Efficient Analysis of Pipeline Models for WCET Computation Stephan - - PowerPoint PPT Presentation

efficient analysis of pipeline models for wcet computation
SMART_READER_LITE
LIVE PREVIEW

Efficient Analysis of Pipeline Models for WCET Computation Stephan - - PowerPoint PPT Presentation

A b s I n t Efficient Analysis of Pipeline Models for WCET Computation Stephan Wilhelm (sw@absint.com) AbsInt GmbH and Saarland University Outline 1. aiTs pipeline analysis 3. Why improve its efficiency? 5. BDD based pipeline analysis 7.


slide-1
SLIDE 1

Efficient Analysis of Pipeline Models for WCET Computation

Stephan Wilhelm (sw@absint.com) AbsInt GmbH and Saarland University

A b s I n t

slide-2
SLIDE 2

Outline

  • 1. aiT’s pipeline analysis
  • 3. Why improve its efficiency?
  • 5. BDD based pipeline analysis
  • 7. Generating pipeline analyzers
  • 9. Status and prospects
slide-3
SLIDE 3

The aiT-Framework

Executable Program Control-Flow Reconstruction Value Analyzer Worst-Case Execution Path Analysis WCET Visualization Pipeline Analyzer Cache Analyzer

slide-4
SLIDE 4

Cache Analysis

Pipeline Analysis

Pipeline Abstraction

b

c e a d

Pipeline Model CFG

instruction A instruction B

a

b

c d e a d e

Timing Anomalies!

State explosion.

slide-5
SLIDE 5

Timing Anomalies

  • local worst case vs. global worst case

– cache hit on global worst-case path (Lundquist/Stenström) – “unintuitive behavior” – no upper bound

  • Caused by features like: caches, out-of-order execution,

branch prediction, ... No local worst-case decisions. State explosion.

slide-6
SLIDE 6

Symbolic Representation of FSM State Sets

a = 1 b = 1 a = 0 b = 0 a = 1 b = 0 a = 0 b = 1

A 2-bit automaton

  • BDD size depends on variable ordering
  • worst case: exponential
  • in most cases better orderings are possible

01 00 01 11 10 00 01

FSM state sets 1 characteristic functions

¬a∧b ¬a

BDD’s

a b 1 a 1 1

slide-7
SLIDE 7

BDD based Pipeline Analysis

00 01 10

00: 01 01: 10 10: 01

11

A 2-bit automaton The program

10 00 01 11

  • 1. Partition the set

10 00 10 00 01 11

  • 2. Compute images

10 10 00 01

  • 3. Apply inputs

10 00 01

  • 4. Compute union

01 11 10 00

Start set

slide-8
SLIDE 8

Model Specification & Validation

  • Documentation is often incorrect/incomplete.
  • Model specification is error-prone.
  • Trace validation is difficult.

Generate analysis from HDL specification.

  • Verilog and VHDL can be compiled into FSM’s

(Cheng, 1994).

– Implemented in VIS model checker.

slide-9
SLIDE 9

Generating Pipeline Analyzers

HDL Pipeline Analyzer Program (CFG) Simplified HDL FSM specification Model Checker Properties Abstraction HDL->FSM compiler

slide-10
SLIDE 10

Status and Prospects

1. Implementation of proof-of-concept.

– For simple examples. – Only one basic block.

2. Integration with DFA framework (PAG). 3. Integration with CFG representation. 4. Specification of a real pipeline.

– Infineon Tricore 2

5. Compare against “classic aiT”.

today future past