Now the emulation is complete, we can enable FEAT_RPRES for the 'max'
CPU type.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the increased precision variation of FRSQRTE. In the
pseudocode this corresponds to the handling of the
"increasedprecision" boolean in the FPRSqrtEstimate() and
RecipSqrtEstimate() functions.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the increased precision variation of FRECPE. In the
pseudocode this corresponds to the handling of the
"increasedprecision" boolean in the FPRecipEstimate() and
RecipEstimate() functions.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
FEAT_RPRES implements an "increased precision" variant of the single
precision FRECPE and FRSQRTE instructions from an 8 bit to a 12
bit mantissa. This applies only when FPCR.AH == 1. Note that the
halfprec and double versions of these insns retain the 8 bit
precision regardless.
In this commit we add all the plumbing to make these instructions
call a new helper function when the increased-precision is in
effect. In the following commit we will provide the actual change
in behaviour in the helpers.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Now that we have completed the handling for FPCR.{AH,FIZ,NEP}, we
can enable FEAT_AFP for '-cpu max', and document that we support it.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Handle FPCR.AH's requirement to not negate the sign of a NaN in SVE
FMLSL (indexed), using the usual trick of negating by XOR when AH=0
and by muladd flags when AH=1.
Since we have the CPUARMState* in the helper anyway, we can
look directly at env->vfp.fpcr and don't need toa pass in the
FPCR.AH value via the SIMD data word.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20250129013857.135256-33-richard.henderson@linaro.org
[PMM: tweaked commit message]
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Handle FPCR.AH's requirement to not negate the sign of a NaN in SVE
FMLSL (indexed), using the usual trick of negating by XOR when AH=0
and by muladd flags when AH=1.
Since we have the CPUARMState* in the helper anyway, we can
look directly at env->vfp.fpcr and don't need toa pass in the
FPCR.AH value via the SIMD data word.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20250129013857.135256-32-richard.henderson@linaro.org
[PMM: commit message tweaked]
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Handle FPCR.AH's requirement to not negate the sign of a NaN
in FMLSL by element and vector, using the usual trick of
negating by XOR when AH=0 and by muladd flags when AH=1.
Since we have the CPUARMState* in the helper anyway, we can
look directly at env->vfp.fpcr and don't need toa pass in the
FPCR.AH value via the SIMD data word.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20250129013857.135256-31-richard.henderson@linaro.org
[PMM: commit message tweaked]
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
The negation step in SVE FCMLA mustn't negate a NaN when FPCR.AH is
set. Use the same approach as we did for A64 FCMLA of passing in
FPCR.AH and using it to select whether to negate by XOR or by the
muladd negate_product flag.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20250129013857.135256-28-richard.henderson@linaro.org
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
The negation step in FCMLA by index mustn't negate a NaN when
FPCR.AH is set. Use the same approach as vector FCMLA of
passing in FPCR.AH and using it to select whether to negate
by XOR or by the muladd negate_product flag.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20250129013857.135256-27-richard.henderson@linaro.org
[PMM: Expanded commit message]
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
The negation step in FCMLA mustn't negate a NaN when FPCR.AH
is set. Handle this by passing FPCR.AH to the helper via the
SIMD data field, and use this to select whether to do the
negation via XOR or via the muladd negate_product flag.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20250129013857.135256-26-richard.henderson@linaro.org
[PMM: Expanded commit message]
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
The negation step in the SVE FTMAD insn mustn't negate a NaN when
FPCR.AH is set. Pass FPCR.AH to the helper via the SIMD data field,
so we can select the correct behaviour.
Because the operand is known to be negative, negating the operand
is the same as taking the absolute value. Defer this to the muladd
operation via flags, so that it happens after NaN detection, which
is correct for FPCR.AH.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
The negation step in the SVE FTSSEL insn mustn't negate a NaN when
FPCR.AH is set. Pass FPCR.AH to the helper via the SIMD data field
and use that to determine whether to do the negation.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Handle the FPCR.AH "don't negate the sign of a NaN" semantics fro the
SVE FMLS (vector) insns, by providing new helpers for the AH=1 case
which end up passing fpcr_ah = true to the do_fmla_zpzzz_* functions
that do the work.
The float*_muladd functions have a flags argument that can
perform optional negation of various operand. We don't use
that for "normal" arm fmla, because the muladd flags are not
applied when an input is a NaN. But since FEAT_AFP does not
negate NaNs, this behaviour is exactly what we need.
The non-AH helpers pass in a zero flags argument and control the
negation via the neg1 and neg3 arguments; the AH helpers always pass
in neg1 and neg3 as zero and control the negation via the flags
argument. This allows us to avoid conditional branches within the
inner loop.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Handle the FPCR.AH "don't negate the sign of a NaN" semantics
in FMLS (vector), by implementing a new set of helpers for
the AH=1 case.
The float_muladd_negate_product flag produces the same result
as negating either of the multiplication operands, assuming
neither of the operands are NaNs. But since FEAT_AFP does not
negate NaNs, this behaviour is exactly what we need.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Handle the FPCR.AH "don't negate the sign of a NaN" semantics in FMLS
(indexed). We do this by creating 6 new helpers, which allow us to
do the negation either by XOR (for AH=0) or by muladd flags
(for AH=1).
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
[PMM: Mostly from RTH's patch; error in index order into fns[][]
fixed]
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Handle the FPCR.AH "don't negate the sign of a NaN" semantics
in the vector versions of FRECPS and FRSQRTS, by implementing
new vector wrappers that call the _ah_ scalar helpers.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Handle the FPCR.AH semantics that we do not change the sign of an
input NaN in the FRECPS and FRSQRTS scalar insns, by providing
new helper functions that do the CHS part of the operation
differently.
Since the extra helper functions would be very repetitive if written
out longhand, we condense them and the existing non-AH helpers into
being emitted via macros.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
The negation steps in FCADD must honour FPCR.AH's "don't change the
sign of a NaN" semantics. Implement this by encoding FPCR.AH into
the SIMD data field passed to the helper and using that to decide
whether to negate the values.
The construction of neg_imag and neg_real were done to make it easy
to apply both in parallel with two simple logical operations. This
changed with FPCR.AH, which is more complex than that. Switch to
an approach closer to the pseudocode, where we extract the rot
parameter from the SIMD data word and negate the appropriate
input value.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
The negation steps in FCADD must honour FPCR.AH's "don't change the
sign of a NaN" semantics. Implement this in the same way we did for
the base ASIMD FCADD, by encoding FPCR.AH into the SIMD data field
passed to the helper and using that to decide whether to negate the
values.
The construction of neg_imag and neg_real were done to make it easy
to apply both in parallel with two simple logical operations. This
changed with FPCR.AH, which is more complex than that. Switch to
an approach that follows the pseudocode more closely, by extracting
the 'rot=1' parameter from the SIMD data field and changing the
sign of the appropriate input value.
Note that there was a naming issue with neg_imag and neg_real.
They were named backward, with neg_imag being non-zero for rot=1,
and vice versa. This was combined with reversed usage within the
loop, so that the negation in the end turned out correct.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Make the SVE FABD insn honour the FPCR.AH "don't negate the sign
of a NaN" semantics.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Make SVE FABS honour the FPCR.AH "don't negate the sign of a NaN"
semantics.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Make SVE FNEG honour the FPCR.AH "don't negate the sign of a NaN"
semantics.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Split the handling of vector FABD so that it calls a different set
of helpers when FPCR.AH is 1, which implement the "no negation of
the sign of a NaN" semantics.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
FPCR.AH == 1 mandates that taking the absolute value of a NaN should
not change its sign bit. This means we can no longer use
gen_vfp_abs*() everywhere but must instead generate slightly more
complex code when FPCR.AH is set.
Implement these semantics for scalar FABS and FABD. This change also
affects all other instructions whose psuedocode calls FPAbs(); we
will extend the change to those instructions in following commits.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
FPCR.AH == 1 mandates that negation of a NaN value should not flip
its sign bit. This means we can no longer use gen_vfp_neg*()
everywhere but must instead generate slightly more complex code when
FPCR.AH is set.
Make this change for the scalar FNEG and for those places in
translate-a64.c which were previously directly calling
gen_vfp_neg*().
This change in semantics also affects any other instruction whose
pseudocode calls FPNeg(); in following commits we extend this
change to the other affected instructions.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the FPCR.AH semantics for the SVE FMAX and FMIN
operations that take two vector operands.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the FPCR.AH semantics for the SVE FMAX and FMIN operations
that take an immediate as the second operand.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the FPCR.AH semantics for the SVE FMAXV and FMINV
vector-reduction-to-scalar max/min operations.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the FPCR.AH semantics for the pairwise floating
point minimum/maximum insns FMINP and FMAXP.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the FPCR.AH semantics for FMAXV and FMINV. These are the
"recursively reduce all lanes of a vector to a scalar result" insns;
we just need to use the _ah_ helper for the reduction step when
FPCR.AH == 1.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Implement the FPCR.AH == 1 semantics for vector FMIN/FMAX, by
creating new _ah_ versions of the gvec helpers which invoke the
scalar fmin_ah and fmax_ah helpers on each element.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
When FPCR.AH == 1, floating point FMIN and FMAX have some odd special
cases:
* comparing two zeroes (even of different sign) or comparing a NaN
with anything always returns the second argument (possibly
squashed to zero)
* denormal outputs are not squashed to zero regardless of FZ or FZ16
Implement these semantics in new helper functions and select them at
translate time if FPCR.AH is 1 for the scalar FMAX and FMIN insns.
(We will convert the other FMAX and FMIN insns in subsequent
commits.)
Note that FMINNM and FMAXNM are not affected.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
do_fp3_scalar_idx() is used only for the FMUL and FMULX scalar by
element instructions; these both need to merge the result with the Rn
register when FPCR.NEP is set.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Unlike the other users of do_2misc_narrow_scalar(), FCVTXN (scalar)
is always double-to-single and must honour FPCR.NEP. Implement this
directly in a trans function rather than using
do_2misc_narrow_scalar().
We still need gen_fcvtxn_sd() and the f_scalar_fcvtxn[] array for
the FCVTXN (vector) insn, so we move those down in the file to
where they are used.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Handle FPCR.NEP merging for scalar FABS and FNEG; this requires
an extra parameter to do_fp1_scalar_int(), since FMOV scalar
does not have the merging behaviour.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Handle FPCR.NEP in the operations handled by do_cvtf_scalar().
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Handle FPCR.NEP for the 1-input scalar operations.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Currently we implement BFCVT scalar via do_fp1_scalar(). This works
even though BFCVT is a narrowing operation from 32 to 16 bits,
because we can use write_fp_sreg() for float16. However, FPCR.NEP
support requires that we use write_fp_hreg_merging() for float16
outputs, so we can't continue to borrow the non-narrowing
do_fp1_scalar() function for this. Split out trans_BFCVT_s()
into its own implementation that honours FPCR.NEP.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Handle FPCR.NEP for the 3-input scalar operations which use
do_fmla_scalar_idx() and do_fmadd(), by making them call the
appropriate write_fp_*reg_merging() functions.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
For FEAT_AFP's FPCR.NEP bit, we need to programmatically change the
behaviour of the writeback of the result for most SIMD scalar
operations, so that instead of zeroing the upper part of the result
register it merges the upper elements from one of the input
registers.
Provide new functions write_fp_*reg_merging() which can be used
instead of the existing write_fp_*reg() functions when we want this
"merge the result with one of the input registers if FPCR.NEP is
enabled" handling, and use them in do_fp3_scalar_with_fpsttype().
Note that (as documented in the description of the FPCR.NEP bit)
which input register to use as the merge source varies by
instruction: for these 2-input scalar operations, the comparison
instructions take from Rm, not Rn.
We'll extend this to also provide the merging behaviour for
the remaining scalar insns in subsequent commits.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
For FEAT_AFP, we want to emit different code when FPCR.NEP is set, so
that instead of zeroing the high elements of a vector register when
we write the output of a scalar operation to it, we instead merge in
those elements from one of the source registers. Since this affects
the generated code, we need to put FPCR.NEP into the TBFLAGS.
FPCR.NEP is treated as 0 when in streaming SVE mode and FEAT_SME_FA64
is not implemented or not enabled; we can implement this logic in
rebuild_hflags_a64().
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
When FPCR.AH is 1, use FPST_FPCR_AH for:
* AdvSIMD BFMLALB, BFMLALT
* SVE BFMLALB, BFMLALT, BFMLSLB, BFMLSLT
so that they get the required behaviour changes.
We do this by making gen_gvec_op4_fpst() take an ARMFPStatusFlavour
rather than a bool is_fp16; existing callsites now select
FPST_FPCR_F16_A64 vs FPST_FPCR_A64 themselves rather than passing in
the boolean.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
When FPCR.AH is 1, use FPST_FPCR_AH for:
* AdvSIMD BFCVT, BFCVTN, BFCVTN2
* SVE BFCVT, BFCVTNT
so that they get the required behaviour changes.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
For the instructions FRECPE, FRECPS, FRECPX, FRSQRTE, FRSQRTS, use
FPST_FPCR_AH or FPST_FPCR_AH_F16 when FPCR.AH is 1, so that they get
the required behaviour changes.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
When FPCR.AH is 1, the behaviour of some instructions changes:
* AdvSIMD BFCVT, BFCVTN, BFCVTN2, BFMLALB, BFMLALT
* SVE BFCVT, BFCVTNT, BFMLALB, BFMLALT, BFMLSLB, BFMLSLT
* SME BFCVT, BFCVTN, BFMLAL, BFMLSL (these are all in SME2 which
QEMU does not yet implement)
* FRECPE, FRECPS, FRECPX, FRSQRTE, FRSQRTS
The behaviour change is:
* the instructions do not update the FPSR cumulative exception flags
* trapped floating point exceptions are disabled (a no-op for QEMU,
which doesn't implement FPCR.{IDE,IXE,UFE,OFE,DZE,IOE})
* rounding is always round-to-nearest-even regardless of FPCR.RMode
* denormalized inputs and outputs are always flushed to zero, as if
FPCR.{FZ,FIZ} is {1,1}
* FPCR.FZ16 is still honoured for half-precision inputs
(See the Arm ARM DDI0487L.a section A1.5.9.)
We can provide all these behaviours with another pair of float_status fields
which we use only for these insns, when FPCR.AH is 1. These float_status
fields will always have:
* flush_to_zero and flush_inputs_to_zero set for the non-F16 field
* rounding mode set to round-to-nearest-even
and so the only FPCR fields they need to honour are DN and FZ16.
In this commit we only define the new fp_status fields and give them
the required behaviour when FPSR is updated. In subsequent commits
we will arrange to use this new fp_status field for the instructions
that should be affected by FPCR.AH in this way.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
We are going to need to generate different code in some cases when
FPCR.AH is 1. For example:
* Floating point neg and abs must not flip the sign bit of NaNs
* some insns (FRECPE, FRECPS, FRECPX, FRSQRTE, FRSQRTS, and various
BFCVT and BFM bfloat16 ops) need to use a different float_status
to the usual one
Encode FPCR.AH into the A64 tbflags, so we can refer to it at
translate time.
Because we now have a bit in FPCR that affects codegen, we can't mark
the AArch64 FPCR register as being SUPPRESS_TB_END any more; writes
to it will now end the TB and trigger a regeneration of hflags.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
When FPCR.AH = 1, some of the cumulative exception flags in the FPSR
behave slightly differently for A64 operations:
* IDC is set when a denormal input is used without flushing
* IXC (Inexact) is set when an output denormal is flushed to zero
Update vfp_get_fpsr_from_host() to do this.
Note that because half-precision operations never set IDC, we now
need to add float_flag_input_denormal_used to the set we mask out of
fp_status_f16_a64.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
When FPCR.AH is set, various behaviours of AArch64 floating point
operations which are controlled by softfloat config settings change:
* tininess and ftz detection before/after rounding
* NaN propagation order
* result of 0 * Inf + NaN
* default NaN value
When the guest changes the value of the AH bit, switch these config
settings on the fp_status_a64 and fp_status_f16_a64 float_status
fields.
This requires us to make the arm_set_default_fp_behaviours() function
global, since we now need to call it from cpu.c and vfp_helper.c; we
move it to vfp_helper.c so it can be next to the new
arm_set_ah_fp_behaviours().
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Part of FEAT_AFP is the new control bit FPCR.FIZ. This bit affects
flushing of single and double precision denormal inputs to zero for
AArch64 floating point instructions. (For half-precision, the
existing FPCR.FZ16 control remains the only one.)
FPCR.FIZ differs from FPCR.FZ in that if we flush an input denormal
only because of FPCR.FIZ then we should *not* set the cumulative
exception bit FPSR.IDC.
FEAT_AFP also defines that in AArch64 the existing FPCR.FZ only
applies when FPCR.AH is 0.
We can implement this by setting the "flush inputs to zero" state
appropriately when FPCR is written, and by not reflecting the
float_flag_input_denormal status flag into FPSR reads when it is the
result only of FPSR.FIZ.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
