Re: remaining sql/json patches - Mailing list pgsql-hackers

On Wed, Oct 4, 2023 at 10:26 PM Amit Langote <amitlangote09@gmail.com> wrote:
> On Tue, Oct 3, 2023 at 10:11 PM Amit Langote <amitlangote09@gmail.com> wrote:
> > On Mon, Oct 2, 2023 at 2:26 PM Amit Langote <amitlangote09@gmail.com> wrote:
> > > On Mon, Oct 2, 2023 at 1:24 PM Amit Langote <amitlangote09@gmail.com> wrote:
> > > > Pushed this 30 min ago (no email on -committers yet!) and am looking
> > > > at the following llvm crash reported by buildfarm animal pogona [1]:
> > > > This seems to me to be complaining about the following addition:
> > > >
> > > > +                   {
> > > > +                       Oid         ioparam = op->d.iocoerce.typioparam;
> > > > +                       LLVMValueRef v_params[6];
> > > > +                       LLVMValueRef v_success;
> > > > +
> > > > +                       v_params[0] = l_ptr_const(op->d.iocoerce.finfo_in,
> > > > +                                                 l_ptr(StructFmgrInfo));
> > > > +                       v_params[1] = v_output;
> > > > +                       v_params[2] = l_oid_const(lc, ioparam);
> > > > +                       v_params[3] = l_int32_const(lc, -1);
> > > > +                       v_params[4] = l_ptr_const(op->d.iocoerce.escontext,
> > > > +
> > > > l_ptr(StructErrorSaveContext));
> > > >
> > > > -                   LLVMBuildStore(b, v_retval, v_resvaluep);
> > > > +                       /*
> > > > +                        * InputFunctionCallSafe() will write directly into
> > > > +                        * *op->resvalue.
> > > > +                        */
> > > > +                       v_params[5] = v_resvaluep;
> > > > +
> > > > +                       v_success = LLVMBuildCall(b, llvm_pg_func(mod,
> > > > "InputFunctionCallSafe"),
> > > > +                                                 v_params, lengthof(v_params),
> > > > +                                                 "funccall_iocoerce_in_safe");
> > > > +
> > > > +                       /*
> > > > +                        * Return null if InputFunctionCallSafe() encountered
> > > > +                        * an error.
> > > > +                        */
> > > > +                       v_resnullp = LLVMBuildICmp(b, LLVMIntEQ, v_success,
> > > > +                                                  l_sbool_const(0), "");
> > > > +                   }
> > >
> ...I haven't yet pinpointed down
> which of the LLVM's asserts it is, nor have I been able to walk
> through LLVM source code using gdb to figure what the new code is
> doing wrong.  Maybe I'm still missing a trick or two...

I finally managed to analyze the crash by getting the correct LLVM build.

So the following bits are the culprits:

1. v_output needed to be converted from being reference to a Datum to
be reference to char * as follows before passing to
InputFunctionCallSafe():

-                       v_params[1] = v_output;
+                       v_params[1] = LLVMBuildIntToPtr(b, v_output,
+
l_ptr(LLVMInt8TypeInContext(lc)),
+                                                       "");

2. Assignment of op->d.iocoerce.escontext needed to be changed like this:

                        v_params[4] = l_ptr_const(op->d.iocoerce.escontext,
-
l_ptr(StructErrorSaveContext));
+                                                 l_ptr(StructNode));

3. v_success needed to be "zero-extended" to match in type with
whatever s_bool_const() produces, as follows:

+                       v_success = LLVMBuildZExt(b, v_success,
TypeStorageBool, "");
                        v_resnullp = LLVMBuildICmp(b, LLVMIntEQ, v_success,
                                                   l_sbool_const(0), "");

No more crashes with the above fixes.

Attached shows the delta against the patch I reverted.  I'll push the
fixed up version on Monday.

--
Thanks, Amit Langote
EDB: http://www.enterprisedb.com