A user found a bug where an array of packed arrays was printed
incorrectly. The bug here is that the packed array has a bit stride,
but the outer array does not -- and should not. However,
update_static_array_size does not distinguish between an array of
packed arrays and a multi-dimensional packed array, and for the
latter, only the innermost array will end up with a stride.
This patch fixes the problem by adding a flag to indicate whether a
given array type is a constituent of a multi-dimensional array.
Currently, every internal_error call must be passed __FILE__/__LINE__
explicitly, like:
internal_error (__FILE__, __LINE__, "foo %d", var);
The need to pass in explicit __FILE__/__LINE__ is there probably
because the function predates widespread and portable variadic macros
availability. We can use variadic macros nowadays, and in fact, we
already use them in several places, including the related
gdb_assert_not_reached.
So this patch renames the internal_error function to something else,
and then reimplements internal_error as a variadic macro that expands
__FILE__/__LINE__ itself.
The result is that we now should call internal_error like so:
internal_error ("foo %d", var);
Likewise for internal_warning.
The patch adjusts all calls sites. 99% of the adjustments were done
with a perl/sed script.
The non-mechanical changes are in gdbsupport/errors.h,
gdbsupport/gdb_assert.h, and gdb/gdbarch.py.
Approved-By: Simon Marchi <simon.marchi@efficios.com>
Change-Id: Ia6f372c11550ca876829e8fd85048f4502bdcf06
Add the `length` and `set_length` methods on `struct type`, in order to remove
the `TYPE_LENGTH` macro. In this patch, the macro is changed to use the
getter, so all the call sites of the macro that are used as a setter are
changed to use the setter method directly. The next patch will remove the
macro completely.
Change-Id: Id1090244f15c9856969b9be5006aefe8d8897ca4
Add the `target_type` and `set_target_type` methods on `struct type`, in order
to remove the `TYPE_TARGET_TYPE` macro. In this patch, the macro is changed to
use the getter, so all the call sites of the macro that are used as a setter
are changed to use the setter method directly. The next patch will remove the
macro completely.
Change-Id: I85ce24d847763badd34fdee3e14b8c8c14cb3161
gdbarch implements its own registry-like approach. This patch changes
it to instead use registry.h. It's a rather large patch but largely
uninteresting -- it's mostly a straightforward conversion from the old
approach to the new one.
The main benefit of this change is that it introduces type safety to
the gdbarch registry. It also removes a bunch of code.
One possible drawback is that, previously, the gdbarch registry
differentiated between pre- and post-initialization setup. This
doesn't seem very important to me, though.
This rewrites registry.h, removing all the macros and replacing it
with relatively ordinary template classes. The result is less code
than the previous setup. It replaces large macros with a relatively
straightforward C++ class, and now manages its own cleanup.
The existing type-safe "key" class is replaced with the equivalent
template class. This approach ended up requiring relatively few
changes to the users of the registry code in gdb -- code using the key
system just required a small change to the key's declaration.
All existing users of the old C-like API are now converted to use the
type-safe API. This mostly involved changing explicit deletion
functions to be an operator() in a deleter class.
The old "save/free" two-phase process is removed, and replaced with a
single "free" phase. No existing code used both phases.
The old "free" callbacks took a parameter for the enclosing container
object. However, this wasn't truly needed and is removed here as
well.
When an objfile is destroyed, types that are still in use and
allocated on that objfile are copied. A temporary hash map is created
during this process, and it is allocated on the destroyed objfile's
obstack -- which normally is fine, as that is going to be destroyed
shortly anyway.
However, this approach requires that the objfile be passed to registry
destruction, and this won't be possible in the rewritten registry.
This patch changes the copied type hash table to simply use the heap
instead. It also removes the 'objfile' parameter from
copy_type_recursive, to make this all more clear.
This patch also fixes an apparent bug in copy_type_recursive.
Previously it was copying the dynamic property list to the dying
objfile's obstack:
- = copy_dynamic_prop_list (&objfile->objfile_obstack,
However I think this is incorrect -- that obstack is about to be
destroyed.
PR exp/20630 points out a simple way to cause an assertion failure in
copy_type -- but this was found in the wild a few times as well.
copy_type only works for objfile-owned types, but there isn't a deep
reason for this. This patch fixes the bug by updating copy_type to
work for any sort of type.
Better would perhaps be to finally implement type GC, but I still
haven't attempted this.
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=20630
If a variable is passed to function in FORTRAN as an argument the
variable is treated as an array with rank zero. GDB currently does
not support the case for assumed rank 0. This patch provides support
for assumed rank 0 and updates the testcase as well.
Without patch:
Breakpoint 1, arank::sub1 (a=<error reading variable:
failed to resolve dynamic array rank>) at assumedrank.f90:11
11 PRINT *, RANK(a)
(gdb) p a
failed to resolve dynamic array rank
(gdb) p rank(a)
failed to resolve dynamic array rank
With patch:
Breakpoint 1, arank::sub1 (a=0) at assumedrank.f90:11
11 PRINT *, RANK(a)
(gdb) p a
$1 = 0
(gdb) p rank(a)
$2 = 0
Commit:
commit df7a7bdd97
Date: Thu Mar 17 18:56:23 2022 +0000
gdb: add support for Fortran's ASSUMED RANK arrays
Added support for Fortran assumed rank arrays. Unfortunately, this
commit contained a bug that means though GDB can correctly calculate
the rank of an assumed rank array, GDB can't fetch the contents of an
assumed rank array.
The history of this patch can be seen on the mailing list here:
https://sourceware.org/pipermail/gdb-patches/2022-January/185306.html
The patches that were finally committed can be found here:
https://sourceware.org/pipermail/gdb-patches/2022-March/186906.html
The original patches did support fetching the array contents, it was
only the later series that introduced the regression.
The problem is that when calculating the array rank the result is a
count of the number of ranks, i.e. this is a 1 based result, 1, 2, 3,
etc.
In contrast, when computing the details of any particular rank the
value passed to the DWARF expression evaluator should be a 0 based
rank offset, i.e. a 0 based number, 0, 1, 2, etc.
In the patches that were originally merged, this was not the case, and
we were passing the 1 based rank number to the expression evaluator,
e.g. passing 1 when we should pass 0, 2 when we should pass 1, etc.
As a result the DWARF expression evaluator was reading the
wrong (undefined) memory, and returning garbage results.
In this commit I have extended the test case to cover checking the
array contents, I've then ensured we make use of the correct rank
value, and extended some comments, and added or adjusted some asserts
as appropriate.
This patch adds a new dynamic property DYN_PROP_RANK, this property is
read from the DW_AT_rank attribute and stored within the type just
like other dynamic properties.
As arrays with dynamic ranks make use of a single
DW_TAG_generic_subrange to represent all ranks of the array, support
for this tag has been added to dwarf2/read.c.
The final piece of this puzzle is to add support in gdbtypes.c so that
we can resolve an array type with dynamic rank. To do this the
existing resolve_dynamic_array_or_string function is split into two,
there's a new resolve_dynamic_array_or_string_1 core that is
responsible for resolving each rank of the array, while the now outer
resolve_dynamic_array_or_string is responsible for figuring out the
array rank (which might require resolving a dynamic property) and then
calling the inner core.
The resolve_dynamic_range function now takes a rank, which is passed
on to the dwarf expression evaluator. This rank will only be used in
the case where the array itself has dynamic rank, but we now pass the
rank in all cases, this should be harmless if the rank is not needed.
The only small nit is that resolve_dynamic_type_internal actually
handles resolving dynamic ranges itself, which now obviously requires
us to pass a rank value. But what rank value to use? In the end I
just passed '1' through here as a sane default, my thinking is that if
we are in resolve_dynamic_type_internal to resolve a range, then the
range isn't part of an array with dynamic rank, and so the range
should actually be using the rank value at all.
An alternative approach would be to make the rank value a
gdb::optional, however, this ends up adding a bunch of complexity to
the code (e.g. having to conditionally build the array to pass to
dwarf2_evaluate_property, and handling the 'rank - 1' in
resolve_dynamic_array_or_string_1) so I haven't done that, but could,
if people think that would be a better approach.
Finally, support for assumed rank arrays was only fixed very recently
in gcc, so you'll need the latest gcc in order to run the tests for
this.
Here's an example test program:
PROGRAM arank
REAL :: a1(10)
CALL sub1(a1)
CONTAINS
SUBROUTINE sub1(a)
REAL :: a(..)
PRINT *, RANK(a)
END SUBROUTINE sub1
END PROGRAM arank
Compiler Version:
gcc (GCC) 12.0.0 20211122 (experimental)
Compilation command:
gfortran assumedrank.f90 -gdwarf-5 -o assumedrank
Without Patch:
gdb -q assumedrank
Reading symbols from assumedrank...
(gdb) break sub1
Breakpoint 1 at 0x4006ff: file assumedrank.f90, line 10.
(gdb) run
Starting program: /home/rupesh/STAGING-BUILD-2787/bin/assumedrank
Breakpoint 1, arank::sub1 (a=<unknown type in /home/rupesh/STAGING-BUILD-2787
/bin/assumedrank, CU 0x0, DIE 0xd5>) at assumedrank.f90:10
10 PRINT *, RANK(a)
(gdb) print RANK(a)
'a' has unknown type; cast it to its declared type
With patch:
gdb -q assumedrank
Reading symbols from assumedrank...
(gdb) break sub1
Breakpoint 1 at 0x4006ff: file assumedrank.f90, line 10.
(gdb) run
Starting program: /home/rupesh/STAGING-BUILD-2787/bin/assumedrank
Breakpoint 1, arank::sub1 (a=...) at assumedrank.f90:10
10 PRINT *, RANK(a)
(gdb) print RANK(a)
$1 = 1
(gdb) ptype a
type = real(kind=4) (10)
(gdb)
Co-Authored-By: Andrew Burgess <aburgess@redhat.com>
When we need to evaluate a DWARF expression in order to resolve some
dynamic property of a type we call the dwarf2_evaluate_property
function, which is declared in gdb/dwarf/loc.h and defined in
gdb/dwarf/loc.c.
Currently, this function takes (amongst other things) an argument of
type property_addr_info called addr_stack and a boolean called
push_initial_value. When push_initial_value then the top value of
addr_stack is pushed onto the dwarf expression evaluation stack before
the expression is evaluated.
So far this has worked fine, as the only two cases we needed to handle
are the case the DWARF expression doesn't require the object
address (what the top of addr_stack represents), and the case where
the DWARF expression does require the address.
In the next commit this is going to change. As we add support for
Fortran assumed rank arrays, we need to start resolving the dynamic
properties of arrays. To do this, we need to push the array rank onto
the dwarf expression evaluation stack before the expression is
evaluated.
This commit is a refactoring commit aimed at making it easier to
support Fortran assumed rank arrays. Instead of passing a boolean,
and using this to decide if we should push the object address or not,
we instead pass an array (view) of values that should be pushed to the
dwarf expression evaluation stack.
In the couple of places where we previously passed push_initial_value
as true (mostly this was defaulting to false), we now have to pass the
address from the addr_stack as an item in the array view.
In the next commit, when we want to handle passing the array rank,
this will easily be supported too.
There should be no user visible changes after this commit.
It is better to rename floatformats_ia64_quad to floatformats_ieee_quad
to reflect the reality, and then we can clean up the related code.
As Tom Tromey said [1]:
These files are maintained in gcc and then imported into the
binutils-gdb repository, so any changes to them will have to
be proposed there first.
the related changes have been merged into gcc master now [2], it is time
to do it for gdb.
[1] https://sourceware.org/pipermail/gdb-patches/2022-March/186569.html
[2] https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=b2dff6b2d9d6
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Now that filtered and unfiltered output can be treated identically, we
can unify the printf family of functions. This is done under the name
"gdb_printf". Most of this patch was written by script.
Now that filtered and unfiltered output can be treated identically, we
can unify the puts family of functions. This is done under the name
"gdb_puts". Most of this patch was written by script.
PR c++/28901 points out a bug in C++ overload resolution. When
comparing two overloads, one might be better than the other for
certain parameters -- but, if that one also has some invalid
conversion, then it should never be considered the better choice.
Instead, a valid-but-not-apparently-quite-as-good overload should be
preferred.
This patch fixes this problem by changing how overload comparisons are
done. I don't believe it should affect any currently valid overload
resolution; nor should it affect resolutions where all the choices are
equally invalid.
It is possible for a compiler to optimize a function in a such ways that
the function does not follow the calling convention of the target. In
such situation, the compiler can use the DW_AT_calling_convention
attribute with the value DW_CC_nocall to tell the debugger that it is
unsafe to call the function. The DWARF5 standard states, in 3.3.1.1:
> If the value of the calling convention attribute is the constant
> DW_CC_nocall, the subroutine does not obey standard calling
> conventions, and it may not be safe for the debugger to call this
> subroutine.
Non standard calling convention can affect GDB's assumptions in multiple
ways, including how arguments are passed to the function, how values are
returned, and so on. For this reason, it is unsafe for GDB to try to do
the following operations on a function with marked with DW_CC_nocall:
- call / print an expression requiring the function to be evaluated,
- inspect the value a function returns using the 'finish' command,
- force the value returned by a function using the 'return' command.
This patch ensures that if a command which relies on GDB's knowledge of
the target's calling convention is used on a function marked nocall, GDB
prints an appropriate message to the user and does not proceed with the
operation which is unreliable.
Note that it is still possible for someone to use a vendor specific
value for the DW_AT_calling_convention attribute for example to indicate
the use of an alternative calling convention. This commit does not
prevent this, and target dependent code can be adjusted if one wanted to
support multiple calling conventions.
Tested on x86_64-Linux, with no regression observed.
Change-Id: I72970dae68234cb83edbc0cf71aa3d6002a4a540
Add a getter and a setter for a symbol's type. Remove the corresponding
macro and adjust all callers.
Change-Id: Ie1a137744c5bfe1df4d4f9ae5541c5299577c8de
Add a getter and a setter for a compunit_symtab's block line section. Remove
the corresponding macro and adjust all callers.
Change-Id: I3eb1a323388ad55eae8bfa45f5bc4a08dc3df455
An earlier patch of mine, commit 64b7cc50 ("Remove
gdb_print_host_address") inadvertently changed a function in
gdbtypes.c to use printf rather than printf_filtered. This patch
fixes the problem.
This commit brings all the changes made by running gdb/copyright.py
as per GDB's Start of New Year Procedure.
For the avoidance of doubt, all changes in this commits were
performed by the script.
gdb_print_host_address is just a simple wrapper around
fprintf_filtered. However, it is readily replaced in all callers by a
combination of %s and call to host_address_to_string. This also
simplifies the code, so I think it's worthwhile to remove this
function.
Regression tested on x86-64 Fedora 64.
The bug fixed by this [1] patch was caused by an out-of-bounds access to
a value's content. The code gets the value's content (just a pointer)
and then indexes it with a non-sensical index.
This made me think of changing functions that return value contents to
return array_views instead of a plain pointer. This has the advantage
that when GDB is built with _GLIBCXX_DEBUG, accesses to the array_view
are checked, making bugs more apparent / easier to find.
This patch changes the return types of these functions, and updates
callers to call .data() on the result, meaning it's not changing
anything in practice. Additional work will be needed (which can be done
little by little) to make callers propagate the use of array_view and
reap the benefits.
[1] https://sourceware.org/pipermail/gdb-patches/2021-September/182306.html
Change-Id: I5151f888f169e1c36abe2cbc57620110673816f3
This changes gdb to check the index that is passed to type::field.
This caught one bug in the Ada code when running the test suite
(actually I found the bug first, then realized that the check would
have helped), so this patch fixes that as well.
Regression tested on x86-64 Fedora 34.
A customer-reported problem led us to a bug in dynamic type
resolution. resolve_dynamic_struct will recursively call
resolve_dynamic_type_internal, passing it the sub-object for the
particular field being resolved. While it offsets the address here,
it does not also offset the "valaddr" -- the array of bytes describing
the memory.
This patch fixes the bug, by offsetting both. A test case is included
that can be used to reproduce the bug.
Add accessors for the various location values in struct field. This
lets us assert that when we get a location value of a certain kind (say,
bitpos), the field's location indeed contains a value of that kind.
Remove the SET_FIELD_* macros, instead use the new setters directly.
Update the FIELD_* macros used to access field locations to go through
the getters. They will be removed in a subsequent patch.
There are places where the FIELD_* macros are used on call_site_target
structures, because it contains members of the same name (loc_kind and
loc). For now, I have replicated the getters/setters in
call_site_target. But we could perhaps eventually factor them in a
"location" structure that can be used at both places.
Note that the field structure, being zero-initialized, defaults to a
bitpos location with value 0. While writing this patch, I tried to make
it default to an "unset" location, to catch places where we would miss
setting a field's location. However, I found that some places relied on
the default being "bitpos 0", so I left it as-is. This change could
always be done as follow-up work, making these places explicitly set the
"bitpos 0" location.
I found two issues to fix:
- I got some failures in the gdb.base/infcall-nested-structs-c++.exp
test. They were caused by two functions in amd64-tdep.c using
TYPE_FIELD_BITPOS before checking if the location is of the bitpos
kind, which they do indirectly through `field_is_static`. Simply
move getting the bitpos below the field_is_static call.
- I got a failure in gdb.xml/tdesc-regs.exp. It turns out that in
make_gdb_type_enum, we set enum field values using SET_FIELD_BITPOS,
and later access them through FIELD_ENUMVAL. Fix that by using
set_loc_enumval to set the value.
Change-Id: I53d3734916c46457576ba11dd77df4049d2fc1e8
I noticed that some methods in language_defn could use
unique_xmalloc_ptr<char> rather than a plain 'char *'. This patch
implements this change, fixing up the fallout and changing
gdb_demangle to also return this type. In one spot, std::string is
used to simplify some related code, and in another, an auto_obstack is
used to avoid manual management.
Regression tested on x86-64 Fedora 34.
Consider test-case gdb.trace/entry-values.exp with target board
unix/-fPIE/-pie.
Using this command we have an abbreviated version, and can see the correct
@entry values for foo:
...
$ gdb -q -batch outputs/gdb.trace/entry-values/entry-values \
-ex start \
-ex "break foo" \
-ex "set print entry-values both" \
-ex continue
Temporary breakpoint 1 at 0x679
Temporary breakpoint 1, 0x0000555555554679 in main ()
Breakpoint 2 at 0x55555555463e
Breakpoint 2, 0x000055555555463e in foo (i=0, i@entry=2, j=2, j@entry=3)
...
Now, let's try the same again, but run directly to foo rather than stopping at
main:
...
$ gdb -q -batch outputs/gdb.trace/entry-values/entry-values \
-ex "break foo" \
-ex "set print entry-values both" \
-ex run
Breakpoint 1 at 0x63e
Breakpoint 1, 0x000055555555463e in foo (i=0, i@entry=<optimized out>, \
j=2, j@entry=<optimized out>)
...
So, what explains the difference? Noteworthy, this is a dwarf assembly
test-case, with debug info for foo and bar, but not for main.
In the first case:
- we run to main
- this does not trigger expanding debug info, because there's none for main
- we set a breakpoint at foo
- this triggers expanding debug info. Relocated addresses are used in
call_site info (because the exec is started)
- we continue to foo, and manage to find the call_site info
In the second case:
- we set a breakpoint at foo
- this triggers expanding debug info. Unrelocated addresses are used in
call_site info (because the exec is not started)
- we run to foo
- this triggers objfile_relocate1, but it doesn't update the call_site
info addresses
- we don't manage to find the call_site info
We could fix this by adding the missing call_site relocation in
objfile_relocate1.
This solution however is counter-trend in the sense that we're trying to
work towards the situation where when starting two instances of an executable,
we need only one instance of debug information, implying the use of
unrelocated addresses.
So, fix this instead by using unrelocated addresses in call_site info.
Tested on x86_64-linux.
This fixes all remaining unix/-fno-PIE/-no-pie vs unix/-fPIE/-pie
regressions, like f.i. PR24892.
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=24892
Co-Authored-By: Tom de Vries <tdevries@suse.de>
Remove the `TYPE_FIELD_NAME` and `FIELD_NAME` macros, changing all the
call sites to use field::name directly.
Change-Id: I6900ae4e1ffab1396e24fb3298e94bf123826ca6
Add the `name` and `set_name` methods on `struct field`, in order to
remove `FIELD_NAME` and `TYPE_FIELD_NAME` macros. In this patch, the
macros are changed to use `field::name`, so all the call sites that are
used to set the field's name are changed to use `field::set_name`.
The next patch will remove the macros completely.
Note that because of the name clash between the existing field named
`name` and the new method, I renamed the field `m_name`. It is not
private per-se, because we can't make `struct field` a non-POD yet, but
it should be considered private anyway (not accessed outside `struct
field`).
Change-Id: If16ddbca4e0c39d0ff9da420bb5cdebe5b9b0896
I noticed that pointer_type is declared in language.h and defined in
language.c. However, it really has to do with types, so it should
have been in gdbtypes.h all along.
This patch changes it to be a method on struct type. And, I went
through uses of TYPE_IS_REFERENCE and updated many spots to use the
new method as well. (I didn't update ones that were in arch-specific
code, as I couldn't readily test that.)
The case for FIELD_LOC_KIND_DWARF_BLOCK was missing for
switch TYPE_FIELD_LOC_KIND. Thas caused an internal-error
under some circumstances.
Fixes bug 28030.
