230 likes | 344 Vues
Motivation for the Design of an Instruction Set. Consider the invented assembly language instruction add x, y, z y is added to z , and the result is written to x. 4 fields are implied within the machine code for this instruction
E N D
Motivation for the Design of an Instruction Set Consider the invented assembly language instruction add x, y, z y is added to z, and the result is written to x
4 fields are implied within the machine code for this instruction The number of unique instructions determines the size of the op code field 8 bits will be enough. (MIPS has 6.) z y x addop code addresses (assume 32 bits)
The goal: Compute how much time it takes to fetch and execute an instruction. Then, by design, make it faster. P M Fact: memory accesses take much longer than processor operations.
processor operation : memory access 1 : 20 ish For this exercise, use 1 : 10 Each memory access transfers 32 bits of data in parallel (or, at one time). This implies that there are 32 wires connecting P to M.
add x, y, z fetch instruction PC update decode get operands do operation (add) store result requires a memory access
Machine code for this add is more than 96 bits • instruction fetch requires3 memory accesses • 2 operands to load (y and z) requires2 memory accesses • 1 result to store (x) requires1 memory access
using 1:10 ratio, 62 60/62 is spent waiting for memory !
new approach: redesign instructions with 2 operands add x, y does x x + y y x addop code addresses (again, 32 bits)
original instruction: add x, y, z becomes copy x, z add x, y move add + 41 52 = 93
MIPSSPARCALPHAPowerPC instruction set with • 3 operands is a 3-address architecture • 2 operands is a 2-address architecture • 1 operand is a 1-address architecture Intel x86
1 operand ? How to add ? from memory Acc to memory add x does Acc Acc + x
add x, y, z becomes load z # Acc z add y # Acc Acc + y store x # mem Acc an address (again, 32 bits) op code
store load add + + 21 22 21 = 64
But, what if the example includes more than 1 instruction? x (y + z) / 3 On the 3 address architecture: add x, y, z div x, x, 3 3 x x immediate div addresses
on the 3-address architecture add div immediatevalue 3"loaded" atfetch ! + 62 52 = 114
on the 2-address architecture: copy x, z add x, y div x, 3 must load x, but 3 is in the instruction div copy add + + 41 52 42 = 135
immediatevalue 3 is in the instruction on the 1-address architecture: load z add y div 3 store x div store load add + + + 21 22 12 21 = 76
The best parts of the ideas: • Make instructions small, to minimize fetch time • Keep the Accumulator, to avoid extra loads and stores • In fact, we'd like to have several Accumulators. . . all inside the processor call them registers
All modern processors have sets of registers, kept in a register file. • The MIPS R2000 has a set of 3232-bit registers. . .
These architectures are categorized asload/store architectures. • operands for arithmetic and logical instructions only come from and go to registers . . . not memory • other instructions (loads and stores) read from or write to memory
Our invented example: add x, y, z becomes more like load reg1, y load reg2, z add reg3, reg2, reg1 store reg3, x addresses
consider the add instruction add reg3, reg2, reg1 For 32 registers, it only takes 5 bits. instruction size ≈ 8 + 5+ 5 + 5 = 23 bits
add load 14 22