> On Jan 17, 2024, at 1:43 AM, Richard Biener <richard.guent...@gmail.com> > wrote: > > On Wed, Jan 17, 2024 at 7:42 AM Richard Biener > <richard.guent...@gmail.com> wrote: >> >> On Tue, Jan 16, 2024 at 9:26 PM Qing Zhao <qing.z...@oracle.com> wrote: >>> >>> >>> >>>> On Jan 15, 2024, at 4:31 AM, Richard Biener <richard.guent...@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)
Thanks a lot for the hints. One correction first, the LTO bug is not related to -fsanitize=bounds. Deleting -fsanitize=bounds still can repeat the failure. After further debugging into the gimplification phase related with the SAVE_EXPR, I finally locate the place where the unshareing of the expression is missing. This is in the routine “pointer_int_sum” of c-family/c-common.cc: 3330 { 3331 if (!complain && !COMPLETE_TYPE_P (TREE_TYPE (result_type))) 3332 return error_mark_node; 3333 size_exp = size_in_bytes_loc (loc, TREE_TYPE (result_type)); 3334 /* Wrap the pointer expression in a SAVE_EXPR to make sure it 3335 is evaluated first when the size expression may depend 3336 on it for VM types. */ 3337 if (TREE_SIDE_EFFECTS (size_exp) 3338 && TREE_SIDE_EFFECTS (ptrop) 3339 && variably_modified_type_p (TREE_TYPE (ptrop), NULL)) 3340 { 3341 ptrop = save_expr (ptrop); 3342 size_exp = build2 (COMPOUND_EXPR, TREE_TYPE (intop), ptrop, size_exp); 3343 } 3344 } In the above, at line 3333, the tree node, TYPE_SIZE_UNIT (TREE_TYPE(result_type)), is returned directly as the size_exp, (gdb) call debug_tree(size_exp) <mult_expr 0xfffff5a6f910 type <integer_type 0xfffff57c0000 sizetype public unsigned DI size <integer_cst 0xfffff56c0e70 constant 64> unit-size <integer_cst 0xfffff56c0e88 constant 8> align:64 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type 0xfffff57c0000 precision:64 min <integer_cst 0xfffff56c0ea0 0> max <integer_cst 0xfffff56d05e0 18446744073709551615>> side-effects arg:0 <mult_expr 0xfffff5a6f8e8 type <integer_type 0xfffff57c0000 sizetype> side-effects arg:0 <nop_expr 0xfffff56dc540 type <integer_type 0xfffff57c0000 sizetype> side-effects arg:0 <save_expr 0xfffff56dc4c0 type <integer_type 0xfffff57c05e8 int> side-effects arg:0 <parm_decl 0xfffff76b6f80 n1>>> arg:1 <nop_expr 0xfffff56dc600 type <integer_type 0xfffff57c0000 sizetype> side-effects arg:0 <save_expr 0xfffff56dc580 type <integer_type 0xfffff57c05e8 int> side-effects arg:0 <parm_decl 0xfffff76b7000 n2>>>> arg:1 <integer_cst 0xfffff56c10c8 type <integer_type 0xfffff57c0000 sizetype> constant 4>> Without unshare_expr to this size_exp, the above TYPE_SIZE_UNIT node containing SAVE_EXPRs is gimpflified to expressions with SSA_NAME during gimplification. (This is unaccepted by LTO). Adding an unshare_expr (size_exp) resolved this problem. Although I still think that there might be potential issue with the gimpflication of SAVE_EXPRs, I dare not to modify that part of the code. At this moment, I will add unshare_expr to the routine “pointer_int_sum” to workaround this issue. Let me know if you have any comment and suggestion. Thanks a lot. Qing >> >>> >>> Qing >>> >>>> >>>>> Thanks a lot for the help. >>>>> >>>>> Qing