[ImportVerilog] Add variableAssignCallback, allow LHS capture

[Scheremo]

This patch extends ImportVerilog’s capture mechanism to include write-side
(LHS) captures, complementing the existing read-side collection. Tasks and
functions that assign to variables in an enclosing scope now have those refs
added as explicit function parameters during lowering.

• Context API

  • Introduce variableAssignCallback: std::function<void(mlir::Operation *)>.
  • Both read and write callbacks are chained and restored via scope guards.

• Capture collection

  • Reads: keep existing rvalueReadCallback behavior.
  • Writes: new callback fires on moore.blocking_assign,
    moore.nonblocking_assign, and moore.delayed_nonblocking_assign.
    The LHS ref (getDst()) is captured when it originates outside the
    function’s region.

• Lowering changes

  • In Expressions.cpp, creation sites for blocking / (delayed) non-blocking
    assigns invoke variableAssignCallback.
  • In Structure.cpp::convertFunction, install both callbacks to accumulate
    captures/captureIndex for the function, covering both reads and
    writes.
  • Existing finalizeFunctionBodyCaptures retypes the function to include the
    new capture parameters and replaces internal uses with the new block args.

• testLHSTaskCapture in basic.sv:

  • Confirms that a task assigning to a parent variable correctly captures it as
    an additional function argument.
  • Ensures call sites and function bodies are lowered with consistent capture
    semantics even when the function is defined after its use.

[fabianschuiki]

LGTM! I'm wondering if one approach to dealing with function captures in the future would be to lower the functions (without using the callbacks), and to then walk the generated function IR and find all SSA values that are used but defined outside the function. And then add those to the function as arguments.

[Scheremo]

LGTM! I'm wondering if one approach to dealing with function captures in the future would be to lower the functions (without using the callbacks), and to then walk the generated function IR and find all SSA values that are used but defined outside the function. And then add those to the function as arguments.

I don't see why not! 😄 What would the advantage of such an approach be?

[fabianschuiki]

I think the only benefit would be that it removes a source of bugs when we forget to call the callback in some function, or something to that end. Instead of collecting the information separately that then must match the IR (no SSA value uses from outside), that would use the IR itself to detect that.