On Mon, 2015-06-29 at 21:38 -0400, David Malcolm wrote:
> On Mon, 2015-06-29 at 21:11 -0400, David Malcolm wrote:
> > On Mon, 2015-06-29 at 22:25 +0100, Dibyendu Majumdar wrote:
> > > On 29 June 2015 at 22:08, David Malcolm <dmalcolm@redhat.com> wrote:
> > > 
> > > > Looking at src/lvm.c, I see that the signature of luaV_tonumber and that
> > > > lua_Number can be various types, including "double".
> > > >   int luaV_tonumber_ (const TValue *obj, lua_Number *n)
> > > >
> > > 
> > > I am using double as lua_Number.
> > > 
> > > > Does the type signature of luaV_tonumber as compiled by the main
> > > > compiler agree with that supplied by gcc_jit_context_new_function?  If
> > > > one of them is e.g. expecting a float *, but the other is expecting a
> > > > double *, then you might get the symptoms you're seeing.
> > > >
> > > 
> > > I checked this - I think they match.
> > > 
> > > > If that doesn't highlight a cause, maybe you're running into a libgccjit
> > > > bug.  If so, can you generate a reproducer and post it here so I can
> > > > poke at it please?
> > > >
> > > 
> > > Attached is a reproducer.
> > 
> > Thanks; I'm able to run the reproducer, at least to the point of being
> > able to invoke gcc_jit_context_compile on it.
> > 
> > Interestingly, looking at the gimple dump (via
> > GCC_JIT_BOOL_OPTION_DUMP_INITIAL_GIMPLE), I see:
> > 
> >   <float:64> OP_RAVI_TOFLT_n_2_8.0;
> >   <float:64> OP_RAVI_TOFLT_n_2_8.1;
> >   [...snip...]
> >   <float:64> OP_RAVI_TOFLT_n_2_8;
> >   [...snip...]
> > 
> > OP_RAVI_TOFLT_if_not_float_2_10:
> >   OP_RAVI_TOFLT_n_2_8.0 = OP_RAVI_TOFLT_n_2_8;
> >   printf ("number %p = %f before call to luaV_number",
> > &OP_RAVI_TOFLT_n_2_8.0, OP_RAVI_TOFLT_n_2_8);
> >   OP_RAVI_TOFLT_n_2_8.1 = OP_RAVI_TOFLT_n_2_8;
> >   [...snip...]
> >   OP_RAVI_TOFLT_n_2_8.1 = OP_RAVI_TOFLT_n_2_8;
> >   D.392 = 0;
> >   D.393 = base + D.392;
> >   D.405 = luaV_tonumber_ (D.393, &OP_RAVI_TOFLT_n_2_8.1);
> >           ^^^^^^^^^^^^^^ here's the call          ^^^^^
> >   comparison_0_12 = D.405 == 0;
> >   if (comparison_0_12 != 0) goto
> > OP_RAVI_TOFLT_if_conversion_failed_2_13; else goto
> > OP_RAVI_TOFLT_if_conversion_ok_2_14;
> > 
> > OP_RAVI_TOFLT_done_2_11:
> >   D.390 = L->ci;
> >   base = D.390->u.l.base;
> >   D.399 = 16;
> >   D.400 = base + D.399;
> >   L->top = D.400;
> >   D.394 = cl->p;
> >   D.406 = D.394->sizep;
> >   comparison_0_15 = D.406 > 0;
> >   if (comparison_0_15 != 0) goto OP_RETURN_if_sizep_gt_0_3_16; else goto
> > OP_RETURN_else_sizep_gt_0_3_17;
> > 
> > OP_RAVI_TOFLT_if_conversion_failed_2_13:
> >   luaG_runerror (L, "number expected");
> >   goto OP_RAVI_TOFLT_if_conversion_ok_2_14;
> >   OP_RAVI_TOFLT_if_conversion_ok_2_14:
> >   printf ("number ok = %f", OP_RAVI_TOFLT_n_2_8);
> > 
> >   [...snip...]
> > 
> > which, if I'm reading it right, suggests that the local has effectively
> > been split into three locals during the conversion to gimple, and that a
> > ptr to
> >   OP_RAVI_TOFLT_n_2_8.1
> > is passed to luaV_tonumber_, but
> >   OP_RAVI_TOFLT_n_2_8
> > is then used.
> > 
> > This is feeling like a bug at my end; sorry.  I'll continue to
> > investigate.
> 
> Definitely looks like a bug at my end.
> 
> I've created a minimal reproducer, and filed it as
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66700.
> 
> I plan to investigate further tomorrow.
> 
> Sorry about this.

Does the attached patch fix it for you?
(it's currently a bit of a hack; hope to get it more robust tomorrow)