Make the Neon helper routines use the correct FP status from
the CPUEnv rather than using a dummy static one. This means
they will correctly handle denormals and NaNs and will set
FPSCR exception bits properly.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Use the global 'env' variable in the helper functions in iwmmxt_helper.c.
This means we don't need to pass env as an argument to them any more.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Use the global 'env' variable in the helper functions in neon_helper.c.
This means we don't need to pass env as an argument to them any more.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Correct the argument and return types for the float<->int conversion helper
functions so that integer arguments and return values are declared as
uint32_t/uint64_t, not float32/float64. This allows us to remove the
hand-rolled functions which were doing bitwise copies between the types
via unions.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Nathan Froyd <froydnj@codesourcery.com>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Use the new softfloat min/max functions to implement the Neon VMAX
and VMIN instructions. This allows us to get the right behaviour
for NaN and negative zero.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Implement ABD by taking the absolute value of the difference
of the operands (as the ARM ARM specifies) rather than by
flipping the order of the operands to the subtract based
on the results of a comparison. The latter approch gives
the wrong answers for some edge cases like negative zero.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Implementing the floating-point versions of VCLE #0 and VCLT #0 by
doing a GT comparison and inverting the result gives the wrong
result if the input is a NaN. Implement as a GT comparison with the
operands swapped instead.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Fix the helper functions implementing the Neon floating point comparison
ops (VCGE, VCGT, VCEQ, VACGT, VACGE) to return the right answer when
one of the values being compared is a NaN.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Use the softfloat make_float32 and float32_val macros to convert between
softfloat's float32 type and raw uint32_t types, rather than private
conversion functions.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Nathan Froyd <froydnj@codesourcery.com>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Move the allocation and freeing of the TCG temp used for the address for
Neon load/store instructions so that we don't allocate the temporary
until we've done enough decoding to know that the instruction is not
an UNDEF pattern; this avoids leaking the TCG temp in these cases.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Fix several bugs in VLD of single element to all lanes:
The "single element to all lanes" form of VLD1 differs from those for
VLD2, VLD3 and VLD4 in that bit 5 indicates whether the loaded element
should be written to one or two Dregs (rather than being a register
stride). Handle this by special-casing VLD1 rather than trying to
have one loop which deals with both VLD1 and 2/3/4.
Handle VLD4.32 with 16 byte alignment specified, rather than UNDEFfing.
UNDEF for the invalid size and alignment combinations.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
The preferred way to create a constant floating point value is to use
make_float32() rather than doing a runtime int32_to_float32().
Convert the code in the VRSQRTS helper to work this way.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Correct the handling of edge cases for the VRECPS instruction:
* this is a Neon instruction so uses the "standard FPSCR value"
* (zero, inf) is a special case which returns 2.0
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
SMUAD and SMLAD are supposed to set the Q bit if the addition of
the two 16x16 multiply products and optional accumulator overflows
considered as a signed value. However we were only doing this check
for the addition of the accumulator, not when adding the products,
with the effect that we were mishandling the edge case where
both inputs are 0x80008000.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Fix the signed modulo arithmetic helpers for the v6media
instructions (SADD8, SSUB8, SADD16, SSUB16, SASX, SSAX) to set
the GE bits correctly (based on the result of the add or subtract
before it is truncated to 16 bits, not after).
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Decode of Thumb load/store was merging together the cases of 'bit 11==0'
(reg+reg LSL imm) and 'bit 11==1' (reg+imm). This happens to work for
valid instruction patterns but meant that we would not UNDEF for the
cases the architecture mandates that we must. Make the decode actually
look at bit 11 as well as [10..8] so that we UNDEF in the right places.
This change also removes what was a spurious unreachable 'case 8',
and correctly frees TCG temporaries on the illegal-insn codepaths.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
All implementations are now the same, and there is only one caller,
so inline the function there.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Blue Swirl <blauwirbel@gmail.com>
Newer ARM kernels try to probe for whether the CPU has hardware breakpoint
support. For this to work QEMU has to implement a minimal set of the cp14
debug registers. The architecture requires v7 cores to implement debug
and so there is no defined way to report its absence; however in practice
returning a zero DBGDIDR (ie with a reserved value for "debug architecture
version") should cause well-written hw debug users to do the right thing.
We also implement DBGDRAR and DBGDSAR as RAZ, indicating no memory mapped
debug components.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Use the new TCG temporary leak debugging facilities to
check that each ARM instruction does not leak temporaries.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
This commit removes the ad-hoc resource leak checking code from
target-arm. This includes replacing all uses of new_tmp() with
tcg_temp_new_i32() and all uses of dead_tmp() with
tcg_temp_free_i32().
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Implement VA->PA translations by cp15-c7 that went through unchanged
previously.
Signed-off-by: Adam Lackorzynski <adam@os.inf.tu-dresden.de>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
The code for Thumb2 ORNS (or negated and set flags) was trashing
a TCG input register which was needed later for use in calculating
flags, with the effect that the carry flag was always set with
the wrong sense. Fix this by using the TCG orc op instead of
separate not and or ops.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Fix two bugs in the translation of the instructions VMOV sa,sb,rx,ry and
VMOV rx,ry,sa,sb (which copy between a pair of ARM core registers and a
pair of VFP single precision registers):
* An incorrect condition meant these instruction patterns were being
treated as load/store multiple, which resulted in the generation
of bad code and a runtime segfault
* The order of the core register pair was reversed so the values would
go to the wrong registers
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
In v7 of the ARM architecture, WFI (wait for interrupt) is a first-class
instruction, but in previous versions this functionality was provided
via a cp15 coprocessor register. Add correct feature checks to the
decoding of the cp15 WFI instructions so that they behave correctly
for newer cores. In particular, the old 0,c7,c8,2 encoding used on
ARM940 has been reused for VA-to-PA translation in v6 and v7.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Adam Lackorzynski <adam@os.inf.tu-dresden.de>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Now use the same algorithm as described in the ARM ARM.
Signed-off-by: Christophe Lyon <christophe.lyon@st.com>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Now use the same algorithm as described in the ARM ARM.
Signed-off-by: Christophe Lyon <christophe.lyon@st.com>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
These two constants will be used by helper functions such as recpe_f32
and rsqrte_f32.
Signed-off-by: Christophe Lyon <christophe.lyon@st.com>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
For Neon shifts by immediate and narrow, correctly handle the case
where the source registers and the destination registers overlap
(the second pass should use the original register contents, not the
results of the first pass).
This includes a refactoring to pull the size check outside the
loop rather than inside, since there is now very little common
code between the size == 3 and size != 3 case.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Pull the code which decodes narrowing operations as being either
signed/unsigned saturate or plain out into its own function.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Correctly handle VQRSHL of unsigned values by a shift count of the
width of the data type or larger, which must be special-cased in the
qrshl_u* helper functions.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Handle the case of signed VQRSHL by a shift count of the width of the
data type or larger, which must be special cased in the qrshl_s*
helper functions.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Call the normal shift helpers instead of the rounding ones.
Signed-off-by: Christophe Lyon <christophe.lyon@st.com>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Fix value returned by signed 8 and 16 bit qrshl helpers
when the result has saturated.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Fix range of shift amounts which always give 0 as result.
Signed-off-by: Christophe Lyon <christophe.lyon@st.com>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Fix handling of unsigned VRSHL.s8 and .s16 right shifts by the type
width.
Signed-off-by: Christophe Lyon <christophe.lyon@st.com>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Correctly handle VRSHL of signed values by a shift count of the
width of the data type or larger, which must be special-cased in the
rshl_s* helper functions.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Handle cases where adding the rounding constant could overflow in Neon
shift instructions: VRSHR, VRSRA, VQRSHRN, VQRSHRUN, VRSHRN.
Signed-off-by: Christophe Lyon <christophe.lyon@st.com>
[peter.maydell@linaro.org: fix handling of large shifts in rshl_s32,
calculate signed saturated value as other functions do.]
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Move the implementation of the Neon VUZP unzip instruction from inline
code to helper functions. (At 50+ TCG ops it was well over the
recommended limit for coding inline.) The helper implementations also
give the correct answers where the inline implementation did not.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Move the implementation of the Neon VUZP unzip instruction from inline
code to helper functions. (At 50+ TCG ops it was well over the
recommended limit for coding inline.) The helper implementations also
fix the handling of the quadword version of the instruction.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
We handle Thumb Neon data processing instructions by converting them
into the equivalent ARM encoding, as the two are very close. However
the ARM encoding should have bit 28 set, not clear. This wasn't causing
any problems because we don't actually look at that bit during decode;
however it is better to do the conversion correctly to avoid problems
later if we add checks to UNDEF on SBZ/SBO bits.
Signed-off-by: Juha Riihimäki <juha.riihimaki@nokia.com>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
For VQDMLSL, negation has to occur after saturation, not before.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Refactor the handling of VQDMULL so that it is dealt with in
its own if() case rather than together with the accumulating
instructions.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Implement VMULL.P8 (the 32x32->64 version of the polynomial multiply
instruction).
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
These two fields were added as a part of ARMv7 support patch (back in
2007), were never used by any code, so can be dropped.
Signed-off-by: Dmitry Eremin-Solenikov <dbaryshkov@gmail.com>
Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Andrzej Zaborowski <andrew.zaborowski@intel.com>
Remove a stray #include <stdio.h> from the middle of neon_helper.c:
it was harmless but pointless since we include stdio.h at the top
of the file anyway.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
The Neon half-precision conversion operations (VCVT.F16.F32 and
VCVT.F32.F16) use ARM standard floating-point arithmetic, unlike
the VFP versions (VCVTB and VCVTT).
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Silence the NaNs that may result from half-precision conversion,
as we do for the other conversions.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Add a float16 type to softfloat, rather than using bits16 directly.
Also add the missing functions float16_is_quiet_nan(),
float16_is_signaling_nan() and float16_maybe_silence_nan(),
which are needed for the float16 conversion routines.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
Peter Maydell
Paolo Bonzini
Adam Lackorzynski
Christophe Lyon
Juha Riihimäki
Dmitry Eremin-Solenikov