Nowadays, we set computed value's frame id, which is a misuse to me.
The computed value itself doesn't care about frame id, but function
value_computed_funcs (val)->read (or read_pieced_value) cares about
which frame the register is relative to, so 'struct piece_closure' is
a better place to fit frame id.
This patch adds a frame id in 'struct piece_closure', and use it
instead of using computed value's frame id.
gdb:
2016-11-28 Yao Qi <yao.qi@linaro.org>
* dwarf2loc.c (struct piece_closure) <frame_id>: New field.
(allocate_piece_closure): Add new parameter 'frame' and set
closure's frame_id field accordingly.
(read_pieced_value): Get frame from closure instead of value.
(dwarf2_evaluate_loc_desc_full): Remove code getting frame id.
Don't set value's frame id.
The function copy_bitwise used for copying DWARF pieces can potentially
be invoked for large chunks of data. For instance, consider a large
struct one of whose members is currently located in a register. In this
case copy_bitwise would still copy the data bitwise in a loop, which is
much slower than necessary.
This change uses memcpy for the large part instead, if possible.
gdb/ChangeLog:
* dwarf2loc.c (copy_bitwise): Use memcpy for the middle part, if
it is byte-aligned.
This adds a unit test for the copy_bitwise function in dwarf2loc.c.
With the old (broken) version of copy_bitwise this test would generate
the following failure message:
(gdb) maintenance selftest
Self test failed: copy_bitwise 11000000 != 10000000 (7+2 -> 0)
gdb/ChangeLog:
2016-11-24 Andreas Arnez <arnez@linux.vnet.ibm.com>
Pedro Alves <palves@redhat.com>
* dwarf2loc.c (bits_to_str, check_copy_bitwise)
(copy_bitwise_tests): New functions.
(_initialize_dwarf2loc): Register the new function
copy_bitwise_tests as a unit test.
* selftest.c (run_self_tests): Improve the failure message's
wording and formatting.
When the user writes or reads a variable whose location is described
with DWARF pieces (DW_OP_piece or DW_OP_bit_piece), GDB's helper
function copy_bitwise is invoked for each piece. The implementation of
this function has a bug that may result in a corrupted copy, depending
on alignment and bit size. (Full-byte copies are not affected.)
This rewrites copy_bitwise, replacing its algorithm by a fixed version,
and adding an appropriate test case. Without the fix the new test case
fails, e.g.:
print def_t
$2 = {a = 0, b = 4177919}
(gdb) FAIL: gdb.dwarf2/nonvar-access.exp: print def_t
Written in binary, the wrong result above looks like this:
01111111011111111111111
Which means that two zero bits have sneaked into the copy of the
original all-one bit pattern. The test uses this simple all-one value
in order to avoid another GDB bug that causes the DWARF piece of a
DW_OP_stack_value to be taken from the wrong end on big-endian
architectures.
gdb/ChangeLog:
* dwarf2loc.c (extract_bits_primitive): Remove.
(extract_bits): Remove.
(copy_bitwise): Rewrite. Fixes a possible corruption that may
occur for non-byte-aligned copies.
gdb/testsuite/ChangeLog:
* gdb.dwarf2/nonvar-access.exp: Add a test for accessing
non-byte-aligned bit fields.
The VALUE_FRAME_ID macro provides access to a member in struct value
that's used to hold the frame id that's used when determining a
register's value or when assigning to a register. The underlying
member has a long and obscure name. I won't refer to it here, but
will simply refer to VALUE_FRAME_ID as if it's the struct value member
instead of being a convenient macro.
At the moment, without this patch in place, VALUE_FRAME_ID is set in
value_of_register_lazy() and several other locations to hold the frame
id of the frame passed to those functions.
VALUE_FRAME_ID is used in the lval_register case of
value_fetch_lazy(). To fetch the register's value, it calls
get_frame_register_value() which, in turn, calls
frame_unwind_register_value() with frame->next.
A python based unwinder may wish to determine the value of a register
or evaluate an expression containing a register. When it does this,
value_fetch_lazy() will be called under some circumstances. It will
attempt to determine the frame id associated with the frame passed to
it. In so doing, it will end up back in the frame sniffer of the very
same python unwinder that's attempting to learn the value of a
register as part of the sniffing operation. This recursion is not
desirable.
As noted above, when value_fetch_lazy() wants to fetch a register's
value, it does so (indirectly) by unwinding from frame->next.
With this in mind, a solution suggests itself: Change VALUE_FRAME_ID
to hold the frame id associated with the next frame. Then, when it
comes time to obtain the value associated with the register, we can
simply unwind from the frame corresponding to the frame id stored in
VALUE_FRAME_ID. This neatly avoids the python unwinder recursion
problem by changing when the "next" operation occurs. Instead of the
"next" operation occuring when the register value is fetched, it
occurs earlier on when assigning a frame id to VALUE_FRAME_ID.
(Thanks to Pedro for this suggestion.)
This patch implements this idea.
It builds on the patch "Distinguish sentinel frame from null frame".
Without that work in place, it's necessary to check for null_id at
several places and then obtain the sentinel frame.
It also renames most occurences of VALUE_FRAME_ID to
VALUE_NEXT_FRAME_ID to reflect the new meaning of this field.
There are several uses of VALUE_FRAME_ID which were not changed. In
each case, the original meaning of VALUE_FRAME_ID is required to get
correct results. In all but one of these uses, either
put_frame_register_bytes() or get_frame_register_bytes() is being
called with the frame value obtained from VALUE_FRAME_ID. Both of
these functions perform some unwinding by performing a "->next"
operation on the frame passed to it. If we were to use the new
VALUE_NEXT_FRAME_ID macro, this would effectively do two "->next"
operations, which is not what we want.
The VALUE_FRAME_ID macro has been redefined in terms of
VALUE_NEXT_FRAME_ID. It simply fetches the previous frame's id,
providing this id as the value of the macro.
gdb/ChangeLog:
* value.h (VALUE_FRAME_ID): Rename to VALUE_NEXT_FRAME_ID. Update
comment. Create new VALUE_FRAME_ID which is defined in terms of
VALUE_NEXT_FRAME_ID.
(deprecated_value_frame_id_hack): Rename to
deprecated_value_next_frame_id_hack.
* dwarf2loc.c, findvar.c, frame-unwind.c, sentinel-frame.c,
valarith.c, valops.c, value.c: Adjust nearly all occurences of
VALUE_FRAME_ID to VALUE_NEXT_FRAME_ID. Add comments for those
which did not change.
* value.c (struct value): Rename frame_id field to next_frame_id.
Update comment.
(deprecated_value_frame_id_hack): Rename to
deprecated_value_next_frame_id_hack.
(value_fetch_lazy): Call frame_unwind_register_value()
instead of get_frame_register_value().
* frame.c (get_prev_frame_id_by_id): New function.
* frame.h (get_prev_frame_id_by_id): Declare.
* dwarf2loc.c (dwarf2_evaluate_loc_desc_full): Make
VALUE_NEXT_FRAME_ID refer to the next frame.
* findvar.c (value_of_register_lazy): Likewise.
(default_value_from_register): Likewise.
(value_from_register): Likewise.
* frame_unwind.c (frame_unwind_got_optimized): Likewise.
* sentinel-frame.c (sentinel_frame_prev_register): Likewise.
* value.h (VALUE_FRAME_ID): Update comment describing this macro.
This fixes some regressions found in the patch to convert
dwarf_expr_context to use methods. Specifically:
* get_base_type could erroneously throw; this was rewritten to move
the size checks into the only spot needing them.
* Previously the "symbol needs frame" implementation reused th
"cfa" function for the get_frame_pc slot; this reimplements
it under the correct name.
* Not enough members were saved and restored in one implementation
of push_dwarf_reg_entry_value; this patch fixes this oversight
and also takes the opportunity to remove an extraneous structure
definition.
2016-11-02 Tom Tromey <tom@tromey.com>
* dwarf2loc.c (dwarf_evaluate_loc_desc::get_base_type): Rename
from impl_get_base_type. Rewrite.
(struct dwarf_expr_baton): Remove.
(dwarf_evaluate_loc_desc::push_dwarf_reg_entry_value): Save and
restore more fields.
(symbol_needs_eval_context::get_frame_pc): New method.
* dwarf2expr.h (dwarf_expr_context::get_base_type): Now public,
virtual.
(dwarf_expr_context::impl_get_base_type): Remove.
* dwarf2expr.c (dwarf_expr_context::get_base_type): Remove.
This patch converts the function pointers in dwarf_expr_context_funcs
into methods on dwarf_expr_context, and then updates the various
implementations and callers to follow.
NB this patch uses "override" (which caught a couple of renaming bugs
during development) -- but this is C++11, so this patch at least has
to wait for Pedro's patch that adds the OVERRIDE macro.
After this patch it would be possible to do one more, that makes
various members of dwarf_expr_context "protected"; but I haven't done
this.
2016-10-21 Tom Tromey <tom@tromey.com>
* dwarf2loc.c (struct dwarf_expr_context_funcs): Don't declare.
(dwarf_expr_read_addr_from_reg, dwarf_expr_get_reg_value)
(dwarf_expr_read_mem, dwarf_expr_frame_base): Rename; turn into
methods.
(get_frame_pc_for_per_cu_dwarf_call): New function.
(dwarf_expr_frame_cfa, dwarf_expr_frame_pc)
(dwarf_expr_tls_address): Rename; turn into methods.
(per_cu_dwarf_call): Remove arguments. Use
get_frame_pc_for_per_cu_dwarf_call.
(dwarf_evaluate_loc_desc): New class.
(dwarf_expr_dwarf_call, dwarf_expr_context)
(dwarf_expr_push_dwarf_reg_entry_value)
(dwarf_expr_get_addr_index, dwarf_expr_get_obj_addr): Rename; turn
into methods.
(dwarf_expr_ctx_funcs): Remove.
(dwarf2_evaluate_loc_desc_full): Update.
(dwarf2_locexpr_baton_eval): Update.
(symbol_needs_eval_context): New class.
(symbol_needs_read_addr_from_reg, symbol_needs_get_reg_value)
(symbol_needs_read_mem, symbol_needs_frame_base)
(symbol_needs_frame_cfa, symbol_needs_tls_address)
(symbol_needs_dwarf_call, needs_dwarf_reg_entry_value): Rename;
turn into methods.
(needs_get_addr_index, needs_get_obj_addr): Remove; turn into
methods.
(symbol_needs_ctx_funcs): Remove.
(dwarf2_loc_desc_get_symbol_read_needs): Update.
* dwarf2expr.h (struct dwarf_expr_context_funcs): Remove; turn
contents into methods.
(struct dwarf_expr_context) <baton, funcs>: Remove.
<read_addr_from_reg, get_reg_value, read_mem, get_frame_base,
get_frame_cfa, get_frame_pc, get_tls_address, dwarf_call,
impl_get_base_type, push_dwarf_block_entry_value, get_addr_index,
get_object_address>: Declare new methods.
(ctx_no_get_frame_base, ctx_no_get_frame_cfa)
(ctx_no_get_frame_pc, ctx_no_get_tls_address, ctx_no_dwarf_call)
(ctx_no_get_base_type, ctx_no_push_dwarf_reg_entry_value)
(ctx_no_get_addr_index): Don't declare.
* dwarf2expr.c (get_base_type): Use impl_get_base_type.
(execute_stack_op): Update.
(ctx_no_get_frame_base, ctx_no_get_frame_cfa)
(ctx_no_get_frame_pc, ctx_no_get_tls_address, ctx_no_dwarf_call)
(ctx_no_get_base_type, ctx_no_push_dwarf_reg_entry_value)
(ctx_no_get_addr_index): Remove; now methods on
dwarf_expr_context.
* dwarf2-frame.c (read_addr_from_reg): Take a frame_info, not a
baton.
(class dwarf_expr_executor): New class.
(get_reg_value, read_mem): Rename, turn into methods.
(execute_stack_op): Use dwarf_expr_executor.
This converts various DWARF expr functions to be members on
dwarf_expr_context, then fixes up the various users. This results in
somewhat less wordy code and sets the stage for the next patch.
2016-10-21 Tom Tromey <tom@tromey.com>
* dwarf2loc.c (per_cu_dwarf_call)
(dwarf_expr_push_dwarf_reg_entry_value)
(dwarf2_evaluate_loc_desc_full, dwarf2_locexpr_baton_eval)
(needs_dwarf_reg_entry_value)
(dwarf2_loc_desc_get_symbol_read_needs): Update.
* dwarf2expr.h (dwarf_expr_context) <push_address, eval, fetch,
fetch_address, fetch_in_stack_memory, address_type, grow_stack,
push, stack_empty_p, add_piece, get_base_type, execute_stack_op,
pop>: New method declarations.
(dwarf_expr_push_address, dwarf_expr_eval, dwarf_expr_fetch)
(dwarf_expr_fetch_address, dwarf_expr_fetch_in_stack_memory):
Don't declare.
* dwarf2expr.c (address_type, grow_stack, push, push_address)
(pop, fetch, fetch_address, fetch_in_stack_memory)
(stack_empty_p, add_piece, eval, get_base_type)
(execute_stack_op): Rename. Turn into methods.
* dwarf2-frame.c (execute_stack_op): Update.
This is the first step in the conversion of dwarf_expr_ctx to a C++
class. This conversion is done in steps to make the patches, and the
reviews, a bit simpler. This patch changes dwarf_expr_ctx to be
stack-allocated and removes the associated cleanup.
2016-10-21 Tom Tromey <tom@tromey.com>
* dwarf2loc.c (dwarf2_evaluate_loc_desc_full): Stack-allocate
dwarf_expr_context. Remove cleanups.
(dwarf2_locexpr_baton_eval)
(dwarf2_loc_desc_get_symbol_read_needs): Likewise.
* dwarf2expr.h (dwarf_expr_context, ~dwarf_expr_context): Add
constructors and destructors.
(new_dwarf_expr_context, free_dwarf_expr_context)
(make_cleanup_free_dwarf_expr_context): Don't declare.
* dwarf2-frame.c (execute_stack_op): Stack-allocate
dwarf_expr_context. Remove cleanups.
(dwarf_expr_context): Rename from new_dwarf_expr_context. Turn
into constructor.
(free_dwarf_expr_context, free_dwarf_expr_context_cleanup):
Remove.
(~dwarf_expr_context): Rename from
make_cleanup_free_dwarf_expr_context. Turn into destructor.
This removes some cleanups and manual allocation handling in
dwarf2loc.c with std::vector. Note that this patch has a case where
the vector would normally fall into the "use gdb::unique_ptr"
guidelines -- but here because the vector is immediately initialized,
I moved the initialization into the constructor call, for further
code savings.
2016-10-21 Tom Tromey <tom@tromey.com>
* dwarf2loc.c: Include <vector>.
(read_pieced_value, write_pieced_value)
(dwarf2_compile_expr_to_ax): Use std::vector.
Currently gdb supports DW_OP_GNU_push_tls_address, but not
DW_OP_form_tls_address. I think it would be better if the toolchain
as a whole moved to using the standard opcode, and the prerequisite to
this is getting gdb to recognize it.
GCC can sometimes emit DW_OP_form_tls_address for emultls targets. As
far as I know, nobody has ever tried this with gdb (since it wouldn't
work at all).
I don't think there's a major drawback to using a single opcode for
all targets, because computing the location of a thread-local is
already target specific.
This is PR gdb/11616.
I don't know how to write a test case for this; though it's worth
noting that there aren't explicit tests for DW_OP_GNU_push_tls_address
either -- and if I change GCC, these paths will be tested to the same
extent they are now.
2016-09-02 Tom Tromey <tom@tromey.com>
PR gdb/11616:
* dwarf2read.c (decode_locdesc): Handle DW_OP_form_tls_address.
* dwarf2loc.c (dwarf2_compile_expr_to_ax): Handle
DW_OP_form_tls_address.
(locexpr_describe_location_piece): Likewise.
* dwarf2expr.h (struct dwarf_expr_context_funcs): Update comment.
* dwarf2expr.c (execute_stack_op): Handle DW_OP_form_tls_address.
(ctx_no_get_tls_address): Mention DW_OP_form_tls_address.
* compile/compile-loc2c.c (struct insn_info): Update comment.
(compute_stack_depth_worker): Handle DW_OP_form_tls_address.
PR python/20190 arose from an exception I noticed when trying to use
the Python unwinder for Spider Monkey in Firefox.
The problem is that the unwinder wants to examine the value of a
thread-local variable. However, sympy_value rejects this because
symbol_read_needs_frame returns true for a TLS variable.
This problem arose once before, though in a different context:
https://sourceware.org/bugzilla/show_bug.cgi?id=11803
At the time Pedro and Daniel pointed out a simpler way to fix that bug
(see links in 20190 if you are interested); but for this new bug I
couldn't think of a similar fix and ended up implementing Daniel's
other suggestion:
https://sourceware.org/ml/gdb-patches/2010-07/msg00393.html
That is, this patch makes it possible to detect whether a symbol needs
a specific frame, or whether it just needs the inferior to have
registers.
Built and regtested on x86-64 Fedora 24.
2016-07-26 Tom Tromey <tom@tromey.com>
* symtab.c (register_symbol_computed_impl): Update.
PR python/20190:
* value.h (symbol_read_needs): Declare.
(symbol_read_needs_frame): Add comment.
* symtab.h (struct symbol_computed_ops) <read_variable>: Update
comment.
<get_symbol_read_needs>: Rename. Change return type.
* findvar.c (symbol_read_needs): New function.
(symbol_read_needs_frame): Rewrite.
(default_read_var_value): Use symbol_read_needs.
* dwarf2loc.c (struct symbol_needs_baton): Rename.
<needs>: Renamed from needs_frame. Changed type.
(needs_frame_read_addr_from_reg, symbol_needs_get_reg_value)
(symbol_needs_read_mem, symbol_needs_frame_base)
(symbol_needs_frame_cfa, symbol_needs_tls_address)
(symbol_needs_dwarf_call): Rename.
(needs_dwarf_reg_entry_value): Update.
(symbol_needs_ctx_funcs, dwarf2_loc_desc_get_symbol_read_needs):
Rename and update.
(locexpr_get_symbol_read_needs, loclist_symbol_needs): Likewise.
(dwarf2_locexpr_funcs, dwarf2_loclist_funcs): Update.
* defs.h (enum symbol_needs_kind): New.
2016-07-26 Tom Tromey <tom@tromey.com>
PR python/20190:
* gdb.threads/tls.exp (check_thread_local): Add python symbol
test.
GDB computes structure byte offsets using a 32 bit integer. And,
first it computes the offset in bits and then converts to bytes. The
result is that any offset that if 512K bytes or larger overflows.
This patch changes GDB to use LONGEST for such calculations.
PR gdb/17520 Structure offset wrong when 1/4 GB or greater.
* c-lang.h: Change all parameters, variables, and struct or union
members used as struct or union fie3ld offsets from int to
LONGEST.
* c-valprint.c: Likewise.
* cp-abi.c: Likewise.
* cp-abi.h: Likewise.
* cp-valprint.c: Likewise.
* d-valprint.c: Likewise.
* dwarf2loc.c: Likewise.
* eval.c: Likewise.
* extension-priv.h: Likewise.
* extension.c: Likewise.
* extension.h: Likewise.
* findvar.c: Likewise.
* gdbtypes.h: Likewise.
* gnu-v2-abi.c: Likewise.
* gnu-v3-abi.c: Likewise.
* go-valprint.c: Likewise.
* guile/guile-internal.h: Likewise.
* guile/scm-pretty-print.c: Likewise.
* jv-valprint.c Likewise.
* opencl-lang.c: Likewise.
* p-lang.h: Likewise.
* python/py-prettyprint.c: Likewise.
* python/python-internal.h: Likewise.
* spu-tdep.c: Likewise.
* typeprint.c: Likewise.
* valarith.c: Likewise.
* valops.c: Likewise.
* valprint.c: Likewise.
* valprint.h: Likewise.
* value.c: Likewise.
* value.h: Likewise.
* p-valprint.c: Likewise.
* c-typeprint.c (c_type_print_base): When printing offset, use
plongest, not %d.
* gdbtypes.c (recursive_dump_type): Ditto.
https://sourceware.org/bugzilla/show_bug.cgi?id=19893
I've traced the main source of the problem to pieced_value_funcs.coerce_ref not being
implemented. Since gdb always assumes references are implemented as pointers, this
causes it to think that it's dealing with a NULL pointer, thus breaking any operations
involving synthetic references.
What I did here was implementing pieced_value_funcs.coerce_ref using some of the synthetic
pointer handling code from indirect_pieced_value, as Pedro suggested. I also made a few
adjustments to the reference printing code so that it correctly shows either the address
of the referenced value or (if it's non-addressable) the "<synthetic pointer>" string.
I also wrote some unit tests based on Dwarf::assemble; these took a while to make
because in most cases I needed a synthetic reference to a physical variable. Additionally,
I started working on a unit test for classes that have a vtable, but ran into a few issues
so that'll probably go in a future patch. One thing that should definitely be fixed is that
proc function_range (called for MACRO_AT_func) will always try to compile/link using gcc
with the default options instead of g++, thus breaking C++ compilations that require e.g. libstdc++.
gdb/ChangeLog:
* dwarf2loc.c (coerce_pieced_ref, indirect_synthetic_pointer,
fetch_const_value_from_synthetic_pointer): New functions.
(indirect_pieced_value): Move lower half to indirect_synthetic_pointer.
(pieced_value_funcs): Implement coerce_ref.
* valops.c (value_addr): Call coerce_ref for synthetic references.
* valprint.c (valprint_check_validity): Return true for synthetic
references. Also, don't show "<synthetic pointer>" if they reference
addressable values.
(generic_val_print_ref): Handle synthetic references. Also move some
code to print_ref_address.
(print_ref_address, get_value_addr_contents): New functions.
gdb/testsuite/ChangeLog:
* gdb.dwarf2/implref.exp: Rename to...
* gdb.dwarf2/implref-const.exp: ...this. Also add more test statements.
* gdb.dwarf2/implref-array.c: New file.
* gdb.dwarf2/implref-array.exp: Likewise.
* gdb.dwarf2/implref-global.c: Likewise.
* gdb.dwarf2/implref-global.exp: Likewise.
* gdb.dwarf2/implref-struct.c: Likewise.
* gdb.dwarf2/implref-struct.exp: Likewise.
gdb/ChangeLog:
* dwarf2loc.c (dwarf2_find_location_expression): For DWO files still
add base_offset.
gdb/testsuite/ChangeLog:
* lib/dwarf.exp (build_executable_from_fission_assembler): Pass
$options when building executable.
* gdb.dwarf2/fission-loclists-pie.c: New file.
* gdb.dwarf2/fission-loclists-pie.exp: New file.
This patch fixes the following failures for rl78-elf:
FAIL: gdb.base/vla-datatypes.exp: print int_vla
FAIL: gdb.base/vla-datatypes.exp: print unsigned_int_vla
FAIL: gdb.base/vla-datatypes.exp: print double_vla
FAIL: gdb.base/vla-datatypes.exp: print float_vla
FAIL: gdb.base/vla-datatypes.exp: print long_vla
FAIL: gdb.base/vla-datatypes.exp: print unsigned_long_vla
FAIL: gdb.base/vla-datatypes.exp: print char_vla
FAIL: gdb.base/vla-datatypes.exp: print short_vla
FAIL: gdb.base/vla-datatypes.exp: print unsigned_short_vla
FAIL: gdb.base/vla-datatypes.exp: print unsigned_char_vla
FAIL: gdb.base/vla-datatypes.exp: print foo_vla
FAIL: gdb.base/vla-datatypes.exp: print bar_vla
FAIL: gdb.base/vla-datatypes.exp: print vla_struct_object
FAIL: gdb.base/vla-datatypes.exp: print vla_union_object
FAIL: gdb.base/vla-ptr.exp: print td_vla
FAIL: gdb.mi/mi-vla-c99.exp: evaluate complete vla
The first failure in this bunch occurs due to printing an incorrect
result for a variable length array:
print int_vla
$1 = {-1, -1, -1, -1, -1}
The result should actually be this:
$1 = {0, 2, 4, 6, 8}
When I started examining this bug, I found that printing an
individual array element worked correctly. E.g. "print int_vla[2]"
resulted in 4 being printed. I have not looked closely to see why
this is the case.
I found that evaluation of the location expression for int_vla was
causing problems. This is the relevant DWARF entry for int_vla:
<2><15a>: Abbrev Number: 10 (DW_TAG_variable)
<15b> DW_AT_name : (indirect string, offset: 0xbf): int_vla
<15f> DW_AT_decl_file : 1
<160> DW_AT_decl_line : 35
<161> DW_AT_type : <0x393>
<165> DW_AT_location : 4 byte block: 86 7a 94 2 (DW_OP_breg22 (r22): -6; DW_OP_deref_size: 2)
I found that DW_OP_breg22 was providing a correct result.
DW_OP_deref_size was fetching the correct value from memory. However,
the value being fetched should be considered a pointer.
DW_OP_deref_size zero extends the fetched value prior to pushing
it onto the evaluation stack. (The DWARF-4 document specifies this
action; so GDB is faithfully implementing the DWARF-4 specification.)
However, zero extending the pointer is not sufficient for converting
that value to an address for rl78 and (perhaps) other architectures
which define a `pointer_to_address' method. (I suspect that m32c
would have the same problem.)
Ideally, we would perform the pointer to address conversion in
DW_OP_deref_size. We don't, however, know the type of the object
that the address refers to in DW_OP_deref_size. I can't think
of a way to infer the type at that point in the code.
Before proceeding, I should note that there are two other DWARF
operations that could be used in place of DW_OP_deref_size. One of
these is DW_OP_GNU_deref_type. Current GDB implements this operation,
but as is obvious from the name, it is non-standard DWARF. The other
operation is DW_OP_xderef_size. Even though it's part of DWARF-2
through DWARF-4 specifications, it's not presently implemented in GDB.
Present day GCC does not output dwarf expressions containing this
operation either. [Of the two, I like DW_OP_GNU_deref_type better.
Using it avoids the need to specify an "address space identifier".
(GCC, GDB, and other non-free tools all need to agree on the meanings
of these identifiers.)]
Back to the bug analysis...
The closest consumer of the DW_OP_deref_size result is the
DWARF_VALUE_MEMORY case in dwarf2_evaluate_loc_desc_full. At that
location, we do know the object type to which the address is intended
to refer. I added code to perform a pointer to address conversion at
this location. (See the patch.)
I do have some misgivings regarding this patch. As noted earlier, it
would really be better to perform the pointer to address conversion in
DW_OP_deref_size. I can't, however, think of a way to make this work.
Changing GCC to output one of the other aforementioned operations might
be preferable but, as noted earlier, these solutions have problems as
well. Long term, I think it'd be good to have something like
DW_OP_GNU_deref_type become part of the standard. If that can't or
won't happen, we'll need to implement DW_OP_xderef_size.
But until that happens, this patch will work for expressions in which
DW_OP_deref_size occurs last. It should even work for dereferences
followed by adding an offset. I don't think it'll work for more than
one dereference in the same expression.
gdb/ChangeLog:
* dwarf2loc.c (dwarf2_evaluate_loc_desc_full): Perform a pointer
to address conversion for DWARF_VALUE_MEMORY.
We have noticed that GDB would sometimes crash trying to print
from a nested function the value of a variable declared in an
enclosing scope. This appears to be target dependent, although
that correlation might only be fortuitious. We noticed the issue
on x86_64-darwin, x86-vxworks6 and x86-solaris. The investigation
was done on Darwin.
This is a new feature that was introduced by:
commit 63e43d3aed
Date: Thu Feb 5 17:00:06 2015 +0100
DWARF: handle non-local references in nested functions
We can reproduce the problem with one of the testcases that was
added with the patch (gdb.base/nested-subp1.exp), where we have...
18 int
19 foo (int i1)
20 {
21 int
22 nested (int i2)
23 {
[...]
27 return i1 * i2; /* STOP */
28 }
... After building the example program, and running until line 27,
try printing the value of "i1":
% gdb gdb.base/nested-subp1
(gdb) break foo.c:27
(gdb) run
Breakpoint 1, nested (i2=2) at /[...]/nested-subp1.c:27
27 return i1 * i2; /* STOP */
(gdb) p i1
[1] 73090 segmentation fault ../gdb -q gdb.base/nested-subp1
Ooops!
What happens is that, because the reference is non-local, we are trying
to follow the function's static link, which does...
/* If we don't know how to compute FRAME's base address, don't give up:
maybe the frame we are looking for is upper in the stace frame. */
if (framefunc != NULL
&& SYMBOL_BLOCK_OPS (framefunc)->get_frame_base != NULL
&& (SYMBOL_BLOCK_OPS (framefunc)->get_frame_base (framefunc, frame)
== upper_frame_base))
... or, in other words, calls the get_frame_base "method" of
framefunc's struct symbol_block_ops data. This resolves to
the block_op_get_frame_base function.
Looking at the function's implementation, we see:
struct dwarf2_locexpr_baton *dlbaton;
[...]
dlbaton = SYMBOL_LOCATION_BATON (framefunc);
[...]
result = dwarf2_evaluate_loc_desc (type, frame, start, length,
dlbaton->per_cu);
^^^^^^^^^^^^^^^
Printing dlbaton->per_cu gives a value that seems fairly bogus for
a memory address (0x60). Because of it, dwarf2_evaluate_loc_desc
then crashes trying to dereference it.
What's different on Darwin compared to Linux is that the function's
frame base is encoded using the following form:
.byte 0x40 # uleb128 0x40; (DW_AT_frame_base)
.byte 0x6 # uleb128 0x6; (DW_FORM_data4)
... and so dwarf2_symbol_mark_computed ends up creating
a SYMBOL_LOCATION_BATON as a struct dwarf2_loclist_baton:
if (attr_form_is_section_offset (attr)
/* .debug_loc{,.dwo} may not exist at all, or the offset may be outside
the section. If so, fall through to the complaint in the
other branch. */
&& DW_UNSND (attr) < dwarf2_section_size (objfile, section))
{
struct dwarf2_loclist_baton *baton;
[...]
SYMBOL_LOCATION_BATON (sym) = baton;
However, if you look more closely at block_op_get_frame_base's
implementation, you'll notice that the function extracts the
symbol's SYMBOL_LOCATION_BATON as a dwarf2_locexpr_baton
(a DWARF _expression_ rather than a _location list_).
That's why we end up decoding the DLBATON improperly, and thus
pass a random dlbaton->per_cu when calling dwarf2_evaluate_loc_desc.
This works on x86_64-linux, because we indeed have the frame base
described using a different form:
.uleb128 0x40 # (DW_AT_frame_base)
.uleb128 0x18 # (DW_FORM_exprloc)
This patch fixes the issue by doing what we do for most (if not all)
other such methods: providing one implementation each for loc-list,
and loc-expr. Both implementations are nearly identical, so perhaps
we might later want to improve this. But this patch first tries to
fix the crash first, leaving the design issue for later.
gdb/ChangeLog:
* dwarf2loc.c (locexpr_get_frame_base): Renames
block_op_get_frame_base.
(dwarf2_block_frame_base_locexpr_funcs): Replace reference to
block_op_get_frame_base by reference to locexpr_get_frame_base.
(loclist_get_frame_base): New function, near identical copy of
locexpr_get_frame_base.
(dwarf2_block_frame_base_loclist_funcs): Replace reference to
block_op_get_frame_base by reference to loclist_get_frame_base.
Tested on x86_64-darwin (AdaCore testsuite), and x86_64-linux
(official testsuite).
GDB's current behavior when dealing with non-local references in the
context of nested fuctions is approximative:
- code using valops.c:value_of_variable read the first available stack
frame that holds the corresponding variable (whereas there can be
multiple candidates for this);
- code directly relying on read_var_value will instead read non-local
variables in frames where they are not even defined.
This change adds the necessary context to symbol reads (to get the block
they belong to) and to blocks (the static link property, if any) so that
GDB can make the proper decisions when dealing with non-local varibale
references.
gdb/ChangeLog:
* ada-lang.c (ada_read_var_value): Add a var_block argument
and pass it to default_read_var_value.
* block.c (block_static_link): New accessor.
* block.h (block_static_link): Declare it.
* buildsym.c (finish_block_internal): Add a static_link
argument. If there is a static link, associate it to the new
block.
(finish_block): Add a static link argument and pass it to
finish_block_internal.
(end_symtab_get_static_block): Update calls to finish_block and
to finish_block_internal.
(end_symtab_with_blockvector): Update call to
finish_block_internal.
* buildsym.h: Forward-declare struct dynamic_prop.
(struct context_stack): Add a static_link field.
(finish_block): Add a static link argument.
* c-exp.y: Remove an obsolete comment (evaluation of variables
already start from the selected frame, and now they climb *up*
the call stack) and propagate the block information to the
produced expression.
* d-exp.y: Likewise.
* f-exp.y: Likewise.
* go-exp.y: Likewise.
* jv-exp.y: Likewise.
* m2-exp.y: Likewise.
* p-exp.y: Likewise.
* coffread.c (coff_symtab_read): Update calls to finish_block.
* dbxread.c (process_one_symbol): Likewise.
* xcoffread.c (read_xcoff_symtab): Likewise.
* compile/compile-c-symbols.c (convert_one_symbol): Promote the
"sym" parameter to struct block_symbol, update its uses and pass
its block to calls to read_var_value.
(convert_symbol_sym): Update the calls to convert_one_symbol.
* compile/compile-loc2c.c (do_compile_dwarf_expr_to_c): Update
call to read_var_value.
* dwarf2loc.c (block_op_get_frame_base): New.
(dwarf2_block_frame_base_locexpr_funcs): Implement the
get_frame_base method.
(dwarf2_block_frame_base_loclist_funcs): Likewise.
(dwarf2locexpr_baton_eval): Add a frame argument and use it
instead of the selected frame in order to evaluate the
expression.
(dwarf2_evaluate_property): Add a frame argument. Update call
to dwarf2_locexpr_baton_eval to provide a frame in available and
to handle the absence of address stack.
* dwarf2loc.h (dwarf2_evaluate_property): Add a frame argument.
* dwarf2read.c (attr_to_dynamic_prop): Add a forward
declaration.
(read_func_scope): Record any available static link description.
Update call to finish_block.
(read_lexical_block_scope): Update call to finish_block.
* findvar.c (follow_static_link): New.
(get_hosting_frame): New.
(default_read_var_value): Add a var_block argument. Use
get_hosting_frame to handle non-local references.
(read_var_value): Add a var_block argument and pass it to the
LA_READ_VAR_VALUE method.
* gdbtypes.c (resolve_dynamic_range): Update calls to
dwarf2_evaluate_property.
(resolve_dynamic_type_internal): Likewise.
* guile/scm-frame.c (gdbscm_frame_read_var): Update call to
read_var_value, passing it the block coming from symbol lookup.
* guile/scm-symbol.c (gdbscm_symbol_value): Update call to
read_var_value (TODO).
* infcmd.c (finish_command_continuation): Update call to
read_var_value, passing it the block coming from symbol lookup.
* infrun.c (insert_exception_resume_breakpoint): Likewise.
* language.h (struct language_defn): Add a var_block argument to
the LA_READ_VAR_VALUE method.
* objfiles.c (struct static_link_htab_entry): New.
(static_link_htab_entry_hash): New.
(static_link_htab_entry_eq): New.
(objfile_register_static_link): New.
(objfile_lookup_static_link): New.
(free_objfile): Free the STATIC_LINKS hashed map if needed.
* objfiles.h: Include hashtab.h.
(struct objfile): Add a static_links field.
(objfile_register_static_link): New.
(objfile_lookup_static_link): New.
* printcmd.c (print_variable_and_value): Update call to
read_var_value.
* python/py-finishbreakpoint.c (bpfinishpy_init): Likewise.
* python/py-frame.c (frapy_read_var): Update call to
read_var_value, passing it the block coming from symbol lookup.
* python/py-framefilter.c (extract_sym): Add a sym_block
parameter and set the pointed value to NULL (TODO).
(enumerate_args): Update call to extract_sym.
(enumerate_locals): Update calls to extract_sym and to
read_var_value.
* python/py-symbol.c (sympy_value): Update call to
read_var_value (TODO).
* stack.c (read_frame_local): Update call to read_var_value.
(read_frame_arg): Likewise.
(return_command): Likewise.
* symtab.h (struct symbol_block_ops): Add a get_frame_base
method.
(struct symbol): Add a block field.
(SYMBOL_BLOCK): New accessor.
* valops.c (value_of_variable): Remove frame/block handling and
pass the block argument to read_var_value, which does this job
now.
(value_struct_elt_for_reference): Update calls to
read_var_value.
(value_of_this): Pass the block found to read_var_value.
* value.h (read_var_value): Add a var_block argument.
(default_read_var_value): Likewise.
gdb/testsuite/ChangeLog:
* gdb.base/nested-subp1.exp: New file.
* gdb.base/nested-subp1.c: New file.
* gdb.base/nested-subp2.exp: New file.
* gdb.base/nested-subp2.c: New file.
* gdb.base/nested-subp3.exp: New file.
* gdb.base/nested-subp3.c: New file.
This patch is mostly extracted from Pedro's C++ branch. It adds explicit
casts from integer to enum types, where it is really the intention to do
so. This could be because we are ...
* iterating on enum values (we need to iterate on an equivalent integer)
* converting from a value read from bytes (dwarf attribute, agent
expression opcode) to the equivalent enum
* reading the equivalent integer value from another language (Python/Guile)
An exception to that is the casts in regcache.c. It seems to me like
struct regcache's register_status field could be a pointer to an array of
enum register_status. Doing so would waste a bit of memory (4 bytes
used by the enum vs 1 byte used by the current signed char, for each
register). If we switch to C++11 one day, we can define the underlying
type of an enum type, so we could have the best of both worlds.
gdb/ChangeLog:
* arm-tdep.c (set_fp_model_sfunc): Add cast from integer to enum.
(arm_set_abi): Likewise.
* ax-general.c (ax_print): Likewise.
* c-exp.y (exp : string_exp): Likewise.
* compile/compile-loc2c.c (compute_stack_depth_worker): Likewise.
(do_compile_dwarf_expr_to_c): Likewise.
* cp-name-parser.y (demangler_special : DEMANGLER_SPECIAL start):
Likewise.
* dwarf2expr.c (execute_stack_op): Likewise.
* dwarf2loc.c (dwarf2_compile_expr_to_ax): Likewise.
(disassemble_dwarf_expression): Likewise.
* dwarf2read.c (dwarf2_add_member_fn): Likewise.
(read_array_order): Likewise.
(abbrev_table_read_table): Likewise.
(read_attribute_value): Likewise.
(skip_unknown_opcode): Likewise.
(dwarf_decode_macro_bytes): Likewise.
(dwarf_decode_macros): Likewise.
* eval.c (value_f90_subarray): Likewise.
* guile/scm-param.c (gdbscm_make_parameter): Likewise.
* i386-linux-tdep.c (i386_canonicalize_syscall): Likewise.
* infrun.c (handle_command): Likewise.
* memory-map.c (memory_map_start_memory): Likewise.
* osabi.c (set_osabi): Likewise.
* parse.c (operator_length_standard): Likewise.
* ppc-linux-tdep.c (ppc_canonicalize_syscall): Likewise, and use
single return point.
* python/py-frame.c (gdbpy_frame_stop_reason_string): Likewise.
* python/py-symbol.c (gdbpy_lookup_symbol): Likewise.
(gdbpy_lookup_global_symbol): Likewise.
* record-full.c (record_full_restore): Likewise.
* regcache.c (regcache_register_status): Likewise.
(regcache_raw_read): Likewise.
(regcache_cooked_read): Likewise.
* rs6000-tdep.c (powerpc_set_vector_abi): Likewise.
* symtab.c (initialize_ordinary_address_classes): Likewise.
* target-debug.h (target_debug_print_signals): Likewise.
* utils.c (do_restore_current_language): Likewise.
Initially there is some chain (let's say the longest one
but that doe snot matter). Consequently its elements from the middle are
being removed and there remains only some few unambiguous top and bottom ones.
The original idea why the comparison should be sharp ("<") was that if there
are multiple chains like (0xaddr show jmp instruction address):
main(0x100) -> a(0x200) -> d(0x400)
main(0x100) -> a(0x200) -> c(0x300) -> d(0x400)
then - such situation cannot exist - if two jmp instructions in "a" have the
same address they must also jump to the same address (*).
(*) jump to a computed address would be never considered for the DWARF
tail-call records.
So there could be:
main(0x100) -> a(0x200) -> d(0x400)
main(0x100) -> a(0x270) -> c(0x300) -> d(0x400)
But then "a" frame itself is ambiguous and it must not be displayed.
I did not realize that there can be self-tail-call:
main(0x100) -> a(0x200) -> d(0x400)
main(0x100) -> a(0x280) -> a(0x200) -> d(0x400)
which intersects to:
main(0x100) -> <???>? -> a(0x200) -> d(0x400)
And so if the first chain was chosen the
main(0x100) -> a(0x200) -> d(0x400)
then the final intersection has callers+callees==length.
> for example, if CALLERS is 3 and
> CALLEES is 2, what does the chain look like?
main(0x100) -> x(0x150) -> y(0x200) -> <???>? -> a(0x200) -> d(0x400)
And if LENGTH is 7 then:
call_site[0] = main(0x100)
call_site[1] = x(0x150)
call_site[2] = y(0x200)
call_site[3] = garbage
call_site[4] = garbage
call_site[5] = a(0x200)
call_site[6] = d(0x400)
gdb/ChangeLog
2015-06-01 Andreas Schwab <schwab@linux-m68k.org>
Jan Kratochvil <jan.kratochvil@redhat.com>
PR symtab/18392
* dwarf2-frame-tailcall.c (pretended_chain_levels): Correct
assertion.
* dwarf2loc.c (chain_candidate): Likewise.
gdb/testsuite/ChangeLog
2015-06-01 Jan Kratochvil <jan.kratochvil@redhat.com>
PR symtab/18392
* gdb.arch/amd64-tailcall-self.S: New file.
* gdb.arch/amd64-tailcall-self.c: New file.
* gdb.arch/amd64-tailcall-self.exp: New file.
This is the second part of enhancing the debugger to print the value
of arrays of records whose size is variable when only standard DWARF
info is available (no GNAT encoding). For instance:
subtype Small_Type is Integer range 0 .. 10;
type Record_Type (I : Small_Type := 0) is record
S : String (1 .. I);
end record;
type Array_Type is array (Integer range <>) of Record_Type;
A1 : Array_Type := (1 => (I => 0, S => <>),
2 => (I => 1, S => "A"),
3 => (I => 2, S => "AB"));
Currently, GDB prints the following output:
(gdb) p a1
$1 = (
The error happens while the ada-valprint module is trying to print
the value of an element of our array. Because of the fact that
the array's element (type Record_Type) has a variant size, the DWARF
info for our array provide the array's stride:
<1><749>: Abbrev Number: 10 (DW_TAG_array_type)
<74a> DW_AT_name : (indirect string, offset: 0xb6d): pck__T18s
<74e> DW_AT_byte_stride : 16
<74f> DW_AT_type : <0x6ea>
And because our array has a stride, ada-valprint treats it the same
way as packed arrays (see ada-valprint.c::ada_val_print_array):
if (TYPE_FIELD_BITSIZE (type, 0) > 0)
val_print_packed_array_elements (type, valaddr, offset_aligned,
0, stream, recurse,
original_value, options);
The first thing that we should notice in the call above is that
the "valaddr" buffer and the associated offset (OFFSET_ALIGNED)
is passed, but that the corresponding array's address is not.
This can be explained by looking inside val_print_packed_array_elements,
where we see that the function unpacks each element of our array from
the buffer alone (ada_value_primitive_packed_val), and then prints
the resulting artificial value instead:
v0 = ada_value_primitive_packed_val (NULL, valaddr + offset,
(i0 * bitsize) / HOST_CHAR_BIT,
(i0 * bitsize) % HOST_CHAR_BIT,
bitsize, elttype);
[...]
val_print (elttype, value_contents_for_printing (v0),
value_embedded_offset (v0), 0, stream,
recurse + 1, v0, &opts, current_language);
Of particular interest, here, is the fact that we call val_print
with a null address, which is OK, since we're providing a buffer
instead (value_contents_for_printing). Also, providing an address
might not always possible, since packing could place elements at
boundaries that are not byte-aligned.
Things go south when val_print tries to see if there is a pretty-printer
that could be applied. In particular, one of the first things that
the Python pretty-printer does is to create a value using our buffer,
and the given address, which in this case is null (see call to
value_from_contents_and_address in gdbpy_apply_val_pretty_printer).
value_from_contents_and_address, in turn immediately tries to resolve
the type, using the given address, which is null. But, because our
array element is a record containing an array whose bound is the value
of one of its elements (the "s" component), the debugging info for
the array's upper bound is a reference...
<3><71a>: Abbrev Number: 7 (DW_TAG_subrange_type)
<71b> DW_AT_type : <0x724>
<71f> DW_AT_upper_bound : <0x703>
... to component "i" of our record...
<2><703>: Abbrev Number: 5 (DW_TAG_member)
<704> DW_AT_name : i
<706> DW_AT_decl_file : 2
<707> DW_AT_decl_line : 6
<708> DW_AT_type : <0x6d1>
<70c> DW_AT_data_member_location: 0
... where that component is located at offset 0 of the start
of the record. dwarf2_evaluate_property correctly determines
the offset where to load the value of the bound from, but then
tries to read that value from inferior memory using the address
that was given, which is null. See case PROP_ADDR_OFFSET in
dwarf2_evaluate_property:
val = value_at (baton->offset_info.type,
pinfo->addr + baton->offset_info.offset);
This triggers a memory error, which then causes the printing to terminate.
Since there are going to be situations where providing an address
alone is not going to be sufficient (packed arrays where array elements
are not stored at byte boundaries), this patch fixes the issue by
enhancing the type resolution to take both address and data. This
follows the same principle as the val_print module, where both
address and buffer ("valaddr") can be passed as arguments. If the data
has already been fetched from inferior memory (or provided by the
debugging info in some form -- Eg a constant), then use that data
instead of reading it from inferior memory.
Note that this should also be a good step towards being able to handle
dynamic types whose value is stored outside of inferior memory
(Eg: in a register).
With this patch, GDB isn't able to print all of A1, but does perform
a little better:
(gdb) p a1
$1 = ((i => 0, s => , (i => 1, s => , (i => 2, s => )
There is another issue which is independent of this one, and will
therefore be patched separately.
gdb/ChangeLog:
* dwarf2loc.h (struct property_addr_info): Add "valaddr" field.
* dwarf2loc.c (dwarf2_evaluate_property): Add handling of
pinfo->valaddr.
* gdbtypes.h (resolve_dynamic_type): Add "valaddr" parameter.
* gdbtypes.c (resolve_dynamic_struct): Set pinfo.valaddr.
(resolve_dynamic_type_internal): Set pinfo.valaddr.
Add handling of addr_stack->valaddr.
(resolve_dynamic_type): Add "valaddr" parameter.
Set pinfo.valaddr field.
* ada-lang.c (ada_discrete_type_high_bound): Update call to
resolve_dynamic_type.
(ada_discrete_type_low_bound): Likewise.
* findvar.c (default_read_var_value): Likewise.
* value.c (value_from_contents_and_address): Likewise.
This patch splits the TRY_CATCH macro into three, so that we go from
this:
~~~
volatile gdb_exception ex;
TRY_CATCH (ex, RETURN_MASK_ERROR)
{
}
if (ex.reason < 0)
{
}
~~~
to this:
~~~
TRY
{
}
CATCH (ex, RETURN_MASK_ERROR)
{
}
END_CATCH
~~~
Thus, we'll be getting rid of the local volatile exception object, and
declaring the caught exception in the catch block.
This allows reimplementing TRY/CATCH in terms of C++ exceptions when
building in C++ mode, while still allowing to build GDB in C mode
(using setjmp/longjmp), as a transition step.
TBC, after this patch, is it _not_ valid to have code between the TRY
and the CATCH blocks, like:
TRY
{
}
// some code here.
CATCH (ex, RETURN_MASK_ERROR)
{
}
END_CATCH
Just like it isn't valid to do that with C++'s native try/catch.
By switching to creating the exception object inside the CATCH block
scope, we can get rid of all the explicitly allocated volatile
exception objects all over the tree, and map the CATCH block more
directly to C++'s catch blocks.
The majority of the TRY_CATCH -> TRY+CATCH+END_CATCH conversion was
done with a script, rerun from scratch at every rebase, no manual
editing involved. After the mechanical conversion, a few places
needed manual intervention, to fix preexisting cases where we were
using the exception object outside of the TRY_CATCH block, and cases
where we were using "else" after a 'if (ex.reason) < 0)' [a CATCH
after this patch]. The result was folded into this patch so that GDB
still builds at each incremental step.
END_CATCH is necessary for two reasons:
First, because we name the exception object in the CATCH block, which
requires creating a scope, which in turn must be closed somewhere.
Declaring the exception variable in the initializer field of a for
block, like:
#define CATCH(EXCEPTION, mask) \
for (struct gdb_exception EXCEPTION; \
exceptions_state_mc_catch (&EXCEPTION, MASK); \
EXCEPTION = exception_none)
would avoid needing END_CATCH, but alas, in C mode, we build with C90,
which doesn't allow mixed declarations and code.
Second, because when TRY/CATCH are wired to real C++ try/catch, as
long as we need to handle cleanup chains, even if there's no CATCH
block that wants to catch the exception, we need for stop at every
frame in the unwind chain and run cleanups, then rethrow. That will
be done in END_CATCH.
After we require C++, we'll still need TRY/CATCH/END_CATCH until
cleanups are completely phased out -- TRY/CATCH in C++ mode will
save/restore the current cleanup chain, like in C mode, and END_CATCH
catches otherwise uncaugh exceptions, runs cleanups and rethrows, so
that C++ cleanups and exceptions can coexist.
IMO, this still makes the TRY/CATCH code look a bit more like a
newcomer would expect, so IMO worth it even if we weren't considering
C++.
gdb/ChangeLog.
2015-03-07 Pedro Alves <palves@redhat.com>
* common/common-exceptions.c (struct catcher) <exception>: No
longer a pointer to volatile exception. Now an exception value.
<mask>: Delete field.
(exceptions_state_mc_init): Remove all parameters. Adjust.
(exceptions_state_mc): No longer pop the catcher here.
(exceptions_state_mc_catch): New function.
(throw_exception): Adjust.
* common/common-exceptions.h (exceptions_state_mc_init): Remove
all parameters.
(exceptions_state_mc_catch): Declare.
(TRY_CATCH): Rename to ...
(TRY): ... this. Remove EXCEPTION and MASK parameters.
(CATCH, END_CATCH): New.
All callers adjusted.
gdb/gdbserver/ChangeLog:
2015-03-07 Pedro Alves <palves@redhat.com>
Adjust all callers of TRY_CATCH to use TRY/CATCH/END_CATCH
instead.
In C, we can forward declare static structure instances. That doesn't
work in C++ though. C++ treats these as definitions. So then the
compiler complains about symbol redefinition, like:
src/gdb/elfread.c:1569:29: error: redefinition of ‘const sym_fns elf_sym_fns_lazy_psyms’
src/gdb/elfread.c:53:29: error: ‘const sym_fns elf_sym_fns_lazy_psyms’ previously declared here
The intent of static here is naturally to avoid making these objects
visible outside the compilation unit. The equivalent in C++ would be
to instead define the objects in the anonymous namespace. But given
that it's desirable to leave the codebase compiling as both C and C++
for a while, this just makes the objects extern.
(base_breakpoint_ops is already declared in breakpoint.h, so we can
just remove the forward declare from breakpoint.c)
gdb/ChangeLog:
2015-02-11 Tom Tromey <tromey@redhat.com>
Pedro Alves <palves@redhat.com>
* breakpoint.c (base_breakpoint_ops): Delete.
* dwarf2loc.c (dwarf_expr_ctx_funcs): Make extern.
* elfread.c (elf_sym_fns_gdb_index, elf_sym_fns_lazy_psyms): Make extern.
* guile/guile.c (guile_extension_script_ops, guile_extension_ops): Make extern.
* ppcnbsd-tdep.c (ppcnbsd2_sigtramp): Make extern.
* python/py-arch.c (arch_object_type): Make extern.
* python/py-block.c (block_syms_iterator_object_type): Make extern.
* python/py-bpevent.c (breakpoint_event_object_type): Make extern.
* python/py-cmd.c (cmdpy_object_type): Make extern.
* python/py-continueevent.c (continue_event_object_type)
* python/py-event.h (GDBPY_NEW_EVENT_TYPE): Remove 'qual'
parameter. Update all callers.
* python/py-evtregistry.c (eventregistry_object_type): Make extern.
* python/py-exitedevent.c (exited_event_object_type): Make extern.
* python/py-finishbreakpoint.c (finish_breakpoint_object_type): Make extern.
* python/py-function.c (fnpy_object_type): Make extern.
* python/py-inferior.c (inferior_object_type, membuf_object_type): Make extern.
* python/py-infevents.c (call_pre_event_object_type)
(inferior_call_post_event_object_type).
(memory_changed_event_object_type): Make extern.
* python/py-infthread.c (thread_object_type): Make extern.
* python/py-lazy-string.c (lazy_string_object_type): Make extern.
* python/py-linetable.c (linetable_entry_object_type)
(linetable_object_type, ltpy_iterator_object_type): Make extern.
* python/py-newobjfileevent.c (new_objfile_event_object_type)
(clear_objfiles_event_object_type): Make extern.
* python/py-objfile.c (objfile_object_type): Make extern.
* python/py-param.c (parmpy_object_type): Make extern.
* python/py-progspace.c (pspace_object_type): Make extern.
* python/py-signalevent.c (signal_event_object_type): Make extern.
* python/py-symtab.c (symtab_object_type, sal_object_type): Make extern.
* python/py-type.c (type_object_type, field_object_type)
(type_iterator_object_type): Make extern.
* python/python.c (python_extension_script_ops)
(python_extension_ops): Make extern.
* stap-probe.c (stap_probe_ops): Make extern.
Consider the following declarations:
type Array_Type is array (Integer range <>) of Integer;
type Record_Type (N : Integer) is record
A : Array_Type (1 .. N);
end record;
R : Record_Type := Get (10);
It defines what Ada programers call a "discriminated record", where
"N" is a component of that record called a "discriminant", and where
"A" is a component defined as an array type whose upper bound is
equal to the value of the discriminant.
So far, we rely on a number of fairly complex GNAT-specific encodings
to handle this situation. This patch is to enhance GDB to be able to
print this record in the case where the compiler has been modified
to replace those encodings by pure DWARF constructs.
In particular, the debugging information generated for the record above
looks like the following. "R" is a record..
.uleb128 0x10 # (DIE (0x13e) DW_TAG_structure_type)
.long .LASF17 # DW_AT_name: "foo__record_type"
... whose is is of course dynamic (not our concern here)...
.uleb128 0xd # DW_AT_byte_size
.byte 0x97 # DW_OP_push_object_address
.byte 0x94 # DW_OP_deref_size
.byte 0x4
.byte 0x99 # DW_OP_call4
.long 0x19b
.byte 0x23 # DW_OP_plus_uconst
.uleb128 0x7
.byte 0x9 # DW_OP_const1s
.byte 0xfc
.byte 0x1a # DW_OP_and
.byte 0x1 # DW_AT_decl_file (foo.adb)
.byte 0x6 # DW_AT_decl_line
... and then has 2 members, fist "n" (our discriminant);
.uleb128 0x11 # (DIE (0x153) DW_TAG_member)
.ascii "n\0" # DW_AT_name
.byte 0x1 # DW_AT_decl_file (foo.adb)
.byte 0x6 # DW_AT_decl_line
.long 0x194 # DW_AT_type
.byte 0 # DW_AT_data_member_location
... and "A"...
.uleb128 0x11 # (DIE (0x181) DW_TAG_member)
.ascii "a\0" # DW_AT_name
.long 0x15d # DW_AT_type
.byte 0x4 # DW_AT_data_member_location
... which is an array ...
.uleb128 0x12 # (DIE (0x15d) DW_TAG_array_type)
.long .LASF18 # DW_AT_name: "foo__record_type__T4b"
.long 0x194 # DW_AT_type
... whose lower bound is implicitly 1, and the upper bound
a reference to DIE 0x153 = "N":
.uleb128 0x13 # (DIE (0x16a) DW_TAG_subrange_type)
.long 0x174 # DW_AT_type
.long 0x153 # DW_AT_upper_bound
This patch enhanced GDB to understand references to other DIEs
where the DIE's address is at an offset of its enclosing type.
The difficulty was that the address used to resolve the array's
type (R's address + 4 bytes) is different from the address used
as the base to compute N's address (an offset to R's address).
We're solving this issue by using a stack of addresses rather
than a single address when trying to resolve a type. Each address
in the stack corresponds to each containing level. For instance,
if resolving the field of a struct, the stack should contain
the address of the field at the top, and then the address of
the struct. That way, if the field makes a reference to an object
of the struct, we can retrieve the address of that struct, and
properly resolve the dynamic property references that struct.
gdb/ChangeLog:
* gdbtypes.h (struct dynamic_prop): New PROP_ADDR_OFFSET enum
kind.
* gdbtypes.c (resolve_dynamic_type_internal): Replace "addr"
parameter by "addr_stack" parameter.
(resolve_dynamic_range): Replace "addr" parameter by
"stack_addr" parameter. Update function documentation.
Update code accordingly.
(resolve_dynamic_array, resolve_dynamic_union)
(resolve_dynamic_struct, resolve_dynamic_type_internal): Likewise.
(resolve_dynamic_type): Update code, following the changes made
to resolve_dynamic_type_internal's interface.
* dwarf2loc.h (struct property_addr_info): New.
(dwarf2_evaluate_property): Replace "address" parameter
by "addr_stack" parameter. Adjust function documentation.
(struct dwarf2_offset_baton): New.
(struct dwarf2_property_baton): Update documentation of
field "referenced_type" to be more general. New field
"offset_info" in union data field.
* dwarf2loc.c (dwarf2_evaluate_property): Replace "address"
parameter by "addr_stack" parameter. Adjust code accordingly.
Add support for PROP_ADDR_OFFSET properties.
* dwarf2read.c (attr_to_dynamic_prop): Add support for
DW_AT_data_member_location attributes as well. Use case
statements instead of if/else condition.
gdb/testsuite/ChangeLog:
* gdb.ada/disc_arr_bound: New testcase.
Tested on x86_64-linux, no regression.
I see the error message "access outside bounds of object referenced
via synthetic pointer" in the two fails below of mips gdb testing
print d[-2]^M
access outside bounds of object referenced via synthetic pointer^M
(gdb) FAIL: gdb.dwarf2/implptrconst.exp: print d[-2]
(gdb) print/d p[-1]^M
access outside bounds of object referenced via synthetic pointer^M
(gdb) FAIL: gdb.dwarf2/implptrpiece.exp: print/d p[-1]
in the first test, 'd[-2]' is processed by GDB as '* (&d[-2])'. 'd'
is a synthetic pointer, so its value is zero, the address of 'd[-2]'
is -2. In dwarf2loc.c:indirect_pieced_value,
/* This is an offset requested by GDB, such as value subscripts.
However, due to how synthetic pointers are implemented, this is
always presented to us as a pointer type. This means we have to
sign-extend it manually as appropriate. */
byte_offset = value_as_address (value);
if (TYPE_LENGTH (value_type (value)) < sizeof (LONGEST))
byte_offset = gdb_sign_extend (byte_offset,
8 * TYPE_LENGTH (value_type (value)));
byte_offset += piece->v.ptr.offset;
We know that the value is really an offset instead of address, so the
fix is to extract the value as an (signed) offset.
gdb:
2015-01-08 Pedro Alves <palves@redhat.com>
Yao Qi <yao@codesourcery.com>
* dwarf2loc.c (indirect_pieced_value): Don't call
gdb_sign_extend. Call extract_signed_integer instead.
* utils.c (gdb_sign_extend): Remove.
* utils.h (gdb_sign_extend): Remove declaration.
This final patch adds the new "compile" command and subcommands, and
all the machinery needed to make it work.
A shared library supplied by gcc is used for all communications with
gcc. Types and most aspects of symbols are provided directly by gdb
to the compiler using this library.
gdb provides some information about the user's code using plain text.
Macros are emitted this way, and DWARF location expressions (and
bounds for VLA) are compiled to C code.
This hybrid approach was taken because, on the one hand, it is better
to provide global declarations and such on demand; but on the other
hand, for local variables, translating DWARF location expressions to C
was much simpler than exporting a full compiler API to gdb -- the same
result, only easier to implement, understand, and debug.
In the ordinary mode, the user's expression is wrapped in a dummy
function. After compilation, gdb inserts the resulting object code
into the inferior, then calls this function.
Access to local variables is provided by noting which registers are
used by location expressions, and passing a structure of register
values into the function. Writes to registers are supported by
copying out these values after the function returns.
This approach was taken so that we could eventually implement other
more interesting features based on this same infrastructure; for
example, we're planning to investigate inferior-side breakpoint
conditions.
gdb/ChangeLog
2014-12-12 Phil Muldoon <pmuldoon@redhat.com>
Jan Kratochvil <jan.kratochvil@redhat.com>
Tom Tromey <tromey@redhat.com>
* NEWS: Update.
* symtab.h (struct symbol_computed_ops) <generate_c_location>: New
field.
* p-lang.c (pascal_language_defn): Update.
* opencl-lang.c (opencl_language_defn): Update.
* objc-lang.c (objc_language_defn): Update.
* m2-lang.c (m2_language_defn): Update.
* language.h (struct language_defn) <la_get_compile_instance,
la_compute_program>: New fields.
* language.c (unknown_language_defn, auto_language_defn)
(local_language_defn): Update.
* jv-lang.c (java_language_defn): Update.
* go-lang.c (go_language_defn): Update.
* f-lang.c (f_language_defn): Update.
* dwarf2loc.h (dwarf2_compile_property_to_c): Declare.
* dwarf2loc.c (dwarf2_compile_property_to_c)
(locexpr_generate_c_location, loclist_generate_c_location): New
functions.
(dwarf2_locexpr_funcs, dwarf2_loclist_funcs): Update.
* defs.h (enum compile_i_scope_types): New.
(enum command_control_type) <compile_control>: New constant.
(struct command_line) <control_u>: New field.
* d-lang.c (d_language_defn): Update.
* compile/compile.c: New file.
* compile/compile-c-support.c: New file.
* compile/compile-c-symbols.c: New file.
* compile/compile-c-types.c: New file.
* compile/compile.h: New file.
* compile/compile-internal.h: New file.
* compile/compile-loc2c.c: New file.
* compile/compile-object-load.c: New file.
* compile/compile-object-load.h: New file.
* compile/compile-object-run.c: New file.
* compile/compile-object-run.h: New file.
* cli/cli-script.c (multi_line_command_p, print_command_lines)
(execute_control_command, process_next_line)
(recurse_read_control_structure): Handle compile_control.
* c-lang.h (c_get_compile_context, c_compute_program): Declare.
* c-lang.c (c_language_defn, cplus_language_defn)
(asm_language_defn, minimal_language_defn): Update.
* ada-lang.c (ada_language_defn): Update.
* Makefile.in (SUBDIR_GCC_COMPILE_OBS, SUBDIR_GCC_COMPILE_SRCS):
New variables.
(SFILES): Add SUBDIR_GCC_COMPILE_SRCS.
(HFILES_NO_SRCDIR): Add compile.h.
(COMMON_OBS): Add SUBDIR_GCC_COMPILE_OBS.
(INIT_FILES): Add SUBDIR_GCC_COMPILE_SRCS.
(compile.o, compile-c-types.o, compile-c-symbols.o)
(compile-object-load.o, compile-object-run.o, compile-loc2c.o)
(compile-c-support.o): New targets.
gdb/doc/ChangeLog
2014-12-12 Phil Muldoon <pmuldoon@redhat.com>
Jan Kratochvil <jan.kratochvil@redhat.com>
* gdb.texinfo (Altering): Update.
(Compiling and Injecting Code): New node.
gdb/testsuite/ChangeLog
2014-12-12 Phil Muldoon <pmuldoon@redhat.com>
Jan Kratochvil <jan.kratochvil@redhat.com>
Tom Tromey <tromey@redhat.com>
* configure.ac: Add gdb.compile/.
* configure: Regenerate.
* gdb.compile/Makefile.in: New file.
* gdb.compile/compile-ops.exp: New file.
* gdb.compile/compile-ops.c: New file.
* gdb.compile/compile-tls.c: New file.
* gdb.compile/compile-tls.exp: New file.
* gdb.compile/compile-constvar.S: New file.
* gdb.compile/compile-constvar.c: New file.
* gdb.compile/compile-mod.c: New file.
* gdb.compile/compile-nodebug.c: New file.
* gdb.compile/compile-setjmp-mod.c: New file.
* gdb.compile/compile-setjmp.c: New file.
* gdb.compile/compile-setjmp.exp: New file.
* gdb.compile/compile-shlib.c: New file.
* gdb.compile/compile.c: New file.
* gdb.compile/compile.exp: New file.
* lib/gdb.exp (skip_compile_feature_tests): New proc.
This exports a utility function, dwarf2_reg_to_regnum_or_error, that
was previously private to dwarf2loc.c.
gdb/ChangeLog
2014-12-12 Jan Kratochvil <jan.kratochvil@redhat.com>
* dwarf2loc.h (dwarf2_reg_to_regnum_or_error): Declare.
* dwarf2loc.c (dwarf2_reg_to_regnum_or_error): Rename from
translate_register. Now public.
(dwarf2_compile_expr_to_ax): Update.
This exports dwarf_expr_frame_base_1 so that other code can use it.
gdb/ChangeLog
2014-12-12 Tom Tromey <tromey@redhat.com>
Jan Kratochvil <jan.kratochvil@redhat.com>
* dwarf2loc.c (dwarf_expr_frame_base_1): Remove declaration.
(dwarf_expr_frame_base): Update caller.
(dwarf_expr_frame_base_1): Rename to ...
(func_get_frame_base_dwarf_block): ... this and make it public.
(dwarf2_compile_expr_to_ax, locexpr_describe_location_piece): Update
callers.
* dwarf2loc.h (func_get_frame_base_dwarf_block): New declaration.
This removes dwarf2_compile_expr_to_ax, replacing it with a utility
function that fetches the CFA data and adding the code to actually
compile to an agent expression directly into
dwarf2_compile_expr_to_ax. This refactoring lets a later patch reuse
the new dwarf2_fetch_cfa_info.
gdb/ChangeLog
2014-12-12 Tom Tromey <tromey@redhat.com>
* dwarf2loc.c (dwarf2_compile_expr_to_ax) <DW_OP_call_frame_cfa>:
Update.
* dwarf2-frame.c (dwarf2_fetch_cfa_info): New function, based on
dwarf2_compile_cfa_to_ax.
(dwarf2_compile_cfa_to_ax): Remove.
* dwarf2-frame.h (dwarf2_fetch_cfa_info): Declare.
(dwarf2_compile_cfa_to_ax): Remove.
1. Background information
The MIPS architecture, as originally designed and implemented in
mid-1980s has a uniform instruction word size that is 4 bytes, naturally
aligned. As such all MIPS instructions are located at addresses that
have their bits #1 and #0 set to zeroes, and any attempt to execute an
instruction from an address that has any of the two bits set to one
causes an address error exception. This may for example happen when a
jump-register instruction is executed whose register value used as the
jump target has any of these bits set.
Then in mid 1990s LSI sought a way to improve code density for their
TinyRISC family of MIPS cores and invented an alternatively encoded
instruction set in a joint effort with MIPS Technologies (then a
subsidiary of SGI). The new instruction set has been named the MIPS16
ASE (Application-Specific Extension) and uses a variable instruction
word size, which is 2 bytes (as the name of the ASE suggests) for most,
but there are a couple of exceptions that take 4 bytes, and then most of
the 2-byte instructions can be treated with a 2-byte extension prefix to
expand the range of the immediate operands used.
As a result instructions are no longer 4-byte aligned, instead they are
aligned to a multiple of 2. That left the bit #0 still unused for code
references, be it for the standard MIPS (i.e. as originally invented) or
for the MIPS16 instruction set, and based on that observation a clever
trick was invented that on one hand allowed the processor to be
seamlessly switched between the two instruction sets at any time at the
run time while on the other avoided the introduction of any special
control register to do that.
So it is the bit #0 of the instruction address that was chosen as the
selector and named the ISA bit. Any instruction executed at an even
address is interpreted as a standard MIPS instruction (the address still
has to have its bit #1 clear), any instruction executed at an odd
address is interpreted as a MIPS16 instruction.
To switch between modes ordinary jump instructions are used, such as
used for function calls and returns, specifically the bit #0 of the
source register used in jump-register instructions selects the execution
(ISA) mode for the following piece of code to be interpreted in.
Additionally new jump-immediate instructions were added that flipped the
ISA bit to select the opposite mode upon execution. They were
considered necessary to avoid the need to make register jumps in all
cases as the original jump-immediate instructions provided no way to
change the bit #0 at all.
This was all important for cases where standard MIPS and MIPS16 code had
to be mixed, either for compatibility with the existing binary code base
or to access resources not reachable from MIPS16 code (the MIPS16
instruction set only provides access to general-purpose registers, and
not for example floating-point unit registers or privileged coprocessor
0 registers) -- pieces of code in the opposite mode can be executed as
ordinary subroutine calls.
A similar approach has been more recently adopted for the MIPS16
replacement instruction set defined as the so called microMIPS ASE.
This is another instruction set encoding introduced to the MIPS
architecture. Just like the MIPS16 ASE, the microMIPS instruction set
uses a variable-length encoding, where each instruction takes a multiple
of 2 bytes. The ISA bit has been reused and for microMIPS-capable
processors selects between the standard MIPS and the microMIPS mode
instead.
2. Statement of the problem
To put it shortly, MIPS16 and microMIPS code pointers used by GDB are
different to these observed at the run time. This results in the same
expressions being evaluated producing different results in GDB and in
the program being debugged. Obviously it's the results obtained at the
run time that are correct (they define how the program behaves) and
therefore by definition the results obtained in GDB are incorrect.
A bit longer description will record that obviously at the run time the
ISA bit has to be set correctly (refer to background information above
if unsure why so) or the program will not run as expected. This is
recorded in all the executable file structures used at the run time: the
dynamic symbol table (but not always the static one!), the GOT, and
obviously in all the addresses embedded in code or data of the program
itself, calculated by applying the appropriate relocations at the static
link time.
While a program is being processed by GDB, the ISA bit is stripped off
from any code addresses, presumably to make them the same as the
respective raw memory byte address used by the processor to access the
instruction in the instruction fetch access cycle. This stripping is
actually performed outside GDB proper, in BFD, specifically
_bfd_mips_elf_symbol_processing (elfxx-mips.c, see the piece of code at
the very bottom of that function, starting with an: "If this is an
odd-valued function symbol, assume it's a MIPS16 or microMIPS one."
comment).
This function is also responsible for symbol table dumps made by
`objdump' too, so you'll never see the ISA bit reported there by that
tool, you need to use `readelf'.
This is however unlike what is ever done at the run time, the ISA bit
once present is never stripped off, for example a cast like this:
(short *) main
will not strip the ISA bit off and if the resulting pointer is intended
to be used to access instructions as data, for example for software
instruction decoding (like for fault recovery or emulation in a signal
handler) or for self-modifying code then the bit still has to be
stripped off by an explicit AND operation.
This is probably best illustrated with a simple real program example.
Let's consider the following simple program:
$ cat foobar.c
int __attribute__ ((mips16)) foo (void)
{
return 1;
}
int __attribute__ ((mips16)) bar (void)
{
return 2;
}
int __attribute__ ((nomips16)) foo32 (void)
{
return 3;
}
int (*foo32p) (void) = foo32;
int (*foop) (void) = foo;
int fooi = (int) foo;
int
main (void)
{
return foop ();
}
$
This is plain C with no odd tricks, except from the instruction mode
attributes. They are not necessary to trigger this problem, I just put
them here so that the program can be contained in a single source file
and to make it obvious which function is MIPS16 code and which is not.
Let's try it with Linux, so that everyone can repeat this experiment:
$ mips-linux-gnu-gcc -mips16 -g -O2 -o foobar foobar.c
$
Let's have a look at some interesting symbols:
$ mips-linux-gnu-readelf -s foobar | egrep 'table|foo|bar'
Symbol table '.dynsym' contains 7 entries:
Symbol table '.symtab' contains 95 entries:
55: 00000000 0 FILE LOCAL DEFAULT ABS foobar.c
66: 0040068c 4 FUNC GLOBAL DEFAULT [MIPS16] 12 bar
68: 00410848 4 OBJECT GLOBAL DEFAULT 21 foo32p
70: 00410844 4 OBJECT GLOBAL DEFAULT 21 foop
78: 00400684 8 FUNC GLOBAL DEFAULT 12 foo32
80: 00400680 4 FUNC GLOBAL DEFAULT [MIPS16] 12 foo
88: 00410840 4 OBJECT GLOBAL DEFAULT 21 fooi
$
Hmm, no sight of the ISA bit, but notice how foo and bar (but not
foo32!) have been marked as MIPS16 functions (ELF symbol structure's
`st_other' field is used for that).
So let's try to run and poke at this program with GDB. I'll be using a
native system for simplicity (I'll be using ellipses here and there to
remove unrelated clutter):
$ ./foobar
$ echo $?
1
$
So far, so good.
$ gdb ./foobar
[...]
(gdb) break main
Breakpoint 1 at 0x400490: file foobar.c, line 23.
(gdb) run
Starting program: .../foobar
Breakpoint 1, main () at foobar.c:23
23 return foop ();
(gdb)
Yay, it worked! OK, so let's poke at it:
(gdb) print main
$1 = {int (void)} 0x400490 <main>
(gdb) print foo32
$2 = {int (void)} 0x400684 <foo32>
(gdb) print foo32p
$3 = (int (*)(void)) 0x400684 <foo32>
(gdb) print bar
$4 = {int (void)} 0x40068c <bar>
(gdb) print foo
$5 = {int (void)} 0x400680 <foo>
(gdb) print foop
$6 = (int (*)(void)) 0x400681 <foo>
(gdb)
A-ha! Here's the difference and finally the ISA bit!
(gdb) print /x fooi
$7 = 0x400681
(gdb) p/x $pc
p/x $pc
$8 = 0x400491
(gdb)
And here as well...
(gdb) advance foo
foo () at foobar.c:4
4 }
(gdb) disassemble
Dump of assembler code for function foo:
0x00400680 <+0>: jr ra
0x00400682 <+2>: li v0,1
End of assembler dump.
(gdb) finish
Run till exit from #0 foo () at foobar.c:4
main () at foobar.c:24
24 }
Value returned is $9 = 1
(gdb) continue
Continuing.
[Inferior 1 (process 14103) exited with code 01]
(gdb)
So let's be a bit inquisitive...
(gdb) run
Starting program: .../foobar
Breakpoint 1, main () at foobar.c:23
23 return foop ();
(gdb)
Actually we do not like to run foo here at all. Let's run bar instead!
(gdb) set foop = bar
(gdb) print foop
$10 = (int (*)(void)) 0x40068c <bar>
(gdb)
Hmm, no ISA bit. Is it going to work?
(gdb) advance bar
bar () at foobar.c:9
9 }
(gdb) p/x $pc
$11 = 0x40068c
(gdb) disassemble
Dump of assembler code for function bar:
=> 0x0040068c <+0>: jr ra
0x0040068e <+2>: li v0,2
End of assembler dump.
(gdb) finish
Run till exit from #0 bar () at foobar.c:9
Program received signal SIGILL, Illegal instruction.
bar () at foobar.c:9
9 }
(gdb)
Oops!
(gdb) p/x $pc
$12 = 0x40068c
(gdb)
We're still there!
(gdb) continue
Continuing.
Program terminated with signal SIGILL, Illegal instruction.
The program no longer exists.
(gdb)
So let's try something else:
(gdb) run
Starting program: .../foobar
Breakpoint 1, main () at foobar.c:23
23 return foop ();
(gdb) set foop = foo
(gdb) advance foo
foo () at foobar.c:4
4 }
(gdb) disassemble
Dump of assembler code for function foo:
=> 0x00400680 <+0>: jr ra
0x00400682 <+2>: li v0,1
End of assembler dump.
(gdb) finish
Run till exit from #0 foo () at foobar.c:4
Program received signal SIGILL, Illegal instruction.
foo () at foobar.c:4
4 }
(gdb) continue
Continuing.
Program terminated with signal SIGILL, Illegal instruction.
The program no longer exists.
(gdb)
The same problem!
(gdb) run
Starting program:
/net/build2-lucid-cs/scratch/macro/mips-linux-fsf-gcc/isa-bit/foobar
Breakpoint 1, main () at foobar.c:23
23 return foop ();
(gdb) set foop = foo32
(gdb) advance foo32
foo32 () at foobar.c:14
14 }
(gdb) disassemble
Dump of assembler code for function foo32:
=> 0x00400684 <+0>: jr ra
0x00400688 <+4>: li v0,3
End of assembler dump.
(gdb) finish
Run till exit from #0 foo32 () at foobar.c:14
main () at foobar.c:24
24 }
Value returned is $14 = 3
(gdb) continue
Continuing.
[Inferior 1 (process 14113) exited with code 03]
(gdb)
That did work though, so it's the ISA bit only!
(gdb) quit
Enough!
That's the tip of the iceberg only though. So let's rebuild the
executable with some dynamic symbols:
$ mips-linux-gnu-gcc -mips16 -Wl,--export-dynamic -g -O2 -o foobar-dyn foobar.c
$ mips-linux-gnu-readelf -s foobar-dyn | egrep 'table|foo|bar'
Symbol table '.dynsym' contains 32 entries:
6: 004009cd 4 FUNC GLOBAL DEFAULT 12 bar
8: 00410b88 4 OBJECT GLOBAL DEFAULT 21 foo32p
9: 00410b84 4 OBJECT GLOBAL DEFAULT 21 foop
15: 004009c4 8 FUNC GLOBAL DEFAULT 12 foo32
17: 004009c1 4 FUNC GLOBAL DEFAULT 12 foo
25: 00410b80 4 OBJECT GLOBAL DEFAULT 21 fooi
Symbol table '.symtab' contains 95 entries:
55: 00000000 0 FILE LOCAL DEFAULT ABS foobar.c
69: 004009cd 4 FUNC GLOBAL DEFAULT 12 bar
71: 00410b88 4 OBJECT GLOBAL DEFAULT 21 foo32p
72: 00410b84 4 OBJECT GLOBAL DEFAULT 21 foop
79: 004009c4 8 FUNC GLOBAL DEFAULT 12 foo32
81: 004009c1 4 FUNC GLOBAL DEFAULT 12 foo
89: 00410b80 4 OBJECT GLOBAL DEFAULT 21 fooi
$
OK, now the ISA bit is there for a change, but the MIPS16 `st_other'
attribute gone, hmm... What does `objdump' do then:
$ mips-linux-gnu-objdump -Tt foobar-dyn | egrep 'SYMBOL|foo|bar'
foobar-dyn: file format elf32-tradbigmips
SYMBOL TABLE:
00000000 l df *ABS* 00000000 foobar.c
004009cc g F .text 00000004 0xf0 bar
00410b88 g O .data 00000004 foo32p
00410b84 g O .data 00000004 foop
004009c4 g F .text 00000008 foo32
004009c0 g F .text 00000004 0xf0 foo
00410b80 g O .data 00000004 fooi
DYNAMIC SYMBOL TABLE:
004009cc g DF .text 00000004 Base 0xf0 bar
00410b88 g DO .data 00000004 Base foo32p
00410b84 g DO .data 00000004 Base foop
004009c4 g DF .text 00000008 Base foo32
004009c0 g DF .text 00000004 Base 0xf0 foo
00410b80 g DO .data 00000004 Base fooi
$
Hmm, the attribute (0xf0, printed raw) is back, and the ISA bit gone
again.
Let's have a look at some DWARF-2 records GDB uses (I'll be stripping
off a lot here for brevity) -- debug info:
$ mips-linux-gnu-readelf -wi foobar
Contents of the .debug_info section:
[...]
Compilation Unit @ offset 0x88:
Length: 0xbb (32-bit)
Version: 4
Abbrev Offset: 62
Pointer Size: 4
<0><93>: Abbrev Number: 1 (DW_TAG_compile_unit)
<94> DW_AT_producer : (indirect string, offset: 0x19e): GNU C 4.8.0 20120513 (experimental) -meb -mips16 -march=mips32r2 -mhard-float -mllsc -mplt -mno-synci -mno-shared -mabi=32 -g -O2
<98> DW_AT_language : 1 (ANSI C)
<99> DW_AT_name : (indirect string, offset: 0x190): foobar.c
<9d> DW_AT_comp_dir : (indirect string, offset: 0x225): [...]
<a1> DW_AT_ranges : 0x0
<a5> DW_AT_low_pc : 0x0
<a9> DW_AT_stmt_list : 0x27
<1><ad>: Abbrev Number: 2 (DW_TAG_subprogram)
<ae> DW_AT_external : 1
<ae> DW_AT_name : foo
<b2> DW_AT_decl_file : 1
<b3> DW_AT_decl_line : 1
<b4> DW_AT_prototyped : 1
<b4> DW_AT_type : <0xc2>
<b8> DW_AT_low_pc : 0x400680
<bc> DW_AT_high_pc : 0x400684
<c0> DW_AT_frame_base : 1 byte block: 9c (DW_OP_call_frame_cfa)
<c2> DW_AT_GNU_all_call_sites: 1
<1><c2>: Abbrev Number: 3 (DW_TAG_base_type)
<c3> DW_AT_byte_size : 4
<c4> DW_AT_encoding : 5 (signed)
<c5> DW_AT_name : int
<1><c9>: Abbrev Number: 4 (DW_TAG_subprogram)
<ca> DW_AT_external : 1
<ca> DW_AT_name : (indirect string, offset: 0x18a): foo32
<ce> DW_AT_decl_file : 1
<cf> DW_AT_decl_line : 11
<d0> DW_AT_prototyped : 1
<d0> DW_AT_type : <0xc2>
<d4> DW_AT_low_pc : 0x400684
<d8> DW_AT_high_pc : 0x40068c
<dc> DW_AT_frame_base : 1 byte block: 9c (DW_OP_call_frame_cfa)
<de> DW_AT_GNU_all_call_sites: 1
<1><de>: Abbrev Number: 2 (DW_TAG_subprogram)
<df> DW_AT_external : 1
<df> DW_AT_name : bar
<e3> DW_AT_decl_file : 1
<e4> DW_AT_decl_line : 6
<e5> DW_AT_prototyped : 1
<e5> DW_AT_type : <0xc2>
<e9> DW_AT_low_pc : 0x40068c
<ed> DW_AT_high_pc : 0x400690
<f1> DW_AT_frame_base : 1 byte block: 9c (DW_OP_call_frame_cfa)
<f3> DW_AT_GNU_all_call_sites: 1
<1><f3>: Abbrev Number: 5 (DW_TAG_subprogram)
<f4> DW_AT_external : 1
<f4> DW_AT_name : (indirect string, offset: 0x199): main
<f8> DW_AT_decl_file : 1
<f9> DW_AT_decl_line : 21
<fa> DW_AT_prototyped : 1
<fa> DW_AT_type : <0xc2>
<fe> DW_AT_low_pc : 0x400490
<102> DW_AT_high_pc : 0x4004a4
<106> DW_AT_frame_base : 1 byte block: 9c (DW_OP_call_frame_cfa)
<108> DW_AT_GNU_all_tail_call_sites: 1
[...]
$
-- no sign of the ISA bit anywhere -- frame info:
$ mips-linux-gnu-readelf -wf foobar
[...]
Contents of the .debug_frame section:
00000000 0000000c ffffffff CIE
Version: 1
Augmentation: ""
Code alignment factor: 1
Data alignment factor: -4
Return address column: 31
DW_CFA_def_cfa_register: r29
DW_CFA_nop
00000010 0000000c 00000000 FDE cie=00000000 pc=00400680..00400684
00000020 0000000c 00000000 FDE cie=00000000 pc=00400684..0040068c
00000030 0000000c 00000000 FDE cie=00000000 pc=0040068c..00400690
00000040 00000018 00000000 FDE cie=00000000 pc=00400490..004004a4
DW_CFA_advance_loc: 6 to 00400496
DW_CFA_def_cfa_offset: 32
DW_CFA_offset: r31 at cfa-4
DW_CFA_advance_loc: 6 to 0040049c
DW_CFA_restore: r31
DW_CFA_def_cfa_offset: 0
DW_CFA_nop
DW_CFA_nop
DW_CFA_nop
[...]
$
-- no sign of the ISA bit anywhere -- range info (GDB doesn't use arange):
$ mips-linux-gnu-readelf -wR foobar
Contents of the .debug_ranges section:
Offset Begin End
00000000 00400680 00400690
00000000 00400490 004004a4
00000000 <End of list>
$
-- no sign of the ISA bit anywhere -- line info:
$ mips-linux-gnu-readelf -wl foobar
Raw dump of debug contents of section .debug_line:
[...]
Offset: 0x27
Length: 78
DWARF Version: 2
Prologue Length: 31
Minimum Instruction Length: 1
Initial value of 'is_stmt': 1
Line Base: -5
Line Range: 14
Opcode Base: 13
Opcodes:
Opcode 1 has 0 args
Opcode 2 has 1 args
Opcode 3 has 1 args
Opcode 4 has 1 args
Opcode 5 has 1 args
Opcode 6 has 0 args
Opcode 7 has 0 args
Opcode 8 has 0 args
Opcode 9 has 1 args
Opcode 10 has 0 args
Opcode 11 has 0 args
Opcode 12 has 1 args
The Directory Table is empty.
The File Name Table:
Entry Dir Time Size Name
1 0 0 0 foobar.c
Line Number Statements:
Extended opcode 2: set Address to 0x400681
Special opcode 6: advance Address by 0 to 0x400681 and Line by 1 to 2
Special opcode 7: advance Address by 0 to 0x400681 and Line by 2 to 4
Special opcode 55: advance Address by 3 to 0x400684 and Line by 8 to 12
Special opcode 7: advance Address by 0 to 0x400684 and Line by 2 to 14
Advance Line by -7 to 7
Special opcode 131: advance Address by 9 to 0x40068d and Line by 0 to 7
Special opcode 7: advance Address by 0 to 0x40068d and Line by 2 to 9
Advance PC by 3 to 0x400690
Extended opcode 1: End of Sequence
Extended opcode 2: set Address to 0x400491
Advance Line by 21 to 22
Copy
Special opcode 6: advance Address by 0 to 0x400491 and Line by 1 to 23
Special opcode 60: advance Address by 4 to 0x400495 and Line by -1 to 22
Special opcode 34: advance Address by 2 to 0x400497 and Line by 1 to 23
Special opcode 62: advance Address by 4 to 0x40049b and Line by 1 to 24
Special opcode 32: advance Address by 2 to 0x40049d and Line by -1 to 23
Special opcode 6: advance Address by 0 to 0x40049d and Line by 1 to 24
Advance PC by 7 to 0x4004a4
Extended opcode 1: End of Sequence
[...]
-- a-ha, the ISA bit is there! However it's not always right for some
reason, I don't have a small test case to show it, but here's an excerpt
from MIPS16 libc, a prologue of a function:
00019630 <__libc_init_first>:
19630: e8a0 jrc ra
19632: 6500 nop
00019634 <_init>:
19634: f000 6a11 li v0,17
19638: f7d8 0b08 la v1,15e00 <_DYNAMIC+0x15c54>
1963c: f400 3240 sll v0,16
19640: e269 addu v0,v1
19642: 659a move gp,v0
19644: 64f6 save 48,ra,s0-s1
19646: 671c move s0,gp
19648: d204 sw v0,16(sp)
1964a: f352 984c lw v0,-27828(s0)
1964e: 6724 move s1,a0
and the corresponding DWARF-2 line info:
Line Number Statements:
Extended opcode 2: set Address to 0x19631
Advance Line by 44 to 45
Copy
Special opcode 8: advance Address by 0 to 0x19631 and Line by 3 to 48
Special opcode 66: advance Address by 4 to 0x19635 and Line by 5 to 53
Advance PC by constant 17 to 0x19646
Special opcode 25: advance Address by 1 to 0x19647 and Line by 6 to 59
Advance Line by -6 to 53
Special opcode 33: advance Address by 2 to 0x19649 and Line by 0 to 53
Special opcode 39: advance Address by 2 to 0x1964b and Line by 6 to 59
Advance Line by -6 to 53
Special opcode 61: advance Address by 4 to 0x1964f and Line by 0 to 53
-- see that "Advance PC by constant 17" there? It clears the ISA bit,
however code at 0x19646 is not standard MIPS code at all. For some
reason the constant is always 17, I've never seen DW_LNS_const_add_pc
used with any other value -- is that a binutils bug or what?
3. Solution:
I think we should retain the value of the ISA bit in code references,
that is effectively treat them as cookies as they indeed are (although
trivially calculated) rather than raw memory byte addresses.
In a perfect world both the static symbol table and the respective
DWARF-2 records should be fixed to include the ISA bit in all the cases.
I think however that this is infeasible.
All the uses of `_bfd_mips_elf_symbol_processing' can not necessarily be
tracked down. This function is used by `elf_slurp_symbol_table' that in
turn is used by `bfd_canonicalize_symtab' and
`bfd_canonicalize_dynamic_symtab', which are public interfaces.
Similarly DWARF-2 records are used outside GDB, one notable if a bit
questionable is the exception unwinder (libgcc/unwind-dw2.c) -- I have
identified at least bits in `execute_cfa_program' and
`uw_frame_state_for', both around the calls to `_Unwind_IsSignalFrame',
that would need an update as they effectively flip the ISA bit freely;
see also the comment about MASK_RETURN_ADDR in gcc/config/mips/mips.h.
But there may be more places. Any change in how DWARF-2 records are
produced would require an update there and would cause compatibility
problems with libgcc.a binaries already distributed; given that this is
a static library a complex change involving function renames would
likely be required.
I propose therefore to accept the existing inconsistencies and deal with
them entirely within GDB. I have figured out that the ISA bit lost in
various places can still be recovered as long as we have symbol
information -- that'll have the `st_other' attribute correctly set to
one of standard MIPS/MIPS16/microMIPS encoding.
Here's the resulting change. It adds a couple of new `gdbarch' hooks,
one to update symbol information with the ISA bit lost in
`_bfd_mips_elf_symbol_processing', and two other ones to adjust DWARF-2
records as they're processed. The ISA bit is set in each address
handled according to information retrieved from the symbol table for the
symbol spanning the address if any; limits are adjusted based on the
address they point to related to the respective base address.
Additionally minimal symbol information has to be adjusted accordingly
in its gdbarch hook.
With these changes in place some complications with ISA bit juggling in
the PC that never fully worked can be removed from the MIPS backend.
Conversely, the generic dynamic linker event special breakpoint symbol
handler has to be updated to call the minimal symbol gdbarch hook to
record that the symbol is a MIPS16 or microMIPS address if applicable or
the breakpoint will be set at the wrong address and either fail to work
or cause SIGTRAPs (this is because the symbol is handled early on and
bypasses regular symbol processing).
4. Results obtained
The change fixes the example above -- to repeat only the crucial steps:
(gdb) break main
Breakpoint 1 at 0x400491: file foobar.c, line 23.
(gdb) run
Starting program: .../foobar
Breakpoint 1, main () at foobar.c:23
23 return foop ();
(gdb) print foo
$1 = {int (void)} 0x400681 <foo>
(gdb) set foop = bar
(gdb) advance bar
bar () at foobar.c:9
9 }
(gdb) disassemble
Dump of assembler code for function bar:
=> 0x0040068d <+0>: jr ra
0x0040068f <+2>: li v0,2
End of assembler dump.
(gdb) finish
Run till exit from #0 bar () at foobar.c:9
main () at foobar.c:24
24 }
Value returned is $2 = 2
(gdb) continue
Continuing.
[Inferior 1 (process 14128) exited with code 02]
(gdb)
-- excellent!
The change removes about 90 failures per MIPS16 multilib in mips-sde-elf
testing too, results for MIPS16 are now similar to that for standard
MIPS; microMIPS results are a bit worse because of host-I/O problems in
QEMU used instead of MIPSsim for microMIPS testing only:
=== gdb Summary ===
# of expected passes 14299
# of unexpected failures 187
# of expected failures 56
# of known failures 58
# of unresolved testcases 11
# of untested testcases 52
# of unsupported tests 174
MIPS16:
=== gdb Summary ===
# of expected passes 14298
# of unexpected failures 187
# of unexpected successes 2
# of expected failures 54
# of known failures 58
# of unresolved testcases 12
# of untested testcases 52
# of unsupported tests 174
microMIPS:
=== gdb Summary ===
# of expected passes 14149
# of unexpected failures 201
# of unexpected successes 2
# of expected failures 54
# of known failures 58
# of unresolved testcases 7
# of untested testcases 53
# of unsupported tests 175
2014-12-12 Maciej W. Rozycki <macro@codesourcery.com>
Maciej W. Rozycki <macro@mips.com>
Pedro Alves <pedro@codesourcery.com>
gdb/
* gdbarch.sh (elf_make_msymbol_special): Change type to `F',
remove `predefault' and `invalid_p' initializers.
(make_symbol_special): New architecture method.
(adjust_dwarf2_addr, adjust_dwarf2_line): Likewise.
(objfile, symbol): New declarations.
* arch-utils.h (default_elf_make_msymbol_special): Remove
prototype.
(default_make_symbol_special): New prototype.
(default_adjust_dwarf2_addr): Likewise.
(default_adjust_dwarf2_line): Likewise.
* mips-tdep.h (mips_unmake_compact_addr): New prototype.
* arch-utils.c (default_elf_make_msymbol_special): Remove
function.
(default_make_symbol_special): New function.
(default_adjust_dwarf2_addr): Likewise.
(default_adjust_dwarf2_line): Likewise.
* dwarf2-frame.c (decode_frame_entry_1): Call
`gdbarch_adjust_dwarf2_addr'.
* dwarf2loc.c (dwarf2_find_location_expression): Likewise.
* dwarf2read.c (create_addrmap_from_index): Likewise.
(process_psymtab_comp_unit_reader): Likewise.
(add_partial_symbol): Likewise.
(add_partial_subprogram): Likewise.
(process_full_comp_unit): Likewise.
(read_file_scope): Likewise.
(read_func_scope): Likewise. Call `gdbarch_make_symbol_special'.
(read_lexical_block_scope): Call `gdbarch_adjust_dwarf2_addr'.
(read_call_site_scope): Likewise.
(dwarf2_ranges_read): Likewise.
(dwarf2_record_block_ranges): Likewise.
(read_attribute_value): Likewise.
(dwarf_decode_lines_1): Call `gdbarch_adjust_dwarf2_line'.
(new_symbol_full): Call `gdbarch_adjust_dwarf2_addr'.
* elfread.c (elf_symtab_read): Don't call
`gdbarch_elf_make_msymbol_special' if unset.
* mips-linux-tdep.c (micromips_linux_sigframe_validate): Strip
the ISA bit from the PC.
* mips-tdep.c (mips_unmake_compact_addr): New function.
(mips_elf_make_msymbol_special): Set the ISA bit in the symbol's
address appropriately.
(mips_make_symbol_special): New function.
(mips_pc_is_mips): Set the ISA bit before symbol lookup.
(mips_pc_is_mips16): Likewise.
(mips_pc_is_micromips): Likewise.
(mips_pc_isa): Likewise.
(mips_adjust_dwarf2_addr): New function.
(mips_adjust_dwarf2_line): Likewise.
(mips_read_pc, mips_unwind_pc): Keep the ISA bit.
(mips_addr_bits_remove): Likewise.
(mips_skip_trampoline_code): Likewise.
(mips_write_pc): Don't set the ISA bit.
(mips_eabi_push_dummy_call): Likewise.
(mips_o64_push_dummy_call): Likewise.
(mips_gdbarch_init): Install `mips_make_symbol_special',
`mips_adjust_dwarf2_addr' and `mips_adjust_dwarf2_line' gdbarch
handlers.
* solib.c (gdb_bfd_lookup_symbol_from_symtab): Get
target-specific symbol address adjustments.
* gdbarch.h: Regenerate.
* gdbarch.c: Regenerate.
2014-12-12 Maciej W. Rozycki <macro@codesourcery.com>
gdb/testsuite/
* gdb.base/func-ptrs.c: New file.
* gdb.base/func-ptrs.exp: New file.
During armv7b testing gdb.base/store.exp test was failling with
'GDB internal error' with the following message:
Temporary breakpoint 1, wack_double (u=
../../binutils-gdb/gdb/regcache.c:177: internal-error: register_size: Assertion `regnum >= 0 && regnum < (gdbarch_num_regs (gdbarch) + gdbarch_num_pseudo_regs (gdbarch))' failed.
A problem internal to GDB has been detected,
further debugging may prove unreliable.
It turns out that compiler generated DWARF with non-existent
register numbers. The compiler issue is present in both little endian
(armv7) and big endian (armv7b) (it is separate issue). Here is
example for one of formal parameters of wack_double function:
<2><792>: Abbrev Number: 10 (DW_TAG_formal_parameter)
<793> DW_AT_name : u
<795> DW_AT_decl_file : 1
<796> DW_AT_decl_line : 115
<797> DW_AT_type : <0x57c>
<79b> DW_AT_location : 6 byte block: 6d 93 4 6c 93 4 (DW_OP_reg29 (r29); DW_OP_piece: 4; DW_OP_reg28 (r28); DW_OP_piece: 4)
In both big and little endian cases gdbarch_dwarf2_reg_to_regnum
returns -1 which is stored into gdb_regnum. But it causes severe
problem only in big endian case because in read_pieced_value and
write_pieced_value functions BFD_ENDIAN_BIG related processing
happen regardless of gdb_regnum value, for example register_size
function is called and in case of gdb_regnum=-1, it cause
'GDB internal error' and crash.
Solution is to move BFD_ENDIAN_BIG related processing under
(gdb_regnum != -1) branch of processing.
gdb/ChangeLog:
2014-11-02 Victor Kamensky <victor.kamensky@linaro.org>
* dwarf2loc.c (read_pieced_value): Do big endian
processing only if gdb_regnum is not -1.
(write_pieced_value): Ditto.
This fixes PR symtab/14604, PR symtab/14605, and Jan's test at
https://sourceware.org/ml/gdb-patches/2014-07/msg00158.html, in a tree
with bddbbed reverted:
2014-07-22 Pedro Alves <palves@redhat.com>
* value.c (allocate_optimized_out_value): Don't mark value as
non-lazy.
The PRs are about variables described by the DWARF as being split over
multiple registers using DWARF piece information, but some of those
registers being marked as optimised out (not saved) by a later frame.
GDB currently incorrectly mishandles these partially-optimized-out
values.
Even though we can usually tell from the debug info whether a local or
global is optimized out, handling the case of a local living in a
register that was not saved in a frame requires fetching the variable.
GDB also needs to fetch a value to tell whether parts of it are
"<unavailable>". Given this, it's not worth it to try to avoid
fetching lazy optimized-out values based on debug info alone.
So this patch makes GDB track which chunks of a value's contents are
optimized out like it tracks <unavailable> contents. That is, it
makes value->optimized_out be a bit range vector instead of a boolean,
and removes the struct lval_funcs check_validity and check_any_valid
hooks.
Unlike Andrew's series which this is based on (at
https://sourceware.org/ml/gdb-patches/2013-08/msg00300.html, note some
pieces have gone in since), this doesn't merge optimized out and
unavailable contents validity/availability behind a single interface,
nor does it merge the bit range vectors themselves (at least yet).
While it may be desirable to have a single entry point that returns
existence of contents irrespective of what may make them
invalid/unavailable, several places want to treat optimized out /
unavailable / etc. differently, so each spot that potentially could
use it will need to be careful considered on case-by-case basis, and
best done as a separate change.
This fixes Jan's test, because value_available_contents_eq wasn't
considering optimized out value contents. It does now, and because of
that it's been renamed to value_contents_eq.
A new intro comment is added to value.h describing "<optimized out>",
"<not saved>" and "<unavailable>" values.
gdb/
PR symtab/14604
PR symtab/14605
* ada-lang.c (coerce_unspec_val_to_type): Use
value_contents_copy_raw.
* ada-valprint.c (val_print_packed_array_elements): Adjust.
* c-valprint.c (c_val_print): Use value_bits_any_optimized_out.
* cp-valprint.c (cp_print_value_fields): Let the common printing
code handle optimized out values.
(cp_print_value_fields_rtti): Use value_bits_any_optimized_out.
* d-valprint.c (dynamic_array_type): Use
value_bits_any_optimized_out.
* dwarf2loc.c (entry_data_value_funcs): Remove check_validity and
check_any_valid fields.
(check_pieced_value_bits): Delete and inline ...
(check_pieced_synthetic_pointer): ... here.
(check_pieced_value_validity): Delete.
(check_pieced_value_invalid): Delete.
(pieced_value_funcs): Remove check_validity and check_any_valid
fields.
(read_pieced_value): Use mark_value_bits_optimized_out.
(write_pieced_value): Switch to use
mark_value_bytes_optimized_out.
(dwarf2_evaluate_loc_desc_full): Copy the value contents instead
of assuming the whole value is optimized out.
* findvar.c (read_frame_register_value): Remove special handling
of optimized out registers.
(value_from_register): Use mark_value_bytes_optimized_out.
* frame-unwind.c (frame_unwind_got_optimized): Use
mark_value_bytes_optimized_out.
* jv-valprint.c (java_value_print): Adjust.
(java_print_value_fields): Let the common printing code handle
optimized out values.
* mips-tdep.c (mips_print_register): Remove special handling of
optimized out registers.
* opencl-lang.c (lval_func_check_validity): Delete.
(lval_func_check_any_valid): Delete.
(opencl_value_funcs): Remove check_validity and check_any_valid
fields.
* p-valprint.c (pascal_object_print_value_fields): Let the common
printing code handle optimized out values.
* stack.c (read_frame_arg): Remove special handling of optimized
out values. Fetch both VAL and ENTRYVAL before comparing
contents. Adjust to value_available_contents_eq rename.
* valprint.c (valprint_check_validity)
(val_print_scalar_formatted): Use value_bits_any_optimized_out.
(val_print_array_elements): Adjust.
* value.c (struct value) <optimized_out>: Now a VEC(range_s).
(value_bits_any_optimized_out): New function.
(value_entirely_covered_by_range_vector): New function, factored
out from value_entirely_unavailable.
(value_entirely_unavailable): Reimplement.
(value_entirely_optimized_out): New function.
(insert_into_bit_range_vector): New function, factored out from
mark_value_bits_unavailable.
(mark_value_bits_unavailable): Reimplement.
(struct ranges_and_idx): New struct.
(find_first_range_overlap_and_match): New function, factored out
from value_available_contents_bits_eq.
(value_available_contents_bits_eq): Rename to ...
(value_contents_bits_eq): ... this. Check both unavailable
contents and optimized out contents.
(value_available_contents_eq): Rename to ...
(value_contents_eq): ... this.
(allocate_value_lazy): Remove reference to the old optimized_out
boolean.
(allocate_optimized_out_value): Use
mark_value_bytes_optimized_out.
(require_not_optimized_out): Adjust to check whether the
optimized_out vec is empty.
(ranges_copy_adjusted): New function, factored out from
value_contents_copy_raw.
(value_contents_copy_raw): Also copy the optimized out ranges.
Assert the destination ranges aren't optimized out.
(value_contents_copy): Update comment, remove call to
require_not_optimized_out.
(value_contents_equal): Adjust to check whether the optimized_out
vec is empty.
(set_value_optimized_out, value_optimized_out_const): Delete.
(mark_value_bytes_optimized_out, mark_value_bits_optimized_out):
New functions.
(value_entirely_optimized_out, value_bits_valid): Delete.
(value_copy): Take a VEC copy of the 'optimized_out' field.
(value_primitive_field): Remove special handling of optimized out.
(value_fetch_lazy): Assert that lazy values have no unavailable
regions. Use value_bits_any_optimized_out. Remove some special
handling for optimized out values.
* value.h: Add intro comment about <optimized out> and
<unavailable>.
(struct lval_funcs): Remove check_validity and check_any_valid
fields.
(set_value_optimized_out, value_optimized_out_const): Remove.
(mark_value_bytes_optimized_out, mark_value_bits_optimized_out):
New declarations.
(value_bits_any_optimized_out): New declaration.
(value_bits_valid): Delete declaration.
(value_available_contents_eq): Rename to ...
(value_contents_eq): ... this, and extend comments.
gdb/testsuite/
PR symtab/14604
PR symtab/14605
* gdb.dwarf2/dw2-op-out-param.exp: Remove kfail branches and use
gdb_test.
I cannot reproduce any wrong case having the code removed.
I just do not find it correct to have it disabled. But at the same time I do
like much / I do not find correct the code myself. It is a bit problematic to
have struct value describing a memory content which is no longer present
there.
What happens there:
------------------------------------------------------------------------------
volatile int vv;
static __attribute__((noinline)) int
bar (int &ref) {
ref = 20;
vv++; /* break-here */
return ref;
}
int main (void) {
int var = 10;
return bar (var);
}
------------------------------------------------------------------------------
<4><c7>: Abbrev Number: 13 (DW_TAG_GNU_call_site_parameter)
<c8> DW_AT_location : 1 byte block: 55 (DW_OP_reg5 (rdi))
<ca> DW_AT_GNU_call_site_value: 2 byte block: 91 74 (DW_OP_fbreg: -12)
<cd> DW_AT_GNU_call_site_data_value: 1 byte block: 3a (DW_OP_lit10)
------------------------------------------------------------------------------
gdb -ex 'b value_addr' -ex r --args ../gdb ./1 -ex 'watch vv' -ex r -ex 'p &ref@entry'
->
6 return ref;
bar (ref=@0x7fffffffd944: 20, ref@entry=@0x7fffffffd944: 10) at 1.C:25
------------------------------------------------------------------------------
At /* break-here */ struct value variable 'ref' is TYPE_CODE_REF.
With FSF GDB HEAD:
(gdb) x/gx arg1.contents
0x6004000a4ad0: 0x00007fffffffd944
(gdb) p ((struct value *)arg1.location.computed.closure).lval
$1 = lval_memory
(gdb) p/x ((struct value *)arg1.location.computed.closure).location.address
$3 = 0x7fffffffd944
With your #if0-ed code:
(gdb) x/gx arg1.contents
0x6004000a4ad0: 0x00007fffffffd944
(gdb) p ((struct value *)arg1.location.computed.closure).lval
$8 = not_lval
(gdb) p/x ((struct value *)arg1.location.computed.closure).location.address
$9 = 0x0
I do not see how to access
((struct value *)arg1.location.computed.closure).location.address
from GDB CLI. Trying
(gdb) p &ref@entry
will invoke value_addr()'s:
if (TYPE_CODE (type) == TYPE_CODE_REF)
/* Copy the value, but change the type from (T&) to (T*). We
keep the same location information, which is efficient, and
allows &(&X) to get the location containing the reference. */
and therefore the address gets fetched already from
arg1.contents
and not from
((struct value *)arg1.location.computed.closure).location.address
.
And for any other type than TYPE_CODE_REF this code you removed does not get
executed at all. This DW_AT_GNU_call_site_data_value DWARF was meant
primarily for Fortran but with -O0 entry values do not get produced
and with -Og and higher Fortran always optimizes out the passing by reference.
If you do not like the removed code there I am OK with removing it as I do not
know how to make it's use reproducible for user anyway. In the worst case
- if there really is some way how to exploit it - one should just get
Attempt to take address of value not located in memory.
instead of some wrong value and it may be easy to fix then.
gdb/
2014-07-22 Jan Kratochvil <jan.kratochvil@redhat.com>
* dwarf2loc.c (value_of_dwarf_reg_entry): Remove setting value address
for reference entry value target data value.
Message-ID: <20140720150727.GA18488@host2.jankratochvil.net>
This patch fixes a problem that prevented use of the Dwarf unwinders on SPU,
because dwarf2-frame.c common code did not support the situation where the
stack and/or frame pointer is maintained in a *vector* register. This is
because read_addr_from_reg is hard-coded to assume that such pointers can
be read from registers via a simple get_frame_register / unpack_pointer
operation.
Now, there *is* a routine address_from_register that calls into the
appropriate tdep routines to handle pointer values in "weird" registers
like on SPU, but it turns out I cannot simply change dwarf2-frame.c to
use address_from_register. This is because address_from_register uses
value_from_register to create a (temporary) value, and that routine
at some point calls get_frame_id in order to set up that value's
VALUE_FRAME_ID entry.
However, the dwarf2-frame.c read_addr_from_reg routine will be called
during early unwinding (to unwind the frame's CFA), at which point the
frame's ID is not actually known yet! This would cause an assert.
On the other hand, we may notice that VALUE_FRAME_ID is only needed in the
value returned by value_from_register if that value is later used as an
lvalue. But this is obviously never done to the temporary value used in
address_from_register. So, if we could change address_from_register to
not call value_from_register but instead accept constructing a value
that doesn't have VALUE_FRAME_ID set, things should be fine.
To do that, we can change the value_from_register callback to accept
a FRAME_ID instead of a FRAME; the only existing uses of the FRAME
argument were either to extract its frame ID, or its gdbarch. (To
keep a way of getting at the latter, we also change the callback's
type from "f" to "m".) Together with the required follow-on changes
in the existing value_from_register implementations (including the
default one), this seems to fix the problem.
As another minor interface cleanup, I've removed the explicit TYPE
argument from address_from_register. This routine really always
uses a default pointer type, and in the new implementation it -to
some extent- relies on that fact, in that it will now no longer
handle types that require gdbarch_convert_register_p handling.
gdb:
2014-04-17 Ulrich Weigand <uweigand@de.ibm.com>
* gdbarch.sh (value_from_register): Make class "m" instead of "f".
Replace FRAME argument with FRAME_ID.
* gdbarch.c, gdbarch.h: Regenerate.
* findvar.c (default_value_from_register): Add GDBARCH argument;
replace FRAME by FRAME_ID. No longer call get_frame_id.
(value_from_register): Update call to gdbarch_value_from_register.
* value.h (default_value_from_register): Update prototype.
* s390-linux-tdep.c (s390_value_from_register): Update interface
and call to default_value_from_register.
* spu-tdep.c (spu_value_from_register): Likewise.
* findvar.c (address_from_register): Remove TYPE argument.
Do not call value_from_register; use gdbarch_value_from_register
with null_frame_id instead.
* value.h (address_from_register): Update prototype.
* dwarf2-frame.c (read_addr_from_reg): Use address_from_register.
* dwarf2loc.c (dwarf_expr_read_addr_from_reg): Update for
address_from_register interface change.
The dwarf standard allow certain attributes to be expressed as dwarf
expressions rather than constants. For instance upper-/lowerbound attributes.
In case of a c99 variable length array the upperbound is a dynamic attribute.
With this change c99 vla behave the same as with static arrays.
1| void foo (size_t n) {
2| int ary[n];
3| memset(ary, 0, sizeof(ary));
4| }
(gdb) print ary
$1 = {0 <repeats 42 times>}
gdb/ChangeLog:
* dwarf2loc.c (dwarf2_locexpr_baton_eval): New function.
(dwarf2_evaluate_property): New function.
* dwarf2loc.h (dwarf2_evaluate_property): New function prototype.
* dwarf2read.c (attr_to_dynamic_prop): New function.
(read_subrange_type): Use attr_to_dynamic_prop to read high bound
attribute.
* gdbtypes.c: Include dwarf2loc.h.
(is_dynamic_type): New function.
(resolve_dynamic_type): New function.
(resolve_dynamic_bounds): New function.
(get_type_length): New function.
(check_typedef): Use get_type_length to compute type length.
* gdbtypes.h (TYPE_HIGH_BOUND_KIND): New macro.
(TYPE_LOW_BOUND_KIND): New macro.
(is_dynamic_type): New function prototype.
* value.c (value_from_contents_and_address): Call resolve_dynamic_type
to resolve dynamic properties of the type. Update comment.
* valops.c (get_value_at, value_at, value_at_lazy): Update comment.
Yao Qi