Superscalar Processors

7.1 Introduction 7.2 Parallel decoding 7.3 Superscalar instruction issue 7.4 Shelving 7.5 Register renaming 7.6 Parallel execution 7.7 Preserving the sequential consistency of instruction execution 7.8 Preserving the sequential consistency of exception processing 7.9 Implementation of superscalar CISC processors using a superscalar RISC core 7.10 Case studies of superscalar processors

Superscalar Processors vs. VLIW

TECH
CH01

Computer Science

Superscalar Processor: Intro

Parallel Issue Parallel Execution {Hardware} Dynamic Instruction Scheduling Currently the predominant class of processors
Pentium PowerPC UltraSparc AMD K5HP PA7100DEC

Emergence and spread of superscalar processors

Evolution of superscalar processor

Specific tasks of superscalar processing

Parallel decoding {and Dependencies check}

What need to be done

Decoding and Pre-decoding

Superscalar processors tend to use 2 and sometimes even 3 or more pipeline cycles for decoding and issuing instructions >> Pre-decoding:
shifts a part of the decode task up into loading phase resulting of pre-decoding
f the instruction class f the type of resources required for the execution f in some processor (e.g. UltraSparc), branch target addresses calculation as well

the results are stored by attaching 4-7 bits

+ shortens the overall cycle time or reduces the number of cycles needed

The principle of perdecoding

Number of perdecode bits used

Specific tasks of superscalar processing: Issue

7.3 Superscalar instruction issue

How and when to send the instruction(s) to EU(s)

Issue policies

Instruction issue policies of superscalar processors: ---Performance, tread-----

Issue rate {How many instructions/cycle}

CISC about 2 RISC:

Issue policies: Handing Issue Blockages

Issue stopped by True dependency

True dependency (Blocked: need to wait)

Issue order of instructions

Aligned vs. unaligned issue

Issue policies: Use of Shelving

Direct Issue

The principle of shelving: Indirect Issue

Design space of shelving

Scope of shelving

Layout of shelving buffers

Implementation of shelving buffer

Basic variants of shelving buffers

Using a combined buffer for shelving, renaming, and reordering

Number of shelving buffer entries

Number of read and write ports

how many instructions may be written into (input ports) or read out from (output parts) a particular shelving buffer in a cycle depend on individual, group, or central reservation stations

Shelving: Operand fetch policy

7.4.4 Operand fetch policies

Operand fetch during instruction issue

Operand fetch during instruction dispatch

Reg. file

Reg. file

Shelving: Instruction dispatch Scheme //

7.4.5 instruction dispatch scheme

- Dispatch policy
Selection Rule
Specifies when instructions are considered executable e.g. Dataflow principle of operation
f Those instructions whose operands are available are executable.

Dispatch policy: Dispatch order

Arbitration Rule
Needed when more instructions are eligible for execution than can be disseminated. e.g. choose the oldest instruction.

Dispatch order
Determines whether a non-executable instruction prevents all subsequent instructions from being dispatched.

Trend of Dispatch order

-Dispatch rate (instructions/cycle)

Maximum issue rate <= Maximum dispatch rates >> issue rate reaches max more often than dispatch rates

- Scheme for checking the availability of operands: The principle of scoreboarding

Schemes for checking the availability of operand

Operands fetched during dispatch or during issue

Use of multiple buses for updaing multiple individual reservation strations

Interal data paths of the powerpc 604

42

-Treatment of an empty reservation station

7.4.6 Detail Example of Shelving

Issuing the following instruction
cycle i: mul r1, r2, r3 cycle i+1: ad r2, r3, r5 ad r3, r4, r6
f format: Rs1, Rs2, Rd

Example overview

Cycle i: Issue of the mul instruction into the reservation station and fetching of the corresponding operands

Cycle i+1: Checking for executable instructions and dispatching of the mul instruction

Cycle i+1 (2nd phase): Issue of the subsequent two ad instructions into the reservation station

Cycle i+2: Checking for executable instruction (mul not yet completed)

Cycle i+3: Updating the FX register file with the result of the mul instruction

Cycle i+3 (2nd phase): Checking for executable instructions and dispatching the older ad instruction

=Instruction Issue policies:Register Renaming

Register Remaining and dependency

three-operand instruction format e.g. Rd, Rs1, Rs2 False dependency (WAW)
mul r2, , add r2, ,
f two different rename buffer have to allocated

Choronology of introduction of renaming (high complexity, Sparc64 used 371K transistors that is more than i386)

True data dependency (RAW)

mul r2, , ad , r2,
f rename to e.g. f mul p12, , f ad , p12, .

Static or Dynamic Renaming

>Design space of register renaming

-Scope of register renaming

-Layout of rename buffers

-Type of rename buffers

Rename buffers hold intermediate results

Each time a Destination register is referred to, a new rename register is allocated to it. Final results are stored in the Architectural Register file Access both rename buffer and architectural register file to find the latest data,
if found in both, the data content in rename buffer (the intermediate result) is chosen.

When an instruction completed (retired),

(ROB) {retire only in strict program sequence} the correspond rename buffer entry is writing into the architectural register file (as a result modifying the actual program state) the correspond rename buffer entry can be de-allocated

-Number of rename buffers

-Basic mechanisms used for accessing rename buffers

Rename buffers with associative access (latter e.g.) Rename buffers with indexed access
(always corresponds to the most recent instance of renaming)

-Operand fetch policies and Rename Rate

rename bound: fetch operands during renaming (during instruction issue) dispatch bound: fetch operand during dispatching Rename Rate
the maximum number of renames per cycle equals the issue rate: to avoid bottlenecks.

7.5.8 Detailed example of renaming

renaming:
mul r2, r0, r1 ad r3, r1, r2 sub r2, r0, r1

format:
op Rd, Rs1, Rs2

Assume:
separate rename register file, associative access, and operand fetching during renaming

Structure of the rename buffers and their supposed initial contents Latest bit: the most recent rename 1, previous 0

Renaming steps
Allocation of a free rename register to a destination register Accessing valid source register value or a register value that is not yet available Re-allocation of destination register Updating a particular rename buffer with a computed result De-allocation of a rename buffer that is no longer needed.

Allocation of a new rename buffer to destination register (circular buffer: Head and Tail) (before allocation)

(After allocation) of a destination register

Accessing abailable register values

Accessing a register value that is not yet available

3 is the index

Re-allocate of r2 (a destination register)

Updating the rename buffers with computed result of {mul r2, r0, r1} (register 2 with the result 0)

Deallocation of the rename buffer no. 0 (ROB retires instructions) (update tail pointer)

7.6 Parallel Execution

Executing several instruction in parallel
instructions will generally be finished in out-of-program order

to finish
operation of the instruction is accomplished, except for writing back the result into
f the architectural register or f memory location specified, and/or f updating the status bits

to complete
writing back the results

to retire (ROB)
write back the results, and delete the completed instruction from the last ROB entry

7.7 Preserving Sequential Consistency of instruction execution // Multiple EUs operating in parallel, the overall instruction execution should >> mimic sequential execution
the order in which instruction are completed the order in which memory is accessed

Sequential consistency models

Consistency relate to instruction completions or memory access

Trend and performance

Allows the reordering of memory access

it permits load/store reordering
either loads can be performed before pending stores, or vice versa a load can be performed before pending stores only IF none of the preceding stores has the same target address as the load

Using Re-Order Buffer (ROB) for Preserving: The order in which instruction are <completed>
1. Instruction are written into the ROB in strict program order:
One new entry is allocated for each active instruction

2. Each entry indicates the status of the corresponding instruction

issued (i), in execution (x), already finished (f)

it makes Speculative loads or stores feasible

When addresses of pending stores are not yet available, speculative loads avoid delaying memory accesses, perform the load anywhere. When store addresses have been computed, they are compared against the addresses of all younger loads. Re-load is needed if any hit is found.

3. An instruction is allowed to retire only if it has finished and all previous instruction are already retired.
retiring in strict program order only retiring instructions are permitted to complete, that is, to update the program state:
f by writing their result into the referenced architectural register or memory

it allows cache misses to be hidden

if a cache miss, it allows loads to be performed before the missed load; or it allows stores to be performed before the missed store.

Principle of the ROB {Circular Buffer}

Introduction of ROBs in commercial superscalar processors 7.61

Use ROB for speculative execution

Guess the outcome of a branch and execution the path
before the condition is ready

Design space of ROBs

1. Each entry is extended to include a speculative status field

indicating whether the corresponding instruction has been executed speculatively

2. speculatively executed instruction are not allow to retire

before the related condition is resolved

3. After the related condition is resolved,

if the guess turn out to be right, the instruction can retire in order. if the guess is wrong, the speculative instructions are marked to be cancelled. Then, instruction execution continue with the correct instructions.

Basic layout of ROBs

ROB implementation details

7.8 Preserving the Sequential consistency of exception processing When instructions are executed in parallel,
interrupt request, which are caused by exceptions arising in instruction <execution>, are also generated out of order.

Sequential consistency of exception processing

If the requests are acted upon immediately,

the requests are handled in different order than in a sequential operation processor called imprecise interrupts

Precise interrupts: handling the interrupts in consistent with the state of a sequential processor

Use ROB for preserving sequential order of interrupt requests

Interrupts generated in connection with instruction execution
can handled at the correct point in the execution, by accepting interrupt requests only when the related instruction becomes the next to retire.

7.9 Implementation of superscalar CISC processors using superscalar RISC core CISC instructions are first converted into RISC-like instructions <during decoding>.
Simple CISC register-to-register instructions are converted to single RISC operation (1-to-1) CISC ALU instructions referring to memory are converted to two or more RISC operations (1-to-(2-4))
f SUB EAX, [EDI]
converted to e.g.

More complex CISC instructions are converted to long sequences of RISC operations (1-to-(more than 4))

f MOV EBX, [EDI] f SUB EAX, EBX

On average one CISC instruction is converted to 1.5-2 RISC operations

The peinciple of superscalar CISC execution using a superscalar RISC core

PentiumPro: Decoding/converting CISC instructions to RISC operations (are done in program order)

Case Studies: R10000 Core part of the micro-architecture of the R10000

Case Studies: PowerPC 620

67

Case Studies: PentiumPro Core part of the micro-architecture

PentiumPro Long pipeline: Layout of the FX and load pipelines