Alan Hayward [Thu, 1 Mar 2018 11:04:00 +0000 (11:04 +0000)]
[PATCH v3 6/8] Create xml from target descriptions
This patch adds a print_xml_feature visitor class which turns a C
target description into xml. Both gdb and gdbserver can do this.
An accept function is added to gdbserver tdesc to allow it to use
vistor classes.
Tests are added to maintenance_check_xml_descriptions which takes
each pair of tested descriptions, turns them both into xml, then back
again, and confirms the descriptions still match.
Alan.
2018-02-26 Alan Hayward <alan.hayward@arm.com>
gdb/
* common/tdesc.c (print_xml_feature::visit_post): Add xml parsing.
(print_xml_feature::visit_pre): Likewise.
(print_xml_feature::visit_post): Likewise.
(print_xml_feature::visit): Likewise.
(print_xml_feature::visit): Likewise.
(print_xml_feature::visit): Likewise.
(print_xml_feature::visit): Likewise.
* common/tdesc.h (print_xml_feature): Add new class.
* regformats/regdat.sh: obtain xml.
* target-descriptions.c (struct target_desc): Add xmltarget.
(print_xml_feature::visit_pre): Add xml vistor.
(tdesc_get_features_xml): Add function to get xml.
(maintenance_check_xml_descriptions): Test xml generation.
* xml-tdesc.c (target_read_description_xml_string): Add function.
* xml-tdesc.h (target_read_description_xml_string): Add declaration.
gdbserver/
* tdesc.c (void target_desc::accept): Fill in function.
(tdesc_get_features_xml): Remove old xml creation.
(print_xml_feature::visit_pre): Add xml vistor.
Alan Hayward [Thu, 1 Mar 2018 10:59:58 +0000 (10:59 +0000)]
[PATCH v3 2/8] Commonise tdesc_reg
This patch commonises tdesc_reg and makes use of it in gdbserver tdesc.
gdbserver tdesc_create_reg is changed to create a tdesc_reg instead of
a reg_defs entry. The vector of tdesc_reg is held inside tdesc_feature.
However, other modules in gdbserver directly access the reg_defs structure.
To work around this, init_target_desc fills in reg_defs by parsing the
tdesc_reg vector.
The long term goal is to remove reg_defs, replacing with accessor funcs.
I wanted to make tdesc_create_reg common, but I cannot do that until
the next patch.
I also had to commonise tdesc_element_visitor and tdesc_element.
Alan Hayward [Thu, 1 Mar 2018 10:58:49 +0000 (10:58 +0000)]
[PATCH V3 1/8] Move tdesc header funcs to c file
Move the destructor and equals operator for gdbserver tdesc from the .h
to the .c file. Both functions are too long to be inlined and make the
header look messy. Patch does not change any functionality.
2018-03-01 Alan Hayward <alan.hayward@arm.com>
gdbserver/
* tdesc.c (target_desc::~target_desc): Move to here.
(target_desc::operator==): Likewise.
* tdesc.h (target_desc::~target_desc): Move from here.
(target_desc::operator==): Likewise.
Change order of error message printed when gdbserver can't find CWD
I forgot to address Pedro's comment about my last patch and change the
order of the message printed when getcwd returns NULL on gdbserver.
This obvious commit does it.
Simon mentioned on IRC that, after the startup-with-shell feature has
been implemented on gdbserver, it is not possible to specify a
filename-only binary, like:
$ gdbserver :1234 a.out
/bin/bash: line 0: exec: a.out: not found
During startup program exited with code 127.
Exiting
This happens on systems where the current directory "." is not listed
in the PATH environment variable. Although including "." in the PATH
variable is a possible workaround, this can be considered a regression
because before startup-with-shell it was possible to use only the
filename (due to reason that gdbserver used "exec*" directly).
The idea of the patch is to verify if the program path provided by the
user (or by the remote protocol) contains a directory separator
character. If it doesn't, it means we're dealing with a filename-only
binary, so we call "gdb_abspath" to properly expand it and transform
it into a full path. Otherwise, we leave the program path untouched.
This mimicks the behaviour seen on GDB (look at "openp" and
"attach_inferior", for example).
I am also submitting a testcase which exercises the scenario described
above. This test requires gdbserver to be executed in a different CWD
than the original, so I also created a helper function, "with_cwd" (on
testsuite/lib/gdb.exp), which takes care of cd'ing into and out of the
specified dir.
Built and regtested on BuildBot, without regressions.
gdb/ChangeLog:
2018-02-28 Sergio Durigan Junior <sergiodj@redhat.com>
Simon Marchi <simon.marchi@polymtl.ca>
* common/common-utils.c: Include "sys/stat.h".
(is_regular_file): Move here from "source.c"; change return
type to "bool".
* common/common-utils.h (is_regular_file): New prototype.
* common/pathstuff.c (contains_dir_separator): New function.
* common/pathstuff.h (contains_dir_separator): New prototype.
* source.c: Don't include "sys/stat.h".
(is_regular_file): Move to "common/common-utils.c".
* server.c: Include "filenames.h" and "pathstuff.h".
(program_name): Delete variable.
(program_path): New anonymous class.
(get_exec_wrapper): Use "program_path" instead of
"program_name".
(handle_v_run): Likewise.
(captured_main): Likewise.
(process_serial_event): Likewise.
This commit moves the path manipulation routines found on utils.c to a
new common/pathstuff.c, and updates the Makefile.in's accordingly.
The routines moved are "gdb_realpath", "gdb_realpath_keepfile" and
"gdb_abspath".
This will be needed because gdbserver will have to call "gdb_abspath"
on my next patch, which implements a way to expand the path of the
inferior provided by the user in order to allow specifying just the
binary name when starting gdbserver, like:
$ gdbserver :1234 a.out
With the recent addition of the startup-with-shell feature on
gdbserver, this scenario doesn't work anymore if the user doesn't have
the current directory listed in the PATH variable.
I had to do a minor adjustment on "gdb_abspath" because we don't have
access to "tilde_expand" on gdbserver, so now the function is using
"gdb_tilde_expand" instead. Otherwise, the code is the same.
Regression tested on the BuildBot, without regressions.
I removed gdb_is_target_remote_prompt, but did not realize it was used
in mi_is_target_remote. This makes the gdb.mi/mi-nonstop.exp crash, for
example:
ERROR: (DejaGnu) proc "gdb_is_target_remote_prompt {[(]gdb[)]
}" does not exist.
The error code is TCL LOOKUP COMMAND gdb_is_target_remote_prompt
The info on the error is:
invalid command name "gdb_is_target_remote_prompt"
while executing
"::tcl_unknown gdb_is_target_remote_prompt {[(]gdb[)]
}"
("uplevel" body line 1)
invoked from within
"uplevel 1 ::tcl_unknown $args"
This patch restores it.
gdb/testsuite/ChangeLog:
* lib/gdb.exp (gdb_is_target_1): Add prompt_regexp parameter and
use it.
(gdb_is_target_remote_prompt): New proc.
(gdb_is_target_remote): Use gdb_is_target_remote_prompt.
(gdb_is_target_native): Pass prompt parameter to
gdb_is_target_1.
Nick Clifton [Wed, 28 Feb 2018 11:50:49 +0000 (11:50 +0000)]
Catch integer overflows/underflows when parsing corrupt DWARF FORM blocks.
PR 22895
PR 22893
* dwarf2.c (read_n_bytes): Replace size parameter with dwarf_block
pointer. Drop unused abfd parameter. Check the size of the block
before initialising the data field. Return the end pointer if the
size is invalid.
(read_attribute_value): Adjust invocations of read_n_bytes.
Douglas B Rupp [Wed, 28 Feb 2018 11:44:26 +0000 (11:44 +0000)]
Remove any trailing directory separator from ld_canon_sysroot and adjust ld_canon_sysroot_len.
The problem was the ld expect libiberty/lrealpath() to always return a
path, but in some cases it returns a prefix. It seemed like too much of
an earthquake to propose a fix to lrealpath.
* ldmain.c (main): Remove directory separator, if any, from the
end of the canonicalized sysroot.
Alan Modra [Wed, 28 Feb 2018 06:44:24 +0000 (17:14 +1030)]
Nonsense error messages on invalid aout string offset
translate_symbol_table returns false on detecting an out of range name
string offset, hooray for error checking, but doesn't set bfd_error or
print a useful error. bfd_error therefore contains whatever it had
previously, in my testing, bfd_error_system_call. So the error
printed depended on errno.
PR 22887
* aoutx.h (translate_symbol_table): Print an error message and set
bfd_error on finding an invalid name string offset.
John Baldwin [Wed, 25 Oct 2017 04:06:00 +0000 (21:06 -0700)]
Workaround a FreeBSD ptrace() bug with clearing thread events.
When multiple threads within a process wish to report STOPPED events
from wait(), the kernel picks one thread event as the thread event to
report. The chosen thread event is retrieved via PT_LWPINFO by
passing the process ID as the request pid. If multiple events are
pending, then the subsequent wait() after resuming a process will
report another STOPPED event after resuming the process to handle the
next thread event and so on.
A single thread event is cleared as a side effect of resuming the
process with PT_CONTINUE, PT_STEP, etc. In older kernels, however,
the request pid was used to select which thread's event was cleared
rather than always clearing the event that was just reported. To
avoid clearing the event of the wrong LWP, always pass the process ID
instead of an LWP ID to PT_CONTINUE or PT_SYSCALL.
In the case of stepping, the process ID cannot be used with PT_STEP
since it would step the thread that reported an event which may not be
the thread indicated by PTID. For stepping, use PT_SETSTEP to enable
stepping on the desired thread before resuming the process via
PT_CONTINUE instead of using PT_STEP.
This manifested as a failure in the
gdb.threads/continue-pending-status.exp test. Specifically, if thread
2 reported a breakpoint and the test thus switched to thread 3 before
continuing, thread 3's event (if any) was discarded and thread 2's
breakpoint remained pending and was reported a second time as a
duplicate event. As a result, the PC was decremented twice for the
same breakpoint resulting in an illegal instruction fault on x86.
gdb/ChangeLog:
* fbsd-nat.c (fbsd_resume): Use PT_SETSTEP for stepping and a
wildcard process pid for super_resume for kernels with a
specific bug.
H.J. Lu [Tue, 27 Feb 2018 22:45:48 +0000 (14:45 -0800)]
gas: Rename .nop directive to .nops
Since directives of NO_PSEUDO_DOT targets don't have the leading '.'
and "nop" can be a valid instruction, rename .nop directive to .nops
to avoid conflict.
Phil Muldoon [Tue, 27 Feb 2018 20:21:23 +0000 (12:21 -0800)]
Update get_args documentation
This patch adds argument compilation documentation, expanding on the
already existing comments, giving a more thorough explanation of
the source of the arguments used in the final argument string.
H.J. Lu [Tue, 27 Feb 2018 19:34:20 +0000 (11:34 -0800)]
ld: Add --enable-separate-code
This patch adds --enable-separate-code to ld configure to turn on
-z separate-code by default and enables it by default for Linux/x86.
This avoids mixing code pages with data to improve cache performance
as well as security.
To reduce x86-64 executable and shared object sizes, the maximum page
size is reduced from 2MB to 4KB when -z separate-code is turned on by
default. Note: -z max-page-size= can be used to set the maximum page
size.
We compared SPEC CPU 2017 performance before and after this change on
Skylake server. There are no any significant performance changes.
Everything is mostly below +/-1%.
bfd/
* config.in: Regenerated.
* configure: Likewise.
* configure.ac: Add --enable-separate-code.
(DEFAULT_LD_Z_SEPARATE_CODE): New AC_DEFINE_UNQUOTED. Default
to 1 for Linux/x86 targets,
* elf64-x86-64.c (ELF_MAXPAGESIZE): Set to 0x1000 if
DEFAULT_LD_Z_SEPARATE_CODE is 1.
ld/
* NEWS: Mention --enable-separate-code.
* configure.ac: Add --enable-separate-code.
(DEFAULT_LD_Z_SEPARATE_CODE): New AC_DEFINE_UNQUOTED.
* configure.tgt: Default ac_default_ld_z_separate_code to 1 for
Linux/x86 targets.
* config.in: Regenerated.
* configure: Likewise.
* emultempl/elf32.em (gld${EMULATION_NAME}_before_parse): Set
link_info.separate_code DEFAULT_LD_Z_SEPARATE_CODE.
Among the macros to manipulate an arm_feature_set structure is the
ARM_FEATURE_COPY which copy the value of a feature set into another.
This can be achieved with a simple assignement which most of the
existing code does. This patch removes the last 2 uses of that macro and
remove the macro altogether.
2018-02-27 Thomas Preud'homme <thomas.preudhomme@arm.com>
This optimization is enabled with -O, -O2 and -Os.
2. Since 0xb0 mov with 32-bit destination registers zero-extends 32-bit
immediate to 64-bit destination register, we can use it to encode 64-bit
mov with 32-bit immediates. This optimization is enabled with -O, -O2
and -Os.
3. Since the upper bits of destination registers of VEX128 and EVEX128
instructions are extended to zero, if all bits of destination registers
of AVX256 or AVX512 instructions are zero, we can use VEX128 or EVEX128
encoding to encode AVX256 or AVX512 instructions. When 2 source
registers are identical, AVX256 and AVX512 andn and xor instructions:
VOP %reg, %reg, %dest_reg
can be encoded with
VOP128 %reg, %reg, %dest_reg
This optimization is enabled with -O2 and -Os.
4. 16-bit, 32-bit and 64-bit register tests with immediate may be
encoded as 8-bit register test with immediate. This optimization is
enabled with -Os.
This patch does:
1. Add {nooptimize} pseudo prefix to disable instruction size
optimization.
2. Add optimize to i386_opcode_modifier to tell assembler that encoding
of an instruction may be optimized.
binutils PR 22875
* objcopy.c (copy_object): Check the error status after marking symbols used in relocations.
* testsuite/binutils-all/strip-13.s: New test source file.
* testsuite/binutils-all/strip-13.s: New test driver file.
* testsuite/binutils-all/objcopy.exp: Run the new test.
Jim Wilson [Mon, 26 Feb 2018 22:05:13 +0000 (14:05 -0800)]
IA-64: Fix linker error with --no-keep-memory.
bfd/
PR 15904
* elfnn-ia64.c (elfNN_ia64_relax_section): After ia64_elf_relax_brl
call, set changed_contents and changed_relocs. Likewise after
successful ia64_elf_relax_br call.
Simon Marchi [Mon, 26 Feb 2018 19:30:43 +0000 (14:30 -0500)]
Add test for load command
There doesn't seem to by any test for the load command. I suggest to
add this test, so that we can have a minimum of confidence we don't
break it completely while refactoring the code that implements it.
gdb/testsuite/ChangeLog:
* gdb.base/load-command.c: New file.
* gdb.base/load-command.exp: New file.
* lib/gdb.exp (gdb_is_target_remote_prompt): Rename to...
(gdb_is_target_1): ...this, and generalize for other targets
than just remote.
(gdb_is_target_remote): Use gdb_is_target_1.
(gdb_is_target_native): use gdb_is_target_1.
MIPS: Don't use a 32-bit BFD architecture with a 64-bit ABI
Select `bfd_mach_mips4000', which corresponds to the MIPS III ISA, the
earlies with 64-bit support, whenever a 32-bit BFD architecture has been
chosen to use with a 64-bit ABI. The situation can happen in a few
cases:
1. When the user has used `set architecture' or `set mips abi' commands
to override automatic selection and then starts a debug session by
requesting to run, attach or connect to a target.
2. In native debugging when reattaching to a previously debugged process
where the program to be debugged has been since discarded, as
observed with:
FAIL: gdb.base/attach.exp: attach2, with no file (GDB internal error)
in n32 and n64 regression testing.
3. In remote debugging with a non-XML debug stub when discarding the
program to be debugged while connected to the remote target, as
observed with:
FAIL: gdb.base/break-unload-file.exp: cmdline: always-inserted on: break: file (GDB internal error)
in n32 and n64 regression testing.
In the latter two cases the ABI, quite rightfully, is retained while the
program to be debugged is discarded. This is because in that case the
ABI previously determined is carried over along with `gdbarch' in use,
which is retained. The BFD architecture is however discarded and the
default then applies, because it is not attached to `gdbarch'.
In all these cases we trip with an internal error message as follows:
.../gdb/mips-tdep.c:766: internal-error: bad register size
A problem internal to GDB has been detected,
further debugging may prove unreliable.
Quit this debugging session? (y or n) n
This is a bug, please report it. For instructions, see:
<http://www.gnu.org/software/gdb/bugs/>.
coming from `mips_pseudo_register_read', because the raw register width
inferred from the BFD architecture turns out to be 4 for the general
registers while the cooked register width inferred from the ABI in
effect is 8.
We do not hit this internal error in remote debugging with an XML debug
stub, because in that case raw register width information is passed by
the stub along with the XML target description.
Ultimately I think we ought to make the BFD architecture sticky like the
ABI, however in the interim this simple fix will do, removing the error
across all three cases. The case where the user has used `set mips abi'
or `set architecture' commands has to be handled anyway, and although a
more sophisticated solution could be envisaged, such as reporting an
error with the respective `set' command, I think this is too much of a
corner case to bother.
gdb/
* mips-tdep.c (mips_gdbarch_init): Don't use a 32-bit BFD
architecture with a 64-bit ABI.
MIPS: Reorder ABI determination ahead of target description loading
Move ABI determination code ahead of target description loading so that
architecture information can be adjusted according to the ABI selected,
and then used in OS dependent register information initialization needed
for target description processing. No functional change.
gdb/
* gdb/mips-tdep.c (mips_gdbarch_init): Reorder ABI determination
ahead of target description loading.
Tom Tromey [Wed, 26 Apr 2017 04:33:50 +0000 (22:33 -0600)]
Change frame_filter_flags to use DEF_ENUM_FLAGS_TYPE
This changes frame_filter_flags to use DEF_ENUM_FLAGS_TYPE, and
updates all the uses. It also changes the enum constants to use <<,
as suggested by Sergio.
ChangeLog
2018-02-26 Tom Tromey <tom@tromey.com>
* stack.c (backtrace_command_1): Update.
* python/python-internal.h (gdbpy_apply_frame_filter): Change type
of "flags".
* python/py-framefilter.c (py_print_frame)
(gdbpy_apply_frame_filter): Change type of "flags".
* mi/mi-cmd-stack.c (mi_apply_ext_lang_frame_filter): Change type
of "flags".
(mi_cmd_stack_list_frames, mi_cmd_stack_list_locals)
(mi_cmd_stack_list_args, mi_cmd_stack_list_variables): Update.
* extension.h (enum frame_filter_flag): Rename from
frame_filter_flags.
(frame_filter_flags): Define using DEF_ENUM_FLAGS_TYPE.
(apply_ext_lang_frame_filter): Change type of "flags".
* extension.c (apply_ext_lang_frame_filter): Change type of
"flags".
* extension-priv.h (struct extension_language_ops)
<apply_frame_filter>: Change type of "flags".
Tom Tromey [Mon, 20 Jun 2016 16:36:29 +0000 (10:36 -0600)]
Make "bt N" print correct number of frames when using a frame filter
PR python/16497 notes that using "bt" with a positive argument prints
the wrong number of frames when a frame filter is in use. Also, in this
case, the non-frame-filter path will print a message about "More stack
frames" when there are more; but this is not done in the frame-filter
case.
The first problem is that backtrace_command_1 passes the wrong value
to apply_ext_lang_frame_filter -- that function takes the final
frame's number as an argument, but backtrace_command_1 passes the
count, which is off by one.
The solution to the second problem is to have the C stack-printing
code stop at the correct number of frames and then print the message.
Tested using the buildbot.
ChangeLog
2018-02-26 Tom Tromey <tom@tromey.com>
PR python/16497:
* stack.c (backtrace_command_1): Set PRINT_MORE_FRAMES flag. Fix
off-by-one in py_end computation.
* python/py-framefilter.c (gdbpy_apply_frame_filter): Handle
PRINT_MORE_FRAMES.
* extension.h (enum frame_filter_flags) <PRINT_MORE_FRAMES>: New
constant.
Tom Tromey [Wed, 10 Jan 2018 20:53:18 +0000 (13:53 -0700)]
Handle DW_TAG_variant_part and DW_TAG_variant
This changes dwarf2read to understand DW_TAG_variant_part and
DW_TAG_variant.
Note that DW_AT_discr_list is not handled. I did not need this for
Rust. I imagine this should not be too hard to add later, should
someone need it. Meanwhile I have gdb emit a complaint if it is seen.
There is a lurking issue concerning the placement of the discriminant
in the DWARF. For Rust, I ended up following the letter of the
standard and having the discriminant be a child of the
DW_TAG_variant_part. However, GCC's Ada support does not do this.
Pierre-Marie filed this with the DWARF committee:
http://dwarfstd.org/ShowIssue.php?issue=180123.1
However as that is read-only, if you have comments you might consider
adding them to the GCC bug:
Finally, there is a DWARF extension lurking in here. In Rust, a
univariant enum will not have a discriminant. However, in order to
unify the representation of all data-carrying enums, I've made LLVM
(and my forthcoming rustc patch) emit a univariant enum using a
DW_TAG_variant with a single variant part and without DW_AT_discr.
The lack of this DW_AT_discr is the extension. I will submit an issue
on dwarfstd.org about this.
2018-02-26 Tom Tromey <tom@tromey.com>
* dwarf2read.c (struct variant_field): New.
(struct nextfield) <variant>: New field.
(dwarf2_add_field): Handle DW_TAG_variant_part.
(dwarf2_attach_fields_to_type): Attach a discriminant_info to a
discriminated union.
(read_structure_type): Handle DW_TAG_variant_part.
(handle_struct_member_die): New function, extracted from
process_structure_scope. Handle DW_TAG_variant.
(process_structure_scope): Handle discriminated unions. Call
handle_struct_member_die.
2018-02-26 Tom Tromey <tom@tromey.com>
* gdb.dwarf2/variant.c: New file.
* gdb.dwarf2/variant.exp: New file.
Tom Tromey [Fri, 9 Feb 2018 20:31:51 +0000 (13:31 -0700)]
Convert Rust to use discriminated unions
A Rust enum is, essentially, a discriminated union. Currently the
Rust language support handles Rust enums locally, in rust-lang.c.
However, because I am changing the Rust compiler to use
DW_TAG_variant* to represent enums, it seemed better to have a single
internal representation for Rust enums in gdb.
This patch implements this idea by moving the current Rust enum
handling code to dwarf2read. This allows the simplification of some
parts of rust-lang.c as well.
2018-02-26 Tom Tromey <tom@tromey.com>
* rust-lang.h (rust_last_path_segment): Declare.
* rust-lang.c (rust_last_path_segment): Now public. Change
contract.
(struct disr_info): Remove.
(RUST_ENUM_PREFIX, RUST_ENCODED_ENUM_REAL)
(RUST_ENCODED_ENUM_HIDDEN, rust_union_is_untagged)
(rust_get_disr_info, rust_tuple_variant_type_p): Remove.
(rust_enum_p, rust_enum_variant): New function.
(rust_underscore_fields): Remove "offset" parameter.
(rust_print_enum): New function.
(rust_val_print) <TYPE_CODE_UNION>: Remove enum code.
<TYPE_CODE_STRUCT>: Call rust_print_enum when appropriate.
(rust_print_struct_def): Add "for_rust_enum" parameter. Handle
enums.
(rust_internal_print_type): New function, from rust_print_type.
Remove enum code.
(rust_print_type): Call rust_internal_print_type.
(rust_evaluate_subexp) <STRUCTOP_ANONYMOUS, STRUCTOP_STRUCT>:
Update enum handling.
* dwarf2read.c (struct dwarf2_cu) <rust_unions>: New field.
(rust_fully_qualify, alloc_discriminant_info, quirk_rust_enum)
(rust_union_quirks): New functions.
(process_full_comp_unit, process_full_type_unit): Call
rust_union_quirks.
(process_structure_scope): Update rust_unions if necessary.
2018-02-26 Tom Tromey <tom@tromey.com>
* gdb.rust/simple.exp: Accept more possible results in enum test.
Tom Tromey [Fri, 9 Feb 2018 20:31:42 +0000 (13:31 -0700)]
Initial support for variant parts
This adds some initial support for variant parts to gdbtypes.h. A
variant part is represented as a union. The union has a flag
indicating that it has a discriminant, and information about the
discriminant is attached using the dynamic property system.
2018-02-26 Tom Tromey <tom@tromey.com>
* value.h (value_union_variant): Declare.
* valops.c (value_union_variant): New function.
* gdbtypes.h (TYPE_FLAG_DISCRIMINATED_UNION): New macro.
(struct discriminant_info): New.
(enum dynamic_prop_node_kind) <DYN_PROP_DISCRIMINATED>: New
enumerator.
(struct main_type) <flag_discriminated_union>: New field.
Tom Tromey [Wed, 21 Feb 2018 17:36:55 +0000 (10:36 -0700)]
Sign-extend non-bit-fields in unpack_bits_as_long
unpack_bits_as_long is documented as sign-extending its result when
the type is signed. However, it was only doing sign-extension in the
case where the field was a bitfield -- that is, not when the "bitsize"
parameter was 0, indicating the size should be taken from the type.
Also, unpack_bits_as_long was incorrectly computing the shift for
big-endian architectures for the non-bitfield case.
This patch fixes these bugs in a straightforward way. A new selftest
is included.
2018-02-26 Tom Tromey <tom@tromey.com>
* Makefile.in (SUBDIR_UNITTESTS_SRCS): Add
unittests/unpack-selftests.c.
* unittests/unpack-selftests.c: New file.
* value.c (unpack_bits_as_long): Fix bugs in non-bitfield cases.
Yao Qi [Mon, 26 Feb 2018 15:38:01 +0000 (15:38 +0000)]
Move read_partial_die to partial_die_info::read
gdb:
2018-02-26 Yao Qi <yao.qi@linaro.org>
* dwarf2read.c (struct partial_die_info) <read>: New method.
(read_partial_die): Remove the declaration.
(load_partial_dies): Update.
(partial_die_info::partial_die_info):
(read_partial_die): Change it to partial_die_info::read.
Yao Qi [Mon, 26 Feb 2018 15:38:01 +0000 (15:38 +0000)]
Class-fy partial_die_info
This patch is to class-fy partial_die_info. Two things special here,
- disable assignment operator, but keep copy ctor, which is used in
load_partial_dies,
- have a private ctor which is only used by dwarf2_cu::find_partial_die,
I don't want other code use it, so make it private,
gdb:
2018-02-26 Yao Qi <yao.qi@linaro.org>
* dwarf2read.c (struct partial_die_info): Add ctor, delete
assignment operator.
(load_partial_dies): Use ctor and copy ctor.
(read_partial_die): Update.
(dwarf2_cu::find_partial_die): Use ctor.
Yao Qi [Mon, 26 Feb 2018 15:38:01 +0000 (15:38 +0000)]
Change find_partial_die_in_comp_unit to dwarf2_cu::find_partial_die
This patch changes find_partial_die_in_comp_unit to a dwarf2_cu method
find_partial_die.
gdb:
2018-02-26 Yao Qi <yao.qi@linaro.org>
* dwarf2read.c (struct dwarf2_cu) <find_partial_die>: New method.
(find_partial_die_in_comp_unit): Change it to
dwarf2_cu::find_partial_die.
(find_partial_die): Update.
the code was written in a way that spaces are allocated on necessary on
cu->comp_unit_obstack. I want to class-fy partial_die_info, but
partial_die_info ctor can't follow XOBNEW immediately, so this patch
rewrite this loop to:
we create a local variable pdi, if we need it, call XOBNEW, and copy.
This also reduce one XOBNEW call. I measured the number of XOBNEW call in
load_partial_dies when gdb reads dwarf2read.o, without this patch, it is
18827, and with this patch, it is 18826.
gdb:
2018-026-26 Yao Qi <yao.qi@linaro.org>
* dwarf2read.c (load_partial_dies): Move the location of XOBNEW.
Alan Modra [Sun, 25 Feb 2018 21:46:24 +0000 (08:16 +1030)]
Segfault on phdrs allocated but not created
A number of ARM tests show results like:
ld-new: no address assigned to the veneers output section .gnu.sgstubs
ld-new: can not size stub section: invalid operation
ld-new: warning: cannot find entry symbol _start; defaulting to 0000000000010074
ld-new: tmpdir/cmse-veneers-no-gnu_sgstubs: warning: allocated section `.text' not in segment
Segmentation fault
PASS: Secure gateway veneers: no .gnu.sgstubs section
The PASS despite the segv shows the danger of using "#..." to match
error output. This patch doesn't try to fix the testuite, but does
prevent the segfault.
* elf.c (assign_file_positions_except_relocs): Don't segfault
when actual number of phdrs is less than allocated. Delete
outdated comment. Formatting.
Alan Modra [Fri, 23 Feb 2018 23:58:12 +0000 (10:28 +1030)]
ld error/warning messages
This patch standardizes messages in ld, to better conform to the GNU
coding standard. Besides issues of capitalization and full-stops,
I've
- Split up help messages for target options, so that adding a new
option does not mean loss of translation for all the others.
- Embedded tabs have been removed, since a user might have tab stops
set at other than 8 char intervals.
- Added missing program name (%P). ld isn't the compiler.
- Put %F and %X first (and removed %X if %F was present). These can
go anywhere, but look silly in the m%Fiddle of a message, and
choosing "%P%F:" in some messages but "%F%P:" in others leads to the
likelihood of duplication in ld.pot. Besides, the colon belongs
with %P.
Alan Modra [Sat, 24 Feb 2018 08:20:21 +0000 (18:50 +1030)]
ld message translation for emultempl and emulparams
Most of the emultempl/*.em error/warning messages do not currently
appear in ld.pot, and fixing that is not simply a matter of adding
missing files to POTFILES. The difficulty is the shell-script
quoting, for example aarch64elf.em:PARSE_AND_LIST_OPTIONS "Don'\''t".
I suppose you could avoid contractions like "don't" but I'm unsure as
to whether gettext could handle everything in the emulparams scripts.
The right thing to do is feed the generated C files to gettext, as
this patch does. The patch just copies what is already done in bfd/
for generated files.
* Makefile.am (EMULATION_FILES, POTFILES): Delete.
(SRC_POTFILES, BLD_POTFILES): Define.
(po/POTFILES.in): Delete rule.
(po/SRC-POTFILES.in, po/BLD-POTFILES.in): New rules.
* configure.ac: Add AC_CONFIG_COMMANDS to create po/Makefile.
* po/Make-in (DISTFILES): Remove POTFILES.in, add SRC-POTFILES.in
and BLD-POTFILES.in.
(POTFILES): Delete.
(SRC-POTFILES, BLD-POTFILES): Define place marker.
(ld.pot): Build from SRC-POTFILES plus BLD-POTFILES.
(distclean, maintainer-clean): Tidy up new files.
(POTFILES, POTFILES.in): Delete rules.
(SRC-POTFILES, BLD-POTFILES, SRC-POTFILES.in, BLD-POTFILES.in):
New rules.
(Makefile): Depend on SRC-POTFILES and BLD-POTFILES.
* po/SRC-POTFILES.in: Rename from po/POTFILES.in.
* po/BLD-POTFILES.in: New file.
* Makefile.in: Regenerate.
* configure: Regenerate.
Alan Modra [Wed, 21 Feb 2018 02:34:48 +0000 (13:04 +1030)]
ARM and AArch64 messages
More standarization of error/warning messages. The ARM backend has
some multi-line error messages, which don't fit particularly well with
the GNU coding standard insistence that error messages shouldn't be
capitalized or have a full-stop. I've replaced the line breaks with
semicolons.
So for instance the following
system/path/to/ld: myobject.o: invalid special symbol `foo'.
system/path/to/ld: It must be a global or weak function symbol.
becomes
system/path/to/ld: myobject.o: invalid special symbol `foo'; it must be a global or weak function symbol
Alan Modra [Wed, 21 Feb 2018 11:17:07 +0000 (21:47 +1030)]
unrecognized/unsupported reloc message
It must get boring translating all the variants we have of
unrecognized/unsupported/invalid/unexpected reloc number. This patch
cuts down on the number of variations.
Alan Modra [Tue, 20 Feb 2018 08:59:09 +0000 (19:29 +1030)]
PPC error/warning messages
The GNU coding standard says error messages should be of the form
program:sourcefile:lineno: message
or
program: message
and
"The string message should not begin with a capital letter when it
follows a program name and/or file name, because that isn’t the
beginning of a sentence. (The sentence conceptually starts at the
beginning of the line.) Also, it should not end with a period."
This patch does that for ppc, and removes some British spelling.
I've also switched some error output from using the linker callback
einfo to _bfd_error_handler, due to improved compilation time
argument checking now done for the latter function.
bfd/
* elf32-ppc.c: Standardize error/warning messages. Use
_bfd_error_handler rather than einfo when einfo features not used.
* elf64-ppc.c: Likewise.
ld/
* testsuite/ld-powerpc/attr-gnu-12-21.d: Update.
* testsuite/ld-powerpc/attr-gnu-4-12.d: Update.
* testsuite/ld-powerpc/attr-gnu-4-13.d: Update.
* testsuite/ld-powerpc/attr-gnu-4-21.d: Update.
* testsuite/ld-powerpc/attr-gnu-4-23.d: Update.
* testsuite/ld-powerpc/attr-gnu-4-31.d: Update.
* testsuite/ld-powerpc/attr-gnu-4-32.d: Update.
* testsuite/ld-powerpc/attr-gnu-8-23.d: Update.
The issue is that they have an older gcc (not c++11 by default?) that
doesn't emit the enum underlying type information. When the
enum type is printed by ptype, it looks like this:
enum S10::E10 {S10::A10, S10::B10, S10::C10};
instead of this on older gccs:
enum S10::E10 : unsigned int {S10::A10, S10::B10, S10::C10};
The regex that matches this is in cp_test_ptype_class, and is
enum $nested_name (: (unsigned )?int)? \{
If the "unsigned int" portion is not present, then it requires the
string to have two spaces between the enum name and opening bracket.
The fix is simply to move the trailing space inside the ? group.
gdb/testsuite/ChangeLog:
* lib/cp-support.exp (cp_test_ptype_class): Move space inside
parentheses.
Tom Tromey [Wed, 21 Feb 2018 23:53:56 +0000 (16:53 -0700)]
Remove cleanups from check_fast_tracepoint_sals
This changes the gdbarch fast_tracepoint_valid_at method to use a
std::string as its out parameter, and then updates all the uses. This
allows removing a cleanup from breakpoint.c.
Regression tested by the buildbot.
ChangeLog
2018-02-24 Tom Tromey <tom@tromey.com>
* i386-tdep.c (i386_fast_tracepoint_valid_at): "msg" now a
std::string.
* gdbarch.sh (fast_tracepoint_valid_at): Change "msg" to a
std::string*.
* gdbarch.c: Rebuild.
* gdbarch.h: Rebuild.
* breakpoint.c (check_fast_tracepoint_sals): Use std::string.
* arch-utils.h (default_fast_tracepoint_valid_at): Update.
* arch-utils.c (default_fast_tracepoint_valid_at): "msg" now a
std::string*.
Fix a commit 883fd55ab104 ("Record nested types") issue:
ERROR: tcl error sourcing .../gdb/testsuite/gdb.cp/nested-types.exp.
ERROR: can't read "actual_linejj": no such variable
while executing
"append txt " definition: $actual_linejj""
(procedure "cp_test_ptype_class" line 324)
invoked from within
"cp_test_ptype_class $name "ptype $name (limit = $limit)" $key $name $children" (procedure "test_nested_limit" line 28)
invoked from within
"test_nested_limit -1 false"
(file ".../gdb/testsuite/gdb.cp/nested-types.exp" line 310)
invoked from within
"source .../gdb/testsuite/gdb.cp/nested-types.exp"
("uplevel" body line 1)
invoked from within
"uplevel #0 source .../gdb/testsuite/gdb.cp/nested-types.exp"
invoked from within
"catch "uplevel #0 source $test_file_name""
testcase .../gdb/testsuite/gdb.cp/nested-types.exp completed in 9 seconds
caused by $actual_line having been accidentally referred to as
$actual_linejj in one place.
gdb/testsuite/
* lib/cp-support.exp (cp_test_ptype_class): Fix a typo in the
name of a variable: $actual_linejj -> $actual_line.
Simon Marchi [Fri, 23 Feb 2018 18:03:33 +0000 (13:03 -0500)]
dwarf: Make sect_offset 64-bits
Does anybody have an opinion about this? It would be nice to unbreak
the "default" build with clang (i.e. without passing special -Wno-error=
flags).
Here's a version rebased on today's master.
From 47d28075117fa2ddb93584ec50881e33777a85e5 Mon Sep 17 00:00:00 2001
From: Simon Marchi <simon.marchi@ericsson.com>
Date: Sat, 30 Dec 2017 22:48:18 -0500
Subject: [PATCH] dwarf: Make sect_offset 64-bits
Compiling with Clang 6 shows these errors:
/home/emaisin/src/binutils-gdb/gdb/dwarf2read.c:26610:43: error: result of comparison of constant 4294967296 with expression of type 'typename std::underlying_type<sect_offset>::type' (a
ka 'unsigned int') is always false [-Werror,-Wtautological-constant-out-of-range-compare]
if (to_underlying (per_cu.sect_off) >= (static_cast<uint64_t> (1) << 32))
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/emaisin/src/binutils-gdb/gdb/dwarf2read.c:26618:43: error: result of comparison of constant 4294967296 with expression of type 'typename std::underlying_type<sect_offset>::type' (a
ka 'unsigned int') is always false [-Werror,-Wtautological-constant-out-of-range-compare]
if (to_underlying (per_cu.sect_off) >= (static_cast<uint64_t> (1) << 32))
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The code in question checks if there is any offset exceeding 32 bits,
and therefore if we need to use the 64-bit DWARF format when writing the
.debug_names section. The type we use currently to represent section
offsets is an unsigned int (32-bits), which means a value of this type
will never exceed 32 bits, hence the errors above.
There are many signs that we want to support 64-bits DWARF (although I
haven't tested), such as:
- We correctly read initial length fields (read_initial_length)
- We take that into account when reading offsets (read_offset_1)
- The check_dwarf64_offsets function
However, I don't see how it can work if sect_offset is a 32-bits type.
Every time we record a section offset, we risk truncating the value.
And if a file uses the 64-bit DWARF format, it's most likely because
there are such offset values that overflow 32 bits.
Because of this, I think the way forward is to change sect_offset to be
a uint64_t. It will be able to represent any offset, regardless of the
bitness of the DWARF info.
Sriraman Tallam [Thu, 22 Feb 2018 21:56:46 +0000 (13:56 -0800)]
New plugin interface to get list of symbols wrapped with --wrap option.
2018-02-22 Sriraman Tallam <tmsriram@google.com>
* plugin.cc (get_wrap_symbols): New plugin interface.
(load): Add get_wrap_symbols to transfer vector.
* plugin-api.h (ld_plugin_get_wrap_symbols): New plugin interface.
* testsuite/plugin_test.c (onload): Call and check get_wrap_symbols
interface.
* testsuite/plugin_test_wrap_symbols.sh: New test script.
* testsuite/plugin_test_wrap_symbols_1.cc: New file.
* testsuite/plugin_test_wrap_symbols_2.cc: New file.
* testsuite/Makefile.am (plugin_test_wrap_symbols): New test.
* testsuite/Makefile.in: Regenerate.
Diagnose when trying to assemble conditional FP16 vmovx and vins
This patch makes GAS emit a warning when trying to assemble the Armv8.2
FP16 instructions VMOVX and VINS with condition codes. The Armv8-A
Reference Manual specifies these instructions without conditional codes
and says that if they are found in an IT block that they are CONSTRAINED
UNPREDICABLE.
gas/ChangeLog:
2018-02-22 Andre Vieira <andre.simoesdiasvieira@arm.com>
* config/tc-arm.c (do_neon_movhf): If conditional error out when in arm
mode and emit warning in thumb mode.
* testsuite/gas/arm/armv8-2-fp16-scalar-bad.s: Add new tests.
* testsuite/gas/arm/armv8-2-fp16-scalar-bad.l: Idem.
H.J. Lu [Thu, 22 Feb 2018 14:18:27 +0000 (06:18 -0800)]
x86: Add {rex} pseudo prefix
Add {rex} pseudo prefix to generate a REX byte for integer and legacy
vector instructions if possible. Note that this differs from the rex
prefix which generates REX prefix unconditionally.
gas/
* config/tc-i386.c (_i386_insn): Add rex_encoding.
(md_assemble): When i.rex_encoding is true, generate a REX byte
if possible.
(parse_insn): Set i.rex_encoding for {rex}.
* doc/c-i386.texi: Document {rex}.
* testsuite/gas/i386/x86-64-pseudos.s: Add {rex} tests.
* testsuite/gas/i386/x86-64-pseudos.d: Updated.
Tom Tromey [Fri, 16 Feb 2018 05:41:03 +0000 (22:41 -0700)]
Remove a cleanup from parse_expression_for_completion
This removes a cleanup from parse_expression_for_completion, by
changing various expression-completion functions to use
gdb::unique_xmalloc_ptry rather than explicit malloc+free.
Regression tested by the buildbot.
gdb/ChangeLog
2018-02-21 Tom Tromey <tom@tromey.com>
* value.h: (extract_field_op): Update.
* eval.c (extract_field_op): Return a const char *.
* expression.h (parse_expression_for_completion): Update.
* completer.c (complete_expression): Update.
(add_struct_fields): Make fieldname const.
* parse.c (expout_completion_name): Now a unique_xmalloc_ptr.
(mark_completion_tag, parse_exp_in_context_1): Update.
(parse_expression_for_completion): Change "name" to
unique_xmalloc_ptr*.
Yao Qi [Wed, 21 Feb 2018 11:20:03 +0000 (11:20 +0000)]
Remove regcache::m_readonly_p
Now, m_readonly_p is always false, so we can remove it, and regcache no
longer includes pseudo registers. Some regcache methods are lift up to
its parent class, like reg_buffer or detached_regcache.
gdb:
2018-02-21 Yao Qi <yao.qi@linaro.org>
* regcache.c (regcache::regcache): Update.
(regcache::invalidate): Move it to detached_regcache::invalidate.
(get_thread_arch_aspace_regcache): Update.
(regcache::raw_update): Update.
(regcache::cooked_read): Remove some code.
(regcache::cooked_read_value): Likewise.
(regcache::raw_write): Remove assert on m_readonly_p.
(regcache::raw_supply_integer): Move it to
detached_regcache::raw_supply_integer.
(regcache::raw_supply_zeroed): Likewise.
* regcache.h (detached_regcache) <raw_supply_integer>: New
declaration.
<raw_supply_zeroed, invalidate>: Likewise.
(regcache) <raw_supply_integer, raw_supply_zeroed>: Removed.
<invalidate>: Likewise.
<m_readonly_p>: Removed.
Yao Qi [Wed, 21 Feb 2018 11:20:03 +0000 (11:20 +0000)]
No longer create readonly regcache
Nowadays, we create a readonly regcache in get_return_value, and pass it
to gdbarch_return_value to get the return value. In theory, we can pass a
readable_regcache instance and get the return value, because we don't need
to modify the regcache. Unfortunately, gdbarch_return_value is designed
to multiplex regcache, according to READBUF and WRITEBUF.
# If READBUF is not NULL, extract the return value and save it in this
# buffer.
#
# If WRITEBUF is not NULL, it contains a return value which will be
# stored into the appropriate register.
In fact, gdbarch_return_value should be split to three functions, 1) only
return return_value_convention, 2) pass regcache_readonly and readbuf, 3)
pass regcache and writebuf. These changes are out of the scope of this
patch series, so I pass regcache to gdbarch_return_value even for read,
and trust each gdbarch backend doesn't modify regcache.
gdb:
2018-02-21 Yao Qi <yao.qi@linaro.org>
* infcmd.c (get_return_value): Let stop_regs point to
get_current_regcache.
* regcache.c (regcache::regcache): Remove.
(register_dump_reg_buffer): New class.
(regcache_print): Adjust.
* regcache.h (regcache): Remove constructors.
Yao Qi [Wed, 21 Feb 2018 11:20:03 +0000 (11:20 +0000)]
Replace regcache::dump with class register_dump
Nowadays, we need to dump registers contents from "readwrite" regcache and
"readonly" regcache,
if (target_has_registers)
get_current_regcache ()->dump (out, what_to_dump);
else
{
/* For the benefit of "maint print registers" & co when
debugging an executable, allow dumping a regcache even when
there is no thread selected / no registers. */
regcache dummy_regs (target_gdbarch ());
dummy_regs.dump (out, what_to_dump);
}
since we'll have two different types/classes for "readwrite" regcache and
"readonly" regcache, we have to move dump method to their parent class,
reg_buffer. However, the functionality of "dump" looks unnecessary to
reg_buffer (because some dump modes like regcache_dump_none,
regcache_dump_remote and regcache_dump_groups don't need reg_buffer at
all, they need gdbarch to do the dump), so I decide to move "dump" into a
separate classes, and each sub-class is about each mode of dump.
gdb:
2018-02-21 Yao Qi <yao.qi@linaro.org>
* regcache.c (class register_dump): New class.
(register_dump_regcache, register_dump_none): New class.
(register_dump_remote, register_dump_groups): New class.
(regcache_print): Update.
* regcache.h (regcache_dump_what): Move it to regcache.c.
(regcache) <dump>: Remove.
Yao Qi [Wed, 21 Feb 2018 11:20:03 +0000 (11:20 +0000)]
Class detached_regcache
jit.c uses the regcache in a slightly different way, the regcache dosen't
write through to target, but it has read and write methods. If I apply
regcache in record-full.c, it has the similar use pattern. This patch
adds a new class detached_regcache, a register buffer, but can be
read and written.
Since jit.c doesn't want to write registers through to target, it uses
regcache as a readonly regcache (because only readonly regcache
disconnects from the target), but it adds a hole in regcache
(raw_set_cached_value) in order to modify a readonly regcache. This patch
fixes this hole completely.
regcache inherits detached_regcache, and detached_regcache inherits
readable_regcache. The ideal design is that both detached_regcache and
readable_regcache inherit reg_buffer, and regcache inherit
detached_regcache and regcache_read (virtual inheritance). I concern
about the performance overhead of virtual inheritance, so I don't do it in
the patch.
gdb:
2018-02-21 Yao Qi <yao.qi@linaro.org>
* jit.c (struct jit_unwind_private) <regcache>: Change its type to
reg_buffer_rw *.
(jit_unwind_reg_set_impl): Call raw_supply.
(jit_frame_sniffer): Use reg_buffer_rw.
* record-full.c (record_full_core_regbuf): Change its type.
(record_full_core_open_1): Use reg_buffer_rw.
(record_full_close): Likewise.
(record_full_core_fetch_registers): Use regcache->raw_supply.
(record_full_core_store_registers): Likewise.
* regcache.c (regcache::get_register_status): Move it to
reg_buffer.
(regcache_raw_set_cached_value): Remove.
(regcache::raw_set_cached_value): Remove.
(regcache::raw_write): Call raw_supply.
(regcache::raw_supply): Move it to reg_buffer_rw.
* regcache.h (regcache_raw_set_cached_value): Remove.
(reg_buffer_rw): New class.
Yao Qi [Wed, 21 Feb 2018 11:20:03 +0000 (11:20 +0000)]
class readable_regcache and pass readable_regcache to gdbarch pseudo_register_read and pseudo_register_read_value
pseudo registers are either from raw registers or memory, so
gdbarch methods pseudo_register_read and pseudo_register_read_value
should have regcache object which only have read methods. In other
words, we should disallow writing to regcache in these two gdbarch
methods. In order to apply this restriction, this patch adds a new
class readable_regcache, derived from reg_buffer, and it only has
raw_read and cooked_read methods. regcache is derived from
readable_regcache. This patch also passes readable_regcache instead of
regcache to gdbarch methods pseudo_register_read and
pseudo_register_read_value.
This patch moves raw_read* and cooked_read* methods to readable_regcache,
which is straightforward. One thing not straightforward is that I split
regcache::xfer_part to readable_regcache::read_part and regcache::write_part,
because readable_regcache can only have methods to read.
readable_regcache is an abstract base class, and it has a pure virtual
function raw_update, because I don't want readable_regcache know where
these raw registers are from. They can be from either the target
(readwrite regcache) or the regcache itself (readonly regcache).
Yao Qi [Wed, 21 Feb 2018 11:20:02 +0000 (11:20 +0000)]
Class reg_buffer
This patch adds a new class reg_buffer, and regcache inherits it. Class
reg_buffer is a very simple class, which has the buffer for register
contents and status only. It doesn't have any methods to set contents and
status, and it is expected that its children classes can inherit it and
add different access methods.
Another reason I keep class reg_buffer so simple is that I think
reg_buffer can be even reused in other classes which need to record the
registers contents and status, like frame cache for example.
gdb:
2018-02-21 Yao Qi <yao.qi@linaro.org>
* regcache.c (regcache::regcache): Call reg_buffer ctor.
(regcache::arch): Move it to reg_buffer::arch.
(regcache::register_buffer): Likewise.
(regcache::assert_regnum): Likewise.
(regcache::num_raw_registers): Likewise.
* regcache.h (reg_buffer): New class.
(regcache): Inherit reg_buffer.