Re: HELP: Questions on unshare_expr

[Richard Biener <richard.guenther@gmail.com>]

On Wed, Jan 17, 2024 at 7:42 AM Richard Biener
<richard.guenther@gmail.com> wrote:
>
> On Tue, Jan 16, 2024 at 9:26 PM Qing Zhao <qing.zhao@oracle.com> wrote:
> >
> >
> >
> > > On Jan 15, 2024, at 4:31 AM, Richard Biener <richard.guenther@gmail.com> wrote:
> > >
> > >> All my questions for unshare_expr relate to a  LTO bug that I currently stuck with
> > >> when using .ACCESS_WITH_SIZE in bound sanitizer (only with -flto, without -flto, no issue):
> > >>
> > >> [opc@qinzhao-aarch64-ol8 gcc]$ sh t
> > >> during IPA pass: modref
> > >> t.c:20:1: internal compiler error: tree code ‘ssa_name’ is not supported in LTO streams
> > >> 0x14c3993 lto_write_tree
> > >>        ../../latest-gcc-write/gcc/lto-streamer-out.cc:561
> > >> 0x14c3aeb lto_output_tree_1
> > >>
> > >> And the value of the tree node that triggered the ICE is:
> > >> (gdb) call debug_tree(expr)
> > >> <ssa_name 0xfffff5761e60 type <error_mark 0xfffff56c0e58>
> > >>    nothrow
> > >>    def_stmt
> > >>    version:13 in-free-list>
> > >>
> > >> Is there any good way to debug LTO bug?
> > >
> > > This happens usually when you have a VLA type and its type fields are not
> > > properly gimplified which usually happens because the frontend fails to
> > > insert a gimplification point for it (a DECL_EXPR).
> >
> > I found an old gcc bug
> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97172
> > ICE: tree code ‘ssa_name’ is not supported in LTO streams since r11-3303-g6450f07388f9fe57
> >
> > Which is very similar to the bug I am having right now.
> >
> > After further study, I suspect that the issue I am having right now with the LTO streaming also
> > relate to “unshare_expr”, “save_expr”, and the combination of these two, I suspect that
> > the current gcc cannot handle the combination of these two correctly for my case.
> >
> > My testing case is:
> >
> > #include <stdlib.h>
> > void __attribute__((__noinline__)) setup_and_test_vla (int n1, int n2, int m)
> > {
> >    struct foo {
> >        int n;
> >        int p[][n2][n1] __attribute__((counted_by(n)));
> >    } *f;
> >
> >    f = (struct foo *) malloc (sizeof(struct foo) + m*sizeof(int[n2][n1]));
> >    f->n = m;
> >    f->p[m][n2][n1]=1;
> >    return;
> > }
> >
> > int main(int argc, char *argv[])
> > {
> >   setup_and_test_vla (10, 11, 20);
> >   return 0;
> > }
> >
> > Failed with
> > my_gcc -Os -fsanitize=bounds -flto
> >
> > If changing either n1 or n2 to a constant, the testing passed.
> > If deleting -flto, the testing passed too.
> >
> > I double checked my code per the suggestions provided by you and Jakub in this
> > email thread, and I think the code should be fine.
> >
> > The code is following:
> >
> > =====
> > 504 /* Instrument array bounds for INDIRECT_REFs whose pointers are
> > 505    POINTER_PLUS_EXPRs of calls to .ACCESS_WITH_SIZE. We create special
> > 506    builtins that gets expanded in the sanopt pass, and make an array
> > 507    dimension of it.  ARRAY is the pointer to the base of the array,
> > 508    which is a call to .ACCESS_WITH_SIZE, *OFFSET is the offset to the
> > 509    beginning of array.
> > 510    Return NULL_TREE if no instrumentation is emitted.  */
> > 511
> > 512 tree
> > 513 ubsan_instrument_bounds_indirect_ref (location_t loc, tree array, tree *offset)
> > 514 {
> > 515   if (!is_access_with_size_p (array))
> > 516     return NULL_TREE;
> > 517   tree bound = get_bound_from_access_with_size (array);
> > 518   /* The type of the call to .ACCESS_WITH_SIZE is a pointer type to
> > 519      the element of the array.  */
> > 520   tree element_size = TYPE_SIZE_UNIT (TREE_TYPE (TREE_TYPE (array)));
> > 521   gcc_assert (bound);
> > 522
> > 523   /* Given the offset, and the size of each element, the index can be
> > 524      computed as: offset/element_size.  */
> > 525   *offset = save_expr (*offset);
> > 526   tree index = fold_build2 (EXACT_DIV_EXPR,
> > 527                            sizetype, *offset,
> > 528                            unshare_expr (element_size));
> > 529   /* Create a "(T *) 0" tree node to describe the original array type.
> > 530      We get the original array type from the first argument of the call to
> > 531      .ACCESS_WITH_SIZE (REF, COUNTED_BY_REF, 1, num_bytes, -1).
> > 532
> > 533      Originally, REF is a COMPONENT_REF with the original array type,
> > 534      it was converted to a pointer to an ADDR_EXPR, and the ADDR_EXPR's
> > 535      first operand is the original COMPONENT_REF.  */
> > 536   tree ref = CALL_EXPR_ARG (array, 0);
> > 537   tree array_type
> > 538     = unshare_expr (TREE_TYPE (TREE_OPERAND (TREE_OPERAND(ref, 0), 0)));
> > 539   tree zero_with_type = build_int_cst (build_pointer_type (array_type), 0);
> > 540   return build_call_expr_internal_loc (loc, IFN_UBSAN_BOUNDS,
> > 541                                        void_type_node, 3, zero_with_type,
> > 542                                        index, bound);
> > 543 }
> >
> > =====
> >
> > Inside gdb, the guilty IR failed in LTO streaming is from the above line 520:
> > TYPE_SIZE_UNIT (TREE_TYPE (TREE_TYPE (array))),
> >
> > When I use this tree node as an operand of the expression at line 526, I added
> > unshare_expr.
> >
> > However, I still see the guilty IR as in gdb:
> >
> >             unit-size <mult_expr 0xfffff5aabf90 type <integer_type 0xfffff57c0000 sizetype>
> >                 side-effects
> >                 arg:0 <mult_expr 0xfffff5aabf68 type <integer_type 0xfffff57c0000 sizetype>
> >
> >                     arg:0 <ssa_name 0xfffff5761e18 type <error_mark 0xfffff56c0e58>
> >                         nothrow
> >                         def_stmt
> >                         version:12 in-free-list>
> >                     arg:1 <ssa_name 0xfffff5761e60 type <error_mark 0xfffff56c0e58>
> >                         nothrow
> >                         def_stmt
> >                         version:13 in-free-list>>
> >                 arg:1 <integer_cst 0xfffff56c10c8 constant 4>>
> >
> >
> > I have been stuck with this bug for quite some time.
> > Any help is helpful.
>
> The above hasn't been gimplified correctly, you'd instead see
> a D.1234 in there, not an expression with SSA names.  That happens
> when the frontend fails to emit a DECL_EXPR for a decl with this
> type.

.. which then also results in missing unsharing of this expression
(so the SSA names leak in)

>
> >
> > Qing
> >
> > >
> > >> Thanks a lot for the help.
> > >>
> > >> Qing
> > >>
> >