Add D.U.M.A., the Electric Fence successor.
1 Taken from http://sources.redhat.com/ml/crossgcc/2004-02/msg00104.html
2 Author: addsub@eyou.com
5 Fixes http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/testsuite/gcc.c-torture/execute/920501-8.c
6 and makes printf("%f", 1.0) work.
8 Lennert Buytenhek wrote in http://sources.redhat.com/ml/crossgcc/2004-09/msg00115.html :
9 It ... fixes the 'printf("%f\n", 0.5); prints 0.000000' and general 'floating point
10 is broken' on my big-endian hardfloat FPA ARM platform. ...
11 It's definitely needed for hardfloat. So I'd think it's needed for
12 big-endian systems in any case, and for VFP on little-endian systems
13 too. Someone would have to verify that though.
15 Lennert Buytenhek wrote in http://sources.redhat.com/ml/crossgcc/2004-09/msg00123.html
16 I just had a look at glibc-20040830, and [this patch] is still needed and useful
17 for this version. glibc-20040830 out-of-the-box still contains the
18 following wrong assumptions:
19 - sysdeps/arm/bits/endian.h: float word order is big endian (which it is
21 - sysdeps/arm/gmp-mparam.h: IEEE doubles are mixed endian (which they
22 are not on big endian systems, neither on vfp systems)
23 - sysdeps/arm/ieee754.h: IEEE doubles are in little endian byte order
24 (which they are not on big endian systems)
25 [This patch] seems the right solution for all of these issues.
27 Dimitry Andric wrote in http://sources.redhat.com/ml/crossgcc/2004-09/msg00132.html :
28 It's even needed for glibc CVS, AFAICS.
29 The patch hunk which modifies glibc.new/sysdeps/arm/bits/endian.h
30 (currently at version 1.4) is only needed for proper VFP operation.
31 But the hunk which modifies sysdeps/arm/gmp-mparam.h, and the hunk
32 that deletes sysdeps/arm/ieee754.h (yes, this IS correct), are needed
33 for proper operation of *any* FP model on big endian ARM.
35 See also discussion in followups to
36 http://sources.redhat.com/ml/crossgcc/2004-05/msg00245.html)
38 Message-ID: <276985760.37584@eyou.com>
39 Received: from unknown (HELO eyou.com) (172.16.2.2)
40 by 0.0.0.0 with SMTP; Tue, 17 Feb 2004 10:42:40 +0800
41 Received: (qmail 8238 invoked by uid 65534); 17 Feb 2004 10:42:38 +0800
42 Date: 17 Feb 2004 10:42:38 +0800
43 Message-ID: <20040217104238.8237.qmail@eyou.com>
44 From: "add" <addsub@eyou.com>
46 Reply-To: "add" <addsub@eyou.com>
47 Subject: Re: problem while building arm vfp softfloat gcc `
49 Hi, Dan, This is a patch I applied to my glibc-2.3.2, then my softfloat
50 toolchain can printf("%f\n",1.0). So you may have a try of this
53 diff -uNrp glibc.old/sysdeps/arm/bits/endian.h glibc.new/sysdeps/arm/bits/endian.h
54 --- glibc.old/sysdeps/arm/bits/endian.h 1999-04-12 11:59:13.000000000 -0400
55 +++ glibc.new/sysdeps/arm/bits/endian.h 2004-02-12 09:15:13.000000000 -0500
58 #define __BYTE_ORDER __LITTLE_ENDIAN
62 +#define __FLOAT_WORD_ORDER __BYTE_ORDER
64 #define __FLOAT_WORD_ORDER __BIG_ENDIAN
66 diff -uNrp glibc.old/sysdeps/arm/gmp-mparam.h glibc.new/sysdeps/arm/gmp-mparam.h
67 --- glibc.old/sysdeps/arm/gmp-mparam.h 2001-07-07 15:21:19.000000000 -0400
68 +++ glibc.new/sysdeps/arm/gmp-mparam.h 2004-02-12 09:15:13.000000000 -0500
69 @@ -26,5 +26,13 @@ MA 02111-1307, USA. */
70 #define BITS_PER_SHORTINT 16
71 #define BITS_PER_CHAR 8
73 -#define IEEE_DOUBLE_BIG_ENDIAN 0
74 -#define IEEE_DOUBLE_MIXED_ENDIAN 1
75 +#if defined(__ARMEB__)
76 +# define IEEE_DOUBLE_MIXED_ENDIAN 0
77 +# define IEEE_DOUBLE_BIG_ENDIAN 1
78 +#elif defined(__VFP_FP__)
79 +# define IEEE_DOUBLE_MIXED_ENDIAN 0
80 +# define IEEE_DOUBLE_BIG_ENDIAN 0
82 +# define IEEE_DOUBLE_BIG_ENDIAN 0
83 +# define IEEE_DOUBLE_MIXED_ENDIAN 1
85 diff -uNrp glibc.old/sysdeps/arm/ieee754.h glibc.new/sysdeps/arm/ieee754.h
86 --- glibc.old/sysdeps/arm/ieee754.h 2001-07-07 15:21:19.000000000 -0400
87 +++ glibc.new/sysdeps/arm/ieee754.h 1969-12-31 19:00:00.000000000 -0500
89 -/* Copyright (C) 1992, 1995, 1996, 1998 Free Software Foundation, Inc.
90 - This file is part of the GNU C Library.
92 - The GNU C Library is free software; you can redistribute it and/or
93 - modify it under the terms of the GNU Lesser General Public
94 - License as published by the Free Software Foundation; either
95 - version 2.1 of the License, or (at your option) any later version.
97 - The GNU C Library is distributed in the hope that it will be useful,
98 - but WITHOUT ANY WARRANTY; without even the implied warranty of
99 - MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
100 - Lesser General Public License for more details.
102 - You should have received a copy of the GNU Lesser General Public
103 - License along with the GNU C Library; if not, write to the Free
104 - Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
109 -#define _IEEE754_H 1
110 -#include <features.h>
120 - /* This is the IEEE 754 single-precision format. */
123 - unsigned int mantissa:23;
124 - unsigned int exponent:8;
125 - unsigned int negative:1;
128 - /* This format makes it easier to see if a NaN is a signalling NaN. */
131 - unsigned int mantissa:22;
132 - unsigned int quiet_nan:1;
133 - unsigned int exponent:8;
134 - unsigned int negative:1;
138 -#define IEEE754_FLOAT_BIAS 0x7f /* Added to exponent. */
141 -union ieee754_double
145 - /* This is the IEEE 754 double-precision format. */
148 - unsigned int mantissa0:20;
149 - unsigned int exponent:11;
150 - unsigned int negative:1;
151 - unsigned int mantissa1:32;
154 - /* This format makes it easier to see if a NaN is a signalling NaN. */
157 - unsigned int mantissa0:19;
158 - unsigned int quiet_nan:1;
159 - unsigned int exponent:11;
160 - unsigned int negative:1;
161 - unsigned int mantissa1:32;
165 -#define IEEE754_DOUBLE_BIAS 0x3ff /* Added to exponent. */
168 -/* The following two structures are correct for `new' floating point systems but
169 - wrong for the old FPPC. The only solution seems to be to avoid their use on
172 -union ieee854_long_double
176 - /* This is the IEEE 854 double-extended-precision format. */
179 - unsigned int exponent:15;
180 - unsigned int empty:16;
181 - unsigned int negative:1;
182 - unsigned int mantissa1:32;
183 - unsigned int mantissa0:32;
186 - /* This is for NaNs in the IEEE 854 double-extended-precision format. */
189 - unsigned int exponent:15;
190 - unsigned int empty:16;
191 - unsigned int negative:1;
192 - unsigned int mantissa1:32;
193 - unsigned int mantissa0:30;
194 - unsigned int quiet_nan:1;
195 - unsigned int one:1;
199 -#define IEEE854_LONG_DOUBLE_BIAS 0x3fff
203 -#endif /* ieee754.h */
205 Signed-off-by: Robert P. J. Day <rpjday@mindspring.com>