MIR Codegen: Assembler Label & Jump Resolution

[gakonst]

Summary

Implement correct label, jump, and PUSH-width resolution so scheduled EVM instruction sequences become final bytecode.

Parent issue: #687

Context

Currently, jumps in the codegen are placeholders. We need a proper assembler that:

1. Assigns symbolic labels to basic blocks
2. Computes concrete byte offsets
3. Handles variable-width PUSH instructions (the target offset affects the PUSH size, which affects all subsequent offsets)

Tasks

Label abstraction

• Represent basic block labels as symbolic IDs
• JUMP/JUMPI instructions target labels, not numeric offsets
• Track label definitions and references

Two-pass assembly

• Pass 1: Linearize instructions, assign tentative offsets, track label positions
• Compute sizes of variable-width instructions (PUSHn)
• Pass 2: Re-run if PUSH widths changed due to offset size
• Iterate until fixed-point (typically 2-3 iterations max)

Jump resolution

• Replace each label reference with concrete offset bytes
• Emit correct PUSHn followed by JUMP/JUMPI
• Handle forward and backward jumps

JUMPDEST handling

• Ensure every jump target has preceding JUMPDEST
• Maintain mapping label → JUMPDEST offset
• Insert JUMPDEST opcodes during linearization

Example

; Symbolic assembly
  PUSH label_else
  JUMPI
  PUSH 42
  PUSH label_end
  JUMP
label_else:
  JUMPDEST
  PUSH 0
label_end:
  JUMPDEST
  RETURN

; After resolution (assuming offsets)
  PUSH1 0x0A    ; offset of label_else
  JUMPI
  PUSH1 0x2A    ; 42
  PUSH1 0x0C    ; offset of label_end  
  JUMP
  JUMPDEST      ; 0x0A: label_else
  PUSH1 0x00
  JUMPDEST      ; 0x0C: label_end
  RETURN

Patterns to follow

From Venom:

• Structured assembler with labels and separate patching step

From Sonatina:

• Clean separation between IR and assembly, symbol resolution passes

Acceptance Criteria

• Assembly tests: branches, loops, function dispatch have correct JUMP destinations
• Round-trip tests: assemble → disassemble → compare structure
• No invalid jumps or missing JUMPDESTs under EVM verification
• Variable-width PUSH correctly sized (PUSH1 vs PUSH2 vs PUSH3)

Estimated Complexity

Small-Medium - Well-understood algorithm but requires careful offset tracking

Dependencies

• MIR codegen skeleton (done)
• MIR Codegen: Stack Model & Basic Scheduling (≤16 values) #696 or MIR Codegen: Full Stack Scheduling with Spilling #697 (Stack scheduling) for realistic instruction streams

[0xpanicError]

can I take this one?