• DocumentCode
    2532888
  • Title

    BiDIFS and CoDEBTS: Exploiting Compiler-Processor Interaction for Performance Enhancement

  • Author

    Aditya, Pratyush ; Kulkarni, Parag

  • fYear
    2008
  • fDate
    19-21 Nov. 2008
  • Firstpage
    1
  • Lastpage
    6
  • Abstract
    This paper presents a novel approach towards the architecture of a 32-bit RISC pipeline and a compiler for it, which aim to improve performance by ldquoreducing the stalls caused by branch hazardsrdquo. It presents two techniques - Bimodal Dual-Instruction Fetch Stage (BiDIFS) and Compiler Dependent Branch Target Stack (CoDeBTS). CoDeBTS is a prediction scheme which give close to 100% accurate prediction for the specific types of branches that they are suggested for, whereas BiDIFS is a novel architecture that has no branch penalties. Both the techniques aim to exploit the advantages of modern compilers coupled with architectures specifically targeted for being able to use the valuable information that is gathered by the compiler in its various phases especially semantic analysis. Compiler Processor interaction techniques have enormous potential for improving both performance and reducing power consumption, here we propose just two of them. Both the architectures are proposed, implemented and evaluated.
  • Keywords
    microprocessor chips; parallel architectures; pipeline processing; program compilers; reduced instruction set computing; storage management; BiDIFS; CoDEBTS; RISC pipeline; bimodal dual-instruction fetch stage; branch hazard; branch penalties; branch prediction; compiler dependent branch target stack; compiler-processor interaction; performance enhancement; prediction scheme; semantic analysis; word length 32 bit; Computer aided instruction; Concurrent computing; Coupled mode analysis; Decoding; Energy consumption; Hardware; Hazards; Information analysis; Pipelines; Reduced instruction set computing;
  • fLanguage
    English
  • Publisher
    ieee
  • Conference_Titel
    TENCON 2008 - 2008 IEEE Region 10 Conference
  • Conference_Location
    Hyderabad
  • Print_ISBN
    978-1-4244-2408-5
  • Electronic_ISBN
    978-1-4244-2409-2
  • Type

    conf

  • DOI
    10.1109/TENCON.2008.4766860
  • Filename
    4766860