The x87 fpatan emulation is currently based around conversion to
double. This is inherently unsuitable for a good emulation of any
floatx80 operation. Reimplement using the soft-float operations, as
for other such instructions.
Backports commit ff57bb7b63267dabd60f88354c8c29ea5e1eb3ec from qemu
The x87 fyl2x emulation is currently based around conversion to
double. This is inherently unsuitable for a good emulation of any
floatx80 operation. Reimplement using the soft-float operations,
building on top of the reimplementation of fyl2xp1 and factoring out
code to be shared between the two instructions.
The included test assumes that the result in round-to-nearest mode
should always be one of the two closest floating-point numbers to the
mathematically exact result (including that it should be exact, in the
exact cases which cover more cases than for fyl2xp1).
Backports commit 1f18a1e6ab8368a4eab2d22894d3b2ae75250cd3 from qemu
The x87 fyl2xp1 emulation is currently based around conversion to
double. This is inherently unsuitable for a good emulation of any
floatx80 operation, even before considering that it is a particularly
naive implementation using double (adding 1 then using log rather than
attempting a better emulation using log1p).
Reimplement using the soft-float operations, as was done for f2xm1; as
in that case, m68k has related operations but not exactly this one and
it seemed safest to implement directly rather than reusing the m68k
code to avoid accumulation of errors.
A test is included with many randomly generated inputs. The
assumption of the test is that the result in round-to-nearest mode
should always be one of the two closest floating-point numbers to the
mathematical value of y * log2(x + 1); the implementation aims to do
somewhat better than that (about 70 correct bits before rounding). I
haven't investigated how accurate hardware is.
Intel manuals describe a narrower range of valid arguments to this
instruction than AMD manuals. The implementation accepts the wider
range (it's needed anyway for the core code to be reusable in a
subsequent patch reimplementing fyl2x), but the test only has inputs
in the narrower range so that it's valid on hardware that may reject
or produce poor results for inputs outside that range.
Code in the previous implementation that sets C2 for some out-of-range
arguments is not carried forward to the new implementation; C2 is
undefined for this instruction and I suspect that code was just
cut-and-pasted from the trigonometric instructions (fcos, fptan, fsin,
fsincos) where C2 *is* defined to be set for out-of-range arguments.
Backports commit 5eebc49d2d0aa5fc7e90eeac97533051bb7b72fa from qemu
The x87 fprem and fprem1 emulation is currently based around
conversion to double, which is inherently unsuitable for a good
emulation of any floatx80 operation. Reimplement using the soft-float
floatx80 remainder operations.
Backports commit 5ef396e2ba865f34a4766dbd60c739fb4bcb4fcc from qemu
Both x87 and m68k need the low parts of the quotient for their
remainder operations. Arrange for floatx80_modrem to track those bits
and return them via a pointer.
The architectures using float32_rem and float64_rem do not appear to
need this information, so the *_rem interface is left unchanged and
the information returned only from floatx80_modrem. The logic used to
determine the low 7 bits of the quotient for m68k
(target/m68k/fpu_helper.c:make_quotient) appears completely bogus (it
looks at the result of converting the remainder to integer, the
quotient having been discarded by that point); this patch does not
change that, but the m68k maintainers may wish to do so.
Backports commit 445810ec915687d37b8ae0ef8d7340ab4a153efa from qemu
The floatx80 remainder implementation unnecessarily sets the high bit
of bSig explicitly. By that point in the function, arguments that are
invalid, zero, infinity or NaN have already been handled and
subnormals have been through normalizeFloatx80Subnormal, so the high
bit will already be set. Remove the unnecessary code.
Backports commit 566601f1f9d972e44214696d3cb320e6c18880aa from qemu
The floatx80 remainder implementation sometimes returns the numerator
unchanged when the denominator is sufficiently larger than the
numerator. But if the value to be returned unchanged is a
pseudo-denormal, that is incorrect. Fix it to normalize the numerator
in that case.
Backports commit b662495dca0a2a36008cf8def91e2566519ed3f2 from qemu
The floatx80 remainder implementation ignores the high bit of the
significand when checking whether an operand (numerator) with zero
exponent is zero. This means it mishandles a pseudo-denormal
representation of 0x1p-16382L by treating it as zero. Fix this by
checking the whole significand instead.
Backports commit 499a2f7b554a295cfc10f8cd026d9b20a38fe664 from qemu
The m68k-specific softfloat code includes a function floatx80_mod that
is extremely similar to floatx80_rem, but computing the remainder
based on truncating the quotient toward zero rather than rounding it
to nearest integer. This is also useful for emulating the x87 fprem
and fprem1 instructions. Change the floatx80_rem implementation into
floatx80_modrem that can perform either operation, with both
floatx80_rem and floatx80_mod as thin wrappers available for all
targets.
There does not appear to be any use for the _mod operation for other
floating-point formats in QEMU (the only other architectures using
_rem at all are linux-user/arm/nwfpe, for FPA emulation, and openrisc,
for instructions that have been removed in the latest version of the
architecture), so no change is made to the code for other formats.
Backports commit 6b8b0136ab3018e4b552b485f808bf66bcf19ead from qemu
The x87 f2xm1 emulation is currently based around conversion to
double. This is inherently unsuitable for a good emulation of any
floatx80 operation, even before considering that it is a particularly
naive implementation using double (computing with pow and then
subtracting 1 rather than attempting a better emulation using expm1).
Reimplement using the soft-float operations, including additions and
multiplications with higher precision where appropriate to limit
accumulation of errors. I considered reusing some of the m68k code
for transcendental operations, but the instructions don't generally
correspond exactly to x87 operations (for example, m68k has 2^x and
e^x - 1, but not 2^x - 1); to avoid possible accumulation of errors
from applying multiple such operations each rounding to floatx80
precision, I wrote a direct implementation of 2^x - 1 instead. It
would be possible in principle to make the implementation more
efficient by doing the intermediate operations directly with
significands, signs and exponents and not packing / unpacking floatx80
format for each operation, but that would make it significantly more
complicated and it's not clear that's worthwhile; the m68k emulation
doesn't try to do that.
A test is included with many randomly generated inputs. The
assumption of the test is that the result in round-to-nearest mode
should always be one of the two closest floating-point numbers to the
mathematical value of 2^x - 1; the implementation aims to do somewhat
better than that (about 70 correct bits before rounding). I haven't
investigated how accurate hardware is.
Backports commit eca30647fc078f4d9ed1b455bd67960f99dbeb7a from qemu
In commit cfdb2c0c95ae9205b0 ("target/arm: Vectorize SABA/UABA") we
replaced the old handling of SABA/UABA with a vectorized implementation
which returns early rather than falling into the loop-ever-elements
code. We forgot to delete the part of the old looping code that
did the accumulate step, and Coverity correctly warns (CID 1428955)
that this code is now dead. Delete it.
Fixes: cfdb2c0c95ae9205b0
Backports commit ced7e8edb282765685d2ba0206a11f8692d8ec1c from qemu
Since commit ba3e7926691ed3 it has been unnecessary for target code
to call gen_io_end() after an IO instruction in icount mode; it is
sufficient to call gen_io_start() before it and to force the end of
the TB.
Many now-unnecessary calls to gen_io_end() were removed in commit
9e9b10c6491153b, but some were missed or accidentally added later.
Remove unneeded calls from the arm target:
* the call in the handling of exception-return-via-LDM is
unnecessary, and the code is already forcing end-of-TB
* the call in the VFP access check code is more complicated:
we weren't ending the TB, so we need to add the code to
force that by setting DISAS_UPDATE
* the doc comment for ARM_CP_IO doesn't need to mention
gen_io_end() any more
Backports commit 55c812b74289863c348449135812027d188f040a from qemu
The functions neon_element_offset(), neon_load_element(),
neon_load_element64(), neon_store_element() and
neon_store_element64() are used only in the translate-neon.inc.c
file, so move their definitions there.
Since the .inc.c file is #included in translate.c this doesn't make
much difference currently, but it's a more logical place to put the
functions and it might be helpful if we ever decide to try to make
the .inc.c files genuinely separate compilation units.
Backports commit 6fb5787898aab6aa04887fed9cf3220dd4c3f36a from qemu
Convert the Neon VTRN insn to decodetree. This is the last insn in the
Neon data-processing group, so we can remove all the now-unused old
decoder framework.
It's possible that there's a more efficient implementation of
VTRN, but for this conversion we just copy the existing approach.
Backports commit d4366190f84fe89cc5d46da995dac1e7d541b98e from qemu
Convert the Neon VSWP insn to decodetree. Since the new implementation
doesn't have to share a pass-loop with the other 2-reg-misc operations
we can implement the swap with 64-bit accesses rather than 32-bits
(which brings us into line with the pseudocode and is more efficient).
Backports commit 8ab3a227a0f13f0ff85846f36f7c466769aef4fc from qemu
Convert the Neon 2-reg-misc VRINT insns to decodetree.
Giving these insns their own do_vrint() function allows us
to change the rounding mode just once at the start and end
rather than doing it for every element in the vector.
Backports commit 128123ea34e9e6afe4842aefcb9cf84b9642ac22 from qemu
Convert the Neon 2-reg-misc insns which are implemented with
simple calls to functions that take the input, output and
fpstatus pointer.
Backports commit 3e96b205286dfb8bbf363229709e4f8648fce379 from qemu
Convert the Neon VQABS and VQNEG insns to decodetree.
Since these are the only ones which need cpu_env passing to
the helper, we wrap the helper rather than creating a whole
new do_2misc_env() function.
Backports commit 4936f38abe6db0a9d23fd04e4cb0cf4d51cff174 from qemu
Convert the remaining ops in the Neon 2-reg-misc group which
can be implemented simply with our do_2misc() helper.
Backports commit 84eae770af69c37a92496a4c4248875c070d5ee3 from qemu
Make gen_swap_half() take a source and destination TCGv_i32 rather
than modifying the input TCGv_i32; we're going to want to be able to
use it with the more flexible function signature, and this also
brings it into line with other functions like gen_rev16() and
gen_revsh().
Backports commit 8ec3de7018a8198624aae49eef5568256114a829 from qemu
All the other typedefs like these spell "Op" with a lowercase 'p';
remane the NeonGenTwoSingleOPFn and NeonGenTwoDoubleOPFn typedefs to
match.
Backports commit 5de3fd045be11b74cd0fbf36c6d4fb8387d5463b from qemu
The NeonGenOneOpFn typedef breaks with the pattern of the other
NeonGen*Fn typedefs, because it is a TCGv_i64 -> TCGv_i64 operation
but it does not have '64' in its name. Rename it to NeonGenOne64OpFn,
so that the old name is available for a TCGv_i32 -> TCGv_i32 operation
(which we will need in a subsequent commit).
Backports commit 039f4e809ad2772fb33de4511ff68a485d875618 from qemu
Convert to decodetree the insns in the Neon 2-reg-misc grouping which
we implement using gvec.
Backports commit 75153179e9928775d5333243ea4b278f438d75ae from qemu
Convert the Neon insns in the 2-reg-misc group which are
VCVT between f32 and f16 to decodetree.
Backports commit 654a517355e249435505ae5ff14a7520410cf7a4 from qemu
Convert the Neon narrowing moves VMQNV, VQMOVN, VQMOVUN in the 2-reg-misc
group to decodetree.
Backports commit 3882bdacb0ad548864b9f2582a32bb5c785e3165 from qemu
Convert the pairwise ops VPADDL and VPADAL in the 2-reg-misc grouping
to decodetree.
At this point we can get rid of the weird CPU_V001 #define that was
used to avoid having to explicitly list all the arguments being
passed to some TCG gen/helper functions.
Backports commit 6106af3aa2304fccee91a3a90138352b0c2af998 from qemu
Call the helper_hyp_tlb_flush() function on hfence instructions which
will generate an illegal insruction execption if we don't have
permission to flush the Hypervisor level TLBs.
Backports commit 2761db5fc20943bbd606b6fd49640ac000398de6 from qemu
The function that makes NaN-boxing when a 32-bit value is assigned
to a 64-bit FP register is split out to a helper gen_nanbox_fpr().
Then it is applied in translating of the FLW instruction.
Backports commit 354908cee1f7ff761b5fedbdb6376c378c10f941 from qemu
When building with clang version 10.0.0-4ubuntu1, we get:
CC lm32-softmmu/fpu/softfloat.o
fpu/softfloat.c:3365:13: error: bitwise negation of a boolean expression; did you mean logical negation? [-Werror,-Wbool-operation]
absZ &= ~ ( ( ( roundBits ^ 0x40 ) == 0 ) & roundNearestEven );
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
fpu/softfloat.c:3423:18: error: bitwise negation of a boolean expression; did you mean logical negation? [-Werror,-Wbool-operation]
absZ0 &= ~ ( ( (uint64_t) ( absZ1<<1 ) == 0 ) & roundNearestEven );
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
...
fpu/softfloat.c:4273:18: error: bitwise negation of a boolean expression; did you mean logical negation? [-Werror,-Wbool-operation]
zSig1 &= ~ ( ( zSig2 + zSig2 == 0 ) & roundNearestEven );
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Fix by rewriting the fishy bitwise AND of two bools as an int.
Backports commit 4066288694c3bdd175df813cad675a3b5191956b from qemu
Convert the Neon VDUP (scalar) insn to decodetree. (Note that we
can't call this just "VDUP" as we used that already in vfp.decode for
the "VDUP (general purpose register" insn.)
Backports commit 9aaa23c2ae18e6fb9a291b81baf91341db76dfa0 from qemu
Convert the Neon VTBL, VTBX instructions to decodetree. The actual
implementation of the insn is copied across to the new trans function
unchanged except for renaming 'tmp5' to 'tmp4'.
Backports commit 54e96c744b70a5d19f14b212a579dd3be8fcaad9 from qemu
Convert the Neon VEXT insn to decodetree. Rather than keeping the
old implementation which used fixed temporaries cpu_V0 and cpu_V1
and did the extraction with by-hand shift and logic ops, we use
the TCG extract2 insn.
We don't need to special case 0 or 8 immediates any more as the
optimizer is smart enough to throw away the dead code.
Backports commit 0aad761fb0aed40c99039eacac470cbd03d07019 from qemu
Convert the Neon 2-reg-scalar long multiplies to decodetree.
These are the last instructions in the group.
Backports commit 77e576a9281825fc170f3b3af83f47e110549b5c from qemu
