Once the Slang/MLIR frontend is self-contained and no longer consumes solc's IRs, the inline-assembly body is available to us as Yul-dialect MLIR — which lets us analyze it instead of blindly trusting the ("memory-safe") annotation the way solc does, turning a silent, test-invisible corruption class into compile-time diagnostics.
We should establish how much of the memory-safe contract is statically verifiable in our frontend: which rules reduce to tractable dataflow, where the undecidability boundary lies, what severity policy fits, and above all the false-positive profile on real projects, which is the make-or-break metric.
Once the Slang/MLIR frontend is self-contained and no longer consumes solc's IRs, the inline-assembly body is available to us as Yul-dialect MLIR — which lets us analyze it instead of blindly trusting the ("memory-safe") annotation the way solc does, turning a silent, test-invisible corruption class into compile-time diagnostics.
We should establish how much of the memory-safe contract is statically verifiable in our frontend: which rules reduce to tractable dataflow, where the undecidability boundary lies, what severity policy fits, and above all the false-positive profile on real projects, which is the make-or-break metric.