meshletcodec: Simplify triangle decoding tail

When decoding triangles into 24-bit representation, the tail needs to
write 4 or 8 bytes. We were processing this as either a 8 or 4 byte
write; however, that results in a more complex branching structure than
writing 4 or 4+4 bytes does; and in practice, doing two writes (second -
conditionally) seems to be a little bit faster.

Author: zeux

src/meshletcodec.cpp

@@ -731,17 +731,15 @@ static const unsigned char* decodeTrianglesSimd(unsigned char* triangles, const
 #if defined(SIMD_SSE)
 		__m128i r = _mm_srli_si128(state, 9);
 
-		if ((triangle_count & 3) == 1)
-			*reinterpret_cast<unaligned_int*>(tail) = _mm_cvtsi128_si32(r);
-		else
-			_mm_storel_epi64(reinterpret_cast<__m128i*>(tail), r);
+		*reinterpret_cast<unaligned_int*>(tail) = _mm_cvtsi128_si32(r);
+		if ((triangle_count & 3) > 1)
+			*reinterpret_cast<unaligned_int*>(tail + 4) = _mm_extract_epi32(r, 1);
 #elif defined(SIMD_NEON)
 		uint8x16_t r = vextq_u8(state, vdupq_n_u8(0), 9);
 
-		if ((triangle_count & 3) == 1)
-			vst1q_lane_u32(reinterpret_cast<unsigned int*>(tail), vreinterpretq_u32_u8(r), 0);
-		else
-			vst1_u8(tail, vget_low_u8(r));
+		vst1q_lane_u32(reinterpret_cast<unsigned int*>(tail), vreinterpretq_u32_u8(r), 0);
+		if ((triangle_count & 3) > 1)
+			vst1q_lane_u32(reinterpret_cast<unsigned int*>(tail + 4), vreinterpretq_u32_u8(r), 1);
 #endif
 	}