This is mostly useful for removing SHLs by constant 1, 2, 3, which this
commit implements in one place.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@7652 8ced0084-cf51-0410-be5f-012b33b47a6e
This makes sure they are below 2G on 64bit systems,
allowing us to continue using the simpler 32bit accesses.
Fixes issue 4357 and issue 4397.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@7465 8ced0084-cf51-0410-be5f-012b33b47a6e
In assembly, these are 'ld $ax0.d,$ax1.r,@$arS with their n,m and nm variants,
which have been special cased for S==3. The regular 'ld can be decomposed
into lrri $ax0.d,@$arS and lrri $ax1.r,@$ar3, while the S==3 case decomposes
to lrri $axR.h,@$arD and lrri $axR.l,@$ar3. The latter variant will be
disassembled to 'ldax $axR,@$arD after this change. The assembler recognizes
both the new 'ldax variant and the old 'ld with @$ar3 but the disassembler
only outputs 'ldax. Besides the readability, this allows for more correct
register use analysis(when it's done).
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@7413 8ced0084-cf51-0410-be5f-012b33b47a6e
Reading the instruction from memory at the given address and using that
as index in the opTable leads to more correct results than using the
address as index into opTable.
Also assert when trying to execute a bad instruction instead of crashing
without warning.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@7384 8ced0084-cf51-0410-be5f-012b33b47a6e
A repeat count of zero still executed the instruction once. Fixes Zelda Ucode
audio crackling.
Fixes DSPLLE audio in: Luigi Mansion, Pikmin, Super Mario Galaxy,
Zelda CE, Zelda Four Swords, Zelda WW, Zelda TP.
(Thanks LM for testing)
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@7370 8ced0084-cf51-0410-be5f-012b33b47a6e
Uses the last sample from the ppc buffer to fill the samples the ppc
didn't deliver data for, avoids clicking on underruns.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@7338 8ced0084-cf51-0410-be5f-012b33b47a6e
No speed difference, probably because our basic blocks are
too small to contain more than one access to any register
or used too seldom for multiple accesses to make a difference.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@7260 8ced0084-cf51-0410-be5f-012b33b47a6e
Was made possible by the move of Core/DSPCommon to Core/Core.
This should not give a significant speed difference, but frees
another register for register allocator use on x64 and
generally simplifies the code.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@7252 8ced0084-cf51-0410-be5f-012b33b47a6e
In addition protect against their use on 32 bit and the use of [ABCD]H
together with a REX prefix on 64 bit.
This assumes that the customOp parameter of WriteREX and operandReg of
OpArg always are registers, and thus needs to give something valid to
WriteREX when that is not the case (WriteShift).
In addition to the patch i sent to the ML, there are a few changes to the
error reporting(mostly whitespace).
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@7202 8ced0084-cf51-0410-be5f-012b33b47a6e
so VideoOGL always is in the same state when starting a guest program.
Also constify the RasterFont, while we are at it.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@7166 8ced0084-cf51-0410-be5f-012b33b47a6e
Looks like compilers tend to use EBX for parameters if not told otherwise and
don't bother to update SP in leaf functions, so PUSH/POP kill local variables.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@7151 8ced0084-cf51-0410-be5f-012b33b47a6e
stacks), addition of a register cache(currently accumulators on 64bit only).
It feels a bit slower, but the numbers are about the same.
Caching accumulators doesn't change speed either.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6848 8ced0084-cf51-0410-be5f-012b33b47a6e
reads the high bits from the mailbox registers. It is probably waiting for
the dsp to read the data from the cpu-to-dsp mailbox or for the dsp to
write to the dsp-to-cpu mailbox.
This about removes DSP::Read16 from lle profiles where it previously used
up to 2% of all system time. Also speeds up games quiet a bit.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6719 8ced0084-cf51-0410-be5f-012b33b47a6e
functions accidentally added. Fixed the jitted ar register arithmetic.
Added a CMakeList.txt for the UnitTests, but did not add the subdirectory
to Source/CMakeLists.txt.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6687 8ced0084-cf51-0410-be5f-012b33b47a6e
(acc and ax) and product register with one read/write.
Gives a minuscule speedup of not more than 4%. In exchange, breaks all
your out-of-tree changes to dsp. Tests are not building again, yet.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6680 8ced0084-cf51-0410-be5f-012b33b47a6e
assembler for JIT. Replace JIT ToMask() with a different variant. Remove
superfluous zeroWriteBackLog calls(added by me).
Core/Common: Don't bother creating a string and calling into a Logs trigger()
when there is noone listening. Change AtomicLoadAcquire for gcc to just
make the compiler not reorder memory accesses around it instead of doing
a full memory barrier, per the comment in the win32 variant.
Core/AudioCommon: Fix a use of uninitialized variable inside libalsa.
Microbenchmarking results for ToMask variants:(1 000 000 000 iterations):
cpu\variant| shifts | bit scan
intel mobile [email protected] | 5.5s | 4.0s
amd athlon64x2@3GHz | 6.1s | 6.4s
(including some constant overhead identical to both variants)
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6667 8ced0084-cf51-0410-be5f-012b33b47a6e
ESI is not available as callee save register on linux 64, so use R12, which
is available on windows 64, too. The only callee save register still available on
all platforms is EBP, but i guess we don't want to fiddle with that one.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6492 8ced0084-cf51-0410-be5f-012b33b47a6e
Also add the mask for ADPCM predictor, like r6480 does.
Tested 64 bit linux/windows. At least compiles for 32 bit windows, but
with these binaries most of my guest software fails with
DSPLLE(interpreted and jit) when it gets to do actual work.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6484 8ced0084-cf51-0410-be5f-012b33b47a6e
change the unit test to catch this bug if it turns up again and two more
zeroWriteBackLog() calls.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6356 8ced0084-cf51-0410-be5f-012b33b47a6e
* Make writeToBackLog private to DSPIntExtOps.cpp
(JIT variants of 'l and 'ln are disabled and broken as is)
* Make zeroing of the backlog conditional on doing an interpreter fallback and
do it at a few more places
* Fix selection of cleanup for extended opcodes.
* Fix the DSP unit tests to correctly emit the function prolog/epilog
(else EBX wouldn't be saved)
* Add a few more DSP unit tests
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6325 8ced0084-cf51-0410-be5f-012b33b47a6e
call it for every instruction, like in DSPInterpreter::Step. Fix the analysis lookups
for DSP_IDLE_SKIP and DSP_LOOP_END.
After this patch, the block_size and the return value from the block always match,
except when an internal exception is detected. This will change in a later patch,
so we actually get some benefits from checking if we need to call HandleLoop().
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6280 8ced0084-cf51-0410-be5f-012b33b47a6e
It is pure luck that we did get a fresh (thus zeroed) memory area when not putting DSPLLE on thread. ClearIRAM() is supposed to only clear the non-static part.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6278 8ced0084-cf51-0410-be5f-012b33b47a6e
The block_size seems to be not reliable, even after trying to more closely match
the interpreter by looking at the analysis for (addr+opcode->size-1) for detecting
the end of the block.
Since we need to "calculate" this number shortly before returning to
RunForCycles, it seemed logical to use the (up to now) unused return value
of the blocks.
Improves SMG2 here.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6266 8ced0084-cf51-0410-be5f-012b33b47a6e
IsSameMemArea check in the rest of the file.
Looks like this is simply a copy-paste error, that i dutifully "fixed" for 64 bits.
No idea if this fixes anything, but it brings JIT more in line with the interpreter.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6263 8ced0084-cf51-0410-be5f-012b33b47a6e
pointer cannot be reached by 32bit displacement.
RAX is caller save and non-parameter in all 64 bit ABIs, so it should be safe
to clobber it on function call.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6248 8ced0084-cf51-0410-be5f-012b33b47a6e
This is the safe variant that does not assume other
functions use R11 for &g_dsp.r, which they actually do.
So there is a lot unnecessary reloading of the register
getting emitted.
There are a few small changes exploiting pointers known
at emission time and making use of MComplex math.
Also renames m_LLEplaying to m_AIplaying in Core/AudioCommon.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6245 8ced0084-cf51-0410-be5f-012b33b47a6e
This should eliminate the crackling with alsa and pulseaudio backends and replace it
with much nicer pauses. This is only interesting for audio backends that do not
respect Mixer::GetNumSamples() and should not impact users able to run the DSPLLE at
full speed.
The cost of this is an added LLE audio latency of about 0.06 s in the continuous
playback case. If that is too much, lower the low watermark.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6239 8ced0084-cf51-0410-be5f-012b33b47a6e
If you see this alert, dolphin will probably crash in a few moments.
Found in the course of fiddling with DSPLLEs JIT, which puts its data
structures and static code somewhere above 0x7fff00000000 on my machine, making
it unreachable via 32bit displacements from the created code at ~ 0x40000000.
Fixed all the offending places in DSPLLEs JIT to emit register indirect accesses,
only to find out that the generated code is slower than the interpreter and does
work just as good(or bad). Oh well, back to DSPHLE it is then.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6229 8ced0084-cf51-0410-be5f-012b33b47a6e
final rendering of (at least) Sonic Unleashed.
This should be (nearly) invisible in Z16 depth copies and in games, but there is
some chance that it collides with other such border cases. There probably is some
room to decrease the adjustment but 9999999/10000000 is not enough. A static offset
may be an option, too. I don't know if the game can/does set something like that.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6225 8ced0084-cf51-0410-be5f-012b33b47a6e
eating each others events.
Manifests itself in GPU thread waiting forever for a reply in GL library code(and thus
a frozen emulation). Most of the time, this can be resolved by creating more events
(clicking, changing focus), but sometimes it stays stuck.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6218 8ced0084-cf51-0410-be5f-012b33b47a6e
The theoretical result is slightly different to the original shader because the
final adjustment is to the range [0/255..255/255] instead of
[0/16777215..16777215/16777215].
The real result is vastly different on some gpus that were giving incorrect results
of bits 23..16 (y-component) wrapping around while bits 31..24 (z-component)
stayed the same, and bits 31..24 changing while in the middle of the value range
for bits 23..16 for large depth values.
This should fix issue #3123.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6217 8ced0084-cf51-0410-be5f-012b33b47a6e
I am not sure i am understanding what the pipeline really does, and more so what
the GC/WII expects here. If my comments are incorrect, please let me know.
This was tested with MP2:E, ZWW on r6075.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6078 8ced0084-cf51-0410-be5f-012b33b47a6e
(those originating from a thread other than the gpu thread)
glxMakeCurrent is supposed to error out when the context is
still current in another thread. NVIDIA handles that, but mesa3d errors.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6054 8ced0084-cf51-0410-be5f-012b33b47a6e
All visuals with parameter GLX_SAMPLE_BUFFERS_ARB = 0 have
GLX_SAMPLES_ARB = 0, so the old code would only ever choose
visuals with GLX_SAMPLE_BUFFERS_ARB>=1 and GLX_SAMPLES_ARB>=1.
Some drivers do not export any multisampling capable visuals,
so dolphin fails to find a matching visual and falls back to
the single buffered visual(which does not work for some yet
unknown reason). Fixes issue #2647.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@5565 8ced0084-cf51-0410-be5f-012b33b47a6e
This should give the exact same shader code except for the cases of
SELA == SELB and SELC == 1, the latter only for the color combinder.
In these two cases the results of the shader should not change in any way.
It would have been possible to optimize the SELC == HALF case too, if
the actual factor had been 0.5f instead of 0.4980392f.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@5472 8ced0084-cf51-0410-be5f-012b33b47a6e
strcasecmp is provided by Source/Core/Common/Src/CommonFuncs.h
and vswprintf is overloaded in C++ to take the same parameters as
vsnwprintf, according to MSDN.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@5470 8ced0084-cf51-0410-be5f-012b33b47a6e
The real hardware probably always sends the result of its computations
to the alpha test/fog units, in parallel to its registers, but those
only take the result of the last texenv stage.
This makes grass show up in ZTP, where the last texenv stage oddly writes to
reg #3 instead of #0.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@5462 8ced0084-cf51-0410-be5f-012b33b47a6e
This removes some bad behaviour, like the audio thread busy waiting inside ALSA code,
missing begins of voice samples(due to DSP_HLE code not getting called often enough).
Turns out my understanding of ALSA API was lacking.
Tested with pulse plugin and directly on my intel-hda/realtek hardware.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@5461 8ced0084-cf51-0410-be5f-012b33b47a6e
The fix is asking alsa for a prefered "hardware" buffer size, so alsa does not fall
back to pulseaudios offering of ~1 minute. Additionally, alsa busy waits when pulseaudio
is used, sw:avail_min is set and hw:buffer_size is small.
git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@5439 8ced0084-cf51-0410-be5f-012b33b47a6e
pierre