This commit is the result of the following actions:
- Running gdb/copyright.py to update all of the copyright headers to
include 2024,
- Manually updating a few files the copyright.py script told me to
update, these files had copyright headers embedded within the
file,
- Regenerating gdbsupport/Makefile.in to refresh it's copyright
date,
- Using grep to find other files that still mentioned 2023. If
these files were updated last year from 2022 to 2023 then I've
updated them this year to 2024.
I'm sure I've probably missed some dates. Feel free to fix them up as
you spot them.
This came up testing the CRC optimization work from Mariam@RAU.
Basically to optimize some CRC loops into table lookups or carryless
multiplies, we may need to do a bit reflection, which on the mcore
processor is done using a rotate instruction.
Unfortunately the simulator implementation of rotates has the exact same
problem as we saw with right shifts. The input value may have been sign
extended from 32 to 64 bits. When we rotate the extended value, we get
those sign extension bits and thus the wrong result.
The fix is the same. Rather than using a "long", use a uint32_t for the
type of the temporary. This fixes a handful of tests in the GCC testsuite:
If the value to be shifted has the sign bit set, the sign
bit would get copied into bits 32..63 of the temporary. Those
would then be right shifted into the final value giving an
incorrect final result.
This was observed with upcoming GCC improvements which eliminate
unnecessary extensions.
I was looking for cases where a GCC patch under evaluation would cause test
results to change. Quite surprisingly the mcore-elf port showed test
differences. After a fair amount of digging my conclusion was the sequences
before/after the patch should have been semantically the same.
Of course if the code is supposed to behave the same, then that points to
problems elsewhere (assembler, linker, simulator). Sure enough the mcore
simulator was mis-handling the sign extension instructions. The simulator
implementation of sextb is via paired shift-by-24 operations. Similarly the
simulator implements sexth via paired shift-by-16 operations.
The temporary holding the value was declared as a "long" thus this approach
worked fine for hosts with a 32 bit wide long and failed miserably for hosts
with a 64 bit wide long.
This patch makes the shift count automatically adjust based on the size of the
temporary. It includes a simple test for sextb and sexth. I have _not_ done a
full audit of the mcore simulator for more 32->64 bit issues.
This also fixes 443 execution tests in the GCC testsuite
The verbose argument has always been an int treated as a bool, so
convert it to an explicit bool. Further, update the API docs to
match the reality that the verbose value is actually used by some
of the internal modules.
Add explicit arch-specific modules.c rules to keep the build from
generating an incorrect common/modules.c. Otherwise the pattern
rules would cascade such that it'd look for $arch/modules.o which
turned into common/modules.c which triggered the gen rule.
My local testing of this code didn't catch this bug because of how
Automake manages .Po (dependency files) in incremental builds -- it
was adding extra rules that override the pattern rules which caused
the build to generate correct modules.c files. But when building
from a cold cache, the pattern rules would force common/modules.c to
be used leading to crashes at runtime.
This makes sure the arch-specific modules.c wildcard is matched and
not the common/%.c so that we compile it correctly. It also makes
sure each subdir has depdir logic enabled.
This simplifies the build logic and avoids an Automake bug where the
common_libcommon_a_OBJECTS variable isn't set in the arch libsim.a
DEPENDENCIES for targets that, alphabetically, come before "common".
We aren't affected by that bug with the current code, but as we move
things out of SIM_ALL_RECURSIVE_DEPS and rely on finer dependencies,
we will trip over it.
The objects are still compiled in the subdir, but the creation of the
archive itself is in the top-level. This is a required step before we
can move compilation itself up, and makes it easier to review.
The downside is that each object compile is a recursive make instead of
a single one. On my 4 core system, it adds ~100msec to the build per
port, so it's not great, but it shouldn't be a big deal. This will go
away of course once the top-level compiles objects.
This commit is the result of running the gdb/copyright.py script,
which automated the update of the copyright year range for all
source files managed by the GDB project to be updated to include
year 2023.
All the runtimes were only initializing a single CPU. When SMP is
enabled, things quickly crash as none of the other CPU structs are
setup. Change the default from 0 to the compile time value.
There's no need for these settings to be in sim-main.h which is shared
with common/ sim code, so move it all out to a new header which only
this port will include.
Not all arches include this in sim-main.h, and the ones that do don't
actually use bfd defines in the sim-main.h header. Prune it to make
sim-main.h simpler so we can kill it off entirely in the future.
We add the include to the files that utilize e.g. bfd_vma though.
This is a 32-bit architecture with 32-bit registers, so replace the
custom "word" long int typedef with an explicit int32_t. This is
a correctness fix since long will be 64-bits on most 64-bit hosts.
Automake will run each subdir individually before moving on to the next
one. This means that the linking phase, a single threaded process, will
not run in parallel with anything else. When we have to link ~32 ports,
that's 32 link steps that don't take advantage of parallel systems. On
my really old 4-core system, this cuts a multi-target build from ~60 sec
to ~30 sec. We eventually want to move all compile+link steps to this
common dir anyways, so might as well move linking now for a nice speedup.
We use noinst_PROGRAMS instead of bin_PROGRAMS because we're taking care
of the install ourselves rather than letting automake process it.
When reading/writing arbitrary data to the system's memory, the unsigned
char pointer type doesn't make that much sense. Switch it to void so we
align a bit with standard C library read/write functions, and to avoid
having to sprinkle casts everywhere.
This commit brings all the changes made by running gdb/copyright.py
as per GDB's Start of New Year Procedure.
For the avoidance of doubt, all changes in this commits were
performed by the script.
Use the new target-newlib-syscall module. This is needed to merge all
the architectures into a single build, and mcore has a custom syscall
table for its newlib/libgloss port.
We use the program argv to both find the program to run (argv[0]) and
to hold the arguments to the program. Most of the time this is fine,
but if we want to let programs specify argv[0] independently (which is
possible in standard *NIX programs), this double duty doesn't work.
So let's split the path to the program to run out into a separate
field by itself. This simplifies the various sim_open funcs too.
By itself, this code is more of a logical cleanup than something that
is super useful. But it will open up customization of argv[0] in a
follow up commit. Split the changes to make it easier to review.
Now that ChangeLog entries are no longer used for sim patches,
this commit renames all relevant sim ChangeLog to ChangeLog-2021,
similar to what we would do in the context of the "Start of New
Year" procedure.
The purpose of this change is to avoid people merging ChangeLog
entries by mistake when applying existing commits that they are
currently working on.
Also throw in a .gitignore entry to keep people from adding new
ChangeLog files anywhere in the sim tree.
These ports only use the pieces that have been unified, so we can
merge them into the common configure script and get rid of their
unique one entirely.
We still compile & link separate run programs, and have dedicated
subdir Makefiles, but the configure script portion is merged.
The sim-hardware configure option allows builders to select a set of
device models to enable. But this seems like unnecessary overkill:
the existence of individual device models doesn't affect performance
at all as they are only enabled at runtime if the config uses them,
and individually these are all <5KB a piece. Stripping off a total
of ~50KB from a ~1MB binary doesn't seem useful, and it's extremely
unlikely anyone will ever bother.
So let's simplify the configure/make logic by turning sim-hardware
into a boolean option like many of the other sim options. Any ports
that have unique device models will declare them in their Makefile
instead of at configure time. This will allow us to (eventually)
unify the setting into the common dir.
Move these options up to the common dir so we only test & export
them once across all ports. This takes a page from the cgen maint
logic to make $(MAINT) work for non-automake Makefiles which will
allow us to merge it together.
Move these options up to the common dir so we only test & export
them once across all ports. It also enables -Werror usage on the
common files we've been pulling out of arch subdirs.
For the ports that still don't build with -Werror, rather than disable
the flag at configure time, do it at make time. This will allow us to
unify these tests in the common sim configure script.
The sim-basics.h is too big and includes too many things. This leads
to some arch's sim-main.h having circular loop issues with defs, and
makes it hard to separate out common objects from arch-specific defs.
By splitting up sim-basics.h and killing off sim-main.h, it'll make
it easier to separate out the two.
The m4 macro has 2 args: the "wire" settings (which represents the
hardwired port behavior), and the default settings (which are used
if nothing else is specified). If none are specified, the arch is
expected to support both, and the value will be probed based on the
user runtime options or the input program.
Only two arches today set the default value (bpf & mips). We can
probably let this go as it only shows up in one scenario: the sim
is invoked, but with no inputs, and no user endian selection. This
means bpf will not behave like the other arches: an error is shown
and forces the user to make a choice. If an input program is used
though, we'll still switch the default to that. This allows us to
remove the WITH_DEFAULT_TARGET_BYTE_ORDER setting.
For the ports that set a "wire" endian, move it to the runtime init
of the respective sim_open calls. This allows us to change the
WITH_TARGET_BYTE_ORDER to purely a user-selected configure setting
if they want to force a specific endianness.
With all the endian logic moved to runtime selection, we can move
the configure call up to the common dir so we only process it once
across all ports.
The ppc arch was picking the wire endian based on the target used,
but since we weren't doing that for other biendian arches, we can
let this go too. We'll rely on the input selecting the endian, or
make the user decide.
All of the settings in here are handled by the common top-level
config.h, so drop the individual arch-config.h files entirely.
This will also help guarantee that we don't add any new arch
specific defines that would affect common code which will help
with the effort of unifying them.
