In the neon_padd/pmax/pmin helpers for float16, a cut-and-paste error
meant we were using the H4() address swizzler macro rather than the
H2() which is required for 2-byte data. This had no effect on
little-endian hosts but meant we put the result data into the
destination Dreg in the wrong order on big-endian hosts.
Backports 552714c0812a10e5cff239bd29928e5fcb8d8b3b
In both cases, we can sink the write-back and perform
the accumulate into the normal destination temps
Backports 9f1a5f93c2dd345dc6c8fe86ed14bf1485056f6e
The only uses of this function are for loading VFP
double-precision values, and nothing to do with NEON.
Backports b38b96ca90827012ab8eb045c1337cea83a54c4b
The only uses of this function are for loading VFP
single-precision values, and nothing to do with NEON.
Backports 21c1c0e50b73c580c6bfc8f2314d1b6a14793561
We can then use this to improve VMOV (scalar to gp) and
VMOV (gp to scalar) so that we simply perform the memory
operation that we wanted, rather than inserting or
extracting from a 32-bit quantity.
These were the last uses of neon_load/store_reg, so remove them.
Backports 4d5fa5a80ac28f34b8497be1e85371272413a12e
Model these off the aa64 read/write_vec_element functions.
Use it within translate-neon.c.inc. The new functions do
not allocate or free temps, so this rearranges the calling
code a bit.
Backports a712266f5d5a36d04b22fe69fa15592d62bed019
This function makes it clear that we're talking about the whole
register, and not the 32-bit piece at index 0. This fixes a bug
when running on a big-endian host.
Backports 015ee81a4c06b644969f621fd9965cc6372b879e
If the M-profile low-overhead-branch extension is implemented, FPSCR
bits [18:16] are a new field LTPSIZE. If MVE is not implemented
(currently always true for us) then this field always reads as 4 and
ignores writes.
These bits used to be the vector-length field for the old
short-vector extension, so we need to take care that they are not
misinterpreted as setting vec_len. We do this with a rearrangement
of the vfp_set_fpscr() code that deals with vec_len, vec_stride
and also the QC bit; this obviates the need for the M-profile
only masking step that we used to have at the start of the function.
We provide a new field in CPUState for LTPSIZE, even though this
will always be 4, in preparation for MVE, so we don't have to
come back later and split it out of the vfp.xregs[FPSCR] value.
(This state struct field will be saved and restored as part of
the FPSCR value via the vmstate_fpscr in machine.c.)
Backports 8128c8e8cc9489a8387c74075974f86dc0222e7f
M-profile CPUs with half-precision floating point support should
be able to write to FPSCR.FZ16, but an M-profile specific masking
of the value at the top of vfp_set_fpscr() currently prevents that.
This is not yet an active bug because we have no M-profile
FP16 CPUs, but needs to be fixed before we can add any.
The bits that the masking is effectively preventing from being
set are the A-profile only short-vector Len and Stride fields,
plus the Neon QC bit. Rearrange the order of the function so
that those fields are handled earlier and only under a suitable
guard; this allows us to drop the M-profile specific masking,
making FZ16 writeable.
This change also makes the QC bit correctly RAZ/WI for older
no-Neon A-profile cores.
This refactoring also paves the way for the low-overhead-branch
LTPSIZE field, which uses some of the bits that are used for
A-profile Stride and Len.
Backports commit d31e2ce68d56f5bcc83831497e5fe4b8a7e18e85
v8.1M's "low-overhead-loop" extension has three instructions
for looping:
* DLS (start of a do-loop)
* WLS (start of a while-loop)
* LE (end of a loop)
The loop-start instructions are both simple operations to start a
loop whose iteration count (if any) is in LR. The loop-end
instruction handles "decrement iteration count and jump back to loop
start"; it also caches the information about the branch back to the
start of the loop to improve performance of the branch on subsequent
iterations.
As with the branch-future instructions, the architecture permits an
implementation to discard the LO_BRANCH_INFO cache at any time, and
QEMU takes the IMPDEF option to never set it in the first place
(equivalent to discarding it immediately), because for us a "real"
implementation would be unnecessary complexity.
(This implementation only provides the simple looping constructs; the
vector extension MVE (Helium) adds some extra variants to handle
looping across vectors. We'll add those later when we implement
MVE.)
Backports commit b7226369721896ab9ef71544e4fe95b40710e05a
v8.1M implements a new 'branch future' feature, which is a
set of instructions that request the CPU to perform a branch
"in the future", when it reaches a particular execution address.
In hardware, the expected implementation is that the information
about the branch location and destination is cached and then
acted upon when execution reaches the specified address.
However the architecture permits an implementation to discard
this cached information at any point, and so guest code must
always include a normal branch insn at the branch point as
a fallback. In particular, an implementation is specifically
permitted to treat all BF insns as NOPs (which is equivalent
to discarding the cached information immediately).
For QEMU, implementing this caching of branch information
would be complicated and would not improve the speed of
execution at all, so we make the IMPDEF choice to implement
all BF insns as NOPs.
Backports commit 05903f036edba8e3ed940cc215b8e27fb49265b9
The BLX immediate insn in the Thumb encoding always performs
a switch from Thumb to Arm state. This would be totally useless
in M-profile which has no Arm decoder, and so the instruction
does not exist at all there. Make the encoding UNDEF for M-profile.
(This part of the encoding space is used for the branch-future
and low-overhead-loop insns in v8.1M.)
Backports 920f04fa3ea789f8f85a52cee5395b8887b56cf7
The t32 decode has a group which represents a set of insns
which overlap with B_cond_thumb because they have [25:23]=111
(which is an invalid condition code field for the branch insn).
This group is currently defined using the {} overlap-OK syntax,
but it is almost entirely non-overlapping patterns. Switch
it over to use a non-overlapping group.
For this to be valid syntactically, CPS must move into the same
overlapping-group as the hint insns (CPS vs hints was the
only actual use of the overlap facility for the group).
The non-overlapping subgroup for CLREX/DSB/DMB/ISB/SB is no longer
necessary and so we can remove it (promoting those insns to
be members of the parent group).
Backports 45f11876ae86128bdee27e0b089045de43cc88e4
From v8.1M, disabled-coprocessor handling changes slightly:
* coprocessors 8, 9, 14 and 15 are also governed by the
cp10 enable bit, like cp11
* an extra range of instruction patterns is considered
to be inside the coprocessor space
We previously marked these up with TODO comments; implement the
correct behaviour.
Unfortunately there is no ID register field which indicates this
behaviour. We could in theory test an unrelated ID register which
indicates guaranteed-to-be-in-v8.1M behaviour like ID_ISAR0.CmpBranch
>= 3 (low-overhead-loops), but it seems better to simply define a new
ARM_FEATURE_V8_1M feature flag and use it for this and other
new-in-v8.1M behaviour that isn't identifiable from the ID registers.
Backports commit 5d2555a1fe7370feeb1efbbf276a653040910017
For nested groups like:
{
[
pattern 1
pattern 2
]
pattern 3
}
the intended behaviour is that patterns 1 and 2 must not
overlap with each other; if the insn matches neither then
we fall through to pattern 3 as the next thing in the
outer overlapping group.
Currently we generate incorrect code for this situation,
because in the code path for a failed match inside the
inner non-overlapping group we generate a "return" statement,
which causes decode to stop entirely rather than continuing
to the next thing in the outer group.
Generate a "break" instead, so that decode flow behaves
as required for this nested group case.
Backports 514101c0b931f0a11a40d29d26af1cc40482f951
Unlike many other bits in HCR_EL2, the description for this
bit does not contain the phrase "if ... this field behaves
as 0 for all purposes other than", so do not squash the bit
in arm_hcr_el2_eff.
Instead, replicate the E2H+TGE test in the two places that
require it.
Backports 4301acd7d7d455792ea873ced75c0b5d653618b1
The reporting in AArch64.TagCheckFail only depends on PSTATE.EL,
and not the AccType of the operation. There are two guest
visible problems that affect LDTR and STTR because of this:
(1) Selecting TCF0 vs TCF1 to decide on reporting,
(2) Report "data abort same el" not "data abort lower el".
Backports 50244cc76abcac3296cff3d84826f5ff71808c80
We already have the full ARMMMUIdx as computed from the
function parameter.
For the purpose of regime_has_2_ranges, we can ignore any
difference between AccType_Normal and AccType_Unpriv, which
would be the only difference between the passed mmu_idx
and arm_mmu_idx_el.
Backports 4aedfc0f633fd06dd2a5dc8ffa93f4c56117e37f
For AArch32, unlike the VCVT of integer to float, which honours the
rounding mode specified by the FPSCR, VCVT of fixed-point to float is
always round-to-nearest. (AArch64 fixed-point-to-float conversions
always honour the FPCR rounding mode.)
Implement this by providing _round_to_nearest versions of the
relevant helpers which set the rounding mode temporarily when making
the call to the underlying softfloat function.
We only need to change the VFP VCVT instructions, because the
standard- FPSCR value used by the Neon VCVT is always set to
round-to-nearest, so we don't need to do the extra work of saving
and restoring the rounding mode.
Backports commit 61db12d9f9eb36761edba4d9a414cd8dd34c512b
The SMLAD instruction is supposed to:
* signed multiply Rn[15:0] * Rm[15:0]
* signed multiply Rn[31:16] * Rm[31:16]
* perform a signed addition of the products and Ra
* set Rd to the low 32 bits of the theoretical
infinite-precision result
* set the Q flag if the sign-extension of Rd
would differ from the infinite-precision result
(ie on overflow)
Our current implementation doesn't quite do this, though: it performs
an addition of the products setting Q on overflow, and then it adds
Ra, again possibly setting Q. This sometimes incorrectly sets Q when
the architecturally mandated only-check-for-overflow-once algorithm
does not. For instance:
r1 = 0x80008000; r2 = 0x80008000; r3 = 0xffffffff
smlad r0, r1, r2, r3
This is (-32768 * -32768) + (-32768 * -32768) - 1
The products are both 0x4000_0000, so when added together as 32-bit
signed numbers they overflow (and QEMU sets Q), but because the
addition of Ra == -1 brings the total back down to 0x7fff_ffff
there is no overflow for the complete operation and setting Q is
incorrect.
Fix this edge case by resorting to 64-bit arithmetic for the
case where we need to add three values together.
Backports commit 5288145d716338ace0f83e3ff05c4d07715bb4f4
QEMU supports a 48-bit physical address range, but we don't currently
expose it in the '-cpu max' ID registers (you get the same range as
Cortex-A57, which is 44 bits).
Set the ID_AA64MMFR0.PARange field to indicate 48 bits.
Backports d1b6b7017572e8d82f26eb827a1dba0e8cf3cae6
When the two arguments are identical, this can be reduced to
dup_vec or to mov_vec from a tcg_constant_vec.
Backports commit 1dc4fe70128db05237a00eda6eb15e2b44deb31f
The definition of INDEX_op_dupi_vec is that it operates on
units of tcg_target_ulong -- in this case 32 bits. It does
not work to use this for a uint64_t value that happens to be
small enough to fit in tcg_target_ulong.
Backports a5b30d950c42b14bc9da24d1e68add6538d23336
The previous change wrongly stated that 32-bit avx2 should have
used VPBROADCASTW. But that's a 16-bit broadcast and we want a
32-bit broadcast.
Backports f80d09b599a5e0fd7f44653f23b04104cb703f7a
These are easier to set and test when they have their own fields.
Reduce the size of alias_index and sort_index to 4 bits, which is
sufficient for TCG_MAX_OP_ARGS. This leaves only the bits indicating
constants within the ct field.
Move all initialization to allocation time, rather than init
individual fields in process_op_defs.
Backports bc2b17e6ea582ef3ade2bdca750de269c674c915
This wasn't actually used for anything, really. All variable
operands must accept registers, and which are indicated by the
set in TCGArgConstraint.regs.
Backports commit 74a117906b87ff9220e4baae5a7431d6f4eadd45
This uses an existing hole in the TCGArgConstraint structure
and will be convenient for keeping the data in one place.
Backports 66792f90f14fef18b25a168922877a367ecdca05
With larger vector sizes, it turns out oprsz == maxsz, and we only
need to represent mismatch for oprsz <= 32. We do, however, need
to represent larger oprsz and do so without reducing SIMD_DATA_BITS.
Reduce the size of the oprsz field and increase the maxsz field.
Steal the oprsz value of 24 to indicate equality with maxsz.
Backports e2e7168a214b0ed98dc357bba96816486a289762
While converting to gen_gvec_ool_zzzp, we lost passing
a->esz as the data argument to the function.
Backports commit dd701fafe55a78e655d4823d29226d92250a6b56
The M-profile definition of the MVFR1 ID register differs slightly
from the A-profile one, and in particular the check for "does the CPU
support fp16 arithmetic" is not the same.
We don't currently implement any M-profile CPUs with fp16 arithmetic,
so this is not yet a visible bug, but correcting the logic now
disarms this beartrap for when we eventually do.
Backports commit dfc523a84b06b6a4b583ed4c29d24fd980dd37a0
Move the id_pfr0 and id_pfr1 fields into the ARMISARegisters
sub-struct. We're going to want id_pfr1 for an isar_features
check, and moving both at the same time avoids an odd
inconsistency.
Changes other than the ones to cpu.h and kvm64.c made
automatically with:
perl -p -i -e 's/cpu->id_pfr/cpu->isar.id_pfr/' target/arm/*.c hw/intc/armv7m_nvic.c
Backports commit 8a130a7be6e222965641e1fd9469fd3ee752c7d4
The ARM_FEATURE_PXN bit indicates whether the CPU supports the PXN
bit in short-descriptor translation table format descriptors. This
is indicated by ID_MMFR0.VMSA being at least 0b0100. Replace the
feature bit with an ID register check, in line with our preference
for ID register checks over feature bits.
Backports commit 0ae0326b984e77a55c224b7863071bd3d8951231
Per Intel SDM vol 1, 13.2, if CPUID.1:ECX.XSAVE[bit 26] is 0, the
processor provides no further enumeration through CPUID function 0DH.
QEMU does not do this for "-cpu host,-xsave".
Backports 19ca8285fcd61a8f60f2f44f789a561e0958e8e6
It turns out that some hosts have a default malloc alignment less
than that required for vectors.
We assume that, with compiler annotation on CPUArchState, that we
can properly align the vector portion of the guest state. Fix the
alignment of the allocation by using qemu_memalloc when required.
This check was backwards when introduced in commit
033614c47de78409ad3fb39bb7bd1483b71c6789:
target/arm: Filter cycle counter based on PMCCFILTR_EL0
Backports commit db1f3afb17269cf2bd86c222e1bced748487ef71
The VCMLA and VCADD insns have a size field which is 0 for fp16
and 1 for fp32 (note that this is the reverse of the Neon 3-same
encoding!). Convert it to MO_* values in decode for consistency.
Backports d186a4854c04e9832907b0b4240a47731da20993
Convert the insns using the 2reg_vcvt and 2reg_vcvt_f16 formats
to pass the size through to the trans function as a MO_* value
rather than the '0==f32, 1==f16' used in the fp 3-same encodings.
Backports commit 0ae715c658a02af1834b63563c56112a6d8842cb
