Currently all ports have to declare sim_cpu themselves in their
sim-main.h and then embed the common sim_cpu_base in it. This
dynamic makes it impossible to share common object code among
multiple ports because the core data structure is always different.
Let's invert this relationship: common code declares sim_cpu, and
the port uses the new arch_data field for its per-cpu state.
This is the first in a series of changes: it adds a define to select
between the old & new layouts, then converts all the ports that don't
need custom state over to the new layout. This includes mn10300 that,
while it defines custom fields in its cpu struct, never uses them.
These headers define the register numbers for each port to implement
the sim_fetch_register & sim_store_register interfaces. While gdb
uses these, the APIs are part of the sim, not gdb. Move the headers
out of the gdb/ include namespace and into sim/ instead.
Simplify the build by moving the generation of these files from
build-time (via dgen.c that we have to compile & execute on the
build system) to maintainer/release mode (via spreg-gen.py that
we only ever execute when the spreg table actually changes). It
speeds up the build process and makes it easier for us to reason
about & review changes to the code generator.
The tool is renamed from "dgen" because it's hardcoded to only
generated spreg files. It isn't a generalized tool for creating
lookup tables.
The core device has an attach address method as the root of the tree
which calls out to the sim API. But it doesn't have a corresponding
detach method which means we just crash if anything tries to detach
itself from the core. In practice, the m68hc11 is the only model
that actually tries to detach itself on the fly, so no one noticed
earlier.
With this in place, we can delete the existing detach code from the
m68hc11 model since it defaults to "passthru" callback which will in
turn call the dv-core detach, and they have the same behavior -- call
the sim core API to detach from the address space.
Bug: https://sourceware.org/PR25211
If the counter for LOOP instruction is provided by a register with
value zero, then the instruction must cause a PC jump directly to the
loop end. But in that particular case simulator must not initialize
its internal loop variables, because loop body will not be executed.
Instead, simulator must obtain the loop's end address directly from
the LOOP instruction.
Signed-off-by: Dimitar Dimitrov <dimitar@dinux.eu>
The parser for boolean rules fails to skip over the , separator in
the options which makes it hang forever. No dc files in the tree
use boolean rules atm which is why no one noticed.
To make it clear this is an input to the igen tool, rename it with an
igen extension. This matches the other files in the ppc dir (altivec
& e500 igen files), and the other igen ports (mips, mn10300, v850).
Nothing passes this to dgen, and even if it did, nothing would happen
because the generated spreg.[ch] files don't include any references
back to the original data table. So drop it to simplify.
Since we know we'll return 0 by default, we don't have to output case
statements for readonly or length fields whose values are also zero.
This is the most common case by far and thus generates a much smaller
switch table in the end.
Instead of writing:
case 1:
return 1;
case 2:
return 1;
...etc...
Output a single return so we get:
case 1:
case 2:
case ...
return 1;
This saves ~100 lines of code. Hopefully the compiler was already
smart enough to optimize to the same code, but if not, this probably
helps there too :).
This saves a single line for the same result. By itself, it's not
interesting, but we can further optimize the generated output and
completely omit the switch table in some cases. Which we'll do in
follow up commits.
When merging ppc configure checks into the top-level, these 2 funcs
were accidentally dropped (probably due to incorrect resolution of
conflicts). Restore them since the ppc code utilizes them both.
This controls only one thing: how to call mkdir(). The gnulib code
already has a mkdir module that provides this exact logic for us, so
punt the code entirely.
This macro expansion was missing a set of outer-most parenthesis which
some compilers would complain about depending on how the macro is used.
This is just standard good macro hygiene too.
We've never run these helper programs directly. The igen program
includes the relevant source files directly and runs the code that
way. So stop wasting developer CPU time linking programs that are
never run. We leave the rules in place for people who need to test
and debug the specific bits of code every now & then.
The intention of this code seems to be to indicate that this insn
should not be used and produces undefined behavior, so instead of
setting registers to bogus values, call Unpredictable. This fixes
build warnings due to 32-bit/64-bit type conversions, and outputs
a log message for users at runtime instead of silent corruption.
Bug: https://sourceware.org/PR29276
We've been using this only to set the default word size to 32. We
can easily move this into the makefile via a -D compiler flag and
clean up the build logic quite a bit.
We've been using this only to set the default word size to 32. We
can easily move this into the makefile via a -D compiler flag and
clean up the build logic quite a bit.
We've been using this only to set the default word size to 32. We
can easily move this into the makefile via a -D compiler flag and
clean up the build logic quite a bit.
We've been using this only to set the default word size to 64. We
can easily move this into the makefile via a -D compiler flag and
clean up the build logic quite a bit.
We've been using this only to set the default word size to 32-vs-64
based on the $target. We can easily merge this with the top-level
configure script to clean things up a bit.
The install code was using $SUBDIRS to track all enabled arches. This
works, but isn't great if we want to add a subdir that isn't an arch
port, or as we merge the subdirs into the top-level. Create a new var
explicitly to track the list of enabled arches instead.
This was needed when the install step was run in subdirs, but now
that we process that entirely in the top-level, we don't need to
pass this down, so drop it.
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.
We still have to maintain custom install rules due to how we rename
arch-specific files with an arch prefix in their name, but we can at
least unify the logic in the common dir.
