ob64 0.4.0 → 0.5.0

data/vendor/libbase64/lib/arch/generic/32/dec_loop.c

@@ -0,0 +1,86 @@
+static inline int
+dec_loop_generic_32_inner (const uint8_t **s, uint8_t **o, size_t *rounds)
+{
+	const uint32_t str
+		= base64_table_dec_32bit_d0[(*s)[0]]
+		| base64_table_dec_32bit_d1[(*s)[1]]
+		| base64_table_dec_32bit_d2[(*s)[2]]
+		| base64_table_dec_32bit_d3[(*s)[3]];
+
+#if BASE64_LITTLE_ENDIAN
+
+	// LUTs for little-endian set MSB in case of invalid character:
+	if (str & UINT32_C(0x80000000)) {
+		return 0;
+	}
+#else
+	// LUTs for big-endian set LSB in case of invalid character:
+	if (str & UINT32_C(1)) {
+		return 0;
+	}
+#endif
+	// Store the output:
+	memcpy(*o, &str, sizeof (str));
+
+	*s += 4;
+	*o += 3;
+	*rounds -= 1;
+
+	return 1;
+}
+
+static inline void
+dec_loop_generic_32 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	if (*slen < 8) {
+		return;
+	}
+
+	// Process blocks of 4 bytes per round. Because one extra zero byte is
+	// written after the output, ensure that there will be at least 4 bytes
+	// of input data left to cover the gap. (Two data bytes and up to two
+	// end-of-string markers.)
+	size_t rounds = (*slen - 4) / 4;
+
+	*slen -= rounds * 4;	// 4 bytes consumed per round
+	*olen += rounds * 3;	// 3 bytes produced per round
+
+	do {
+		if (rounds >= 8) {
+			if (dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds)) {
+				continue;
+			}
+			break;
+		}
+		if (rounds >= 4) {
+			if (dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds)) {
+				continue;
+			}
+			break;
+		}
+		if (rounds >= 2) {
+			if (dec_loop_generic_32_inner(s, o, &rounds) &&
+			    dec_loop_generic_32_inner(s, o, &rounds)) {
+				continue;
+			}
+			break;
+		}
+		dec_loop_generic_32_inner(s, o, &rounds);
+		break;
+
+	} while (rounds > 0);
+
+	// Adjust for any rounds that were skipped:
+	*slen += rounds * 4;
+	*olen -= rounds * 3;
+}

data/vendor/libbase64/lib/arch/generic/32/enc_loop.c

@@ -0,0 +1,73 @@
+static inline void
+enc_loop_generic_32_inner (const uint8_t **s, uint8_t **o)
+{
+	uint32_t src;
+
+	// Load input:
+	memcpy(&src, *s, sizeof (src));
+
+	// Reorder to 32-bit big-endian, if not already in that format. The
+	// workset must be in big-endian, otherwise the shifted bits do not
+	// carry over properly among adjacent bytes:
+	src = BASE64_HTOBE32(src);
+
+	// Two indices for the 12-bit lookup table:
+	const size_t index0 = (src >> 20) & 0xFFFU;
+	const size_t index1 = (src >>  8) & 0xFFFU;
+
+	// Table lookup and store:
+	memcpy(*o + 0, base64_table_enc_12bit + index0, 2);
+	memcpy(*o + 2, base64_table_enc_12bit + index1, 2);
+
+	*s += 3;
+	*o += 4;
+}
+
+static inline void
+enc_loop_generic_32 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	if (*slen < 4) {
+		return;
+	}
+
+	// Process blocks of 3 bytes at a time. Because blocks are loaded 4
+	// bytes at a time, ensure that there will be at least one remaining
+	// byte after the last round, so that the final read will not pass
+	// beyond the bounds of the input buffer:
+	size_t rounds = (*slen - 1) / 3;
+
+	*slen -= rounds * 3;	// 3 bytes consumed per round
+	*olen += rounds * 4;	// 4 bytes produced per round
+
+	do {
+		if (rounds >= 8) {
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			rounds -= 8;
+			continue;
+		}
+		if (rounds >= 4) {
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			rounds -= 4;
+			continue;
+		}
+		if (rounds >= 2) {
+			enc_loop_generic_32_inner(s, o);
+			enc_loop_generic_32_inner(s, o);
+			rounds -= 2;
+			continue;
+		}
+		enc_loop_generic_32_inner(s, o);
+		break;
+
+	} while (rounds > 0);
+}

data/vendor/libbase64/lib/arch/generic/64/enc_loop.c

@@ -0,0 +1,77 @@
+static inline void
+enc_loop_generic_64_inner (const uint8_t **s, uint8_t **o)
+{
+	uint64_t src;
+
+	// Load input:
+	memcpy(&src, *s, sizeof (src));
+
+	// Reorder to 64-bit big-endian, if not already in that format. The
+	// workset must be in big-endian, otherwise the shifted bits do not
+	// carry over properly among adjacent bytes:
+	src = BASE64_HTOBE64(src);
+
+	// Four indices for the 12-bit lookup table:
+	const size_t index0 = (src >> 52) & 0xFFFU;
+	const size_t index1 = (src >> 40) & 0xFFFU;
+	const size_t index2 = (src >> 28) & 0xFFFU;
+	const size_t index3 = (src >> 16) & 0xFFFU;
+
+	// Table lookup and store:
+	memcpy(*o + 0, base64_table_enc_12bit + index0, 2);
+	memcpy(*o + 2, base64_table_enc_12bit + index1, 2);
+	memcpy(*o + 4, base64_table_enc_12bit + index2, 2);
+	memcpy(*o + 6, base64_table_enc_12bit + index3, 2);
+
+	*s += 6;
+	*o += 8;
+}
+
+static inline void
+enc_loop_generic_64 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	if (*slen < 8) {
+		return;
+	}
+
+	// Process blocks of 6 bytes at a time. Because blocks are loaded 8
+	// bytes at a time, ensure that there will be at least 2 remaining
+	// bytes after the last round, so that the final read will not pass
+	// beyond the bounds of the input buffer:
+	size_t rounds = (*slen - 2) / 6;
+
+	*slen -= rounds * 6;	// 6 bytes consumed per round
+	*olen += rounds * 8;	// 8 bytes produced per round
+
+	do {
+		if (rounds >= 8) {
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			rounds -= 8;
+			continue;
+		}
+		if (rounds >= 4) {
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			rounds -= 4;
+			continue;
+		}
+		if (rounds >= 2) {
+			enc_loop_generic_64_inner(s, o);
+			enc_loop_generic_64_inner(s, o);
+			rounds -= 2;
+			continue;
+		}
+		enc_loop_generic_64_inner(s, o);
+		break;
+
+	} while (rounds > 0);
+}

data/vendor/libbase64/lib/arch/generic/codec.c

@@ -0,0 +1,39 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <string.h>
+
+#include "../../../include/libbase64.h"
+#include "../../tables/tables.h"
+#include "../../codecs.h"
+#include "config.h"
+#include "../../env.h"
+
+#if BASE64_WORDSIZE == 32
+#  include "32/enc_loop.c"
+#elif BASE64_WORDSIZE == 64
+#  include "64/enc_loop.c"
+#endif
+
+#if BASE64_WORDSIZE >= 32
+#  include "32/dec_loop.c"
+#endif
+
+BASE64_ENC_FUNCTION(plain)
+{
+	#include "enc_head.c"
+#if BASE64_WORDSIZE == 32
+	enc_loop_generic_32(&s, &slen, &o, &olen);
+#elif BASE64_WORDSIZE == 64
+	enc_loop_generic_64(&s, &slen, &o, &olen);
+#endif
+	#include "enc_tail.c"
+}
+
+BASE64_DEC_FUNCTION(plain)
+{
+	#include "dec_head.c"
+#if BASE64_WORDSIZE >= 32
+	dec_loop_generic_32(&s, &slen, &o, &olen);
+#endif
+	#include "dec_tail.c"
+}

data/vendor/libbase64/lib/arch/generic/dec_head.c

@@ -0,0 +1,37 @@
+int ret = 0;
+const uint8_t *s = (const uint8_t *) src;
+uint8_t *o = (uint8_t *) out;
+uint8_t q;
+
+// Use local temporaries to avoid cache thrashing:
+size_t olen = 0;
+size_t slen = srclen;
+struct base64_state st;
+st.eof = state->eof;
+st.bytes = state->bytes;
+st.carry = state->carry;
+
+// If we previously saw an EOF or an invalid character, bail out:
+if (st.eof) {
+	*outlen = 0;
+	ret = 0;
+	// If there was a trailing '=' to check, check it:
+	if (slen && (st.eof == BASE64_AEOF)) {
+		state->bytes = 0;
+		state->eof = BASE64_EOF;
+		ret = ((base64_table_dec_8bit[*s++] == 254) && (slen == 1)) ? 1 : 0;
+	}
+	return ret;
+}
+
+// Turn four 6-bit numbers into three bytes:
+// out[0] = 11111122
+// out[1] = 22223333
+// out[2] = 33444444
+
+// Duff's device again:
+switch (st.bytes)
+{
+	for (;;)
+	{
+	case 0:

data/vendor/libbase64/lib/arch/generic/dec_tail.c

@@ -0,0 +1,91 @@
+		if (slen-- == 0) {
+			ret = 1;
+			break;
+		}
+		if ((q = base64_table_dec_8bit[*s++]) >= 254) {
+			st.eof = BASE64_EOF;
+			// Treat character '=' as invalid for byte 0:
+			break;
+		}
+		st.carry = q << 2;
+		st.bytes++;
+
+		// Deliberate fallthrough:
+		BASE64_FALLTHROUGH
+
+	case 1:	if (slen-- == 0) {
+			ret = 1;
+			break;
+		}
+		if ((q = base64_table_dec_8bit[*s++]) >= 254) {
+			st.eof = BASE64_EOF;
+			// Treat character '=' as invalid for byte 1:
+			break;
+		}
+		*o++ = st.carry | (q >> 4);
+		st.carry = q << 4;
+		st.bytes++;
+		olen++;
+
+		// Deliberate fallthrough:
+		BASE64_FALLTHROUGH
+
+	case 2:	if (slen-- == 0) {
+			ret = 1;
+			break;
+		}
+		if ((q = base64_table_dec_8bit[*s++]) >= 254) {
+			st.bytes++;
+			// When q == 254, the input char is '='.
+			// Check if next byte is also '=':
+			if (q == 254) {
+				if (slen-- != 0) {
+					st.bytes = 0;
+					// EOF:
+					st.eof = BASE64_EOF;
+					q = base64_table_dec_8bit[*s++];
+					ret = ((q == 254) && (slen == 0)) ? 1 : 0;
+					break;
+				}
+				else {
+					// Almost EOF
+					st.eof = BASE64_AEOF;
+					ret = 1;
+					break;
+				}
+			}
+			// If we get here, there was an error:
+			break;
+		}
+		*o++ = st.carry | (q >> 2);
+		st.carry = q << 6;
+		st.bytes++;
+		olen++;
+
+		// Deliberate fallthrough:
+		BASE64_FALLTHROUGH
+
+	case 3:	if (slen-- == 0) {
+			ret = 1;
+			break;
+		}
+		if ((q = base64_table_dec_8bit[*s++]) >= 254) {
+			st.bytes = 0;
+			st.eof = BASE64_EOF;
+			// When q == 254, the input char is '='. Return 1 and EOF.
+			// When q == 255, the input char is invalid. Return 0 and EOF.
+			ret = ((q == 254) && (slen == 0)) ? 1 : 0;
+			break;
+		}
+		*o++ = st.carry | q;
+		st.carry = 0;
+		st.bytes = 0;
+		olen++;
+	}
+}
+
+state->eof = st.eof;
+state->bytes = st.bytes;
+state->carry = st.carry;
+*outlen = olen;
+return ret;

data/vendor/libbase64/lib/arch/generic/enc_head.c

@@ -0,0 +1,24 @@
+// Assume that *out is large enough to contain the output.
+// Theoretically it should be 4/3 the length of src.
+const uint8_t *s = (const uint8_t *) src;
+uint8_t *o = (uint8_t *) out;
+
+// Use local temporaries to avoid cache thrashing:
+size_t olen = 0;
+size_t slen = srclen;
+struct base64_state st;
+st.bytes = state->bytes;
+st.carry = state->carry;
+
+// Turn three bytes into four 6-bit numbers:
+// in[0] = 00111111
+// in[1] = 00112222
+// in[2] = 00222233
+// in[3] = 00333333
+
+// Duff's device, a for() loop inside a switch() statement. Legal!
+switch (st.bytes)
+{
+	for (;;)
+	{
+	case 0:

data/vendor/libbase64/lib/arch/generic/enc_tail.c

@@ -0,0 +1,34 @@
+		if (slen-- == 0) {
+			break;
+		}
+		*o++ = base64_table_enc_6bit[*s >> 2];
+		st.carry = (*s++ << 4) & 0x30;
+		st.bytes++;
+		olen += 1;
+
+		// Deliberate fallthrough:
+		BASE64_FALLTHROUGH
+
+	case 1:	if (slen-- == 0) {
+			break;
+		}
+		*o++ = base64_table_enc_6bit[st.carry | (*s >> 4)];
+		st.carry = (*s++ << 2) & 0x3C;
+		st.bytes++;
+		olen += 1;
+
+		// Deliberate fallthrough:
+		BASE64_FALLTHROUGH
+
+	case 2:	if (slen-- == 0) {
+			break;
+		}
+		*o++ = base64_table_enc_6bit[st.carry | (*s >> 6)];
+		*o++ = base64_table_enc_6bit[*s++ & 0x3F];
+		st.bytes = 0;
+		olen += 2;
+	}
+}
+state->bytes = st.bytes;
+state->carry = st.carry;
+*outlen = olen;

data/vendor/libbase64/lib/arch/neon32/codec.c

@@ -0,0 +1,72 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <string.h>
+
+#include "../../../include/libbase64.h"
+#include "../../tables/tables.h"
+#include "../../codecs.h"
+#include "config.h"
+#include "../../env.h"
+
+#ifdef __arm__
+#  if (defined(__ARM_NEON__) || defined(__ARM_NEON)) && HAVE_NEON32
+#    define BASE64_USE_NEON32
+#  endif
+#endif
+
+#ifdef BASE64_USE_NEON32
+#include <arm_neon.h>
+
+static inline uint8x16_t
+vqtbl1q_u8 (const uint8x16_t lut, const uint8x16_t indices)
+{
+	// NEON32 only supports 64-bit wide lookups in 128-bit tables. Emulate
+	// the NEON64 `vqtbl1q_u8` intrinsic to do 128-bit wide lookups.
+	uint8x8x2_t lut2;
+	uint8x8x2_t result;
+
+	lut2.val[0] = vget_low_u8(lut);
+	lut2.val[1] = vget_high_u8(lut);
+
+	result.val[0] = vtbl2_u8(lut2, vget_low_u8(indices));
+	result.val[1] = vtbl2_u8(lut2, vget_high_u8(indices));
+
+	return vcombine_u8(result.val[0], result.val[1]);
+}
+
+#include "../generic/32/dec_loop.c"
+#include "../generic/32/enc_loop.c"
+#include "dec_loop.c"
+#include "enc_reshuffle.c"
+#include "enc_translate.c"
+#include "enc_loop.c"
+
+#endif	// BASE64_USE_NEON32
+
+// Stride size is so large on these NEON 32-bit functions
+// (48 bytes encode, 32 bytes decode) that we inline the
+// uint32 codec to stay performant on smaller inputs.
+
+BASE64_ENC_FUNCTION(neon32)
+{
+#ifdef BASE64_USE_NEON32
+	#include "../generic/enc_head.c"
+	enc_loop_neon32(&s, &slen, &o, &olen);
+	enc_loop_generic_32(&s, &slen, &o, &olen);
+	#include "../generic/enc_tail.c"
+#else
+	BASE64_ENC_STUB
+#endif
+}
+
+BASE64_DEC_FUNCTION(neon32)
+{
+#ifdef BASE64_USE_NEON32
+	#include "../generic/dec_head.c"
+	dec_loop_neon32(&s, &slen, &o, &olen);
+	dec_loop_generic_32(&s, &slen, &o, &olen);
+	#include "../generic/dec_tail.c"
+#else
+	BASE64_DEC_STUB
+#endif
+}

data/vendor/libbase64/lib/arch/neon32/dec_loop.c

@@ -0,0 +1,106 @@
+static inline int
+is_nonzero (const uint8x16_t v)
+{
+	uint64_t u64;
+	const uint64x2_t v64 = vreinterpretq_u64_u8(v);
+	const uint32x2_t v32 = vqmovn_u64(v64);
+
+	vst1_u64(&u64, vreinterpret_u64_u32(v32));
+	return u64 != 0;
+}
+
+static inline uint8x16_t
+delta_lookup (const uint8x16_t v)
+{
+	const uint8x8_t lut = {
+		0, 16, 19, 4, (uint8_t) -65, (uint8_t) -65, (uint8_t) -71, (uint8_t) -71,
+	};
+
+	return vcombine_u8(
+		vtbl1_u8(lut, vget_low_u8(v)),
+		vtbl1_u8(lut, vget_high_u8(v)));
+}
+
+static inline uint8x16_t
+dec_loop_neon32_lane (uint8x16_t *lane)
+{
+	// See the SSSE3 decoder for an explanation of the algorithm.
+	const uint8x16_t lut_lo = {
+		0x15, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11, 0x11,
+		0x11, 0x11, 0x13, 0x1A, 0x1B, 0x1B, 0x1B, 0x1A
+	};
+
+	const uint8x16_t lut_hi = {
+		0x10, 0x10, 0x01, 0x02, 0x04, 0x08, 0x04, 0x08,
+		0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10
+	};
+
+	const uint8x16_t mask_0F = vdupq_n_u8(0x0F);
+	const uint8x16_t mask_2F = vdupq_n_u8(0x2F);
+
+	const uint8x16_t hi_nibbles = vshrq_n_u8(*lane, 4);
+	const uint8x16_t lo_nibbles = vandq_u8(*lane, mask_0F);
+	const uint8x16_t eq_2F      = vceqq_u8(*lane, mask_2F);
+
+	const uint8x16_t hi = vqtbl1q_u8(lut_hi, hi_nibbles);
+	const uint8x16_t lo = vqtbl1q_u8(lut_lo, lo_nibbles);
+
+	// Now simply add the delta values to the input:
+	*lane = vaddq_u8(*lane, delta_lookup(vaddq_u8(eq_2F, hi_nibbles)));
+
+	// Return the validity mask:
+	return vandq_u8(lo, hi);
+}
+
+static inline void
+dec_loop_neon32 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	if (*slen < 64) {
+		return;
+	}
+
+	// Process blocks of 64 bytes per round. Unlike the SSE codecs, no
+	// extra trailing zero bytes are written, so it is not necessary to
+	// reserve extra input bytes:
+	size_t rounds = *slen / 64;
+
+	*slen -= rounds * 64;	// 64 bytes consumed per round
+	*olen += rounds * 48;	// 48 bytes produced per round
+
+	do {
+		uint8x16x3_t dec;
+
+		// Load 64 bytes and deinterleave:
+		uint8x16x4_t str = vld4q_u8(*s);
+
+		// Decode each lane, collect a mask of invalid inputs:
+		const uint8x16_t classified
+			= dec_loop_neon32_lane(&str.val[0])
+			| dec_loop_neon32_lane(&str.val[1])
+			| dec_loop_neon32_lane(&str.val[2])
+			| dec_loop_neon32_lane(&str.val[3]);
+
+		// Check for invalid input: if any of the delta values are
+		// zero, fall back on bytewise code to do error checking and
+		// reporting:
+		if (is_nonzero(classified)) {
+			break;
+		}
+
+		// Compress four bytes into three:
+		dec.val[0] = vorrq_u8(vshlq_n_u8(str.val[0], 2), vshrq_n_u8(str.val[1], 4));
+		dec.val[1] = vorrq_u8(vshlq_n_u8(str.val[1], 4), vshrq_n_u8(str.val[2], 2));
+		dec.val[2] = vorrq_u8(vshlq_n_u8(str.val[2], 6), str.val[3]);
+
+		// Interleave and store decoded result:
+		vst3q_u8(*o, dec);
+
+		*s += 64;
+		*o += 48;
+
+	} while (--rounds > 0);
+
+	// Adjust for any rounds that were skipped:
+	*slen += rounds * 64;
+	*olen -= rounds * 48;
+}

data/vendor/libbase64/lib/arch/neon32/enc_loop.c

@@ -0,0 +1,58 @@
+static inline void
+enc_loop_neon32_inner (const uint8_t **s, uint8_t **o)
+{
+	// Load 48 bytes and deinterleave:
+	uint8x16x3_t src = vld3q_u8(*s);
+
+	// Reshuffle:
+	uint8x16x4_t out = enc_reshuffle(src);
+
+	// Translate reshuffled bytes to the Base64 alphabet:
+	out = enc_translate(out);
+
+	// Interleave and store output:
+	vst4q_u8(*o, out);
+
+	*s += 48;
+	*o += 64;
+}
+
+static inline void
+enc_loop_neon32 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
+{
+	size_t rounds = *slen / 48;
+
+	*slen -= rounds * 48;	// 48 bytes consumed per round
+	*olen += rounds * 64;	// 64 bytes produced per round
+
+	while (rounds > 0) {
+		if (rounds >= 8) {
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			rounds -= 8;
+			continue;
+		}
+		if (rounds >= 4) {
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			rounds -= 4;
+			continue;
+		}
+		if (rounds >= 2) {
+			enc_loop_neon32_inner(s, o);
+			enc_loop_neon32_inner(s, o);
+			rounds -= 2;
+			continue;
+		}
+		enc_loop_neon32_inner(s, o);
+		break;
+	}
+}