[1/2] Fix behaviour of 'is_binary128' in __printf_fp

Message ID 20180629041819.19721-2-gabriel@inconstante.eti.br
State New
Headers show
Series
  • Fix for printf_size and IEEE binary128 implementation
Related show

Commit Message

Gabriel F. T. Gomes June 29, 2018, 4:18 a.m.
Should this have an associated bugzilla entry?

-- 8< --
Both 'is_long_double' and 'is_binary128' (members of the 'printf_info'
structure) are used in __printf_fp to select the type of the
floating-point argument it receives in its 'args' parameter.  When the
argument is of double type, 'is_long_double' should be set to zero.
Similarly, when the argument is of long double type, 'is_long_double'
should be set to one.  More recently (since the addition of the
_Float128 API), the floating-point argument can also be of _Float128
type, then 'is_binary128' should be set to one.

In current __printf_fp, when 'is_binary128' is one, the floating-point
argument is treated as if it was of _Float128 type, regardless of the
value of 'is_long_double'.  This is not a problem for strfromd, because
strfromd always sets 'is_binary128' to zero.  However, it could be a
problem for users of printf_size, who could be expecting that
'is_long_double == 0' meant that the 'args' parameter was of double
type, regardless of the value of 'is_binary128'.

This patch modifies this behaviour in __printf_fp and adjusts
strfromf128 so that it always sets 'is_long_double' to one.

Tested for powerpc64le and x86_64.

	* stdio-common/printf_fp.c (__printf_fp_l): Also check the value
	of 'is_long_double' before reading the argument into a variable
	of _Float128 type.
	* stdio-common/printf_size.c (__printf_size): Likewise.
	* stdlib/strfrom-skeleton.c [__HAVE_DISTINCT_FLOAT128] (STRFROM):
	If 'is_binary128' is one, also set 'is_long_double' to one.

	* stdio-common/Makefile (test-srcs): Add tst-isbinary128
	(tests-special) Add $(objpfx)tst-isbinary128.out.
	($(objpfx)tst-isbinary128.out): New build and run rule.
	* stdio-common/tst-isbinary128.c: New file.
	* stdio-common/tst-isbinary128.sh: Likewise.
---
 stdio-common/Makefile           |  9 +++++--
 stdio-common/printf_fp.c        |  2 +-
 stdio-common/printf_size.c      |  2 +-
 stdio-common/tst-isbinary128.c  | 59 +++++++++++++++++++++++++++++++++++++++++
 stdio-common/tst-isbinary128.sh | 38 ++++++++++++++++++++++++++
 stdlib/strfrom-skeleton.c       |  6 ++++-
 6 files changed, 111 insertions(+), 5 deletions(-)
 create mode 100644 stdio-common/tst-isbinary128.c
 create mode 100644 stdio-common/tst-isbinary128.sh

Comments

Joseph Myers June 29, 2018, 4:24 p.m. | #1
On Fri, 29 Jun 2018, Gabriel F. T. Gomes wrote:

> Should this have an associated bugzilla entry?
> 
> -- 8< --
> Both 'is_long_double' and 'is_binary128' (members of the 'printf_info'
> structure) are used in __printf_fp to select the type of the
> floating-point argument it receives in its 'args' parameter.  When the
> argument is of double type, 'is_long_double' should be set to zero.
> Similarly, when the argument is of long double type, 'is_long_double'
> should be set to one.  More recently (since the addition of the
> _Float128 API), the floating-point argument can also be of _Float128
> type, then 'is_binary128' should be set to one.
> 
> In current __printf_fp, when 'is_binary128' is one, the floating-point
> argument is treated as if it was of _Float128 type, regardless of the
> value of 'is_long_double'.  This is not a problem for strfromd, because
> strfromd always sets 'is_binary128' to zero.  However, it could be a
> problem for users of printf_size, who could be expecting that
> 'is_long_double == 0' meant that the 'args' parameter was of double
> type, regardless of the value of 'is_binary128'.

Well, the main use case of printf_size is being registered as a printf 
handler, not users creating their own printf_info structure.  I'm not 
convinced that calling printf_size with any form of printf_info structure 
that cannot be created by printf functions is valid.

It's certainly true that adding support for float128 versions of printf 
means a documentation patch is required to document is_binary128 in 
stdio.texi.  But why should the semantics be that is_long_double == 0 
overrides is_binary128 rather than is_binary128 being independent of 
is_long_double?  I'd have expected the documentation to say that 
is_binary128 is set alongside is_long_double in the case where printf has 
the 'L' modifier and is called from code with binary128 long double, on a 
platform where that's the third long double format supported by glibc, 
i.e. powerpc64le, but that printf will never set is_binary128 without 
is_long_double (and thus what printf_size does with that combination is 
basically irrelevant, but I'd think it natural for it to make is_binary128 
== 1 take precedence over is_long_double).

(So handlers could potentially, on powerpc64le, themselves deal with both 
ibm128 and binary128 long double, but if not trying to do so - if the 
handler is always built with the same long double format as the printf 
caller - they could just ignore is_binary128.)

Patch

diff --git a/stdio-common/Makefile b/stdio-common/Makefile
index 738a3cead0..94b8b10357 100644
--- a/stdio-common/Makefile
+++ b/stdio-common/Makefile
@@ -63,13 +63,14 @@  tests := tstscanf test_rdwr test-popen tstgetln test-fseek \
 	 tst-vfprintf-mbs-prec \
 	 tst-scanf-round \
 
-test-srcs = tst-unbputc tst-printf
+test-srcs = tst-unbputc tst-printf tst-isbinary128
 
 ifeq ($(run-built-tests),yes)
 tests-special += $(objpfx)tst-unbputc.out $(objpfx)tst-printf.out \
 		 $(objpfx)tst-printf-bz18872-mem.out \
 		 $(objpfx)tst-setvbuf1-cmp.out \
-		 $(objpfx)tst-vfprintf-width-prec-mem.out
+		 $(objpfx)tst-vfprintf-width-prec-mem.out \
+		 $(objpfx)tst-isbinary128.out
 generated += tst-printf-bz18872.c tst-printf-bz18872.mtrace \
 	     tst-printf-bz18872-mem.out \
 	     tst-vfprintf-width-prec.mtrace tst-vfprintf-width-prec-mem.out
@@ -103,6 +104,10 @@  $(objpfx)tst-printf.out: tst-printf.sh $(objpfx)tst-printf
 	$(SHELL) $< $(common-objpfx) '$(test-program-prefix)' > $@; \
 	$(evaluate-test)
 
+$(objpfx)tst-isbinary128.out: tst-isbinary128.sh $(objpfx)tst-isbinary128
+	$(SHELL) $^ '$(test-program-prefix)' $@; \
+	$(evaluate-test)
+
 # We generate this source because it requires a printf invocation with
 # 10K arguments.
 $(objpfx)tst-printf-bz18872.c: tst-printf-bz18872.sh
diff --git a/stdio-common/printf_fp.c b/stdio-common/printf_fp.c
index 66ab59bad2..3f09581810 100644
--- a/stdio-common/printf_fp.c
+++ b/stdio-common/printf_fp.c
@@ -375,7 +375,7 @@  __printf_fp_l (FILE *fp, locale_t loc,
 
   /* Fetch the argument value.	*/
 #if __HAVE_DISTINCT_FLOAT128
-  if (info->is_binary128)
+  if (info->is_binary128 && info->is_long_double)
     PRINTF_FP_FETCH (_Float128, fpnum.f128, float128, FLT128_MANT_DIG)
   else
 #endif
diff --git a/stdio-common/printf_size.c b/stdio-common/printf_size.c
index 7e073c50d4..4655276f2b 100644
--- a/stdio-common/printf_size.c
+++ b/stdio-common/printf_size.c
@@ -142,7 +142,7 @@  __printf_size (FILE *fp, const struct printf_info *info,
 
   /* Fetch the argument value.	*/
 #if __HAVE_DISTINCT_FLOAT128
-  if (info->is_binary128)
+  if (info->is_binary128 && info->is_long_double)
     PRINTF_SIZE_FETCH (_Float128, fpnum.f128)
   else
 #endif
diff --git a/stdio-common/tst-isbinary128.c b/stdio-common/tst-isbinary128.c
new file mode 100644
index 0000000000..a7ee358d5c
--- /dev/null
+++ b/stdio-common/tst-isbinary128.c
@@ -0,0 +1,59 @@ 
+/* Test for the behaviour of 'is_binary128' in printf_size.
+   Copyright (C) 2018 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library 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
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <printf.h>
+#include <stdio.h>
+#include <string.h>
+
+#include <support/check.h>
+
+static int
+do_test (void)
+{
+  double d = 2000;
+  double *dptr = &d;
+  long double ld = 4000;
+  long double *ldptr = & ld;
+  struct printf_info info;
+
+  memset (&info, 0, sizeof (info));
+  info.spec = L'f';
+
+  /* First, call printf_size with 'is_long_double' and 'is_binary128'
+     both zeroed.  */
+  printf_size (stdout, &info, (void *) &dptr);
+
+  /* Then test that setting 'is_binary128' to one still prints double
+     correctly.  */
+  info.is_binary128 = 1;
+  printf_size (stdout, &info, (void *) &dptr);
+
+  /* Finally, check that long double values are also printed correctly,
+     provided that 'is_binary128' is zeroed.  */
+  info.is_long_double = 1;
+  info.is_binary128 = 0;
+  printf_size (stdout, &info, (void *) &ldptr);
+
+  /* Setting both 'is_long_double' and 'is_binary128' to one is out of
+     the scope of this test, because such configuration is only valid
+     when _Float128 and long double are ABI-distinct (which is not
+     always true and this is an arch-independent test).  */
+  return 0;
+}
+
+#include <support/test-driver.c>
diff --git a/stdio-common/tst-isbinary128.sh b/stdio-common/tst-isbinary128.sh
new file mode 100644
index 0000000000..f729e55ff7
--- /dev/null
+++ b/stdio-common/tst-isbinary128.sh
@@ -0,0 +1,38 @@ 
+#!/bin/sh
+# Test for the behaviour of 'is_binary128' in printf_size.
+# Copyright (C) 2018 Free Software Foundation, Inc.
+# This file is part of the GNU C Library.
+
+# The GNU C Library is free software; you can redistribute it and/or
+# modify it under the terms of the GNU Lesser General Public
+# License as published by the Free Software Foundation; either
+# version 2.1 of the License, or (at your option) any later version.
+
+# The GNU C Library 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
+# Lesser General Public License for more details.
+
+# You should have received a copy of the GNU Lesser General Public
+# License along with the GNU C Library; if not, see
+# <http://www.gnu.org/licenses/>.
+
+set -e
+
+test_program=$1; shift
+test_program_prefix=$1; shift
+test_program_output=$1; shift
+
+status=0
+
+${test_program_prefix} \
+  ${test_program} \
+  > ${test_program_output} || status=1
+
+echo -n "2k2k4k" | cmp - ${test_program_output} > /dev/null 2>&1 ||
+{
+  status=1
+  echo "*** output comparison failed"
+}
+
+exit $status
diff --git a/stdlib/strfrom-skeleton.c b/stdlib/strfrom-skeleton.c
index 2840512cae..5d6e44d997 100644
--- a/stdlib/strfrom-skeleton.c
+++ b/stdlib/strfrom-skeleton.c
@@ -133,9 +133,13 @@  STRFROM (char *dest, size_t size, const char *format, FLOAT f)
   info.is_long_double = __builtin_types_compatible_p (FLOAT, long double);
 
   /* Similarly, the function strfromf128 passes a floating-point number in
-     _Float128 format to printf_fp.  */
+     _Float128 format to __printf_fp.  Setting is_binary128 alone is not
+     enough, because when is_long_double is zero, __printf_fp always treats
+     the floating-point number as double (regardless of is_binary128).  */
 #if __HAVE_DISTINCT_FLOAT128
   info.is_binary128 = __builtin_types_compatible_p (FLOAT, _Float128);
+  if (info.is_binary128)
+    info.is_long_double = 1;
 #endif
 
   /* Set info according to the format string.  */