search

willi19 / swpp202301-compiler-team6

MIT License
0 stars 0 forks source link

Summary of external meetings on 5/20 #31

Open sharaelong opened 1 year ago

sharaelong commented 1 year ago

  1. Deep performance analysis

    • Cost file format is strange(??)
    • instcombine pass gives powerful performance than LICMPass, but only bubblesort case optimization harms performance.

  2. Furious optimization idea

    • small values in one 64bit memory
    • Prerequisites: recognize continuous two similar operations in loop or elsewhere

  3. Oracle pass idea

    • Extract some loop from code, push into oracle.
    • Or extract multiple 'small' loops and distinguish 'mode' inside oracle.
    • Before call oracle, store variable into memory. After call oracle, load variable from memory.

  4. Final results We decided to implement function inline pass, oracle pass, free-remove pass, and advance of arithmetic pass. Maybe more pass will be implemented in sprint 3.

  5. Merge NFC PR firstly.