Implement more simple 2-operand floating point MVE insns.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Implement the MVE VADD (floating-point) insn. Handling of this is
similar to the 2-operand integer insns, except that we must take care
to only update the floating point exception status if the least
significant bit of the predicate mask for each element is active.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Implement the MVE interleaving load/store functions VLD2, VLD4, VST2
and VST4. VLD2 loads 16 bytes of data from memory and writes to 2
consecutive Qregs; VLD4 loads 16 bytes of data from memory and writes
to 4 consecutive Qregs. The 'pattern' field in the encoding
determines the offset into memory which is accessed and also which
elements in the Qregs are written to. (The intention is that a
sequence of four consecutive VLD4 with different pattern values
performs a complete de-interleaving load of 64 bytes into all
elements of the 4 Qregs.) VST2 and VST4 do the same, but for stores.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE VLDR/VSTR insns which do scatter-gather using base
addresses from Qm plus or minus an immediate offset (possibly with
writeback). Note that writeback is not predicated but it does have
to honour ECI state, so we have to add an eci_mask check to the
VSTR_SG macros (the VLDR_SG macros already needed this to be able
to distinguish "skip beat" from "set predicated element to 0").
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE gather-loads and scatter-stores which
form the address by adding a base value from a scalar
register to an offset in each element of a vector.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE VCTP insn, which sets the VPR.P0 predicate bits so
as to predicate any element at index Rn or greater is predicated. As
with VPNOT, this insn itself is predicable and subject to beatwise
execution.
The calculation of the mask is the same as is used to determine
ltpmask in mve_element_mask(), but we precalculate masklen in
generated code to avoid having to have 4 helpers specialized by size.
We put the decode line in with the low-overhead-loop insns in
t32.decode because it's logically part of that collection of insn
patterns, even though it is an MVE only insn.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE VPNOT insn, which inverts the bits in VPR.P0
(subject to both predication and to beatwise execution).
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE VMAXA and VMINA insns, which take the absolute
value of the signed elements in the input vector and then accumulate
the unsigned max or min into the destination vector.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE 1-operand saturating operations VQABS and VQNEG.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE saturating doubling multiply accumulate insns
VQDMLAH, VQRDMLAH, VQDMLASH and VQRDMLASH. These perform a multiply,
double, add the accumulator shifted by the element size, possibly
round, saturate to twice the element size, then take the high half of
the result. The *MLAH insns do vector * scalar + vector, and the
*MLASH insns do vector * vector + scalar.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE VMLA insn, which multiplies a vector by a scalar
and accumulates into another vector.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE VMLADAV and VMLSLDAV insns. Like the VMLALDAV and
VMLSLDAV insns already implemented, these accumulate multiplied
vector elements; but they accumulate a 32-bit result rather than a
64-bit one.
Note that these encodings overlap with what would be RdaHi=0b111 for
VMLALDAV, VMLSLDAV, VRMLALDAVH and VRMLSLDAVH.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE narrowing move insns VMOVN, VQMOVN and VQMOVUN.
These take a double-width input, narrow it (possibly saturating) and
store the result to either the top or bottom half of the output
element.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE VABAV insn, which computes absolute differences
between elements of two vectors and accumulates the result into
a general purpose register.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE integer min/max across vector insns
VMAXV, VMINV, VMAXAV and VMINAV, which find the maximum
from the vector elements and a general purpose register,
and store the maximum back into the general purpose
register.
These insns overlap with VRMLALDAVH (they use what would
be RdaHi=0b110).
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE instructions which perform shifts by a scalar.
These are VSHL T2, VRSHL T2, VQSHL T1 and VQRSHL T2. They take the
shift amount in a general purpose register and shift every element in
the vector by that amount.
Mostly we can reuse the helper functions for shift-by-immediate; we
do need two new helpers for VQRSHL.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE VMLAS insn, which multiplies a vector by a vector
and adds a scalar.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE VPSEL insn, which sets each byte of the destination
vector Qd to the byte from either Qn or Qm depending on the value of
the corresponding bit in VPR.P0.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE integer vector comparison instructions that compare
each element against a scalar from a general purpose register. These
are "VCMP (vector)" encodings T4, T5 and T6 and "VPT (vector)"
encodings T4, T5 and T6.
We have to move the decodetree pattern for VPST, because it
overlaps with VCMP T4 with size = 0b11.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE integer vector comparison instructions. These are
"VCMP (vector)" encodings T1, T2 and T3, and "VPT (vector)" encodings
T1, T2 and T3.
These insns compare corresponding elements in each vector, and update
the VPR.P0 predicate bits with the results of the comparison. VPT
also sets the VPR.MASK01 and VPR.MASK23 fields -- it is effectively
"VCMP then VPST".
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE incrementing/decrementing dup insns VIDUP, VDDUP,
VIWDUP and VDWDUP. These fill the elements of a vector with
successively incrementing values, starting at the offset specified in
a general purpose register. The final value of the offset is written
back to this register. The wrapping variants take a second general
purpose register which specifies the point where the count should
wrap back to 0.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE VMULL (polynomial) insn. Unlike Neon, this comes
in two flavours: 8x8->16 and a 16x16->32. Also unlike Neon, the
inputs are in either the low or the high half of each double-width
element.
The assembler for this insn indicates the size with "P8" or "P16",
encoded into bit 28 as size = 0 or 1. We choose to follow the
same encoding as VQDMULL and decode this into a->size as MO_16
or MO_32 indicating the size of the result elements. This then
carries through to the helper function names where it then
matches up with the existing pmull_h() which does an 8x8->16
operation and a new pmull_w() which does the 16x16->32.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
For vector loads, predicated elements are zeroed, instead of
retaining their previous values (as happens for most data
processing operations). This means we need to distinguish
"beat not executed due to ECI" (don't touch destination
element) from "beat executed but predicated out" (zero
destination element).
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
We were not paying attention to the ECI state when advancing the VPT
state. Architecturally, VPT state advance happens for every beat
(see the pseudocode VPTAdvance()), so on every beat the 4 bits of
VPR.P0 corresponding to the current beat are inverted if required,
and at the end of beats 1 and 3 the VPR MASK fields are updated.
This means that if the ECI state says we should not be executing all
4 beats then we need to skip some of the updating of the VPR that we
currently do in mve_advance_vpt().
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
In some situations we need a mask telling us which parts of the
vector correspond to beats that are not being executed because of
ECI, separately from the combined "which bytes are predicated away"
mask. Factor this mask calculation out of mve_element_mask() into
its own function.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
In mve_element_mask(), we calculate a mask for tail predication which
should have a number of 1 bits based on the value of LR. However,
our MAKE_64BIT_MASK() macro has undefined behaviour when passed a
zero length. Special case this to give the all-zeroes mask we
require.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
We got an edge case wrong in the 48-bit SQRSHRL implementation: if
the shift is to the right, although it always makes the result
smaller than the input value it might not be within the 48-bit range
the result is supposed to be if the input had some bits in [63..48]
set and the shift didn't bring all of those within the [47..0] range.
Handle this similarly to the way we already do for this case in
do_uqrshl48_d(): extend the calculated result from 48 bits,
and return that if not saturating or if it doesn't change the
result; otherwise fall through to return a saturated value.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
In do_sqrshl48_d() and do_uqrshl48_d() we got some of the edge
cases wrong and failed to saturate correctly:
(1) In do_sqrshl48_d() we used the same code that do_shrshl_bhs()
does to obtain the saturated most-negative and most-positive 48-bit
signed values for the large-shift-left case. This gives (1 << 47)
for saturate-to-most-negative, but we weren't sign-extending this
value to the 64-bit output as the pseudocode requires.
(2) For left shifts by less than 48, we copied the "8/16 bit" code
from do_sqrshl_bhs() and do_uqrshl_bhs(). This doesn't do the right
thing because it assumes the C type we're working with is at least
twice the number of bits we're saturating to (so that a shift left by
bits-1 can't shift anything off the top of the value). This isn't
true for bits == 48, so we would incorrectly return 0 rather than the
most-positive value for situations like "shift (1 << 44) right by
20". Instead check for saturation by doing the shift and signextend
and then testing whether shifting back left again gives the original
value.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
In the MVE helpers for the narrowing operations (DO_VSHRN and
DO_VSHRN_SAT) we were using the wrong bits of the predicate mask for
the 'top' versions of the insn. This is because the loop works over
the double-sized input elements and shifts the predicate mask by that
many bits each time, but when we write out the half-sized output we
must look at the mask bits for whichever half of the element we are
writing to.
Correct this by shifting the whole mask right by ESIZE bits for the
'top' insns. This allows us also to simplify the saturation bit
checking (where we had noticed that we needed to look at a different
mask bit for the 'top' insn.)
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
A cut-and-paste error meant we handled signed VADDV like
unsigned VADDV; fix the type used.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
In the MVE shift-and-insert insns, we special case VSLI by 0
and VSRI by <dt>. VSRI by <dt> means "don't update the destination",
which is what we've implemented. However VSLI by 0 is "set
destination to the input", so we don't want to use the same
special-casing that we do for VSRI by <dt>.
Since the generic logic gives the right answer for a shift
by 0, just use that.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the MVE shifts by register, which perform
shifts on a single general-purpose register.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210628135835.6690-19-peter.maydell@linaro.org
Implement the MVE shifts by immediate, which perform shifts
on a single general-purpose register.
These patterns overlap with the long-shift-by-immediates,
so we have to rearrange the grouping a little here.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210628135835.6690-18-peter.maydell@linaro.org
Implement the MVE long shifts by register, which perform shifts on a
pair of general-purpose registers treated as a 64-bit quantity, with
the shift count in another general-purpose register, which might be
either positive or negative.
Like the long-shifts-by-immediate, these encodings sit in the space
that was previously the UNPREDICTABLE MOVS/ORRS with Rm==13,15.
Because LSLL_rr and ASRL_rr overlap with both MOV_rxri/ORR_rrri and
also with CSEL (as one of the previously-UNPREDICTABLE Rm==13 cases),
we have to move the CSEL pattern into the same decodetree group.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210628135835.6690-17-peter.maydell@linaro.org
The MVE extension to v8.1M includes some new shift instructions which
sit entirely within the non-coprocessor part of the encoding space
and which operate only on general-purpose registers. They take up
the space which was previously UNPREDICTABLE MOVS and ORRS encodings
with Rm == 13 or 15.
Implement the long shifts by immediate, which perform shifts on a
pair of general-purpose registers treated as a 64-bit quantity, with
an immediate shift count between 1 and 32.
Awkwardly, because the MOVS and ORRS trans functions do not UNDEF for
the Rm==13,15 case, we need to explicitly emit code to UNDEF for the
cases where v8.1M now requires that. (Trying to change MOVS and ORRS
is too difficult, because the functions that generate the code are
shared between a dozen different kinds of arithmetic or logical
instruction for all A32, T16 and T32 encodings, and for some insns
and some encodings Rm==13,15 are valid.)
We make the helper functions we need for UQSHLL and SQSHLL take
a 32-bit value which the helper casts to int8_t because we'll need
these helpers also for the shift-by-register insns, where the shift
count might be < 0 or > 32.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210628135835.6690-16-peter.maydell@linaro.org
Implement the MVE VADDLV insn; this is similar to VADDV, except
that it accumulates 32-bit elements into a 64-bit accumulator
stored in a pair of general-purpose registers.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210628135835.6690-15-peter.maydell@linaro.org
Implement the MVE VSHLC insn, which performs a shift left of the
entire vector with carry in bits provided from a general purpose
register and carry out bits written back to that register.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210628135835.6690-14-peter.maydell@linaro.org
Implement the MVE saturating shift-right-and-narrow insns
VQSHRN, VQSHRUN, VQRSHRN and VQRSHRUN.
do_srshr() is borrowed from sve_helper.c.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210628135835.6690-13-peter.maydell@linaro.org
Implement the MVE shift-right-and-narrow insn VSHRN and VRSHRN.
do_urshr() is borrowed from sve_helper.c.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210628135835.6690-12-peter.maydell@linaro.org
Implement the MVE VSRI and VSLI insns, which perform a
shift-and-insert operation.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210628135835.6690-11-peter.maydell@linaro.org
Implement the MVE VHLL (vector shift left long) insn. This has two
encodings: the T1 encoding is the usual shift-by-immediate format,
and the T2 encoding is a special case where the shift count is always
equal to the element size.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210628135835.6690-10-peter.maydell@linaro.org
Implement the MVE vector shift right by immediate insns VSHRI and
VRSHRI. As with Neon, we implement these by using helper functions
which perform left shifts but allow negative shift counts to indicate
right shifts.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210628135835.6690-9-peter.maydell@linaro.org
Implement the MVE shift-vector-left-by-immediate insns VSHL, VQSHL
and VQSHLU.
The size-and-immediate encoding here is the same as Neon, and we
handle it the same way neon-dp.decode does.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210628135835.6690-8-peter.maydell@linaro.org
Implement the MVE logical-immediate insns (VMOV, VMVN,
VORR and VBIC). These have essentially the same encoding
as their Neon equivalents, and we implement the decode
in the same way.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210628135835.6690-7-peter.maydell@linaro.org
The initial implementation of the MVE VRMLALDAVH and VRMLSLDAVH
insns had some bugs:
* the 32x32 multiply of elements was being done as 32x32->32,
not 32x32->64
* we were incorrectly maintaining the accumulator in its full
72-bit form across all 4 beats of the insn; in the pseudocode
it is squashed back into the 64 bits of the RdaHi:RdaLo
registers after each beat
In particular, fixing the second of these allows us to recast
the implementation to avoid 128-bit arithmetic entirely.
Since the element size here is always 4, we can also drop the
parameterization of ESIZE to make the code a little more readable.
Suggested-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210628135835.6690-3-peter.maydell@linaro.org
Implement the MVE VADDV insn, which performs an addition
across vector lanes.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210617121628.20116-44-peter.maydell@linaro.org
Implement the MVE VHCADD insn, which is similar to VCADD
but performs a halving step. This one overlaps with VADC.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210617121628.20116-43-peter.maydell@linaro.org
Implement the MVE VCADD insn, which performs a complex add with
rotate. Note that the size=0b11 encoding is VSBC.
The architecture grants some leeway for the "destination and Vm
source overlap" case for the size MO_32 case, but we choose not to
make use of it, instead always calculating all 16 bytes worth of
results before setting the destination register.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210617121628.20116-42-peter.maydell@linaro.org
Implement the MVE VADC and VSBC insns. These perform an
add-with-carry or subtract-with-carry of the 32-bit elements in each
lane of the input vectors, where the carry-out of each add is the
carry-in of the next. The initial carry input is either 1 or is from
FPSCR.C; the carry out at the end is written back to FPSCR.C.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210617121628.20116-41-peter.maydell@linaro.org
Implement the MVE VRHADD insn, which performs a rounded halving
addition.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210617121628.20116-40-peter.maydell@linaro.org
