My original implementation for LDI32 pseudo does not conform to
the TI ABI. I wrongly documented my TI PRU ELF object files inspection,
which got propagated into my binutils implementation.
Issue was exposed when running the GCC ABI testsuite against TI toolchain.
According to TI ABI, LDI32 must use first LDI instruction to load
the MSB 16bits, and second LDI instruction for the LSB 16bits.
This patch will break binary compatibility with previously released
binutils versions for PRU. Still, I think it is better to fix
binutils to conform to the chip vendor ABI.
bfd * elf32-pru.c (pru_elf32_do_ldi32_relocate): Make LDI32 relocation
conformant to TI ABI.
(pru_elf32_relax_section): Likewise.
(pru_elf_relax_delete_bytes): Fix offsets for new LDI32 code.
* elf32-pru.c (pru_elf32_do_ldi32_relocate): Ignore addend.
(pru_elf32_pmem_relocate): Trap - should not get here.
(pru_elf32_relocate_section): Add support for REL relocations.
(elf_info_to_howto_rel): Enable REL.
(elf_backend_may_use_rel_p): Likewise.
(elf_backend_may_use_rela_p): Likewise.
(elf_backend_default_use_rela_p): Likewise.
gas * config/tc-pru.c (md_apply_fix): Make LDI32 relocation conformant
to TI ABI.
(pru_assemble_arg_i): Likewise.
(output_insn_ldi32): Likewise.
* testsuite/gas/pru/ldi.d: Update test for the now fixed LDI32.
* gas/config/tc-pru.c (pru_assemble_arg_b): Check imm8 operand range.
* gas/testsuite/gas/pru/illegal2.l: New test.
* gas/testsuite/gas/pru/illegal2.s: New test.
* gas/testsuite/gas/pru/pru.exp: Register new illegal2 test.
ld * scripttempl/pru.sc: Add LD sections to allow linking TI
toolchain object files.
* scripttempl/pru.sc: Switch to init_array.
* testsuite/ld-pru/ldi32.d: Update LDI32 test to conform to TI ABI.
* testsuite/ld-pru/norelax_ldi32-data.d: Likewise.
* testsuite/ld-pru/norelax_ldi32-dis.d: Likewise.
* testsuite/ld-pru/relax_ldi32-data.d: Likewise.
* testsuite/ld-pru/relax_ldi32-dis.d: Likewise.
When building with x86_64-w64-mingw32-g++ (to test cross-compiling for
Windows), I get this error:
unittests/string_view-selftests.o: In function `selftests::string_view::inserters_2::test05(unsigned long long)':
/home/emaisin/src/binutils-gdb/gdb/unittests/basic_string_view/inserters/char/2.cc:60: undefined reference to `std::basic_ofstream<char, std::char_traits<char> >::rpl_close()'
This is caused by gnulib redefining "close" as "rpl_close", and
therefore messing up the declaration of basic_ofstream in the libstdc++
header. The solution would be to use gnulib namespaces [1]. Until we
use them across GDB, we can use them locally in files that are
problematic, like this one.
gdb/ChangeLog:
* unittests/string_view-selftests.c: Define GNULIB_NAMESPACE.
For GNU/Linux on x86-64, if the target is using the xsave format for
passing the floating-point information from the inferior then there
currently exists a bug relating to the x87 control registers, and the
mxcsr register.
The xsave format allows different floating-point features to be lazily
enabled, a bit in the xsave format tells GDB which floating-point
features have been enabled, and which have not.
Currently in GDB, when reading the floating point state, we check the
xsave bit flags, if the feature is enabled then we read the feature
from the xsave buffer, and if the feature is not enabled, then we
supply the default value from within GDB.
Within GDB, when writing the floating point state, we first fetch the
xsave state from the target and then, for any feature that is not yet
enabled, we write the default values into the xsave buffer. Next we
compare the regcache value with the value in the xsave buffer, and, if
the value has changed we update the value in the xsave buffer, and
mark the feature enabled in the xsave bit flags.
The problem then, is that the x87 control registers were not following
this pattern. We assumed that these registers were always written out
by the kernel, and we always wrote them out to the xsave buffer (but
didn't enabled the feature). The result of this is that if the kernel
had not yet enabled the x87 feature then within GDB we would see
random values for the x87 floating point control registers, and if the
user tried to modify one of these register, that modification would be
lost.
Finally, the mxcsr register was also broken in the same way as the x87
control registers. The added complexity with this case is that the
mxcsr register is part of both the avx and sse floating point feature
set. When reading or writing this register we need to check that at
least one of these features is enabled.
This bug was present in native GDB, and within gdbserver. Both are
fixed with this commit.
gdb/ChangeLog:
* common/x86-xstate.h (I387_FCTRL_INIT_VAL): New constant.
(I387_MXCSR_INIT_VAL): New constant.
* amd64-tdep.c (amd64_supply_xsave): Only read state from xsave
buffer if it was supplied by the inferior.
* i387-tdep.c (i387_supply_fsave): Use I387_MXCSR_INIT_VAL.
(i387_xsave_get_clear_bv): New function.
(i387_supply_xsave): Only read x87 control registers from the
xsave buffer if the feature is enabled, and the state will have
been written, otherwise, provide a suitable default.
(i387_collect_xsave): Pre-clear all registers in xsave buffer,
including x87 control registers. Update control registers if they
have changed from the default value, and mark features as enabled
as required.
* i387-tdep.h (i387_xsave_get_clear_bv): Declare.
gdb/gdbserver/ChangeLog:
* i387-fp.c (i387_cache_to_xsave): Only write x87 control
registers to the cache if their values have changed.
(i387_xsave_to_cache): Provide default values for x87 control
registers when these features are available, but disabled.
* regcache.c (supply_register_by_name_zeroed): New function.
* regcache.h (supply_register_by_name_zeroed): Declare new
function.
gdb/testsuite/ChangeLog:
* gdb.arch/amd64-init-x87-values.S: New file.
* gdb.arch/amd64-init-x87-values.exp: New file.
Another fuzzing fix. I think it's reasonable to simply strip out any
group section that is too weird for objcopy to handle.
PR 23141
* objcopy.c (is_strip_section): Strip groups without a valid
signature symbol.
Defining SPE2_OPCD_SEGS as 13 discounts the possibility that we'd
ever look up spe2_opcd_indices[14..16], which I think is possible.
Extend that array to size 16+1, using the macros we use to index the
array. Similarly use the index macros for PPC_OPCD_SEGS and
VLE_OPCD_SEGS.
* ppc-dis.c (PPC_OPCD_SEGS): Define using PPC_OP.
(VLE_OPCD_SEGS, SPE2_OPCD_SEGS): Similarly, using macros used to
partition opcode space for index lookup.
The formatting of the output of the "info spu event" command changed, causing
spurious test suite failures. Use phex instead of phex_nz to get back the
expected format, and fix emission of new line characters.
gdb/ChangeLog:
2018-05-08 Ulrich Weigand <uweigand@de.ibm.com>
* spu-tdep.c (info_spu_event_command): Fix output formatting.
PR 22809
* elf.c (bfd_elf_get_str_section): Check for an excessively large
string section.
* elf-attrs.c (_bfd_elf_parse_attributes): Issue an error if the
attribute section is larger than the size of the file.
This patch replaces AddrPrefixOp0 with AddrPrefixOpReg to indicate that
the size of register operand is controlled by the address size prefix.
This will be used by Intel MOVDIRI and MOVDIR64B instructions later.
gas/
* config/tc-i386.c (process_suffix): Check addrprefixopreg
instead of addrprefixop0.
opcodes/
* i386-gen.c (opcode_modifiers): Replace AddrPrefixOp0 with
AddrPrefixOpReg.
* i386-opc.h (AddrPrefixOp0): Renamed to ...
(AddrPrefixOpReg): This.
(i386_opcode_modifier): Rename addrprefixop0 to addrprefixopreg.
* i386-opc.tbl: Replace AddrPrefixOp0 with AddrPrefixOpReg.
This adds -Wduplicated-cond to warnings.m4. This caught one bug.
I tried adding -Wduplicated-branches as well, but it results in some
spurious failures from code like this in cgen.h:
#define CGEN_ATTR_TYPE(n) \
struct { unsigned int bool_; \
CGEN_ATTR_VALUE_TYPE nonbool[(n) ? (n) : 1]; }
This will trigger a warning if passed n==1, which seems like a
perfectly valid thing to do; and there were other issues like this as
well.
ChangeLog
2018-05-07 Tom Tromey <tom@tromey.com>
* configure: Rebuild.
* warning.m4 (AM_GDB_WARNINGS): Add -Wduplicated-cond.
gdbserver/ChangeLog
2018-05-07 Tom Tromey <tom@tromey.com>
* configure: Rebuild.
-Wduplicated-cond pointed out that arm_record_vfp_data_proc_insn
checks "opc1 == 0x0b" twice. I filed this a while ago as
PR tdep/20362.
Based on the ARM instruction manual at
https://www.scss.tcd.ie/~waldroj/3d1/arm_arm.pdf, I think the
instruction decoding in this function has two bugs.
First, opc1 is computed as:
opc1 = bits (arm_insn_r->arm_insn, 20, 23);
[...]
opc1 = opc1 & 0x04;
This means that tests like:
else if (opc1 == 0x01)
can never be true.
In the ARM manual, "opc1" corresponds to these bits:
name bit
r 20
q 21
D 22
p 23
... where the D bit is not used for VFP instruction decoding.
So, I believe this code should use ~0x04 instead.
Second, VDIV is recognized by the bits "pqrs" being equal to "1000".
This tranlates to opc1 == 0x08 -- not 0x0b. Note that pqrs==1001 is
an undefined encoding, which is probably why opc2 is not checked here;
this code doesn't seem to really deal with undefined encodings in
general, so I've left that as is.
I don't have an ARM machine or any reasonable way to test this.
ChangeLog
2018-05-07 Tom Tromey <tom@tromey.com>
PR tdep/20362:
* arm-tdep.c (arm_record_vfp_data_proc_insn): Properly mask off D
bit. Use correct value for VDIV.
Avoids gcc pr85623 for these calls.
* cofflink.c (_bfd_coff_link_input_bfd): Use memcmp rather than
strncmp when checking for ".bf" special symbol.
* prXXigen.c (_bfd_XXi_swap_scnhdr_out): Make pe_required_section_flags
section name a char array, remove sentinal known_sections entry,
and adjust loop over known_sections to suit. Use memcmp rather
than strncmp.
This adds -Wimplicit-fallthrough to the set of default warnings.
2018-05-04 Tom Tromey <tom@tromey.com>
* configure: Rebuild.
* warning.m4 (AM_GDB_WARNINGS): Add -Wimplicit-fallthrough.
gdbserver/ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
* configure: Rebuild.
This adds a "break" at the end of the RECORD_SYS_RECVFROM case in
record_linux_system_call. This seemed correct to me.
2018-05-04 Tom Tromey <tom@tromey.com>
* linux-record.c (record_linux_system_call) <case
RECORD_SYS_RECVFROM>: Add "break".
This adds a "break" to a couple of spots where it was erroneously
omitted. I think these are the two (potential) real bugs caught by
this series.
2018-05-04 Tom Tromey <tom@tromey.com>
* mi/mi-main.c (mi_cmd_trace_frame_collected) <REGISTERS_FORMAT>:
Add missing "break".
* mi/mi-cmd-stack.c (mi_cmd_stack_list_locals) <NO_FRAME_FILTERS>:
Add missing "break".
This adds two fall-through comments in rs6000-tdep.c. I looked at the
PPC instruction manual and convinced myself that this was correct.
And, this isn't a semantic change. However, close review would still
be good.
2018-05-04 Tom Tromey <tom@tromey.com>
* rs6000-tdep.c (ppc_process_record_op4)
(ppc_process_record_op63): Add fall-through comment.
This adds a fall-through comment in i386-tdep.c. I was not sure what
to do here, so I elected to preserve the status quo. In review, John
Baldwin pointed out that: "I believe this is correct based on the diff
that added the special cases for xgetbv and xsetbv as previously ldgt
and lidt were treated the same".
gdb/ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
* i386-tdep.c (i386_process_record): Add fall-through comment.
This adds a fall-through comment to stabsread.c. I skimmed the stabs
manual a bit and it seems that 'p' and 'P' are similar enough that
this makes sense. Also, stabs is mostly deprecated, and the code has
been this way for a long time, so it seemed safest to keep the status
quo.
ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
* stabsread.c (define_symbol) <case 'p'>: Add fall-through
comment.
This patch fixes the subset of -Wimplicit-fallthrough warnings that I
considered obvious. In most cases it was obvious from context that
falling through was desired; here I added the appropriate comment. In
a couple of cases it seemed clear that a "break" was missing.
ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
* riscv-tdep.c (riscv_isa_xlen): Add fall-through comment.
* utils.c (can_dump_core) <LIMIT_CUR>: Add fall-through comment.
* eval.c (fetch_subexp_value) <MEMORY_ERROR>: Add fall-through
comment.
* d-valprint.c (d_val_print) <TYPE_CODE_STRUCT>: Add fall-through
comment.
* coffread.c (coff_symtab_read) <C_LABEL>: Add fall-through
comment.
This patch adds a missing ATTRIBUTE_NORETURN. This lets
-Wimplicit-fallthrough recognize that a given case does not fall
through.
ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
* dwarf2loc.c (unimplemented): Add ATTRIBUTE_NORETURN.
This changes the Python API so that breakpoint commands can be set by
writing to the "commands" attribute.
ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
PR python/22731:
* NEWS: Mention that breakpoint commands are writable.
* python/py-breakpoint.c (bppy_set_commands): New function.
(breakpoint_object_getset) <"commands">: Use it.
doc/ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
PR python/22731:
* python.texi (Breakpoints In Python): Mention that "commands" is
writable.
testsuite/ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
PR python/22731:
* gdb.python/py-breakpoint.exp: Test setting breakpoint commands.
This changes some functions in cli-script.c to use function_view
rather than a function pointer and closure argument. This simplifies
the code a bit and is useful in a subsequent patch.
ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
* tracepoint.c (actions_command): Update.
* mi/mi-cmd-break.c (mi_command_line_array)
(mi_command_line_array_cnt, mi_command_line_array_ptr)
(mi_read_next_line): Remove.
(mi_cmd_break_commands): Update.
* cli/cli-script.h (read_command_lines, read_command_lines_1): Use
function_view.
* cli/cli-script.c (get_command_line): Update.
(process_next_line): Use function_view. Constify.
(recurse_read_control_structure, read_command_lines)
(read_command_lines_1): Change argument types to function_view.
(do_define_command, document_command): Update.
* breakpoint.h (check_tracepoint_command): Don't declare.
* breakpoint.c (check_tracepoint_command): Remove.
(commands_command_1, create_tracepoint_from_upload): Update.
PR gdb/11750 concerns defining a command inside a user commnad, like:
define outer
define inner
echo hi\n
end
end
This patch adds this capability to gdb.
ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
PR gdb/11750:
* cli/cli-script.h (enum command_control_type) <define_control>:
New constant.
* cli/cli-script.c (multi_line_command_p): Handle define_control.
(build_command_line, execute_control_command_1)
(process_next_line): Likewise.
(do_define_command): New function, extracted from define_command.
(define_command): Use it.
testsuite/ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
PR gdb/11750:
* gdb.base/define.exp: Test defining a user command inside a user
command.
* gdb.base/commands.exp (define_if_without_arg_test): Test "define".
The prompt argument to read_command_lines can be const. This patch
makes this change, and also removes some fixed-sized buffers in favor
of using string_printf.
ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
* tracepoint.c (actions_command): Update.
* cli/cli-script.h (read_command_lines): Update.
* cli/cli-script.c (read_command_lines): Constify prompt_arg.
(MAX_TMPBUF): Remove define.
(define_command): Use string_printf.
(document_command): Likewise.
* breakpoint.c (commands_command_1): Update.
I noticed some code in execute_control_command_1 that could be
simplified by making print_command_trace a printf-like function. This
patch makes this change.
ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
* top.c (execute_command): Update.
* cli/cli-script.h (print_command_lines): Now varargs.
* cli/cli-script.c (print_command_lines): Now varargs.
(execute_control_command_1) <case while_control, case if_control>:
Update.
Currently command lines are reference counted using shared_ptr only
when attached to breakpoints. This patch changes gdb to use
shared_ptr in commands as well. This allows for the removal of
copy_command_lines.
Note that the change to execute_user_command explicitly makes a new
reference to the command line. This will be used in a later patch.
This simplifies struct command_line based on the observation that a
given command can have at most two child bodies: an "if" can have both
"then" and "else" parts. Perhaps the names I've chosen for the
replacements here are not very good -- your input requested.
ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
* tracepoint.c (all_tracepoint_actions): Rename from
all_tracepoint_actions_and_cleanup. Change return type.
(actions_command, encode_actions_1, encode_actions)
(trace_dump_actions, tdump_command): Update.
* remote.c (remote_download_command_source): Update.
* python/python.c (gdbpy_eval_from_control_command)
(python_command, python_interactive_command): Update.
* mi/mi-cmd-break.c (mi_cmd_break_commands): Update.
* guile/guile.c (guile_command)
(gdbscm_eval_from_control_command, guile_command): Update.
* compile/compile.c (compile_code_command)
(compile_print_command, compile_to_object): Update.
* cli/cli-script.h (struct command_lines_deleter): New.
(counted_command_line): New typedef.
(struct command_line): Add constructor, destructor.
<body_list>: Remove.
<body_list_0, body_list_1>: New members.
(command_line_up): Remove typedef.
(read_command_lines, read_command_lines_1, get_command_line):
Update.
(copy_command_lines): Don't declare.
* cli/cli-script.c (build_command_line): Use "new".
(get_command_line): Return counted_command_line.
(print_command_lines, execute_user_command)
(execute_control_command_1, while_command, if_command): Update.
(realloc_body_list): Remove.
(process_next_line, recurse_read_control_structure): Update.
(read_command_lines, read_command_lines_1): Return counted_command_line.
(free_command_lines): Use "delete".
(copy_command_lines): Remove.
(define_command, document_command, show_user_1): Update.
* cli/cli-decode.h (struct cmd_list_element) <user_commands>: Now
a counted_command_line.
* breakpoint.h (counted_command_line): Remove typedef.
(breakpoint_set_commands): Update.
* breakpoint.c (check_no_tracepoint_commands)
(validate_commands_for_breakpoint): Update.
(breakpoint_set_commands): Change commands to be a
counted_command_line.
(commands_command_1, update_dprintf_command_list)
(create_tracepoint_from_upload): Update.
This adds a constructor and destructor to cmd_list_element and changes
it to be allocated with new. This will be useful in a subsequent
patch.
ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
* cli/cli-decode.h (cmd_list_element): New constructor.
(~cmd_list_element): New destructor.
(struct cmd_list_element): Add initializers.
* cli/cli-decode.c (do_add_cmd): Use "new".
(delete_cmd): Use "delete".
Some unaligned watchpoints were currently missed.
On old kernels as specified in
kernel RFE: aarch64: ptrace: BAS: Support any contiguous range (edit)
https://sourceware.org/bugzilla/show_bug.cgi?id=20207
after this patch some other unaligned watchpoints will get reported as false
positives.
With new kernels all the watchpoints should work exactly.
There may be a regresion that it now less merges watchpoints so that with
multiple overlapping watchpoints it may run out of the 4 hardware watchpoint
registers. But as discussed in the original thread GDB needs some generic
watchpoints merging framework to be used by all the target specific code.
Even current FSF GDB code does not merge it perfectly. Also with the more
precise watchpoints one can technically merge them less. And I do not think
it matters too much to improve mergeability only for old kernels.
Still even on new kernels some better merging logic would make sense.
There remains one issue:
kernel-4.15.14-300.fc27.armv7hl
FAIL: gdb.base/watchpoint-unaligned.exp: continue
FAIL: gdb.base/watchpoint-unaligned.exp: continue
(gdb) continue
Continuing.
Unexpected error setting watchpoint: Invalid argument.
(gdb) FAIL: gdb.base/watchpoint-unaligned.exp: continue
But that looks as a kernel bug to me.
(1) It is not a regression by this patch.
(2) It is unrelated to this patch.
gdb/ChangeLog
2018-05-04 Jan Kratochvil <jan.kratochvil@redhat.com>
Pedro Alves <palves@redhat.com>
PR breakpoints/19806 and support for PR external/20207.
* NEWS: Mention Aarch64 watchpoint improvements.
* aarch64-linux-nat.c (aarch64_linux_stopped_data_address): Fix missed
watchpoints and PR external/20207 watchpoints.
* nat/aarch64-linux-hw-point.c
(kernel_supports_any_contiguous_range): New.
(aarch64_watchpoint_offset): New.
(aarch64_watchpoint_length): Support PR external/20207 watchpoints.
(aarch64_point_encode_ctrl_reg): New parameter offset, new asserts.
(aarch64_point_is_aligned): Support PR external/20207 watchpoints.
(aarch64_align_watchpoint): New parameters aligned_offset_p and
next_addr_orig_p. Support PR external/20207 watchpoints.
(aarch64_downgrade_regs): New.
(aarch64_dr_state_insert_one_point): New parameters offset and
addr_orig.
(aarch64_dr_state_remove_one_point): Likewise.
(aarch64_handle_breakpoint): Update caller.
(aarch64_handle_aligned_watchpoint): Likewise.
(aarch64_handle_unaligned_watchpoint): Support addr_orig and
aligned_offset.
(aarch64_linux_set_debug_regs): Remove const from state. Call
aarch64_downgrade_regs.
(aarch64_show_debug_reg_state): Print also dr_addr_orig_wp.
* nat/aarch64-linux-hw-point.h (DR_CONTROL_LENGTH): Rename to ...
(DR_CONTROL_MASK): ... this.
(struct aarch64_debug_reg_state): New field dr_addr_orig_wp.
(unsigned int aarch64_watchpoint_offset): New prototype.
(aarch64_linux_set_debug_regs): Remove const from state.
* utils.c (align_up, align_down): Move to ...
* common/common-utils.c (align_up, align_down): ... here.
* utils.h (align_up, align_down): Move to ...
* common/common-utils.h (align_up, align_down): ... here.
gdb/gdbserver/ChangeLog
2018-05-04 Jan Kratochvil <jan.kratochvil@redhat.com>
Pedro Alves <palves@redhat.com>
* linux-aarch64-low.c (aarch64_stopped_data_address):
Likewise.
gdb/testsuite/ChangeLog
2018-05-04 Jan Kratochvil <jan.kratochvil@redhat.com>
Pedro Alves <palves@redhat.com>
PR breakpoints/19806 and support for PR external/20207.
* gdb.base/watchpoint-unaligned.c: New file.
* gdb.base/watchpoint-unaligned.exp: New file.
Add prefixes or suffixes to some test names to make them unique.
Replace a send_gdb/gdb_expect with a gdb_test, and make the test name
unique.
Remove test of 'help maint' as this is already covered by a later call
to test_prefix_command_help.
Removed test of 'help maint info' and add a new call to
test_prefix_command_help instead.
gdb/testsuite/ChangeLog:
* gdb.base/maint.exp: Make test names unique, use
test_prefix_command_help to test 'help maint info', and remove
repeated test of 'help maint'.
Consider the code in the gdb.ada/array_return.exp testcase, which
defines a function returning an array of 2 integers:
type Data_Small is array (1 .. 2) of Integer;
function Create_Small return Data_Small;
When doing a "finish" from inside function Create_Small, we expect
GDB to tell us that the return value was "(1, 1)". However, it currently
prints the wrong value:
(gdb) finish
Run till exit from #0 pck.create_small () at /[...]/pck.adb:5
p () at /[...]/p.adb:10
10 Large := Create_Large;
Value returned is $1 = (0, 0)
This is a regression which I traced back to the following commit...
| commit 1933fd8ee01ad2e74a9c6341bc40f54962a8f889
| Date: Fri May 19 03:06:19 2017 -0700
| Subject: gdb: fix TYPE_CODE_ARRAY handling in sparc targets
... which, despite what the subject says, is not really about
TYPE_CODE_ARRAY handling, which is a bit of an implementation detail,
but about the GNU vectors extension.
The author of the patch equated TYPE_CODE_ARRAY with vectors, which
is not correct. Vectors are TYPE_CODE_ARRAY types with the TYPE_VECTOR
flag set. So at the very minimum, the patch should have been checking
for both TYPE_CODE_ARRAY and TYPE_VECTOR.
But, that's not the only thing that did not seem right to me. When
looking at the ABI, and at the summary of the implementation in GCC
of the calling conventions for that architecture:
size argument return value
small integer <4 int. reg. int. reg.
word 4 int. reg. int. reg.
double word 8 int. reg. int. reg.
_Complex small integer <8 int. reg. int. reg.
_Complex word 8 int. reg. int. reg.
_Complex double word 16 memory int. reg.
vector integer <=8 int. reg. FP reg.
vector integer >8 memory memory
float 4 int. reg. FP reg.
double 8 int. reg. FP reg.
long double 16 memory memory
_Complex float 8 memory FP reg.
_Complex double 16 memory FP reg.
_Complex long double 32 memory FP reg.
vector float any memory memory
aggregate any memory memory
The nice thing about the patch above is that it nicely factorized
the code that determines how arguments are passed/returns. The bad
news is that the implementation, particularly for the handling of
arrays and vectors, doesn't seem to match the summary above. Hence,
the regression we observed.
So what I did was review and re-implement some of the predicate functions
according to the summary above. Because dejagnu crashes all our Solaris
machines real bad, I can't run the dejagnu testsuite there. So what I did
was test the patch with AdaCore's testsuite against leon3-elf, no
regression. I verified that this fixes the regression above while
at the same time still passing gdb.base/gnu_vector.exp (I transposed
that testcase to our testsuite), which is the testcase that was cited
in the commit above as seeing some FAIL->PASS improvements.
This patch also removes one assertion...
gdb_assert (sparc_integral_or_pointer_p (type)
|| (TYPE_CODE (type) == TYPE_CODE_ARRAY && len <= 8));
... because that assertion is really the "negative" of the other conditions
written in the same "if, else if, else [assert]" block in this function.
To me, this assertion forces us to maintain two versions of the same code,
and is an unnecessary burden. In particular, the above is not the
correct condition, and the ABI summary table above shows that we need
a more complex condition to describe the situations where we expect
arguments to be passed by register.
gdb/ChangeLog:
* sparc-tdep.c (sparc_structure_return_p): Re-implement to
match the ABI as summarized in GCC's gcc/config/sparc/sparc.c.
(sparc_arg_by_memory_p): Renamed from sparc_arg_on_registers_p.
Re-implement to match the ABI as summarized in GCC's
gcc/config/sparc/sparc.c. All callers updated.
(sparc32_store_arguments): Remove assertion.
This changes decode_format to use skip_spaces, and changes printcmd.c
not to include tui.h, which apparently is not needed.
2018-05-04 Tom Tromey <tom@tromey.com>
* printcmd.c: Don't include tui.h.
(decode_format): Use skip_spaces.
About the 'x' command, the manual says:
If you use <RET> to repeat the 'x' command, the repeat count N is
used again; the other arguments default as for successive uses of
'x'.
However, PR gdb/22619 points out that this does not work.
This patch fixes the problem.
ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
PR gdb/22619:
* printcmd.c (last_count): New global.
(x_command): Use saved count when repeating.
testsuite/ChangeLog
2018-05-04 Tom Tromey <tom@tromey.com>
PR gdb/22619:
* gdb.base/long_long.exp (gdb_test_long_long): Add test for repeat
behavior.
