[PATCH] Kill __errno_location PLT slot in --with-tls --with-__thread build

[PATCH] Kill __errno_location PLT slot in --with-tls --with-__thread build

[Jakub Jelinek]

Hi!

There is an unnecessary __errno_location .plt slot because s_ldexp*.os
is used by both libc.so and libm.so, thus compiled
with -DNOT_IN_libc -DIS_IN_libm (which means it uses __errno_location).
Other fix could be to do something like what is already used for
rtld-%.os for math/Makefile(calls) routines, but I'm afraid it could slow
down the build by adding lots of new rules to sysd-rules.

2002-11-15  Jakub Jelinek  <jakub@redhat.com>

	* math/Makefile (libm-calls): Change s_ldexp to m_ldexp.
	* sysdeps/generic/m_ldexp.c: New file.
	* sysdeps/generic/m_ldexpf.c: New file.
	* sysdeps/generic/m_ldexpl.c: New file.
	* sysdeps/ia64/fpu/m_ldexp.c: New file.
	* sysdeps/ia64/fpu/m_ldexpf.c: New file.
	* sysdeps/ia64/fpu/m_ldexpl.c: New file.

--- libc/math/Makefile.jj	2002-09-10 10:48:12.000000000 +0200
+++ libc/math/Makefile	2002-11-15 10:41:35.000000000 +0100
@@ -46,7 +46,7 @@ libm-calls = e_acos e_acosh e_asin e_ata
 	     e_rem_pio2 e_remainder e_scalb e_sinh e_sqrt e_gamma_r	\
 	     k_cos k_rem_pio2 k_sin k_tan s_asinh s_atan s_cbrt		\
 	     s_ceil s_copysign s_cos s_erf s_expm1 s_fabs s_finite	\
-	     s_floor s_frexp s_ilogb s_ldexp s_log1p s_logb		\
+	     s_floor s_frexp s_ilogb m_ldexp s_log1p s_logb		\
 	     s_modf s_nextafter s_nexttoward s_rint s_scalbn s_scalbln	\
 	     s_significand s_sin s_tan s_tanh w_acos w_acosh w_asin	\
 	     w_atan2 w_atanh w_cosh w_drem w_exp w_exp2 w_exp10 w_fmod	\
--- libc/sysdeps/generic/m_ldexp.c.jj	2002-11-15 10:42:35.000000000 +0100
+++ libc/sysdeps/generic/m_ldexp.c	2002-11-15 10:42:35.000000000 +0100
@@ -0,0 +1 @@
+#include <s_ldexp.c>
--- libc/sysdeps/generic/m_ldexpf.c.jj	2002-11-15 10:42:35.000000000 +0100
+++ libc/sysdeps/generic/m_ldexpf.c	2002-11-15 10:42:35.000000000 +0100
@@ -0,0 +1 @@
+#include <s_ldexpf.c>
--- libc/sysdeps/generic/m_ldexpl.c.jj	2002-11-15 10:42:35.000000000 +0100
+++ libc/sysdeps/generic/m_ldexpl.c	2002-11-15 10:42:35.000000000 +0100
@@ -0,0 +1 @@
+#include <s_ldexpl.c>
--- libc/sysdeps/ia64/fpu/m_ldexp.c.jj	2002-11-15 10:44:41.000000000 +0100
+++ libc/sysdeps/ia64/fpu/m_ldexp.c	2002-11-15 10:44:41.000000000 +0100
@@ -0,0 +1 @@
+/* m_ldexp.c not needed */
--- libc/sysdeps/ia64/fpu/m_ldexpf.c.jj	2002-11-15 10:44:41.000000000 +0100
+++ libc/sysdeps/ia64/fpu/m_ldexpf.c	2002-11-15 10:44:41.000000000 +0100
@@ -0,0 +1 @@
+/* m_ldexpf.c not needed */
--- libc/sysdeps/ia64/fpu/m_ldexpl.c.jj	2002-11-15 10:44:41.000000000 +0100
+++ libc/sysdeps/ia64/fpu/m_ldexpl.c	2002-11-15 10:44:41.000000000 +0100
@@ -0,0 +1 @@
+/* m_ldexpl.c not needed */

	Jakub

Re: [PATCH] Kill __errno_location PLT slot in --with-tls --with-__thread build

[Roland McGrath]

There is only this tiny set of duplicated routines and should never be any
more, unlike the rtld case.  So I think this kludge is fine enough.

Re: [PATCH] Kill __errno_location PLT slot in --with-tls --with-__thread build

[Roland McGrath]

The deal with ia64 is just that its implementations don't use __errno_location
directly and so don't have the issue, right?  There is no harm in compiling
those files again for libm, so unless I am missing something I will omit
the ia64-specific bits to minimize the maintenance hair.

Re: [PATCH] Kill __errno_location PLT slot in --with-tls --with-__thread build

[Roland McGrath]

In your changes, on ia64 now there will be no ldexp{,f,l} in libm at all.
Was that the plan?  I think that will break versioned dependencies.

I did some pure makefile hackery to generate the m_foo.? source files (two
more rules per dir in sysd-rules) that dtrt for .c vs .S generically.  It
needs nothing in sysdeps, and won't require any more files anywhere if we
need to do the same for another of those modules.

Re: [PATCH] Kill __errno_location PLT slot in --with-tls --with-__thread build

[Roland McGrath]

Duh, good point.