Re: [Tarantool-patches] [PATCH] jit: fix cdatanum addressing for GC64 mode on x86

[Igor Munkin <imun@tarantool.org>]

Sergos,

Thanks for you review!

On 14.10.20, Sergey Ostanevich wrote:
> Hi!
> 
> Thanks for the patch, I got one question below.
> 
> Regards,
> Sergos
> 
> On 14 окт 16:53, Igor Munkin wrote:
> > This patch fixes the regression introduced in scope of
> > 5f6775ae0e141422193ad9b492806834064027ca ('core: introduce various
> > platform metrics'). As a result of the patch <cdatanum> displacement is
> > misencoded when GC64 mode is enabled.
> > 
> > In X86 long mode 32-bit displacement is encoded either via SIB byte or
> > is addressed relatively to RIP register value. The first approach is
> > used in JIT for 32-bit addresses (i.e. when GC64 mode is disabled), but
> > doesn't work for 64-bit ones. As a result all addresses to GG_State
> > contents to be "hardcoded" on the trace are encoded relatively to
> > RID_DISPATCH register (i.e. callee-safe R14 register) containing global
> > dispatch table. For this purpose this register is not used by the JIT
> > register allocator in GC64 build and not spoiled throughout LuaJIT VM
> > cycle (and therefore trace execution).
> > 
> > NB: Since R14 is the additional GRP, the <add> instruction ought to be
> > REX-prefixed.
> > 
> > Follows up tarantool/tarantool#5187
> > 
> > Reported-by: Vladislav Shpilevoy <v.shpilevoy@tarantool.org>
> > Signed-off-by: Igor Munkin <imun@tarantool.org>
> > ---
> > 
> > Branch: https://github.com/tarantool/luajit/compare/imun/gh-5187-fix-disp-encoding-on-gc64
> > 
> > Unforunately, CI is red, but those failures relates to the known build
> > issues. Nevertheless I tested the patch manually on tntmac04 and faced
> > no failures.
> > 
> >  src/lj_asm_x86.h | 5 +++++
> >  1 file changed, 5 insertions(+)
> > 
> > diff --git a/src/lj_asm_x86.h b/src/lj_asm_x86.h
> > index 959fc2d..767bf6f 100644
> > --- a/src/lj_asm_x86.h
> > +++ b/src/lj_asm_x86.h
> > @@ -1837,8 +1837,13 @@ static void asm_cnew(ASMState *as, IRIns *ir)
> >  
> >    /* Increment cdatanum counter by address directly. */
> >    emit_i8(as, 1);
> > +#if LJ_GC64
> > +  emit_rmro(as, XO_ARITHi8, XOg_ADD|REX_64, RID_DISPATCH,
> > +	    dispofs(as, &J2G(as->J)->gc.cdatanum));
> 
> Should we cast the disp to 32bit? Here

IIRC, in function calls, the arguments are converted to the types of the
corresponding parameters. <ofs> parameter in <emit_rmro> is int32_t
type, so I guess an explicit cast is not obligatory here, isn't it?

> https://wiki.osdev.org/X86-64_Instruction_Encoding#Displacement
> I see only a disp32. 

However, as you've already mentioned offline the *valid* dispofs values
fit 32-bit integers since the size of GG_State equals to 6344 bytes.

I surmise, these explicit casts around relate to the old dark times when
various compilers were not so good, so Mike had to add such casts
everywhere. I checked the machine code generated by GCC on my machine
and see no difference between two versions: with or without the cast.

> 
> > +#else
> >    emit_rmro(as, XO_ARITHi8, XOg_ADD, RID_NONE,
> >  	    ptr2addr(&J2G(as->J)->gc.cdatanum));
> > +#endif
> >    /* Combine initialization of marked, gct and ctypeid. */
> >    emit_movtomro(as, RID_ECX, RID_RET, offsetof(GCcdata, marked));
> >    emit_gri(as, XG_ARITHi(XOg_OR), RID_ECX,
> > -- 
> > 2.25.0
> > 

-- 
Best regards,
IM