Re: gcc 15 "array subscript 0" warning at level -O3 - Mailing list pgsql-hackers

Hi,

On 2025-04-25 13:37:15 -0400, Tom Lane wrote:
> Whilst poking at Erik Rijkers' nearby report, I found that
> Fedora 42's gcc 15.0.1 will produce this complaint if you
> select optimization level -O3:
> 
> In file included from ../../../../src/include/access/htup_details.h:22,
>                  from pl_exec.c:21:
> In function 'assign_simple_var',
>     inlined from 'exec_set_found' at pl_exec.c:8609:2:
> ../../../../src/include/varatt.h:230:36: warning: array subscript 0 is outside array bounds of 'char[0]'
[-Warray-bounds=]
>   230 |         (((varattrib_1b_e *) (PTR))->va_tag)
>       |                                    ^
> ../../../../src/include/varatt.h:94:12: note: in definition of macro 'VARTAG_IS_EXPANDED'
>    94 |         (((tag) & ~1) == VARTAG_EXPANDED_RO)
>       |            ^~~
> ../../../../src/include/varatt.h:284:57: note: in expansion of macro 'VARTAG_1B_E'
>   284 | #define VARTAG_EXTERNAL(PTR)                            VARTAG_1B_E(PTR)
>       |                                                         ^~~~~~~~~~~
> ../../../../src/include/varatt.h:301:57: note: in expansion of macro 'VARTAG_EXTERNAL'
>   301 |         (VARATT_IS_EXTERNAL(PTR) && !VARTAG_IS_EXPANDED(VARTAG_EXTERNAL(PTR)))
>       |                                                         ^~~~~~~~~~~~~~~
> pl_exec.c:8797:17: note: in expansion of macro 'VARATT_IS_EXTERNAL_NON_EXPANDED'
>  8797 |                 VARATT_IS_EXTERNAL_NON_EXPANDED(DatumGetPointer(newvalue)))
>       |                 ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> In function 'exec_set_found':
> cc1: note: source object is likely at address zero

FWIW, I've seen this even before GCC 15.


> Buildfarm member serinus has been producing the identical warning for
> some time.  I'd been ignoring that because it runs "experimental gcc",
> but I guess the experiment has leaked out to production distros.
> 
> What seems to be happening here is that after inlining
> assign_simple_var into exec_set_found, the compiler decides that
> "newvalue" might be zero (since it's a BoolGetDatum result),
> and then it warns -- in a rather strange way -- about the
> potential null dereference.

I don't think it actually is complaining about a null dereference - it thinks
we're interpreting a boolean as a pointer (for which it obviously is not wide
enough)


> The dereference is not reachable
> because of the preceding "var->datatype->typlen == -1" check,
> but that's not stopping the optimizer from bitching.

> I experimented with modifying exec_set_found thus:
> 
>     var = (PLpgSQL_var *) (estate->datums[estate->found_varno]);
> +    Assert(var->datatype->typlen == 1);
>     assign_simple_var(estate, var, BoolGetDatum(state), false, false);
> 
> which should be OK since we're expecting the "found" variable to
> be boolean.  That does silence the warning, but of course only
> in --enable-cassert builds.

One way to address this is outlined here:

https://postgr.es/m/20230316172818.x6375uvheom3ibt2%40awork3.anarazel.de
https://postgr.es/m/20240207203138.sknifhlppdtgtxnk%40awork3.anarazel.de

I've been wondering about adding wrapping something like that in a
pg_assume(expr) or such.

Greetings,

Andres Freund