Arm64 Baseline Porting

ARM64 Baseline Porting Constraints

This skill provides the minimum correctness guarantees that every x64→ARM64 migration must satisfy, regardless of whether a specific migration spec matches.

When invoked via /dispatcher-skill <id> --mode llm-freeform, apply ALL rules below before and during the porting attempt.

Sections 1–9 apply to hand-written ARM64 assembly output. Sections 10–11 apply to all freeform output (asm or C/intrinsics). Section 12 applies specifically to C/intrinsics output. Section 13 is the verification gate — split by output language.

1. Calling Convention (Windows ARM64 ABI)

2. Stack Alignment

• SP MUST be 16-byte aligned at all times
• Stack allocation MUST use SUB SP, SP, #N where N is a multiple of 16
• Paired save/restore: use STP/LDP with pre-index or signed offset
• No red zone on Windows ARM64 — do not access below SP

3. Flag Discipline

• ARM64 arithmetic instructions (ADD, SUB) do NOT set flags by default
• MUST use flag-setting variants (ADDS, SUBS, CMP, TST) before any conditional branch
• Every B.cond MUST have an explicit flag-setting instruction preceding it
• No implicit flag setting from MOV, LDR, STR, or logical shifts

4. Immediate Encoding

• ADD/SUB: 12-bit immediate (0–4095), optionally shifted left by 12
• MOV: 16-bit immediate per MOVZ/MOVK; large constants need MOVZ+MOVK sequence
• Bitwise (AND/ORR/EOR): bitmask immediate (not arbitrary 64-bit values)
• If x64 code uses a large immediate, split via MOVZ/MOVK or load from literal pool

5. Memory Access

• Unaligned access: generally supported on ARM64 but slower; prefer aligned access
• No x86-style scaled-index [base + index*scale + disp] — use shifted register offset: LDR X0, [X1, X2, LSL #3]
• Maximum LDR/STR immediate offset: unsigned 12-bit scaled by access size
• For large offsets: compute address in a temp register first

6. SIMD / NEON Constraints

• Maximum native vector width: 128-bit (Q registers / V registers)
• No 256-bit or 512-bit native registers; split AVX/AVX-512 into multiple 128-bit ops
• Horizontal reduction across a full vector: prefer ADDV / vaddvq_* (full reduction). FADDP is pairwise only — use it only when you actually want pairwise behavior (e.g. iterative tree reduction)
• Movemask analogue: there is no direct _mm_movemask_epi8. Common replacements: vaddvq_u8 on a comparison-mask vector with bit-weights, or vshrn_n_u16 to compress mask lanes — choose per the matched intrinsic spec
• Use LD1/ST1 for potentially unaligned vector loads/stores
• Floating-point: IEEE 754 compliant; denormals may be flushed to zero depending on FPCR

7. Branch and Control Flow

8. Prolog / Epilog Pattern

Prolog template:

STP X29, X30, [SP, #-frame_size]!   ; save FP + LR, allocate frame
MOV X29, SP                          ; establish frame pointer
STP X19, X20, [SP, #16]             ; save callee-saved registers
STP X21, X22, [SP, #32]             ; (as many pairs as needed)

Epilog template:

LDP X21, X22, [SP, #32]             ; restore callee-saved registers
LDP X19, X20, [SP, #16]
LDP X29, X30, [SP], #frame_size     ; restore FP + LR, deallocate frame
RET

9. Function Returns and Indirect Calls

• Return via RET (uses X30) — never by jumping to an address loaded from the stack
• Indirect call: BLR Xn (saves return in X30); indirect tail-call: BR Xn
• Pointer authentication (PAC) may be enabled — if so, use PACIASP/AUTIASP or RETAA/RETAB per project convention

9a. Register-Aliasing Hazards (the most common silent correctness/crash bug)

Wn and Xn are the same physical register; a 32-bit Wn write zero-extends and destroys bits 63:32 of Xn. This single class of bug recurs constantly in ported NEON kernels, as a wrong result or (more often) a crash far from its cause:

• In-function: never load a narrow value into Wn (LDRB Wn, ADD Wn,...) while Xn still holds a live 64-bit pointer. Build addresses in a different register. (e.g. LDRB W9,[X2] then ADD X9,X2,#1 loses the byte just loaded into W9 — use X10 for the pointer.)
• Across a BL: a helper must NOT return its result in, or scratch, a register whose Xn-form the caller keeps live across the call (a pointer it advances between calls). Return results in a dedicated clobber GPR (e.g. X9–X15), never X0 when X0 is a reused argument pointer. The fault PC lands at the caller's next load/store, not at the helper.
• A doc-comment like "returns in W0, preserves X0" is self-contradictory — W0 IS X0. Trust the register algebra.
• SIMD analogue: a .2s/.4h/.8b/D-form write zeroes the upper 64 bits of the V register — don't reuse a .4s accumulator for a D-form tail.
• Small inputs / single-iteration paths mask it (the value is consumed before the clobbering reuse); exercise the looped/larger-size path.

9b. Stale Flags Before CSEL/CSET/B.cond

NZCV holds the result of the most recent flag-setting instruction, not the one your condition name implies. A CSEL ...,NE placed after an intervening CMP/CSET/branch selects on the prior comparison and compiles cleanly. Re-establish flags with a fresh CMP/TST immediately before each conditional select/set/branch whose condition differs from the last one set.

9c. Partial Loads Leave Stale Lanes

A narrow load (LD1 {Vn.S}[0], LDR Sn/Bn) writes only the addressed lanes; the rest of the register keeps stale contents. A later full-width op (.8B/.16B/.8H) folds that garbage into the result. Pre-zero with MOVI Vn.8B,#0 before the partial load, or use a zero-extending load. A comment claiming "the upper lanes are 0" is not a guarantee.

10. Memory Ordering (applies to ALL output: asm AND C/intrinsics)

ARM64 has a weakly-ordered memory model; x86 has TSO (total store order). Code that synchronized correctly on x86 by accident often breaks silently on ARM64.

• Plain loads/stores can be reordered across each other and across atomics
• volatile does NOT imply ordering on ARM64 (MSVC's /volatile:ms semantics are x86-only and non-portable)
• For inter-thread synchronization, use acquire/release atomics or explicit barriers:
  • Acquire load: LDAR (asm) / __atomic_load_n(..., __ATOMIC_ACQUIRE) or std::atomic::load(memory_order_acquire) (C/C++)
  • Release store: STLR (asm) / __atomic_store_n(..., __ATOMIC_RELEASE) (C/C++)
  • Full fence: DMB ISH (asm) / __atomic_thread_fence(__ATOMIC_SEQ_CST) (C/C++)
• _Interlocked* MSVC intrinsics: on ARM64, the un-suffixed forms imply sequentially consistent ordering; if porting performance-sensitive code consider _*_acq / _*_rel variants or __atomic_* builtins
• LSE atomics (LDADD, SWPAL, CAS*) are faster than LDXR/STXR retry loops but require ARMv8.1-A. Detect with IsProcessorFeaturePresent(PF_ARM_V81_ATOMIC_INSTRUCTIONS_AVAILABLE) at runtime, or compile with /arch:armv8.1 / -march=armv8.1-a if the deployment target guarantees it

11. ARM64EC Awareness

ARM64EC (Emulation Compatible) is a hybrid ABI on Windows that allows ARM64 code to interop with x64 code in the same process. Freeform output for an ARM64EC target is NOT the same as classic ARM64.

• Detect target: _M_ARM64EC is defined on ARM64EC; _M_ARM64 is defined on both classic ARM64 and ARM64EC. Check _M_ARM64EC first
• ARM64EC functions called from x64 must use the EC ABI shim — do NOT emit raw ARM64 calling-convention asm in an ARM64EC translation unit unless you understand the entry/exit thunks
• Header guards: #if defined(_M_ARM64) || defined(_M_ARM64EC) is the right pattern for "ARM64-family"; use #if defined(_M_ARM64) && !defined(_M_ARM64EC) to mean "classic ARM64 only"
• If you don't know whether the target is classic ARM64 or ARM64EC, say so explicitly in a code comment and prefer C/intrinsics over hand-written asm — the compiler emits correct EC thunks; hand-written asm doesn't

12. C / Intrinsics Freeform Output Rules

When the freeform output is C/C++ (not assembly):

• Header: #include <arm_neon.h> for NEON; #include <arm_acle.h> for CRC32, AES, SHA, atomics, etc.
• Target macro guard: wrap NEON code in #if defined(_M_ARM64) || defined(_M_ARM64EC) || defined(__aarch64__) (or use the project's existing macro convention if visible in the porting context)
• NEON feature macros: __ARM_NEON (always 1 on ARM64), __ARM_FEATURE_CRC32, __ARM_FEATURE_CRYPTO — gate optional features on these
• Bit-scan / byte-swap mappings:
  • _BitScanForward / _tzcnt_u32 → __builtin_ctz (clang/gcc) or _CountTrailingZeros (MSVC)
  • _BitScanReverse / _lzcnt_u32 → __builtin_clz / _CountLeadingZeros
  • _byteswap_ulong → __builtin_bswap32 or vector vrev32q_u8
  • __popcnt → __builtin_popcount (scalar) or vcntq_u8 + reduce (vector)
• Do NOT leave any _mm_* / _mm256_* / __m128i / __m256i token in the output. If you cannot port a specific intrinsic, comment-out the call and emit a TODO(arm64): note rather than emitting wrong code
• Do NOT use inline __asm blocks unless the matched spec explicitly requires it — MSVC ARM64 does not support inline assembly anyway; use intrinsics or a separate .s file
• Type fidelity: pick the NEON type whose lane width and signedness match the x86 op. _mm_add_epi32 → vaddq_s32 on int32x4_t, NOT vaddq_u32. Wrong signedness compiles fine but breaks comparisons and saturation
• Short-buffer / tail guard (crash risk). A SIMD kernel whose main loop consumes a fixed stride (16/32/64 bytes) MUST guard the entry: if (len < STRIDE) return <scalar-fallback>(...) BEFORE any unconditional vld1q_u8(ptr) + len -= STRIDE. x86 source usually has this short-input branch; porting only the hot loop drops it. Two failures compound: the unconditional vector load reads out of bounds, AND len -= STRIDE on an unsigned size_t underflows to a near-SIZE_MAX value, turning the next while (len >= STRIDE) into an unbounded OOB walk → segfault. Watch especially for a separate "fold the initial state in" branch (e.g. if (crc != 0) { load 16; len -= 16; }) that also loads unconditionally. Test every kernel at len ∈ {0,1,STRIDE-1,STRIDE,STRIDE+1} with a non-default seed — large aligned buffers never exercise this path.
• MSVC ARM64 intrinsic portability (compile failures). GCC/Clang accept things MSVC's ARM64 headers reject; verify on MSVC, not just clang:
  • Aggregate brace-init of NEON vectors fails on MSVC (uint32x4_t v = {a,b,c,d} → C2078). Build constant vectors from a plain array: vld1q_u32((const uint32_t[]){...}) or a named tmp[] + vld1q_u32(tmp).
  • poly64_t is a 64-bit integer on GCC/Clang but the opaque __n64 on MSVC, so (poly64_t)some_u64 casts only compile on the former. For vmull_p64, the operand form also diverges: ACLE takes a scalar lane (vgetq_lane_p64(vreinterpretq_p64_u8(x), n)), MSVC takes a vector lane (vget_low_p64 / vget_high_p64). Split such macros on #if defined(_MSC_VER).
  • poly128_t is unnamed on MSVC — only use the 128-bit poly result through vreinterpretq_u8_p128(), never as a named type. <arm_acle.h> does not exist on MSVC (NEON/crypto/SHA3 intrinsics come from <arm64_neon.h> via <arm_neon.h>); guard the acle include with #if !defined(_MSC_VER).
• #include "*.c" template reuse defeats incremental builds. This skilltree's kernels often #include "<base>.c" to share static inline helpers across variant files. Build systems (CMake/MSBuild) do NOT track #include of a .c as a dependency, so editing the included base file does NOT recompile the including TUs — they keep an inlined STALE copy of the helper (e.g. a fix in the base appears to "not take"). After editing a file that is #included by siblings, force-rebuild those siblings (touch them, or clean-build) before trusting a test result. Suppress duplicate external symbols from the included .c with a #ifndef <BASE>_HELPERS_ONLY guard around its non-inline entry points, defined by each includer.

13. Verification Checklist

For ALL output

• Memory ordering preserved: any _Interlocked* / atomic / lock primitive maps to acquire/release-correct ARM64 form
• Target macro guard distinguishes ARM64EC from classic ARM64 if the code path differs
• No x86-only intrinsic / mnemonic / register name leaked into output

For ARM64 assembly output

• SP remains 16-byte aligned after all stack operations
• All callee-saved registers used are saved in prolog and restored in epilog
• X18 is never written or used as scratch
• Every conditional branch has an explicit flag-setting source
• No CSEL/CSET/B.cond relies on stale NZCV — flags are freshly set for its condition (no CMP carried across intervening flag-setting work)
• No Wn write clobbers a live 64-bit pointer/value in Xn; no BL-reached helper returns in / scratches a register whose Xn the caller keeps live across the call
• Partial/narrow loads (LD1 {Vn.S}[0], LDR Sn) zero the unloaded lanes before any full-width op reads them; D-form writes don't silently clear an accumulator's upper lanes
• B.cond ranges fit ±1 MB; long-range conditionals use a trampoline
• Large immediates are properly split (MOVZ+MOVK)
• No memory access below SP (no red zone)
• Vector operations stay within 128-bit NEON width
• Function returns via RET (using X30), not by jumping to address on stack

For C / intrinsics output

• <arm_neon.h> (and <arm_acle.h> if needed) included
• Code is wrapped in an ARM64 target-macro guard
• No _mm_* / _mm256_* / __m128i / __m256i tokens remain
• NEON lane width and signedness match the original x86 intrinsic semantics
• Vector operations stay within 128-bit (no implicit AVX-style 256-bit assumptions)
• Bit-scan / byte-swap / popcount use ARM64-portable builtins, not MSVC x86-only ones
• Any unsupported intrinsic is flagged with a TODO(arm64): comment, not silently mistranslated
• Fixed-stride SIMD kernels guard len < STRIDE (and any unconditional initial-state load) with a scalar fallback; tested at len ∈ {0,1,STRIDE±1} with a non-default seed (prevents OOB + size_t underflow → segfault)
• Compiles on MSVC ARM64, not only clang/gcc: no NEON-vector brace-init, no (poly64_t)/poly128_t named-type use, vmull_p64 lane-extraction split on _MSC_VER, <arm_acle.h> guarded with #if !defined(_MSC_VER)
• If kernels #include "*.c" to share helpers: includers force-rebuilt after editing the base, and base's external entry points guarded by a *_HELPERS_ONLY macro to avoid duplicate symbols