bwahaha

decode.cpp

    		LZNA_DEC_INT val = 0;

		lzna_simd_vec xvec = vdupq_n_u16((U16) x);
        
		lzna_simd_vec cumfreq0 = vld1q_u16(&cumfreq[0]);
		lzna_simd_vec cumfreq1 = vld1q_u16(&cumfreq[8]);
		
        	// do compares
		lzna_simd_vec gt0 = vcgtq_u16(cumfreq0, xvec);
		lzna_simd_vec gt1 = vcgtq_u16(cumfreq1, xvec);

#if 0 && defined(__aarch64__)
		val = 15 + vaddvq_s16(U16toS16(gt0)) + vaddvq_s16(U16toS16(gt1));
#else
		val  = (cumfreq[ 4] > x) ? 0 : 1;
		val += (cumfreq[ 8] > x) ? 0 : 1;
		val += (cumfreq[12] > x) ? 0 : 1;
		val *= 4;

		const U16 *cffine = cumfreq + val;
		val += (cffine[ 1] > x) ? 0 : 1;
		val += (cffine[ 2] > x) ? 0 : 1;
		val += (cffine[ 3] > x) ? 0 : 1;
#endif
        
		assert(val >= 0 && val < nsyms);
        
    		U32 start = cumfreq[val];
    		U32 range = getcumfreq(val + 1) - start;