Certain Hexagon instructions suppress changes to the result
when the product of fma() is a true zero.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
All uses have been convered to float*_muladd_scalbn.
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
We currently have a flag, float_muladd_halve_result, to scale
the result by 2**-1. Extend this to handle arbitrary scaling.
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Replace the "index" selecting between A and B with a result variable
of the proper type. This improves clarity within the function.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Message-id: 20241203203949.483774-12-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Move the fractional comparison to the end of the
float_2nan_prop_x87 case. This is not required for
any other 2nan propagation rule. Reorganize the
x87 case itself to break out of the switch when the
fractional comparison is not required.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Message-id: 20241203203949.483774-11-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Remember if there was an SNaN, and use that to simplify
float_2nan_prop_s_{ab,ba} to only the snan component.
Then, fall through to the corresponding
float_2nan_prop_{ab,ba} case to handle any remaining
nans, which must be quiet.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Message-id: 20241203203949.483774-10-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Inline pickNaN into its only caller. This makes one assert
redundant with the immediately preceding IF.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Message-id: 20241203203949.483774-9-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
While all indices into val[] should be in [0-2], the mask
applied is two bits. To help static analysis see there is
no possibility of read beyond the end of the array, pad the
array to 4 entries, with the final being (implicitly) NULL.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Message-id: 20241203203949.483774-6-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Assign the pointer return value to 'a' directly,
rather than going through an intermediary index.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Message-id: 20241203203949.483774-5-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Remove "3" as a special case for which and simply
branch to return the desired value.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Message-id: 20241203203949.483774-4-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Inline pickNaNMulAdd into its only caller. This makes
one assert redundant with the immediately preceding IF.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Message-id: 20241203203949.483774-3-richard.henderson@linaro.org
[PMM: keep comment from old code in new location]
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
The new implementation of pickNaNMulAdd() will find it convenient
to know whether at least one of the three arguments to the muladd
was a signaling NaN. We already calculate that in the caller,
so pass it in as a new bool have_snan.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20241202131347.498124-15-peter.maydell@linaro.org
If the target sets default_nan_mode then we're always going to return
the default NaN, and pickNaNMulAdd() no longer has any side effects.
For consistency with pickNaN(), check for default_nan_mode before
calling pickNaNMulAdd().
When we convert pickNaNMulAdd() to allow runtime selection of the NaN
propagation rule, this means we won't have to make the targets which
use default_nan_mode also set a propagation rule.
Since RiscV always uses default_nan_mode, this allows us to remove
its ifdef case from pickNaNMulAdd().
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20241202131347.498124-3-peter.maydell@linaro.org
For IEEE fused multiply-add, the (0 * inf) + NaN case should raise
Invalid for the multiplication of 0 by infinity. Currently we handle
this in the per-architecture ifdef ladder in pickNaNMulAdd().
However, since this isn't really architecture specific we can hoist
it up to the generic code.
For the cases where the infzero test in pickNaNMulAdd was
returning 2, we can delete the check entirely and allow the
code to fall into the normal pick-a-NaN handling, because this
will return 2 anyway (input 'c' being the only NaN in this case).
For the cases where infzero was returning 3 to indicate "return
the default NaN", we must retain that "return 3".
For Arm, this looks like it might be a behaviour change because we
used to set float_flag_invalid | float_flag_invalid_imz only if C is
a quiet NaN. However, it is not, because Arm target code never looks
at float_flag_invalid_imz, and for the (0 * inf) + SNaN case we
already raised float_flag_invalid via the "abc_mask &
float_cmask_snan" check in pick_nan_muladd.
For any target architecture using the "default implementation" at the
bottom of the ifdef, this is a behaviour change but will be fixing a
bug (where we failed to raise the Invalid exception for (0 * inf +
QNaN). The architectures using the default case are:
* hppa
* i386
* sh4
* tricore
The x86, Tricore and SH4 CPU architecture manuals are clear that this
should have raised Invalid; HPPA is a bit vaguer but still seems
clear enough.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20241202131347.498124-2-peter.maydell@linaro.org
This patch is part of a series that moves towards a consistent use of
g_assert_not_reached() rather than an ad hoc mix of different
assertion mechanisms.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Pierrick Bouvier <pierrick.bouvier@linaro.org>
Message-ID: <20240919044641.386068-24-pierrick.bouvier@linaro.org>
Signed-off-by: Thomas Huth <thuth@redhat.com>
Motorola treats denormals with explicit integer bit set as
having unbiased exponent 0, unlike Intel which treats it as
having unbiased exponent 1 (more like all other IEEE formats
that have no explicit integer bit).
Add a flag on FloatFmt to differentiate the behaviour.
Reported-by: Keith Packard <keithp@keithp.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Add versions of float64_to_int* which do not saturate the result.
Reviewed-by: Christoph Muellner <christoph.muellner@vrull.eu>
Tested-by: Christoph Muellner <christoph.muellner@vrull.eu>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Message-Id: <20230527141910.1885950-2-richard.henderson@linaro.org>
logB(0) should raise divideByZero exception from IEEE 754-2008 spec 7.3
Suggested-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Song Gao <gaosong@loongson.cn>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20220930024510.800005-4-gaosong@loongson.cn>
Added the possibility of recalculating a result if it overflows or
underflows, if the result overflow and the rebias bool is true then the
intermediate result should have 3/4 of the total range subtracted from
the exponent. The same for underflow but it should be added to the
exponent of the intermediate number instead.
Signed-off-by: Lucas Mateus Castro (alqotel) <lucas.araujo@eldorado.org.br>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20220805141522.412864-2-lucas.araujo@eldorado.org.br>
Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com>
As the return type is FloatRelation, it's clearer to
use the type for 'cmp' within the function.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Message-Id: <20220401132240.79730-3-richard.henderson@linaro.org>
PowerPC has this flag, and it's easier to compute it here
than after the fact.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20211119160502.17432-8-richard.henderson@linaro.org>
Signed-off-by: Cédric Le Goater <clg@kaod.org>
PowerPC has this flag, and it's easier to compute it here
than after the fact.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20211119160502.17432-7-richard.henderson@linaro.org>
Signed-off-by: Cédric Le Goater <clg@kaod.org>
PowerPC has this flag, and it's easier to compute it here
than after the fact.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20211119160502.17432-6-richard.henderson@linaro.org>
Signed-off-by: Cédric Le Goater <clg@kaod.org>
PowerPC has these flags, and it's easier to compute them here
than after the fact.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20211119160502.17432-5-richard.henderson@linaro.org>
Signed-off-by: Cédric Le Goater <clg@kaod.org>
PowerPC has this flag, and it's easier to compute it here
than after the fact.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20211119160502.17432-4-richard.henderson@linaro.org>
Signed-off-by: Cédric Le Goater <clg@kaod.org>
PowerPC has this flag, and it's easier to compute it here
than after the fact.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20211119160502.17432-3-richard.henderson@linaro.org>
Signed-off-by: Cédric Le Goater <clg@kaod.org>
For "fmax/fmin ft0, ft1, ft2" and if one of the inputs is sNaN,
The original logic:
Return NaN and set invalid flag if ft1 == sNaN || ft2 == sNan.
The alternative path:
Set invalid flag if ft1 == sNaN || ft2 == sNaN.
Return NaN only if ft1 == NaN && ft2 == NaN.
The IEEE 754 spec allows both implementation and some architecture such
as riscv choose different defintions in two spec versions.
(riscv-spec-v2.2 use original version, riscv-spec-20191213 changes to
alternative)
Signed-off-by: Chih-Min Chao <chihmin.chao@sifive.com>
Signed-off-by: Frank Chang <frank.chang@sifive.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20211021160847.2748577-2-frank.chang@sifive.com>
Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
Rename to parts$N_modrem. This was the last use of a lot
of the legacy infrastructure, so remove it as required.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Rename to parts$N_log2. Though this is partly a ruse, since I do not
believe the code will succeed for float128 without work. Which is ok
for now, because we do not need this for more than float32 and float64.
Since berkeley-testfloat-3 doesn't support log2, compare float64_log2
vs the system log2. Fix the errors for inputs near 1.0:
test: 3ff00000000000b0 +0x1.00000000000b0p+0
sf: 3d2fa00000000000 +0x1.fa00000000000p-45
libm: 3d2fbd422b1bd36f +0x1.fbd422b1bd36fp-45
Error in fraction: 32170028290927 ulp
test: 3feec24f6770b100 +0x1.ec24f6770b100p-1
sf: bfad3740d13c9ec0 -0x1.d3740d13c9ec0p-5
libm: bfad3740d13c9e98 -0x1.d3740d13c9e98p-5
Error in fraction: 40 ulp
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
With floatx80_precision_x, the rounding happens across
the break between words. Notice this case with
frac_lsb = round_mask + 1 -> 0
and check the bits in frac_hi as needed.
In addition, since frac_shift == 0, we won't implicitly clear
round_mask via the right-shift, so explicitly clear those bits.
This fixes rounding for floatx80_precision_[sd].
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Remove frac_lsb, frac_lsbm1, roundeven_mask. Compute
these from round_mask in parts$N_uncanon_normal.
With floatx80, round_mask will not be tied to frac_shift.
Everything else is easily computable.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
We will need to treat the non-normal cases of floatx80 specially,
so split out the normal case that we can reuse.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Rename to parts$N_sqrt.
Reimplement float128_sqrt with FloatParts128.
Reimplement with the inverse sqrt newton-raphson algorithm from musl.
This is significantly faster than even the berkeley sqrt n-r algorithm,
because it does not use division instructions, only multiplication.
Ordinarily, changing algorithms at the same time as migrating code is
a bad idea, but this is the only way I found that didn't break one of
the routines at the same time.
Tested-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Rename to parts$N_scalbn.
Reimplement float128_scalbn with FloatParts128.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Rename to parts$N_compare. Rename all of the intermediate
functions to ftype_do_compare. Rename the hard-float functions
to ftype_hs_compare. Convert float128 to FloatParts128.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Rename to parts$N_minmax. Combine 3 bool arguments to a bitmask.
Introduce ftype_minmax functions as a common optimization point.
Fold bfloat16 expansions into the same macro as the other types.
Reviewed-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Rename to parts$N_uint_to_float.
Reimplement uint64_to_float128 with FloatParts128.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Rename to parts$N_sint_to_float.
Reimplement int{32,64}_to_float128 with FloatParts128.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Rename to parts$N_float_to_uint. Reimplement
float128_to_uint{32,64}{_round_to_zero} with FloatParts128.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
For Arm BFDOT and BFMMLA, we need a version of round-to-odd
that overflows to infinity, instead of the max normal number.
Cc: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20210525225817.400336-6-richard.henderson@linaro.org
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Rename to parts$N_float_to_sint. Reimplement
float128_to_int{32,64}{_round_to_zero} with FloatParts128.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
At the same time, convert to pointers, split out
parts$N_round_to_int_normal, define a macro for
parts_round_to_int using QEMU_GENERIC.
This necessarily meant some rearrangement to the
rount_to_{,u}int_and_pack routines, so go ahead and
convert to parts_round_to_int_normal, which in turn
allows cleaning up of the raised exception handling.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Rename to parts$N_div.
Implement float128_div with FloatParts128.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Rename to parts$N_muladd.
Implement float128_muladd with FloatParts128.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Rename to parts$N_mul.
Reimplement float128_mul with FloatParts128.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
In preparation for implementing multiple sizes. Rename to parts_addsub,
split out parts_add/sub_normal for future reuse with muladd.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
At the same time, convert to pointers, renaming to parts$N_uncanon,
and define a macro for parts_uncanon using QEMU_GENERIC.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
At the same time, convert to pointers, rename to parts$N_canonicalize
and define a macro for parts_canonicalize using QEMU_GENERIC.
Rearrange the cases to recognize float_class_normal as
early as possible.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
At the same time, convert to pointers, rename to pick_nan_muladd$N
and define a macro for pick_nan_muladd using QEMU_GENERIC.
Reviewed-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
