This adds a bit to the BTC to store whether the corresponding branch
instruction was taken last time it was encountered. That lets us pass
a not-taken prediction down to decode1, which for backwards direct
branches inhibits it from redirecting fetch to the target of the
branch. This increases coremark by about 2%.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This implements most of the architected PMU events. The ones missing
are mostly the ones that depend on which level of the cache hierarchy
data is fetched from. The events implemented here, and their raw
event codes, are:
Floating-point operation completed (100f4)
Load completed (100fc)
Store completed (200f0)
Icache miss (200fc)
ITLB miss (100f6)
ITLB miss resolved (400fc)
Dcache load miss (400f0)
Dcache load miss resolved (300f8)
Dcache store miss (300f0)
DTLB miss (300fc)
DTLB miss resolved (200f6)
No instruction available and none being executed (100f8)
Instruction dispatched (200f2, 300f2, 400f2)
Taken branch instruction completed (200fa)
Branch mispredicted (400f6)
External interrupt taken (200f8)
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This makes the icache snoop writes to memory in the same way that the
dcache does, thus making DMA cache-coherent for the icache as well as
the dcache.
This also simplifies the logic for the WAIT_ACK state by removing the
stbs_done variable, since is_last_row(r.store_row, r.end_row_ix) can
only be true when stbs_done is true.
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 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>
The ability to stop the core using the debug interface has been broken
since commit bb4332b7e6b5 ("Remove fetch2 pipeline stage"), which
removed a statement that cleared the valid bit on instructions when
their stop_mark was 1.
Fix this by clearing r.req coming out of fetch1 when r.stop_mark = 1.
This has the effect of making i_out.valid be 0 from the icache. We
also fix a bug in icache.vhdl where it was not honouring i_in.req when
use_previous = 1.
It turns out that the logic in fetch1.vhdl to handle stopping and
restarting was not correct, with the effect that stopping the core
would leave NIA pointing to the last instruction executed, not the
next instruction to be executed. In fact the state machine is
unnecessary and the whole thing can be simplified enormously - we
need to increment NIA whenever stop_in = 0 in the previous cycle.
Fixes: bb4332b7e6b5 ("Remove fetch2 pipeline stage")
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Big-endian mode affects both instruction fetches and data accesses.
For instruction fetches, we byte-swap each word read from memory when
writing it into the icache data RAM, and use a tag bit to indicate
whether each cache line contains instructions in BE or LE form.
For data accesses, we simply need to invert the existing byte_reverse
signal in BE mode. The only thing to be careful of is to get the sign
bit from the correct place when doing a sign-extending load that
crosses two doublewords of memory.
For now, interrupts unconditionally set MSR[LE]. We will need some
sort of interrupt-little-endian bit somewhere, perhaps in LPCR.
This also fixes a debug report statement in fetch1.vhdl.
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 does the PLRU update based on r.hit_valid and r.hit_way rather
than req_is_hit and req_hit_way, which means there is now a register
between the TLB and cache tag lookup and the PLRU update, which
should help with timing.
As a result, the PLRU victim way selection becomes valid one cycle
later, in the cycle when r.state = CLR_TAG. So we have to use the
PLRU output directly in the CLR_TAG state and r.store_way in the
WAIT_ACK state.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
The icache can now detect a hit on a line being refilled from memory,
as we have an array of individual valid bits per row for the line
that is currently being loaded. This enables the request that
initiated the refill to be satisfied earlier, and also enables
following requests to the same cache line to be satisfied before the
line is completely refilled. Furthermore, the refill now starts
at the row that is needed. This should reduce the latency for an
icache miss.
We now get a 'sequential' indication from fetch1, and use that to know
when we can deliver an instruction word using the other half of the
64-bit doubleword that was read last cycle. This doesn't make much
difference at the moment, but it frees up cycles where we could test
whether the next line is present in the cache so that we could
prefetch it if not.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
The fetch2 stage existed primarily to provide a stash buffer for the
output of icache when a stall occurred. However, we can get the same
effect -- of having the input to decode1 stay unchanged on a stall
cycle -- by using the read enable of the BRAMs in icache, and by
adding logic to keep the outputs unchanged on a clock cycle when
stall_in = 1. This reduces branch and interrupt latency by one
cycle.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This logs 256 bits of data per cycle to a ring buffer in BRAM. The
data collected can be read out through 2 new SPRs or through the
debug interface.
The new SPRs are LOG_ADDR (724) and LOG_DATA (725). LOG_ADDR contains
the buffer write pointer in the upper 32 bits (in units of entries,
i.e. 32 bytes) and the read pointer in the lower 32 bits (in units of
doublewords, i.e. 8 bytes). Reading LOG_DATA gives the doubleword
from the buffer at the read pointer and increments the read pointer.
Setting bit 31 of LOG_ADDR inhibits the trace log system from writing
to the log buffer, so the contents are stable and can be read.
There are two new debug addresses which function similarly to the
LOG_ADDR and LOG_DATA SPRs. The log is frozen while either or both of
the LOG_ADDR SPR bit 31 or the debug LOG_ADDR register bit 31 are set.
The buffer defaults to 2048 entries, i.e. 64kB. The size is set by
the LOG_LENGTH generic on the core_debug module. Software can
determine the length of the buffer because the length is ORed into the
buffer write pointer in the upper 32 bits of LOG_ADDR. Hence the
length of the buffer can be calculated as 1 << (31 - clz(LOG_ADDR)).
There is a program to format the log entries in a somewhat readable
fashion in scripts/fmt_log/fmt_log.c. The log_entry struct in that
file describes the layout of the bits in the log entries.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This stores the output of the PLRU big mux and clears the
tags and valid bits on the next cycle.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
icbi currently just resets the icache. This has some nasty side
effects such as also clearing the TLB, but also the wishbone interface.
That means that any ongoing cycle will be dropped.
However, most of our slaves don't handle that well and will continue
sending acks for already issued requests.
Under some circumstances we can thus restart an icache load and get
spurious ack/data from the wishbone left over from the "cancelled"
sequence.
This has broken booting Linux for me.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The global wr_en signal is causing Vivado to generate two TDP (True Dual Port)
block RAMs instead of one SDP (Simple Dual Port) for each cache way. Remove
it and instead apply a AND to the individual byte write enables.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Slbia (with IH=7) is used in the Linux kernel to flush the ERATs
(our iTLB/dTLB), so make it do that.
This moves the logic to work out whether to flush a single entry
or the whole TLB from dcache and icache into mmu. We now invalidate
all dTLB and iTLB entries when the AP (actual pagesize) field of
RB is non-zero on a tlbie[l], as well as when IS is non-zero.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
This adds a direct-mapped TLB to the icache, with 64 entries by default.
Execute1 now sends a "virt_mode" signal from MSR[IR] to fetch1 along
with redirects to indicate whether instruction addresses should be
translated through the TLB, and fetch1 sends that on to icache.
Similarly a "priv_mode" signal is sent to indicate the privilege
mode for instruction fetches. This means that changes to MSR[IR]
or MSR[PR] don't take effect until the next redirect, meaning an
isync, rfid, branch, etc.
The icache uses a hash of the effective address (i.e. next instruction
address) to index the TLB. The hash is an XOR of three fields of the
address; with a 64-entry TLB, the fields are bits 12--17, 18--23 and
24--29 of the address. TLB invalidations simply invalidate the
indexed TLB entry without checking the contents.
If the icache detects a TLB miss with virt_mode=1, it will send a
fetch_failed indication through fetch2 to decode1, which will turn it
into a special OP_FETCH_FAILED opcode with unit=LDST. That will get
sent down to loadstore1 which will currently just raise a Instruction
Storage Interrupt (0x400) exception.
One bit in the PTE obtained from the TLB is used to check whether an
instruction access is allowed -- the privilege bit (bit 3). If bit 3
is 1 and priv_mode=0, then a fetch_failed indication is sent down to
fetch2 and to decode1, which generates an OP_FETCH_FAILED. Any PTEs
with PTE bit 0 (EAA[3]) clear or bit 8 (R) clear should not be put
into the iTLB since such PTEs would not allow execution by any
context.
Tlbie operations get sent from mmu to icache over a new connection.
Unfortunately the privileged instruction tests are broken for now.
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
The icache would still spit out an instruction which could
cause a 0x700 instead of a reset.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
All that needs to be changed now is the size in wishbone_types.vhdl
and the address decoder in soc.vhdl
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
We used the variable "way" in the wrong state in the cache when
updating a line valid bit after the end of the wishbone transactions,
we need to use the latched "store_way".
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This might slightly increase the logic in synthesis but avoids
us looking at uninitialized tags when not servicing an active
request
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This adds support for set associativity to the icache. It can still
be direct mapped by setting NUM_WAYS to 1.
The replacement policy uses a simple tree-PLRU for each set.
This is only lightly tested, tests pass but I have to double check
that we are using the ways effectively and not creating duplicates.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
We only ever access the cache memory for at most the wishbone bus
width at a time. So having the BRAMs organized as a cache-line-wide
port is a waste of resources.
Instead, use a wishbone-wide memory and store a line as consecutive
rows in the BRAM.
This significantly improves BRAM usage in the FPGA as we can now use
more rows in the BRAM blocks. It also saves a few LUTs and muxes.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
The goal is to have the icache fit in BRAM by latching the output
into a register. In order to avoid timing issues , we need to give
the BRAM a full cycle on reads, and thus we souce the BRAM address
directly from fetch1 latched NIA.
(Note: This will be problematic if/when we want to hash the address,
we'll probably be better off having fetch1 latch a fully hashed address
along with the normal one, so the icache can use the former to address
the BRAM and pass the latter along)
One difficulty is that we cannot really stall the icache without adding
more combo logic that would break the "one full cycle" BRAM model. This
means that on stalls from decode, by the time we stall fetch1, it has
already gone to the next address, which the icache is already latching.
We work around this by having a "stash" buffer in fetch2 that will stash
away the icache output on a stall, and override the output of the icache
with the content of the stash buffer when unstalling.
This requires a rewrite of the stop/step debug logic as well. We now
do most of the hard work in fetch1 which makes more sense.
Note: Vivado is still not inferring an built-in output register for the
BRAMs. I don't want to add another cycle... I don't fully understand why
it wouldn't be able to treat current_row as such but clearly it won't. At
least the timing seems good enough now for 100Mhz, possibly more.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Paul Mackerras
Anton Blanchard