toolchain/45837: perl is miscompiled on sparc64-current

[martin%NetBSD.org@localhost]

>Number:         45837
>Category:       toolchain
>Synopsis:       perl is miscompiled on sparc64-current
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    toolchain-manager
>State:          open
>Class:          sw-bug
>Submitter-Id:   net
>Arrival-Date:   Sun Jan 15 08:05:00 +0000 2012
>Originator:     Martin Husemann
>Release:        NetBSD 5.99.59
>Organization:
The NetBSD Foundation, Inc.
>Environment:
System: NetBSD nelly.aprisoft.de 5.99.59 NetBSD 5.99.59 (NELLY.MP) #186: Fri 
Jan 13 12:43:00 CET 2012 
martin%emmas.aprisoft.de@localhost:/nelly/usr/src/sys/arch/sparc64/compile/NELLY.MP
 sparc64
Architecture: sparc64
Machine: sparc64
>Description:

When perl compiles a perl regexp and midway finds a UTF8 encoded part inside
the pattern it dies with a core dump due to a NULL pointer dereference.

The problem hits in perl regcomp.s line 4563:

   4560     /****************** LONG JUMP TARGET HERE***********************/
   4561     /* Longjmp back to here if have to switch in midstream to utf8 */
   4562     if (! RExC_orig_utf8) {
   4563         JMPENV_PUSH(jump_ret);
   4564         used_setjump = TRUE;
   4565     }

The macro JMPENV_PUSH is used in excepiton handling/unwinding and does
this:

#define JMPENV_PUSH(v) \
    STMT_START {                                                        \
        DEBUG_l({                                                       \
            int i = 0; JMPENV *p = PL_top_env;                          \
            while (p) { i++; p = p->je_prev; }                          \
            Perl_deb(aTHX_ "JUMPENV_PUSH level=%d at %s:%d\n",          \
                         i,  __FILE__, __LINE__);})                     \
        cur_env.je_prev = PL_top_env;                                   \
        OP_REG_TO_MEM;                                                  \
        cur_env.je_ret = PerlProc_setjmp(cur_env.je_buf, SCOPE_SAVES_SIGNAL_M$
        OP_MEM_TO_REG;                                                  \
        PL_top_env = &cur_env;                                          \
        cur_env.je_mustcatch = FALSE;                                   \
        (v) = cur_env.je_ret;                                           \
    } STMT_END


Finally PerlProc_setjmp is a sigsetjmp(). Now the compiler generates this
code for the above invocation:

   0x4058f89c <Perl_re_compile+988>:    ldx  [ %i0 + 0x270 ], %g1
   0x4058f8a0 <Perl_re_compile+992>:    add  %fp, 0x69f, %o0
   0x4058f8a4 <Perl_re_compile+996>:    clr  %o1
   0x4058f8a8 <Perl_re_compile+1000>:   call  0x40780520 <__sigsetjmp14@plt>
   0x4058f8ac <Perl_re_compile+1004>:   stx  %g1, [ %fp + 0x697 ]
   0x4058f8b0 <Perl_re_compile+1008>:   add  %fp, 0x697, %g1
   0x4058f8b4 <Perl_re_compile+1012>:   st  %o0, [ %fp + 0x717 ]
   0x4058f8b8 <Perl_re_compile+1016>:   andcc  %o0, 0xff, %o1
=> 0x4058f8bc <Perl_re_compile+1020>:   stx  %g1, [ %i0 + 0x270 ]

All fine, but after calling sigsetjmp, %i0 should be considered clobbered
as it is caller-saved. Still it just uses it in the marked instructions
and dies - because it is NULL now.

Am I missing something?

>How-To-Repeat:

Try to build mail/spamassasin from pkgsrc, watch perl dump core in the
middle of the build.

>Fix:
n/a