summary refs log tree commit diff
path: root/bits
diff options
context:
space:
mode:
authorJoseph Myers <joseph@codesourcery.com>2016-11-23 00:28:30 +0000
committerJoseph Myers <joseph@codesourcery.com>2016-11-23 00:28:30 +0000
commitb0216d3e4d98a3528bad428c22ff96fcbcc102a4 (patch)
tree226efd1328173ea09755cc75af10d41ba84909b5 /bits
parent510abe7b945ddab6f4497e7c097cff677286bb4d (diff)
downloadglibc-b0216d3e4d98a3528bad428c22ff96fcbcc102a4.tar.gz
glibc-b0216d3e4d98a3528bad428c22ff96fcbcc102a4.tar.xz
glibc-b0216d3e4d98a3528bad428c22ff96fcbcc102a4.zip
Fix default float_t definition (bug 20855).
The default (top-level) version of bits/mathdef.h defines float_t to
double.  It is used on ColdFire, MicroBlaze, Nios II and SH3, all of
which define FLT_EVAL_METHOD to 0, so float_t should be float (and C11
requires a certain correspondence between these typedefs and
FLT_EVAL_METHOD values).

I proposed fixing this default in
<https://sourceware.org/ml/libc-alpha/2015-01/msg00499.html>, with no
objections from architecture maintainers, and this patch makes that
fix.  As noted in the NEWS entry added, this might affect the ABIs of
non-glibc libraries (ImageMagick has been mentioned in gcc-patches
discussion of the S/390 case - which is unaffected by this patch), but
as noted in my previous message, affected libraries would have
problems with -mfpmath=sse anyway on 32-bit x86.

A (compilation) testcase is added to verify the required
correspondence of typedefs to FLT_EVAL_METHOD values.  This test is
built with -fexcess-precision=standard to avoid any issues with GCC 7
on S/390 providing a more accurate FLT_EVAL_METHOD definition in the
default (no excess precision) mode.  (This will also be usable to test
a fix for the recently reported bug about these typedefs on x86_64
-mfpmath=387, as architecture-specific tests can be added that

It is entirely possible that the fixed default makes some
architecture-specific versions of bits/mathdef.h semantically
equivalent to the default version and so no longer required.  I don't
intend to investigate that separately from the refactoring I proposed
in <https://sourceware.org/ml/libc-alpha/2016-11/msg00745.html>, which
will create as few header variants as possible for each group of
definitions.

Tested (compilation only) with build-many-glibcs.py.

	[BZ #20855]
	* bits/mathdef.h (float_t): Define to float.
	* math/test-flt-eval-method.c: New file.
	* math/Makefile (tests): Add test-flt-eval-method.
	(CFLAGS-test-flt-eval-method.c): New variable.
Diffstat (limited to 'bits')
-rw-r--r--bits/mathdef.h8
1 files changed, 4 insertions, 4 deletions
diff --git a/bits/mathdef.h b/bits/mathdef.h
index c5d5f8c0a8..d557d08582 100644
--- a/bits/mathdef.h
+++ b/bits/mathdef.h
@@ -22,10 +22,10 @@
 #if defined  __USE_ISOC99 && defined _MATH_H && !defined _MATH_H_MATHDEF
 # define _MATH_H_MATHDEF	1
 
-/* Normally, there is no long double type and the `float' and `double'
-   expressions are evaluated as `double'.  */
-typedef double float_t;		/* `float' expressions are evaluated as
-				   `double'.  */
+/* Expressions are evaluated with the range and precision of their
+   type.  */
+typedef float float_t;		/* `float' expressions are evaluated as
+				   `float'.  */
 typedef double double_t;	/* `double' expressions are evaluated as
 				   `double'.  */