about summary refs log tree commit diff
path: root/sysdeps/powerpc
diff options
context:
space:
mode:
authorJoseph Myers <joseph@codesourcery.com>2018-01-09 22:34:35 +0000
committerJoseph Myers <joseph@codesourcery.com>2018-01-09 22:34:35 +0000
commite134ca175318e0f1c04b4e857f38efae091c35ce (patch)
tree5d4d40fd1f09333f79901108038d670fdcf5ee32 /sysdeps/powerpc
parent311ba8dc4416467947eff2ab327854f124226309 (diff)
downloadglibc-e134ca175318e0f1c04b4e857f38efae091c35ce.tar.gz
glibc-e134ca175318e0f1c04b4e857f38efae091c35ce.tar.xz
glibc-e134ca175318e0f1c04b4e857f38efae091c35ce.zip
Fix powerpc-nofpu complex long double functions spurious "invalid" exception (bug 22687).
For soft-float powerpc, various _Complex long double functions
generate spurious "invalid" exceptions, even with a compiler with GCC
bug 64811 fixed.

The problem is GCC's built-in fabsl expansion.  Various files are
already built with -fno-builtin-fabsl because in this case (IBM long
double, for soft-float or e500v1) a fallback fabsl expansion based on
comparisons is used, which can produce the wrong sign of a zero
result.  Those comparisons can also produce spurious exceptions for
NaN arguments.  Furthermore, __builtin_fpclassify implemently uses
__builtin_fabsl, and is unaffected by -fno-builtin-fabsl, and the
fpclassify macro uses __builtin_fpclassify in the absence of
-fsignaling-nans.  Thus, this patch arranges for the problem files
using fpclassify to be built with -fsignaling-nans in this case, to
avoid spurious exceptions from fpclassify.

Tested for powerpc (soft-float).

	[BZ #22687]
	* sysdeps/powerpc/nofpu/Makefile (CFLAGS-s_cacosl.c): New
	variable.
	(CFLAGS-s_cacoshl.c): Likewise.
	(CFLAGS-s_casinhl.c): Likewise.
	(CFLAGS-s_catanl.c): Likewise.
	(CFLAGS-s_catanhl.c): Likewise.
	(CFLAGS-s_cexpl.c): Likewise.
	(CFLAGS-s_ccoshl.c): Add -fsignaling-nans.
	(CFLAGS-s_csinhl.c): Likewise.
	(CFLAGS-s_clogl.c): Likewise.
	(CFLAGS-s_clog10l.c): Likewise.
	(CFLAGS-s_csinl.c): Likewise.
	(CFLAGS-s_csqrtl.c): Likewise.
Diffstat (limited to 'sysdeps/powerpc')
-rw-r--r--sysdeps/powerpc/nofpu/Makefile26
1 files changed, 20 insertions, 6 deletions
diff --git a/sysdeps/powerpc/nofpu/Makefile b/sysdeps/powerpc/nofpu/Makefile
index ed163219be..09ecdf1dcc 100644
--- a/sysdeps/powerpc/nofpu/Makefile
+++ b/sysdeps/powerpc/nofpu/Makefile
@@ -15,15 +15,29 @@ CPPFLAGS += -I../soft-fp/
 # gcc-4.1.1 and may be too late for gcc-4.2.  So we need these flags
 # until the fix in a gcc release and glibc drops support for earlier
 # versions of gcc.
+#
+# This bug can cause wrong signs of zero results from fabsl, and
+# spurious "invalid" exceptions from comparisons with NaN in the
+# default code sequence generated.  The fpclassify expansion used by
+# __builtin_fpclassify, which is used in the absence of
+# -fsignaling-nans, uses __builtin_fabsl, and is not affected by
+# -fno-builtin-fabsl, so requiring -fsignaling-nans for some affected
+# source files.
 CFLAGS-e_atan2l.c += -fno-builtin-fabsl
 CFLAGS-e_hypotl.c += -fno-builtin-fabsl
 CFLAGS-e_powl.c += -fno-builtin-fabsl
-CFLAGS-s_ccoshl.c += -fno-builtin-fabsl
-CFLAGS-s_csinhl.c += -fno-builtin-fabsl
-CFLAGS-s_clogl.c += -fno-builtin-fabsl
-CFLAGS-s_clog10l.c += -fno-builtin-fabsl
-CFLAGS-s_csinl.c += -fno-builtin-fabsl
-CFLAGS-s_csqrtl.c += -fno-builtin-fabsl
+CFLAGS-s_cacosl.c += -fsignaling-nans
+CFLAGS-s_cacoshl.c += -fsignaling-nans
+CFLAGS-s_casinhl.c += -fsignaling-nans
+CFLAGS-s_catanl.c += -fsignaling-nans
+CFLAGS-s_catanhl.c += -fsignaling-nans
+CFLAGS-s_ccoshl.c += -fno-builtin-fabsl -fsignaling-nans
+CFLAGS-s_cexpl.c += -fsignaling-nans
+CFLAGS-s_csinhl.c += -fno-builtin-fabsl -fsignaling-nans
+CFLAGS-s_clogl.c += -fno-builtin-fabsl -fsignaling-nans
+CFLAGS-s_clog10l.c += -fno-builtin-fabsl -fsignaling-nans
+CFLAGS-s_csinl.c += -fno-builtin-fabsl -fsignaling-nans
+CFLAGS-s_csqrtl.c += -fno-builtin-fabsl -fsignaling-nans
 CFLAGS-w_acosl_compat.c += -fno-builtin-fabsl
 CFLAGS-w_asinl_compat.c += -fno-builtin-fabsl
 CFLAGS-w_atanhl_compat.c += -fno-builtin-fabsl