The present invention relates to a prefetch unit for use in a computer
In a computer system, instructions are typically fetched from a program
memory, decoded and supplied to an execution unit where they are executed to run
the program stored in the program memory. If more than one execution unit is provided,
it is possible to arrange for very high speed instruction execution. In order
to take advantage of this, it is clearly necessary to be able to supply decoded
instructions to the execution unit at a sufficient rate. Presently, access times
to memory cannot match execution speeds, and therefore several machine cycles are
needed to access each new instruction from memory. Thus, there can be a severe
performance degradation because the fetches from memory cannot match the rate at
which instructions can be executed by the execution units.
According to the present invention there is provided a prefetch buffer
for holding instructions in a processor having a memory and an instruction decode
unit, the prefetch buffer comprising:
- a plurality of storage locations, each having the same bit capacity (2n bits)
and arranged in groups with the same number p of storage locations in each group;
- a write port for selectively writing words of bit length n x p from the memory
into respective groups of the prefetch buffer;
- read circuitry for reading instructions out of the prefetch buffer in dependence
on an instruction mode of the processor, said instruction mode controlling the
number of storage locations which are read during a machine cycle; and
- means for indicating when all storage locations in a group have been read so
that a fetch signal can be issued to fetch a next word from the memory into the
storage locations of that group.
In the described embodiment, each storage location has a capacity
of 32 bits, and are arranged in groups of four such that each group has a capacity
for a 128 bit word read out of memory on a memory fetch. In the described embodiment,
four groups of storage locations are provided in the prefetch buffer, thus allowing
for up to four successive memory accesses even if the first word has not yet been
either received or executed. Moreover, because the processor supports more than
one instruction mode, the time which it takes to read all storage locations in
a group in terms of machine cycles can vary. According to the invention, the indicating
means allow for a next word to be fetched from memory when all storage locations
in a group have been read, however many machine cycles that has taken. Thus, memory
latency is hidden through this mechanism.
According to a first instruction mode, one storage location is read
out during each machine cycle to provide a pair of 16 bit instructions to the decode
unit (referred to herein as GP16 mode).
According to a second instruction mode, two storage locations are
read during each machine cycle to provide two 32 bit instructions to the decode
unit (referred to herein as GP32 mode).
According to a third instruction mode, four storage locations are
read out during each machine cycle to provide four instructions each of 32 bits
to the decode unit (referred to herein as VLIW (Very Long Instruction Word) mode).
In the described embodiment, the indicating means comprises a set
of flags, each group having a flag associated therewith which is set to indicate
that all storage locations in the associated group have been read so as to initiate
a subsequent memory fetch.
The invention also provides a prefetch unit comprising a prefetch
buffer as hereinabove defined and control circuitry arranged to monitor the indicating
means and to issue a fetch signal to memory to fetch the next word into the prefetch
buffer when all storage locations in a group have been read. The control circuitry
can include an aligner for controlling a read pointer determining the storage locations
to be read in a next machine cycle.
For a better understanding of the present invention and to show how
the same may be carried into effect, reference will now be made by way of example
to the accompanying drawings in which:-
- Figure 1 is a block diagram of a prefetch unit;
- Figure 2 illustrates the different instruction modes of the processor;
- Figure 3 illustrates the organisation of a prefetch buffer; and
- Figure 4 is a schematic diagram illustrating the operation of the prefetch
Figure 1 is a block diagram of a prefetch unit 2 for a processor,
the prefetch unit 2 comprising a prefetch buffer 4 with associated control bits
6 and control circuitry comprising a prefetcher 8 and an aligner 10. The prefetcher
8 is connected to a program memory 12 and is responsible for initiating memory
accesses to the program memory 12 using memory access control signals 14a,14b.
The address in memory to which a fetch is initiated is held in a prefetch program
counter 16 in the prefetcher 8. Control of the prefetch program counter is not
discussed herein, but it can be assumed that fetches are initiated from memory
in accordance with a sequence of instructions to be executed by the processor.
That is, the prefetch program counter may be incremented each time as a sequence
of adjacent instructions is fetched, or it may change according to branches, traps,
interrupts etc. Responsive to a memory fetch initiated by the prefetcher, instruction
words are supplied from the program memory 12 to the prefetch buffer 4 as represented
by data-in path 18.
The aligner 10 controls reading of instructions from the prefetch
buffer to a decoder 20 along data-out path 22. To do this, the aligner issues
and is responsive to prefetcher align (PFAL)/decoder control signals 24a,24b. The
aligner 10 has an align program counter 26 which keeps track of how many instructions
have been dispatched to the decoder 20 in each machine cycle, and a state machine
28 which generates a read pointer RP for controlling the prefetch buffer in a manner
which is described in more detail hereinafter.
Instructions in the program memory 12 can have a length of 16 bits
or 32 bits. The prefetch buffer supports three different instruction modes as described
with reference to Figure 2 as follows. The instruction mode is held in a process
status register (PSR) 3 and can be changed. Change mode signals chmd1,chmd2 are
issued by the decoder 20 responsive to a change in instruction mode.
According to a first instruction mode, a pair of 16 bit instructions
are supplied during each machine cycle to the decoder 20 from the prefetch buffer
4. This pair is denoted slot0,slot1 in bit sequences w0,w1 etc. This is referred
to herein as GP16 mode.
According to a second instruction mode, two instructions each having
a length of 32 bits are supplied to the decoder from the prefetch buffer in each
machine cycle, for example w0,w1 in CYCLE 0. This mode is referred to herein as
According to a third instruction mode, four instructions w0,w1,w2,w3
each of 32 bits in length are supplied to the decoder in each machine cycle. This
is referred to herein as VLIW.
In all modes, each fetch operation initiated to the program memory
12 retrieves an instruction word of 128 bits in length.
Thus, in GP16 mode, the instruction word comprises eight 16 bit instructions,
paired as slot0,slot1 for each machine cycle. In GP32 and VLIW mode, the instruction
word comprises four 32 bit instructions.
The organisation of the prefetch buffer 4 is illustrated in Figure
3. In diagrammatic terms, the prefetch buffer can be considered to have four successive
lines L0 to L3, each having a capacity of 128 bits. There is a single write port
WPO having a width of 128 bits which receives data from the program memory via
the data-in path 18 and an input latch FF-in and writes it into the selected line
under the control of a write pointer WP [3:0]. Each line comprises four storage
locations each having a capacity of 32 bits and each of which is shown diagrammatically
divided into two 16 bit sections for the purposes of explanation. The storage
locations are denoted F0 to F15. Each line in Figure 3 is referred to herein as
a group of storage locations and has the capacity for one 128 bit line from memory.
This allows up to four successive memory accesses to be made, even if the first
instruction word has not been received or executed by the processor. While the
instruction word in storage locations F0 to F3 is being decoded and subsequently
executed, memory fetches can continue to be implemented into the storage locations
F4 to F7, F8 to F11 and F12 to F15 until the buffer is full. By the time that
a memory fetch has been made into the last group F12 to F15, it is most likely
that the first group F0 to F3 will have been completely read out into the decoder
and will thus be ready to receive a subsequent instruction word from memory. The
number of cycles required to decode an instruction word in each group varies depending
on the instruction mode of the machine in a manner which will be described in more
detail in the following. Nevertheless, a minimum of one cycle is required for reading
and decoding, and therefore the use of the prefetch buffer hides memory latency.
In order to save a cycle when the prefetch buffer is empty or flushed
after a branch, data can bypass the prefetch buffer through a bypass circuitry
BS. As described in more detail later, the bypass circuitry is implemented as a
plurality of multiplexors (MUX0 to MUX3 in Figure 4).
Figure 4 is a more detailed diagram of the prefetch buffer and its
associated read circuitry. The storage locations F0 to F15 are illustrated aligned
vertically for the purposes of explanation.
The control bits 6 described above in Figure 1 include empty flags
EF1 to EF4 which indicate when a complete 128 bit line of storage locations is
empty such that a subsequent memory fetch can be initiated. When a fetch is instituted
from memory, and data has been received by the prefetch buffer, the empty flag
is cleared to indicate that those storage locations are now full.
Reading from the prefetch buffer will now be described with reference
to the schematic diagram of Figure 4. The prefetch buffer includes four read ports
RP1,RP2,RP3 and RP4. These read ports each take the form of multiplexors each capable
of connecting selected ones of the storage locations F0 to F15 to a 32 bit output,
pf-buf-out1,2,3 or 4. However, the read ports are not identical. The first read
port RP1 has sixteen inputs each of which is connected to a respective storage
location F0 to F15 and each of which can be connected to the output pf-buf-out1.
The second read port RP2 has eight inputs which are respectively connected to
storage locations F1,F3,F5,F7,F9,F11,F13,F15 to selectively connect the contents
of those storage locations to the output pf-buf-out2.
The third read port RP3 has four inputs connected to storage locations
F2,F6,F10 and F14 for selectively connecting the contents of those storage locations
to the output pf-buf-out3. The fourth read port RP4 also has four inputs which
are connected to storage locations F3,F7,F11 and F15 for selectively connecting
the contents of those storage locations to the output pf-buf-out4.
The read ports RP1 to RP4 are controlled by the read pointer RP from
the aligner 10 in dependence on the instruction mode of the machine and the consequential
number of machine cycles required for decoding each instruction word.
Alternatively, for instructions supplied directly from memory along
data-in path 18, the control of instructions supplied to the decoder in dependence
on the instruction mode and machine cycles is additionally controllable by multiplexors
MUX0,MUX1,MUX2 and MUX3. These receive at their input respective bits of the 128
bit data-in path 18 to supply a 32 bit sequence to each multiplexor in each machine
cycle as described in the following.
The selection of which instructions within the instruction word are
supplied to the decoder 20 is made on dependence on the instruction mode as described
in the following. In Figure 3, the symbols w0 to w3 are used on different input
lines of the multiplexors MUX0 to MUX3 to represent different 32 bit sequences,
as in Figure 4. The definition of each 32 bit sequence depends on the instruction
mode, but bits of the data-in path are always allocated as w0 [0:31], w1 [32:63],
w2 [64:95], w3 [96:127]. The inputs to the multiplexors are individually labelled
so as to distinguish between them. That is, in GP16 mode, on the first decode cycle,
cycle 0, the first sequence w0 is supplied to the decoder 20. This presents a pair
of 16 bit instructions, slot0,slot1 (w0) for simultaneous decoding by the decoder
20. On the next cycle, cycle 1, the sequence w1 is supplied, presenting the next
pair of 16 bit instructions slot0,slot1 (w1) for decoding. In GP16 mode, the read
port RP1 and the multiplexor MUX0 are the only read devices which are used and
the control of the word which is supplied to the decoder is made by the multiplexor
MUX0 under the control of signal mux-ctrl0, and the read pointer RP. If the signal
mux-ctrl0 selects the read port output pf-buf-out1, the read pointer selects inputs
F0 to F3 over four successive cycles CYCLE0 to CYCLE3 to read out successively
w0 to w4. Once storage location F3 has been read out, the read port counter will
reset the read port RP1 so that it reads out from storage locations F4 to F7 over
the next four cycles. If the buffer is not in use, the first instruction pair
w0 is read out by the multiplexor MUX0. That is in cycle 0, input M00 of the multiplexor
MUX0 is selected. Meanwhile, the 128 bit line is loaded into the first location
of the prefetch buffer and the read pointer points to the next location to be
read out by the decoder. Therefore on cycle 1, the next instruction pair w1 is
read out by the multiplexor MUX0 by selecting pf-buf-out1
In GP32 mode, in the first machine cycle the first two instructions
w0,w1 are presented to the decoder 20. In the subsequent cycle, cycle 1, the next
two instructions w2,w3 are presented to the decoder. This utilises read ports RP1
and RP2 and the multiplexors MUX0 and MUX1. If the signal mux-ctr10 is set to
pf-buf-out1, and mux-ctrl1 to pf-buf-out2, then the read pointer RP is set to F0
for RP1 and F1 for RP2 in cycle 0. In cycle 1, it is changed to F2 and F3 respectively.
Instructions are then read over the next two cycles from the next group of storage
locations F4 to F7 by altering the setting of the read ports RP1 and RP2 responsive
to the read pointer RP. Alternatively, when read from the data-in path 18, in the
first cycle, the first input M10 of the multiplexor MUX1 is set to read w1 (bits
31 to 63) and the first input M00 of the multiplexor MUX0 is set to read w0 (bits
0 to 31). Thus, instructions w0 and w1 are presented to the decoder 20 in CYCLE
0. Meanwhile, the 128 bit line is loaded into the prefetch buffer so that in the
subsequent cycle, CYCLE 1, w2 and w3 are read from the buffer by selecting pf-buf-out1
In VLIW mode, four 32 bit instructions W0 to W3 (slot0 to slot3)
are supplied simultaneously to the decoder 20 in each machine cycle, e.g. CYCLE
0. The multiplexors MUX2 and MUX3 are set according to the control signals mux-ctrl2
and mux-ctrl3 respectively to allow the instruction words w2 and w3 to be read
either from the buffer or from the data-in path 18. In other respects, the settings
of RP1 and RP2, MUX0 and MUX1 are as in GP32 mode. However, in the subsequent cycle,
e.g. CYCLE 1 in VLIW mode, it will be noticed that the instruction words w2 and
w3 which would have been remaining in GP32 mode have now been read out. Therefore,
the read pointer RP can immediately move on to the next set of storage locations
F4 to F7 to read out the subsequent VLIW instruction word containing the next four
Data is passed from the multiplexors MUX0 to MUX3 to respective output
flip-flops FF0 to FF3 via a set of control gates labelled GC1, GC2 and GS0 to GS3.
The control gates GC1,GC2 are responsive to change mode signals chmd1,chmd2 respectively
which indicate to the prefetch unit that there has been a change in the instruction
mode in which the machine is operating. The control gates GS0 to GS3 are responsive
to respective stop signals stop to stop to prevent any new data from entering
the decoder from that output flip-flop. These effectively allow the decoder to
be stalled. In a stop condition, the outputs of the flip-flops are recirculated
to the input of its associated control switch to prevent unnecessary operation
of the subsequent decoder.
Operation of the prefetch unit responsive to the change mode signals
chmd1 and chmd2 will now be described. The output flip-flop FF0 is connected to
a single 32 bit decoder and to two 16 bit decoders. When the machine is in GP16
mode, the outputs of the two 16 bit decoders are selected for the instruction pair
supplied to the flip-flop FF0. When the machine is in GP32 mode, the output of
the 32 bit decoder is selected. The remaining flip-flops FF1 to FF3 are each connected
to respective 32 bit decoders.
A first change mode signal chmd1 signals a change of machine instruction
mode from GP32 to GP16. If the machine had been operating in GP32 mode, consider
the situation at the end of cycle 0 which reference to Figure 2. Instructions w0
and w1 will have been supplied via the flip-flops FF0 and FF1 to the respective
32 bit decoders of the decoder 20. However, the change in instruction mode now
implies that the 32 bit sequence which was formerly to be considered as the second
instruction W1 in cycle 0 of GP32 mode, in fact contains a pair of 16 bit instructions
as denoted in cycle 1 of GP16 mode. Thus, the output of the 32 bit decoder connected
to the flip-flop FF1 needs to be ignored, and the 32 bit sequence w1 needs to be
reapplied to the two 16 bit decoders connected to the output flip-flop F0. This
is achieved by the recirculation line 42 from the output of the flip-flop FF1 to
the input of the control gate CG1.
Conversely, control signal chmd2 denotes a change of instruction
mode from GP16 to GP32. Consider again the effect at the end of cycle 0 with reference
to Figure 2. The instruction pair denoted w0 has just been decoded in GP16 mode,
and the expectation is that the machine will now wait for the next instruction
pair w1. However, in GP32 mode, that word w1 represents a single instruction and
the change mode signal chmd2 allows it to be applied directly through the control
gate GC2 to the output flip-flop FF1 so that it can be applied directly to the
input of the dedicated 32 bit decoder connected to the output of the flip-flop
FF1. This allows the instruction w1 to be decoded as a single 32 bit instruction.
In the next cycle, instructions w2 and w3 can be transmitted normally as indicated
by cycle 1 in GP32 mode in Figure 2.
It will be clear from the above that the number of cycles needed
to read all four storage locations in a group depends on the instruction mode.
That is, in GP16 mode, four cycles are needed, in GP32 two cycles are needed and
VLIW one cycle is needed. When all the storage locations F0 to F3 in the first
group have been read, the first empty flag EF1 is cleared to empty.
The aligner controls the setting and clearing of the "empty" flags
using information from the read pointer. The aligner detects when the read pointer
goes from one line (128 bits) to the next. When this occurs, the "empty" flag corresponding
to the page which has just been read is set.
The state of an empty flag being cleared is detected by the prefetcher
8 along line 48 and a fetch is initiated to the next prefetch address in the prefetch
program counter 16. Thus, the next instruction line is fetched from memory and
the write pointer WP is set to write it into storage locations F0 to F3. In the
meantime, the read pointer has moved to the second group F4 to F7 to read and decode
instructions of that group. When those storage locations are empty, the empty flag
EF2 is cleared, a next memory fetch is initiated by the prefetcher 8 and the read
pointer moves onto the group F8 to F11. As can readily be seen, the prefetch buffers
masks a latency of memory fetches of at least three cycles in the VLIW mode, and
a greater number of cycles in GP32 and GP16 mode. Signals are supplied from the
decoder along line 24b to the aligner 10 indicating what mode the decoder is operating
in so that the aligner can adjust the align program counter 26 accordingly and
keep track of the next instructions to be decoded so that the read pointer RP can
correctly be issued by the state machine 28.