* [committed][low-mem] Fix DW_OP_GNU_implicit_pointer handling
2019-01-01 0:00 [PATCH][low-mem] Fix DW_OP_GNU_implicit_pointer handling Tom de Vries
@ 2019-01-01 0:00 ` Tom de Vries
0 siblings, 0 replies; 2+ messages in thread
From: Tom de Vries @ 2019-01-01 0:00 UTC (permalink / raw)
To: dwz, jakub
[-- Attachment #1: Type: text/plain, Size: 1163 bytes --]
On 13-02-19 18:16, Tom de Vries wrote:
> Hi,
>
> If dwz is processing a file f containing a DIE d in compilation unit cu_d,
> with DW_AT_location pointing to a loclist with one of the locexprs containing
> a DW_OP_GNU_implicit_pointer reference to a DIE not in cu_d (which can be the
> case in dwz output), and dwz is processing in low-mem mode, then we may see
> this message:
> ...
> dwz: f: Couldn't find DIE referenced by DW_OP_GNU_implicit_pointer
> ...
>
> The message is emitted in read_exprloc while we're calculating a checksum for
> the DW_OP_GNU_implicit_pointer reference. The reason we can't find the DIE
> it's referencing, is that in low-mem mode, we can only count on the current
> compilation unit and DW_FORM_ref_addr references to be available.
>
> Fix this by silencing the message in low-mem mode.
>
After understanding that low-mem mode is a pure run-time vs memory usage
trade-off and should not affect dwz output, I realized that the fix
described above is incorrect, given that it still would error out in
low-mem mode where regular mode would give a result.
Fixed as described in rationale, and committed as below.
Thanks,
- Tom
[-- Attachment #2: 0003-low-mem-Fix-DW_OP_GNU_implicit_pointer-handling.patch --]
[-- Type: text/x-patch, Size: 11815 bytes --]
[low-mem] Fix DW_OP_GNU_implicit_pointer handling
When twice dwz-ing a cc1 build with bootstrap-lto in low-mem mode, we run
into this error:
...
$ dwz -l0 cc1
$ dwz -l0 cc1
dwz: cc1: Couldn't find DIE referenced by DW_OP_GNU_implicit_pointer
...
The error is triggered as follows. There's a DIE 0x369186:
...
<12><369186>: Abbrev Number: 3 (DW_TAG_formal_parameter)
<369187> DW_AT_abstract_origin: <0xecc97>
<36918b> DW_AT_location : 0x4ee259 (location list)
<36918f> DW_AT_GNU_locviews: 0x4ee255
...
referring to location list 0x4ee259:
...
004ee255 v000000000000004 v000000000000001 location view pair
004ee257 v000000000000000 v000000000000000 location view pair
004ee259 v000000000000004 v000000000000001 views at 004ee255 for:
0000000000876020 000000000087602b (DW_OP_GNU_implicit_pointer: \
<0x800> 0)
004ee271 v000000000000000 v000000000000000 views at 004ee257 for:
0000000000587ca8 0000000000587cb0 (DW_OP_GNU_implicit_pointer: \
<0x800> 0)
004ee289 <End of list>
...
which refer to DIE 0x800 in a partial unit 0x7f7:
...
<0><7f7>: Abbrev Number: 5 (DW_TAG_partial_unit)
<7f8> DW_AT_stmt_list : 0xd1db8
<7fc> DW_AT_comp_dir : (indirect string, offset: 0x2183): \
/home/mliska/Programming/gcc/objdir/gcc
<1><800>: Abbrev Number: 112 (DW_TAG_dwarf_procedure)
<801> DW_AT_location : 21 byte block: 9e 13 66 69 72 73 74 5f 69 6d \
6d 5f 75 73 65 5f 73 74 6d 74 0 \
(DW_OP_implicit_value 19 byte block: 66 69 72 \
73 74 5f 69 6d 6d 5f 75 73 65 5f 73 74 6d 74 0 )
...
When calculating the checksum for DIE 0x369186 in checksum_die, we call
read_loclist for loc_list 0x4ee259 and read_exprloc for
DW_OP_GNU_implicit_pointer 0x800 0, after which we run into the error for
addr == 0x800 here:
...
ref = off_htab_lookup (NULL, addr);
if (ref == NULL)
{
error (0, 0, "%s: Couldn't find DIE referenced by %s",
dso->filename, get_DW_OP_str (op));
return 1;
}
...
In read_debug_info in regular mode, the .debug_info section is processed one
CU at a time. Each CU is parsed, after which the checksums and duplicates are
calculated, but the parsed CU remains in memory when processing the next CU.
So, when parsing the partial unit containing the DIE at 0x800, the DIE's added
to off_htab, and when reading the CU containing DIE 0x369186, calculating the
checksum for DIE 0x369186 in checksum_die and trying to find the DIE at 0x800
in read_exprloc, we succeed because DIE 0x800 is still present in off_htab.
In read_debug_info in low-mem mode, also the .debug_info section is processed
one CU at a time. Each CU is parsed, after which the checksums and duplicates
are calculated, but after that the amount of parsed CU that remains in memory
is minimized by remove_unneeded and collapse_children. The inter-cu
referenced DIEs however need to be kept in memory, because they may be
referenced from CUs yet to process. But there's no way of knowing which DIEs
are inter-cu referenced, before all CUs have been parsed. To solve this
problem, in low-mem mode read_debug_info is split up into two phases which
each read the .debug_info section once. In phase 1, for each DIE that is
inter-cu referenced, a dummy die is allocated, which is merged during phase 2
with the actual DIE when it's being parsed. This way in phase 2 we can always
know if a DIE is inter-cu referenced, and consequently keep it in memory.
However, in phase 1 we currently only handle inter-cu refs of one type:
DW_FORM_ref_addr refs from DIE attributes, and consequently no dummy DIE is
generated during phase 1 for DIE 0x800, so it's not kept during phase 2, which
leads to the error.
Fix this by also creating dummy DIEs in phase 1 for inter-cu references in
expr_locs. [ And due to the fact that we're not keeping track of the CUs of
dummy DIEs, we cannot detect when references which could be inter-cu, are in
fact intra-cu, so we also add dummy DIEs for those. ]
2019-04-29 Tom de Vries <tdevries@suse.de>
PR dwz/24204
* dwz.c (read_exprloc_low_mem_phase1, read_loclist_low_mem_phase1)
(add_locexpr_dummy_dies): New function.
(read_debug_info): Call add_locexpr_dummy_dies.
---
dwz.c | 297 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 297 insertions(+)
diff --git a/dwz.c b/dwz.c
index 758f604..aa28a7f 100644
--- a/dwz.c
+++ b/dwz.c
@@ -1722,6 +1722,297 @@ add_dummy_die (dw_cu_ref cu, unsigned int offset)
*slot = (void *) ref;
}
+/* Add dummy DIEs for expr_loc at PTR. */
+static int
+read_exprloc_low_mem_phase1 (DSO *dso, dw_die_ref die, unsigned char *ptr,
+ size_t len)
+{
+ unsigned char *end = ptr + len;
+ unsigned char op;
+ GElf_Addr addr;
+ dw_cu_ref cu;
+
+ while (ptr < end)
+ {
+ op = *ptr++;
+ switch (op)
+ {
+ case DW_OP_addr:
+ ptr += ptr_size;
+ break;
+ case DW_OP_deref:
+ case DW_OP_dup:
+ case DW_OP_drop:
+ case DW_OP_over:
+ case DW_OP_swap:
+ case DW_OP_rot:
+ case DW_OP_xderef:
+ case DW_OP_abs:
+ case DW_OP_and:
+ case DW_OP_div:
+ case DW_OP_minus:
+ case DW_OP_mod:
+ case DW_OP_mul:
+ case DW_OP_neg:
+ case DW_OP_not:
+ case DW_OP_or:
+ case DW_OP_plus:
+ case DW_OP_shl:
+ case DW_OP_shr:
+ case DW_OP_shra:
+ case DW_OP_xor:
+ case DW_OP_eq:
+ case DW_OP_ge:
+ case DW_OP_gt:
+ case DW_OP_le:
+ case DW_OP_lt:
+ case DW_OP_ne:
+ case DW_OP_lit0 ... DW_OP_lit31:
+ case DW_OP_reg0 ... DW_OP_reg31:
+ case DW_OP_nop:
+ case DW_OP_push_object_address:
+ case DW_OP_form_tls_address:
+ case DW_OP_call_frame_cfa:
+ case DW_OP_stack_value:
+ case DW_OP_GNU_push_tls_address:
+ case DW_OP_GNU_uninit:
+ break;
+ case DW_OP_const1u:
+ case DW_OP_pick:
+ case DW_OP_deref_size:
+ case DW_OP_xderef_size:
+ case DW_OP_const1s:
+ ++ptr;
+ break;
+ case DW_OP_const2u:
+ case DW_OP_const2s:
+ case DW_OP_skip:
+ case DW_OP_bra:
+ ptr += 2;
+ break;
+ case DW_OP_call2:
+ case DW_OP_call4:
+ case DW_OP_GNU_parameter_ref:
+ if (op == DW_OP_call2)
+ read_16 (ptr);
+ else
+ read_32 (ptr);
+ break;
+ case DW_OP_const4u:
+ case DW_OP_const4s:
+ ptr += 4;
+ break;
+ case DW_OP_call_ref:
+ case DW_OP_GNU_implicit_pointer:
+ case DW_OP_GNU_variable_value:
+ cu = die_cu (die);
+ addr = read_size (ptr, cu->cu_version == 2 ? ptr_size : 4);
+ if (cu->cu_version == 2)
+ ptr += ptr_size;
+ else
+ ptr += 4;
+ /* Adding a dummy DIE ref to mark an intercu reference is only
+ necessary if die_cu (ref) != cu, but we don't track cu's during
+ low-mem phase1. */
+ add_dummy_die (cu, addr);
+ if (op == DW_OP_GNU_implicit_pointer)
+ read_uleb128 (ptr);
+ break;
+ case DW_OP_const8u:
+ case DW_OP_const8s:
+ ptr += 8;
+ break;
+ case DW_OP_constu:
+ case DW_OP_plus_uconst:
+ case DW_OP_regx:
+ case DW_OP_piece:
+ case DW_OP_consts:
+ case DW_OP_breg0 ... DW_OP_breg31:
+ case DW_OP_fbreg:
+ read_uleb128 (ptr);
+ break;
+ case DW_OP_bregx:
+ case DW_OP_bit_piece:
+ read_uleb128 (ptr);
+ read_uleb128 (ptr);
+ break;
+ case DW_OP_implicit_value:
+ {
+ uint32_t leni = read_uleb128 (ptr);
+ ptr += leni;
+ }
+ break;
+ case DW_OP_GNU_entry_value:
+ {
+ uint32_t leni = read_uleb128 (ptr);
+ if ((uint64_t) (end - ptr) < leni)
+ {
+ error (0, 0, "%s: DWARF DW_OP_GNU_entry_value with too large"
+ " length", dso->filename);
+ return 1;
+ }
+ if (read_exprloc_low_mem_phase1 (dso, die, ptr, leni))
+ return 1;
+ ptr += leni;
+ }
+ break;
+ case DW_OP_GNU_convert:
+ case DW_OP_GNU_reinterpret:
+ read_uleb128 (ptr);
+ break;
+ case DW_OP_GNU_regval_type:
+ read_uleb128 (ptr);
+ read_uleb128 (ptr);
+ break;
+ case DW_OP_GNU_const_type:
+ read_uleb128 (ptr);
+ ptr += *ptr + 1;
+ break;
+ case DW_OP_GNU_deref_type:
+ ++ptr;
+ read_uleb128 (ptr);
+ break;
+ default:
+ error (0, 0, "%s: Unknown DWARF %s",
+ dso->filename, get_DW_OP_str (op));
+ return 1;
+ }
+ }
+
+ return 0;
+}
+
+/* Add dummy DIEs for loclist at OFFSET. */
+static int
+read_loclist_low_mem_phase1 (DSO *dso, dw_die_ref die, GElf_Addr offset)
+{
+ unsigned char *ptr, *endsec;
+ GElf_Addr low, high;
+ size_t len;
+
+ ptr = debug_sections[DEBUG_LOC].data;
+ if (ptr == NULL)
+ {
+ error (0, 0, "%s: loclistptr attribute, yet no .debug_loc section",
+ dso->filename);
+ return 1;
+ }
+ if (offset >= debug_sections[DEBUG_LOC].size)
+ {
+ error (0, 0,
+ "%s: loclistptr offset %Ld outside of .debug_loc section",
+ dso->filename, (long long) offset);
+ return 1;
+ }
+ endsec = ptr + debug_sections[DEBUG_LOC].size;
+ ptr += offset;
+ while (ptr < endsec)
+ {
+ low = read_size (ptr, ptr_size);
+ high = read_size (ptr + ptr_size, ptr_size);
+ ptr += 2 * ptr_size;
+ if (low == 0 && high == 0)
+ break;
+
+ if (low == ~ (GElf_Addr) 0 || (ptr_size == 4 && low == 0xffffffff))
+ continue;
+
+ len = read_16 (ptr);
+ if (unlikely (!(ptr + len <= endsec)))
+ {
+ error (0, 0,
+ "%s: locexpr length 0x%Lx exceeds .debug_loc section",
+ dso->filename, (long long) len);
+ return 1;
+ }
+
+ if (read_exprloc_low_mem_phase1 (dso, die, ptr, len))
+ return 1;
+
+ ptr += len;
+ }
+
+ return 0;
+}
+
+/* Add dummy dies for loc_exprs and loc_lists referenced from DIE. */
+static int
+add_locexpr_dummy_dies (DSO *dso, dw_cu_ref cu, dw_die_ref die,
+ unsigned char *ptr, uint32_t form, unsigned int attr,
+ size_t len)
+{
+ if (form == DW_FORM_block1)
+ {
+ switch (attr)
+ {
+ case DW_AT_frame_base:
+ case DW_AT_location:
+ case DW_AT_data_member_location:
+ case DW_AT_vtable_elem_location:
+ case DW_AT_byte_size:
+ case DW_AT_bit_offset:
+ case DW_AT_bit_size:
+ case DW_AT_string_length:
+ case DW_AT_lower_bound:
+ case DW_AT_return_addr:
+ case DW_AT_bit_stride:
+ case DW_AT_upper_bound:
+ case DW_AT_count:
+ case DW_AT_segment:
+ case DW_AT_static_link:
+ case DW_AT_use_location:
+ case DW_AT_allocated:
+ case DW_AT_associated:
+ case DW_AT_data_location:
+ case DW_AT_byte_stride:
+ case DW_AT_GNU_call_site_value:
+ case DW_AT_GNU_call_site_data_value:
+ case DW_AT_GNU_call_site_target:
+ case DW_AT_GNU_call_site_target_clobbered:
+ if (read_exprloc_low_mem_phase1 (dso, die, ptr, len))
+ return 1;
+ default:
+ break;
+ }
+
+ return 0;
+ }
+
+ if (form == DW_FORM_exprloc)
+ return read_exprloc_low_mem_phase1 (dso, die, ptr, len);
+
+ switch (attr)
+ {
+ case DW_AT_location:
+ case DW_AT_string_length:
+ case DW_AT_return_addr:
+ case DW_AT_data_member_location:
+ case DW_AT_frame_base:
+ case DW_AT_segment:
+ case DW_AT_static_link:
+ case DW_AT_use_location:
+ case DW_AT_vtable_elem_location:
+ if ((cu->cu_version < 4 && form == DW_FORM_data4)
+ || form == DW_FORM_sec_offset)
+ {
+ if (read_loclist_low_mem_phase1 (dso, die, do_read_32 (ptr)))
+ return 1;
+ break;
+ }
+ else if (cu->cu_version < 4 && form == DW_FORM_data8)
+ {
+ if (read_loclist_low_mem_phase1 (dso, die, do_read_64 (ptr)))
+ return 1;
+ break;
+ }
+ break;
+ default:
+ break;
+ }
+
+ return 0;
+}
+
/* Structure recording a portion of .debug_loc section that will need
adjusting. */
struct debug_loc_adjust
@@ -4819,6 +5110,12 @@ read_debug_info (DSO *dso, int kind)
goto fail;
}
}
+
+ if (unlikely (low_mem_phase1)
+ && add_locexpr_dummy_dies (dso, cu, die, ptr, form,
+ t->attr[i].attr, len))
+ goto fail;
+
switch (form)
{
case DW_FORM_ref_addr:
^ permalink raw reply [flat|nested] 2+ messages in thread