This patch fixes PR 18605 which is about incorrectly decoding media
instructions in software single step.
gdb:
2015-06-30 Yao Qi <yao.qi@linaro.org>
PR tdep/18605
* arm-tdep.c (arm_get_next_pc_raw): Break for media
instructions.
This change turns on dwarf2 unwinding in rx-tdep.c. I found it
necessary to add rx_dwarf_reg_to_regnum in order to cause PC to be
mapped correctly.
gdb/ChangeLog:
* rx-tdep.c (RX_PSW_REGNUM): New enum constant.
(rx_dwarf_reg_to_regnum): New function.
(rx_gdbarch_init): Register rx_dwarf_reg_to_regnum. Use dwarf2
unwinding.
The crash reason is that the insn to be moved to stub may be a
relocation spot, so instead of placing the origin insn (that is insn
before-relocation) to the stub, I have to place the relocated one.
Note the relocation involved is non-pc-relative, so it is safe to move
the relocated insn.
gold/ChangeLog:
2015-06-29 Han Shen <shenhan@google.com>
* AArch64.cc (Erratum_stub::Insn_utilities): New typedef.
(Erratum_stub::update_erratum_insn): New method.
(Stub_table::relocate_stubs): Modified to place relocated insn.
(AArch64_relobj::fix_errata): Modified gold_assert.
Refs:
https://sourceware.org/ml/gdb/2015-03/msg00024.htmlhttps://sourceware.org/ml/gdb/2015-06/msg00005.html
On GNU/Linux, if an infcall spawns a thread, that thread ends up with
stuck running state. This happens because:
- when linux-nat.c detects a new thread, it marks them as running,
and does not report anything to the core.
- we skip finish_thread_state when the thread that is running the
infcall stops.
As result, that new thread ends up with stuck "running" state, even
though it really is stopped.
On Windows, _all_ threads end up stuck in running state, not just the
one that was spawned. That happens because when a new thread is
detected, unlike linux-nat.c, windows-nat.c reports
TARGET_WAITKIND_SPURIOUS to infrun. It's the fact that that event
does not cause a user-visible stop that triggers the problem. When
the target is re-resumed, we call set_running with a wildcard ptid,
which marks all thread as running. That set_running is not suppressed
because the (leader) thread being resumed does not have in_infcall
set. Later, when the infcall finally finishes successfully, nothing
marks all threads back to stopped.
We can trigger the same problem on all targets by having a thread
other than the one that is running the infcall report a breakpoint hit
to infrun, and then have that breakpoint not cause a stop. That's
what the included test does.
The fix is to stop GDB from suppressing the set_running calls while
doing an infcall, and then set the threads back to stopped when the
call finishes, iff they were originally stopped before the infcall
started. (Note the MI *running/*stopped event suppression isn't
affected.)
Tested on x86_64 GNU/Linux.
gdb/ChangeLog:
2015-06-29 Pedro Alves <palves@redhat.com>
PR threads/18127
* infcall.c (run_inferior_call): On infcall success, if the thread
was marked stopped before, reset it back to stopped.
* infrun.c (resume): Don't suppress the set_running calls when
doing an infcall.
(normal_stop): Only discard the finish_thread_state cleanup if the
infcall succeeded.
gdb/testsuite/ChangeLog:
2015-06-29 Pedro Alves <palves@redhat.com>
PR threads/18127
* gdb.threads/hand-call-new-thread.c: New file.
* gdb.threads/hand-call-new-thread.c: New file.
This patch lets GDBServer handle software breakpoints instead of relying
on GDB.
gdb/gdbserver/ChangeLog:
* linux-aarch64-low.c (aarch64_supports_z_point_type): Enable for
Z_PACKET_SW_BP.
GDB uses a "brk #0" instruction to perform a software breakpoint while
GDBServer uses an illegal instruction. Both instructions should match.
When enabling support for the 'Z0' packet, we let GDBServer insert the
breakpoint instruction instead of GDB. And in case of permanent
breakpoints for example, GDB will check if a breakpoint is inserted in the
inferior with `program_breakpoint_here_p (gdbarch, address)', and
compare the instruction read from the inferior with the breakpoint
instruction.
On AArch64, instructions are always little endian so we need to
represent it as an array of bytes, as done in aarch64-tdep.c.
gdb/gdbserver/ChangeLog:
* linux-aarch64-low.c: Remove comment about endianness.
(aarch64_breakpoint): Change type to gdb_byte[]. Set to "brk #0".
(aarch64_breakpoint_at): Change type of insn to gdb_byte[]. Use
memcmp.
* Makefile.in (SIM_EXTRA_CFLAGS): Add -lm.
* armdefs.h (ARMdval, ARMfval): New types.
(ARM_VFP_reg): New union.
(struct ARMul_State): Add VFP_Reg and FPSCR fields.
(VFP_fval, VFP_uword, VFP_sword, VFP_dval, VFP_dword): Accessor
macros for the new VFP_Reg field.
* armemu.c (handle_v6_insn): Add code to handle MOVW, MOVT,
QADD16, QASX, QSAX, QSUB16, QADD8, QSUB8, UADD16, USUB16, UADD8,
USUB8, SEL, REV, REV16, RBIT, BFC, BFI, SBFX and UBFX
instructions.
(handle_VFP_move): New function.
(ARMul_Emulate16): Add checks for newly supported v6
instructions. Add support for VMRS, VMOV and MRC instructions.
(Multiply64): Allow nRdHi == nRm and/or nRdLo == nRm when
operating in v6 mode.
* armemu.h (t_resolved): Define.
* armsupp.c: Include math.h.
(handle_VFP_xfer): New function. Handles VMOV, VSTM, VSTR, VPUSH,
VSTM, VLDM and VPOP instructions.
(ARMul_LDC): Test for co-processor 10 or 11 and pass call to the
new handle_VFP_xfer function.
(ARMul_STC): Likewise.
(handle_VFP_op): New function. Handles VMLA, VMLS, VNMLA, VNMLS,
VNMUL, VMUL, VADD, VSUB, VDIV, VMOV, VABS, VNEG, VSQRT, VCMP,
VCMPE and VCVT instructions.
(ARMul_CDP): Test for co-processor 10 or 11 and pass call to the
new handle_VFP_op function.
* thumbemu.c (tBIT, tBITS, ntBIT, ntBITS): New macros.
(test_cond): New function. Tests a condition and returns non-zero
if the condition has been met.
(handle_IT_block): New function.
(in_IT_block): New function.
(IT_block_allow): New function.
(ThumbExpandImm): New function.
(handle_T2_insn): New function. Handles T2 thumb instructions.
(handle_v6_thumb_insn): Add next_instr and pc parameters.
(ARMul_ThumbDecode): Add support for IT blocks. Add support for
v6 instructions.
* wrapper.c (sim_create_inferior): Detect a thumb address and call
SETT appropriately.
Last year a patch was submitted/approved/commited to eliminate
symbol_matches_domain which was causing this problem. It was later reverted
because it introduced a (severe) performance regression.
Recap:
(gdb) list
1 enum e {A,B,C} e;
2 int main (void) { return 0; }
3
(gdb) p e
Attempt to use a type name as an expression
The parser attempts to find a symbol named "e" of VAR_DOMAIN.
This gets passed down through lookup_symbol and (eventually) into
block_lookup_symbol_primary, which iterates over the block's dictionary
of symbols:
for (sym = dict_iter_name_first (block->dict, name, &dict_iter);
sym != NULL;
sym = dict_iter_name_next (name, &dict_iter))
{
if (symbol_matches_domain (SYMBOL_LANGUAGE (sym),
SYMBOL_DOMAIN (sym), domain))
return sym;
}
The problem here is that we have a symbol named "e" in both STRUCT_DOMAIN
and VAR_DOMAIN, and for languages like C++, Java, and Ada, where a tag name
may be used as an implicit typedef of the type, symbol_matches_domain ignores
the difference between VAR_DOMAIN and STRUCT_DOMAIN. As it happens, the
STRUCT_DOMAIN symbol is found first, considered a match, and that symbol is
returned to the parser, eliciting the (now dreaded) error message.
Since this bug exists specifically because we have both STRUCT and VAR_DOMAIN
symbols in a given block/CU, this patch rather simply/naively changes
block_lookup_symbol_primary so that it continues to search for an exact
domain match on the symbol if symbol_matches_domain returns a symbol
which does not exactly match the requested domain.
This "fixes" the immediate problem, but admittedly might uncover other,
related bugs. [Paranoia?] However, it causes no regressions (functional
or performance) in the test suite. A similar change has been made
to block_lookup_symbol for other cases in which this bug might appear.
The tests from the previous submission have been resurrected and updated.
However since we can still be given a matching symbol with a different domain
than requested, we cannot say that a symbol "was not found." The error
messages today will still be the (dreaded) "Attempt to use a type name..."
ChangeLog
PR 16253
* block.c (block_lookup_symbol): For non-function blocks,
continue to search for a symbol with an exact domain match
Otherwise, return any previously found "best domain" symbol.
(block_lookup_symbol_primary): Likewise.
testsuite/ChangeLog
PR 16253
* gdb.cp/var-tag-2.cc: New file.
* gdb.cp/var-tag-3.cc: New file.
* gdb.cp/var-tag-4.cc: New file.
* gdb.cp/var-tag.cc: New file.
* gdb.cp/var-tag.exp: New file.
This patch implements the new option "history remove-duplicates", which
controls the removal of duplicate history entries ("off" by default).
The motivation for this option is to be able to reduce the prevalence of
basic commands such as "up" and "down" in the history file. These
common commands crowd out more unique commands in the history file (when
the history file has a fixed size), and they make navigation of the
history file via ^P, ^N and ^R more inconvenient.
The option takes an integer denoting the number of history entries to
look back at for a history entry that is a duplicate of the latest one.
"history remove-duplicates 1" is equivalent to bash's ignoredups option,
and "history remove-duplicates unlimited" is equivalent to bash's
erasedups option.
[ I decided to go with this integer approach instead of a tri-state enum
because it's slightly more flexible and seemingly more intuitive than
leave/erase/ignore. ]
gdb/ChangeLog:
* NEWS: Mention the new option "history remove-duplicates".
* top.c (history_remove_duplicates): New static variable.
(show_history_remove_duplicates): New static function.
(gdb_add_history): Conditionally remove duplicate history
entries.
(init_main): Add "history remove-duplicates" option.
gdb/doc/ChangeLog:
* gdb.texinfo (Command History): Document the new option
"history remove-duplicates".
gdb/testsuite/ChangeLog:
* gdb.base/history-duplicates.exp: New test.
The implementation is pretty straightforward, with the only caveat being
that the "src", "cmd", "next" and "prev" entries get delibrately added
to the completion list even when the TUI has not yet been initialized
(i.e. has never been enabled during the session), since invoking the
"focus" command with these arguments already works when the TUI has not
yet been initialized.
gdb/ChangeLog:
* tui/tui-win.c (focus_completer): New static function.
(_initialize_tui_win): Set the completion function of the
"focus" command to focus_completer.
gdb/testsuite/ChangeLog:
* gdb.base/completion.exp: Test the completion of the "focus"
command.
GDB tries to skip prologue for .S files according to .debug_line but it then
places the breakpoint to a location where it is never hit.
This is because #defines in .S files cause prologue skipping which is
completely inappropriate, for s390x:
glibc/sysdeps/unix/syscall-template.S
78:/* This is a "normal" system call stub: if there is an error,
79: it returns -1 and sets errno. */
80:
81:T_PSEUDO (SYSCALL_SYMBOL, SYSCALL_NAME, SYSCALL_NARGS)
82: ret
00000000000f4210 T __select
Line Number Statements:
Extended opcode 2: set Address to 0xf41c8
Advance Line by 80 to 81
Copy
Advance PC by 102 to 0xf422e
Special opcode 6: advance Address by 0 to 0xf422e and Line by 1 to 82
Special opcode 34: advance Address by 2 to 0xf4230 and Line by 1 to 83
Advance PC by 38 to 0xf4256
Extended opcode 1: End of Sequence
Compilation Unit @ offset 0x28b3e0:
<0><28b3eb>: Abbrev Number: 1 (DW_TAG_compile_unit)
<28b3ec> DW_AT_stmt_list : 0x7b439
<28b3f0> DW_AT_low_pc : 0xf41c8
<28b3f8> DW_AT_high_pc : 0xf4256
<28b400> DW_AT_name : ../sysdeps/unix/syscall-template.S
<28b423> DW_AT_comp_dir : /usr/src/debug////////glibc-2.17-c758a686/misc
<28b452> DW_AT_producer : GNU AS 2.23.52.0.1
<28b465> DW_AT_language : 32769 (MIPS assembler)
without debuginfo or with debuginfo and the fix - correct address:
(gdb) b select
Breakpoint 1 at 0xf4210
It is also where .dynsym+.symtab point to:
00000000000f4210 T __select
00000000000f4210 W select
with debuginfo, without the fix:
(gdb) b select
Breakpoint 1 at 0xf41c8: file ../sysdeps/unix/syscall-template.S, line 81.
One part is to behave for asm files similar way like for 'locations_valid':
/* Symtab has been compiled with both optimizations and debug info so that
GDB may stop skipping prologues as variables locations are valid already
at function entry points. */
unsigned int locations_valid : 1;
The other part is to extend the 'locations_valid'-like functionality more.
Both minsym_found and find_function_start_sal need to be patched, otherwise
their addresses do not match and GDB regresses on ppc64:
gdb/ChangeLog
2015-06-26 Jan Kratochvil <jan.kratochvil@redhat.com>
* linespec.c (minsym_found): Reset sal.PC for COMPUNIT_LOCATIONS_VALID
and language_asm..
* symtab.c (find_function_start_sal): Likewise.
gdb/testsuite/ChangeLog
2015-06-26 Jan Kratochvil <jan.kratochvil@redhat.com>
* gdb.arch/amd64-prologue-skip.S: New file.
* gdb.arch/amd64-prologue-skip.exp: New file.
_bfd_elf_assign_file_positions_for_non_load will set the sh_size
field for .shstrtab section. There is need to initialize it in
_bfd_elf_compute_section_file_positions.
* elf.c (_bfd_elf_compute_section_file_positions): Don't
initialize sh_size for .shstrtab here.
Since not all ELF targets support -shared, we limit ld-elf/strtab to
linux and gnu targets.
* ld-elf/strtab.d: Only run on *-*-linux* and *-*-gnu* targets.
This patch uses ELF strtab with GC and suffix merging support to create
ELF .strtab section. There is some small memory overhead to use ELF
strtab:
==14928== HEAP SUMMARY:
==14928== in use at exit: 3,276,318 bytes in 679 blocks
==14928== total heap usage: 1,544 allocs, 865 frees, 15,259,146 bytes allocated
vs.
==14936== HEAP SUMMARY:
==14936== in use at exit: 3,276,318 bytes in 679 blocks
==14936== total heap usage: 1,532 allocs, 853 frees, 15,026,402 bytes allocated
when running:
./ld-new -m elf_x86_64 -o tmpdir/ld-partial.o -r ldgram.o ldlex-wrapper.o lexsup.o ldlang.o mri.o ldctor.o ldmain.o plugin.o ldwrite.o ldexp.o ldemul.o ldver.o ldmisc.o ldfile.o ldcref.o eelf_x86_64.o eelf32_x86_64.o eelf_i386.o eelf_iamcu.o ei386linux.o eelf_l1om.o eelf_k1om.o ldbuildid.o
The results are
[32] .strtab STRTAB 0+ 3beff8 00407a 00 0 0 1
vs
[32] .strtab STRTAB 0+ 3beff8 0041d8 00 0 0 1
It reduces the .strtab size by 350 bytes, about 2%
Saving on libc.so from glibc is much more since libc.so has many alias
symbols with the same suffix. For x32 glibc,
[82] .strtab STRTAB 0+ 81b348 0159e7 00 0 0 1
vs
[82] .strtab STRTAB 0+ 81b8bc 019e72 00 0 0 1
It reduces the .strtab size by 17547 bytes, about 16%.
bfd/
PR gas/18451
* elf-bfd.h (elf_sym_strtab): New.
(elf_link_hash_table): Add strtabcount, strtabsize and
strtab.
(_bfd_elf_stringtab_init): Removed.
* elf.c (_bfd_elf_stringtab_init): Removed.
(_bfd_elf_compute_section_file_positions): Replace
bfd_strtab_hash/_bfd_elf_stringtab_init/_bfd_stringtab_free/
_bfd_stringtab_size with
elf_strtab_hash/_bfd_elf_strtab_init/_bfd_elf_strtab_free/
_bfd_elf_strtab_size. Use _bfd_elf_strtab_add,
_bfd_elf_strtab_finalize and _bfd_elf_strtab_offset to get
st_name.
(swap_out_syms): Likewise.
* elflink.c (elf_final_link_info): Replace bfd_strtab_hash
with elf_strtab_hash. Remove symbuf, symbuf_count,
symbuf_size and shndxbuf_size.
(elf_link_flush_output_syms): Removed.
(elf_link_output_sym): Renamed to ...
(elf_link_output_symstrtab): This. Replace _bfd_stringtab_add
with _bfd_elf_strtab_add. Don't flush symbols to the file nor
swap out symbols.
(elf_link_swap_symbols_out): New.
(elf_link_output_extsym): Replace elf_link_output_sym with
elf_link_output_symstrtab.
(elf_link_input_bfd): Likewise.
(elf_final_link_free): Replace _bfd_stringtab_free with
_bfd_elf_strtab_free. Remove symbuf.
(bfd_elf_final_link): Replace _bfd_elf_stringtab_init with
_bfd_elf_strtab_init. Don't set symbuf, symbuf_count,
symbuf_size nor shndxbuf_size. Initialize strtabsize and
strtab. Initialize symshndxbuf to -1 when number of sections
>= 64K. Replace elf_link_output_sym/elf_link_output_sym with
elf_link_output_symstrtab/elf_link_output_symstrtab. Don't
call elf_link_flush_output_syms. Call _bfd_elf_strtab_finalize
and elf_link_swap_symbols_out. Replace _bfd_stringtab_size
and _bfd_stringtab_emit with _bfd_elf_strtab_size and
_bfd_elf_strtab_emit.
gas/testsuite/
PR gas/18451
* gas/elf/elf.exp: Run strtab.
* gas/elf/strtab.d: New file.
* gas/elf/strtab.s: Likewise.
ld/testsuite/
PR gas/18451
* ld-elf/strtab.d: New file.
* ld-elf/strtab.s: Likewise.
Some parts of solib_find_1 should only operate if the sysroot
is nonempty after processing, but the logic that checked this
happened before trailing slashes were stripped so empty but
non-NULL sysroots were possible. This commit moves the logic
so it correctly notices all empty sysroots.
gdb/ChangeLog:
* solib.c (solib_find_1): Set local variable sysroot to NULL if
it is the empty string after trailing slashes have been stripped.
Commit 2f0c68f23b added an extra @section
.cfi_lsda:
@subsection @code{.cfi_lsda @var{encoding} [, @var{exp}]}
+@section @code{.cfi_lsda @var{encoding} [, @var{exp}]}
@code{.cfi_lsda} defines LSDA and its encoding.
It shouldn't be there.
* doc/as.texinfo (.cfi_lsda): Remove the extra @section.
Valgrind reports memory leaking from build_id_to_debug_bfd:
==7261== 88 bytes in 2 blocks are definitely lost in loss record 31,319 of 35,132
==7261== at 0x4A06BCF: malloc (vg_replace_malloc.c:296)
==7261== by 0x32CA88A9B9: strdup (strdup.c:42)
==7261== by 0xFE62AB: lrealpath (lrealpath.c:88)
==7261== by 0x7F7AD6: build_id_to_debug_bfd (build-id.c:116)
==7261== by 0x7F7BB5: find_separate_debug_file_by_buildid (build-id.c:149)
==7261== by 0x6D9382: elf_symfile_read (elfread.c:1348)
==7261== by 0x777F02: read_symbols (symfile.c:875)
==7261== by 0x778505: syms_from_objfile_1 (symfile.c:1078)
==7261== by 0x778548: syms_from_objfile (symfile.c:1094)
==7261== by 0x778746: symbol_file_add_with_addrs (symfile.c:1191)
==7261== by 0x77893B: symbol_file_add_from_bfd (symfile.c:1280)
==7261== by 0x8E51E3: solib_read_symbols (solib.c:706)
==7261== by 0x8E58AF: solib_add (solib.c:1029)
This occurs because commit 1be5090b in bfd, addressing PR 11983, started
taking a copy of the input filename instead of directly caching it. It
appears that this code was never updated to reflect that API change.
This simple patch creates a cleanup to free the return value for lrealpath.
gdb/ChangeLog
* build-id.c (build_id_to_debug_bfd): Add cleanup to free
return value from lrealpath.
Since the .plt section and DT_PLTGOT are used by prelink to undo
prelinking for dynamic relocations, we must keep them even if there is
no PLT relocation. This patch reverted commit a3747075a.
bfd/
* elf32-i386.c (elf_i386_allocate_dynrelocs): Always allocate
space for the first .plt entry.
(elf_i386_size_dynamic_sections): Always add DT_PLTGOT for .plt
section. Add DT_PLTRELSZ, DT_PLTREL and DT_JMPREL only if
there are PLT relocations.
* elf64-x86-64.c (elf_x86_64_allocate_dynrelocs): Always
allocate space for the first .plt entry.
(elf_x86_64_size_dynamic_sections): Always add DT_PLTGOT for
.plt section. Add DT_PLTRELSZ, DT_PLTREL and DT_JMPREL only if
there are PLT relocations.
ld/testsuite/
* ld-i386/i386.exp: Run pltgot-2 for Linux targets.
* ld-x86-64/x86-64.exp: Likewise.
* ld-i386/pltgot-1.d: Updated.
* ld-x86-64/pltgot-1.d: Likewise.
* ld-i386/pltgot-2.d: New file.
* ld-x86-64/pltgot-2.d: Likewise.
When converting "mov foo@GOTPCREL(%rip), %reg" to "lea foo(%rip), %reg"
with R_X86_64_PC32 relocation, it may overflow if the target section
is more than 2GB away. This patch estimates distance between mov
instruction and the target section. We convert R_X86_64_GOTPCREL to
R_X86_64_PC32 only if their distance is less than 2GB.
PR ld/18591
* elf64-x86-64.c (elf_x86_64_convert_mov_to_lea): Don't convert
R_X86_64_GOTPCREL to R_X86_64_PC32 if it will cause relocation
overflow.
The bfin/msp430 ports already had trace logic set up for reading/writing
cpu registers, albeit using different unrelated levels (core & vpu). Add
a proper register class for these and for other ports.
The default gdb sysroot now sets itself to "target:". This works for
most remote targets, but when using the simulator, this causes problems
as the sim will attempt to search for that path.
Update the remote-sim logic to skip this leading prefix when it is found
so that the sysroot isn't passed in as an invalid value.
