deflate-ruby 0.1.0

data/ext/deflate_ruby/libdeflate/lib/ht_matchfinder.h

@@ -0,0 +1,234 @@
+/*
+ * ht_matchfinder.h - Lempel-Ziv matchfinding with a hash table
+ *
+ * Copyright 2022 Eric Biggers
+ *
+ * Permission is hereby granted, free of charge, to any person
+ * obtaining a copy of this software and associated documentation
+ * files (the "Software"), to deal in the Software without
+ * restriction, including without limitation the rights to use,
+ * copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following
+ * conditions:
+ *
+ * The above copyright notice and this permission notice shall be
+ * included in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
+ * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
+ * HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+ * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * ---------------------------------------------------------------------------
+ *
+ * This is a Hash Table (ht) matchfinder.
+ *
+ * This is a variant of the Hash Chains (hc) matchfinder that is optimized for
+ * very fast compression.  The ht_matchfinder stores the hash chains inline in
+ * the hash table, whereas the hc_matchfinder stores them in a separate array.
+ * Storing the hash chains inline is the faster method when max_search_depth
+ * (the maximum chain length) is very small.  It is not appropriate when
+ * max_search_depth is larger, as then it uses too much memory.
+ *
+ * Due to its focus on speed, the ht_matchfinder doesn't support length 3
+ * matches.  It also doesn't allow max_search_depth to vary at runtime; it is
+ * fixed at build time as HT_MATCHFINDER_BUCKET_SIZE.
+ *
+ * See hc_matchfinder.h for more information.
+ */
+
+#ifndef LIB_HT_MATCHFINDER_H
+#define LIB_HT_MATCHFINDER_H
+
+#include "matchfinder_common.h"
+
+#define HT_MATCHFINDER_HASH_ORDER	15
+#define HT_MATCHFINDER_BUCKET_SIZE	2
+
+#define HT_MATCHFINDER_MIN_MATCH_LEN	4
+/* Minimum value of max_len for ht_matchfinder_longest_match() */
+#define HT_MATCHFINDER_REQUIRED_NBYTES	5
+
+struct MATCHFINDER_ALIGNED ht_matchfinder {
+	mf_pos_t hash_tab[1UL << HT_MATCHFINDER_HASH_ORDER]
+			 [HT_MATCHFINDER_BUCKET_SIZE];
+};
+
+static forceinline void
+ht_matchfinder_init(struct ht_matchfinder *mf)
+{
+	STATIC_ASSERT(sizeof(*mf) % MATCHFINDER_SIZE_ALIGNMENT == 0);
+
+	matchfinder_init((mf_pos_t *)mf, sizeof(*mf));
+}
+
+static forceinline void
+ht_matchfinder_slide_window(struct ht_matchfinder *mf)
+{
+	matchfinder_rebase((mf_pos_t *)mf, sizeof(*mf));
+}
+
+/* Note: max_len must be >= HT_MATCHFINDER_REQUIRED_NBYTES */
+static forceinline u32
+ht_matchfinder_longest_match(struct ht_matchfinder * const mf,
+			     const u8 ** const in_base_p,
+			     const u8 * const in_next,
+			     const u32 max_len,
+			     const u32 nice_len,
+			     u32 * const next_hash,
+			     u32 * const offset_ret)
+{
+	u32 best_len = 0;
+	const u8 *best_matchptr = in_next;
+	u32 cur_pos = in_next - *in_base_p;
+	const u8 *in_base;
+	mf_pos_t cutoff;
+	u32 hash;
+	u32 seq;
+	mf_pos_t cur_node;
+	const u8 *matchptr;
+#if HT_MATCHFINDER_BUCKET_SIZE > 1
+	mf_pos_t to_insert;
+	u32 len;
+#endif
+#if HT_MATCHFINDER_BUCKET_SIZE > 2
+	int i;
+#endif
+
+	/* This is assumed throughout this function. */
+	STATIC_ASSERT(HT_MATCHFINDER_MIN_MATCH_LEN == 4);
+
+	if (cur_pos == MATCHFINDER_WINDOW_SIZE) {
+		ht_matchfinder_slide_window(mf);
+		*in_base_p += MATCHFINDER_WINDOW_SIZE;
+		cur_pos = 0;
+	}
+	in_base = *in_base_p;
+	cutoff = cur_pos - MATCHFINDER_WINDOW_SIZE;
+
+	hash = *next_hash;
+	STATIC_ASSERT(HT_MATCHFINDER_REQUIRED_NBYTES == 5);
+	*next_hash = lz_hash(get_unaligned_le32(in_next + 1),
+			     HT_MATCHFINDER_HASH_ORDER);
+	seq = load_u32_unaligned(in_next);
+	prefetchw(&mf->hash_tab[*next_hash]);
+#if HT_MATCHFINDER_BUCKET_SIZE == 1
+	/* Hand-unrolled version for BUCKET_SIZE == 1 */
+	cur_node = mf->hash_tab[hash][0];
+	mf->hash_tab[hash][0] = cur_pos;
+	if (cur_node <= cutoff)
+		goto out;
+	matchptr = &in_base[cur_node];
+	if (load_u32_unaligned(matchptr) == seq) {
+		best_len = lz_extend(in_next, matchptr, 4, max_len);
+		best_matchptr = matchptr;
+	}
+#elif HT_MATCHFINDER_BUCKET_SIZE == 2
+	/*
+	 * Hand-unrolled version for BUCKET_SIZE == 2.  The logic here also
+	 * differs slightly in that it copies the first entry to the second even
+	 * if nice_len is reached on the first, as this can be slightly faster.
+	 */
+	cur_node = mf->hash_tab[hash][0];
+	mf->hash_tab[hash][0] = cur_pos;
+	if (cur_node <= cutoff)
+		goto out;
+	matchptr = &in_base[cur_node];
+
+	to_insert = cur_node;
+	cur_node = mf->hash_tab[hash][1];
+	mf->hash_tab[hash][1] = to_insert;
+
+	if (load_u32_unaligned(matchptr) == seq) {
+		best_len = lz_extend(in_next, matchptr, 4, max_len);
+		best_matchptr = matchptr;
+		if (cur_node <= cutoff || best_len >= nice_len)
+			goto out;
+		matchptr = &in_base[cur_node];
+		if (load_u32_unaligned(matchptr) == seq &&
+		    load_u32_unaligned(matchptr + best_len - 3) ==
+		    load_u32_unaligned(in_next + best_len - 3)) {
+			len = lz_extend(in_next, matchptr, 4, max_len);
+			if (len > best_len) {
+				best_len = len;
+				best_matchptr = matchptr;
+			}
+		}
+	} else {
+		if (cur_node <= cutoff)
+			goto out;
+		matchptr = &in_base[cur_node];
+		if (load_u32_unaligned(matchptr) == seq) {
+			best_len = lz_extend(in_next, matchptr, 4, max_len);
+			best_matchptr = matchptr;
+		}
+	}
+#else
+	/* Generic version for HT_MATCHFINDER_BUCKET_SIZE > 2 */
+	to_insert = cur_pos;
+	for (i = 0; i < HT_MATCHFINDER_BUCKET_SIZE; i++) {
+		cur_node = mf->hash_tab[hash][i];
+		mf->hash_tab[hash][i] = to_insert;
+		if (cur_node <= cutoff)
+			goto out;
+		matchptr = &in_base[cur_node];
+		if (load_u32_unaligned(matchptr) == seq) {
+			len = lz_extend(in_next, matchptr, 4, max_len);
+			if (len > best_len) {
+				best_len = len;
+				best_matchptr = matchptr;
+				if (best_len >= nice_len)
+					goto out;
+			}
+		}
+		to_insert = cur_node;
+	}
+#endif
+out:
+	*offset_ret = in_next - best_matchptr;
+	return best_len;
+}
+
+static forceinline void
+ht_matchfinder_skip_bytes(struct ht_matchfinder * const mf,
+			  const u8 ** const in_base_p,
+			  const u8 *in_next,
+			  const u8 * const in_end,
+			  const u32 count,
+			  u32 * const next_hash)
+{
+	s32 cur_pos = in_next - *in_base_p;
+	u32 hash;
+	u32 remaining = count;
+	int i;
+
+	if (unlikely(count + HT_MATCHFINDER_REQUIRED_NBYTES > in_end - in_next))
+		return;
+
+	if (cur_pos + count - 1 >= MATCHFINDER_WINDOW_SIZE) {
+		ht_matchfinder_slide_window(mf);
+		*in_base_p += MATCHFINDER_WINDOW_SIZE;
+		cur_pos -= MATCHFINDER_WINDOW_SIZE;
+	}
+
+	hash = *next_hash;
+	do {
+		for (i = HT_MATCHFINDER_BUCKET_SIZE - 1; i > 0; i--)
+			mf->hash_tab[hash][i] = mf->hash_tab[hash][i - 1];
+		mf->hash_tab[hash][0] = cur_pos;
+
+		hash = lz_hash(get_unaligned_le32(++in_next),
+			       HT_MATCHFINDER_HASH_ORDER);
+		cur_pos++;
+	} while (--remaining);
+
+	prefetchw(&mf->hash_tab[hash]);
+	*next_hash = hash;
+}
+
+#endif /* LIB_HT_MATCHFINDER_H */

data/ext/deflate_ruby/libdeflate/lib/lib_common.h

@@ -0,0 +1,106 @@
+/*
+ * lib_common.h - internal header included by all library code
+ */
+
+#ifndef LIB_LIB_COMMON_H
+#define LIB_LIB_COMMON_H
+
+#ifdef LIBDEFLATE_H
+ /*
+  * When building the library, LIBDEFLATEAPI needs to be defined properly before
+  * including libdeflate.h.
+  */
+#  error "lib_common.h must always be included before libdeflate.h"
+#endif
+
+#if defined(LIBDEFLATE_DLL) && (defined(_WIN32) || defined(__CYGWIN__))
+#  define LIBDEFLATE_EXPORT_SYM  __declspec(dllexport)
+#elif defined(__GNUC__)
+#  define LIBDEFLATE_EXPORT_SYM  __attribute__((visibility("default")))
+#else
+#  define LIBDEFLATE_EXPORT_SYM
+#endif
+
+/*
+ * On i386, gcc assumes that the stack is 16-byte aligned at function entry.
+ * However, some compilers (e.g. MSVC) and programming languages (e.g. Delphi)
+ * only guarantee 4-byte alignment when calling functions.  This is mainly an
+ * issue on Windows, but it has been seen on Linux too.  Work around this ABI
+ * incompatibility by realigning the stack pointer when entering libdeflate.
+ * This prevents crashes in SSE/AVX code.
+ */
+#if defined(__GNUC__) && defined(__i386__)
+#  define LIBDEFLATE_ALIGN_STACK  __attribute__((force_align_arg_pointer))
+#else
+#  define LIBDEFLATE_ALIGN_STACK
+#endif
+
+#define LIBDEFLATEAPI	LIBDEFLATE_EXPORT_SYM LIBDEFLATE_ALIGN_STACK
+
+#include "../common_defs.h"
+
+typedef void *(*malloc_func_t)(size_t);
+typedef void (*free_func_t)(void *);
+
+extern malloc_func_t libdeflate_default_malloc_func;
+extern free_func_t libdeflate_default_free_func;
+
+void *libdeflate_aligned_malloc(malloc_func_t malloc_func,
+				size_t alignment, size_t size);
+void libdeflate_aligned_free(free_func_t free_func, void *ptr);
+
+#ifdef FREESTANDING
+/*
+ * With -ffreestanding, <string.h> may be missing, and we must provide
+ * implementations of memset(), memcpy(), memmove(), and memcmp().
+ * See https://gcc.gnu.org/onlinedocs/gcc/Standards.html
+ *
+ * Also, -ffreestanding disables interpreting calls to these functions as
+ * built-ins.  E.g., calling memcpy(&v, p, WORDBYTES) will make a function call,
+ * not be optimized to a single load instruction.  For performance reasons we
+ * don't want that.  So, declare these functions as macros that expand to the
+ * corresponding built-ins.  This approach is recommended in the gcc man page.
+ * We still need the actual function definitions in case gcc calls them.
+ */
+void *memset(void *s, int c, size_t n);
+#define memset(s, c, n)		__builtin_memset((s), (c), (n))
+
+void *memcpy(void *dest, const void *src, size_t n);
+#define memcpy(dest, src, n)	__builtin_memcpy((dest), (src), (n))
+
+void *memmove(void *dest, const void *src, size_t n);
+#define memmove(dest, src, n)	__builtin_memmove((dest), (src), (n))
+
+int memcmp(const void *s1, const void *s2, size_t n);
+#define memcmp(s1, s2, n)	__builtin_memcmp((s1), (s2), (n))
+
+#undef LIBDEFLATE_ENABLE_ASSERTIONS
+#else
+#  include <string.h>
+   /*
+    * To prevent false positive static analyzer warnings, ensure that assertions
+    * are visible to the static analyzer.
+    */
+#  ifdef __clang_analyzer__
+#    define LIBDEFLATE_ENABLE_ASSERTIONS
+#  endif
+#endif
+
+/*
+ * Runtime assertion support.  Don't enable this in production builds; it may
+ * hurt performance significantly.
+ */
+#ifdef LIBDEFLATE_ENABLE_ASSERTIONS
+NORETURN void
+libdeflate_assertion_failed(const char *expr, const char *file, int line);
+#define ASSERT(expr) { if (unlikely(!(expr))) \
+	libdeflate_assertion_failed(#expr, __FILE__, __LINE__); }
+#else
+#define ASSERT(expr) (void)(expr)
+#endif
+
+#define CONCAT_IMPL(a, b)	a##b
+#define CONCAT(a, b)		CONCAT_IMPL(a, b)
+#define ADD_SUFFIX(name)	CONCAT(name, SUFFIX)
+
+#endif /* LIB_LIB_COMMON_H */

data/ext/deflate_ruby/libdeflate/lib/matchfinder_common.h

@@ -0,0 +1,224 @@
+/*
+ * matchfinder_common.h - common code for Lempel-Ziv matchfinding
+ */
+
+#ifndef LIB_MATCHFINDER_COMMON_H
+#define LIB_MATCHFINDER_COMMON_H
+
+#include "lib_common.h"
+
+#ifndef MATCHFINDER_WINDOW_ORDER
+#  error "MATCHFINDER_WINDOW_ORDER must be defined!"
+#endif
+
+/*
+ * Given a 32-bit value that was loaded with the platform's native endianness,
+ * return a 32-bit value whose high-order 8 bits are 0 and whose low-order 24
+ * bits contain the first 3 bytes, arranged in octets in a platform-dependent
+ * order, at the memory location from which the input 32-bit value was loaded.
+ */
+static forceinline u32
+loaded_u32_to_u24(u32 v)
+{
+	if (CPU_IS_LITTLE_ENDIAN())
+		return v & 0xFFFFFF;
+	else
+		return v >> 8;
+}
+
+/*
+ * Load the next 3 bytes from @p into the 24 low-order bits of a 32-bit value.
+ * The order in which the 3 bytes will be arranged as octets in the 24 bits is
+ * platform-dependent.  At least 4 bytes (not 3) must be available at @p.
+ */
+static forceinline u32
+load_u24_unaligned(const u8 *p)
+{
+#if UNALIGNED_ACCESS_IS_FAST
+	return loaded_u32_to_u24(load_u32_unaligned(p));
+#else
+	if (CPU_IS_LITTLE_ENDIAN())
+		return ((u32)p[0] << 0) | ((u32)p[1] << 8) | ((u32)p[2] << 16);
+	else
+		return ((u32)p[2] << 0) | ((u32)p[1] << 8) | ((u32)p[0] << 16);
+#endif
+}
+
+#define MATCHFINDER_WINDOW_SIZE (1UL << MATCHFINDER_WINDOW_ORDER)
+
+typedef s16 mf_pos_t;
+
+#define MATCHFINDER_INITVAL ((mf_pos_t)-MATCHFINDER_WINDOW_SIZE)
+
+/*
+ * This is the memory address alignment, in bytes, required for the matchfinder
+ * buffers by the architecture-specific implementations of matchfinder_init()
+ * and matchfinder_rebase().  "Matchfinder buffer" means an entire struct
+ * hc_matchfinder, bt_matchfinder, or ht_matchfinder; the next_tab field of
+ * struct hc_matchfinder; or the child_tab field of struct bt_matchfinder.
+ *
+ * This affects how the entire 'struct deflate_compressor' is allocated, since
+ * the matchfinder structures are embedded inside it.
+ *
+ * Currently the maximum memory address alignment required is 32 bytes, needed
+ * by the AVX-2 matchfinder functions.
+ */
+#define MATCHFINDER_MEM_ALIGNMENT	32
+
+/*
+ * This declares a size, in bytes, that is guaranteed to divide the sizes of the
+ * matchfinder buffers (where "matchfinder buffers" is as defined for
+ * MATCHFINDER_MEM_ALIGNMENT).  The architecture-specific implementations of
+ * matchfinder_init() and matchfinder_rebase() take advantage of this value.
+ *
+ * Currently the maximum size alignment required is 128 bytes, needed by
+ * the AVX-2 matchfinder functions.  However, the RISC-V Vector Extension
+ * matchfinder functions can, in principle, take advantage of a larger size
+ * alignment.  Therefore, we set this to 1024, which still easily divides the
+ * actual sizes that result from the current matchfinder struct definitions.
+ * This value can safely be changed to any power of two that is >= 128.
+ */
+#define MATCHFINDER_SIZE_ALIGNMENT	1024
+
+#undef matchfinder_init
+#undef matchfinder_rebase
+#ifdef _aligned_attribute
+#  define MATCHFINDER_ALIGNED _aligned_attribute(MATCHFINDER_MEM_ALIGNMENT)
+#  if defined(ARCH_ARM32) || defined(ARCH_ARM64)
+#    include "arm/matchfinder_impl.h"
+#  elif defined(ARCH_RISCV)
+#    include "riscv/matchfinder_impl.h"
+#  elif defined(ARCH_X86_32) || defined(ARCH_X86_64)
+#    include "x86/matchfinder_impl.h"
+#  endif
+#else
+#  define MATCHFINDER_ALIGNED
+#endif
+
+/*
+ * Initialize the hash table portion of the matchfinder.
+ *
+ * Essentially, this is an optimized memset().
+ *
+ * 'data' must be aligned to a MATCHFINDER_MEM_ALIGNMENT boundary, and
+ * 'size' must be a multiple of MATCHFINDER_SIZE_ALIGNMENT.
+ */
+#ifndef matchfinder_init
+static forceinline void
+matchfinder_init(mf_pos_t *data, size_t size)
+{
+	size_t num_entries = size / sizeof(*data);
+	size_t i;
+
+	for (i = 0; i < num_entries; i++)
+		data[i] = MATCHFINDER_INITVAL;
+}
+#endif
+
+/*
+ * Slide the matchfinder by MATCHFINDER_WINDOW_SIZE bytes.
+ *
+ * This must be called just after each MATCHFINDER_WINDOW_SIZE bytes have been
+ * run through the matchfinder.
+ *
+ * This subtracts MATCHFINDER_WINDOW_SIZE bytes from each entry in the given
+ * array, making the entries be relative to the current position rather than the
+ * position MATCHFINDER_WINDOW_SIZE bytes prior.  To avoid integer underflows,
+ * entries that would become less than -MATCHFINDER_WINDOW_SIZE stay at
+ * -MATCHFINDER_WINDOW_SIZE, keeping them permanently out of bounds.
+ *
+ * The given array must contain all matchfinder data that is position-relative:
+ * the hash table(s) as well as any hash chain or binary tree links.  Its
+ * address must be aligned to a MATCHFINDER_MEM_ALIGNMENT boundary, and its size
+ * must be a multiple of MATCHFINDER_SIZE_ALIGNMENT.
+ */
+#ifndef matchfinder_rebase
+static forceinline void
+matchfinder_rebase(mf_pos_t *data, size_t size)
+{
+	size_t num_entries = size / sizeof(*data);
+	size_t i;
+
+	if (MATCHFINDER_WINDOW_SIZE == 32768) {
+		/*
+		 * Branchless version for 32768-byte windows.  Clear all bits if
+		 * the value was already negative, then set the sign bit.  This
+		 * is equivalent to subtracting 32768 with signed saturation.
+		 */
+		for (i = 0; i < num_entries; i++)
+			data[i] = 0x8000 | (data[i] & ~(data[i] >> 15));
+	} else {
+		for (i = 0; i < num_entries; i++) {
+			if (data[i] >= 0)
+				data[i] -= (mf_pos_t)-MATCHFINDER_WINDOW_SIZE;
+			else
+				data[i] = (mf_pos_t)-MATCHFINDER_WINDOW_SIZE;
+		}
+	}
+}
+#endif
+
+/*
+ * The hash function: given a sequence prefix held in the low-order bits of a
+ * 32-bit value, multiply by a carefully-chosen large constant.  Discard any
+ * bits of the product that don't fit in a 32-bit value, but take the
+ * next-highest @num_bits bits of the product as the hash value, as those have
+ * the most randomness.
+ */
+static forceinline u32
+lz_hash(u32 seq, unsigned num_bits)
+{
+	return (u32)(seq * 0x1E35A7BD) >> (32 - num_bits);
+}
+
+/*
+ * Return the number of bytes at @matchptr that match the bytes at @strptr, up
+ * to a maximum of @max_len.  Initially, @start_len bytes are matched.
+ */
+static forceinline unsigned
+lz_extend(const u8 * const strptr, const u8 * const matchptr,
+	  const unsigned start_len, const unsigned max_len)
+{
+	unsigned len = start_len;
+	machine_word_t v_word;
+
+	if (UNALIGNED_ACCESS_IS_FAST) {
+
+		if (likely(max_len - len >= 4 * WORDBYTES)) {
+
+		#define COMPARE_WORD_STEP				\
+			v_word = load_word_unaligned(&matchptr[len]) ^	\
+				 load_word_unaligned(&strptr[len]);	\
+			if (v_word != 0)				\
+				goto word_differs;			\
+			len += WORDBYTES;				\
+
+			COMPARE_WORD_STEP
+			COMPARE_WORD_STEP
+			COMPARE_WORD_STEP
+			COMPARE_WORD_STEP
+		#undef COMPARE_WORD_STEP
+		}
+
+		while (len + WORDBYTES <= max_len) {
+			v_word = load_word_unaligned(&matchptr[len]) ^
+				 load_word_unaligned(&strptr[len]);
+			if (v_word != 0)
+				goto word_differs;
+			len += WORDBYTES;
+		}
+	}
+
+	while (len < max_len && matchptr[len] == strptr[len])
+		len++;
+	return len;
+
+word_differs:
+	if (CPU_IS_LITTLE_ENDIAN())
+		len += (bsfw(v_word) >> 3);
+	else
+		len += (WORDBITS - 1 - bsrw(v_word)) >> 3;
+	return len;
+}
+
+#endif /* LIB_MATCHFINDER_COMMON_H */

data/ext/deflate_ruby/libdeflate/lib/riscv/matchfinder_impl.h

@@ -0,0 +1,97 @@
+/*
+ * riscv/matchfinder_impl.h - RISC-V implementations of matchfinder functions
+ *
+ * Copyright 2024 Eric Biggers
+ *
+ * Permission is hereby granted, free of charge, to any person
+ * obtaining a copy of this software and associated documentation
+ * files (the "Software"), to deal in the Software without
+ * restriction, including without limitation the rights to use,
+ * copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following
+ * conditions:
+ *
+ * The above copyright notice and this permission notice shall be
+ * included in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
+ * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
+ * HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+ * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#ifndef LIB_RISCV_MATCHFINDER_IMPL_H
+#define LIB_RISCV_MATCHFINDER_IMPL_H
+
+#if defined(ARCH_RISCV) && defined(__riscv_vector)
+#include <riscv_vector.h>
+
+/*
+ * Return the maximum number of 16-bit (mf_pos_t) elements that fit in 8 RISC-V
+ * vector registers and also evenly divide the sizes of the matchfinder buffers.
+ */
+static forceinline size_t
+riscv_matchfinder_vl(void)
+{
+	const size_t vl = __riscv_vsetvlmax_e16m8();
+
+	STATIC_ASSERT(sizeof(mf_pos_t) == sizeof(s16));
+	/*
+	 * MATCHFINDER_SIZE_ALIGNMENT is a power of 2, as is 'vl' because the
+	 * RISC-V Vector Extension requires that the vector register length
+	 * (VLEN) be a power of 2.  Thus, a simple MIN() gives the correct
+	 * answer here; rounding to a power of 2 is not required.
+	 */
+	STATIC_ASSERT((MATCHFINDER_SIZE_ALIGNMENT &
+		       (MATCHFINDER_SIZE_ALIGNMENT - 1)) == 0);
+	ASSERT((vl & (vl - 1)) == 0);
+	return MIN(vl, MATCHFINDER_SIZE_ALIGNMENT / sizeof(mf_pos_t));
+}
+
+/* matchfinder_init() optimized using the RISC-V Vector Extension */
+static forceinline void
+matchfinder_init_rvv(mf_pos_t *p, size_t size)
+{
+	const size_t vl = riscv_matchfinder_vl();
+	const vint16m8_t v = __riscv_vmv_v_x_i16m8(MATCHFINDER_INITVAL, vl);
+
+	ASSERT(size > 0 && size % (vl * sizeof(p[0])) == 0);
+	do {
+		__riscv_vse16_v_i16m8(p, v, vl);
+		p += vl;
+		size -= vl * sizeof(p[0]);
+	} while (size != 0);
+}
+#define matchfinder_init matchfinder_init_rvv
+
+/* matchfinder_rebase() optimized using the RISC-V Vector Extension */
+static forceinline void
+matchfinder_rebase_rvv(mf_pos_t *p, size_t size)
+{
+	const size_t vl = riscv_matchfinder_vl();
+
+	ASSERT(size > 0 && size % (vl * sizeof(p[0])) == 0);
+	do {
+		vint16m8_t v = __riscv_vle16_v_i16m8(p, vl);
+
+		/*
+		 * This should generate the vsadd.vx instruction
+		 * (Vector Saturating Add, integer vector-scalar)
+		 */
+		v = __riscv_vsadd_vx_i16m8(v, (s16)-MATCHFINDER_WINDOW_SIZE,
+					   vl);
+		__riscv_vse16_v_i16m8(p, v, vl);
+		p += vl;
+		size -= vl * sizeof(p[0]);
+	} while (size != 0);
+}
+#define matchfinder_rebase matchfinder_rebase_rvv
+
+#endif /* ARCH_RISCV && __riscv_vector */
+
+#endif /* LIB_RISCV_MATCHFINDER_IMPL_H */