Patchwork [v3,build] Clear hardware capabilities on libstdc++.so with Sun as

login
register
mail settings
Submitter Rainer Orth
Date Sept. 12, 2012, 9:26 a.m.
Message ID <yddehm7twv9.fsf@manam.CeBiTec.Uni-Bielefeld.DE>
Download mbox | patch
Permalink /patch/183304/
State New
Headers show

Comments

Rainer Orth - Sept. 12, 2012, 9:26 a.m.
Since the use of rdrand was introduced in src/c++11/random.cc, all
execution tests involving libstdc++.so.6 fail on Solaris 10 and 11/x86
with a sufficiently recent native assembler that supports rdrand: either
Solaris 10/x86 patch 119961-11 or Solaris 11.1 builds (haven't checked
which one).  The problem is that as tags src/c++11/random.o as needing
RDRAND support:

% file random.o
random.o:       ELF 32-bit LSB relocatable 80386 Version 1 [CMOV RDRAND]

which is propagated to libstdc++.so.6 and causes the runtime linker to
fail the execution if the hardware doesn't support that hardware
capability, although rdrand is executed only if the cpuid indicates the
support is present.

The usual solution so far has been to clear the hardware capability
using a linker map (as in libitm, cf. libitm/clearcap.map).
Unfortunately, this doesn't work here: as can be seen with elfdump,
RDRAND is set in a second mask (CA_SUNW_HW_2) since all bits in
CA_SUNW_HW_1 are already used:

% elfdump -H random.o

Capabilities Section:  .SUNW_cap

 Object Capabilities:
     index  tag               value
       [0]  CA_SUNW_HW_1     0x20  [ CMOV ]
       [1]  CA_SUNW_HW_2     0x1  [ RDRAND ]

The old (v1) linker map syntax has no support for clearing that
bit/mask, and while the v2 map syntax does, we cannot universally assume
it's present on Solaris 10: while recent linker patches include it,
older ones don't and ld and as can be updated independently.

So I've settled for a different solution instead: Sun assemblers with
hardware capability support also have a -nH switch to suppress their
generation, thus I'm testing for that and use it if possible.

This is exactly what the following patch does.

Bootstrapped without regressions on i386-pc-solaris2.1[01] with affected
versions of Sun as and gas 2.22.  Results with gas are unchanged, with
Sun as all failures are gone.  x86_64-unknown-linux-gnu bootstrap is
running to make sure nothing breaks there.

Ok for mainline if that passes?

	Rainer


2012-09-11  Rainer Orth  <ro@CeBiTec.Uni-Bielefeld.DE>

	* acinclude.m4 (GLIBCXX_CHECK_ASSEMBLER_HWCAP): Define.
	* configure.ac: Call GLIBCXX_CHECK_ASSEMBLER_HWCAP.
	* fragment.am (CONFIG_CXXFLAGS): Add $(HWCAP_FLAGS).
	* configure: Regenerate.
	* Makefile.in: Regenerate.
	* doc/Makefile.in: Regenerate.
	* include/Makefile.in: Regenerate.
	* libsupc++/Makefile.in: Regenerate.
	* po/Makefile.in: Regenerate.
	* python/Makefile.in: Regenerate.
	* src/Makefile.in: Regenerate.
	* src/c++11/Makefile.in: Regenerate.
	* src/c++98/Makefile.in: Regenerate.
	* testsuite/Makefile.in: Regenerate.
Paolo Bonzini - Sept. 12, 2012, 11:27 a.m.
Il 12/09/2012 11:26, Rainer Orth ha scritto:
> Since the use of rdrand was introduced in src/c++11/random.cc, all
> execution tests involving libstdc++.so.6 fail on Solaris 10 and 11/x86
> with a sufficiently recent native assembler that supports rdrand: either
> Solaris 10/x86 patch 119961-11 or Solaris 11.1 builds (haven't checked
> which one).  The problem is that as tags src/c++11/random.o as needing
> RDRAND support:
> 
> % file random.o
> random.o:       ELF 32-bit LSB relocatable 80386 Version 1 [CMOV RDRAND]
> 
> which is propagated to libstdc++.so.6 and causes the runtime linker to
> fail the execution if the hardware doesn't support that hardware
> capability, although rdrand is executed only if the cpuid indicates the
> support is present.
> 
> The usual solution so far has been to clear the hardware capability
> using a linker map (as in libitm, cf. libitm/clearcap.map).
> Unfortunately, this doesn't work here: as can be seen with elfdump,
> RDRAND is set in a second mask (CA_SUNW_HW_2) since all bits in
> CA_SUNW_HW_1 are already used:
> 
> % elfdump -H random.o
> 
> Capabilities Section:  .SUNW_cap
> 
>  Object Capabilities:
>      index  tag               value
>        [0]  CA_SUNW_HW_1     0x20  [ CMOV ]
>        [1]  CA_SUNW_HW_2     0x1  [ RDRAND ]
> 
> The old (v1) linker map syntax has no support for clearing that
> bit/mask, and while the v2 map syntax does, we cannot universally assume
> it's present on Solaris 10: while recent linker patches include it,
> older ones don't and ld and as can be updated independently.
> 
> So I've settled for a different solution instead: Sun assemblers with
> hardware capability support also have a -nH switch to suppress their
> generation, thus I'm testing for that and use it if possible.
> 
> This is exactly what the following patch does.
> 
> Bootstrapped without regressions on i386-pc-solaris2.1[01] with affected
> versions of Sun as and gas 2.22.  Results with gas are unchanged, with
> Sun as all failures are gone.  x86_64-unknown-linux-gnu bootstrap is
> running to make sure nothing breaks there.
> 
> Ok for mainline if that passes?
> 
> 	Rainer
> 
> 
> 2012-09-11  Rainer Orth  <ro@CeBiTec.Uni-Bielefeld.DE>
> 
> 	* acinclude.m4 (GLIBCXX_CHECK_ASSEMBLER_HWCAP): Define.
> 	* configure.ac: Call GLIBCXX_CHECK_ASSEMBLER_HWCAP.
> 	* fragment.am (CONFIG_CXXFLAGS): Add $(HWCAP_FLAGS).
> 	* configure: Regenerate.
> 	* Makefile.in: Regenerate.
> 	* doc/Makefile.in: Regenerate.
> 	* include/Makefile.in: Regenerate.
> 	* libsupc++/Makefile.in: Regenerate.
> 	* po/Makefile.in: Regenerate.
> 	* python/Makefile.in: Regenerate.
> 	* src/Makefile.in: Regenerate.
> 	* src/c++11/Makefile.in: Regenerate.
> 	* src/c++98/Makefile.in: Regenerate.
> 	* testsuite/Makefile.in: Regenerate.
> 
> 
> 
> 

Ok.

Paolo

Patch

# HG changeset patch
# Parent cc6aab46be72c37bfdfccd786ed5c332a7ce4cd9
Clear hardware capabilities on libstdc++.so with Sun as

diff --git a/libstdc++-v3/acinclude.m4 b/libstdc++-v3/acinclude.m4
--- a/libstdc++-v3/acinclude.m4
+++ b/libstdc++-v3/acinclude.m4
@@ -169,6 +169,32 @@  AC_DEFUN([GLIBCXX_CHECK_COMPILER_FEATURE
 
 
 dnl
+dnl Check if the assembler used supports disabling generation of hardware
+dnl capabilities.  This is only supported by Sun as at the moment.
+dnl
+dnl Defines:
+dnl  HWCAP_FLAGS='-Wa,-nH' if possible.
+dnl
+AC_DEFUN([GLIBCXX_CHECK_ASSEMBLER_HWCAP], [
+  test -z "$HWCAP_FLAGS" && HWCAP_FLAGS=''
+
+  ac_save_CFLAGS="$CFLAGS"
+  CFLAGS="$CFLAGS -Wa,-nH"
+
+  AC_MSG_CHECKING([for as that supports -Wa,-nH])
+  AC_TRY_COMPILE([], [return 0;], [ac_hwcap_flags=yes],[ac_hwcap_flags=no])
+  if test "$ac_hwcap_flags" = "yes"; then
+    HWCAP_FLAGS="-Wa,-nH $HWCAP_FLAGS"
+  fi
+  AC_MSG_RESULT($ac_hwcap_flags)
+
+  CFLAGS="$ac_save_CFLAGS"
+
+  AC_SUBST(HWCAP_FLAGS)
+])
+
+
+dnl
 dnl If GNU ld is in use, check to see if tricky linker opts can be used.  If
 dnl the native linker is in use, all variables will be defined to something
 dnl safe (like an empty string).
diff --git a/libstdc++-v3/configure.ac b/libstdc++-v3/configure.ac
--- a/libstdc++-v3/configure.ac
+++ b/libstdc++-v3/configure.ac
@@ -333,6 +333,9 @@  case "$target" in
 esac
 GLIBCXX_CONDITIONAL(GLIBCXX_LDBL_COMPAT, test $ac_ldbl_compat = yes)
 
+# Check if assembler supports disabling hardware capability support.
+GLIBCXX_CHECK_ASSEMBLER_HWCAP
+
 # Check if assembler supports rdrand opcode.
 GLIBCXX_CHECK_X86_RDRAND
 
diff --git a/libstdc++-v3/fragment.am b/libstdc++-v3/fragment.am
--- a/libstdc++-v3/fragment.am
+++ b/libstdc++-v3/fragment.am
@@ -22,7 +22,7 @@  endif
 # These bits are all figured out from configure.  Look in acinclude.m4
 # or configure.ac to see how they are set.  See GLIBCXX_EXPORT_FLAGS.
 CONFIG_CXXFLAGS = \
-	$(SECTION_FLAGS) $(EXTRA_CXX_FLAGS) -frandom-seed=$@
+	$(SECTION_FLAGS) $(HWCAP_FLAGS) $(EXTRA_CXX_FLAGS) -frandom-seed=$@
 WARN_CXXFLAGS = \
 	$(WARN_FLAGS) $(WERROR_FLAG) -fdiagnostics-show-location=once