Hey Zhongwei,

If you want to take a look at ARM64 frame layout you can find it in
vm_arm64.dasc, around line 90 (or in lj_frame.h, since they have to
match).

Now, SPS_FIXED basically says to JIT that it can use part of the
interpreter's frame. If we take a look at 32-bit ARM (vm_arm.dasc:60),
we can see that it has 2 temporary slots at the top of the stack, which
don't need to be preserved. But on ARM64, there is no temporary slot on
top of stack, which (I would say) means that SPS_FIXED should be 0.

ARM64 does actually have one 8-byte temporary slot, but it's deep in
stack (sp+184), which doesn't make it usable for JIT. I guess that we
could rearrange the frame and put it to the top of the frame, but I'm
not sure how smart is that. Mike Pall probably put it there for a
reason.

Anyway, there might be more to this than what I said. I encourage you
to take a better look

Stefan

On Mon, 1 Aug 2016 09:02:47 +0800
Zhongwei Yao <zhongwei.yao@linaro.org> wrote:

> Oh, sorry! I didn't update it to my latest test case:(
>
> Here is:
>
> ======
> x = 0
>
> for i = 1, 100 do
>   x = tostring(i)
> end
>
> y = "101"
>
> assert(x == y, "Got " .. x .. ", expect " .. y)
> ======
>
> It runs into Segment Fault instead of giving assertion error.
>
> I'll take a look at the SPS_FIXED related issue.
>
> On 29 July 2016 at 22:05, Charles Baylis <charles.baylis@linaro.org>
> wrote:
>
> > On 29 July 2016 at 10:12, Zhongwei Yao <zhongwei.yao@linaro.org>
> > wrote:
> > > Hi, all,
> > > I run into Segment Fault in:
> > >
> > > ======
> > >
> > > x = 0
> > >
> > > for i = 1, 100 do
> > >   x = tostring(i)
> > > end
> > >
> > > y = "101"
> > >
> > > ======
> > >
> > > And it turns out the JITed code overwrites native stack position
> > > at trace head. It can be walked around by set SPS_FIXED to 0.
> > >
> > > The JITed code is doing register spill and assuming there is "2
> > > fixed
> > spill
> > > slots" (defined by SPS_FIXED). But there seems no "fixed spill
> > > slots" in arm64 interpreter implementation. Is it true? Could
> > > anyone provide some
> > info
> > > of aarch64 frame layout?
> > >
> > > Any input will help! Thanks.
> >
> > I've tried your test case here, and it works ok for me. I added
> > print(x) and print(y) and I get 100 and 101 as expected. Do you have
> > an old tree or some local patches?
> >
> > I added the SPS_FIXED values very early in the port, but they are
> > only copy-and-pasted, so are quite likely to be wrong. Any
> > investigation you can do is valuable :)
> >
> > There are a few comments about LuaJIT frame layout in lj_frame.h,
> > but I haven't found any clear explanations about how it really
> > works, and there a few hard-coded constants in the interpreter
> > which may be tied to the frame layout.
> >
> > Charles
> >
>
>
>