[BUG mm] "fixed" i386 memcpy inlining buggy

[Christophe Saout]

Hi Denis,

the new i386 memcpy macro is a ticking timebomb.

I've been debugging a new mISDN crash, just to find out that a memcpy
was not inlined correctly.

Andrew, you should drop the fix-i386-memcpy.patch (or have it fixed).

This source code:

        mISDN_pid_t     pid;
	[...]
        memcpy(&st->mgr->pid, &pid, sizeof(mISDN_pid_t));

was compiled as:

        lea    0xffffffa4(%ebp),%esi    <---- %esi is loaded
(       add    $0x10,%ebx                      )
(       mov    %ebx,%eax                       ) something else
(       call   1613 <test_stack_protocol+0x83> ) %esi preserved
        mov    0xffffffa0(%ebp),%edx
        mov    0x74(%edx),%edi          <---- %edi is loaded
        add    $0x20,%edi                     offset in structure added
!       mov    $0x14,%esi        !!!!!! <---- %esi overwritten!
        mov    %esi,%ecx                <---- %ecx loaded
        repz movsl %ds:(%esi),%es:(%edi)

Apparently the compiled decided that the value 0x14 could be reused
afterwards (which it does for an inlined memset of the same size some
instructions below) and clobbers %esi.

Looking at the macro:

                __asm__ __volatile__(
                        ""
                        : "=&D" (edi), "=&S" (esi)
                        : "0" ((long) to),"1" ((long) from)
                        : "memory"
                );
        }
        if (n >= 5*4) {
                /* large block: use rep prefix */
                int ecx;
                __asm__ __volatile__(
                        "rep ; movsl"
                        : "=&c" (ecx)

it seems obvious that the compiled assumes it can reuse %esi and %edi
for something else between the two __asm__ sections. These should
probably be merged.

Attachment: signature.asc
Description: This is a digitally signed message part