Re: [patch RFC] SH: -fstack-protector support

[Jakub Jelinek <jakub at redhat dot com>]

On Wed, Jul 20, 2005 at 02:57:34PM +0100, Joern RENNECKE wrote:
> >It seems a typo in the stack_protect_test documentation.
> >
> > This pattern, if defined, compares a @code{Pmode} value from the
> > memory in operand 1 with the memory in operand 0 without leaving the
> > value in a register afterward and branches to operand 2 if the values
> > weren't equal.
> >     ^^^
> >The last line should be "were equal", shouldn't it?  i386 and ppc
> >implementations do so.
> > 
> >
> That is a likely explanation, but till we get clarification from Jacub, we
> also have to consider the possibility that it was a bug in the original
> implementation that was copied to the ppc port.

operand2 is label before the call to __stack_chk_fail{,_local}, so of course
the pattern needs to jump to that label if the 2 values (canary value
in __stack_chk_guard resp. in TLS area and the stack slot to which
that value was saved in the prologue) are NOT equal.
So IMHO the documentation is correct and what i386 and ppc (and other
platforms) implement is also ok, say on i386 a trivial testcase
with -m32 -fstack-protector -O2 -fpic leads to:
void bar (char *) __attribute__((visibility ("hidden")));
void foo (void)
{
  char buf[64];
  bar (buf);
}

.globl foo
        .type   foo, @function
foo:
        pushl   %ebp
        movl    %esp, %ebp
        subl    $100, %esp
        movl    %gs:20, %eax
        movl    %eax, -4(%ebp)
        xorl    %eax, %eax
        leal    -68(%ebp), %eax
        pushl   %eax
        call    bar
        addl    $16, %esp
        movl    -4(%ebp), %eax
        xorl    %gs:20, %eax
        jne     .L5
        leave
        ret
.L5:
        call    __stack_chk_fail_local
        .size   foo, .-foo

	Jakub