And in order to avoid a double dereference in the dispatcher, directly store
the normalEntry in the map.
The index to the block map becomes ((((DR<<1) | IR) << 30) | (address >> 2)).
This has been chosen since the msr bits change less often than the address,
thus we keep nearby entries together.
Also do not call the C dispatcher in case the assembly dispatcher didn't
find a block, since it wouldn't find a block either due to the 1:1 mapping,
except when falling back to the non shm segment lookup table.
This bug has been lurking in the code ever since I added the discard
functionality. It doesn't seem to be triggered all that often,
and on top of that the emitted code only runs conditionally, so I'm not
sure if people have been affected by this bug in practice or not.
We need to check for the address of the *previous* instruction, because
checking fifoWriteAddresses happens not at the end of the instruction
that triggered it but at the start of the next instruction.
Whenever JitBaseBlockCache::Clear() got called, it threw away the memory mapping for the fast block map and created a new one. This new mapping typically got mapped at the same address at the old one, but this is not guaranteed. The pointer to the mapping gets embedded in the generated dispatcher code in Jit64AsmRoutineManager::Generate(), which is only called once on game boot, so if the new mapping ended up at a different address than the old one, the pointer in the ASM pointed at garbage, leading to a crash.
This fixes the issue by guaranteeing that the new mapping is mapped at the same address.
While both fastmem and the BLR optimization depend on fault handling,
the BLR optimization doesn't depend on fastmem, and there are cases
where you might want the BLR optimization but not fastmem. For me
personally, it's useful when I try to use a debugger on Android and have
to disable fastmem so I don't get SIGSEGVs all the time, but it would be
especially useful for iOS users.
This fixes a problem where changing the JIT debug settings on
Android while a game was running wouldn't cause the changed settings
to apply to code blocks that already had been compiled.
Modify PPCSTATE_OFF and PPCSTATE_OFF_ARRAY macros when using GCC to
avoid useless log spam. Specifically, use a consteval lambda with gcc
_Pragma statements to disable the -Winvalid-offsetof warning inside the
macros.
Each successful build (and many failing ones) on the Android buildbot
generates almost 300 cases of -Winvalid-offsetof, resulting in thousands
of lines of log spam per build. In addition to bloating the log filesize
these spurious warnings make it harder to find actual warnings.
These warnings are generated by calls to the macros PPCSTATE_OFF and
PPCSTATE_OFF_ARRAY, which in turn are used by many other macros used by
the JIT. The ultimate cause is that offsetof is only conditionally
supported on non-standard-layout types, which includes the PowerPCState
struct.
To address potential questions of whether there's a better way to handle
this:
The obvious solution would be to modify PowerPCState so that it does
have a standard layout. This is unfortunately impractical.
To have a standard layout a type can only contain other types with
standard layouts. None of the stl containers are guaranteed to have
standard layouts, and PowerPCState contains a std::tuple and std::array.
PowerPCState also contains a PowerPC::Cache and InstructionCache which
themselves contain std:arrays and std::vectors.
Furthermore InstructionCache derives from Cache, and a derived class can
only have standard layout if at most one class in its hierarchy has a
non-static data member, but both classes have such members. Making
InstructionCache have a standard layout would require duplicating all
the functionality of Cache so it no longer derived from it, as well as
replacing the stl containers. This might require having a raw pointer to
said containers, with the manual memory management that implies.
All of that would be much more disruptive than would be justified to get
rid of some warnings (however annoying they might be). This is
compounded by the fact that PowerPCState hasn't had a standard layout
for a long time, if ever, and if the PPCSTATE_OFF macros weren't working
reliably it would have become obvious a long time ago.
As to why I picked the lambda solution over other potential changes:
- Keeping the define as-is and wrapping some gcc #pragmas around it
doesn't work because the pragmas don't get included when the define is
substituted to the call site.
- Keeping the define as a non-lambda expression and using inline
_Pragma() statements would ideally be better and works fine for msvc,
but fails for GCC with "'#pragma' is not allowed here".
- Turning off -Winvalid-offsetof globally for gcc would work, but there
might be other contexts where offsetof is problematic and GCC seems to
be the only compiler warning about it.
In theory, our config system supports calling Set from any thread. But
because we have config callbacks that call RunAsCPUThread, it's a lot
more restricted in practice. Calling Set from any thread other than the
host thread or the CPU thread is formally thread unsafe, and calling Set
on the host thread while the CPU thread is showing a panic alert causes
a deadlock. This is especially a problem because 04072f0 made the
"Ignore for this session" button in panic alerts call Set.
Because so many of our config callbacks want their code to run on the
CPU thread, I thought it would make sense to have a centralized way to
move execution to the CPU thread for config callbacks. To solve the
deadlock problem, this new way is non-blocking. This means that threads
other than the CPU thread might continue executing before the CPU thread
is informed of the new config, but I don't think there's any problem
with that.
Intends to fix https://bugs.dolphin-emu.org/issues/13108.
Should fix https://bugs.dolphin-emu.org/issues/13178.
Jit64 doesn't need to do this because it stores PC directly into
ppcState instead of first storing it in a register.
Now block link nearcode is back to a length of three instructions.
Unfortunately, the code I'm adding to Jit.cpp ends up being a bit messy
because we need to handle the case of already being in farcode...
Jumping between linked blocks currently works as follows: First, at the
end of the first block, we check if the downcount is greater than zero.
If it is, we jump to the `normalEntry` of the block. So far so good. But
if the downcount wasn't greater than zero, we jump to the `checkedEntry`
of the block, which checks the downcount *again* and then jumps to
`do_timing` if it's less than zero (which seems like an off by one error
- Jit64 doesn't do anything like this). This second check is rather
redundant. Let's jump to `do_timing` where we previously jumped to
`checkedEntry`.
Jit64 doesn't check the downcount on block entry. See 5236dc3.
We want to have a low number of instructions between the LDP and the
MADD so that the MADD can start immediately after the LDP finishes
even if we're on a lower-end in-order CPU.
Replace the bool parameter force5bytes in J, JMP, and J_CC with an enum
class Jump::Short/Near. Many callers set that parameter to the literal
'true', which was unclear if you didn't already know what it did.
The base DebugInterface now depends on the Core's CPUThreadGuard, and
utilities in Common shouldn't be depending on Core facilities. So, we
can move this into the core library instead.
We can utilize the fact that if swapping two instructions causes another
swap to become possible, that swap involves one of the two instructions
we just swapped. There's need to search for new swap opportunities as
far back as the beginning of the block; we can just go one step back.