Remove the now unused TCG globals cpu_F0s, cpu_F0d, cpu_F1s, cpu_F1d.
cpu_M0 is still used by the iwmmxt code, and cpu_V0 and
cpu_V1 are used by both iwmmxt and Neon.
Backports commit d9eea52c67c04c58ecceba6ffe5a93d1d02051fa from qemu
Remove some old constructns from NEON_2RM_VCVT_F16_F32 code:
* don't use CPU_F0s
* don't use tcg_gen_st_f32
Backports commit b66f6b9981004bbf120b8d17c20f92785179bdf2 from qemu
Remove some old constructs from NEON_2RM_VCVT_F16_F32 code:
* don't use cpu_F0s
* don't use tcg_gen_ld_f32
Backports commit 58f2682eee738e8890f9cfe858e0f4f68b00d45d from qemu
Stop using cpu_F0s for the Neon f32/s32 VCVT operations.
Since this is the last user of cpu_F0s in the Neon 2rm-op
loop, we can remove the handling code for it too.
Backports commit 60737ed5785b9c1c6f1c85575dfdd1e9eec91878 from qemu
Where Neon instructions are floating point operations, we
mostly use the old VFP utility functions like gen_vfp_abs()
which work on the TCG globals cpu_F0s and cpu_F1s. The
Neon for-each-element loop conditionally loads the inputs
into either a plain old TCG temporary for most operations
or into cpu_F0s for float operations, and similarly stores
back either cpu_F0s or the temporary.
Switch NEON_2RM_VABS_F away from using cpu_F0s, and
update neon_2rm_is_float_op() accordingly.
Backports commit fd8a68cdcf81d70eebf866a132e9780d4108da9c from qemu
The AArch32 VMOV (immediate) instruction uses the same VFP encoded
immediate format we already handle in vfp_expand_imm(). Use that
function rather than hand-decoding it.
Backports commit 9bee50b498410ed6466018b26464d7384c7879e9 from qemu
We want to use vfp_expand_imm() in the AArch32 VFP decode;
move it from the a64-only header/source file to the
AArch32 one (which is always compiled even for AArch64).
Backports commit d6a092d479333b5f20a647a912a31b0102d37335 from qemu
For VFP short vectors, the VFP registers are divided into a
series of banks: for single-precision these are s0-s7, s8-s15,
s16-s23 and s24-s31; for double-precision they are d0-d3,
d4-d7, ... d28-d31. Some banks are "scalar" meaning that
use of a register within them triggers a pure-scalar or
mixed vector-scalar operation rather than a full vector
operation. The scalar banks are s0-s7, d0-d3 and d16-d19.
When using a bank as part of a vector operation, we
iterate through it, increasing the register number by
the specified stride each time, and wrapping around to
the beginning of the bank.
Unfortunately our calculation of the "increment" part of this
was incorrect:
vd = ((vd + delta_d) & (bank_mask - 1)) | (vd & bank_mask)
will only do the intended thing if bank_mask has exactly
one set high bit. For instance for doubles (bank_mask = 0xc),
if we start with vd = 6 and delta_d = 2 then vd is updated
to 12 rather than the intended 4.
This only causes problems in the unlikely case that the
starting register is not the first in its bank: if the
register number doesn't have to wrap around then the
expression happens to give the right answer.
Fix this bug by abstracting out the "check whether register
is in a scalar bank" and "advance register within bank"
operations to utility functions which use the right
bit masking operations
Backports commit 18cf951af9a27ae573a6fa17f9d0c103f7b7679b from qemu
Convert the float-to-integer VCVT instructions to decodetree.
Since these are the last unconverted instructions, we can
delete the old decoder structure entirely now.
Backports commit 3111bfc2da6ba0c8396dc97ca479942d711c6146 from qemu
Convert the VCVT (between floating-point and fixed-point) instructions
to decodetree.
Backports commit e3d6f4290c788e850c64815f0b3e331600a4bcc0 from qemu
Convert the VFP round-to-integer instructions VRINTR, VRINTZ and
VRINTX to decodetree.
These instructions were only introduced as part of the "VFP misc"
additions in v8A, so we check this. The old decoder's implementation
was incorrectly providing them even for v7A CPUs.
Backports commit e25155f55dc4abb427a88dfe58bbbc550fe7d643 from qemu
Convert the VCVTT and VCVTB instructions which convert from
f32 and f64 to f16 to decodetree.
Since we're no longer constrained to the old decoder's style
using cpu_F0s and cpu_F0d we can perform a direct 16 bit
store of the right half of the input single-precision register
rather than doing a load/modify/store sequence on the full
32 bits.
Backports commit cdfd14e86ab0b1ca29a702d13a8e4af2e902a9bf from qemu
Convert the VCVTT, VCVTB instructions that deal with conversion
from half-precision floats to f32 or 64 to decodetree.
Since we're no longer constrained to the old decoder's style
using cpu_F0s and cpu_F0d we can perform a direct 16 bit
load of the right half of the input single-precision register
rather than loading the full 32 bits and then doing a
separate shift or sign-extension.
Backports commit b623d803dda805f07aadcbf098961fde27315c19 from qemu
Convert the VFP comparison instructions to decodetree.
Note that comparison instructions should not honour the VFP
short-vector length and stride information: they are scalar-only
operations. This applies to all the 2-operand instructions except
for VMOV, VABS, VNEG and VSQRT. (In the old decoder this is
implemented via the "if (op == 15 && rn > 3) { veclen = 0; }" check.)
Backports commit 386bba2368842fc74388a3c1651c6c0c0c70adbd from qemu
Convert the VFP VABS instruction to decodetree.
Unlike the 3-op versions, we don't pass fpst to the VFPGen2OpSPFn or
VFPGen2OpDPFn because none of the operations which use this format
and support short vectors will need it.
Backports commit 90287e22c987e9840704345ed33d237cbe759dd9 from qemu
Convert the VFP fused multiply-add instructions (VFNMA, VFNMS,
VFMA, VFMS) to decodetree.
Note that in the old decode structure we were implementing
these to honour the VFP vector stride/length. These instructions
were introduced in VFPv4, and in the v7A architecture they
are UNPREDICTABLE if the vector stride or length are non-zero.
In v8A they must UNDEF if stride or length are non-zero, like
all VFP instructions; we choose to UNDEF always.
Backports commit d4893b01d23060845ee3855bc96626e16aad9ab5 from qemu
Convert the VFP VMLA instruction to decodetree.
This is the first of the VFP 3-operand data processing instructions,
so we include in this patch the code which loops over the elements
for an old-style VFP vector operation. The existing code to do this
looping uses the deprecated cpu_F0s/F0d/F1s/F1d TCG globals; since
we are going to be converting instructions one at a time anyway
we can take the opportunity to make the new loop use TCG temporaries,
which means we can do that conversion one operation at a time
rather than needing to do it all in one go.
We include an UNDEF check which was missing in the old code:
short-vector operations (with stride or length non-zero) were
deprecated in v7A and must UNDEF in v8A, so if the MVFR0 FPShVec
field does not indicate that support for short vectors is present
we UNDEF the operations that would use them. (This is a change
of behaviour for Cortex-A7, Cortex-A15 and the v8 CPUs, which
previously were all incorrectly allowing short-vector operations.)
Note that the conversion fixes a bug in the old code for the
case of VFP short-vector "mixed scalar/vector operations". These
happen where the destination register is in a vector bank but
but the second operand is in a scalar bank. For example
vmla.f64 d10, d1, d16 with length 2 stride 2
is equivalent to the pair of scalar operations
vmla.f64 d10, d1, d16
vmla.f64 d8, d3, d16
where the destination and first input register cycle through
their vector but the second input is scalar (d16). In the
old decoder the gen_vfp_F1_mul() operation uses cpu_F1{s,d}
as a temporary output for the multiply, which trashes the
second input operand. For the fully-scalar case (where we
never do a second iteration) and the fully-vector case
(where the loop loads the new second input operand) this
doesn't matter, but for the mixed scalar/vector case we
will end up using the wrong value for later loop iterations.
In the new code we use TCG temporaries and so avoid the bug.
This bug is present for all the multiply-accumulate insns
that operate on short vectors: VMLA, VMLS, VNMLA, VNMLS.
Note 2: the expression used to calculate the next register
number in the vector bank is not in fact correct; we leave
this behaviour unchanged from the old decoder and will
fix this bug later in the series.
Backports commit 266bd25c485597c94209bfdb3891c1d0c573c164 from qemu
Expand out the sequences in the new decoder VLDR/VSTR/VLDM/VSTM trans
functions which perform the memory accesses by going via the TCG
globals cpu_F0s and cpu_F0d, to use local TCG temps instead.
Backports commit 3993d0407dff7233e42f2251db971e126a0497e9 from qemu
Convert the VFP load/store multiple insns to decodetree.
This includes tightening up the UNDEF checking for pre-VFPv3
CPUs which only have D0-D15 : they now UNDEF for any access
to D16-D31, not merely when the smallest register in the
transfer list is in D16-D31.
This conversion does not try to share code between the single
precision and the double precision versions; this looks a bit
duplicative of code, but it leaves the door open for a future
refactoring which gets rid of the use of the "F0" registers
by inlining the various functions like gen_vfp_ld() and
gen_mov_F0_reg() which are hiding "if (dp) { ... } else { ... }"
conditionalisation.
Backports commit fa288de272c5c8a66d5eb683b123706a52bc7ad6 from qemu
Convert the VFP two-register transfer instructions to decodetree
(in the v8 Arm ARM these are the "Advanced SIMD and floating-point
64-bit move" encoding group).
Again, we expand out the sequences involving gen_vfp_msr() and
gen_msr_vfp().
Backports commit 81f681106eabe21c55118a5a41999fb7387fb714 from qemu
Convert the "single-precision" register moves to decodetree:
* VMSR
* VMRS
* VMOV between general purpose register and single precision
Note that the VMSR/VMRS conversions make our handling of
the "should this UNDEF?" checks consistent between the two
instructions:
* VMSR to MVFR0, MVFR1, MVFR2 now UNDEF from EL0
(previously was a nop)
* VMSR to FPSID now UNDEFs from EL0 or if VFPv3 or better
(previously was a nop)
* VMSR to FPINST and FPINST2 now UNDEF if VFPv3 or better
(previously would write to the register, which had no
guest-visible effect because we always UNDEF reads)
We also tighten up the decode: we were previously underdecoding
some SBZ or SBO bits.
The conversion of VMOV_single includes the expansion out of the
gen_mov_F0_vreg()/gen_vfp_mrs() and gen_mov_vreg_F0()/gen_vfp_msr()
sequences into the simpler direct load/store of the TCG temp via
neon_{load,store}_reg32(): we know in the new function that we're
always single-precision, we don't need to use the old-and-deprecated
cpu_F0* TCG globals, and we don't happen to have the declaration of
gen_vfp_msr() and gen_vfp_mrs() at the point in the file where the
new function is.
Backports commit a9ab50011aeda2dd012da99069e078379315ea18 from qemu
Convert the "double-precision" register moves to decodetree:
this covers VMOV scalar-to-gpreg, VMOV gpreg-to-scalar and VDUP.
Note that the conversion process has tightened up a few of the
UNDEF encoding checks: we now correctly forbid:
* VMOV-to-gpr with U:opc1:opc2 == 10x00 or x0x10
* VMOV-from-gpr with opc1:opc2 == 0x10
* VDUP with B:E == 11
* VDUP with Q == 1 and Vn<0> == 1
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
---
The accesses of elements < 32 bits could be improved by doing
direct ld/st of the right size rather than 32-bit read-and-shift
or read-modify-write, but we leave this for later cleanup,
since this series is generally trying to stick to fixing
the decode.
Backports commit 9851ed9269d214c0c6feba960dd14ff09e6c34b4 from qemu
The current VFP code has two different idioms for
loading and storing from the VFP register file:
1 using the gen_mov_F0_vreg() and similar functions,
which load and store to a fixed set of TCG globals
cpu_F0s, CPU_F0d, etc
2 by direct calls to tcg_gen_ld_f64() and friends
We want to phase out idiom 1 (because the use of the
fixed globals is a relic of a much older version of TCG),
but idiom 2 is quite longwinded:
tcg_gen_ld_f64(tmp, cpu_env, vfp_reg_offset(true, reg))
requires us to specify the 64-bitness twice, once in
the function name and once by passing 'true' to
vfp_reg_offset(). There's no guard against accidentally
passing the wrong flag.
Instead, let's move to a convention of accessing 64-bit
registers via the existing neon_load_reg64() and
neon_store_reg64(), and provide new neon_load_reg32()
and neon_store_reg32() for the 32-bit equivalents.
Implement the new functions and use them in the code in
translate-vfp.inc.c. We will convert the rest of the VFP
code as we do the decodetree conversion in subsequent
commits.
Backports commit 160f3b64c5cc4c8a09a1859edc764882ce6ad6bf from qemu
Move the trans_*() functions we've just created from translate.c
to translate-vfp.inc.c. This is pure code motion with no textual
changes (this can be checked with 'git show --color-moved').
Backports commit f7bbb8f31f0761edbf0c64b7ab3c3f49c13612ea from qemu
Convert the VCVTA/VCVTN/VCVTP/VCVTM instructions to decodetree.
trans_VCVT() is temporarily left in translate.c.
Backports commit c2a46a914cd5c38fd0ee57ff0befc1c5bde27bcf from qemu
Convert the VRINTA/VRINTN/VRINTP/VRINTM instructions to decodetree.
Again, trans_VRINT() is temporarily left in translate.c.
Backports commit e3bb599d16e4678b228d80194cee328f894b1ceb from qemu
Convert the VMINNM and VMAXNM instructions to decodetree.
As with VSEL, we leave the trans_VMINMAXNM() function
in translate.c for the moment.
Backports commit f65988a1efdb42f9058db44297591491842e697c from qemu
Convert the VSEL instructions to decodetree.
We leave trans_VSEL() in translate.c for now as this allows
the patch to show just the changes from the old handle_vsel().
In the old code the check for "do D16-D31 exist" was hidden in
the VFP_DREG macro, and assumed that VFPv3 always implied that
D16-D31 exist. In the new code we do the correct ID register test.
This gives identical behaviour for most of our CPUs, and fixes
previously incorrect handling for Cortex-R5F, Cortex-M4 and
Cortex-M33, which all implement VFPv3 or better with only 16
double-precision registers.
Backports commit b3ff4b87b4ae08120a51fe12592725e1dca8a085 from qemu
At the moment our -cpu max for AArch32 supports VFP short-vectors
because we always implement them, even for CPUs which should
not have them. The following commits are going to switch to
using the correct ID-register-check to enable or disable short
vector support, so we need to turn it on explicitly for -cpu max,
because Cortex-A15 doesn't implement it.
We don't enable this for the AArch64 -cpu max, because the v8A
architecture never supports short-vectors.
Backports commit 973751fd798d41402d34f9f705c0c6d1633d0cda from qemu
