feat: Make StepRecord use repr(C) layout

[Velaciela]

Motivation

A fixed-size, Copy, repr(C) StepRecord is a prerequisite for GPU-accelerated witness generation. With a deterministic layout, millions of step records can be bulk-copied (H2D) to the GPU without any serialization. The Copy trait also reduces per-step clone overhead on the CPU side by turning clones into cheap bitwise copies.

Summary

• Refactor StepRecord to implement Copy and use #[repr(C)] for deterministic memory layout, making it easy to bulk H2D copy to GPU for future CUDA witness generation
• Narrow RegIdx type from usize to u8 for compactness (RISC-V has only 32 registers)
• Extract SyscallWitness out of StepRecord into a separate store indexed by u32, since SyscallWitness contains Vec and cannot be Copy

---

• before

pub struct StepRecord {
    cycle: Cycle,
    pc: Change<ByteAddr>,
    pub heap_maxtouch_addr: Change<ByteAddr>,
    pub hint_maxtouch_addr: Change<ByteAddr>,
    pub insn: Instruction,

    rs1: Option<ReadOp>,
    rs2: Option<ReadOp>,
    rd: Option<WriteOp>,

    memory_op: Option<WriteOp>,
    syscall: Option<SyscallWitness>,
}

• after

#[repr(C)]
pub struct StepRecord {
    cycle: Cycle,
    pc: Change<ByteAddr>,
    pub heap_maxtouch_addr: Change<ByteAddr>,
    pub hint_maxtouch_addr: Change<ByteAddr>,
    pub insn: Instruction,

    has_rs1: bool,
    has_rs2: bool,
    has_rd: bool,
    has_memory_op: bool,

    rs1: ReadOp,
    rs2: ReadOp,
    rd: WriteOp,
    memory_op: WriteOp,

    /// Index into the separate syscall witness storage.
    /// `u32::MAX` means no syscall for this step.
    syscall_index: u32,
}

---

Changes

Core data structure changes (ceno_emul)

1. StepRecord is now Copy + repr(C) (136 bytes, 4-byte aligned):

   • Replaced Option<ReadOp> / Option<WriteOp> fields with inline values + has_rs1 / has_rs2 / has_rd / has_memory_op boolean flags
   • Replaced Option<SyscallWitness> with syscall_index: u32 (index into external store, u32::MAX = no syscall)
   • Added manual Default impl with sentinel values

2. RegIdx narrowed from usize to u8 (addr.rs):

   • All register index casts updated across disassemble, rv32im, vm_state, platform
   • Bounds checks use original u32 width before narrowing to avoid silent truncation

3. #[repr(C)] / #[repr(u8)] added to supporting types:

   • ByteAddr, WordAddr, Instruction, InsnKind, MemOp<T>, Change<T>
   • Ensures field ordering matches CUDA struct definitions

4. FullTracer gains syscall_witnesses: Vec<SyscallWitness>:

   • track_syscall() pushes to this vec and stores the index in StepRecord
   • New accessor syscall_witnesses() -> &[SyscallWitness]
   • reset_step_buffer() clears syscall_witnesses, keeping indices shard-local and avoiding cross-shard accumulation

5. Layout verification tests (tracer.rs):

   • test_step_record_is_copy_and_compact: asserts Copy trait and size <= 144 bytes
   • test_step_record_layout_for_gpu: asserts exact byte offsets of every field for CUDA header alignment
   • test_supporting_types_are_copy: asserts ReadOp, WriteOp, Change are Copy

API signature changes

• StepRecord::syscall() now takes &[SyscallWitness] parameter instead of returning from an internal Option
• StepRecord::has_syscall() -> bool added
• StepSource trait gains syscall_witnesses() -> &[SyscallWitness]
• ShardContext gains syscall_witnesses: Arc<Vec<SyscallWitness>> field
• keccak_step() test helper returns (StepRecord, Vec<Instruction>, Vec<SyscallWitness>)

Callers updated (ceno_zkvm, ceno_host)

• All ecall instruction assign_instances methods updated to read syscall witnesses from shard_ctx.syscall_witnesses
• OpFixedRS const generic changed from usize to u8
• ceno_host/tests/test_elf.rs: run() returns (Vec<StepRecord>, Vec<SyscallWitness>)
• e2e.rs: generate_witness propagates syscall witnesses into ShardContext

[hero78119]

a first quick pass with some perf related question

[hero78119]

this heavily calls within tracer better add #[inline(always)] to aligned with above step_record

[hero78119]

this to_vec() step is slightly expensive because of vector clone. Can we follow above line 1293 shard_steps to get only slice, then in the place where need it, just retrieved it by slice[index]?