Transcript Lec4-isa - ECE Users Pages - Georgia Institute of Technology

ECE 4100/6100
Advanced Computer Architecture
Lecture 4 ISA Taxonomy
Prof. Hsien-Hsin Sean Lee
School of Electrical and Computer Engineering
Georgia Institute of Technology
Instruction Set Architecture
• Specification of a microprocessor design
• Interface between user and machine’s functionality
• Good instruction set design principles
–
–
–
–
–
Compatibility
Implementability
Programmability
Usability
Encoding efficiency
2
Main ISA Design Philosophy
• CISC (Complex Instruction Set Computer)
• RISC (Reduced Instruction Set Computer)
• VLIW (Very Long Instruction Word)
• EPIC (Explicitly Parallel Instruction Computer)
3
CISC
• Complex Instruction Set Computers
• Close “semantic gap” between programming and execution
– Smaller code size (memory was expensive!)
– Simplify compilation
• Another state machine (controlled by microcode) inside the
machine
• Example: x86, Intel 432, IBM 360, DEC VAX
4
CISC Example: x86
• MOVSD ;; move a double word, 1-byte instruction
MOVSD // m32[DS:EDI] = m32[DS:ESI]
• REP;; 1-byte prefix to repeat string operations
REP MOVSD // count set up in ECX
LOCK ADD ds:[esi+ecx*2+0x67452301], 0xEFCDAB89 // 13-byte
F0 3E 81 84 4E 01 23 45 67 89 AB CD EF
prefix
[--][--]+disp32
ESI+ECX*2
5
RISC
• Observation made by IBM (John Cocke, Eckert-Mauchly
Award’85, Turing Award’87, Nat’l Medal of Technology’91,
Nat’l Medal of Science’94)
– Few of the available instructions are used
• CISC : “n+1” phenomenon
– Adding an instruction requiring an extra level of decoding
logic can slow down the entire ISA
• Reduced Instruction Set Computer
– Originated at IBM in 1975, a telephone project
• To achieve 12 MIPS (300 calls per sec, 20k inst per call)
• Simple instructions
– IBM 801 in 1978
– More compiler effort to gain performance
6
A Typical RISC
•
•
•
•
•
•
Smaller number of instructions
Fixed format instruction (e.g., 32 bits)
3-address, reg-to-reg arithmetic instructions
Single cycle operation for execution
Load-store architecture
Simple address modes
– Base + displacement
– No indirection
•
•
•
•
Simple branch conditions
Hardwired control (No microcode)
More compiler effort
Examples:
– RISC I and RISC II at Berkeley
– MIPS (Microprocessors without Interlocked Pipe Stage) at Stanford
– IBM RISC Technology, Sun Sparc, HP PA-RISC, ARM
7
RISC Example: MIPS
R-format (Register-Register)
31
26 25
Op
21 20
Rs
16 15
Rt
Rd
11 10
6 5
Shamt
0
add $1, $2, $3
Funct
I-format (Register-Immediate)
31
26 25
Op
21 20
Rs
16 15
0
addi $1, $2, -5
immediate
Rt
I-format (Load/Store)
31
26 25
Op
21 20
Base
16 15
Dest
0
lw $1, 24($9)
immediate
I-format (Branch)
31
26 25
Op
21 20
Rs
16 15
0
immediate
Rt
beq L1, $4, $0
J-format (Jump / Call)
31
26 25
Op
0
target
j L2
8
CISC vs. RISC
CISC
RISC
Variable length instructions
Fixed-length instructions, single-cycle
operation
Abundant instructions and addressing
modes
Fewer instructions and addressing
modes
Long, complex decoding
Simple decoding
Contain mem-to-mem operations
Load/store architecture
Use microcode
No microinstructions, directly decoded
and executed by HW logic
Closer semantic gap (shift complexity
to microcode)
Needs smart compilers, or intelligent
hardware to reorder instructions
IBM 360, DEC VAX, x86, Moto 68030
IBM 801, MIPS, RISC I, IBM POWER,
Sun Sparc
• Some definitions were from the paper by Colwell et al. in 1985
9
CISC vs. RISC (Reality)
CISC
RISC
IBM
370/168
VAX
11/780
Xerox
Dorado
IBM
801
Berkeley
RISC1
Stanford
MIPS
Year
introduced
1973
1978
1978
1980
1981
1983
#
instructions
208
303
270
120
39
55
Microcode
54KB
61KB
17KB
0
0
0
Instruction
size
2 to 6 B
2 to 57 B
1 to 3 B
4B
4B
4B
Execution
model
Reg-reg
Reg-mem
Mem-mem
Reg-reg
Reg-mem
Memmem
Stack
Reg-reg
Reg-reg
Reg-reg
10
Observation and Controversy
•
”Instruction Set and Beyond: Computers, Complexity and Controversy” by Bob
Colwell (Eckert-Mauchly Award, 2005) and gang from CMU, also see response
from RISC camp: Patterson (Eckert-Mauchly Award, 2008) and Hennessy (EckertMauchly Award, 2001)
• CISC/RISC classification should *not* be a dichotomy
• Case in point: MicroVAX-32 by DEC, a single chip implementation
– Subsetting VAX instructions (but still, 175 instructions!)
– Emulate complex instructions
– a RISC or a CISC? (Well, it has variable length instructions, not a ld/st
machine, with a microcode control, have all VAX addressing mode)
• Effective processor design = CISC experiences + RISC tenets
• RISC features are not incompatible or mutually exclusive
– Large register file (w/ register windows)
• RISC/CISC issues are best considered in light of their function-toimplementation level assignment
11
Modern X86 Machine Design
• CISC outfit
• RISC inside
• E.g., Intel P6/Netburst/Core, AMD Athlon/Phenom/Opteron
• Each x86 instruction is decoded into “micro-op” (op) or
“RISC-op” on-the-fly
• Internal microarchitecture resembles RISC design philosophy
• Processor dynamically schedules “ops”
• Compiler’s scheduling is still beneficial
12
Recent ISA Design Trend
• Look at this instruction in MIPS (CISC or RISC?)
CABS.LE.PS $fcc0, $f8, $f10 ;; |y||w| , |x||w|?
• Many complex instructions emerged for new apps
– Viterbi instruction for wireless communication/DSP
– Sum of absolute differences in SSE (PSAD) or other DSP: C = |A-B|
for MPEG (motion estimation)
• In embedded domain, code size is critical
• Reducing programming efforts
• Optimizing performance via
– Specialized hardware (accelerator-based)
– Co-processor (controlled by main processor)
– ISA plug-in (flexible)
13
VLIW
•
•
Very Long Instruction Word
– Originated from microcode compaction
– Coined by Josh Fisher (Eckert-Mauchly Award, 2003)
Compiler will
–
–
•
Perform instruction scheduling (latency-aware)
Pack several independent instructions into a VLIW instruction
Issues
–
–
–
Compatibility
Many nop’s
Very complex compiler
•
•
Information unavailable at static compile time
interprocedural optimization is difficult)
Pioneers
• Culler Scientific
–
•
Multiflow (Fisher)
–
–
•
Led by Prof. Glen J. Culler (National Medal of Technology winner 2000, Berkeley Prof. David Culler’s father)
Led by Josh Fisher (Eckert-Mauchly Award 2003), John O’Donnell, John Ruttenberg, David Papworth, Bob Colwell
(Eckert-Mauchly Award 2005), Geoffery Lowney, etc.
Several Multiflow TRACE were delivered
Cydrome (Rau, Yen’s) in the 80’s
–
–
Led by Bob Rau (Eckert-Mauchly Award 2002), David Yen, Wei Yen, etc.
Had a working prototype
Modern Processors
• Most DSP embrace VLIW (e.g., TI C6x, StarCore, ADI TigerSHARC, etc.)
• Transmeta Crusoe (internal, never released ISA)
14
Intel/HP EPIC
• Explicitly Parallel Instruction Computer
• A kin breed of VLIW (e.g., compiler holding the key to high
performance)
• Some new features
– Stop bits to address compatibility
– ISA enabling data speculation and control speculation (minimum
hardware support needed)
– Fully predicated ISA
– Rotating registers, RSE (not so new, e.g., MRS in RISC I)
• Lots of ideas from Polycyclic architecture (TRW) and Cydrome
by the late Bob Rau (Eckert-Mauchly Award, 2002)
An Itanium Instruction Bundle
ld4 r43=[r38]
add r38=16,r38
br.call.sptk b0=printf# ;;
15
VLIW Tradeoffs
• Plentiful registers, simple encodings, …
• Potentially lower # of transistors than other designs
– Reduced speculation, OoO not needed
– Size efficiencies, price, power consumption
– Is this true for Itanium?
• Drawbacks
– Backward compatibility or upgradeability
– Due to exposed implementation details
• VLIW is orthogonal to other techniques
– Pipeline, SMT, and CMP/Multi-core can be built on top of processors
including VLIW
16
Design Philosophy: VLIW vs. Superscalar
Static _VOID
_DEFUN(_mor_nu),
struct _reent
*ptr _AND
register size_t
{ .
.
.
Same
Normal
Source code
Static _VOID
_DEFUN(_mor_nu),
struct _reent
*ptr _AND
register size_t
{ .
.
.
RISC
Object code
Normal
Compiler
IM1 = I–1
IM2 = I–2
IM3 = I–3
T1 = LOAD .
T3 = 2*T1
.
.
Scheduling and
Operation
Independence:
Recognizing
hardware
Run-time
Compile Time
The same ILP
Hardware in
Both cases
Normal compiler
plus scheduling
and operation
Independence:
Recognizing
software
17
Design Philosophy: VLIW vs. Superscalar
• VLIW
– Requiring less hardware and lower power
– Programs need to be changed to run correctly
when even small changes (not always though)
• Superscalar
– Object-code compatible
•Sequential programs can be presented to different
superscalar implementation of the same ISA
18
Design Philosophy: VLIW vs. Superscalar
19
Superscalar or VLIW?
• Reality: the current world is dominated by …
– X86: Core (quad-issue) & ATOM (dual-issue)
– And ARM (Cortex A8 is a dual-issue; A9 has OOO)
• VLIW is largely embraced by the DSP camp
20
Should we continue to teach this Chapter about ISA?
21