[Valgrind-developers] valgrind: r13346: Cope with old linkers which don't support -Ttext-segment but which do

[<sv...@va...>]

tom	2013-03-28 15:53:21 +0000 (Thu, 28 Mar 2013)

  New Revision: 13346

  Log:
    Cope with old linkers which don't support -Ttext-segment but which do
    generate build-id sections. #317091.  (Mark Wielaard, mj...@re...)

  Modified files:
    trunk/Makefile.tool.am
    trunk/configure.in
    trunk/coregrind/link_tool_exe_linux.in


  Modified: trunk/Makefile.tool.am (+11 -3)
===================================================================
--- trunk/Makefile.tool.am	2013-03-28 10:40:53 +00:00 (rev 13345)
+++ trunk/Makefile.tool.am	2013-03-28 15:53:21 +00:00 (rev 13346)
@@ -26,8 +26,14 @@
 endif
 
 
+# -Wl,--build-id=none is needed when linking tools with a linker that only
+# knows -Ttext and not -Ttext-segment. Without this flag newer ld versions
+# (2.20 and later) create a .note.gnu.build-id at the default text segment
+# address, which of course means the resulting executable
+# is unusable. So we have to tell ld not to generate that, with
+# --build-id=none unless the linker supports -Ttext-segment.
 TOOL_LDFLAGS_COMMON_LINUX = \
-	-static -nodefaultlibs -nostartfiles -u _start
+	-static -nodefaultlibs -nostartfiles -u _start @FLAG_NO_BUILD_ID@
 TOOL_LDFLAGS_COMMON_DARWIN = \
 	-nodefaultlibs -nostartfiles -Wl,-u,__start -Wl,-e,__start
 
@@ -57,10 +63,12 @@
 
 # MIPS Linux default start symbol is __start, not _start like on x86 or amd
 TOOL_LDFLAGS_MIPS32_LINUX = \
-	-static -nodefaultlibs -nostartfiles -u __start @FLAG_M32@
+	-static -nodefaultlibs -nostartfiles -u __start @FLAG_NO_BUILD_ID@ \
+	@FLAG_M32@
 
 TOOL_LDFLAGS_MIPS64_LINUX = \
-	-static -nodefaultlibs -nostartfiles -u __start @FLAG_M64@
+	-static -nodefaultlibs -nostartfiles -u __start @FLAG_NO_BUILD_ID@ \
+	@FLAG_M64@
 
 # On Android we must ask for non-executable stack, not sure why.
 if VGCONF_PLATFORMS_INCLUDE_ARM_LINUX

  Modified: trunk/coregrind/link_tool_exe_linux.in (+12 -4)
===================================================================
--- trunk/coregrind/link_tool_exe_linux.in	2013-03-28 10:40:53 +00:00 (rev 13345)
+++ trunk/coregrind/link_tool_exe_linux.in	2013-03-28 15:53:21 +00:00 (rev 13346)
@@ -27,15 +27,23 @@
 # directly.  It is only run as part of the build process, with 
 # carefully constrained inputs.
 #
-# Linux specific complication:
+# Linux specific complications:
 #
 # - need to support both old GNU ld and gold: use -Ttext= to
 #   set the text segment address if that is all we have. We really
 #   need -Ttext-segment. Otherwise with GNU ld sections or notes
-#   (like the build-id) don't get at the desired address. Luckily
-#   -Ttext does the right thing for gold. So configure checks for
-#   us and sets FLAG_T_TEXT.
+#   (like the build-id) don't get at the desired address. But older
+#   linkers only know about -Ttext, not -Ttext-segment. So configure
+#   checks for us and sets FLAG_T_TEXT.
 #
+# - If all we have is -Ttext, then we need to pass --build-id=none
+#   (that is, -Wl,--build-id=none to gcc) if it accepts it, to ensure
+#   the linker doesn't add a notes section which ends up at the default
+#   load address and so defeats our attempts to keep that address clear
+#   for the client.  However, older linkers don't support this flag,
+#   so it is tested for by configure.in and is shipped to us as part of
+#   argv[2 ..].
+#
 # So: what we actually do:
 # 
 #   pass the specified command to the linker as-is, except, add

  Modified: trunk/configure.in (+32 -0)
===================================================================
--- trunk/configure.in	2013-03-28 10:40:53 +00:00 (rev 13345)
+++ trunk/configure.in	2013-03-28 15:53:21 +00:00 (rev 13346)
@@ -1723,15 +1723,47 @@
 AC_LINK_IFELSE(
 [AC_LANG_PROGRAM([ ], [return 0;])],
 [
+  linker_using_t_text="no"
   AC_SUBST([FLAG_T_TEXT], ["-Ttext-segment"])
   AC_MSG_RESULT([yes])
 ], [
+  linker_using_t_text="yes"
   AC_SUBST([FLAG_T_TEXT], ["-Ttext"])
   AC_MSG_RESULT([no])
 ])
 CFLAGS=$safe_CFLAGS
 
+# If the linker only supports -Ttext (not -Ttext-segment) then we will
+# have to strip any build-id ELF NOTEs from the staticly linked tools.
+# Otherwise the build-id NOTE might end up at the default load address.
+# (Pedantically if the linker is gold then -Ttext is fine, but newer
+# gold versions also support -Ttext-segment. So just assume that unless
+# we can use -Ttext-segment we need to strip the build-id NOTEs.
+if test "x${linker_using_t_text}" == "xyes"; then
+AC_MSG_NOTICE([ld -Ttext used, need to strip build-id NOTEs.])
+# does the linker support -Wl,--build-id=none ?  Note, it's
+# important that we test indirectly via whichever C compiler
+# is selected, rather than testing /usr/bin/ld or whatever
+# directly.
+AC_MSG_CHECKING([if the linker accepts -Wl,--build-id=none])
+safe_CFLAGS=$CFLAGS
+CFLAGS="-Wl,--build-id=none"
 
+AC_LINK_IFELSE(
+[AC_LANG_PROGRAM([ ], [return 0;])],
+[
+  AC_SUBST([FLAG_NO_BUILD_ID], ["-Wl,--build-id=none"])
+  AC_MSG_RESULT([yes])
+], [
+  AC_SUBST([FLAG_NO_BUILD_ID], [""])
+  AC_MSG_RESULT([no])
+])
+else
+AC_MSG_NOTICE([ld -Ttext-segment used, no need to strip build-id NOTEs.])
+AC_SUBST([FLAG_NO_BUILD_ID], [""])
+fi
+CFLAGS=$safe_CFLAGS
+
 # does the ppc assembler support "mtocrf" et al?
 AC_MSG_CHECKING([if ppc32/64 as supports mtocrf/mfocrf])