[mLib] / m4 / mdw-probe-fltfmt.m4

dnl -*-autoconf-*-

### SYNOPSIS
###
###   mdw_PROBE_FLTFMT
###
### DESCRIPTION
###
###   Attempts to determine the target system's floating-point formats.
###   The macros `FLT_FORMAT', `DBL_FORMAT', and `LDBL_FORMAT' are defined if
###   the corresponding formats are recognized.  The values of these macros
###   are a bitmask:
###
###	* `FLTFMT_ORGMASK' is a bitmask covering the `organization' field,
###	  which identifies the organization which defined the format.
###
###	    -- `FLTFMT_IEEE' identifies IEEE 754.
###	    -- `FLTFMT_INTEL' identifies Intel.
###
###     * `FLTFMT_TYPEMASK' is a bitmask covering the `type' field, which
###	  must be interpreted together with the organization field to
###	  determine the format.
###
###	    -- `FLTFMT_IEEE_F32' is the IEEE 754 `binary32' format.
###	    -- `FLTFMT_IEEE_F64' is the IEEE 754 `binary64' format.
###	    -- `FLTFMT_IEEE_F128' is the IEEE 754 `binary128' format.
###	    -- `FLTFMT_INTEL_F80' is the Intel x87 80-bit double-extended
###	       format.
###
###     * `FLTFMT_ENDMASK' is a bitmask covering the `endian' field, which
###	  describes the byte ordering convention used.
###
###	    -- `FLTFMT_LE' means little-endian format.
###
###	    -- `FLTFMT_BE' means big-endian format.
###
###	    -- `FLTFMT_ARME' means `Acorn's ridiculous mixed-endian' format,
###	       used by the ARM FPA, which stored the a `binary64' value as
###	       two 32-bit words, most significant word first, but with the
###	       individual words stored in little-endian byte order.
###
###	  Other conventions exist and support for them may be added later.
###
###   Finally, there are predefined values for relevant combinations of
###   format codes and byte orderings:
###
###     * `FLTFMT_IEEE_F32_LE' and `FLTFMT_IEEE_F32_BE';
###     * `FLTFMT_IEEE_F64_LE' and `FLTFMT_IEEE_F64_BE';
###     * `FLTFMT_IEEE_F128_LE' and `FLTFMT_IEEE_F128_BE';
###     * `FLTFMT_INTEL_F80_LE' and `FLTFMT_INTEL_F80_BE'.
###
###   (I don't know of anything which uses Intel's double-extended format in
###   big-endian order, but it was easy enough to check for.  The IEEE
###   formats are used on ARM64 and z/Architecture with opposite byte order.)
###
###   This macro works correctly when cross-compiling.
###
### LICENSE
###
###   Copyright (c) 2024 Mark Wooding <mdw@distorted.org.uk>
###
###   This program is free software: you can redistribute it and/or modify it
###   under the terms of the GNU General Public License as published by the
###   Free Software Foundation, either version 2 of the License, or (at your
###   option) any later version.
###
###   This program is distributed in the hope that it will be useful, but
###   WITHOUT ANY WARRANTY; without even the implied warranty of
###   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
###   General Public License for more details.
###
###   You should have received a copy of the GNU General Public License along
###   with this program. If not, see <http://www.gnu.org/licenses/>.
###
###   In particular, no exception to the GPL is granted regarding generated
###   `configure' scripts which are the output of Autoconf.

dnl Principle of operation:
dnl
dnl The essential trick here lies in finding floating-point numbers whose
dnl encoding, in various formats of interest, happen to be recognizable
dnl diagnostic text strings.  The structure definitions provide some space
dnl for framing text which allows us to scrape the resulting diagnostic
dnl strings from the object file.
dnl
dnl IEEE formats conveniently don't impose any restrictions on the contents
dnl of the fraction field because there's a hidden bit.  The Intel x87
dnl double-extended format makes the most significant bit explicit, and also
dnl expects normalization, which means that the top bit of the fraction bytes
dnl must be set, so everything gets quite ugly.  Worse, the actual data is 10
dnl bytes long, but it sits in a 16-byte field to force alignment in vectors,
dnl with the result that there are unavoidably zero bytes in our diagnostic
dnl output.  Different tools respond to these differently; e.g., GNU sed(1)
dnl just writes out the zero bytes like they were any other character, while
dnl Busybox sed(1) terminates the output line.  As a result, we have to be
dnl rather more flexible about matching this than I'd really like.  (Messing
dnl about with compiler-specific hacks for structure packing won't help here
dnl because the analysis code still has to cope with compilers which don't
dnl have those hacks.)

# Serial 1
AC_COPYRIGHT([
Portions copyright (c) 2024 Mark Wooding.

This configure script is free software: you can redistribute it and/or
modify it under he terms of the GNU General Public License as published
by the Free Software Foundation, either version 2 of the License, or
(at your option) any later version.])

AC_DEFUN([mdw_PROBE_FLTFMT],
  [AC_CACHE_CHECK([floating-point representations], [mdw_cv_fltfmt],
     [mdw_fltfmt=nil mdw_dblfmt=nil mdw_ldblfmt=nil
      AC_LINK_IFELSE([AC_LANG_SOURCE([

/* The following program is copyright (c) 2024 Mark Wooding.  It is free
 * software: you can redistribute it and/or modify it under the terms of the
 * GNU General Public License as published by the Free Software Foundation,
 * either version 2 of the License, or (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful, but
 * WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
 * or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
 * for more details.
 *
 * You should have received a copy of the GNU General Public License along
 * with this program. If not, see <http://www.gnu.org/licenses/>.
 */

@%:@include <float.h>

@%:@define DEFFLTDIAG(ty, type)						\
  struct fltdiag_@%:@@%:@ty { char top@<:@48@:>@; type x; char tail@<:@4@:>@; }

DEFFLTDIAG(flt, float);

static const struct fltdiag_flt flt_ieee_f32 = {
  "\0\0\0\0\0\0\0\0\n@@@ mdw-probe-fltfmt float ieee-f32 = >",
  781.0352,
  "<\n\0\0"
};
@%:@define DO_FLT_IEEE_F32 DO(flt_ieee_f32)

DEFFLTDIAG(dbl, double);

@%:@if DBL_MAX_10_EXP > 40
static const struct fltdiag_dbl dbl_ieee_f64 = {
  "\0\0\0\0\0\0\0\n@@@ mdw-probe-fltfmt double ieee-f64 = >",
  1.5839800103804824e40,
  "<\n\0\0"
};
@%:@  define DO_DBL_IEEE_F64 DO(dbl_ieee_f64)
@%:@else
@%:@  define DO_DBL_IEEE_F64
@%:@endif

DEFFLTDIAG(ldbl, long double);

@%:@if LDBL_MAX_10_EXP > 40
static const struct fltdiag_ldbl ldbl_ieee_f64 = {
  "\0\0\n@@@ mdw-probe-fltfmt long-double ieee-f64 = >",
  1.5839800103804824e40,
  "<\n\0\0"
};
@%:@  define DO_LDBL_IEEE_F64 DO(ldbl_ieee_f64)
@%:@else
@%:@  define DO_LDBL_IEEE_F64
@%:@endif

@%:@if LDBL_MAX_10_EXP > 1257
static const struct fltdiag_ldbl ldbl_ieee_f128 = {
  "\0\n@@@ mdw-probe-fltfmt long-double ieee-f128 = >",
  1.6487728650847311136108983312706536e+1257L,
  "<\n\0\0"
};
@%:@  define DO_LDBL_IEEE_F128 DO(ldbl_ieee_f128)
@%:@else
@%:@  define DO_LDBL_IEEE_F128
@%:@endif

@%:@if LDBL_MAX_10_EXP > 793
static const struct fltdiag_ldbl ldbl_intel_f80 = {
  "\0\n@@@ mdw-probe-fltfmt long-double intel-f80 = >",
  1.2806567921142816197e+793L,
  "<\n\0\0"
};
@%:@  define DO_LDBL_INTEL_F80 DO(ldbl_intel_f80)
@%:@else
@%:@  define DO_LDBL_INTEL_F80
@%:@endif

@%:@include <stdio.h>
int main(void)
{
@%:@define DO(var) fwrite(&var, sizeof(var), 1, stdout)
  DO_FLT_IEEE_F32;
  DO_DBL_IEEE_F64;
  DO_LDBL_IEEE_F64;
  DO_LDBL_IEEE_F128;
  DO_LDBL_IEEE_F128;
  DO_LDBL_INTEL_F80;
@%:@undef DO
  return (0);
}
])],
	[sed -n "/^@@@ mdw-probe-fltfmt @<:@^ @:>@* @<:@^ @:>@* = >/p" \
		 conftest$EXEEXT >conftest.out
	 while read _at _tag ty fmt _eq diag; do
	   case $ty,$fmt,$diag in
	     "float,ieee-f32,>ABCD<") mdw_fltfmt=ieee-f32-le ;;
	     "float,ieee-f32,>DCBA<") mdw_fltfmt=ieee-f32-be ;;
	     "double,ieee-f64,>ABCDEFGH<") mdw_dblfmt=ieee-f64-le ;;
	     "double,ieee-f64,>EFGHABCD<") mdw_dblfmt=ieee-f64-arme ;;
	     "double,ieee-f64,>HGFEDCBA<") mdw_dblfmt=ieee-f64-be ;;
	     "long-double,ieee-f64,>ABCDEFGH<") mdw_ldblfmt=ieee-f64-le ;;
	     "long-double,ieee-f64,>HGFEDCBA<") mdw_ldblfmt=ieee-f64-be ;;
	     "long-double,ieee-f128,>ABCDEFGHIJKLMNOP<")
	       mdw_ldblfmt=ieee-f128-le ;;
	     "long-double,ieee-f128,>PONMLKJIHGFEDCBA<")
	       mdw_ldblfmt=ieee-f128-be ;;
Commit	Line	Data
b1a20bee MW	1	dnl --autoconf--
	2
	3	### SYNOPSIS
	4	###
	5	### mdw_PROBE_FLTFMT
	6	###
	7	### DESCRIPTION
	8	###
	9	### Attempts to determine the target system's floating-point formats.
	10	### The macros `FLT_FORMAT', `DBL_FORMAT', and `LDBL_FORMAT' are defined if
	11	### the corresponding formats are recognized. The values of these macros
	12	### are a bitmask:
	13	###
	14	### * `FLTFMT_ORGMASK' is a bitmask covering the `organization' field,
	15	### which identifies the organization which defined the format.
	16	###
	17	### -- `FLTFMT_IEEE' identifies IEEE 754.
	18	### -- `FLTFMT_INTEL' identifies Intel.
	19	###
	20	### * `FLTFMT_TYPEMASK' is a bitmask covering the `type' field, which
	21	### must be interpreted together with the organization field to
	22	### determine the format.
	23	###
	24	### -- `FLTFMT_IEEE_F32' is the IEEE 754 `binary32' format.
	25	### -- `FLTFMT_IEEE_F64' is the IEEE 754 `binary64' format.
	26	### -- `FLTFMT_IEEE_F128' is the IEEE 754 `binary128' format.
	27	### -- `FLTFMT_INTEL_F80' is the Intel x87 80-bit double-extended
	28	### format.
	29	###
	30	### * `FLTFMT_ENDMASK' is a bitmask covering the `endian' field, which
	31	### describes the byte ordering convention used.
	32	###
	33	### -- `FLTFMT_LE' means little-endian format.
	34	###
	35	### -- `FLTFMT_BE' means big-endian format.
	36	###
	37	### -- `FLTFMT_ARME' means `Acorn's ridiculous mixed-endian' format,
	38	### used by the ARM FPA, which stored the a `binary64' value as
	39	### two 32-bit words, most significant word first, but with the
	40	### individual words stored in little-endian byte order.
	41	###
	42	### Other conventions exist and support for them may be added later.
	43	###
	44	### Finally, there are predefined values for relevant combinations of
	45	### format codes and byte orderings:
	46	###
	47	### * `FLTFMT_IEEE_F32_LE' and `FLTFMT_IEEE_F32_BE';
	48	### * `FLTFMT_IEEE_F64_LE' and `FLTFMT_IEEE_F64_BE';
	49	### * `FLTFMT_IEEE_F128_LE' and `FLTFMT_IEEE_F128_BE';
	50	### * `FLTFMT_INTEL_F80_LE' and `FLTFMT_INTEL_F80_BE'.
	51	###
	52	### (I don't know of anything which uses Intel's double-extended format in
	53	### big-endian order, but it was easy enough to check for. The IEEE
	54	### formats are used on ARM64 and z/Architecture with opposite byte order.)
	55	###
	56	### This macro works correctly when cross-compiling.
	57	###
	58	### LICENSE
	59	###
	60	### Copyright (c) 2024 Mark Wooding <mdw@distorted.org.uk>
	61	###
	62	### This program is free software: you can redistribute it and/or modify it
	63	### under the terms of the GNU General Public License as published by the
	64	### Free Software Foundation, either version 2 of the License, or (at your
65	### option) any later version.
66	###
67	### This program is distributed in the hope that it will be useful, but
68	### WITHOUT ANY WARRANTY; without even the implied warranty of
69	### MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
70	### General Public License for more details.
71	###
72	### You should have received a copy of the GNU General Public License along
73	### with this program. If not, see <http://www.gnu.org/licenses/>.
74	###
75	### In particular, no exception to the GPL is granted regarding generated
76	### `configure' scripts which are the output of Autoconf.
77
78	dnl Principle of operation:
79	dnl
80	dnl The essential trick here lies in finding floating-point numbers whose
81	dnl encoding, in various formats of interest, happen to be recognizable
82	dnl diagnostic text strings. The structure definitions provide some space
83	dnl for framing text which allows us to scrape the resulting diagnostic
84	dnl strings from the object file.
85	dnl
86	dnl IEEE formats conveniently don't impose any restrictions on the contents
87	dnl of the fraction field because there's a hidden bit. The Intel x87
88	dnl double-extended format makes the most significant bit explicit, and also
89	dnl expects normalization, which means that the top bit of the fraction bytes
90	dnl must be set, so everything gets quite ugly. Worse, the actual data is 10
91	dnl bytes long, but it sits in a 16-byte field to force alignment in vectors,
92	dnl with the result that there are unavoidably zero bytes in our diagnostic
93	dnl output. Different tools respond to these differently; e.g., GNU sed(1)
94	dnl just writes out the zero bytes like they were any other character, while
95	dnl Busybox sed(1) terminates the output line. As a result, we have to be
96	dnl rather more flexible about matching this than I'd really like. (Messing
97	dnl about with compiler-specific hacks for structure packing won't help here
98	dnl because the analysis code still has to cope with compilers which don't
99	dnl have those hacks.)
100
101	# Serial 1
102	AC_COPYRIGHT([
103	Portions copyright (c) 2024 Mark Wooding.
104
105	This configure script is free software: you can redistribute it and/or
106	modify it under he terms of the GNU General Public License as published
107	by the Free Software Foundation, either version 2 of the License, or
108	(at your option) any later version.])
109
110	AC_DEFUN([mdw_PROBE_FLTFMT],
111	[AC_CACHE_CHECK([floating-point representations], [mdw_cv_fltfmt],
112	[mdw_fltfmt=nil mdw_dblfmt=nil mdw_ldblfmt=nil
113	AC_LINK_IFELSE([AC_LANG_SOURCE([
114
115	/* The following program is copyright (c) 2024 Mark Wooding. It is free
116	* software: you can redistribute it and/or modify it under the terms of the
117	* GNU General Public License as published by the Free Software Foundation,
118	* either version 2 of the License, or (at your option) any later version.
119	*
120	* This program is distributed in the hope that it will be useful, but
121	* WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
122	* or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
123	* for more details.
124	*
125	* You should have received a copy of the GNU General Public License along
126	* with this program. If not, see <http://www.gnu.org/licenses/>.
127	*/
128
129	@%:@include <float.h>
130
131	@%:@define DEFFLTDIAG(ty, type) \
132	struct fltdiag_@%:@@%:@ty { char top@<:@48@:>@; type x; char tail@<:@4@:>@; }
133
134	DEFFLTDIAG(flt, float);
135
136	static const struct fltdiag_flt flt_ieee_f32 = {
137	"\0\0\0\0\0\0\0\0\n@@@ mdw-probe-fltfmt float ieee-f32 = >",
138	781.0352,
139	"<\n\0\0"
140	};
141	@%:@define DO_FLT_IEEE_F32 DO(flt_ieee_f32)
142
143	DEFFLTDIAG(dbl, double);
144
145	@%:@if DBL_MAX_10_EXP > 40
146	static const struct fltdiag_dbl dbl_ieee_f64 = {
147	"\0\0\0\0\0\0\0\n@@@ mdw-probe-fltfmt double ieee-f64 = >",
148	1.5839800103804824e40,
149	"<\n\0\0"
150	};
151	@%:@ define DO_DBL_IEEE_F64 DO(dbl_ieee_f64)
152	@%:@else
153	@%:@ define DO_DBL_IEEE_F64
154	@%:@endif
155
156	DEFFLTDIAG(ldbl, long double);
157
158	@%:@if LDBL_MAX_10_EXP > 40
159	static const struct fltdiag_ldbl ldbl_ieee_f64 = {
160	"\0\0\n@@@ mdw-probe-fltfmt long-double ieee-f64 = >",
161	1.5839800103804824e40,
162	"<\n\0\0"
163	};
164	@%:@ define DO_LDBL_IEEE_F64 DO(ldbl_ieee_f64)
165	@%:@else
166	@%:@ define DO_LDBL_IEEE_F64
167	@%:@endif
168
169	@%:@if LDBL_MAX_10_EXP > 1257
170	static const struct fltdiag_ldbl ldbl_ieee_f128 = {
171	"\0\n@@@ mdw-probe-fltfmt long-double ieee-f128 = >",
172	1.6487728650847311136108983312706536e+1257L,
173	"<\n\0\0"
174	};
175	@%:@ define DO_LDBL_IEEE_F128 DO(ldbl_ieee_f128)
176	@%:@else
177	@%:@ define DO_LDBL_IEEE_F128
178	@%:@endif
179
180	@%:@if LDBL_MAX_10_EXP > 793
181	static const struct fltdiag_ldbl ldbl_intel_f80 = {
182	"\0\n@@@ mdw-probe-fltfmt long-double intel-f80 = >",
183	1.2806567921142816197e+793L,
184	"<\n\0\0"
185	};
186	@%:@ define DO_LDBL_INTEL_F80 DO(ldbl_intel_f80)
187	@%:@else
188	@%:@ define DO_LDBL_INTEL_F80
189	@%:@endif
190
191	@%:@include <stdio.h>
192	int main(void)
193	{
194	@%:@define DO(var) fwrite(&var, sizeof(var), 1, stdout)
195	DO_FLT_IEEE_F32;
196	DO_DBL_IEEE_F64;
197	DO_LDBL_IEEE_F64;
198	DO_LDBL_IEEE_F128;
199	DO_LDBL_IEEE_F128;
200	DO_LDBL_INTEL_F80;
201	@%:@undef DO
202	return (0);
203	}
204	])],
205	[sed -n "/^@@@ mdw-probe-fltfmt @<:@^ @:>@* @<:@^ @:>@* = >/p" \
206	conftest$EXEEXT >conftest.out
207	while read _at _tag ty fmt _eq diag; do
208	case $ty,$fmt,$diag in
209	"float,ieee-f32,>ABCD<") mdw_fltfmt=ieee-f32-le ;;
210	"float,ieee-f32,>DCBA<") mdw_fltfmt=ieee-f32-be ;;
211	"double,ieee-f64,>ABCDEFGH<") mdw_dblfmt=ieee-f64-le ;;
212	"double,ieee-f64,>EFGHABCD<") mdw_dblfmt=ieee-f64-arme ;;
213	"double,ieee-f64,>HGFEDCBA<") mdw_dblfmt=ieee-f64-be ;;
214	"long-double,ieee-f64,>ABCDEFGH<") mdw_ldblfmt=ieee-f64-le ;;
215	"long-double,ieee-f64,>HGFEDCBA<") mdw_ldblfmt=ieee-f64-be ;;
216	"long-double,ieee-f128,>ABCDEFGHIJKLMNOP<")
217	mdw_ldblfmt=ieee-f128-le ;;
218	"long-double,ieee-f128,>PONMLKJIHGFEDCBA<")
219	mdw_ldblfmt=ieee-f128-be ;;
220