
-    float32x_ext_type = None
 
     def __init__(self, name, suffix=None, mant_dig=None, condition='1',
-                 order=None, integer=False, complex=False, real_type=None):
+                 order=None, integer=False, complex=False, real_type=None,
+                 floatnx=False):
         """Initialize a Type object, creating any corresponding complex type
         in the process."""
         self.name = name
@@ -102,6 +100,7 @@ class Type(object):
         self.order = order
         self.integer = integer
         self.complex = complex
+        self.floatnx = floatnx
         if complex:
             self.complex_type = self
             self.real_type = real_type
@@ -119,8 +118,6 @@ class Type(object):
                 Type.real_argument_types_list.append(self)
                 if not self.name.startswith('_Float'):
                     Type.standard_real_argument_types_list.append(self)
-                if self.name not in ('float', 'double', 'long double'):
-                    Type.non_standard_real_argument_types_list.append(self)
         if self.order is not None:
             Type.real_types_order[self.order] = self
         if self.name == 'double':
@@ -133,26 +130,28 @@ class Type(object):
             Type.float64_type = self
         if self.name == '_Complex _Float64':
             Type.complex_float64_type = self
+        if self.name == '_Float32x':
+            Type.float32x_type = self
+        if self.name == '_Complex _Float32x':
+            Type.complex_float32x_type = self
         if self.name == '_Float64x':
             Type.float64x_type = self
-        if self.name == 'Float32x_ext':
-            Type.float32x_ext_type = self
 
     @staticmethod
     def create_type(name, suffix=None, mant_dig=None, condition='1', order=None,
                     integer=False, complex_name=None, complex_ok=True,
-                    internal=False):
+                    floatnx=False, internal=False):
         """Create and register a Type object for a real type, creating any
         corresponding complex type in the process."""
         real_type = Type(name, suffix=suffix, mant_dig=mant_dig,
                          condition=condition, order=order, integer=integer,
-                         complex=False)
+                         complex=False, floatnx=floatnx)
         if complex_ok:
             if complex_name is None:
                 complex_name = '_Complex %s' % name
             complex_type = Type(complex_name, condition=condition,
                                 integer=integer, complex=True,
-                                real_type=real_type)
+                                real_type=real_type, floatnx=floatnx)
         else:
             complex_type = None
         real_type.complex_type = complex_type
@@ -160,13 +159,13 @@ class Type(object):
         if complex_type is not None:
             complex_type.register_type(internal)
 
-    def floating_type(self, floatn):
+    def floating_type(self, integer_float32x):
         """Return the corresponding floating type."""
         if self.integer:
-            if floatn:
-                return (Type.complex_float64_type
+            if integer_float32x:
+                return (Type.complex_float32x_type
                         if self.complex
-                        else Type.float64_type)
+                        else Type.float32x_type)
             else:
                 return (Type.complex_double_type
                         if self.complex
@@ -174,9 +173,9 @@ class Type(object):
         else:
             return self
 
-    def real_floating_type(self, floatn):
+    def real_floating_type(self, integer_float32x):
         """Return the corresponding real floating type."""
-        return self.real_type.floating_type(floatn)
+        return self.real_type.floating_type(integer_float32x)
 
     def __str__(self):
         """Return string representation of a type."""
@@ -194,7 +193,8 @@ class Type(object):
                          condition='defined HUGE_VAL_F32', order=(2, 2))
         Type.create_type('_Float32x', 'f32x', 'FLT32X_MANT_DIG',
                          complex_name='__CFLOAT32X',
-                         condition='defined HUGE_VAL_F32X', order=(3, 3))
+                         condition='defined HUGE_VAL_F32X', order=(3, 3),
+                         floatnx=True)
         Type.create_type('double', '', 'DBL_MANT_DIG', order=(4, 4))
         Type.create_type('long double', 'l', 'LDBL_MANT_DIG', order=(5, 7))
         Type.create_type('_Float64', 'f64', 'FLT64_MANT_DIG',
@@ -202,7 +202,8 @@ class Type(object):
                          condition='defined HUGE_VAL_F64', order=(6, 5))
         Type.create_type('_Float64x', 'f64x', 'FLT64X_MANT_DIG',
                          complex_name='__CFLOAT64X',
-                         condition='defined HUGE_VAL_F64X', order=(7, 6))
+                         condition='defined HUGE_VAL_F64X', order=(7, 6),
+                         floatnx=True)
         Type.create_type('_Float128', 'f128', 'FLT128_MANT_DIG',
                          complex_name='__CFLOAT128',
                          condition='defined HUGE_VAL_F128', order=(8, 8))
@@ -235,21 +236,16 @@ class Type(object):
                          complex_name='complex_long_double_Float64x',
                          condition='defined HUGE_VAL_F64X', order=(7, 7),
                          internal=True)
-        # An internal type for the argument type used by f32x*
-        # narrowing macros (_Float64x if available, otherwise
-        # _Float64).
-        Type.create_type('Float32x_ext', None, 'FLT32X_EXT_MANT_DIG',
-                         complex_name='complex_Float32x_ext',
-                         condition='1', internal=True)
 
     @staticmethod
-    def can_combine_types(types, floatn):
+    def can_combine_types(types):
         """Return a C preprocessor conditional for whether the given list of
         types can be used together as type-generic macro arguments."""
         have_long_double = False
         have_float128 = False
+        integer_float32x = any(t.floatnx for t in types)
         for t in types:
-            t = t.real_floating_type(floatn)
+            t = t.real_floating_type(integer_float32x)
             if t.name == 'long double':
                 have_long_double = True
             if t.name == '_Float128' or t.name == '_Float64x':
@@ -262,14 +258,15 @@ class Type(object):
         return '1'
 
     @staticmethod
-    def combine_types(types, floatn):
+    def combine_types(types):
         """Return the result of combining a set of types."""
         have_complex = False
         combined = None
+        integer_float32x = any(t.floatnx for t in types)
         for t in types:
             if t.complex:
                 have_complex = True
-            t = t.real_floating_type(floatn)
+            t = t.real_floating_type(integer_float32x)
             if combined is None:
                 combined = t
             else:
@@ -375,18 +372,8 @@ class Tests(object):
                          '# endif\n')
         float64x_text = if_cond_text([Type.float64x_type.condition],
                                      float64x_text)
-        float32x_ext_text = ('#ifdef HUGE_VAL_F64X\n'
-                             'typedef _Float64x Float32x_ext;\n'
-                             'typedef __CFLOAT64X complex_Float32x_ext;\n'
-                             '# define FLT32X_EXT_MANT_DIG FLT64X_MANT_DIG\n'
-                             '#else\n'
-                             'typedef _Float64 Float32x_ext;\n'
-                             'typedef __CFLOAT64 complex_Float32x_ext;\n'
-                             '# define FLT32X_EXT_MANT_DIG FLT64_MANT_DIG\n'
-                             '#endif\n')
         self.header_list.append(float64_text)
         self.header_list.append(float64x_text)
-        self.header_list.append(float32x_ext_text)
         self.types_seen = set()
         for t in Type.all_types_list:
             self.add_type_var(t.name, t.condition)
@@ -439,39 +426,33 @@ class Tests(object):
             narrowing_std = True
             narrow_cond = '1'
             narrow_args = [Type.double_type, Type.long_double_type]
-            narrow_fallback = Type.double_type
         elif ret == 'double':
             narrowing = True
             narrowing_std = True
             narrow_cond = '1'
             narrow_args = [Type.long_double_type]
-            narrow_fallback = Type.long_double_type
         elif ret.startswith('_Float'):
             narrowing = True
-            narrow_args = []
+            narrow_args_1 = []
+            narrow_args_2 = []
             nret_type = None
-            narrow_fallback = None
             for order, real_type in sorted(Type.real_types_order.items()):
                 if real_type.name == ret:
                     nret_type = real_type
                 elif nret_type and real_type.name.startswith('_Float'):
-                    narrow_args.append(real_type)
-                    if (narrow_fallback is None
-                        and ret.endswith('x') == real_type.name.endswith('x')):
-                        narrow_fallback = real_type
+                    if ret.endswith('x') == real_type.name.endswith('x'):
+                        narrow_args_1.append(real_type)
+                    else:
+                        narrow_args_2.append(real_type)
+            narrow_args = narrow_args_1 + narrow_args_2
             if narrow_args:
                 narrow_cond = ('(%s && (%s))'
                                % (nret_type.condition,
                                   ' || '.join(t.condition
                                               for t in narrow_args)))
-                if narrow_fallback is None:
-                    narrow_fallback = narrow_args[0]
-                if ret == '_Float32x':
-                    narrow_fallback = Type.float32x_ext_type
             else:
                 # No possible argument types, even conditionally.
                 narrow_cond = '0'
-        narrowing_nonstd = narrowing and not narrowing_std
         types = [ret] + args
         for t in types:
             if t != 'c' and t != 'g' and t != 'r' and t != 's':
@@ -530,19 +511,13 @@ class Tests(object):
             if t == 'g' or t == 'c':
                 arg_types.append(Type.argument_types_list)
             elif t == 'r':
-                if narrowing_std:
-                    arg_types.append(Type.standard_real_argument_types_list)
-                elif narrowing:
-                    arg_types.append(
-                        Type.non_standard_real_argument_types_list)
-                else:
-                    arg_types.append(Type.real_argument_types_list)
+                arg_types.append(Type.real_argument_types_list)
             elif t == 's':
                 arg_types.append(Type.standard_real_argument_types_list)
         arg_types_product = list_product(arg_types)
         test_num = 0
         for this_args in arg_types_product:
-            comb_type = Type.combine_types(this_args, narrowing_nonstd)
+            comb_type = Type.combine_types(this_args)
             if narrowing:
                 # As long as there are no integer arguments, and as
                 # long as the chosen argument type is as wide as all
@@ -550,22 +525,22 @@ class Tests(object):
                 # of the macro call do not depend on the exact
                 # function chosen.  In particular, for f32x functions
                 # when _Float64x exists, the chosen type should differ
-                # for _Float32x and _Float64 arguments, but it is not
-                # always possible to distinguish those types before
-                # GCC 7 and the implementation does not attempt to do
-                # so before GCC 8.
+                # for double / _Float32x and _Float64 arguments, but
+                # it is not always possible to distinguish those types
+                # before GCC 7 (resulting in some cases - only real
+                # arguments - where a wider argument type is used,
+                # which is semantically OK, and others - integer
+                # arguments present - where it may not be OK, but is
+                # unavoidable).
                 narrow_mant_dig = comb_type.real_type.mant_dig
                 for arg_type in this_args:
                     if arg_type.integer:
                         narrow_mant_dig = 0
             else:
                 narrow_mant_dig = 0
-            if (narrowing
-                and comb_type not in narrow_args
-                and narrow_fallback is not None):
-                comb_type = narrow_fallback
-            can_comb = Type.can_combine_types(this_args, narrowing_nonstd)
+            can_comb = Type.can_combine_types(this_args)
             all_conds = [t.condition for t in this_args]
+            narrow_args_cond = '(%s)' % ' && '.join(sorted(set(all_conds)))
             all_conds.append(can_comb)
             if narrowing:
                 all_conds.append(narrow_cond)
@@ -579,10 +554,69 @@ class Tests(object):
             test_func_name = 'test_%s_%d' % (macro, test_num)
             test_num += 1
             mant_dig = comb_type.real_type.mant_dig
+            test_mant_dig_comp = ''
+            if (narrowing
+                and comb_type not in narrow_args):
+                # The expected argument type is the first in
+                # narrow_args that can represent all the values of
+                # comb_type (which, for the supported cases, means the
+                # first with mant_dig at least as large as that for
+                # comb_type, provided this isn't the case of an IBM
+                # long double argument with binary128 type from
+                # narrow_args).
+                narrow_extra_conds = []
+                test_mant_dig_list = ['#undef NARROW_MANT_DIG\n#if 0\n']
+                for t in narrow_args:
+                    t_cond = '(%s && %s && %s <= %s && %s)' % (
+                        narrow_args_cond, t.condition, mant_dig, t.mant_dig,
+                        Type.can_combine_types(this_args + [t]))
+                    narrow_extra_conds.append(t_cond)
+                    test_mant_dig_list.append('#elif %s\n'
+                                              '#define NARROW_MANT_DIG %s\n'
+                                              % (t_cond, t.mant_dig))
+                test_mant_dig_list.append('#endif\n')
+                test_mant_dig_comp = ''.join(test_mant_dig_list)
+                all_conds.append('(%s)' % ' || '.join(narrow_extra_conds))
+                # A special case where this logic isn't correct is
+                # where comb_type is the internal long_double_Float64
+                # or long_double_Float64x, which will be detected as
+                # not in narrow_args even if the actual type chosen in
+                # a particular configuration would have been in
+                # narrow_args, so check for that case and handle it
+                # appropriately.  In particular, if long double has
+                # the same format as double and there are long double
+                # and _Float64 arguments, and the macro returns
+                # _Float32x, the function called should be one for
+                # _Float64 arguments, not one for _Float64x arguments
+                # that would arise from this logic.
+                if comb_type.real_type.name == 'long_double_Float64':
+                    comb_type_1 = Type.long_double_type
+                    comb_type_2 = Type.float64_type
+                    comb_type_is_2_cond = 'LDBL_MANT_DIG <= FLT64_MANT_DIG'
+                elif comb_type.real_type.name == 'long_double_Float64x':
+                    comb_type_1 = Type.long_double_type
+                    comb_type_2 = Type.float64x_type
+                    comb_type_is_2_cond = 'LDBL_MANT_DIG < FLT64X_MANT_DIG'
+                else:
+                    comb_type_1 = None
+                    comb_type_2 = None
+                if comb_type_1 is None:
+                    mant_dig = 'NARROW_MANT_DIG'
+                else:
+                    mant_dig = ''
+                    if comb_type_1 in narrow_args:
+                        mant_dig += '!(%s) ? %s : ' % (comb_type_is_2_cond,
+                                                       comb_type_1.mant_dig)
+                    if comb_type_2 in narrow_args:
+                        mant_dig += '%s ? %s : ' % (comb_type_is_2_cond,
+                                                    comb_type_2.mant_dig)
+                    mant_dig += 'NARROW_MANT_DIG'
+                if narrow_mant_dig != 0:
+                    narrow_mant_dig = mant_dig
             test_text = '%s, "%s", "%s", %s, %s' % (test_func_name, func_name,
                                                     test_name, mant_dig,
                                                     narrow_mant_dig)
-            test_text = '    { %s },\n' % test_text
+            test_text = '%s    { %s },\n' % (test_mant_dig_comp, test_text)
             test_text = if_cond_text(all_conds, test_text)
             self.test_array_list.append(test_text)
             call_args = []
@@ -742,7 +776,7 @@ class Tests(object):
                        '          && strcmp (called_func_name,\n'
                        '                     tests[i].func_name) == 0)\n'
                        '        num_pass++;\n'
-                       '#if !__GNUC_PREREQ (8, 0)\n'
+                       '#if !__GNUC_PREREQ (7, 0)\n'
                        '      else if (tests[i].narrow_mant_dig > 0\n'
                        '               && (called_mant_dig\n'
                        '                   >= tests[i].narrow_mant_dig)\n'
@@ -759,6 +793,21 @@ class Tests(object):
                        '                  tests[i].mant_dig,\n'
                        '                  called_func_name, called_mant_dig);\n'
                        '        }\n'
+                       '      else if (tests[i].narrow_mant_dig == 0\n'
+                       '               && strcmp (called_func_name,\n'
+                       '                          tests[i].func_name) == 0)\n'
+                       '        {\n'
+                       '          num_pass++;\n'
+                       '          printf ("Test %zu (%s):\\n"\n'
+                       '                  "  Expected: %s precision %d\\n"\n'
+                       '                  "  Actual: %s precision %d\\n"\n'
+                       '                  "  (unavoidable with old GCC)'
+                       '\\n\\n",\n'
+                       '                  i, tests[i].test_name,\n'
+                       '                  tests[i].func_name,\n'
+                       '                  tests[i].mant_dig,\n'
+                       '                  called_func_name, called_mant_dig);\n'
+                       '        }\n'
                        '#endif\n'
                        '      else\n'
                        '        {\n'