When handling left saturated ashifts with negative immediates, they
should be treated as right ashifts. This matches hardware behavior.
Reported-by: Igor Rayak <igorr@gitatechnologies.com>
The Blackfin port had some TRACE_xxx macros for easily logging trace data.
Use these as a base for common ones that have a simple form and match the
existing sets of helper macros.
The SEARCH insn is an oddball when it comes to parallel usage. It places a
big limit on what other insns it can run in parallel with, but we don't
currently track the amount of state needed to verify this (since no other insn
really requires this). Add a note for now in case we get around to it.
For many of the 32bit dsp shift related insns, we were just ignoring the HLs
field. The hardware does not though and will reject the insn if it's set
incorrectly. Update the sim to match.
We wrote a test case that tries every single 32bit opcode on the hardware
and compared it to the sim. There were a bunch of places in the sim where
we weren't strict enough (requiring certain parts of the opcode be set) so
we were treating a lot of invalid opcodes as valid ones. This sprinkles
out a lot additional checks in the dsp32alu class.
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
Two modifications:
1. The addition of 2013 to the copyright year range for every file;
2. The use of a single year range, instead of potentially multiple
year ranges, as approved by the FSF.
More ASTAT directed fixes, but this time at the dsp32shift insns.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
This improves some of the arithmetic shifts to better match the
hardware (especially wrt ASTAT behavior). We hit areas where
the published documentation is thin so we have to rely on tests
run on the hardware to figure out how things should behave.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
Now that we keep track of the exact parallel insn slot we're in, we can
make sure that the current insn being decoded is valid for that slot.
This brings us much closer to the hardware in flagging invalid parallel
insn combinations.
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
Some insns need to know which slot they're in to determine whether they
are valid. So add an enum for each slot, and check that rather than the
overall insn len. This makes tracking things in the code much clearer.
However, this code is functionally the same, so a follow up patch will
leverage this more to properly flag invalid parallel insn combos.
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
The amod1 helper includes a leading space so it can expand into the empty
string when need be, which means the caller need not add spacing itself.
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
Parallel insns can only do one 32bit, then two 16bits. So if we see
a 2nd 32bit insn after the first 32bit in a parallel insn, abort.
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
This fix the build time warning:
warning: format not a string literal, argument types not checked [-Wformat-nonliteral]
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
From: Robin Getz <robin.getz@analog.com>
Overflow with shift operations happens independently of saturation, but
we have the logic merged. Extend the lshift function so that callers
can tell it when to handle each independently, and then do so when it's
needed.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
The VIT_MAX insns can be used in parallel, so we need to use the store
buffer so we don't clobber the output register before we get a chance
to do a memory store with it.
Reported-by: Kai Iskratsch <kai@stella.at>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
The LSHIFT/ASHIFT insns that work with accumulators lacked AV/AVS
handling in the ASTAT register, so add it to match the hardware.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
If we're shifting accumulators, we don't want to touch the V bit in
ASTAT, so add size checks to the ashiftrt/lshiftrt helpers.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
The logical shift insn does not sign extend before shifting, so
we shouldn't either.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
This too should have been squashed into an earlier change. It covers
a few more cases in the V/VS saturation patch when working with TFU
and FU modes of dsp insns.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
When the shift magnitude exceeds 32 bits, the values rotate around (since
the hardware is actually a barrel shifter). So handle this edge case,
update the corresponding AV bit in ASTAT which was missing previously,
and tweak the AZ setting based on how the hardware behaves.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
The shift magnitude is a 5-bit signed value. When it is between 0 and
15, then we do the requested shift, but when it is outside of that, we
have to do the opposite.
That means we flip between lshift and ashiftrt to match the hardware.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
This reverts the previous commit and does it right. This change got
lost in the shuffle of patches I have pending.
Basically the logic is to change the 32bit saturation (and then masked
to 16bits) to a proper 16bit saturation.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
Some saturation cases with dsp mac insns were not setting the V flag.
So implement that part and split up the logic between the dual macs.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
Our handling of the M_IU/M_TFU modes are missing signed saturation when
the MM flag is set, so add it to match the hardware behavior.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
After testing the hardware with all the different dsp flags, the MM flag
triggers sign extension in all modes. So drop the limited use of it, and
the local custom helper that was also extending unsigned values. We also
can see that the flag checks in the mult/mac insns have the same behavior
with sign extending, so add a helper func to keep the logic the same in
both places.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
When using the IH modifier, we need to first saturate the value before
rounding it, and then further saturate it a bit more. This makes the
sim match the hardware behavior with these insns.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
When the accumulator saturates, it needs to be greater than, but not
equal to, the largest unsigned value as this is what the hardware does.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
A bunch of 32bit insns were not using the store buffer, so when they were
used in parallel insns, they would incorrectly clobber a register early.
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
When most video related insns are used in parallel with Ireg loads, the
DISALGNEXCPT insn behavior is implicitly in effect.
Reported-by: Anton Shokurov <shokurov.anton.v@yandex.ru>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
The current implementation attempts to handle the 16bit sign extension
itself. Unfortunately, it gets it right in some cases. So rather than
fix that logic, just drop it in favor of using 16bit signed casts. Now
gcc will take care of getting the logic right.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
The current handling of the subtraction insn with the RND12 modifier
works when saturation isn't involved. So add handling for this edge
case to match the hardware.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
The 16bit add/sub insns missed setting the VS bit in ASTAT whenever the
V bit was also set.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
The Blackfin PRM says that the top 8 bits of the accumulator must be
cleared when using the VIT_MAX insn, so the sim has followed this spec.
Matching the hardware behavior though when the high bits are not cleared
is easy to do and doesn't break existing behavior, so go for it.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
The Blackfin PRM does not cover this case, but the hardware is clear: even
if the search criteria is not met (and thus a new 16bit value is loaded up
into the accumulator), the accumulator undergoes 16bit sign extension. So
simply reload the low signed 16bits in that case.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
The Acc=-Acc insn can overflow or carry with edge values, so make sure
we update the ASTAT bits accordingly to match the hardware. Also fix
a thinko where we always updated AC0 even when working with A1 regs.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
The hardware allows the byteop[123]p insns to use the same src reg pair,
so remove the combination check in the sim.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
We weren't updating AZ when doing a 16bit add or sub insn. Implement it.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
No point in moving unchanged acc values to the acc regs, and avoid
updating the acc ASTAT bits when only reading. This fixes incorrect
changing of the ASTAT bits when they're only being read.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
The current dsp mult handler does not take care of overflows which turn
values negative (and thus set AN in ASTAT). So implement it.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
The current dsp mult handler does not take care of overflows and updating
the V ASTAT bit. So implement it.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
When testing ASTAT regs, specific bit differences carry a lot more meaning
than when checking the value of a data register. So automatically decode
the bits of the two values and print things out so that people don't have
to manually do it themselves every time.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
The saturation behavior with fract modes differs from non-fract modes.
Signed-off-by: Robin Getz <robin.getz@analog.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
This can boot Das U-Boot and a Linux kernel. It also supports Linux
userspace FLAT and FDPIC (dynamic and static) ELFs.
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
