Steven Bosscher <stevenb.gcc@gmail.com> writes:

> Therefore, an RFA for the attached patch. Bootstrapped&tested on
> powerpc64-unknown-linux-gnu. OK?

Unfortunately, this patch completely broke libjava testing on
i386-pc-solaris2* and x86_64-unknown-linux-gnu: all execution tests fail
with a SEGV:

Program received signal SIGSEGV, Segmentation fault.
0xfd0ddf7d in _Jv_Linker::verify_class (klass=klass@entry=0xfe945d80)
    at /vol/gcc/src/hg/trunk/local/libjava/link.cc:1904
1904      klass->engine->verify(klass);

klass->engine is NULL at this point.  Checking the generated .jcr
sections, I found that a large number of them have sh_addralign 0x20
instead of 0x4 before your patch.  This caused a massive reduction in
the number of _Jv_RegisterClassHookDefault calls:

With the following DTrace one-liner

pid$target::_Jv_RegisterClassHookDefault:entry{ @[ustack()] = count(); }

to investigate ArrayStore.exe execution, I find for the unpatched jc1:

              libgcj.so.13.0.0`_Jv_RegisterClassHookDefault
              libgcj.so.13.0.0`_Jv_RegisterClasses+0x44
              libgcj.so.13.0.0`_ZN4java4lang9Cloneable6class$E
              0xe0ec
             7213

With your patch, I get

              libgcj.so.13.0.0`_Jv_RegisterClassHookDefault
              libgcj.so.13.0.0`_Jv_RegisterNewClasses+0xb7
              ArrayStore.exe`_Utf18
                1

              libgcj.so.13.0.0`_Jv_RegisterClassHookDefault
              libgcj.so.13.0.0`_Jv_RegisterClasses+0x44
              libgcj.so.13.0.0`_ZN4java4lang9Cloneable6class$E
              0xe0ec
               12

instead.

The alignment change happens because LOCAL_ALIGNMENT overrides
DECL_ALIGN.  This can be avoided by setting DECL_USER_ALIGN (which
unfortunately is undocumented).  The result looks better, but still
fails in a different way:

Exception in thread "main" java.lang.NoClassDefFoundError: ArrayStore
   at java.lang.Class.initializeClass(libgcj.so.13.0.0)
Caused by: java.lang.NullPointerException
   at java.lang.Class.initializeClass(libgcj.so.13.0.0)

But the _Jv_RegisterClassHookDefault are almost back to normal:

              libgcj.so.13.0.0`_Jv_RegisterClassHookDefault
              libgcj.so.13.0.0`_Jv_RegisterClasses+0x44
              libgcj.so.13.0.0`_ZN4java4lang9Cloneable6class$E
              0xe0ec
             7212

I could trace this to a change in .jcr section flags: in ArrayStore.exe,
I find

Section Header[18]:  sh_name: .jcr
    sh_addr:      0x8052058       sh_flags:   [ SHF_ALLOC ]
    sh_size:      0x4             sh_type:    [ SHT_PROGBITS ]
    sh_offset:    0x2058          sh_entsize: 0
    sh_link:      0               sh_info:    0
    sh_addralign: 0x4       

Section Header[26]:  sh_name: .jcr
    sh_addr:      0x80625e4       sh_flags:   [ SHF_WRITE SHF_ALLOC ]
    sh_size:      0x4             sh_type:    [ SHT_PROGBITS ]
    sh_offset:    0x25e4          sh_entsize: 0
    sh_link:      0               sh_info:    0
    sh_addralign: 0x4       

The first (without SHF_WRITE set) is from jc1, the second from
crtbegin.o.  The Solaris linker (and probably other ELF linkers which
don't special-case .jcr) merges sections not just by name, but only if
name and attributs (flags, type) match.  Not marking .jcr read-only
fixes that part of the failure and restores libjava testsuite results on
Solaris 11/x86 back to normal (no failures).

Bootstrapped without regressions on i386-pc-solaris2.11.

Ok for mainline?

	Rainer


2012-03-31  Rainer Orth  <ro@CeBiTec.Uni-Bielefeld.DE>

	* class.c (emit_register_classes_in_jcr_section): Set DECL_USER_ALIGN.
	Clear TREE_READONLY.