The global current_state handle to the current simulator state is a
design idea that was half implemented, but never really cleaned up.
The point was to have a global variable pointing to the state so that
funcs could more quickly & easily access the state anywhere. We've
instead moved in the direction of passing state around everywhere and
don't have any intention of moving back.
I also can't find any references to gdb using this variable, or to
cgen related "dump_regs" functions, both of which were used in the
comments related to this code.
When connecting to the simulator in gdb, we don't want it to exit on
us when we pass down unknown/invalid/help/etc... options. Plumb down
the kind argument so we can handle both gdb & psim interfaces.
gdb/ChangeLog:
* top.c (print_gdb_version): Change copyright year in version
message.
gdb/gdbserver/ChangeLog:
* gdbreplay.c (gdbreplay_version): Change copyright year in
version message.
* server.c (gdbserver_version): Likewise.
Per GDB the "Start of New Year Procedure", this patch
- renames the current ChangeLog into ChangeLog-2015;
- starts a new ChangeLog file.
gdb/ChangeLog:
* config/djgpp/fnchange.lst: Add entry for gdb/ChangeLog-2015.
In this GCC commit:
https://gcc.gnu.org/ml/gcc-patches/2015-12/msg00735.html
GCC started emitting dmb instructions with no operands. The intention
was that dmb with no operands should be an alias for 'dmb 0'.
The following patch extends the arc opcodes library to support dmb with
no operands.
opcodes/ChangeLog:
* arc-tbl.h (dmb): Add a no operand version of dmb.
We can leverage the cpu->regs array rather than going through the
function helpers to get nice compact code.
Further, fix up the return values: return -1 when we can't find a
register (and let the caller write out warnings), return 2/4 when
we actually write out that amount, and handle the zero reg.
This array isn't used anywhere, and the init phase actually corrupts
some memory because the array has 18 elements but tries to set the
19th (ZERO) position.
This change tracks the "closed" state of file descriptors 0, 1, and 2,
introducing the function fdbad() to emul_netbsd.c and emul_unix.c.
Note that a function of the same name and purpose exists in
sim/common/callback.c.
This patch eliminates all of the "unresolved testcases" when testing
GDB against the powerpc simulator.
This occurs because the powerpc simulator closes, on behalf of the
testcase, the file descriptors associated with stdin, stdout, and
stderr. GDB still needs these descriptors to communicate with the
user or, in this case, with the testing framework.
Tested on x86_64-pc-linux-gnu native-gdbserver, no new regressions.
gdb/gdbserver/ChangeLog:
* server.c (crc32_table): Delete.
(crc32): Use libiberty's xcrc32 function.
The disasm framework reserves the private_data field for the disassemblers
themselves, not for people who use the disassembler. Instead, there is an
application_data field for callers such as the sim. Switch to it to avoid
random corruption/crashes when the disassemblers use private_data.
Pretty much all targets are using this module already, so add it to the
common list of objects. The only oddball out here is cris and that's
because it supports loading via an offset for all the phdrs. We drop
support for that.
No arch is using this anymore, and we want all new ports using the
hardware framework instead. Punt WITH_DEVICES and the two callbacks
device_io_{read,write}_buffer.
We can also punt the tconfig.h file as no port is using it anymore.
This fixes in-tree builds that get confused by picking up the wrong
one (common/ vs <port>/) caused by commit ae7d0cac8c.
Any port that needs to set up a global define can use their own
sim-main.h file that they must provide regardless.
The bfin port is using the WITH_DEVICES framework for two reasons:
- get access to the cpu making the request (if available)
- check the alignment & size for core & system MMRs
We addressed the first part with commit dea10706e9,
and we handle the second part with this commit. Arguably this is more
correct too because trying to do bad reads/writes directly (when devices
support is disabled) often results in bad memory accesses.
As part of this clean up, we also adjust all of the existing logic that
would reject invalid accesses: the code was relying on the checks never
returning, but that's not the case when things like gdb (via the user's
commands) are making the requests. Thus we'd still end up with bad mem
accesses, or sometimes gdb being hung due to while(1) loops.
Now we can connect (most of) these models into any address and have them
work correctly.
We set up an array of 3 elements and then index into it with a 2bit
value. We check the range before we actually use the pointer, but
the indexing is enough to make asan upset, so just stuff a fourth
value in there to keep things simple.
The bfin port has been using the device callback largely so it could be
passed the cpu when available. Add this logic to the common core code
so all ports get access to the active cpu.
The semantics of these buffer functions are changed slightly in that
errors halt the engine synchronously rather than returning the length
to the caller. We'll probably adjust this in a follow up commit.
The bfin code isn't updated just yet as it has a bit more logic in the
device layer that needs to be unwound at which point we can delete it
entirely.
The only unique thing about mip's sim_{read,write} helpers is the call to
address_translation on the incoming address. When we look closer at that
function though, we see it's just a stub that maps physical to virtual,
and the cache/return values are hardcoded. If we delete this function,
we can then collapse all the callers and drop the custom sim_{read,write}
logic entirely.
Some day we might want to add MMU support, but when we do, we'll want to
have the common layers handle things so all targets benefit.
The point of passing down the cpu to core reads/writes is to signal which
cpu is making the access. For system accesses (such as internal memory
initialization), passing the cpu down doesn't make sense, and in the case
of early init like cris, can cause crashes. Since the cpu isn't fully set
up at this point, if the core code tries to access some fields (like the
PC reg), it'll crash. While cris shouldn't be doing this setup here (it
should be in the inferior stage), we'll deal with that later.
For targets that process argv in sim_create_inferior, improve the code:
- provide more details in the comment
- make the check for when to re-init more robust
- clean out legacy sim_copy_argv code
This will be cleaned up more in the future when we have a common inferior
creation function, but at least help new ports get it right until then.
2015-12-24 Thomas Preud'homme <thomas.preudhomme@arm.com>
ld/testsuite/
* ld-arm/arm-elf.exp: Run new test "Thumb-Thumb farcall v6-M (no
profile)".
* ld-arm/farcall-thumb-thumb-m-no-profile-a.s: New file.
* ld-arm/farcall-thumb-thumb-m-no-profile-b.s: Likewise.
* ld-arm/farcall-thumb-thumb-m-no-profile.d: Likewise.
2015-12-25 Sandra Loosemore <sandra@codesourcery.com>
gdb/testsuite/
* lib/gdb.exp (gdb_test): Update comments to clarify that the
PATTERN argument is optional.
The frv port used the device logic to support a single cache address,
and the comments around that are "these were merely copied from a diff
port and are unused", plus the code to attach the memory is "#if 0".
Just punt it all.
The m32r port was using the device framework to handle two devices: the
cache and uart registers. Both can be implemented in the newer hardware
framework instead which allows us to drop the device logic entirely, as
well as delete the tconfig.h file.
While creating the new uart device model, I also added support for using
stdin to read/write data rather than only supporting sockets.
This has been lightly tested as there doesn't appear to be test coverage
for the code already. If anyone still cares about this port, then they
should (hopefully) file bug reports.
The cris port was using the device framework to handle two addresses when
the --cris-900000xx flag was specified. That can be implemented using the
newer hardware framework instead which allows us to drop the device logic
entirely, as well as delete the tconfig.h file. Basically we create a new
cris_900000xx device model and move the read logic out of devices.c and
into that. The rest of the devices logic was callback to the hardware
framework already.
Much like we autodetect the path to the run program when there is none
set explicitly, do the same for the rvdummy program. Otherwise the
default make check fails to execute the helper properly.
These trace calls don't seem to add anything useful and break the cris
hw tests, so punt them. They were disabled before commit 6d519a4606
but were re-enabled as part of TRACE macro cleanups.