This uses the instruction doubling machinery to convert conditional
branch instructions that update both CTR and LR (e.g., bdnzl, bdnzlrl)
into two instructions, of which the first updates CTR and determines
whether the branch is taken, and the second updates LR and does the
redirect if necessary.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This uses the instruction-doubling machinery to send load with update
instructions down to loadstore1 as two separate ops, rather than
one op with two destinations. This will help to simplify the value
tracking mechanisms.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This implements a cache in fetch1, where each entry stores the address
of a simple branch instruction (b or bc) and the target of the branch.
When fetching sequentially, if the address being fetched matches the
cache entry, then fetching will be redirected to the branch target.
The cache has 1024 entries and is direct-mapped, i.e. indexed by bits
11..2 of the NIA.
The bus from execute1 now carries information about taken and
not-taken simple branches, which fetch1 uses to update the cache.
The cache entry is updated for both taken and not-taken branches, with
the valid bit being set if the branch was taken and cleared if the
branch was not taken.
If fetching is redirected to the branch target then that goes down the
pipe as a predicted-taken branch, and decode1 does not do any static
branch prediction. If fetching is not redirected, then the next
instruction goes down the pipe as normal and decode1 does its static
branch prediction.
In order to make timing, the lookup of the cache is pipelined, so on
each cycle the cache entry for the current NIA + 8 is read. This
means that after a redirect (from decode1 or execute1), only the third
and subsequent sequentially-fetched instructions will be able to be
predicted.
This improves the coremark value on the Arty A7-100 from about 180 to
about 190 (more than 5%).
The BTC is optional. Builds for the Artix 7 35-T part have it off by
default because the extra ~1420 LUTs it takes mean that the design
doesn't fit on the Arty A7-35 board.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This does the addition of NIA plus the branch offset from the
instruction after a clock edge, in order to ease timing, as the path
from the icache RAM through the adder in decode1 to the NIA register
in fetch1 was showing up as a critical path.
This adds one extra cycle of latency when redirecting fetch because of
a predicted-taken branch.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This makes it simpler to work out when to deliver a FPU unavailable
interrupt. This also means we can get rid of the OP_FPLOAD and
OP_FPSTORE insn_type values.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This implements the lq, stq, lqarx and stqcx. instructions.
These instructions all access two consecutive GPRs; for example the
"lq %r6,0(%r3)" instruction will load the doubleword at the address
in R3 into R7 and the doubleword at address R3 + 8 into R6. To cope
with having two GPR sources or destinations, the instruction gets
repeated at the decode2 stage, that is, for each lq/stq/lqarx/stqcx.
coming in from decode1, two instructions get sent out to execute1.
For these instructions, the RS or RT register gets modified on one
of the iterations by setting the LSB of the register number. In LE
mode, the first iteration uses RS|1 or RT|1 and the second iteration
uses RS or RT. In BE mode, this is done the other way around. In
order for decode2 to know what endianness is currently in use, we
pass the big_endian flag down from icache through decode1 to decode2.
This is always in sync with what execute1 is using because only rfid
or an interrupt can change MSR[LE], and those operations all cause
a flush and redirect.
There is now an extra column in the decode tables in decode1 to
indicate whether the instruction needs to be repeated. Decode1 also
enforces the rule that lq with RT = RT and lqarx with RA = RT or
RB = RT are illegal.
Decode2 now passes a 'repeat' flag and a 'second' flag to execute1,
and execute1 passes them on to loadstore1. The 'repeat' flag is set
for both iterations of a repeated instruction, and 'second' is set
on the second iteration. Execute1 does not take asynchronous or
trace interrupts on the second iteration of a repeated instruction.
Loadstore1 uses 'next_addr' for the second iteration of a repeated
load/store so that we access the second doubleword of the memory
operand. Thus loadstore1 accesses the doublewords in increasing
memory order. For 16-byte loads this means that the first iteration
writes GPR RT|1. It is possible that RA = RT|1 (this is a legal
but non-preferred form), meaning that if the memory operand was
misaligned, the first iteration would overwrite RA but then the
second iteration might take a page fault, leading to corrupted state.
To avoid that possibility, 16-byte loads in LE mode take an
alignment interrupt if the operand is not 16-byte aligned. (This
is the case anyway for lqarx, and we enforce it for lq as well.)
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This implements fmadd, fmsub, fnmadd, fnmsub and their
single-precision counterparts. The single-precision versions operate
the same as the double-precision versions until the final rounding and
overflow/underflow steps.
This adds an S register to store the low bits of the product. S
shifts into R on left shifts, and can be negated, but doesn't do any
other arithmetic.
This adds a test for the double-precision versions of these
instructions.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This implements the floating square-root calculation using a table
lookup of the inverse square root approximation, followed by three
iterations of Goldschmidt's algorithm, which gives estimates of both
sqrt(FRB) and 1/sqrt(FRB). Then the residual is calculated as
FRB - R * R and that is multiplied by the 1/sqrt(FRB) estimate to get
an adjustment to R. The residual and the adjustment can be negative,
and since we have an unsigned multiplier, the upper bits can be wrong.
In practice the adjustment fits into an 8-bit signed value, and the
bottom 8 bits of the adjustment product are correct, so we sign-extend
them, divide by 4 (because R is in 10.54 format) and add them to R.
Finally the residual is calculated again and compared to 2*R+1 to see
if a final increment is needed. Then the result is rounded and
written back.
This implements fsqrts as fsqrt, but with rounding to single precision
and underflow/overflow calculation using the single-precision exponent
range. This could be optimized later.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This implements frsqrte by table lookup. We first normalize the input
if necessary and adjust so that the exponent is even, giving us a
mantissa value in the range [1.0, 4.0), which is then used to look up
an entry in a 768-entry table. The 768 entries are appended to the
table for reciprocal estimates, giving a table of 1024 entries in
total. frsqrtes is implemented identically to frsqrte.
The estimate supplied is accurate to 1 part in 1024 or better.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This just returns the value from the inverse lookup table. The result
is accurate to better than one part in 512 (the architecture requires
1/256).
This also adds a simple test, which relies on the particular values in
the inverse lookup table, so it is not a general test.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This implements floating-point division A/B by a process that starts
with normalizing both inputs if necessary. Then an estimate of 1/B
from a lookup table is refined by 3 Newton-Raphson iterations and then
multiplied by A to get a quotient. The remainder is calculated as
A - R * B (where R is the result, i.e. the quotient) and the remainder
is compared to 0 and to B to see whether the quotient needs to be
incremented by 1. The calculations of 1 / B are done with 56 fraction
bits and intermediate results are truncated rather than rounded,
meaning that the final estimate of 1 / B is always correct or a little
bit low, never too high, and thus the calculated quotient is correct
or 1 unit too low. Doing the estimate of 1 / B with sufficient
precision that the quotient is always correct to the last bit without
needing any adjustment would require many more bits of precision.
This implements fdivs by computing a double-precision quotient and
then rounding it to single precision. It would be possible to
optimize this by e.g. doing only 2 iterations of Newton-Raphson and
then doing the remainder calculation and adjustment at single
precision rather than double precision.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This implements the fmul and fmuls instructions.
For fmul[s] with denormalized operands we normalize the inputs
before doing the multiplication, to eliminate the need for doing
count-leading-zeroes on P. This adds 3 or 5 cycles to the
execution time when one or both operands are denormalized.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This implements fctiw, fctiwz, fctiwu, fctiwuz, fctid, fctidz, fctidu
and fctiduz, and adds tests for them.
There are some subtleties around the setting of the inexact (XX) and
invalid conversion (VXCVI) flags in the FPSCR. If the rounded value
ends up being out of range, we need to set VXCVI and not XX. For a
conversion to unsigned word or doubleword of a negative value that
rounds to zero, we need to set XX and not VXCVI.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This brings in the invalid exception for the case of frsp with a
signalling NaN as input, and the need to be able to convert a
signalling NaN to a quiet NaN.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This implements fcfid, fcfidu, fcfids and fcfidus, which convert
64-bit integer values in an FPR into a floating-point value.
This brings in a lot of the datapath that will be needed in
future, including the shifter, adder, mask generator and
count-leading-zeroes logic, along with the machinery for rounding
to single-precision or double-precision, detecting inexact results,
signalling inexact-result exceptions, and updating result flags
in the FPSCR.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This implements fmr, fneg, fabs, fnabs and fcpsgn and adds tests
for them.
This adds logic to unpack and repack floating-point data from the
64-bit packed form (as stored in memory and the register file) into
the unpacked form in the fpr_reg_type record. This is not strictly
necessary for fmr et al., but will be useful for when we do actual
arithmetic.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This adds the skeleton of a floating-point unit and implements the
mffs and mtfsf instructions.
Execute1 sends FP instructions to the FPU and receives busy,
exception, FP interrupt and illegal interrupt signals from it.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This adds code to loadstore1 to convert between single-precision and
double-precision formats, and implements the lfs* and stfs*
instructions. The conversion processes are described in Power ISA
v3.1 Book 1 sections 4.6.2 and 4.6.3.
These conversions take one cycle, so lfs* and stfs* are one cycle
slower than lfd* and stfd*.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This extends the register file so it can hold FPR values, and
implements the FP loads and stores that do not require conversion
between single and double precision.
We now have the FP, FE0 and FE1 bits in MSR. FP loads and stores
cause a FP unavailable interrupt if MSR[FP] = 0.
The FPU facilities are optional and their presence is controlled by
the HAS_FPU generic passed down from the top-level board file. It
defaults to true for all except the A7-35 boards.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Trace interrupts occur when the MSR[TE] field is non-zero and an
instruction other than rfid has been successfully completed. A trace
interrupt occurs before the next instruction is executed or any
asynchronous interrupt is taken.
Since the trace interrupt is defined to set SRR1 bits depending on
whether the traced instruction is a load or an instruction treated as
a load, or a store or an instruction treated as a store, we need to
make sure the treated-as-a-load instructions (icbi, icbt, dcbt, dcbst,
dcbf) and the treated-as-a-store instructions (dcbtst, dcbz) have the
correct opcodes in decode1. Several of them were previously marked as
OP_NOP.
We don't yet implement the SIAR or SDAR registers, which should be set
by trace interrupts.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
In the cases where we need to override the values from the decode ROMs,
we now do that overriding after the clock edge (eating into decode2's
cycle) rather than before. This helps timing a little.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
To avoid adding too much logic, this moves the adder used by OP_ADD
out of the case statement in execute1.vhdl so that the result can
be used by OP_ADDG6S as well.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
These are no-ops that are reserved for future use as performance
hints, so we just need to treat them as no-ops.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
The addex instruction is like adde but uses the XER[OV] bit for the
carry in and out rather than XER[CA].
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This adds a true random number generator for the Xilinx FPGAs which
uses a set of chaotic ring oscillators to generate random bits and
then passes them through a Linear Hybrid Cellular Automaton (LHCA) to
remove bias, as described in "High Speed True Random Number Generators
in Xilinx FPGAs" by Catalin Baetoniu of Xilinx Inc., in:
https://pdfs.semanticscholar.org/83ac/9e9c1bb3dad5180654984604c8d5d8137412.pdf
This requires adding a .xdc file to tell vivado that the combinatorial
loops that form the ring oscillators are intentional. The same
code should work on other FPGAs as well if their tools can be told to
accept the combinatorial loops.
For simulation, the random.vhdl module gets compiled in, which uses
the pseudorand() function to generate random numbers.
Synthesis using yosys uses nonrandom.vhdl, which always signals an
error, causing darn to return 0xffff_ffff_ffff_ffff.
This adds an implementation of the darn instruction. Darn can return
either raw or conditioned random numbers. On Xilinx FPGAs, reading a
raw random number gives the output of the ring oscillators, and
reading a conditioned random number gives the output of the LHCA.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
These instructions use major opcode 4 and have a third GPR input
operand, so we need a decode table for major opcode 4 and some
plumbing to get the RC register operand read.
The multiply-add instructions use the same insn_type_t values as the
regular multiply instructions, and we distinguish in execute1 by
looking at the major opcode. This turns out to be convenient because
we don't have to add any cases in the code that handles the output of
the multiplier, and it frees up some insn_type_t values.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This also removes OP_MCRXR, as the mcrxr instruction was removed in
version 3.0B of the Power ISA, having been phased-out for the server
architecture since v2.02.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Commit d5c8c33bae ("decode1: Reformat to 4-space indentation") resulted
in some rows of major_decode_rom_array being misaligned. This fixes it.
No code change.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This adds "if LOG_LENGTH > 0 generate" to the places in the core
where log output data is latched, so that when LOG_LENGTH = 0 we
don't create the logic to collect the data which won't be stored.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This adds a path to allow the CR result of one instruction to be
forwarded to the next instruction, so that sequences such as
cmp; bc can avoid having a 1-cycle bubble.
Forwarding is not available for dot-form (Rc=1) instructions,
since the CR result for them is calculated in writeback. The
decode.output_cr field is used to identify those instructions
that compute the CR result in execute1.
For some reason, the multiply instructions incorrectly had
output_cr = 1 in the decode tables. This fixes that.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This implements the CFAR SPR as a slow SPR stored in 'ctrl'. Taken
branches and rfid update it to the address of the branch or rfid
instruction.
To simplify the logic, this makes rfid use the branch logic to
generate its redirect (requiring SRR0 to come in to execute1 on
the B input and SRR1 on the A input), and the masking of the bottom
2 bits of NIA is moved to fetch1.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This implements a simple branch predictor in the decode1 stage. If it
sees that the instruction is b or bc and the branch is predicted to be
taken, it sends a flush and redirect upstream (to icache and fetch1)
to redirect fetching to the branch target. The prediction is sent
downstream with the branch instruction, and execute1 now only sends
a flush/redirect upstream if the prediction was wrong. Unconditional
branches are always predicted to be taken, and conditional branches
are predicted to be taken if and only if the offset is negative.
Branches that take the branch address from a register (bclr, bcctr)
are predicted not taken, as we don't have any way to predict the
branch address.
Since we can now have a mflr being executed immediately after a bl
or bcl, we now track the update to LR in the hazard tracker, using
the second write register field that is used to track RA updates for
update-form loads and stores.
For those branches that update LR but don't write any other result
(i.e. that don't decrementer CTR), we now write back LR in the same
cycle as the instruction rather than taking a second cycle for the
LR writeback.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This makes the logic that works out decode.unit and decode.sgl_pipe
for mtspr/mfspr to/from slow SPRs detect the fact that the
instruction is mtspr/mfspr based on a match with the instruction
word rather than looking at v.decode.insn_type. This improves timing
substantially, as the ROM lookup to get v.decode is relatively slow.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This means that the busy signal from execute1 (which can be driven
combinatorially from mmu or dcache) now stops at decode1 and doesn't
go on to icache or fetch1. This helps with timing.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
