Rule. When defining SIMD helper functions that accept vector parameters (such as __m128i, __m256i, or __m512i), pass vector arguments by pointer (e.g., const __m256i * for read-only inputs or __m256i * for mutated/in-out accumulators) rather than by value. Consistently pass all SIMD vector parameters as pointers rather than mixing value and pointer parameters, and pass addresses of array-resident vectors (&vec[i]) directly to pointer-based helper functions.
Why. Platform ABIs and calling conventions (such as 32-bit x86 MSVC on Windows and Windows x64) pass only a limited number of vector parameters in registers (typically 3 or 4). Passing large SIMD vectors by value or passing excess parameters spills arguments onto the stack without guaranteed 16- or 32-byte alignment, causing compiler errors (e.g., MSVC C2719), calling-convention mismatches, and register spills. Passing vector parameters by pointer avoids alignment hazards and reduces register pressure while inlining compilers optimize pointer indirection with zero runtime overhead.
Bad:
// Anti-pattern 1: Passing multiple __m256i vector parameters by value static INLINE __m256i fold_mul_and_sum_avx2(__m256i partiala, __m256i partialb, __m256i const1, __m256i const2) { const __m256i tmp = _mm256_madd_epi16(partiala, const1); return _mm256_add_epi32(tmp, _mm256_madd_epi16(partialb, const2)); } // Anti-pattern 2: Mixing value and pointer vector parameters static INLINE __m128i filter_block_2rows(const __m128i a0, const __m128i b0, const __m128i a1, const __m128i *b1, const __m128i *filter); // Anti-pattern 3: Passing array elements by value res = hadd_four_32_to_64_avx2(sum[0], sum[1], sum[2], sum[3]);
Good:
// Pattern 1: Pass vector inputs/accumulators as const or mutable pointers static INLINE __m256i fold_mul_and_sum_avx2(__m256i *partiala, __m256i *partialb, const __m256i *const1, const __m256i *const2) { const __m256i tmp = _mm256_madd_epi16(*partiala, *const1); return _mm256_add_epi32(tmp, _mm256_madd_epi16(*partialb, *const2)); } // Pattern 2: Consistently pass all vector parameters as pointers static INLINE __m128i filter_block_2rows(const __m128i *a0, const __m128i *b0, const __m128i *a1, const __m128i *b1, const __m128i *filter); // Pattern 3: Pass memory-resident array elements by pointer address res = hadd_four_32_to_64_avx2(&sum[0], &sum[1], &sum[2], &sum[3]);