Now we have removed all the target-specifics from the softfloat code,
we can switch to building it once for the whole system rather than
once per target.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Tested-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20250224111524.1101196-13-peter.maydell@linaro.org
Message-id: 20250217125055.160887-11-peter.maydell@linaro.org
We happen to know that for the PPC target the FP status flags (and in
particular float_flag_inexact) will always be cleared before a
floating point operation, and so can_use_fpu() will always return
false. So we speed things up a little by forcing QEMU_NO_HARDFLOAT
to true on that target.
We would like to build softfloat once for all targets; that means
removing target-specific ifdefs. Remove the check for TARGET_PPC;
this won't change behaviour because can_use_fpu() will see that
float_flag_inexact is clear and take the softfloat path anyway.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20250224111524.1101196-12-peter.maydell@linaro.org
Message-id: 20250217125055.160887-10-peter.maydell@linaro.org
Currently we compile-time set an 'm68k_denormal' flag in the FloatFmt
for floatx80 for m68k. This controls our handling of what the Intel
documentation calls a "pseudo-denormal": a value where the exponent
field is zero and the explicit integer bit is set.
For x86, the x87 FPU is supposed to accept a pseudo-denormal as
input, but never generate one on output. For m68k, these values are
permitted on input and may be produced on output.
Replace the flag in the FloatFmt with a flag indicating whether the
float format has an explicit bit (which will be true for floatx80 for
all targets, and false for every other float type). Then we can gate
the handling of these pseudo-denormals on the setting of a
floatx80_behaviour flag.
As far as I can see from the code we don't actually handle the
x86-mandated "accept on input but don't generate" behaviour, because
the handling in partsN(canonicalize) looked at fmt->m68k_denormal.
So I have added TODO comments to that effect.
This commit doesn't change any behaviour for any target.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Message-id: 20250224111524.1101196-9-peter.maydell@linaro.org
Message-id: 20250217125055.160887-7-peter.maydell@linaro.org
The definition of which floatx80 encodings are invalid is
target-specific. Currently we handle this with an ifdef, but we
would like to defer this decision to runtime. In preparation, pass a
float_status argument to floatx80_invalid_encoding().
We will change the implementation from ifdef to looking at
the status argument in the following commit.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Message-id: 20250224111524.1101196-7-peter.maydell@linaro.org
Currently we hardcode at compile time whether the floatx80 default
Infinity value has the explicit integer bit set or not (x86 sets it;
m68k does not). To be able to compile softfloat once for all targets
we'd like to move this setting to runtime.
Define a new FloatX80Behaviour enum which is a set of flags that
define the target's floatx80 handling. Initially we define just one
flag, for whether the default Infinity has the Integer bit set or
not, but we will expand this in future commits to cover the other
floatx80 target specifics that we currently make compile-time
settings.
Define a new function floatx80_default_inf() which returns the
appropriate default Infinity value of the given sign, and use it in
the code that was previously directly using the compile-time constant
floatx80_infinity_{low,high} values when packing an infinity into a
floatx80.
Since floatx80 is highly unlikely to be supported in any new
architecture, and the existing code is generally written as "default
to like x87, with an ifdef for m68k", we make the default value for
the floatx80 behaviour flags be "what x87 does". This means we only
need to change the m68k target to specify the behaviour flags.
(Other users of floatx80 are the Arm NWFPE emulation, which is
obsolete and probably not actually doing the right thing anyway, and
the PPC xsrqpxp insn. Making the default be "like x87" avoids our
needing to review and test for behaviour changes there.)
We will clean up the remaining uses of the floatx80_infinity global
constant in subsequent commits.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Message-id: 20250224111524.1101196-2-peter.maydell@linaro.org
Message-id: 20250217125055.160887-2-peter.maydell@linaro.org
For the x86 and the Arm FEAT_AFP semantics, we need to be able to
tell the target code that the FPU operation has used an input
denormal. Implement this; when it happens we set the new
float_flag_denormal_input_used.
Note that we only set this when an input denormal is actually used by
the operation: if the operation results in Invalid Operation or
Divide By Zero or the result is a NaN because some other input was a
NaN then we never needed to look at the input denormal and do not set
denormal_input_used.
We mostly do not need to adjust the hardfloat codepaths to deal with
this flag, because almost all hardfloat operations are already gated
on the input not being a denormal, and will fall back to softfloat
for a denormal input. The only exception is the comparison
operations, where we need to add the check for input denormals, which
must now fall back to softfloat where they did not before.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Currently in softfloat we canonicalize input denormals and so the
code that implements floating point operations does not need to care
whether the input value was originally normal or denormal. However,
both x86 and Arm FEAT_AFP require that an exception flag is set if:
* an input is denormal
* that input is not squashed to zero
* that input is actually used in the calculation (e.g. we
did not find the other input was a NaN)
So we need to track that the input was a non-squashed denormal. To
do this we add a new value to the FloatClass enum. In this commit we
add the value and adjust the code everywhere that looks at FloatClass
values so that the new float_class_denormal behaves identically to
float_class_normal. We will add the code that does the "raise a new
float exception flag if an input was an unsquashed denormal and we
used it" in a subsequent commit.
There should be no behavioural change in this commit.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Our float_flag_output_denormal exception flag is set when
the fpu code flushes an output denormal to zero. Rename
it to float_flag_output_denormal_flushed:
* this keeps it parallel with the flag for flushing
input denormals, which we just renamed
* it makes it clearer that it doesn't mean "set when
the output is a denormal"
Commit created with
for f in `git grep -l float_flag_output_denormal`; do sed -i -e 's/float_flag_output_denormal/float_flag_output_denormal_flushed/' $f; done
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20250124162836.2332150-21-peter.maydell@linaro.org
Our float_flag_input_denormal exception flag is set when the fpu code
flushes an input denormal to zero. This is what many guest
architectures (eg classic Arm behaviour) require, but it is not the
only donarmal-related reason we might want to set an exception flag.
The x86 behaviour (which we do not currently model correctly) wants
to see an exception flag when a denormal input is *not* flushed to
zero and is actually used in an arithmetic operation. Arm's FEAT_AFP
also wants these semantics.
Rename float_flag_input_denormal to float_flag_input_denormal_flushed
to make it clearer when it is set and to allow us to add a new
float_flag_input_denormal_used next to it for the x86/FEAT_AFP
semantics.
Commit created with
for f in `git grep -l float_flag_input_denormal`; do sed -i -e 's/float_flag_input_denormal/float_flag_input_denormal_flushed/' $f; done
and manual editing of softfloat-types.h and softfloat.c to clean
up the indentation afterwards and to fix a comment which wasn't
using the full name of the flag.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20250124162836.2332150-20-peter.maydell@linaro.org
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>
Unpacking and repacking the parts may be slightly more work
than we did before, but we get to reuse more code. For a
code path handling exceptional values, this is an improvement.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-id: 20241203203949.483774-8-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
This function is part of the public interface and
is not "specialized" to any target in any way.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Message-id: 20241203203949.483774-7-richard.henderson@linaro.org
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
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>
We missed these functions when upstreaming the bfloat16 support.
Signed-off-by: LIU Zhiwei <zhiwei_liu@linux.alibaba.com>
Message-Id: <20230531065458.2082-1-zhiwei_liu@linux.alibaba.com>
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>
The float32_exp2 function is computing wrong exponent of 2.
For example, with the following set of values {0.1, 2.0, 2.0, -1.0},
the expected output would be {1.071773, 4.000000, 4.000000, 0.500000}.
Instead, the function is computing {1.119102, 3.382044, 3.382044, -0.191022}
Looking at the code, the float32_exp2() attempts to do this
2 3 4 5 n
x x x x x x x
e = 1 + --- + --- + --- + --- + --- + ... + --- + ...
1! 2! 3! 4! 5! n!
But because of the typo it ends up doing
x x x x x x x
e = 1 + --- + --- + --- + --- + --- + ... + --- + ...
1! 2! 3! 4! 5! n!
This is because instead of the xnp which holds the numerator, parts_muladd
is using the xp which is just 'x'. Commit '572c4d862f' refactored this
function, and mistakenly used xp instead of xnp.
Cc: qemu-stable@nongnu.org
Fixes: 572c4d862f "softfloat: Convert float32_exp2 to FloatParts"
Partially-Resolves: https://gitlab.com/qemu-project/qemu/-/issues/1623
Reported-By: Luca Barbato (https://gitlab.com/lu-zero)
Signed-off-by: Shivaprasad G Bhat <sbhat@linux.ibm.com>
Signed-off-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Message-Id: <168304110865.537992.13059030916325018670.stgit@localhost.localdomain>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
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>
Since the caller, partsN_compare, is now exclusively
using FloatRelation, it's clearer to use it here too.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Message-Id: <20220401132240.79730-4-richard.henderson@linaro.org>
The declaration used 'int', while the definition used 'FloatRelation'.
This should have resulted in a compiler error, but mysteriously didn't.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Message-Id: <20220401132240.79730-2-richard.henderson@linaro.org>
Based on parts_sint_to_float, implements int128_to_float128 to convert a
signed 128-bit value received through an Int128 argument.
Signed-off-by: Matheus Ferst <matheus.ferst@eldorado.org.br>
Message-Id: <20220330175932.6995-5-matheus.ferst@eldorado.org.br>
Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Based on parts_uint_to_float, implements uint128_to_float128 to convert
an unsigned 128-bit value received through an Int128 argument.
Signed-off-by: Matheus Ferst <matheus.ferst@eldorado.org.br>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20220330175932.6995-4-matheus.ferst@eldorado.org.br>
Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com>
These variants take a float64 as input, compute the result to
infinite precision (as we do with FloatParts), round the result
to the precision and dynamic range of float32, and then return
the result in the format of float64.
This is the operation PowerPC requires for its float32 operations.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20211119160502.17432-28-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: <20211119160502.17432-8-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>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Thomas Huth <thuth@redhat.com>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Message-Id: <20210614233143.1221879-3-richard.henderson@linaro.org>
Typo in the conversion to FloatParts64.
Fixes: 572c4d862f
Fixes: Coverity CID 1457457
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Message-Id: <20210607223812.110596-1-richard.henderson@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
For the normal case of no additional scaling, this reduces the
profile contribution of int64_to_float64 to the testcase in the
linked issue from 0.81% to 0.04%.
Resolves: https://gitlab.com/qemu-project/qemu/-/issues/134
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
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$N_log2. Though this is partly a ruse, since I do not
believe the code will succeed for float128 without work. Which is ok
for now, because we do not need this for more than float32 and float64.
Since berkeley-testfloat-3 doesn't support log2, compare float64_log2
vs the system log2. Fix the errors for inputs near 1.0:
test: 3ff00000000000b0 +0x1.00000000000b0p+0
sf: 3d2fa00000000000 +0x1.fa00000000000p-45
libm: 3d2fbd422b1bd36f +0x1.fbd422b1bd36fp-45
Error in fraction: 32170028290927 ulp
test: 3feec24f6770b100 +0x1.ec24f6770b100p-1
sf: bfad3740d13c9ec0 -0x1.d3740d13c9ec0p-5
libm: bfad3740d13c9e98 -0x1.d3740d13c9e98p-5
Error in fraction: 40 ulp
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Keep the intermediate results in FloatParts instead of
converting back and forth between float64. Use muladd
instead of separate mul+add.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
This is the last use of commonNaNT and all of the routines
that use it, so remove all of them for Werror.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Since this is the first such, this includes all of the
packing and unpacking routines as well.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Use an enumeration instead of raw 32/64/80 values.
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Remove frac_lsb, frac_lsbm1, roundeven_mask. Compute
these from round_mask in parts$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>
