Samuel Thibault [Sat, 30 May 2020 18:45:30 +0000 (18:45 +0000)]
hurd: unwinding support over signal trampolines
This allows to get full backtrace from signal handlers, otherwise the
backtrace stops at the trampoline that calls the handler.
This needs special knowledge how the trampoline records register context
for the sigreturn call after signal handling.
gdb/ChangeLog:
* i386-gnu-tdep.c: Include "gdbcore.h"
(gnu_sigtramp_code, i386_gnu_sc_reg_offset): New arrays.
(GNU_SIGTRAMP_LEN, GNU_SIGTRAMP_TAIL,
I386_GNU_SIGCONTEXT_THREAD_STATE_OFFSET): New macros
(i386_gnu_sigtramp_start, i386_gnu_sigtramp_p,
i386_gnu_sigcontext_addr): New functions
(i386gnu_init_abi): Register i386_gnu_sigtramp_p,
i386_gnu_sigcontext_addr, and i386_gnu_sc_reg_offset in the gdbarch
tdep.
Samuel Thibault [Sat, 30 May 2020 18:41:30 +0000 (18:41 +0000)]
hurd: add missing include
Fixes
../../gdb/gnu-nat.c:2522:14: error: ‘target_gdbarch’ was not declared in this scope; did you mean ‘target_detach’?
2522 | paddress (target_gdbarch (), memaddr), pulongest (len),
Samuel Thibault [Sat, 30 May 2020 18:38:46 +0000 (18:38 +0000)]
hurd: make function cast stronger
Fixes
process_reply_S.c:104:23: error: function called through a non-compatible type [-Werror]
104 | OutP->RetCode = (*(kern_return_t (*)(mach_port_t, kern_return_t)) S_proc_setmsgport_reply) (In0P->Head.msgh_request_port, In0P-
As the existing comment says, it is in general not safe to drop some
parameters like this, but this is the error handling case, where the
called function does not actually read them, and mig is currently planned
to be used on i386 and x86_64 only, where this is not a problem. As the
existing comment says, fixing it properly would be far from trivial:
we can't just pass 0 for them, as they might not be scalar.
gdb/ChangeLog:
* reply_mig_hack.awk (Error return): Cast function through
void *, to bypass compiler function call check.
Pedro Alves [Sat, 30 May 2020 13:20:10 +0000 (14:20 +0100)]
replace_typedefs: handle templates in namespaces
GDB currently crashes with infinite recursion, if you set a breakpoint
on a function inside a namespace that includes a template on its fully
qualified name, and, the template's name is also used as typedef in
the global scope that expands to a name that includes the template
name in its qualified name. For example, from the testcase added by
this commit:
Results in infinite recursion, with this cycle (started by
cp_canonicalize_string_full) repeating until the stack is exhausted:
...
#1709 0x000000000055533c in inspect_type (info=0x38ff720, ret_comp=0xd83be10, finder=0x0, data=0x0) at /home/pedro/gdb/mygit/src/gdb/cp-support.c:267
#1710 0x0000000000555a6f in replace_typedefs (info=0x38ff720, ret_comp=0xd83be10, finder=0x0, data=0x0) at /home/pedro/gdb/mygit/src/gdb/cp-support.c:475
#1711 0x0000000000555a36 in replace_typedefs (info=0x38ff720, ret_comp=0xd83be70, finder=0x0, data=0x0) at /home/pedro/gdb/mygit/src/gdb/cp-support.c:470
#1712 0x0000000000555800 in replace_typedefs_qualified_name (info=0x38ff720, ret_comp=0xd839470, finder=0x0, data=0x0) at /home/pedro/gdb/mygit/src/gdb/cp-support.c:389
#1713 0x0000000000555a8c in replace_typedefs (info=0x38ff720, ret_comp=0xd839470, finder=0x0, data=0x0) at /home/pedro/gdb/mygit/src/gdb/cp-support.c:479
...
The demangle component tree for that symbol name looks like this:
d_dump tree for NS1::NS2::Templ1<int>::Templ1(NS1::NS2::object*):
typed name
qualified name
name 'NS1'
qualified name
name 'NS2'
qualified name
template <<<<<<<<<<
name 'Templ1'
template argument list
builtin type int
name 'Templ1'
function type
argument list
pointer
qualified name
name 'NS1'
qualified name
name 'NS2'
name 'object'
The recursion starts at replace_typedefs_qualified_name, which doesn't
handle the "template" node, and thus doesn't realize that the template
name actually has the fully qualified name NS1::NS2::Templ1.
replace_typedefs_qualified_name calls into replace_typedefs on the
template node, and that ends up in inspect_type looking up for a
symbol named "Templ1", which finds the global namespace typedef, which
itself expands to NS1::NS2::Templ1. GDB then tries replacing typedefs
in that newly expanded name, which ends up again in
replace_typedefs_qualified_name, trying to expand a fully qualified
name with "NS::NS2::Templ1<unsigned>" in its name, which results in
recursion whenever the template node is reached.
Fix this by teaching replace_typedefs_qualified_name how to handle
template nodes. It needs handling in two places: the first spot
handles the symbol above; the second spot handles a symbol like this,
from the new test:
(gdb) b NS1::NS2::grab_it(NS1::NS2::Templ1<int>*)
d_dump tree for NS1::NS2::grab_it(NS1::NS2::Templ1<int>*):
typed name
qualified name
name 'NS1'
qualified name
name 'NS2'
name 'grab_it'
function type
argument list
pointer
qualified name
name 'NS1'
qualified name
name 'NS2'
template <<<<<<<<
name 'Templ1'
template argument list
builtin type int
What's different in this case is that the template node appears on the
right child node of a qualified name, instead of on the left child.
The testcase includes a test that checks whether template aliases are
correctly replaced by GDB too. That fails with GCC due to GCC PR
95437, which makes GDB not know about a typedef for
"NS1::NS2::AliasTempl<int>". GCC emits a typedef named
"NS1::NS2::AliasTempl" instead, with no template parameter info. The
test passes with Clang (5.0.2 at least). See more details here:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95437
gdb/ChangeLog:
2020-05-30 Pedro Alves <palves@redhat.com>
Simon Marchi [Fri, 29 May 2020 19:15:10 +0000 (15:15 -0400)]
gdb: rename dwarf2_per_objfile variables/fields to per_objfile
While doing the psymtab-sharing patchset, I avoided renaming variables
unnecessarily to avoid adding noise to patches, but I'd like to do it
now. Basically, we have these dwarf2 per-something structures:
I named the instances of dwarf2_per_bfd `per_bfd` and most of instances
of dwarf2_per_cu_data are called `per_cu`. Most pre-existing instances
of dwarf2_per_objfile are named `dwarf2_per_objfile`. For consistency
with the other type, I'd like to rename them to just `per_objfile`. The
`dwarf2_` prefix is superfluous, since it's already clear we are in
dwarf2 code. It also helps reducing the line wrapping by saving 7
precious columns.
gdb/ChangeLog:
* dwarf2/comp-unit.c, dwarf2/comp-unit.h, dwarf2/index-cache.c,
dwarf2/index-cache.h, dwarf2/index-write.c,
dwarf2/index-write.h, dwarf2/line-header.c,
dwarf2/line-header.h, dwarf2/macro.c, dwarf2/macro.h,
dwarf2/read.c, dwarf2/read.h: Rename struct dwarf2_per_objfile
variables and fields from `dwarf2_per_objfile` to just
`per_objfile` throughout.
Gary Benson [Fri, 29 May 2020 16:43:17 +0000 (17:43 +0100)]
Fix build errors in with clang in gdb.compile/compile-cplus.c
Clang fails to compile the file, with the following error:
fatal error: 'iostream' file not found
This prevents the following testcase from executing:
gdb.compile/compile-cplus.exp
The testcase sets additional_flags when building with GCC, which
this commit causes to also be set when building with clang. This
makes the testcase fail to build with a different error:
warning: treating 'c' input as 'c++' when in C++ mode, this behavior
is deprecated [-Wdeprecated]
so this commit adds -Wno-deprecated in two places to sidestep this.
Note that, while allowing the testcase to build, this commit reveals
failures when the testsuite is built using clang.
gdb/testsuite/ChangeLog:
* gdb.compile/compile-cplus.exp (additional_flags): Also
set when building with clang.
(additional_flags, srcfilesoptions): Pass -Wno-deprecated
when building with clang.
Gary Benson [Fri, 29 May 2020 16:15:13 +0000 (17:15 +0100)]
Fix file-not-found error with clang in gdb.arch/i386-{avx,sse}.c
Clang fails to compile two testcases with the following error:
fatal error: 'nat/x86-cpuid.h' file not found
This prevents the following testcases from executing:
gdb.arch/i386-avx.exp
gdb.arch/i386-sse.exp
Both testcases set additional_flags when building with GCC.
This commit causes the additional_flags to also be used when
building with clang. Note that, while fixing the build, this
commit reveals several new failures when using clang to build
the testsuite.
gdb/testsuite/ChangeLog:
* gdb.arch/i386-avx.exp (additional_flags): Also set when
building with clang.
* gdb.arch/i386-sse.exp (additional_flags): Likewise.
Gary Benson [Fri, 29 May 2020 13:03:01 +0000 (14:03 +0100)]
Build two gdb.cp testcases with -Wno-unused-comparison
Clang fails to compile two testcases with the following error:
warning: equality comparison result unused [-Wunused-comparison]
This prevents the following testcases from executing:
gdb.cp/koenig.exp
gdb.cp/operator.exp
This commit builds those testcases with -Wno-unused-comparison, to
avoid the failure. Note that this commit reveals a new failure,
"FAIL: gdb.cp/koenig.exp: p foo (p_union)" when the testsuite is
compiled using clang.
Simon Marchi [Thu, 28 May 2020 19:47:53 +0000 (15:47 -0400)]
gdb: add comment in dwarf_evaluate_loc_desc::push_dwarf_reg_entry_value
Add a comment to clarify why we temporarily override some of the
context's fields, and especially the per_objfile field. A longer
explanation can be found in this previous commit
44486dcf19b ("gdb: use caller objfile in dwarf_evaluate_loc_desc::push_dwarf_reg_entry_value")
Kevin Buettner [Thu, 28 May 2020 03:05:40 +0000 (20:05 -0700)]
Fix Python3.9 related runtime problems
Python3.9b1 is now available on Rawhide. GDB w/ Python 3.9 support
can be built using the configure switch -with-python=/usr/bin/python3.9.
Attempting to run gdb/Python3.9 segfaults on startup:
#0 0x00007ffff7b0582c in PyEval_ReleaseLock () from /lib64/libpython3.9.so.1.0
#1 0x000000000069ccbf in do_start_initialization ()
at worktree-test1/gdb/python/python.c:1789
#2 _initialize_python ()
at worktree-test1/gdb/python/python.c:1877
#3 0x00000000007afb0a in initialize_all_files () at init.c:237
...
Consulting the the documentation...
https://docs.python.org/3/c-api/init.html
...we find that PyEval_ReleaseLock() has been deprecated since version
3.2. It recommends using PyEval_SaveThread or PyEval_ReleaseThread()
instead. In do_start_initialization, in gdb/python/python.c, we
can replace the calls to PyThreadState_Swap() and PyEval_ReleaseLock()
with a single call to PyEval_SaveThread. (Thanks to Keith Seitz
for working this out.)
With that in place, GDB gets a little bit further. It still dies
on startup, but the backtrace is different:
#0 0x00007ffff7b04306 in PyOS_InterruptOccurred ()
from /lib64/libpython3.9.so.1.0
#1 0x0000000000576e86 in check_quit_flag ()
at worktree-test1/gdb/extension.c:776
#2 0x0000000000576f8a in set_active_ext_lang (now_active=now_active@entry=0x983c00 <extension_language_python>)
at worktree-test1/gdb/extension.c:705
#3 0x000000000069d399 in gdbpy_enter::gdbpy_enter (this=0x7fffffffd2d0,
gdbarch=0x0, language=0x0)
at worktree-test1/gdb/python/python.c:211
#4 0x0000000000686e00 in python_new_inferior (inf=0xddeb10)
at worktree-test1/gdb/python/py-inferior.c:251
#5 0x00000000005d9fb9 in std::function<void (inferior*)>::operator()(inferior*) const (__args#0=<optimized out>, this=0xccad20)
at /usr/include/c++/10/bits/std_function.h:617
#6 gdb::observers::observable<inferior*>::notify (args#0=0xddeb10,
this=<optimized out>)
at worktree-test1/gdb/../gdbsupport/observable.h:106
#7 add_inferior_silent (pid=0)
at worktree-test1/gdb/inferior.c:113
#8 0x00000000005dbcb8 in initialize_inferiors ()
at worktree-test1/gdb/inferior.c:947
...
We checked with some Python Developers and were told that we should
acquire the GIL prior to calling any Python C API function. We
definitely don't have the GIL for calls of PyOS_InterruptOccurred().
I moved class_gdbpy_gil earlier in the file and use it in
gdbpy_check_quit_flag() to acquire (and automatically release) the
GIL.
With those changes in place, I was able to run to a GDB prompt. But,
when trying to quit, it segfaulted again due to due to some other
problems with gdbpy_check_quit_flag():
Thread 1 "gdb" received signal SIGSEGV, Segmentation fault.
0x00007ffff7bbab0c in new_threadstate () from /lib64/libpython3.9.so.1.0
(top-gdb) bt 8
#0 0x00007ffff7bbab0c in new_threadstate () from /lib64/libpython3.9.so.1.0
#1 0x00007ffff7afa5ea in PyGILState_Ensure.cold ()
from /lib64/libpython3.9.so.1.0
#2 0x000000000069b58c in gdbpy_gil::gdbpy_gil (this=<synthetic pointer>)
at worktree-test1/gdb/python/python.c:278
#3 gdbpy_check_quit_flag (extlang=<optimized out>)
at worktree-test1/gdb/python/python.c:278
#4 0x0000000000576e96 in check_quit_flag ()
at worktree-test1/gdb/extension.c:776
#5 0x000000000057700c in restore_active_ext_lang (previous=0xe9c050)
at worktree-test1/gdb/extension.c:729
#6 0x000000000088913a in do_my_cleanups (
pmy_chain=0xc31870 <final_cleanup_chain>,
old_chain=0xae5720 <sentinel_cleanup>)
at worktree-test1/gdbsupport/cleanups.cc:131
#7 do_final_cleanups ()
at worktree-test1/gdbsupport/cleanups.cc:143
In this case, we're trying to call a Python C API function after
Py_Finalize() has been called from finalize_python(). I made
finalize_python set gdb_python_initialized to false and then cause
check_quit_flag() to return early when it's false.
With these changes in place, GDB seems to be working again with
Python3.9b1. I think it likely that there are other problems lurking.
I wouldn't be surprised to find that there are other calls into Python
where we don't first make sure that we have the GIL. Further changes
may well be needed.
I see no regressions testing on Rawhide using a GDB built with the
default Python version (3.8.3) versus one built using Python 3.9b1.
I've also tested on Fedora 28, 29, 30, 31, and 32 (all x86_64) using
the default (though updated) system installed versions of Python on
those OSes. This means that I've tested against Python versions
2.7.15, 2.7.17, 2.7.18, 3.7.7, 3.8.2, and 3.8.3. In each case GDB
still builds without problem and shows no regressions after applying
this patch.
gdb/ChangeLog:
2020-MM-DD Kevin Buettner <kevinb@redhat.com>
Keith Seitz <keiths@redhat.com>
* python/python.c (do_start_initialization): For Python 3.9 and
later, call PyEval_SaveThread instead of PyEval_ReleaseLock.
(class gdbpy_gil): Move to earlier in file.
(finalize_python): Set gdb_python_initialized.
(gdbpy_check_quit_flag): Acquire GIL via gdbpy_gil. Return early
when not initialized.
Stephen Casner [Thu, 28 May 2020 17:11:59 +0000 (10:11 -0700)]
Fix all unexpected failures in gas testsuite for pdp11-aout.
These failures were caused by the PDP11's mix of little-endian octets
in shorts but shorts in big endian order for long or quad so regexps
did not match. Also tests used addresses as values in .long which
required BRD_RELOC_32 that was not implemented.
* gas/config/tc-pdp11.c (md_number_to_chars): Implement .quad
* gas/testsuite/gas/all/gas.exp: Select alternate test scripts for
pdp11, skip octa test completely.
* gas/testsuite/gas/all/eqv-dot-pdp11.s: Identical to eqv-dot.s
* gas/testsuite/gas/all/eqv-dot-pdp11.d: Match different octet order.
* gas/testsuite/gas/all/cond-pdp11.l: Match different octet order.
* bfd/pdp11.c: Implement BRD_RELOC_32 to relocate the low 16 bits of
addreses in .long (used in testsuites) and .stab values.
Gary Benson [Thu, 28 May 2020 17:02:55 +0000 (18:02 +0100)]
Fix "enumeration values not handled in switch" error in testsuite
When running the testsuite with clang, gdb.base/sigaltstack.c
fails to compile with the following error:
warning: enumeration values 'LEAF' and 'NR_LEVELS' not handled
in switch [-Wswitch]
This prevents the gdb.base/sigaltstack.exp from executing.
This commit fixes.
gdb/testsuite/ChangeLog:
* gdb.base/sigaltstack.c (catcher): Add default case to switch
statement.
Simon Marchi [Thu, 28 May 2020 15:30:11 +0000 (11:30 -0400)]
gdb: use caller objfile in dwarf_evaluate_loc_desc::push_dwarf_reg_entry_value
In commit
89b07335fe ("Add dwarf2_per_objfile to dwarf_expr_context and dwarf2_frame_cache")
I replaced the offset property of dwarf_expr_context by a per_objfile
property (since we can get the text offset from the objfile). The
previous code in dwarf_evaluate_loc_desc::push_dwarf_reg_entry_value
(dwarf_evaluate_loc_desc derives from dwarf_expr_context) did
temporarily override the offset property while evaluating a DWARF
sub-expression. I speculated that this sub-expression always came from
the same objfile as the outer expression, so I didn't see the need to
temporarily override the per_objfile property in the new code. A later
commit:
It turns out that this is not true. Call sites can refer to function in
another objfile, and therefore the caller's objfile can be different
from the callee's objfile. This can happen when the call site DIE in the
DWARF represents a function call done through a function pointer. The
DIE can't describe statically which function is being called, since it's
variable and not known at compile time. Instead, it provides an
expression that evaluates to the address of the function being called.
In this case, the called function can very well be in a separate
objfile.
Fix this by overriding the per_objfile property while evaluating the
sub-expression.
This was exposed by the gdb.base/catch-load.exp test failing on openSUSE
Tumbleweed with the glibc debug info installed. It was also reported to
fail on Fedora.
When I investigated the problem, the particular call site on which we
did hit the assert was coming from this DIE, in
/usr/lib/debug/lib64/libc-2.31.so-2.31-5.1.x86_64.debug on openSUSE
Tumbleweed:
Gary Benson [Thu, 28 May 2020 15:29:48 +0000 (16:29 +0100)]
Pass -Wno-deprecated-register for gdb.cp that use "register"
Clang fails to compile three testcases with the following error:
warning: 'register' storage class specifier is deprecated and
incompatible with C++17 [-Wdeprecated-register]
This prevents the following testcases from executing:
gdb.cp/classes.exp
gdb.cp/inherit.exp
gdb.cp/misc.exp
This commit builds those testcases with -Wno-deprecated-register, to
avoid the failure. Note that this commit reveals five "wrong access
specifier for typedef" failures in gdb.cp/classes.exp when compiling
the testsuite with clang.
Tom de Vries [Thu, 28 May 2020 15:26:22 +0000 (17:26 +0200)]
[gdb/symtab] Make gold index workaround more precise
There's a PR gold/15646 - "gold-generated .gdb_index has duplicated
symbols that gdb-generated index doesn't", that causes gold to generate
duplicate symbols in the index.
F.i., a namespace N1 declared in a header file can be listed for two CUs that
include the header file:
...
[759] N1:
2 [global type]
3 [global type]
...
This causes a gdb performance problem: f.i. when attempting to set a
breakpoint on a non-existing function N1::misspelled, the symtab for both CUs
will be expanded.
Gdb contains a workaround for this, added in commit 8943b87476 "Work around
gold/15646", that skips duplicate global symbols in the index.
However, the workaround does not check for the symbol kind ("type" in the
example above).
Make the workaround more precise by limiting it to symbol kind "type".
Tested on x86_64-linux, with target boards cc-with-gdb-index and
gold-gdb-index.
gdb/ChangeLog:
2020-05-28 Tom de Vries <tdevries@suse.de>
* dwarf2/read.c (dw2_symtab_iter_next, dw2_expand_marked_cus): Limit
PR gold/15646 workaround to symbol kind "type".
Gary Benson [Thu, 28 May 2020 13:18:36 +0000 (14:18 +0100)]
Fix "'operator new' should not return NULL" errors in testsuite
When running the testsuite with clang, gdb.linespec/cpls-ops.cc
fails to compile with the following errors:
warning: 'operator new' should not return a null pointer unless
it is declared 'throw()' or 'noexcept' [-Wnew-returns-null]
warning: 'operator new[]' should not return a null pointer unless
it is declared 'throw()' or 'noexcept' [-Wnew-returns-null]
This prevents the gdb.linespec/cpls-ops.exp testcase from executing.
This commit fixes.
Alan Modra [Thu, 28 May 2020 09:16:17 +0000 (18:46 +0930)]
PR26044, Some targets can't be compiled with GCC 10 (tilepro)
Since this value is used in fields of type tilepro_pipeline (as
NO_PIPELINE, see tc-tilepro.c) it is appropriate to put it in
the tilepro_pipelen enum. This avoids a warning about converting from
one enum type to another.
PR 26044
* opcode/tilepro.h (TILEPRO_NUM_PIPELINE_ENCODINGS): Move to
tilepro_pipeline enum.
Add a configure option, --enable-textrel-check=[no|yes|warning|error],
to decide what ELF linker should do by default with DT_TEXTREL in an
executable or shared library.
Nick Clifton [Thu, 28 May 2020 09:51:29 +0000 (10:51 +0100)]
Linker: Remove support for -Map= with an empty argument.
* lexsup.c (parse_args): Generate an error if a name is not
provided to the -Map option.
(ld_options): Mention that the -Map option supports a directory
name as an argument.
* NEWS: Remove mention of support for an empty string as an
argument to -Map.
* ld.texi: Likewise.
Stephen Casner [Thu, 28 May 2020 01:40:38 +0000 (18:40 -0700)]
Fix PR gas/26001 (pdp11-*-*)
PR gas/26001
* gas/config/tc-pdp11.c (parse_reg): Distinguish register names from
symbols that begin with a register name.
* gas/testsuite/gas/pdp11/pdp11.exp: Add test of such symbols.
* gas/testsuite/gas/pdp11/pr26001.s: Likewise.
* gas/testsuite/gas/pdp11/pr26001.d: Likewise.
Simon Cook [Wed, 27 May 2020 16:34:58 +0000 (17:34 +0100)]
RISC-V: Fix missing initialization of riscv_csr_extra structs
The next pointer of struct riscv_csr_extra was not always initilized
to NULL or a valid pointer, causing the assembler to attempt to read
through an uninitialized pointer on startup.
gas/ChangeLog:
* config/tc-riscv.c (riscv_init_csr_hash): NULL initilize next
pointer when creating struct riscv_csr_extra.
Pedro Alves [Wed, 27 May 2020 18:59:19 +0000 (19:59 +0100)]
cp-completion-aliases.exp: Use test_gdb_complete_{unique,multiple}
gdb.linespec/cp-completion-aliases.exp is calling
test_gdb_complete_{tab,cmd}_unique and
test_gdb_complete_{tab,cmd}_multiple separately for each use case.
I.e., testing once for TAB completion and once for the "complete"
command. There's no need to do that explicitly and separately, we
have wrapper procedures to do that for us.
gdb/testsuite/ChangeLog:
2020-05-27 Pedro Alves <palves@redhat.com>
* gdb.linespec/cp-completion-aliases.exp: Remove readline_is_used
check. Use test_gdb_complete_unique instead of
test_gdb_complete_tab_unique + test_gdb_complete_cmd_unique. Use
test_gdb_complete_multiple instead of
test_gdb_complete_tab_multiple + test_gdb_complete_cmd_multiple.
Tom Tromey [Wed, 27 May 2020 17:48:18 +0000 (11:48 -0600)]
Use add_partial_symbol in load_partial_dies
An earlier patch added the add_partial_symbol helper function to
dwarf2/read.c. However, a couple of calls to add_psymbol_to_list were
left in place. It turns out that these calls slow down partial symbol
reading, because they still go via the path that tries to needlessly
demangle already-demangled names.
This patch improves the performance of partial symbol reading by
changing this code to use add_partial_symbol instead.
The run previous to this had times of (see the first patch in the
series for an explanation):
gdb 1.64
libxul 1.99
Ada 2.47
This patch improves the times to:
gdb 1.47
libxul 1.89
Ada 2.39
gdb/ChangeLog
2020-05-27 Tom Tromey <tromey@adacore.com>
* dwarf2/read.c (load_partial_dies): Use add_partial_symbol.
Tom Tromey [Wed, 27 May 2020 17:48:18 +0000 (11:48 -0600)]
Inline abbrev lookup
Profiling showed that calls to abbrev_table::lookup_abbrev were "too
visible". As these are just forwarding calls to the hash table, this
patch inlines the lookup. Also, htab_find_with_hash is used, avoiding
another call.
The run previous to this had times of (see the first patch in the
series for an explanation):
gdb 1.69
libxul 2.02
Ada 2.52
This patch improves the times to:
gdb 1.64
libxul 1.99
Ada 2.47
gdb/ChangeLog
2020-05-27 Tom Tromey <tromey@adacore.com>
* dwarf2/abbrev.h (struct abbrev_table) <lookup_abbrev>: Inline.
Use htab_find_with_hash.
<add_abbrev>: Remove "abbrev_number" parameter.
* dwarf2/abbrev.c (abbrev_table::add_abbrev): Remove
"abbrev_number" parameter. Use htab_find_slot_with_hash.
(hash_abbrev): Add comment.
(abbrev_table::lookup_abbrev): Move to header file.
(abbrev_table::read): Update.
Tom Tromey [Wed, 27 May 2020 17:48:18 +0000 (11:48 -0600)]
Lazily compute partial DIE name
Currently the name of a partial DIE is computed eagerly. However, the
name is not always needed. This patch changes partial DIEs to compute
their name lazily, improving performance by avoiding unnecessary name
computations.
The run previous to this had times of (see the first patch in the
series for an explanation):
gdb 1.88
libxul 2.11
Ada 2.60
This patch improves the times to:
gdb 1.69
libxul 2.02
Ada 2.52
gdb/ChangeLog
2020-05-27 Tom Tromey <tromey@adacore.com>
* dwarf2/read.c (struct partial_die_info) <name>: Declare new
method.
<canonical_name>: New member.
<raw_name>: Rename from "name".
(partial_die_info): Initialize canonical_name.
(scan_partial_symbols): Check raw_name.
(partial_die_parent_scope, partial_die_full_name)
(add_partial_symbol, add_partial_subprogram)
(add_partial_enumeration, load_partial_dies): Use "name" method.
(partial_die_info::name): New method.
(partial_die_info::read, guess_partial_die_structure_name)
(partial_die_info::fixup): Update.
Tom Tromey [Wed, 27 May 2020 17:48:18 +0000 (11:48 -0600)]
Attribute method inlining
This inlines a couple of methods on struct attribute, improving the
performance of DWARF partial symbol reading. These methods were
discovered as hot spots using callgrind.
For this patch, and for all the patches in this series, I tested gdb's
performance on three programs:
1. gdb itself -- I built gdb and copied it to /tmp, ensuring that the
same version was used in all tests.
2. The system libxul.so, the main library of Firefox. I installed the
separate debuginfo and ensured that gdb read it.
Hannes Domani [Sun, 24 May 2020 20:59:33 +0000 (22:59 +0200)]
Don't close thread handles provided by WaitForDebugEvent
I sometimes encountered a weird breakpoint failure when using start:
(gdb) start
Temporary breakpoint 2 at 0x40162d: file gdb-25911.c, line 4.
Starting program: C:\src\tests\gdb-25911.exe
Warning:
Cannot insert breakpoint 2.
Cannot access memory at address 0x401628
After trying a lot of combinations, I found a way to reproduce it:
(gdb) file gdb-25987.exe
Reading symbols from gdb-25987.exe...
(gdb) start
Temporary breakpoint 1 at 0x401638: file gdb-25987.cpp, line 13.
Starting program: C:\src\tests\gdb-25987.exe
Temporary breakpoint 1, main () at gdb-25987.cpp:13
13 int main() {
(gdb) c
Continuing.
Program received signal SIGSEGV, Segmentation fault.
MyClass::call (this=0x3d20d0) at gdb-25987.cpp:8
8 *(char*)(nullptr) = 1;
(gdb) kill
Kill the program being debugged? (y or n) y
[Inferior 1 (process 1140) killed]
(gdb) file gdb-25911.exe
Load new symbol table from "gdb-25911.exe"? (y or n) y
Reading symbols from gdb-25911.exe...
(gdb) start
Temporary breakpoint 2 at 0x40162d: file gdb-25911.c, line 4.
Starting program: C:\src\tests\gdb-25911.exe
Warning:
Cannot insert breakpoint 2.
Cannot access memory at address 0x401628
Command aborted.
The actual failure was that ReadProcessMemory used a process handle that
was no longer valid.
And the underlying reason was that the windows_thread_info destructor
closes a thread handle that was provided earlier by WaitForDebugEvent.
But since this is not allowed (and it was actually already closed at this
point, and the handle value reused), this closed another still-needed handle.
gdb/ChangeLog:
2020-05-27 Hannes Domani <ssbssa@yahoo.de>
* nat/windows-nat.c (windows_thread_info::~windows_thread_info):
Don't close thread handle.
Nick Clifton [Wed, 27 May 2020 16:49:17 +0000 (17:49 +0100)]
[PATCH] allow empty string as argument to -Map
* lexsup.c (parse_args): If the map filename is defined but empty
create a name based upon the output file name. If the name is
defined but refers to a directory create a file inside the
directory based on the output file name.
* ld.texi: Document the new feature.
* testsuite/ld-script/map-address.exp: Add test of new feature.
* NEWS: Mention the new feature.
Tom Tromey [Wed, 27 May 2020 15:20:14 +0000 (11:20 -0400)]
Share DWARF partial symtabs
This changes the DWARF reader to share partial symtabs (or indices if
they are available) across objfiles. This has a few parts.
* If multiple objfiles backed by the same BFD can share partial symtabs
(see below), a single dwarf2_per_bfd is created. It is stored in the
per-bfd `dwarf2_per_bfd_bfd_data_key` registry. Multiple
dwarf2_per_objfile objects will point to the same instance. The
lifetime of these dwarf2_per_bfd objects is naturally handled. When
all the objfiles using the BFD are destroyed, the BFD's refount drops
to 0, which triggers the removal of the corresponding dwarf2_per_bfd
object from the registry and its destruction.
* If multiple objfiles backed by the same BFD can't share partial
symtabs (see below), one dwarf2_per_bfd object is created for each
objfile. Each dwarf2_per_objfile will point to their own instance of
dwarf2_per_bfd. These instances of dwarf2_per_bfd are kept in a
per-objfile registry, meaning that when the objfile gets destroyed,
the dwarf2_per_bfd instance gets destroyed as well.
* objfile::partial_symtabs is changed to be a shared_ptr again. This
lets us stash a second reference in dwarf2_per_bfd; if the DWARF
data is being shared, we can simply copy this value to the new
objfile.
* Two dwarf2_per_objfile objects backed by the same BFD may share a
dwarf2_per_bfd instance if:
* No other symbol reader has found symbols, and
* No BFD section rqeuires relocation
gdb/ChangeLog:
YYYY-MM-DD Tom Tromey <tom@tromey.com>
YYYY-MM-DD Simon Marchi <simon.marchi@efficios.com>
* objfiles.h (struct objfile) <partial_symtabs>: Now a
shared_ptr.
* dwarf2/read.h (struct dwarf2_per_objfile) <partial_symtabs>: New
member.
* dwarf2/read.c (dwarf2_per_bfd_bfd_data_key,
dwarf2_per_bfd_objfile_data_key>: New globals.
(dwarf2_has_info): Use shared dwarf2_per_bfd if possible.
(dwarf2_get_section_info): Use get_dwarf2_per_objfile.
(dwarf2_initialize_objfile): Consider cases where per_bfd can be
shared.
(dwarf2_build_psymtabs): Set objfile::partial_symtabs and
short-circuit when sharing.
(dwarf2_build_psymtabs): Set dwarf2_per_objfile::partial_symtabs.
(dwarf2_psymtab::expand_psymtab): Use free_cached_comp_units.
Simon Marchi [Wed, 27 May 2020 15:14:12 +0000 (11:14 -0400)]
Move line_header_hash to dwarf2_per_objfile
The `line_header_hash` field of `struct dwarf2_per_bfd` contains some
`struct line_header` objects. A `struct line_header` objects contains
some `file_entry` objects. A `file_entry` object contains a pointer to
the `symtab` object created from it. The `line_header_hash` is
therefore ultimately objfile-dependent and can't be shared as-is between
objfiles.
Move it from `dwarf2_per_bfd` to `dwarf2_per_objfile`.
Simon Marchi [Wed, 27 May 2020 15:14:11 +0000 (11:14 -0400)]
Make mapped_debug_names independent of objfile
mapped_debug_names currently has a dwarf2_per_objfile field. Since we
want it to become objfile-independent, this field must be removed.
This patch removes it, and then arranges for all methods that needed it
to accept a dwarf2_per_objfile parameter. This trickles down at various
places, like the dw2_debug_names_iterator type.
Ultimately, the objfile only seems to be needed because we might need to
read a string from the string section. For that, we might need to read
in the section, and if it's a relocatable section, the objfile is needed
in order to do the relocation. This pattern happens often (that we to
pass an objfile only because a section might be read). I think it's a
bit ugly, but I don't have a good alternative right now.
Simon Marchi [Wed, 27 May 2020 15:14:11 +0000 (11:14 -0400)]
Replace dwarf2_per_cu_data::cu backlink with per-objfile map
The dwarf2_per_cu_data type is going to become objfile-independent,
while the dwarf2_cu type will stay object-dependent. This patch removes
the backlink from dwarf2_per_cu_data to dwarf2_cu, in favor of the
dwarf2_per_objfile::m_dwarf2_cus map. It maps dwarf2_per_cu_data
objects to the corresponding dwarf2_cu objects for this objfile. If a
CU has been read in in the context of this objfile, then an entry will
be present in the map.
The dwarf2_cu objects that are read in are currently kept in a linked
list rooted in the dwarf2_per_bfd. Except that the dwarf2_cu objects
are not simply linked together, they are interleaved with their
corresponding dwarf2_per_cu_data objects. So if we have CUs A and B
read in, the dwarf2_per_bfd::read_in_chain will point to a chain like
this (DPCD == dwarf2_per_cu_data, DC == dwarf2_cu):
DPCD A -> DC A -> DPCD B -> DC B
Obviously, this can't stay as is, since a same CU can be read in for an
objfile but not read in for another objfile sharing the same BFD, and
the dwarf2_per_cu_data::cu link is removed. This is all replaced by
the dwarf2_per_objfile::m_dwarf2_cus map.
gdb/ChangeLog:
* dwarf2/read.h (struct dwarf2_cu): Forward-declare.
(struct dwarf2_per_bfd) <free_cached_comp_units>: Remove,
move to dwarf2_per_objfile.
<read_in_chain>: Remove.
(struct dwarf2_per_objfile) <get_cu, set_cu, remove_cu,
remove_all_cus, age_comp_units>: New methods.
<m_dwarf2_cus>: New member.
(struct dwarf2_per_cu_data) <cu>: Remove.
* dwarf2/read.c (struct dwarf2_cu) <read_in_chain>: Remove.
(age_cached_comp_units, free_one_cached_comp_unit): Remove,
moved to methods of dwarf2_per_objfile.
(dwarf2_clear_marks): Remove.
(dwarf2_queue_item::~dwarf2_queue_item): Update.
(dwarf2_per_bfd::~dwarf2_per_bfd): Don't free dwarf2_cus.
(dwarf2_per_bfd::free_cached_comp_units): Remove.
(dwarf2_per_objfile::remove_all_cus): New.
(class free_cached_comp_units) <~free_cached_comp_units>:
Update.
(load_cu): Update.
(dw2_do_instantiate_symtab): Adjust.
(fill_in_sig_entry_from_dwo_entry): Adjust.
(cutu_reader::init_tu_and_read_dwo_dies): Update.
(cutu_reader::cutu_reader): Likewise.
(cutu_reader::keep): Use dwarf2_per_objfile::set_cu.
(cutu_reader::cutu_reader): Use dwarf2_per_objfile::get_cu.
(process_psymtab_comp_unit): Use dwarf2_per_objfile::remove_cu
and dwarf2_per_objfile::age_comp_units.
(load_partial_comp_unit): Update.
(maybe_queue_comp_unit): Use dwarf2_per_objfile::get_cu.
(process_queue): Likewise.
(find_partial_die): Use dwarf2_per_objfile::get_cu instead of cu
backlink.
(dwarf2_read_addr_index): Likewise.
(follow_die_offset): Likewise.
(dwarf2_fetch_die_loc_sect_off): Likewise.
(dwarf2_fetch_constant_bytes): Likewise.
(dwarf2_fetch_die_type_sect_off): Likewise.
(follow_die_sig_1): Likewise.
(load_full_type_unit): Likewise.
(read_signatured_type): Likewise.
(dwarf2_cu::dwarf2_cu): Don't set cu field.
(dwarf2_cu::~dwarf2_cu): Remove.
(dwarf2_per_objfile::get_cu): New.
(dwarf2_per_objfile::set_cu): New.
(age_cached_comp_units): Rename to...
(dwarf2_per_objfile::age_comp_units): ... this. Adjust
to std::unordered_map.
(free_one_cached_comp_unit): Rename to...
(dwarf2_per_objfile::remove_cu): ... this. Adjust
to std::unordered_map.
(dwarf2_per_objfile::~dwarf2_per_objfile): New.
(dwarf2_mark_helper): Use dwarf2_per_objfile::get_cu, expect
a dwarf2_per_objfile in data.
(dwarf2_mark): Pass dwarf2_per_objfile in data to htab_traverse.
(dwarf2_clear_marks): Remove.
Simon Marchi [Wed, 27 May 2020 15:14:10 +0000 (11:14 -0400)]
Pass existing_cu object to cutu_reader
It is possible, seemingly for a special case described in
find_partial_die, for cutu_reader to re-use an existing dwarf2_cu
instead of creating a new one. This happens when running this test, for
example:
make check TESTS="gdb.dwarf2/fission-reread.exp"
Right now the, `use_existing_cu` flag tells cutu_reader to use the
dwarf2_cu object at dwarf2_per_cu_data::cu. However, we'll remove that
field, so we need to find another solution.
This situation arises when some caller up the stack has already created
the dwarf2_cu to read a dwarf2_per_cu_data, but needs to re-read it with
some other parameters. Therefore, it's possible to just have that
caller pass down the dwarf2_cu object to use as a `existing_cu`
parameter. If `existing_cu` is NULL, it tells cutu_reader that it needs
to instantiate a new one.
Simon Marchi [Wed, 27 May 2020 15:14:10 +0000 (11:14 -0400)]
Add comp_unit_head to dwarf2_per_cu_data
The per_cu_header_read_in function allows obtaining a filled
comp_unit_head object for a given dwarf2_per_cu_data object. If a
dwarf2_cu object exists for this dwarf2_per_cu_data, then it just
returns a pointer to the comp_unit_head from that dwarf2_cu. Otherwise,
it reads the header into a temporary buffer provided by the caller, and
returns a pointer to that.
Since the dwarf2_per_cu_data::cu link is going to be removed
(dwarf2_per_cu_data will become objfile-independent while dwarf2_cu
stays objfile-dependent), we cannot rely anymore on returning the header
from the dwarf2_cu object.
The not too complex solution implemented by this patch is to keep a copy
of the header in the dwarf2_per_cu_data object, independent from the
copy in dwarf2_cu. The new copy is only used in the addr_size,
offset_size and ref_addr_size methods of dwarf2_per_cu_data.
There's nothing intrinsic to the comp_unit_head object that prevents it
to be shared between two dwarf2_cu objects (belonging to different
objfiles) representing the same CU. In other words, I think we could
eventually get rid of the copy in dwarf2_cu to only keep the one in
dwarf2_per_cu_data. It is not trivial, however, so I have decided not
to do it for the moment.
Simon Marchi [Wed, 27 May 2020 15:14:09 +0000 (11:14 -0400)]
Make load_cu return the loaded dwarf2_cu
In a subsequent patch, the dwarf2_per_cu_data::cu link will be removed.
dwarf2_cu objects will instead need to be looked up from a per-objfile
map, using the dwarf2_per_cu_data object as the key.
To make it easier for some callers, this patch makes load_cu return the
dwarf2_cu it creates. If the caller needs to use the created dwarf2_cu,
it will have it available right away, rather than having to do a map
lookup.
At the same time, this allows changing queue_and_load_all_dwo_tus to
take a dwarf2_cu instead of a dwarf2_per_cu_data.
gdb/ChangeLog:
* dwarf2/read.c (load_cu): Return dwarf2_cu.
(dw2_do_instantiate_symtab): Update.
(queue_and_load_all_dwo_tus): Change parameter from
dwarf2_per_cu_data to dwarf2_cu.
(dwarf2_fetch_die_loc_sect_off): Update.
(dwarf2_fetch_constant_bytes): Update.
(dwarf2_fetch_die_type_sect_off): Update.
Simon Marchi [Wed, 27 May 2020 15:14:09 +0000 (11:14 -0400)]
Pass dwarf2_cu to process_full_{comp,type}_unit
These two functions work on a dwarf2_cu. It is currently obtained from
the per_cu->cu link, which we want to remove. Make them accept the
dwarf2_cu directly as a parameter. This moves the per_cu->cu references
one level up, but that one will be removed too in a subsequent patch.
Tom Tromey [Wed, 27 May 2020 15:19:35 +0000 (11:19 -0400)]
Move signatured_type::type to unshareable object
signatured_type has a link to the "struct type". However, types are
inherently objfile-specific, so once sharing is implemented, this will
be incorrect.
This patch moves the type to a new map in the DWARF unshareable
object.
gdb/ChangeLog:
YYYY-MM-DD Tom Tromey <tom@tromey.com>
YYYY-MM-DD Simon Marchi <simon.marchi@efficios.com>
* dwarf2/read.h (struct dwarf2_per_objfile)
<get_type_for_signatured_type, set_type_for_signatured_type>:
New methods.
<m_type_map>: New member.
(struct signatured_type) <type>: Remove.
* dwarf2/read.c
(dwarf2_per_objfile::get_type_for_signatured_type,
dwarf2_per_objfile::set_type_for_signatured_type): New.
(get_signatured_type): Use new methods.
Tom Tromey [Wed, 27 May 2020 15:19:09 +0000 (11:19 -0400)]
Split type_unit_group
type_unit_group has links to the compunit_symtab and other symtabs.
However, once this object is shared across objfiles, this will no
longer be ok.
This patch introduces a new type_unit_group_unshareable and arranges to
store a map from type unit groups to type_unit_group_unshareable objects
in dwarf2_per_objfile.
gdb/ChangeLog:
YYYY-MM-DD Tom Tromey <tom@tromey.com>
YYYY-MM-DD Simon Marchi <simon.marchi@efficios.com>
* dwarf2/read.h (struct type_unit_group_unshareable): New.
(struct dwarf2_per_objfile) <type_units>: New member.
<get_type_unit_group_unshareable>: New method.
* dwarf2/read.c (struct type_unit_group) <compunit_symtab,
num_symtabs, symtabs>: Remove; move to
type_unit_group_unshareable.
(dwarf2_per_objfile::get_type_unit_group_unshareable): New.
(process_full_type_unit, dwarf2_cu::setup_type_unit_groups)
(dwarf2_cu::setup_type_unit_groups): Use type_unit_group_unshareable.
Simon Marchi [Wed, 27 May 2020 15:14:04 +0000 (11:14 -0400)]
Remove dwarf2_per_cu_data::objfile ()
Since dwarf2_per_cu_data objects are going to become
objfile-independent, the backlink from dwarf2_per_cu_data to one
particular objfile must be removed. Instead, users of
dwarf2_per_cu_data that need an objfile must know from somewhere else in
the context of which objfile they are using this CU.
This also helps remove a dwarf2_per_cu_data::dwarf2_per_objfile
reference (from where the objfile was obtained).
Note that the dwarf2_per_cu_data::objfile method has a special case to
make sure to return the main objfile, if the objfile associated to the
dwarf2_per_cu_data is a separate debug objfile. I don't really know if
this is necessary: I ignored that, and didn't see any regression when
testing with the various Dejagnu boards with separate debug info, so I
presume it wasn't needed. If it turns out this was needed, then we can
have a helper method on the objfile type for that.
Simon Marchi [Wed, 27 May 2020 15:14:01 +0000 (11:14 -0400)]
Remove dwarf2_per_cu_data::text_offset
This method simply returns the text offset of the objfile associated to
the dwarf2_per_cu_data object. Since dwarf2_per_cu_data objects are
going to become objfile-independent, we can't keep this method. This
patch removes it.
Existing callers need to figure out the in the context of which objfile
this is being used, and call text_offset on it. Typically, this comes
from a symbol baton, where we store the corresponding
dwarf2_per_objfile.
gdb/ChangeLog:
* dwarf2/read.h (struct dwarf2_per_cu_data) <text_offset>:
Remove.
* dwarf2/read.c (dwarf2_per_cu_data::text_offset): Remove.
* dwarf2/loc.c (dwarf2_find_location_expression): Update.
(dwarf2_compile_property_to_c): Update.
(dwarf2_compile_expr_to_ax): Add dwarf2_per_objfile parameter,
use text offset from objfile.
(locexpr_tracepoint_var_ref): Update.
(locexpr_generate_c_location): Update.
(loclist_describe_location): Update.
(loclist_tracepoint_var_ref): Update.
* dwarf2/compile.h (compile_dwarf_bounds_to_c): Add
dwarf2_per_objfile parameter.
* dwarf2/loc2c.c (do_compile_dwarf_expr_to_c): Likewise,
use text offset from objfile.
(compile_dwarf_expr_to_c): Add dwarf2_per_objfile parameter.
Simon Marchi [Wed, 27 May 2020 15:14:01 +0000 (11:14 -0400)]
Add dwarf2_per_objfile to dwarf_expr_context and dwarf2_frame_cache
Evaluating DWARF expressions (such as location expressions) requires
knowing about the current objfile. For example, it may call functions
like dwarf2_fetch_die_loc_sect_off, which currently obtain the
dwarf2_per_objfile object it needs from the dwarf2_per_cu_data object.
However, since we are going to remove this
dwarf2_per_cu_data::dwarf2_per_objfile link, these functions will need
to obtain the current dwarf2_per_objfile by parmeter.
If we go up the stack, we see that the DWARF expression contexts
(dwarf_expr_context and the classes that derive from it) need to store
the dwarf2_per_objfile, to be able to pass it to those functions that
will need it.
This patch adds a constructor to all these dwarf_expr_context variants,
accepting a dwarf2_per_objfile parameter. This dwarf2_per_objfile
generally comes from a symbol baton created earlier.
For frame-related expressions, the dwarf2_per_objfile object must be
passed through the dwarf2_frame_cache object. This lead to the
dwarf2_frame_find_fde function returning (by parameter) a
dwarf2_per_objfile object. I then realized that this made the existing
"out_offset" parameter redundant. This offset is
`objfile->text_section_offset ()`, so it can be recomputed from the
dwarf2_per_objfile object at any time. I therefore opted to remove this
output parameter, as well as the offset field of dwarf2_frame_cache.
*Note*, there's one spot I'm particularly unsure about. In
dwarf_evaluate_loc_desc::push_dwarf_reg_entry_value, we would save and
overwrite the offset value in the context, along with a bunch of other
state. This is because we might be about to evaluate something in a
different CU that the current one. If the two CUs are in the same
objfile, then the text_offset is the same, as it's a property of the
objfile. However, if the two CUs are possibly in different objfiles,
then it means the text_offsets are different. It would also mean we
would need to save and restore the dwarf2_per_objfile in the context.
Is that even possible?
gdb/ChangeLog:
* dwarf2/expr.h (struct dwarf_expr_context)
<dwarf_expr_context>: Add dwarf2_per_objfile parameter.
<offset>: Remove.
<per_objfile>: New member.
* dwarf2/expr.c (dwarf_expr_context::dwarf_expr_context): Add
dwarf2_per_objfile parameter. Don't set offset, set
per_objfile.
(dwarf_expr_context::execute_stack_op): Use offset from objfile.
* dwarf2/frame.c (dwarf2_frame_find_fde): Return (by parameter)
a dwarf2_per_objfile object instead of an offset.
(class dwarf_expr_executor) <dwarf_expr_executor>: Add
constructor.
(execute_stack_op): Add dwarf2_per_objfile parameter, pass it
to dwarf2_expr_executor constructor. Don't set offset.
(dwarf2_fetch_cfa_info): Update.
(struct dwarf2_frame_cache) <text_offset>: Remove.
<per_objfile>: New field.
(dwarf2_frame_cache): Update.
(dwarf2_frame_prev_register): Update.
* dwarf2/loc.c (class dwarf_evaluate_loc_desc)
<dwarf_evaluate_loc_desc>: Add constructor.
(dwarf2_evaluate_loc_desc_full): Update.
(dwarf2_locexpr_baton_eval): Update.
(class symbol_needs_eval_context) <symbol_needs_eval_context>:
Add constructor.
(dwarf2_loc_desc_get_symbol_read_needs): Update.
Simon Marchi [Wed, 27 May 2020 15:14:00 +0000 (11:14 -0400)]
Move int type methods out of dwarf2_per_cu_data
These methods rely on the current objfile to create types based on it.
Since dwarf2_per_cu_data is to become objfile-independent, these methods
need to mvoe.
int_type can be in dwarf2_per_objfile, as it only requires knowing about
the objfile.
addr_sized_int_type and addr_type also need to know about the DWARF
address type size, which is CU-specific. The dwarf2_cu objects seems
like a good place for it, as it knows both about the current objfile and
the current CU.
Simon Marchi [Wed, 27 May 2020 15:13:59 +0000 (11:13 -0400)]
Remove reference to dwarf2_per_cu_data::dwarf2_per_objfile in queue_and_load_all_dwo_tus
In this context, we know that per_cu->cu will be set, as there is this
assertion:
gdb_assert (per_cu->cu != NULL)
So in order to remove the dwarf2_per_cu_data::dwarf2_per_objfile
reference in queue_and_load_all_dwo_tus, we can go through per_cu->cu.
This adds a reference to dwarf2_per_cu_data::cu, but it will get removed
eventually, in a subsequent patch.
gdb/ChangeLog:
* dwarf2/read.c (queue_and_load_all_dwo_tus): Access per_objfile
data through per_cu->cu.
Simon Marchi [Wed, 27 May 2020 15:13:59 +0000 (11:13 -0400)]
Pass dwarf2_cu objects to dwo-related functions, instead of dwarf2_per_cu_data
This allows removing references to the
dwarf2_per_cu_data::dwarf2_per_objfile field.
I am not too sure of the code flow here, but ultimately
open_and_init_dwo_file calls create_cus_hash_table, and passes it
per_cu->cu. create_cus_hash_table requires a dwarf2_cu to pass to
cutu_reader, as the "parent_cu".
The dwarf2_per_cu_data::cu link is only set when in a certain context.
It's not easy to convince myself in which situations it's safe to use
it. Instead, if a function is going to use a dwarf2_cu, I think it's
simpler if it takes that object directly. If it needs access to the
corresponding dwarf2_per_cu_data object, then it can used the
dwarf2_cu::per_cu field, which we know is always set.
This patch adds some references to dwarf2_per_cu_data::cu in the
cutu_reader context. In this context, we know this field will be set,
as it's cutu_reader that is responsible for instantiating the dwarf2_cu
and assigning the field.
Simon Marchi [Wed, 27 May 2020 15:13:56 +0000 (11:13 -0400)]
Remove dwarf2_per_cu_data::dwarf2_per_objfile reference in cutu_reader::keep
Here, it should be safe to use dwarf2_per_cu_data->cu->per_objfile, as
we know that dwarf2_per_cu_data->cu will be set at this point.
Note that this adds a reference to dwarf2_per_cu_data::cu, which we'll
want to remove later, but the current focus is to remove references to
dwarf2_per_cu_data::dwarf2_per_objfile. We'll deal with that in a
subsequent patch.
gdb/ChangeLog:
* dwarf2/read.c (cutu_reader::keep): Access dwarf2_per_objfile
object through m_this_cu->cu.
Simon Marchi [Wed, 27 May 2020 15:13:56 +0000 (11:13 -0400)]
Make queue_and_load_dwo_tu receive a dwarf2_cu
queue_and_load_dwo_tu, used as a callback for htab_traverse_noresize,
currently receives a dwarf2_per_cu_data as its `info` user data. It
accesses the current dwarf2_cu object through the dwarf2_per_cu_data::cu field.
This field will be removed, because the dwarf2_per_cu_data will become
objfile-independent, while dwarf_cu will remain objfile-dependent.
To remove references to this field, change queue_and_load_dwo_tu so
that it expects to receive a pointer to the dwarf2_cu as its info
parameter.
A reference to dwarf2_per_cu_data::cu needs to be added, but it will get
removed in a subsequent patch, when this function gets re-worked.
I kept this as a separate patch, because since there's no strong typing
here, it's easy to miss something.
gdb/ChangeLog:
* dwarf2/read.c (queue_and_load_dwo_tu): Expect a dwarf2_cu as
the info parameter.
(queue_and_load_all_dwo_tus): Pass per_cu->cu.
Simon Marchi [Wed, 27 May 2020 15:13:55 +0000 (11:13 -0400)]
Add dwarf2_per_objfile parameter to cutu_reader's constructors
The cutu_reader type is used for reading the CU represented by the
passed dwarf2_per_cu_data object. This reading is done in the context
of a given obfile, which is currently the one associated to the passed
dwarf2_per_cu_data object. Since the dwarf2_per_cu_data type will
become objfile-independent, we will need to pass the objfile separately.
This patch therefore adds a dwarf2_per_objfile parameter to the
cutu_reader constructors, as well as to their callers, up until the
point where we can get the dwarf2_per_objfile object from somewhere
else. In the end, this allows removing the reference to
dwarf2_per_cu_data::dwarf2_per_objfile in cutu_reader::cutu_reader.
A few dwarf2_per_cu_data::dwarf2_per_objfile references are added (e.g.
in dwarf2_fetch_die_type_sect_off). This is temporary, this will be
removed once these functions will get re-worked in subsequent patches.
Simon Marchi [Wed, 27 May 2020 15:13:55 +0000 (11:13 -0400)]
Use bfd_get_filename instead of objfile_name in lookup_dwo_unit
There should be no functional difference, as objfile_name defers to
bfd_get_filename if objfile::obfd is non-NULL, which should be the case
here. This allows to remove a reference to
dwarf2_per_cu_data::dwarf2_per_objfile.
gdb/ChangeLog:
* dwarf2/read.c (lookup_dwo_unit): Use bfd_get_filename instead
of objfile_name.
Simon Marchi [Wed, 27 May 2020 15:13:54 +0000 (11:13 -0400)]
Make dwarf2_get_dwz_file take a dwarf2_per_bfd
This allows removing a per_bfd->dwarf2_per_objfile reference in
get_abbrev_section_for_cu.
This requires saving the bfd in dwarf2_per_bfd. The constructor of
dwarf2_per_bfd already accepts the bfd, so it's just a matter of saving
it in a field.
I replaced uses of objfile_name with bfd_get_filename, which should be
equivalent in this case.
Simon Marchi [Wed, 27 May 2020 15:13:54 +0000 (11:13 -0400)]
Add dwarf2_per_bfd field to dwarf2_per_cu_data
Some code using dwarf2_per_cu_data objects accesses the corresponding
dwarf2_per_bfd using the following pattern:
per_cu->dwarf2_per_objfile->per_bfd
Since dwarf2_per_cu_data objects are going to become
objfile-independent, the dwarf2_per_objfile link must go. To replace
it, add a dwarf2_per_cu_data->per_bfd link. It makes sense to have it
there because the dwarf2_per_cu_data objects belong to the
dwarf2_per_bfd, so this is essentially just a backlink to their owner.
gdb/ChangeLog:
* dwarf2/read.h (struct dwarf2_per_cu_data) <per_bfd>: New
member.
* dwarf2/read.c (dwarf2_per_bfd::allocate_per_cu): Initialize
dwarf2_per_cu_data::per_bfd.
(dwarf2_per_bfd::allocate_signatured_type): Likewise.
(create_type_unit_group): Likewise.
(queue_comp_unit): Remove reference to
per_cu->dwarf2_per_objfile.
(maybe_queue_comp_unit): Likewise.
(fill_in_sig_entry_from_dwo_entry): Assign new field.
(create_cus_hash_table): Assign new field.
Change spots that access the dwarf2_per_objfile object through this
pattern:
dwarf2_cu->per_cu->dwarf2_per_objfile
to
dwarf2_cu->per_objfile
This allows removing many references to
dwarf2_per_cu_data::dwarf2_per_objfile.
Again, I hope the following ChangeLog entry will be fine. I'd rather not
list all the affected functions, as it would be time-consuming and a bit
pointless.
gdb/ChangeLog:
* dwarf2/read.c: Replace
dwarf2_cu->per_cu->dwarf2_per_objfile references with
dwarf2_cu->per_objfile throughout.
Simon Marchi [Wed, 27 May 2020 15:13:52 +0000 (11:13 -0400)]
Remove reference to dwarf2_per_cu_data::dwarf2_per_objfile in dw2_do_instantiate_symtab
This patch begins by removing the per_cu->dwarf2_per_objfile reference
in dw2_do_instantiate_symtab, instead accepting a dwarf2_per_objfile
object as a parameter. It then fixes the fallouts. In this context,
the dwarf2_per_objfile is generally derived from an objfile passed to a
quick_symbol_functions callback.
Simon Marchi [Wed, 27 May 2020 15:13:52 +0000 (11:13 -0400)]
Add dwarf2_per_objfile field to dwarf2_cu
Subsequent patches will make dwarf2_per_cu_data objfile-independent.
This means that the dwarf2_per_cu_data::dwarf2_per_objfile field must
go.
The code using a dwarf2_cu structure currently accesses the current
dwarf2_per_objfile object through dwarf2_cu->per_cu->dwarf2_per_objfile.
Since it's ok for the dwarf2_cu to know about the current objfile (a
dwarf2_cu is always used in the context of a particular objfile), add a
dwarf2_per_objfile field to dwarf2_cu. Upcoming patches will gradually
remove uses of dwarf2_per_cu_data::dwarf2_per_objfile in favor of
dwarf2_cu::dwarf2_per_objfile, until the former can be removed.
Simon Marchi [Wed, 27 May 2020 15:13:51 +0000 (11:13 -0400)]
Move die_type_hash to dwarf2_per_objfile
The die_type_hash field can't be shared between multiple obfiles, as it
holds `struct type` objects, which are objfile-specific. Move it from
dwarf2_per_bfd to dwarf2_per_objfile and update all references.
Simon Marchi [Wed, 27 May 2020 15:15:47 +0000 (11:15 -0400)]
Remove symtab links from dwarf2_psymtab and dwarf2_per_cu_quick_data
The dwarf2_psymtab and dwarf2_per_cu_quick_data types contain a pointer
to a compunit_symtab, which is a pointer to the corresponding full
symtab. The dwarf2_psymtab and dwarf2_per_cu_quick_data objects are
going to become objfile-independent, and possibly shared by multiple
objfiles, whereas compunit_symtab will stay objfile-dependent. This
backlink to the compunit_symtab must therefore be removed.
This patch replaces them with a vector in the dwarf2_per_objfile type,
that serves as a mapping from dwarf2_per_cu_data objects to
compunit_symtab objects, for this particular objfile. The vector is
indexed using the index assigned to the dwarf2_per_cu_data at its
creation.
I removed the get_compunit_symtab, as it appears to bring not much value
over calling dwarf2_per_objfile::get_symtab directly.
gdb/ChangeLog:
YYYY-MM-DD Tom Tromey <tom@tromey.com>
YYYY-MM-DD Simon Marchi <simon.marchi@efficios.com>
Simon Marchi [Wed, 27 May 2020 15:13:50 +0000 (11:13 -0400)]
Split dwarf2_per_objfile into dwarf2_per_objfile and dwarf2_per_bfd
This is the first step of splitting dwarf2_per_objfile in two, one
structure for objfile-independent data (dwarf2_per_bfd) and one for
objfile-dependent data (dwarf2_per_objfile).
The existing dwarf2_per_objfile is renamed dwarf2_per_bfd, and a new
dwarf2_per_objfile type is introduced, which sits "in between" the
objfile and dwarf2_per_bfd.
(*) Note that the dwarf2_per_objfile in the former corresponds to
the dwarf2_per_bfd in the latter.
I've done the minimal amount of changes in this patch: following patches
will incrementally move things that are not actually shareable between
objfiles from dwarf2_per_bfd to dwarf2_per_objfile.
Most references to dwarf2_per_objfile objects are changed to
dwarf2_per_objfile->per_bfd. To avoid many of these replacements, which
would have to be reverted later anyway, I've moved right away the
objfile backlink to the new dwarf2_per_objfile structure in this patch.
I've also moved the read_line_string method, since it references the
objfile backlink, and it's actually not difficult to move.
Once the moves are completed, multiple dwarf2_per_objfile sharing the
same BFD will point to the same single instance of dwarf2_per_bfd (as
long as they don't require relocation).
dwarf2_has_info, where we create these objects, is updated to the new
architecture.
I've had to change the get_gdb_index_contents_ftype typedef and related
functions. The parameter type was changed from dwarf2_per_objfile to
dwarf2_per_bfd, otherwise the template wouldn't work.
Please excuse the terse ChangeLog entry, I have not listed all the
functions where dwarf2_per_objfile has been changed to
dwarf2_per_objfile->per_bfd. It would take a considerable amount of
time and would not really be useful in the end.
gdb/ChangeLog:
* dwarf2/read.h (dwarf2_per_objfile): Rename to dwarf2_per_bfd,
then introduce a new dwarf2_per_objfile type.
<read_line_string>: Move to the new dwarf2_per_objfile type.
<objfile>: Likewise.
(dwarf2_per_bfd): Rename dwarf2_per_objfile to this.
* dwarf2/read.c: Replace references to dwarf2_per_objfile with
dwarf2_per_objfile->per_bfd.
(dwarf2_per_objfile::dwarf2_per_objfile): Rename to...
(dwarf2_per_bfd::dwarf2_per_bfd): ... this.
(dwarf2_per_objfile::free_cached_comp_units): Rename to...
(dwarf2_per_bfd::free_cached_comp_units): ... this.
(dwarf2_has_info): Allocate dwarf2_per_bfd.
(dwarf2_per_objfile::locate_sections): Rename to...
(dwarf2_per_bfd::locate_sections): ... this.
(dwarf2_per_objfile::get_cutu): Rename to...
(dwarf2_per_bfd::get_cutu): ... this.
(dwarf2_per_objfile::get_cu): Rename to...
(dwarf2_per_bfd::get_cu): ... this.
(dwarf2_per_objfile::get_tu): Rename to...
(dwarf2_per_bfd::get_tu): ... this.
(dwarf2_per_objfile::allocate_per_cu): Rename to...
(dwarf2_per_bfd::allocate_per_cu): ... this.
(dwarf2_per_objfile::allocate_signatured_type): Rename to...
(dwarf2_per_bfd::allocate_signatured_type): ... this.
(get_gdb_index_contents_ftype): Change parameter from
dwarf2_per_objfile to dwarf2_per_bfd.
* dwarf2/macro.c, dwarf2/index-write.c: Replace references to
dwarf2_per_objfile with dwarf2_per_objfile->per_bfd.
Tom Tromey [Wed, 27 May 2020 15:13:50 +0000 (11:13 -0400)]
Add dwarf2_per_objfile member to DWARF batons
Various DWARF callbacks expect to be able to fetch the objfile and / or
dwarf2_per_objfile from the DWARF CU object. However, this won't be
possible once sharing is implemented.
Because these objects are related to full symbols (e.g., they are used
to implement location expressions), they can simply store the
dwarf2_per_objfile they need.
This patch adds a per_objfile member to the various "baton" structures
and arranges to set this value when constructing the baton.
gdb/ChangeLog:
YYYY-MM-DD Tom Tromey <tom@tromey.com>
YYYY-MM-DD Simon Marchi <simon.marchi@efficios.com>
* dwarf2/loc.c (struct piece_closure) <per_objfile>: New member.
(allocate_piece_closure): Set "per_objfile" member.
(dwarf2_find_location_expression, dwarf2_locexpr_baton_eval)
(locexpr_describe_location, loclist_describe_location): Use new
member.
* dwarf2/read.c (read_call_site_scope)
(mark_common_block_symbol_computed, attr_to_dynamic_prop)
(dwarf2_const_value_attr, dwarf2_fetch_die_loc_sect_off)
(fill_in_loclist_baton, dwarf2_symbol_mark_computed,
handle_data_member_location): Set per_objfile member.
* dwarf2/loc.h (struct dwarf2_locexpr_baton) <per_objfile>: New
member.
(struct dwarf2_loclist_baton) <per_objfile>: New member.
Tom Tromey [Wed, 27 May 2020 15:13:49 +0000 (11:13 -0400)]
Add dwarf2_per_cu_data::index
Currently, a dwarf2_per_cu_data can hold a link to the corresponding
expanded compunit_symtab. However, the dwarf2_per_cu_data objects are
shared across objfiles, a simple pointer won't work: each objfile
sharing the dwarf2_per_cu_data instance will have a corresponding
compunit_symtab.
Instead, this link will be stored in the dwarf2_per_objfile object
(which will contain the objfile-specific data). To enable this, we add
an index to each dwarf2_per_cu_data and signatured_type. The data
structure in the dwarf2_per_objfile will use this new index to map a
dwarf2_per_cu_data to its corresponding compunit_symtab, for this
objfile.
gdb/ChangeLog:
* dwarf2/read.h (struct dwarf2_per_objfile) <allocate_per_cu,
allocate_signatured_type>: Declare new methods.
<m_num_psymtabs>: New member.
(struct dwarf2_per_cu_data) <index>: New member.
* dwarf2/read.c (dwarf2_per_objfile::allocate_per_cu)
(dwarf2_per_objfile::allocate_signatured_type): New methods.
(create_cu_from_index_list): Use allocate_per_cu.
(create_signatured_type_table_from_index)
(create_signatured_type_table_from_debug_names)
(create_debug_type_hash_table, add_type_unit)
(read_comp_units_from_section): Use allocate_signatured_type.
Tom Tromey [Wed, 27 May 2020 15:13:48 +0000 (11:13 -0400)]
Add "objfile" parameter to two partial_symtab methods
This series will cause partial symtabs to be shared across objfiles.
However, full symtabs and symbols will still be objfile-dependent, so
will be expanded separately for each objfile. So, a debug info reader
will need to know which objfile to consult when expanding a partial
symtab.
This patch adds an objfile parameter to the two relevant methods of
partial_symtab. Current implementations simply ignore them.
Tom Tromey [Wed, 27 May 2020 15:13:48 +0000 (11:13 -0400)]
Introduce dwarf2_per_objfile::obstack
Currently much of the DWARF-related data is stored on the objfile
obstack. This prevents sharing this data across objfiles, so this patch
adds a new obstack to dwarf2_per_objfile. Note that the
dwarf2_per_objfile type is going to become "dwarf2_per_bfd" in a
subsequent patch, which is indeed going to be shared between objfiles.
One way to check whether this is correct is to look at the remaining
uses of objfile_obstack in the DWARF code and note that they all
appear in the "full CU" code paths.
The converse -- storing per-objfile data on the shared obstack -- is
not good, but it is just a memory leak, not a potential
use-after-free. Double-checking this would also be useful, though.
gdb/ChangeLog:
* dwarf2/read.h (struct dwarf2_per_objfile) <obstack>: New
member.
* dwarf2/read.c (delete_file_name_entry): Fix comment.
(create_cu_from_index_list)
(create_signatured_type_table_from_index)
(create_signatured_type_table_from_debug_names)
(dw2_get_file_names_reader, dwarf2_initialize_objfile)
(dwarf2_create_include_psymtab)
(create_debug_type_hash_table, add_type_unit)
(create_type_unit_group, read_comp_units_from_section)
(dwarf2_compute_name, create_cus_hash_table)
(create_dwp_hash_table, create_dwo_unit_in_dwp_v1)
(create_dwo_unit_in_dwp_v2, open_and_init_dwp_file): Use new
obstack.
(dw2_get_real_path): Likewise. Change argument to
dwarf2_per_objfile.
This fixes an instruction mask typo. We should be matching only
ldrd (immediate) and not any other of its variants. As is, it never matches
anything.
With the patch, the instruction mask also allows matching of ldrd (literal),
but the check for SP discards this particular instruction pattern, as it has
a hardcoded PC register.
Luis Machado [Mon, 25 May 2020 16:25:50 +0000 (13:25 -0300)]
Fix some duplicate test names
While doing a testsuite run on aarch64-linux, I noticed a bunch of duplicated
test name results. It annoyed me a little, so I decided to go ahead and fix the
worst offenders.
The following patch brings the duplicate test names down from 461 to 137.
The remaining ones are mostly scattered across the testsuite, with 1 to 3
duplicates per testcase. We can fix those as we go.
projects / thirdparty / binutils-gdb.git / log
