On Mar 3 13:19, Jeremy Drake via Cygwin-patches wrote:
> On Mon, 3 Mar 2025, Corinna Vinschen wrote:
>
> > > diff --git a/winsup/cygwin/path.cc b/winsup/cygwin/path.cc
> > > index 599809f941..49740ac465 100644
> > > --- a/winsup/cygwin/path.cc
> > > +++ b/winsup/cygwin/path.cc
> > > @@ -4539,6 +4539,18 @@ find_fast_cwd_pointer ()
> > > %rcx for the subsequent RtlEnterCriticalSection call. */
> > > lock = (const uint8_t *) memmem ((const char *) use_cwd, 80,
> > > "\x48\x8d\x0d", 3);
> > > + if (lock)
> > > + {
> > > + /* A recent Windows 11 Preview calls `lea rel(rip),%rcx' then
> > > + a `mov` and a `movups` instruction, and only then
> > > + `callq RtlEnterCriticalSection'.
> > > + */
> > > + if (memmem (lock + 7, 8, "\x4c\x89\x78\x10\x0f\x11\x40\xc8", 8))
> >
> > Is it really necessary to check for each and every byte between lea and
> > callq? I wonder if this can't be simpler by simply checking for the
> > '\x48\x8d\x0d` needle and then, instead of just assuming a fixed
> > call_rtl_offset, skip programatically to the next callq 0xe8 byte
> > within the next 16 bytes or so?
>
> I think looking for only a single byte might have too high a probability
> of a false-positive match inside a multi-byte instruction. As you said
>
> > It needs a lot of knowledge of instructons and their respective length,
> > to skip the uninteresting parts.
Yeah, sure. I'm a bit concerned that this expression, testing every
single byte for a fixed value, will only match this very preview build,
is all. Shouldn't we give a little slack for different registers at
least? Kind of like
movq <somereg>, <someoffset>(%rax) 0x4c 0x89 + 2 bytes
movups xmm0, <someoffset>(%rax) 0x0f 0x11 + 2 bytes
?
Thanks,
Corinna
