Appendix A PDF
Appendix A PDF
Appendix A PDF
Appendix A Solutions
(by Rui Ma and Gregory D. Peterson)
A.1
The first challenge of this exercise is to obtain the instruction mix. The instruction
frequencies in Figure A.27 must add to 100, although gap and gcc add to 100.2 and
99.5 percent, respectively, because of rounding error. Because each total must in
reality be 100, we should not attempt to scale the per instruction average frequencies by the shown totals of 100.2 and 99.5. However, in computing the average frequencies to one significant digit to the right of the decimal point, we should be
careful to use an unbiased rounding scheme so that the total of the averaged frequencies is kept as close to 100 as possible. One such scheme is called round to
even, which makes the least significant digit always even. For example, 0.15 rounds
to 0.2, but 0.25 also rounds to 0.2. For a summation of terms, round to even will not
accumulate an error as would, for example, rounding up where 0.15 rounds to 0.2
and 0.25 rounds to 0.3.
For gap and gcc the average instruction frequencies are shown in Figure S.12.
Instruction
load
25.8
store
11.8
add
20.0
sub
2.0
mul
0.8
compare
4.4
load imm
3.6
cond branch
10.7
cond move
0.5
jump
0.8
call
1.1
return
1.1
shift
2.4
and
4.4
or
8.2
xor
2.0
other logical
0.2
Figure S.12 MIPS dynamic instruction mix average for gap and gcc.
The exercise statement gives CPI information in terms of four major instruction
categories, with two subcategories for conditional branches. To compute the
Appendix A Solutions
Effective CPI =
= 1.24
A.2
See the solution for A.1 (above) for a discussion regarding the solution methodology for this exercise.
As with problem A.1, the frequency of ALU instructions is the sum of the frequencies of the add, sub, mul, compare, load imm (remember, this instruction
does not access memory, instead the value to load is encoded in a field within the
instruction itself ), cond move (implemented as an OR instruction between a controlling register and the register with the data to move to the destination register),
shift, and, or, xor, and other logical for a total of 51.1%. The frequency of load/
store instructions is the sum of the frequencies of load and store for a total of
35.0%. The frequency of conditional branches is 11.0%. Finally, the frequency of
jumps is the sum of the frequencies of the jump-type instructions, namely jump,
call, and return, for a total of 2.8%.
CPI = 1.0 51.1% + 1.4 35.0% + 2.0 11.0% 60% + 1.5 11.0% 40% + 1.2 2.8% = 1.23
Instruction
load
24.4
store
10.6
add
21.8
sub
3.8
mul
compare
5.2
load imm
1.6
cond branch
11.0
cond move
1.5
jump
1.2
call
0.8
return
0.8
shift
1.3
and
5.3
or
6.8
xor
3.6
other logical
0.2
Figure S.13 MIPS dynamic instruction mix average for gzip and perlbmk.
A.3
This exercise is similar to A.1 and A.2, but focuses on floating point intensive
programs.
Instruction
Load
9.8
Store
2.4
Add
17.8
Sub
3.0
Mul
0.6
Compare
load imm
5.6
cond branch
1.0
cond move
Jump
Call
Return
Shift
1.0
Appendix A Solutions
And
0.9
Or
4.1
Xor
other logical
load FP
16.5
store FP
11.6
add FP
8.6
sub FP
6.2
mul FP
8.2
div FP
0.2
move reg-reg FP
1.4
compare FP
0.4
cond mov FP
0.4
other FP
0.8
Figure S.14 MIPS dynamic instruction mix for lucas and swim.
ALU instructions: (17.8% + 3.0% + 0.6% + 5.6% + 1.0% + 0.9% + 4.1%) = 33.0%
Load-stores: (9.4% + 2.4%) = 11.8%
Conditional branches: 1.0%
Jumps: 0%
FP add: (8.6% + 6.2%) = 14.8%
Load-store FP: (16.5% + 11.6%) = 28.1%
Other FP: (1.4% + 0.4% + 0.4% + 0.8%) = 3.0%
CPI = 1.0 33.0% + 1.4 11.8% + 2.0 1.0% 60% + 1.5 1.0% 40% + 6.0
8.2% + 4.0 14.8% + 20 0.2% + 1.5 28.1% + 2.0 3.0% = 2.12
A.4
Load
16.0
Store
1.4
Add
30.2
Sub
1.2
Mul
1.2
Compare
3.7
load imm
6.8
cond branch
7.0
cond move
0.2
Jump
Figure S.15 Continued
Call
Return
Shift
0.4
And
Or
1.0
Xor
1.6
other logical
load FP
11.7
store FP
4.4
add FP
3.4
sub FP
1.4
mul FP
6.4
div FP
0.4
move reg-reg FP
0.8
compare FP
0.4
cond mov FP
0.3
other FP
Figure S.15 MIPS dynamic instruction mix for applu and art.
2. B = A + C
= B + C + C
3. D = A B
= (B + C) (B + C + C)
= C
Appendix A Solutions
Copy propagation has increased the work in statement 2 from one addition to
two. It has changed the work in statement 3 from subtraction to negation, possibly a savings. The above suggests that writing optimizing compilers means incorporating sophisticated trade-off analysis capability to control any optimizing
steps, if the best results are to be achieved.
A.6
No solution provided.
A.7
DADD
SW
LD
DSLL
DADDI
LD
LD
DADD
LD
DSLL
DADDI
SD
LD
DADDI
SD
LD
DADDI
BNEZ
R1,R0,R0
7000(R0),R1
R1,7000(R0)
R2,R1,#3
R3,R2,#3000
R4,0(R3)
R5,5000(R0)
R6,R4,R5
R1,7000(R0)
R2,R1,#3
R7,R2,#1000
0(R7),R6
R1,7000(R0)
R1,R1,#1
7000(R0),R1
R1,7000(R0)
R8,R1,#-101
R8,loop
;R0 = 0, initialize i = 0
;store i
;get value of i
;R2 = word offset of B[i]
;add base address of B to R2
;load B[i]
;load C
;B[i] + C
;get value of i
;R2 = word offset of A[i]
;add base address of A to R2
;A[i] B[i] + C
;get value of i
;increment i
;store i
;get value of i
;is counter at 101?
;if not 101, repeat
Figure S.16 MIPS code to implement the C loop without using registers to hold
updated values for future use or to pass values to a subsequent loop iteration.
The number of instructions executed dynamically is the number of initialization instructions plus the number of instructions in the loop times the number
of iterations:
Instructions executed = 2 + (16 101) = 1618
The number of memory-data references is a count of the load and store
instructions executed:
Memory-data references executed = 0 + (8 101) = 808
Since MIPS instructions are 4 bytes in size, code size is the number of
instructions times 4:
Instruction bytes = 4 18 = 72
ex_b_7:
movq
movq
movq
movq
movq
mov
shl
movq
loop:
add
mov
shl
movq
movq
add
movq
cmpq
jae
$0x0,%rax
# rax = 0, initialize i = 0
$0x0,%rbp
# base pointer = 0
%rax,0x1b58(%rbp) # store i to location 7000
($1b58)
0x1388(%rbp),%rdx # load C from 5000 ($1388)
0x1b58(%rbp),%rax # get value of i
%rax,%rbx
# rbx gets copy of i
$0x3,%rbx
# rbx now has i * 8
0x0bb8(%rbx),%rcx # load B[i] (3000 = $0bb8)
to rcx
%rdx,%cdx
# B[i] + C
%rax,%rbx
# rbx gets copy of i
$0x3,%rbx
# rbx now has i * 8
%rcx,0x03e8(%rbx) # A[i] B[i] + C
(base address of A is 1000)
0x1b58(%rbp),%rax # get value of i
$0x1,%rax
# increment i
%rax,0x1b58(%rbp) # save i
$0x0065,%rax
# is counter at 101 ($0065)?
loop
# if not 101, repeat
Figure S.17 x86 code to implement the C loop without using registers to hold
updated values for future use or to pass values to a subsequent loop iteration.
b. This problem is similar to part (a), but with x86-64 instructions instead.
The number of instructions executed dynamically is the number of initialization instructions plus the number of instructions in the loop times the number
of iterations:
Instructions executed = 3 + (13 101) = 1316
A.8
This problem focuses on the challenges related to instruction set encoding. The
length of the instructions are 12 bits. There are 32 registers so the length of the
register field will be 5 bits for each register operand. We use addr[11] to addr[0] to
represent the 12 bits of the address as shown in the tables below.
a. In the first case, we need to support 3 two-address instructions. These can be
encoded as follows:
addr[11:10]
addr[9:5]
addr[4:0]
3 two-address instr.
'00000' to '11111'
'00000' to '11111'
Other instructions
'11'
'00000' to '11111'
'00000' to '11111'
Appendix A Solutions
addr[9:5]
addr[4:0]
3 two-address instr.
'00000' to '11111'
'00000' to '11111'
30 one-address instr.
'11'
'00000' to '11101'
'00000' to '11111'
45 zero-address instr.
'11'
'11110'
'00000' to '11111'
'11'
'11111'
'00000' to '01100'
addr[9:5]
addr[4:0]
3 two-address instr.
'00000' to '11111'
'00000' to '11111'
24 zero-address instr.
'11'
'00000'
'00000' to '10111'
X one-address instr.
'11'
'00001' to '11111'
'00000' to '11111'
10
A.9
a. 1) Stack:
Push A // one address appears in the instruction, code size = 8 bits (opcode) +
64 bits (memory address) = 72 bits;
Push B // one address appears in the instruction, code size = 72 bits;
Add // zero address appears in the instruction, code size = 8 bits;
Pop C // one address appears in the instruction, code size = 72 bits;
Total code size = 72 + 72 + 8 + 72 = 224 bits.
2) Accumulator
Load A // one address appears in the instruction, code size = 8 bits (opcode) +
64 bits (memory address) = 72 bits;
Add B // one address appears in the instruction, code size = 72 bits;
Store C // one address appears in the instruction, code size = 72 bits;
Total code size = 72 + 72 + 8 + 72 = 216 bits.
3) Register-memory
Load R1, A // two addresses appear in the instruction, code size = 8 bits
(opcode) + 6 bits (register address) + 64 bits (memory address) = 78 bits;
Add R3, R1, B // three addresses appear in the instruction, code size = 8 bits
(opcode) + 6 bits (register address) + 6 bits (register address) + 64 bits (memory address) = 84 bits;
Store R3,C // two addresses appear in the instruction, code size = 78 bits;
Total code size = 78 + 84 + 78 = 240 bits.
4) Register-register
Load R1, A // two addresses appear in the instruction, code size = 8 bits
(opcode) + 6 bits (register address) + 64 bits (memory address) = 78 bits;
Load R2, B // two addresses appear in the instruction, code size = 78 bits;
Add R3, R1, R2 // three addresses appear in the instruction, code size = 8 bits
(opcode) + 6 bits (register address) + 6 bits (register address) + 6 bits (register
address) = 26 bits;
Store R3, C // two addresses appear in the instruction, code size = 78 bits;
Total code size = 78 + 78 + 26 + 78 = 260 bits.
b. Assume all the data size are 64 bits. Assume there is no complier optimization.
Appendix A Solutions
11
1) Stack
Code
Destroyed
data
Overhead
Code size
(bits)
Size of mem
data
Push A
no
72
64
Push B
no
72
64
A and B
Add
Pop C
72
64
Push E
no
72
64
Push A
no
72
64
Sub
yes
A and E
Pop D
72
64
Push C
no
yes
72
64
no
yes
72
64
Push D
Add
Pop F
C and D
no
72
64
672
576
Total
The total code size is 672/8 = 84 bytes. The number of bytes of data moved to
or from memory is 576/8 = 72 bytes. There are 3 overhead instructions. The
number of overhead data bytes is 24 bytes.
2) Accumulator
Code
Destroyed
data
Overhead
Load A
Add B
Store C
Load A
Sub E
yes
Store D
Add C
Store F
Total
Code size
(bits)
Size of mem
data
72
64
72
64
72
64
72
64
72
64
72
64
72
64
72
64
576
512
The total code size is 576/8 = 72 bytes. The number of bytes of data moved to
or from memory is 512/8 = 64 bytes. There is 1 overhead instruction. The
number of overhead data bytes is 8 bytes.
12
3) Register-memory
Code
Destroyed
data
Code size
(bits)
Overhead
Size of mem
data
Load R1, A
78
64
84
64
Store R3, C
78
64
84
64
Store R2, D
78
64
84
Store R4, F
Total
78
64
564
384
The total code size is 564/8 = 71 bytes. The number of bytes of data moved to
or from memory is 384/8 = 48 bytes. There is no overhead instruction. The
number of overhead data bytes is 0 bytes.
4) Register-register
Code size
(bits)
Size of mem
data
Load R1, A
78
64
Load R2, B
78
64
Code
Destroyed
Data
Overhead
26
Store R3, C
78
64
Load R4, E
78
64
26
Store R5, D
78
26
Store R6, F
Total
64
78
64
546
384
The total code size is 546/8 = 69 bytes. The number of bytes of data moved to
or from memory is 384/8 = 48 bytes. There is no overhead instruction. The
number of overhead data bytes is 0 bytes.
A.10
Appendix A Solutions
13
This program illustrates how alignment impacts data structures in C/C++. Note that
the struct definition is for C++ because it includes a bool type.
Data size on 32-bit
machine (bytes)
char
bool
int
long
Data type
double
short
float
pointer
14
As with the 32-bit processor, two bytes are added after the bool b so the int c is
aligned and two more are added after the short e so the float f is aligned. Finally,
the final int x has four bytes added to the end to make the struct aligned with
an 8 byte boundary. Hence, the original struct requires 56 bytes on a 64-bit
processor. By reordering the contents of the struct in the same manner as with
the 32-bit machine above, the additional padding requirements are eliminated
and the 48 byte size can be achieved for the struct.
A.12
No solution provided.
A.13
No solution provided.
A.14
No solution provided.
A.15
No solution provided.
A.16
No solution provided.
A.17
No solution provided.
A.18
Load B
Add C
Store A
Add C
Store B
Negate
Add A
Store D
B
Acc + C
Acc
"A + C
Acc
Acc
B + A
Acc
Push B
Push C
Add
Pop A
Push A
Push C
Add
Pop B
Push B
Push A
Sub
Pop D
Mem[B], NTTOS *
Mem[C], NTTOS TOS
TOS + NTTOS, NTTOS *
TOS, TOS *
Mem[A], NTTOS *
Mem[C], NTTOS TOS
TOS + NTTOS, NTTOS *
TOS, TOS *
Mem[B], NTTOS *
Mem[A], NTTOS TOS
TOS NTTOS, NTTOS *
TOS, TOS *
;R1 Mem[B]
;R2 Mem[C]
Appendix A Solutions
Add R3,R1,R2
Add R1,R3,R2
Sub R4,R3,R1
Store A,R3
Store B,R1
Store D,R4
A.19
15
;R3 R1 + R2 = B + C
;R1 R3 + R2 = A + C
;R4 R3 R1 = A B
;Mem[A] R3
;Mem[B] R1
;Mem[D] R4
This problem is similar to A.7, but students are asked to consider how to implement
a looping structure for the different architecture styles as in problem A.18. We
assume A, B, C, and i are held in memory using addresses 1000, 3000, 5000, and
7000 as in problem A.7.
For the register-register case, we can use the MIPS code from A.7. In this case we
assume the locations for each of the variables as with A.7.a.
ex_19_1:
loop:
DADD
SW
LD
DSLL
DADDI
LD
LD
DADD
LD
DSLL
DADDI
SD
LD
DADDI
SD
LD
DADDI
BNEZ
R1,R0,R0
7000(R0),R1
R1,7000(R0)
R2,R1,#3
R3,R2,#3000
R4,0(R3)
R5,5000(R0)
R6,R4,R5
R1,7000(R0)
R2,R1,#3
R7,R2,#1000
0(R7),R6
R1,7000(R0)
R1,R1,#1
7000(R0),R1
R1,7000(R0)
R8,R1,#-101
R8,loop
;R0 = 0, initialize i = 0
;store i
;get value of i
;R2 = word offset of B[i]
;add base address of B to R2
;load B[i]
;load C
;B[i] + C
;get value of i
;R2 = word offset of A[i]
;add base address of A to R2
;A[i] B[i] + C
;get value of i
;increment i
;store i
;get value of i
;is counter at 101?
;if not 101, repeat
For the register-memory case, we can use the x86-64 code as in A.7 b):
ex_7_2:
loop:
movq
movq
movq
$0x0,%rax
$0x0,%rbp
%rax,0x1b58(%rbp)
movq
movq
mov
shl
movq
0x1388(%rbp),%rdx
0x1b58(%rbp),%rax
%rax,%rbx
$0x3,%rbx
0x0bb8(%rbx),%rcx
add
mov
shl
movq
%rdx,%cdx
%rax,%rbx
$0x3,%rbx
%rcx,0x03e8(%rbx)
# rax = 0, initialize i = 0
# base pointer = 0
# store i to location 7000
($1b58)
# load C from 5000 ($1388)
# get value of i
# rbx gets copy of i
# rbx now has i * 8
# load B[i] (3000 = $0bb8)
to rcx
# B[i] + C
# rbx gets copy of i
# rbx now has i * 8
# A[i] B[i] + C (base
address of A is 1000)
16
movq
add
movq
cmpq
jae
0x1b58(%rbp),%rax
$0x1,%rax
%rax,0x1b58(%rbp)
$0x0065,%rax
loop
#
#
#
#
#
get value of i
increment i
save i
is counter at 101 ($0065)?
if not 101, repeat
For the stack machine case, we will add an indirect addressing mode where an
address can placed on the stack. If a pushind (push with indirect addressing) is
used, it will take the value from the top of the stack as the address for the push.
For popind (pop with indirect addressing), the top of the stack is the address
for the target and the next entry on the stack is the value to be saved. We also
assume a jle instruction exists that takes the top three elements from the stack
and junps to the target (top of stack) if the second element is greater than the
third element from the top of the stack (the third is less than or equal to the
second).
ex_7_3:
loop:
push 0
pop i
push i
push 3
shl
pop tmp
push tmp
push 3000
add
pushind
push C
add
push 1000
push tmp
add
popind
push i
add 1
pop i
push i
push 100
push loop
jle
# put 0 on stack
# save i
# put i on stack
# push 3
# shift left by three (multiply by 8)
# temporary storage for i*8
# load i*8
# load address of B
# compute address of B[i]
# push with indirect addressing to get B[i]
# read value of C
# get sum of B[i] + C
# load address of A
# load i*8 (offset for A)
# compute address of A[i]
# pop with indirect addressing to assign sum
to A[i]
# get value of i
# increment by one
# save value of i
# get value of i
# termination value
# push loop target address
# jump to next iteration if i <= 100
For the accumulator machine, we can use the accumulator to compute an address
for indirect addressing when we read a value from memory. Unfortunately, when
we want to compute an address for saving a value to memory we will lose the
value to be saved while computing the address! So A[i] = expression becomes
hard to express. One could either add another register to support indirect addressing, which makes the machine no longer fit into the accumulator category. Alternately, one can employ self-modifying code to compute the A[i] address and then
update the immediate field for direct addressing. To do this, we can assume that
Copyright 2012 Elsevier, Inc. All rights reserved.
Appendix A Solutions
17
each instruction includes two words, one for the opcode and the other for the
immediate value. We can then use loadind in which the value in the accumulator
is used as the address for the load and store as before in which the value in the
accumulator is stored to a direct address (but we modify the address in the code
to patch it to what we need at the time).
ex_7_4:
loop:
Astore:
load 0
store i
load i
shl 3
store tmp
load tmp
add 1000
store Aaddress
load tmp
add 3000
loadind
add C
storeind
load i
add 1
store i
cmp 101
jle
A.20
# put 0 in acc
# save i
# put i in acc
# shift left by three (multiply by 8)
# temporary storage for i*8
# get i*8 in acc
# get address for A[i]
# store to immediate value for
instruction at label Astore
# get i*8 in acc
# compute address of B[i]
# use indirect addressing to get B[i]
# get sum of B[i] + C
# store to location in immediate value
of this instruction
# get value of i
# increment by one
# save value of i
# compare i to 101
# jump to next iteration if i <= 100
The total of the integer average instruction frequencies in Figure A.28 is 99%. The
exercise states that we should assume the remaining 1% of instructions are ALU
instructions that use only registers, and thus are 16 bits in size.
a. Multiple offset sizes are allowed, thus instructions containing offsets will be
of variable size. Data reference instructions and branch-type instructions contain offsets; ALU instructions do not and are of fixed length. To find the average length of an instruction, the frequencies of the instruction types must be
known.
The ALU instructions listed in Figure A.28 are add (19%), sub (3%), compare (5%), cond move (1%), shift (2%), and (4%), or (9%), xor (3%), and the
unlisted miscellaneous instructions (1%). The cond move instruction moves a
value from one register to another and, thus, does not use an offset to form an
effective address and is an ALU instruction. The total frequency of ALU
instructions is 47%. These instructions are all a fixed size of 16 bits. The
remaining instructions, 53%, are of variable size.
Assume that a load immediate (load imm) instruction uses the offset field to
hold the immediate value, and so is of variable size depending on the magnitude of the immediate. The data reference instructions are of variable size and
18
comprise load (26%), store (10%), and load imm (2%), for a total of 38%.
The branch-type instructions include cond branch (12%), jump (1%), call
(1%), and return (1%), for a total of 15%.
For instructions using offsets, the data in Figure A.31 provides frequency
information for offsets of 0, 8, 16, and 24 bits. Figure 2.42 lists the cumulative fraction of references for a given offset magnitude. For example, to determine what fraction of branch-type instructions require an offset of 16 bits
(one sign bit and 15 magnitude bits) find the 15-bit magnitude offset branch
cumulative percentage (98.5%) and subtract the 7-bit cumulative percentage
(85.2%) to yield the fraction (13.3%) of branches requiring more than 7 magnitude bits but less than or equal to 15 magnitude bits, which is the desired
fraction. Note that the lowest magnitude at which the cumulative percentage
is 100% defines an upper limit on the offset size. For data reference instructions a 15-bit magnitude captures all offset values, and so an offset of larger
than 16 bits is not needed. Possible data-reference instruction sizes are 16, 24,
and 32 bits. For branch-type instructions 19 bits of magnitude are necessary
to handle all offset values, so offset sizes as large as 24 bits will be used. Possible branch-type instruction sizes are 16, 24, 32, and 40 bits.
Employing the above, the average lengths are as follows.
ALU length = 16 bits
Data-reference length = 16 0.304 + 24 (0.669 0.304) + 32 (1.0 0.669)
= 24.2 bits
Branch-type length = 16 0.001 + 24 (0.852 0.001) + 32 (0.985 0.852)
= 25.3 bits
Average length = 16 0.47 + 24.2 0.38 + 25.3 0.15
= 20.5 bits
b. With a 24-bit fixed instruction size, the data reference instructions requiring a
16-bit offset and the branch instructions needing either a 16-bit or 24-bit offset cannot be performed by a single instruction. Several alternative implementations are or may be possible.
First, a large-offset load instruction might be eliminated by a compiling optimization. If the value was loaded earlier it may be possible, by placing a
higher priority on that value, to keep it in a register until the subsequent use.
Or, if the value is the stored result of an earlier computation, it may be possible to recompute the value from operands while using no large-offset loads.
Because it might not be possible to transform the program code to accomplish either of these schemes at lower cost than using the large-offset load,
lets assume these techniques are not used.
For stores and branch-type instructions, compilation might be able to rearrange code location so that the store effective address or target instruction
address is closer to the store or branch-type instruction, respectively, and thus
require a smaller offset. Again, lets assume that this is not attempted.
Appendix A Solutions
19
Thus, code needing an offset greater than 8 signed magnitude bits and written
with 24-bit fixed-size instructions will use additional instructions to change
the program counter in multiple steps. If adding the available offset to the
low-order PC register bits is all that is available, then to achieve an offset with
a 15-bit magnitude requires adding up to 28 8-bit offsets (only 7 magnitude
bits each) to the initial PC value. Clearly, this is beyond tedious and performance-robbing.
To avoid this performance penalty we assume that there are two new jump
instructions to help with changing the PC. They are JUMP_Shift7 and
JUMP_Shift14. These instructions add an 8-bit offset not at the low-order bit
positions, but rather at a position shifted left by 7 and 14 bits, respectively,
from the least significant bit. During the addition low order bit positions are
filled with 7 or 14 zeros. With these two instructions and the original JUMP
with unshifted offset, offsets of 8, 15, or 22 total bits including sign can be
accomplished piece wise. First an 8-bit offset is added to the PC using JUMP.
Next, another 8-bit offset is added to the PC using JUMP_Shift7. Finally, a
third 8-bit offset is added to the PC using JUMP_Shift14. If the needed
signed magnitude offset can be represented in 8 bits then the original instructions are sufficient. If up to 15 bits are needed for an offset, then a
JUMP_Shift7 precedes the original instruction, moving the PC to within the
range that can be reached by an 8-bit signed magnitude offset. To keep program instruction layout in memory more sequential a JUMP_Shift7 can be
used to restore the PC to point of the next sequential instruction location following the first JUMP_Shift7. Similarly, if up to 22 offset bits are needed,
then a JUMP_Shift7 followed by a JUMP_Shift14 changes the PC value to
within the range of an 8-bit offset of the intended address. Analogously, a
JUMP_Shift14 and a JUMP_Shift7 would follow the original instruction to
return the PC to the next sequential instruction location.
From Figure A.31 no data reference instruction requires more than a 15-bit
offset, so JUMP_Shift14 is not needed. JUMP_Shift7 instruction pairs are
needed for offsets greater than 7 magnitude bits but less than or equal to
15 bits, or 100% 66.9% = 33.1% of data reference instructions. Similarly,
JUMP_Shift14 and JUMP_Shift7 pairs are needed for 100% 98.1% = 1.9%
of branches, and JUMP_Shift7 pairs alone are needed for 98.1% 85.2% =
12.9% of branches. Thus, the number of fetched instructions increases by a
factor of
(2)(0.331)(0.38) + [(2)(0.129) + (4)(0.019)](0.15) = 30.2%
The number of bytes fetched on a per-instruction basis for the byte-variablesized instructions of part (a) is
(20.5 bits/instr)((1/8)byte/bit) = 2.56 bytes/instr.
Because extra instructions must be added to the original code, we compare
the number of bytes fetched for the fixed length 24-bit instructions on a per
original-instruction basis. The amount is
(1 + 0.302)(24 bits/instr)((1/8) bytes/bit) = 3.91 bytes/original instr.
20
c. With a 24-bit fixed offset size, the ALU instructions are 16 bits long and both
data reference and branch instructions are 40 bits long. The number of
instructions fetched is the same as for byte-variable-sized instructions, so the
per-instruction bytes fetched is
(16 bits/instr)(0.47) + (40 bits/instr)(0.38 + 0.15) = 3.59 bytes/instr.
This is less that the bytes per original instruction fetched using the limited 8bit offset field instruction format. Having an adequately sized offset field to
avoid frequent special handling of large offsets is valuable.
A.21
No solution provided.
A.22
4F
4D
50
55
54
45
52
45
52
55
54
4D
50
43
4F
b.
c. 4F4D, 5055, and 5455. Other misaligned 2-byte words would contain data
from outside the given 64 bits.
d. 52 55 54 4D, 55 54 4D 50, and 54 4D 50 43. Other misaligned 4-byte words
would contain data from outside the given 64 bits.
A.23
Appendix A Solutions
21
3) Cloud computing ISAs should provide support for virtualization. Integer operation and memory management will be important. Power savings and facilities
for easier system monitoring will be important for large data centers. Security
and encryption of user data will be important, so ISAs may provide support for
these concerns.
4) Power consumption is a big concern in embedded computing. Price is also
important. The embedded ISAs may provide facilities for sensing and controlling
devices with special instructions and/or registers, analog to digital and digital to
analog converters, and interfaces to specialized, application-specific functions.
The ISA may omit floating point, virtual memory, caching, or other performance
optimizations to save energy or cost.