Page 1

Compiler Optimizations in the AI Error.

• https://xuhongxu.com

• Research Interest: Software Engineering

• DouBou ( ByteDance’s gpt)

• teXT-TO-SPEECH ENGINE

int max (int a, int b){
if(a > b) reuturn a;
return b;
}

• Intermediate Representations

We can solve any problem by introducing an extra level of indirection.

• Compiler Frontend - parses code into itneremdiate repreented

• Compiler Backend Turns IR into targeted assembly code

  • Optimizations happen here.

    • ASTs are not made for optimizations,

      • hard to trap independencies and too like your code.

  • IR uses specific structure, that you can easily capture data dependencies.

  • tO oPTIMIZE mACHINE ir/asm

• Compiler Middle-End

  • Platform-Independent Optimizations to OPtimize IR

Now adays, keras, and tensorflow, if translated to IR, we lose a lot of info if we optimize at the end.

• Focused on Compiler Middle-End

• IR translated from soruce code.

  • Thus why we store and load the same values repeatedly?

    • When it compilers, all it cares about is correction and translateion as simple as possible.

There exists a translation path that removes store and redundant redundants.

• SSA

  • Step single Assembly

• sgt = step greater than

• SimplifyCFG, eliinates Branching.

  • Works like ternary operators.

• InstCombine

  • Arithimitic optimizations.

Original, line has 20 lines, now we have 1.

• that's fucking insane.

• we note that f has redundant code, how does CLANG handle this?

• Sometiems tho compielrs are just stupid.

• How can we fix this??

• InstantaCombine Pass is continuously evolving, around 50,000 lines of code.

• Detecting Missing PeeopHole OPtimizations

  • Manual INspections

    • Rlies on domain expertise

    • Labor-intensive and not scalable

    • Ad-hoc.

      • Some developers can't generalize.

  • Differential Testing

    • Requires Two langauge implementations

    • Cannot find new optimizations.

  • Superoptimization

    • Comptuationally expensive

    • Mostly supports a limited set of instructions.

• there exist feedback loops to further optimize it !

• To appear in ASLOS '26

• In IR, it may have less lines, but once translated to ASSEMBLY, it could be longer than necessary, thus tehre's stuff and help with that.

Undeifned behaviour

• tHIS PROGRAM CAN DO ANYTHING

  • If source program does implied behaviour

    • Either they are undefined behaviour or targetted error.

the most important part of this research is evaluation.

• We can provide more formal moments for LLM

• For now we only edxtact one basic seqeunce from one block/one label ( research more )

• Consider larger scopes ( cross basic

• Alive2 can timeout because of complexity.

  • I.e. if a lot of inputs SMT solver may brute force and thus time out.

• LLM

• Calling APIs in parallel

• In future, improve efficiency

• Proposing to implementations in CI

  • Still costly.

• Local Models....