configure: workaround for Clang 3.5.0

Clang++ 3.5 on Fedora 22 appears to have difficulty tolerating D_FORTIFY_SOURCE for certain glibc headers, such as stdio.

This interferes, currently, with any arm target build.

Work around this by disabling FORTIFY_SOURCE for clang builds if a problem is observed.

Newer versions of clang such as 3.5.2 (As seen in debian-testing) or 3.7.0 (As seen in Fedora 23 Beta) are unaffected and will not trigger this workaround.

Signed-off-by John Snow <[email protected]>

---

 configure | 25 ++++++++++++++++++++++++-
 1 file changed, 24 insertions(+), 1 deletion(-)

configure +24 -1

index 7a1d08d..7abfcc3 100755
@@ -107,6 +107,11 @@ compile_object() {
   do_cc $QEMU_CFLAGS $local_cflags -c -o $TMPO $TMPC
 }
 
+compile_cxx_object() {
+  local_cflags="$1"
+  do_cxx $QEMU_CXXFLAGS $local_cflags -c -o $TMPO $TMPC
+}
+
 compile_prog() {
   local_cflags="$1"
   local_ldflags="$2"

@@ -4436,13 +4441,31 @@ if ! compile_object "-Werror"; then
 fi
 
 ##########################################
+# Test that we can use FORTIFY_SOURCE,
+# which might break Clang.
+
+if test "$debug" = "no"; then
+  cat > $TMPC << EOF
+#include <cstdio>
+int main(int argc, char*argv[]) {
+  fprintf(stdout, "Hello World\n");
+  return 0;
+}
+EOF
+
+  if ! compile_cxx_object "-O2 -U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=2"; then
+      fortify_source="no";
+  fi
+fi
+
+##########################################
 # End of CC checks
 # After here, no more $cc or $ld runs
 
 if test "$gcov" = "yes" ; then
   CFLAGS="-fprofile-arcs -ftest-coverage -g $CFLAGS"
   LDFLAGS="-fprofile-arcs -ftest-coverage $LDFLAGS"
-elif test "$debug" = "no" ; then
+elif test "$debug" = "no" && test "$fortify_source" != "no" ; then
   CFLAGS="-O2 -U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=2 $CFLAGS"
 fi
 

Quoted 8 lines · John Snow

Clang++ 3.5 on Fedora 22 appears to have difficulty tolerating D_FORTIFY_SOURCE for certain glibc headers, such as stdio.

This interferes, currently, with any arm target build.

Work around this by disabling FORTIFY_SOURCE for clang builds if a problem is observed.

Quoted 6 lines

Newer versions of clang such as 3.5.2 (As seen in debian-testing) or 3.7.0 (As seen in Fedora 23 Beta) are unaffected and will not trigger this workaround.

Signed-off-by John Snow <[email protected]>

---

As a meta-cover-letter, this fix is a little weird in that it will disable FORTIFY_SOURCE (silently!) for non-debug builds. Not great.

(Maybe I could have it fail and print a warning encouraging users to either use --enable-debug or --disable-fortify-source?)

Still, It'd be nice to have Clang builds working out of the box for ARM builds. It just so happens that ARM is the only target that happens to trip this specific unfortunate chain of events.

Is it sane to check for clang-and-arm-targets only? Maybe it's a moot point -- newer (and older, I believe) versions of Clang won't trigger this at all.

Anyway, see these links (Dredged up by Laszlo Ersek, thanks!)

https://llvm.org/bugs/show_bug.cgi?id=7219
https://llvm.org/bugs/show_bug.cgi?id=16821
https://llvm.org/bugs/show_bug.cgi?id=23277#c2

Here's a good takeaway quote:

'Also, _FORTIFY_SOURCE + glibc + clang is not supported and does not work (for instance, it relies on __builtin_va_pack_len and friends, which we have no intention of supporting), so glibc compatibility is unlikely to be a strong motivator for a change here.'

So long story short, we have a weird hacky workaround where we disable FORTIFY_SOURCE for compilers that don't appear to be able to support it.

--js

Quoted 53 lines

 configure | 25 ++++++++++++++++++++++++-
 1 file changed, 24 insertions(+), 1 deletion(-)

configure +24 -1

index 7a1d08d..7abfcc3 100755
@@ -107,6 +107,11 @@ compile_object() {
   do_cc $QEMU_CFLAGS $local_cflags -c -o $TMPO $TMPC
 }
 
+compile_cxx_object() {
+  local_cflags="$1"
+  do_cxx $QEMU_CXXFLAGS $local_cflags -c -o $TMPO $TMPC
+}
+
 compile_prog() {
   local_cflags="$1"
   local_ldflags="$2"

@@ -4436,13 +4441,31 @@ if ! compile_object "-Werror"; then
 fi
 
 ##########################################
+# Test that we can use FORTIFY_SOURCE,
+# which might break Clang.
+
+if test "$debug" = "no"; then
+  cat > $TMPC << EOF
+#include <cstdio>
+int main(int argc, char*argv[]) {
+  fprintf(stdout, "Hello World\n");
+  return 0;
+}
+EOF
+
+  if ! compile_cxx_object "-O2 -U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=2"; then
+      fortify_source="no";
+  fi
+fi
+
+##########################################
 # End of CC checks
 # After here, no more $cc or $ld runs
 
 if test "$gcov" = "yes" ; then
   CFLAGS="-fprofile-arcs -ftest-coverage -g $CFLAGS"
   LDFLAGS="-fprofile-arcs -ftest-coverage $LDFLAGS"
-elif test "$debug" = "no" ; then
+elif test "$debug" = "no" && test "$fortify_source" != "no" ; then
   CFLAGS="-O2 -U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=2 $CFLAGS"
 fi
 

Quoted 6 lines · John Snow

Here's a good takeaway quote:

'Also, _FORTIFY_SOURCE + glibc + clang is not supported and does not work (for instance, it relies on __builtin_va_pack_len and friends, which we have no intention of supporting), so glibc compatibility is unlikely to be a strong motivator for a change here.'

this quote would be compelling enough for me to disable _FORTIFY_SOURCE when clang is seen, no questions asked. The above is a public no-support statement from an apparently core clang developer, so "it happens to build without errors with version X.Y.Z." just don't cut it. A positive claim (bugzilla comment, release note, etc) would be necessary.

... As far as I'm concerned, of course. :)

Quoted 14 lines · Laszlo Ersek

Quoted 6 lines · John Snow

Here's a good takeaway quote:

'Also, _FORTIFY_SOURCE + glibc + clang is not supported and does not work (for instance, it relies on __builtin_va_pack_len and friends, which we have no intention of supporting), so glibc compatibility is unlikely to be a strong motivator for a change here.'

this quote would be compelling enough for me to disable _FORTIFY_SOURCE when clang is seen, no questions asked. The above is a public no-support statement from an apparently core clang developer, so "it happens to build without errors with version X.Y.Z." just don't cut it. A positive claim (bugzilla comment, release note, etc) would be necessary.

Quoted 1 line

... As far as I'm concerned, of course. :)

I think I would agree with that.

thanks
-- PMM