bug#18438: 24.4.50; assertion failed in bidi.c

[Ken Brown <kbrown@cornell.edu>]

On 10/10/2014 11:12 AM, Eli Zaretskii wrote:
> Something vague about the upper 32 bits of the 64-bit registers.
> (Yes, I'm desperate.)

I'm desperate too.  Here's another thought: Suppose this really is a 
thread-safety issue in some way that we don't understand.  Then maybe 
the problem is that the test 'type <= 23' is not atomic in the 
compilation that Aidan and I have been doing.  First 'type' is copied 
from ECX to RBP+0x10, then the latter is tested.  We could make it 
atomic by forcing GCC to directly test ECX <= 23.  We can do this by 
compiling with -Og instead of -O0. (Aidan and I have both been using -O0.)

The resulting disassembly (based on your earlier patch, in 
http://debbugs.gnu.org/cgi/bugreport.cgi?bug=18438#103) is

Dump of assembler code for function bidi_check_type:
    0x00000001004ee9db <+0>:     push   %rbx
    0x00000001004ee9dc <+1>:     sub    $0x40,%rsp
    0x00000001004ee9e0 <+5>:     mov    %ecx,%ebx
    0x00000001004ee9e2 <+7>:     mov    0x543027(%rip),%rax        # 
0x100a31a10 <.refptr.suppress_checking>
    0x00000001004ee9e9 <+14>:    cmpb   $0x0,(%rax)
    0x00000001004ee9ec <+17>:    jne    0x1004eea2f <bidi_check_type+84>
    0x00000001004ee9ee <+19>:    cmp    $0x17,%ecx
    0x00000001004ee9f1 <+22>:    jbe    0x1004eea2f <bidi_check_type+84>
    0x00000001004ee9f3 <+24>:    callq  0x10069fd40 <__getreent>
    0x00000001004ee9f8 <+29>:    mov    0x18(%rax),%rcx
    0x00000001004ee9fc <+33>:    movl   $0x17,0x30(%rsp)
    0x00000001004eea04 <+41>:    movl   $0x0,0x28(%rsp)
    0x00000001004eea0c <+49>:    mov    %ebx,0x20(%rsp)
    0x00000001004eea10 <+53>:    mov    $0x14c,%r9d
    0x00000001004eea16 <+59>:    lea    0x51e713(%rip),%r8        # 
0x100a0d130 <chartab_size+112>
    0x00000001004eea1d <+66>:    lea    0x51e814(%rip),%rdx        # 
0x100a0d238 <chartab_size+376>
    0x00000001004eea24 <+73>:    callq  0x1006a0040 <fprintf>
    0x00000001004eea29 <+78>:    callq  0x10065227d <emacs_abort>
    0x00000001004eea2e <+83>:    nop
    0x00000001004eea2f <+84>:    add    $0x40,%rsp
    0x00000001004eea33 <+88>:    pop    %rbx
    0x00000001004eea34 <+89>:    retq
End of assembler dump.

Do you think this is worth trying (perhaps after Aidan tries your other 
suggestion, involving 64-bit registers)?

Ken