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

[Sergey Ostanevich]

Thanks! 

LGTM.


On 14 окт 21:31, Igor Munkin wrote:
> 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 at tarantool.org>
> > > Signed-off-by: Igor Munkin <imun at 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 thought it should cover the difference from assembly placement to the
_G, but now I got it should cover only the one from dispatch table to 
your gc stuff. 

> 
> 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