Adding upstream version 1.21.8.upstream/1.21.8

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

11 files changed, 2776 insertions, 0 deletions

diff --git a/src/internal/zstd/bits.go b/src/internal/zstd/bits.go
new file mode 100644
index 0000000..c9a2f70
--- /dev/null
+++ b/src/internal/zstd/bits.go
@@ -0,0 +1,130 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package zstd
+
+import (
+	"math/bits"
+)
+
+// block is the data for a single compressed block.
+// The data starts immediately after the 3 byte block header,
+// and is Block_Size bytes long.
+type block []byte
+
+// bitReader reads a bit stream going forward.
+type bitReader struct {
+	r    *Reader // for error reporting
+	data block   // the bits to read
+	off  uint32  // current offset into data
+	bits uint32  // bits ready to be returned
+	cnt  uint32  // number of valid bits in the bits field
+}
+
+// makeBitReader makes a bit reader starting at off.
+func (r *Reader) makeBitReader(data block, off int) bitReader {
+	return bitReader{
+		r:    r,
+		data: data,
+		off:  uint32(off),
+	}
+}
+
+// moreBits is called to read more bits.
+// This ensures that at least 16 bits are available.
+func (br *bitReader) moreBits() error {
+	for br.cnt < 16 {
+		if br.off >= uint32(len(br.data)) {
+			return br.r.makeEOFError(int(br.off))
+		}
+		c := br.data[br.off]
+		br.off++
+		br.bits |= uint32(c) << br.cnt
+		br.cnt += 8
+	}
+	return nil
+}
+
+// val is called to fetch a value of b bits.
+func (br *bitReader) val(b uint8) uint32 {
+	r := br.bits & ((1 << b) - 1)
+	br.bits >>= b
+	br.cnt -= uint32(b)
+	return r
+}
+
+// backup steps back to the last byte we used.
+func (br *bitReader) backup() {
+	for br.cnt >= 8 {
+		br.off--
+		br.cnt -= 8
+	}
+}
+
+// makeError returns an error at the current offset wrapping a string.
+func (br *bitReader) makeError(msg string) error {
+	return br.r.makeError(int(br.off), msg)
+}
+
+// reverseBitReader reads a bit stream in reverse.
+type reverseBitReader struct {
+	r     *Reader // for error reporting
+	data  block   // the bits to read
+	off   uint32  // current offset into data
+	start uint32  // start in data; we read backward to start
+	bits  uint32  // bits ready to be returned
+	cnt   uint32  // number of valid bits in bits field
+}
+
+// makeReverseBitReader makes a reverseBitReader reading backward
+// from off to start. The bitstream starts with a 1 bit in the last
+// byte, at off.
+func (r *Reader) makeReverseBitReader(data block, off, start int) (reverseBitReader, error) {
+	streamStart := data[off]
+	if streamStart == 0 {
+		return reverseBitReader{}, r.makeError(off, "zero byte at reverse bit stream start")
+	}
+	rbr := reverseBitReader{
+		r:     r,
+		data:  data,
+		off:   uint32(off),
+		start: uint32(start),
+		bits:  uint32(streamStart),
+		cnt:   uint32(7 - bits.LeadingZeros8(streamStart)),
+	}
+	return rbr, nil
+}
+
+// val is called to fetch a value of b bits.
+func (rbr *reverseBitReader) val(b uint8) (uint32, error) {
+	if !rbr.fetch(b) {
+		return 0, rbr.r.makeEOFError(int(rbr.off))
+	}
+
+	rbr.cnt -= uint32(b)
+	v := (rbr.bits >> rbr.cnt) & ((1 << b) - 1)
+	return v, nil
+}
+
+// fetch is called to ensure that at least b bits are available.
+// It reports false if this can't be done,
+// in which case only rbr.cnt bits are available.
+func (rbr *reverseBitReader) fetch(b uint8) bool {
+	for rbr.cnt < uint32(b) {
+		if rbr.off <= rbr.start {
+			return false
+		}
+		rbr.off--
+		c := rbr.data[rbr.off]
+		rbr.bits <<= 8
+		rbr.bits |= uint32(c)
+		rbr.cnt += 8
+	}
+	return true
+}
+
+// makeError returns an error at the current offset wrapping a string.
+func (rbr *reverseBitReader) makeError(msg string) error {
+	return rbr.r.makeError(int(rbr.off), msg)
+}
diff --git a/src/internal/zstd/block.go b/src/internal/zstd/block.go
new file mode 100644
index 0000000..bd3040c
--- /dev/null
+++ b/src/internal/zstd/block.go
@@ -0,0 +1,436 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package zstd
+
+import (
+	"io"
+)
+
+// debug can be set in the source to print debug info using println.
+const debug = false
+
+// compressedBlock decompresses a compressed block, storing the decompressed
+// data in r.buffer. The blockSize argument is the compressed size.
+// RFC 3.1.1.3.
+func (r *Reader) compressedBlock(blockSize int) error {
+	if len(r.compressedBuf) >= blockSize {
+		r.compressedBuf = r.compressedBuf[:blockSize]
+	} else {
+		// We know that blockSize <= 128K,
+		// so this won't allocate an enormous amount.
+		need := blockSize - len(r.compressedBuf)
+		r.compressedBuf = append(r.compressedBuf, make([]byte, need)...)
+	}
+
+	if _, err := io.ReadFull(r.r, r.compressedBuf); err != nil {
+		return r.wrapNonEOFError(0, err)
+	}
+
+	data := block(r.compressedBuf)
+	off := 0
+	r.buffer = r.buffer[:0]
+
+	litoff, litbuf, err := r.readLiterals(data, off, r.literals[:0])
+	if err != nil {
+		return err
+	}
+	r.literals = litbuf
+
+	off = litoff
+
+	seqCount, off, err := r.initSeqs(data, off)
+	if err != nil {
+		return err
+	}
+
+	if seqCount == 0 {
+		// No sequences, just literals.
+		if off < len(data) {
+			return r.makeError(off, "extraneous data after no sequences")
+		}
+		if len(litbuf) == 0 {
+			return r.makeError(off, "no sequences and no literals")
+		}
+		r.buffer = append(r.buffer, litbuf...)
+		return nil
+	}
+
+	return r.execSeqs(data, off, litbuf, seqCount)
+}
+
+// seqCode is the kind of sequence codes we have to handle.
+type seqCode int
+
+const (
+	seqLiteral seqCode = iota
+	seqOffset
+	seqMatch
+)
+
+// seqCodeInfoData is the information needed to set up seqTables and
+// seqTableBits for a particular kind of sequence code.
+type seqCodeInfoData struct {
+	predefTable     []fseBaselineEntry // predefined FSE
+	predefTableBits int                // number of bits in predefTable
+	maxSym          int                // max symbol value in FSE
+	maxBits         int                // max bits for FSE
+
+	// toBaseline converts from an FSE table to an FSE baseline table.
+	toBaseline func(*Reader, int, []fseEntry, []fseBaselineEntry) error
+}
+
+// seqCodeInfo is the seqCodeInfoData for each kind of sequence code.
+var seqCodeInfo = [3]seqCodeInfoData{
+	seqLiteral: {
+		predefTable:     predefinedLiteralTable[:],
+		predefTableBits: 6,
+		maxSym:          35,
+		maxBits:         9,
+		toBaseline:      (*Reader).makeLiteralBaselineFSE,
+	},
+	seqOffset: {
+		predefTable:     predefinedOffsetTable[:],
+		predefTableBits: 5,
+		maxSym:          31,
+		maxBits:         8,
+		toBaseline:      (*Reader).makeOffsetBaselineFSE,
+	},
+	seqMatch: {
+		predefTable:     predefinedMatchTable[:],
+		predefTableBits: 6,
+		maxSym:          52,
+		maxBits:         9,
+		toBaseline:      (*Reader).makeMatchBaselineFSE,
+	},
+}
+
+// initSeqs reads the Sequences_Section_Header and sets up the FSE
+// tables used to read the sequence codes. It returns the number of
+// sequences and the new offset. RFC 3.1.1.3.2.1.
+func (r *Reader) initSeqs(data block, off int) (int, int, error) {
+	if off >= len(data) {
+		return 0, 0, r.makeEOFError(off)
+	}
+
+	seqHdr := data[off]
+	off++
+	if seqHdr == 0 {
+		return 0, off, nil
+	}
+
+	var seqCount int
+	if seqHdr < 128 {
+		seqCount = int(seqHdr)
+	} else if seqHdr < 255 {
+		if off >= len(data) {
+			return 0, 0, r.makeEOFError(off)
+		}
+		seqCount = ((int(seqHdr) - 128) << 8) + int(data[off])
+		off++
+	} else {
+		if off+1 >= len(data) {
+			return 0, 0, r.makeEOFError(off)
+		}
+		seqCount = int(data[off]) + (int(data[off+1]) << 8) + 0x7f00
+		off += 2
+	}
+
+	// Read the Symbol_Compression_Modes byte.
+
+	if off >= len(data) {
+		return 0, 0, r.makeEOFError(off)
+	}
+	symMode := data[off]
+	if symMode&3 != 0 {
+		return 0, 0, r.makeError(off, "invalid symbol compression mode")
+	}
+	off++
+
+	// Set up the FSE tables used to decode the sequence codes.
+
+	var err error
+	off, err = r.setSeqTable(data, off, seqLiteral, (symMode>>6)&3)
+	if err != nil {
+		return 0, 0, err
+	}
+
+	off, err = r.setSeqTable(data, off, seqOffset, (symMode>>4)&3)
+	if err != nil {
+		return 0, 0, err
+	}
+
+	off, err = r.setSeqTable(data, off, seqMatch, (symMode>>2)&3)
+	if err != nil {
+		return 0, 0, err
+	}
+
+	return seqCount, off, nil
+}
+
+// setSeqTable uses the Compression_Mode in mode to set up r.seqTables and
+// r.seqTableBits for kind. We store these in the Reader because one of
+// the modes simply reuses the value from the last block in the frame.
+func (r *Reader) setSeqTable(data block, off int, kind seqCode, mode byte) (int, error) {
+	info := &seqCodeInfo[kind]
+	switch mode {
+	case 0:
+		// Predefined_Mode
+		r.seqTables[kind] = info.predefTable
+		r.seqTableBits[kind] = uint8(info.predefTableBits)
+		return off, nil
+
+	case 1:
+		// RLE_Mode
+		if off >= len(data) {
+			return 0, r.makeEOFError(off)
+		}
+		rle := data[off]
+		off++
+
+		// Build a simple baseline table that always returns rle.
+
+		entry := []fseEntry{
+			{
+				sym:  rle,
+				bits: 0,
+				base: 0,
+			},
+		}
+		if cap(r.seqTableBuffers[kind]) == 0 {
+			r.seqTableBuffers[kind] = make([]fseBaselineEntry, 1<<info.maxBits)
+		}
+		r.seqTableBuffers[kind] = r.seqTableBuffers[kind][:1]
+		if err := info.toBaseline(r, off, entry, r.seqTableBuffers[kind]); err != nil {
+			return 0, err
+		}
+
+		r.seqTables[kind] = r.seqTableBuffers[kind]
+		r.seqTableBits[kind] = 0
+		return off, nil
+
+	case 2:
+		// FSE_Compressed_Mode
+		if cap(r.fseScratch) < 1<<info.maxBits {
+			r.fseScratch = make([]fseEntry, 1<<info.maxBits)
+		}
+		r.fseScratch = r.fseScratch[:1<<info.maxBits]
+
+		tableBits, roff, err := r.readFSE(data, off, info.maxSym, info.maxBits, r.fseScratch)
+		if err != nil {
+			return 0, err
+		}
+		r.fseScratch = r.fseScratch[:1<<tableBits]
+
+		if cap(r.seqTableBuffers[kind]) == 0 {
+			r.seqTableBuffers[kind] = make([]fseBaselineEntry, 1<<info.maxBits)
+		}
+		r.seqTableBuffers[kind] = r.seqTableBuffers[kind][:1<<tableBits]
+
+		if err := info.toBaseline(r, roff, r.fseScratch, r.seqTableBuffers[kind]); err != nil {
+			return 0, err
+		}
+
+		r.seqTables[kind] = r.seqTableBuffers[kind]
+		r.seqTableBits[kind] = uint8(tableBits)
+		return roff, nil
+
+	case 3:
+		// Repeat_Mode
+		if len(r.seqTables[kind]) == 0 {
+			return 0, r.makeError(off, "missing repeat sequence FSE table")
+		}
+		return off, nil
+	}
+	panic("unreachable")
+}
+
+// execSeqs reads and executes the sequences. RFC 3.1.1.3.2.1.2.
+func (r *Reader) execSeqs(data block, off int, litbuf []byte, seqCount int) error {
+	// Set up the initial states for the sequence code readers.
+
+	rbr, err := r.makeReverseBitReader(data, len(data)-1, off)
+	if err != nil {
+		return err
+	}
+
+	literalState, err := rbr.val(r.seqTableBits[seqLiteral])
+	if err != nil {
+		return err
+	}
+
+	offsetState, err := rbr.val(r.seqTableBits[seqOffset])
+	if err != nil {
+		return err
+	}
+
+	matchState, err := rbr.val(r.seqTableBits[seqMatch])
+	if err != nil {
+		return err
+	}
+
+	// Read and perform all the sequences. RFC 3.1.1.4.
+
+	seq := 0
+	for seq < seqCount {
+		if len(r.buffer)+len(litbuf) > 128<<10 {
+			return rbr.makeError("uncompressed size too big")
+		}
+
+		ptoffset := &r.seqTables[seqOffset][offsetState]
+		ptmatch := &r.seqTables[seqMatch][matchState]
+		ptliteral := &r.seqTables[seqLiteral][literalState]
+
+		add, err := rbr.val(ptoffset.basebits)
+		if err != nil {
+			return err
+		}
+		offset := ptoffset.baseline + add
+
+		add, err = rbr.val(ptmatch.basebits)
+		if err != nil {
+			return err
+		}
+		match := ptmatch.baseline + add
+
+		add, err = rbr.val(ptliteral.basebits)
+		if err != nil {
+			return err
+		}
+		literal := ptliteral.baseline + add
+
+		// Handle repeat offsets. RFC 3.1.1.5.
+		// See the comment in makeOffsetBaselineFSE.
+		if ptoffset.basebits > 1 {
+			r.repeatedOffset3 = r.repeatedOffset2
+			r.repeatedOffset2 = r.repeatedOffset1
+			r.repeatedOffset1 = offset
+		} else {
+			if literal == 0 {
+				offset++
+			}
+			switch offset {
+			case 1:
+				offset = r.repeatedOffset1
+			case 2:
+				offset = r.repeatedOffset2
+				r.repeatedOffset2 = r.repeatedOffset1
+				r.repeatedOffset1 = offset
+			case 3:
+				offset = r.repeatedOffset3
+				r.repeatedOffset3 = r.repeatedOffset2
+				r.repeatedOffset2 = r.repeatedOffset1
+				r.repeatedOffset1 = offset
+			case 4:
+				offset = r.repeatedOffset1 - 1
+				r.repeatedOffset3 = r.repeatedOffset2
+				r.repeatedOffset2 = r.repeatedOffset1
+				r.repeatedOffset1 = offset
+			}
+		}
+
+		seq++
+		if seq < seqCount {
+			// Update the states.
+			add, err = rbr.val(ptliteral.bits)
+			if err != nil {
+				return err
+			}
+			literalState = uint32(ptliteral.base) + add
+
+			add, err = rbr.val(ptmatch.bits)
+			if err != nil {
+				return err
+			}
+			matchState = uint32(ptmatch.base) + add
+
+			add, err = rbr.val(ptoffset.bits)
+			if err != nil {
+				return err
+			}
+			offsetState = uint32(ptoffset.base) + add
+		}
+
+		// The next sequence is now in literal, offset, match.
+
+		if debug {
+			println("literal", literal, "offset", offset, "match", match)
+		}
+
+		// Copy literal bytes from litbuf.
+		if literal > uint32(len(litbuf)) {
+			return rbr.makeError("literal byte overflow")
+		}
+		if literal > 0 {
+			r.buffer = append(r.buffer, litbuf[:literal]...)
+			litbuf = litbuf[literal:]
+		}
+
+		if match > 0 {
+			if err := r.copyFromWindow(&rbr, offset, match); err != nil {
+				return err
+			}
+		}
+	}
+
+	if len(litbuf) > 0 {
+		r.buffer = append(r.buffer, litbuf...)
+	}
+
+	if rbr.cnt != 0 {
+		return r.makeError(off, "extraneous data after sequences")
+	}
+
+	return nil
+}
+
+// Copy match bytes from the decoded output, or the window, at offset.
+func (r *Reader) copyFromWindow(rbr *reverseBitReader, offset, match uint32) error {
+	if offset == 0 {
+		return rbr.makeError("invalid zero offset")
+	}
+
+	lenBlock := uint32(len(r.buffer))
+	if lenBlock < offset {
+		lenWindow := uint32(len(r.window))
+		windowOffset := offset - lenBlock
+		if windowOffset > lenWindow {
+			return rbr.makeError("offset past window")
+		}
+		from := lenWindow - windowOffset
+		if from+match <= lenWindow {
+			r.buffer = append(r.buffer, r.window[from:from+match]...)
+			return nil
+		}
+		r.buffer = append(r.buffer, r.window[from:]...)
+		copied := lenWindow - from
+		offset -= copied
+		match -= copied
+
+		if offset == 0 && match > 0 {
+			return rbr.makeError("invalid offset")
+		}
+	}
+
+	from := lenBlock - offset
+	if offset >= match {
+		r.buffer = append(r.buffer, r.buffer[from:from+match]...)
+		return nil
+	}
+
+	// We are being asked to copy data that we are adding to the
+	// buffer in the same copy.
+	for match > 0 {
+		var copy uint32
+		if offset >= match {
+			copy = match
+		} else {
+			copy = offset
+		}
+		r.buffer = append(r.buffer, r.buffer[from:from+copy]...)
+		match -= copy
+		from += copy
+	}
+	return nil
+}
diff --git a/src/internal/zstd/fse.go b/src/internal/zstd/fse.go
new file mode 100644
index 0000000..ea661d4
--- /dev/null
+++ b/src/internal/zstd/fse.go
@@ -0,0 +1,437 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package zstd
+
+import (
+	"math/bits"
+)
+
+// fseEntry is one entry in an FSE table.
+type fseEntry struct {
+	sym  uint8  // value that this entry records
+	bits uint8  // number of bits to read to determine next state
+	base uint16 // add those bits to this state to get the next state
+}
+
+// readFSE reads an FSE table from data starting at off.
+// maxSym is the maximum symbol value.
+// maxBits is the maximum number of bits permitted for symbols in the table.
+// The FSE is written into table, which must be at least 1<<maxBits in size.
+// This returns the number of bits in the FSE table and the new offset.
+// RFC 4.1.1.
+func (r *Reader) readFSE(data block, off, maxSym, maxBits int, table []fseEntry) (tableBits, roff int, err error) {
+	br := r.makeBitReader(data, off)
+	if err := br.moreBits(); err != nil {
+		return 0, 0, err
+	}
+
+	accuracyLog := int(br.val(4)) + 5
+	if accuracyLog > maxBits {
+		return 0, 0, br.makeError("FSE accuracy log too large")
+	}
+
+	// The number of remaining probabilities, plus 1.
+	// This determines the number of bits to be read for the next value.
+	remaining := (1 << accuracyLog) + 1
+
+	// The current difference between small and large values,
+	// which depends on the number of remaining values.
+	// Small values use 1 less bit.
+	threshold := 1 << accuracyLog
+
+	// The number of bits needed to compute threshold.
+	bitsNeeded := accuracyLog + 1
+
+	// The next character value.
+	sym := 0
+
+	// Whether the last count was 0.
+	prev0 := false
+
+	var norm [256]int16
+
+	for remaining > 1 && sym <= maxSym {
+		if err := br.moreBits(); err != nil {
+			return 0, 0, err
+		}
+
+		if prev0 {
+			// Previous count was 0, so there is a 2-bit
+			// repeat flag. If the 2-bit flag is 0b11,
+			// it adds 3 and then there is another repeat flag.
+			zsym := sym
+			for (br.bits & 0xfff) == 0xfff {
+				zsym += 3 * 6
+				br.bits >>= 12
+				br.cnt -= 12
+				if err := br.moreBits(); err != nil {
+					return 0, 0, err
+				}
+			}
+			for (br.bits & 3) == 3 {
+				zsym += 3
+				br.bits >>= 2
+				br.cnt -= 2
+				if err := br.moreBits(); err != nil {
+					return 0, 0, err
+				}
+			}
+
+			// We have at least 14 bits here,
+			// no need to call moreBits
+
+			zsym += int(br.val(2))
+
+			if zsym > maxSym {
+				return 0, 0, br.makeError("FSE symbol index overflow")
+			}
+
+			for ; sym < zsym; sym++ {
+				norm[uint8(sym)] = 0
+			}
+
+			prev0 = false
+			continue
+		}
+
+		max := (2*threshold - 1) - remaining
+		var count int
+		if int(br.bits&uint32(threshold-1)) < max {
+			// A small value.
+			count = int(br.bits & uint32((threshold - 1)))
+			br.bits >>= bitsNeeded - 1
+			br.cnt -= uint32(bitsNeeded - 1)
+		} else {
+			// A large value.
+			count = int(br.bits & uint32((2*threshold - 1)))
+			if count >= threshold {
+				count -= max
+			}
+			br.bits >>= bitsNeeded
+			br.cnt -= uint32(bitsNeeded)
+		}
+
+		count--
+		if count >= 0 {
+			remaining -= count
+		} else {
+			remaining--
+		}
+		if sym >= 256 {
+			return 0, 0, br.makeError("FSE sym overflow")
+		}
+		norm[uint8(sym)] = int16(count)
+		sym++
+
+		prev0 = count == 0
+
+		for remaining < threshold {
+			bitsNeeded--
+			threshold >>= 1
+		}
+	}
+
+	if remaining != 1 {
+		return 0, 0, br.makeError("too many symbols in FSE table")
+	}
+
+	for ; sym <= maxSym; sym++ {
+		norm[uint8(sym)] = 0
+	}
+
+	br.backup()
+
+	if err := r.buildFSE(off, norm[:maxSym+1], table, accuracyLog); err != nil {
+		return 0, 0, err
+	}
+
+	return accuracyLog, int(br.off), nil
+}
+
+// buildFSE builds an FSE decoding table from a list of probabilities.
+// The probabilities are in norm. next is scratch space. The number of bits
+// in the table is tableBits.
+func (r *Reader) buildFSE(off int, norm []int16, table []fseEntry, tableBits int) error {
+	tableSize := 1 << tableBits
+	highThreshold := tableSize - 1
+
+	var next [256]uint16
+
+	for i, n := range norm {
+		if n >= 0 {
+			next[uint8(i)] = uint16(n)
+		} else {
+			table[highThreshold].sym = uint8(i)
+			highThreshold--
+			next[uint8(i)] = 1
+		}
+	}
+
+	pos := 0
+	step := (tableSize >> 1) + (tableSize >> 3) + 3
+	mask := tableSize - 1
+	for i, n := range norm {
+		for j := 0; j < int(n); j++ {
+			table[pos].sym = uint8(i)
+			pos = (pos + step) & mask
+			for pos > highThreshold {
+				pos = (pos + step) & mask
+			}
+		}
+	}
+	if pos != 0 {
+		return r.makeError(off, "FSE count error")
+	}
+
+	for i := 0; i < tableSize; i++ {
+		sym := table[i].sym
+		nextState := next[sym]
+		next[sym]++
+
+		if nextState == 0 {
+			return r.makeError(off, "FSE state error")
+		}
+
+		highBit := 15 - bits.LeadingZeros16(nextState)
+
+		bits := tableBits - highBit
+		table[i].bits = uint8(bits)
+		table[i].base = (nextState << bits) - uint16(tableSize)
+	}
+
+	return nil
+}
+
+// fseBaselineEntry is an entry in an FSE baseline table.
+// We use these for literal/match/length values.
+// Those require mapping the symbol to a baseline value,
+// and then reading zero or more bits and adding the value to the baseline.
+// Rather than looking thees up in separate tables,
+// we convert the FSE table to an FSE baseline table.
+type fseBaselineEntry struct {
+	baseline uint32 // baseline for value that this entry represents
+	basebits uint8  // number of bits to read to add to baseline
+	bits     uint8  // number of bits to read to determine next state
+	base     uint16 // add the bits to this base to get the next state
+}
+
+// Given a literal length code, we need to read a number of bits and
+// add that to a baseline. For states 0 to 15 the baseline is the
+// state and the number of bits is zero. RFC 3.1.1.3.2.1.1.
+
+const literalLengthOffset = 16
+
+var literalLengthBase = []uint32{
+	16 | (1 << 24),
+	18 | (1 << 24),
+	20 | (1 << 24),
+	22 | (1 << 24),
+	24 | (2 << 24),
+	28 | (2 << 24),
+	32 | (3 << 24),
+	40 | (3 << 24),
+	48 | (4 << 24),
+	64 | (6 << 24),
+	128 | (7 << 24),
+	256 | (8 << 24),
+	512 | (9 << 24),
+	1024 | (10 << 24),
+	2048 | (11 << 24),
+	4096 | (12 << 24),
+	8192 | (13 << 24),
+	16384 | (14 << 24),
+	32768 | (15 << 24),
+	65536 | (16 << 24),
+}
+
+// makeLiteralBaselineFSE converts the literal length fseTable to baselineTable.
+func (r *Reader) makeLiteralBaselineFSE(off int, fseTable []fseEntry, baselineTable []fseBaselineEntry) error {
+	for i, e := range fseTable {
+		be := fseBaselineEntry{
+			bits: e.bits,
+			base: e.base,
+		}
+		if e.sym < literalLengthOffset {
+			be.baseline = uint32(e.sym)
+			be.basebits = 0
+		} else {
+			if e.sym > 35 {
+				return r.makeError(off, "FSE baseline symbol overflow")
+			}
+			idx := e.sym - literalLengthOffset
+			basebits := literalLengthBase[idx]
+			be.baseline = basebits & 0xffffff
+			be.basebits = uint8(basebits >> 24)
+		}
+		baselineTable[i] = be
+	}
+	return nil
+}
+
+// makeOffsetBaselineFSE converts the offset length fseTable to baselineTable.
+func (r *Reader) makeOffsetBaselineFSE(off int, fseTable []fseEntry, baselineTable []fseBaselineEntry) error {
+	for i, e := range fseTable {
+		be := fseBaselineEntry{
+			bits: e.bits,
+			base: e.base,
+		}
+		if e.sym > 31 {
+			return r.makeError(off, "FSE offset symbol overflow")
+		}
+
+		// The simple way to write this is
+		//     be.baseline = 1 << e.sym
+		//     be.basebits = e.sym
+		// That would give us an offset value that corresponds to
+		// the one described in the RFC. However, for offsets > 3
+		// we have to subtract 3. And for offset values 1, 2, 3
+		// we use a repeated offset.
+		//
+		// The baseline is always a power of 2, and is never 0,
+		// so for those low values we will see one entry that is
+		// baseline 1, basebits 0, and one entry that is baseline 2,
+		// basebits 1. All other entries will have baseline >= 4
+		// basebits >= 2.
+		//
+		// So we can check for RFC offset <= 3 by checking for
+		// basebits <= 1. That means that we can subtract 3 here
+		// and not worry about doing it in the hot loop.
+
+		be.baseline = 1 << e.sym
+		if e.sym >= 2 {
+			be.baseline -= 3
+		}
+		be.basebits = e.sym
+		baselineTable[i] = be
+	}
+	return nil
+}
+
+// Given a match length code, we need to read a number of bits and add
+// that to a baseline. For states 0 to 31 the baseline is state+3 and
+// the number of bits is zero. RFC 3.1.1.3.2.1.1.
+
+const matchLengthOffset = 32
+
+var matchLengthBase = []uint32{
+	35 | (1 << 24),
+	37 | (1 << 24),
+	39 | (1 << 24),
+	41 | (1 << 24),
+	43 | (2 << 24),
+	47 | (2 << 24),
+	51 | (3 << 24),
+	59 | (3 << 24),
+	67 | (4 << 24),
+	83 | (4 << 24),
+	99 | (5 << 24),
+	131 | (7 << 24),
+	259 | (8 << 24),
+	515 | (9 << 24),
+	1027 | (10 << 24),
+	2051 | (11 << 24),
+	4099 | (12 << 24),
+	8195 | (13 << 24),
+	16387 | (14 << 24),
+	32771 | (15 << 24),
+	65539 | (16 << 24),
+}
+
+// makeMatchBaselineFSE converts the match length fseTable to baselineTable.
+func (r *Reader) makeMatchBaselineFSE(off int, fseTable []fseEntry, baselineTable []fseBaselineEntry) error {
+	for i, e := range fseTable {
+		be := fseBaselineEntry{
+			bits: e.bits,
+			base: e.base,
+		}
+		if e.sym < matchLengthOffset {
+			be.baseline = uint32(e.sym) + 3
+			be.basebits = 0
+		} else {
+			if e.sym > 52 {
+				return r.makeError(off, "FSE baseline symbol overflow")
+			}
+			idx := e.sym - matchLengthOffset
+			basebits := matchLengthBase[idx]
+			be.baseline = basebits & 0xffffff
+			be.basebits = uint8(basebits >> 24)
+		}
+		baselineTable[i] = be
+	}
+	return nil
+}
+
+// predefinedLiteralTable is the predefined table to use for literal lengths.
+// Generated from table in RFC 3.1.1.3.2.2.1.
+// Checked by TestPredefinedTables.
+var predefinedLiteralTable = [...]fseBaselineEntry{
+	{0, 0, 4, 0}, {0, 0, 4, 16}, {1, 0, 5, 32},
+	{3, 0, 5, 0}, {4, 0, 5, 0}, {6, 0, 5, 0},
+	{7, 0, 5, 0}, {9, 0, 5, 0}, {10, 0, 5, 0},
+	{12, 0, 5, 0}, {14, 0, 6, 0}, {16, 1, 5, 0},
+	{20, 1, 5, 0}, {22, 1, 5, 0}, {28, 2, 5, 0},
+	{32, 3, 5, 0}, {48, 4, 5, 0}, {64, 6, 5, 32},
+	{128, 7, 5, 0}, {256, 8, 6, 0}, {1024, 10, 6, 0},
+	{4096, 12, 6, 0}, {0, 0, 4, 32}, {1, 0, 4, 0},
+	{2, 0, 5, 0}, {4, 0, 5, 32}, {5, 0, 5, 0},
+	{7, 0, 5, 32}, {8, 0, 5, 0}, {10, 0, 5, 32},
+	{11, 0, 5, 0}, {13, 0, 6, 0}, {16, 1, 5, 32},
+	{18, 1, 5, 0}, {22, 1, 5, 32}, {24, 2, 5, 0},
+	{32, 3, 5, 32}, {40, 3, 5, 0}, {64, 6, 4, 0},
+	{64, 6, 4, 16}, {128, 7, 5, 32}, {512, 9, 6, 0},
+	{2048, 11, 6, 0}, {0, 0, 4, 48}, {1, 0, 4, 16},
+	{2, 0, 5, 32}, {3, 0, 5, 32}, {5, 0, 5, 32},
+	{6, 0, 5, 32}, {8, 0, 5, 32}, {9, 0, 5, 32},
+	{11, 0, 5, 32}, {12, 0, 5, 32}, {15, 0, 6, 0},
+	{18, 1, 5, 32}, {20, 1, 5, 32}, {24, 2, 5, 32},
+	{28, 2, 5, 32}, {40, 3, 5, 32}, {48, 4, 5, 32},
+	{65536, 16, 6, 0}, {32768, 15, 6, 0}, {16384, 14, 6, 0},
+	{8192, 13, 6, 0},
+}
+
+// predefinedOffsetTable is the predefined table to use for offsets.
+// Generated from table in RFC 3.1.1.3.2.2.3.
+// Checked by TestPredefinedTables.
+var predefinedOffsetTable = [...]fseBaselineEntry{
+	{1, 0, 5, 0}, {61, 6, 4, 0}, {509, 9, 5, 0},
+	{32765, 15, 5, 0}, {2097149, 21, 5, 0}, {5, 3, 5, 0},
+	{125, 7, 4, 0}, {4093, 12, 5, 0}, {262141, 18, 5, 0},
+	{8388605, 23, 5, 0}, {29, 5, 5, 0}, {253, 8, 4, 0},
+	{16381, 14, 5, 0}, {1048573, 20, 5, 0}, {1, 2, 5, 0},
+	{125, 7, 4, 16}, {2045, 11, 5, 0}, {131069, 17, 5, 0},
+	{4194301, 22, 5, 0}, {13, 4, 5, 0}, {253, 8, 4, 16},
+	{8189, 13, 5, 0}, {524285, 19, 5, 0}, {2, 1, 5, 0},
+	{61, 6, 4, 16}, {1021, 10, 5, 0}, {65533, 16, 5, 0},
+	{268435453, 28, 5, 0}, {134217725, 27, 5, 0}, {67108861, 26, 5, 0},
+	{33554429, 25, 5, 0}, {16777213, 24, 5, 0},
+}
+
+// predefinedMatchTable is the predefined table to use for match lengths.
+// Generated from table in RFC 3.1.1.3.2.2.2.
+// Checked by TestPredefinedTables.
+var predefinedMatchTable = [...]fseBaselineEntry{
+	{3, 0, 6, 0}, {4, 0, 4, 0}, {5, 0, 5, 32},
+	{6, 0, 5, 0}, {8, 0, 5, 0}, {9, 0, 5, 0},
+	{11, 0, 5, 0}, {13, 0, 6, 0}, {16, 0, 6, 0},
+	{19, 0, 6, 0}, {22, 0, 6, 0}, {25, 0, 6, 0},
+	{28, 0, 6, 0}, {31, 0, 6, 0}, {34, 0, 6, 0},
+	{37, 1, 6, 0}, {41, 1, 6, 0}, {47, 2, 6, 0},
+	{59, 3, 6, 0}, {83, 4, 6, 0}, {131, 7, 6, 0},
+	{515, 9, 6, 0}, {4, 0, 4, 16}, {5, 0, 4, 0},
+	{6, 0, 5, 32}, {7, 0, 5, 0}, {9, 0, 5, 32},
+	{10, 0, 5, 0}, {12, 0, 6, 0}, {15, 0, 6, 0},
+	{18, 0, 6, 0}, {21, 0, 6, 0}, {24, 0, 6, 0},
+	{27, 0, 6, 0}, {30, 0, 6, 0}, {33, 0, 6, 0},
+	{35, 1, 6, 0}, {39, 1, 6, 0}, {43, 2, 6, 0},
+	{51, 3, 6, 0}, {67, 4, 6, 0}, {99, 5, 6, 0},
+	{259, 8, 6, 0}, {4, 0, 4, 32}, {4, 0, 4, 48},
+	{5, 0, 4, 16}, {7, 0, 5, 32}, {8, 0, 5, 32},
+	{10, 0, 5, 32}, {11, 0, 5, 32}, {14, 0, 6, 0},
+	{17, 0, 6, 0}, {20, 0, 6, 0}, {23, 0, 6, 0},
+	{26, 0, 6, 0}, {29, 0, 6, 0}, {32, 0, 6, 0},
+	{65539, 16, 6, 0}, {32771, 15, 6, 0}, {16387, 14, 6, 0},
+	{8195, 13, 6, 0}, {4099, 12, 6, 0}, {2051, 11, 6, 0},
+	{1027, 10, 6, 0},
+}
diff --git a/src/internal/zstd/fse_test.go b/src/internal/zstd/fse_test.go
new file mode 100644
index 0000000..6f106b6
--- /dev/null
+++ b/src/internal/zstd/fse_test.go
@@ -0,0 +1,89 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package zstd
+
+import (
+	"slices"
+	"testing"
+)
+
+// literalPredefinedDistribution is the predefined distribution table
+// for literal lengths. RFC 3.1.1.3.2.2.1.
+var literalPredefinedDistribution = []int16{
+	4, 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1, 1, 1,
+	2, 2, 2, 2, 2, 2, 2, 2, 2, 3, 2, 1, 1, 1, 1, 1,
+	-1, -1, -1, -1,
+}
+
+// offsetPredefinedDistribution is the predefined distribution table
+// for offsets. RFC 3.1.1.3.2.2.3.
+var offsetPredefinedDistribution = []int16{
+	1, 1, 1, 1, 1, 1, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1,
+	1, 1, 1, 1, 1, 1, 1, 1, -1, -1, -1, -1, -1,
+}
+
+// matchPredefinedDistribution is the predefined distribution table
+// for match lengths. RFC 3.1.1.3.2.2.2.
+var matchPredefinedDistribution = []int16{
+	1, 4, 3, 2, 2, 2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1,
+	1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+	1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, -1, -1,
+	-1, -1, -1, -1, -1,
+}
+
+// TestPredefinedTables verifies that we can generate the predefined
+// literal/offset/match tables from the input data in RFC 8878.
+// This serves as a test of the predefined tables, and also of buildFSE
+// and the functions that make baseline FSE tables.
+func TestPredefinedTables(t *testing.T) {
+	tests := []struct {
+		name         string
+		distribution []int16
+		tableBits    int
+		toBaseline   func(*Reader, int, []fseEntry, []fseBaselineEntry) error
+		predef       []fseBaselineEntry
+	}{
+		{
+			name:         "literal",
+			distribution: literalPredefinedDistribution,
+			tableBits:    6,
+			toBaseline:   (*Reader).makeLiteralBaselineFSE,
+			predef:       predefinedLiteralTable[:],
+		},
+		{
+			name:         "offset",
+			distribution: offsetPredefinedDistribution,
+			tableBits:    5,
+			toBaseline:   (*Reader).makeOffsetBaselineFSE,
+			predef:       predefinedOffsetTable[:],
+		},
+		{
+			name:         "match",
+			distribution: matchPredefinedDistribution,
+			tableBits:    6,
+			toBaseline:   (*Reader).makeMatchBaselineFSE,
+			predef:       predefinedMatchTable[:],
+		},
+	}
+	for _, test := range tests {
+		test := test
+		t.Run(test.name, func(t *testing.T) {
+			var r Reader
+			table := make([]fseEntry, 1<<test.tableBits)
+			if err := r.buildFSE(0, test.distribution, table, test.tableBits); err != nil {
+				t.Fatal(err)
+			}
+
+			baselineTable := make([]fseBaselineEntry, len(table))
+			if err := test.toBaseline(&r, 0, table, baselineTable); err != nil {
+				t.Fatal(err)
+			}
+
+			if !slices.Equal(baselineTable, test.predef) {
+				t.Errorf("got %v, want %v", baselineTable, test.predef)
+			}
+		})
+	}
+}
diff --git a/src/internal/zstd/fuzz_test.go b/src/internal/zstd/fuzz_test.go
new file mode 100644
index 0000000..bb6f0a9
--- /dev/null
+++ b/src/internal/zstd/fuzz_test.go
@@ -0,0 +1,140 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package zstd
+
+import (
+	"bytes"
+	"io"
+	"os"
+	"os/exec"
+	"testing"
+)
+
+// badStrings is some inputs that FuzzReader failed on earlier.
+var badStrings = []string{
+	"(\xb5/\xfdd00,\x05\x00\xc4\x0400000000000000000000000000000000000000000000000000000000000000000000000000000 \xa07100000000000000000000000000000000000000000000000000000000000000000000000000aM\x8a2y0B\b",
+	"(\xb5/\xfd00$\x05\x0020 00X70000a70000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000",
+	"(\xb5/\xfd00$\x05\x0020 00B00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000",
+	"(\xb5/\xfd00}\x00\x0020\x00\x9000000000000",
+	"(\xb5/\xfd00}\x00\x00&0\x02\x830!000000000",
+	"(\xb5/\xfd\x1002000$\x05\x0010\xcc0\xa8100000000100000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000",
+	"(\xb5/\xfd\x1002000$\x05\x0000\xcc0\xa8100d\x0000001000000000000000000000000000000000000000000000000000000000000000000000000\x000000000000000000000000000000000000000000000000000000000000000000000000000000",
+	"(\xb5/\xfd001\x00\x0000000000000000000",
+}
+
+// This is a simple fuzzer to see if the decompressor panics.
+func FuzzReader(f *testing.F) {
+	for _, test := range tests {
+		f.Add([]byte(test.compressed))
+	}
+	for _, s := range badStrings {
+		f.Add([]byte(s))
+	}
+	f.Fuzz(func(t *testing.T, b []byte) {
+		r := NewReader(bytes.NewReader(b))
+		io.Copy(io.Discard, r)
+	})
+}
+
+// Fuzz test to verify that what we decompress is what we compress.
+// This isn't a great fuzz test because the fuzzer can't efficiently
+// explore the space of decompressor behavior, since it can't see
+// what the compressor is doing. But it's better than nothing.
+func FuzzDecompressor(f *testing.F) {
+	if _, err := os.Stat("/usr/bin/zstd"); err != nil {
+		f.Skip("skipping because /usr/bin/zstd does not exist")
+	}
+
+	for _, test := range tests {
+		f.Add([]byte(test.uncompressed))
+	}
+
+	// Add some larger data, as that has more interesting compression.
+	f.Add(bytes.Repeat([]byte("abcdefghijklmnop"), 256))
+	var buf bytes.Buffer
+	for i := 0; i < 256; i++ {
+		buf.WriteByte(byte(i))
+	}
+	f.Add(bytes.Repeat(buf.Bytes(), 64))
+	f.Add(bigData(f))
+
+	f.Fuzz(func(t *testing.T, b []byte) {
+		cmd := exec.Command("/usr/bin/zstd", "-z")
+		cmd.Stdin = bytes.NewReader(b)
+		var compressed bytes.Buffer
+		cmd.Stdout = &compressed
+		cmd.Stderr = os.Stderr
+		if err := cmd.Run(); err != nil {
+			t.Errorf("running zstd failed: %v", err)
+		}
+
+		r := NewReader(bytes.NewReader(compressed.Bytes()))
+		got, err := io.ReadAll(r)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if !bytes.Equal(got, b) {
+			showDiffs(t, got, b)
+		}
+	})
+}
+
+// Fuzz test to check that if we can decompress some data,
+// so can zstd, and that we get the same result.
+func FuzzReverse(f *testing.F) {
+	if _, err := os.Stat("/usr/bin/zstd"); err != nil {
+		f.Skip("skipping because /usr/bin/zstd does not exist")
+	}
+
+	for _, test := range tests {
+		f.Add([]byte(test.compressed))
+	}
+
+	// Set a hook to reject some cases where we don't match zstd.
+	fuzzing = true
+	defer func() { fuzzing = false }()
+
+	f.Fuzz(func(t *testing.T, b []byte) {
+		r := NewReader(bytes.NewReader(b))
+		goExp, goErr := io.ReadAll(r)
+
+		cmd := exec.Command("/usr/bin/zstd", "-d")
+		cmd.Stdin = bytes.NewReader(b)
+		var uncompressed bytes.Buffer
+		cmd.Stdout = &uncompressed
+		cmd.Stderr = os.Stderr
+		zstdErr := cmd.Run()
+		zstdExp := uncompressed.Bytes()
+
+		if goErr == nil && zstdErr == nil {
+			if !bytes.Equal(zstdExp, goExp) {
+				showDiffs(t, zstdExp, goExp)
+			}
+		} else {
+			// Ideally we should check that this package and
+			// the zstd program both fail or both succeed,
+			// and that if they both fail one byte sequence
+			// is an exact prefix of the other.
+			// Actually trying this proved to be frustrating,
+			// as the zstd program appears to accept invalid
+			// byte sequences using rules that are difficult
+			// to determine.
+			// So we just check the prefix.
+
+			c := len(goExp)
+			if c > len(zstdExp) {
+				c = len(zstdExp)
+			}
+			goExp = goExp[:c]
+			zstdExp = zstdExp[:c]
+			if !bytes.Equal(goExp, zstdExp) {
+				t.Error("byte mismatch after error")
+				t.Logf("Go error: %v\n", goErr)
+				t.Logf("zstd error: %v\n", zstdErr)
+				showDiffs(t, zstdExp, goExp)
+			}
+		}
+	})
+}
diff --git a/src/internal/zstd/huff.go b/src/internal/zstd/huff.go
new file mode 100644
index 0000000..452e24b
--- /dev/null
+++ b/src/internal/zstd/huff.go
@@ -0,0 +1,204 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package zstd
+
+import (
+	"io"
+	"math/bits"
+)
+
+// maxHuffmanBits is the largest possible Huffman table bits.
+const maxHuffmanBits = 11
+
+// readHuff reads Huffman table from data starting at off into table.
+// Each entry in a Huffman table is a pair of bytes.
+// The high byte is the encoded value. The low byte is the number
+// of bits used to encode that value. We index into the table
+// with a value of size tableBits. A value that requires fewer bits
+// appear in the table multiple times.
+// This returns the number of bits in the Huffman table and the new offset.
+// RFC 4.2.1.
+func (r *Reader) readHuff(data block, off int, table []uint16) (tableBits, roff int, err error) {
+	if off >= len(data) {
+		return 0, 0, r.makeEOFError(off)
+	}
+
+	hdr := data[off]
+	off++
+
+	var weights [256]uint8
+	var count int
+	if hdr < 128 {
+		// The table is compressed using an FSE. RFC 4.2.1.2.
+		if len(r.fseScratch) < 1<<6 {
+			r.fseScratch = make([]fseEntry, 1<<6)
+		}
+		fseBits, noff, err := r.readFSE(data, off, 255, 6, r.fseScratch)
+		if err != nil {
+			return 0, 0, err
+		}
+		fseTable := r.fseScratch
+
+		if off+int(hdr) > len(data) {
+			return 0, 0, r.makeEOFError(off)
+		}
+
+		rbr, err := r.makeReverseBitReader(data, off+int(hdr)-1, noff)
+		if err != nil {
+			return 0, 0, err
+		}
+
+		state1, err := rbr.val(uint8(fseBits))
+		if err != nil {
+			return 0, 0, err
+		}
+
+		state2, err := rbr.val(uint8(fseBits))
+		if err != nil {
+			return 0, 0, err
+		}
+
+		// There are two independent FSE streams, tracked by
+		// state1 and state2. We decode them alternately.
+
+		for {
+			pt := &fseTable[state1]
+			if !rbr.fetch(pt.bits) {
+				if count >= 254 {
+					return 0, 0, rbr.makeError("Huffman count overflow")
+				}
+				weights[count] = pt.sym
+				weights[count+1] = fseTable[state2].sym
+				count += 2
+				break
+			}
+
+			v, err := rbr.val(pt.bits)
+			if err != nil {
+				return 0, 0, err
+			}
+			state1 = uint32(pt.base) + v
+
+			if count >= 255 {
+				return 0, 0, rbr.makeError("Huffman count overflow")
+			}
+
+			weights[count] = pt.sym
+			count++
+
+			pt = &fseTable[state2]
+
+			if !rbr.fetch(pt.bits) {
+				if count >= 254 {
+					return 0, 0, rbr.makeError("Huffman count overflow")
+				}
+				weights[count] = pt.sym
+				weights[count+1] = fseTable[state1].sym
+				count += 2
+				break
+			}
+
+			v, err = rbr.val(pt.bits)
+			if err != nil {
+				return 0, 0, err
+			}
+			state2 = uint32(pt.base) + v
+
+			if count >= 255 {
+				return 0, 0, rbr.makeError("Huffman count overflow")
+			}
+
+			weights[count] = pt.sym
+			count++
+		}
+
+		off += int(hdr)
+	} else {
+		// The table is not compressed. Each weight is 4 bits.
+
+		count = int(hdr) - 127
+		if off+((count+1)/2) >= len(data) {
+			return 0, 0, io.ErrUnexpectedEOF
+		}
+		for i := 0; i < count; i += 2 {
+			b := data[off]
+			off++
+			weights[i] = b >> 4
+			weights[i+1] = b & 0xf
+		}
+	}
+
+	// RFC 4.2.1.3.
+
+	var weightMark [13]uint32
+	weightMask := uint32(0)
+	for _, w := range weights[:count] {
+		if w > 12 {
+			return 0, 0, r.makeError(off, "Huffman weight overflow")
+		}
+		weightMark[w]++
+		if w > 0 {
+			weightMask += 1 << (w - 1)
+		}
+	}
+	if weightMask == 0 {
+		return 0, 0, r.makeError(off, "bad Huffman weights")
+	}
+
+	tableBits = 32 - bits.LeadingZeros32(weightMask)
+	if tableBits > maxHuffmanBits {
+		return 0, 0, r.makeError(off, "bad Huffman weights")
+	}
+
+	if len(table) < 1<<tableBits {
+		return 0, 0, r.makeError(off, "Huffman table too small")
+	}
+
+	// Work out the last weight value, which is omitted because
+	// the weights must sum to a power of two.
+	left := (uint32(1) << tableBits) - weightMask
+	if left == 0 {
+		return 0, 0, r.makeError(off, "bad Huffman weights")
+	}
+	highBit := 31 - bits.LeadingZeros32(left)
+	if uint32(1)<<highBit != left {
+		return 0, 0, r.makeError(off, "bad Huffman weights")
+	}
+	if count >= 256 {
+		return 0, 0, r.makeError(off, "Huffman weight overflow")
+	}
+	weights[count] = uint8(highBit + 1)
+	count++
+	weightMark[highBit+1]++
+
+	if weightMark[1] < 2 || weightMark[1]&1 != 0 {
+		return 0, 0, r.makeError(off, "bad Huffman weights")
+	}
+
+	// Change weightMark from a count of weights to the index of
+	// the first symbol for that weight. We shift the indexes to
+	// also store how many we have seen so far,
+	next := uint32(0)
+	for i := 0; i < tableBits; i++ {
+		cur := next
+		next += weightMark[i+1] << i
+		weightMark[i+1] = cur
+	}
+
+	for i, w := range weights[:count] {
+		if w == 0 {
+			continue
+		}
+		length := uint32(1) << (w - 1)
+		tval := uint16(i)<<8 | (uint16(tableBits) + 1 - uint16(w))
+		start := weightMark[w]
+		for j := uint32(0); j < length; j++ {
+			table[start+j] = tval
+		}
+		weightMark[w] += length
+	}
+
+	return tableBits, off, nil
+}
diff --git a/src/internal/zstd/literals.go b/src/internal/zstd/literals.go
new file mode 100644
index 0000000..b46d668
--- /dev/null
+++ b/src/internal/zstd/literals.go
@@ -0,0 +1,330 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package zstd
+
+import (
+	"encoding/binary"
+)
+
+// readLiterals reads and decompresses the literals from data at off.
+// The literals are appended to outbuf, which is returned.
+// Also returns the new input offset. RFC 3.1.1.3.1.
+func (r *Reader) readLiterals(data block, off int, outbuf []byte) (int, []byte, error) {
+	if off >= len(data) {
+		return 0, nil, r.makeEOFError(off)
+	}
+
+	// Literals section header. RFC 3.1.1.3.1.1.
+	hdr := data[off]
+	off++
+
+	if (hdr&3) == 0 || (hdr&3) == 1 {
+		return r.readRawRLELiterals(data, off, hdr, outbuf)
+	} else {
+		return r.readHuffLiterals(data, off, hdr, outbuf)
+	}
+}
+
+// readRawRLELiterals reads and decompresses a Raw_Literals_Block or
+// a RLE_Literals_Block. RFC 3.1.1.3.1.1.
+func (r *Reader) readRawRLELiterals(data block, off int, hdr byte, outbuf []byte) (int, []byte, error) {
+	raw := (hdr & 3) == 0
+
+	var regeneratedSize int
+	switch (hdr >> 2) & 3 {
+	case 0, 2:
+		regeneratedSize = int(hdr >> 3)
+	case 1:
+		if off >= len(data) {
+			return 0, nil, r.makeEOFError(off)
+		}
+		regeneratedSize = int(hdr>>4) + (int(data[off]) << 4)
+		off++
+	case 3:
+		if off+1 >= len(data) {
+			return 0, nil, r.makeEOFError(off)
+		}
+		regeneratedSize = int(hdr>>4) + (int(data[off]) << 4) + (int(data[off+1]) << 12)
+		off += 2
+	}
+
+	// We are going to use the entire literal block in the output.
+	// The maximum size of one decompressed block is 128K,
+	// so we can't have more literals than that.
+	if regeneratedSize > 128<<10 {
+		return 0, nil, r.makeError(off, "literal size too large")
+	}
+
+	if raw {
+		// RFC 3.1.1.3.1.2.
+		if off+regeneratedSize > len(data) {
+			return 0, nil, r.makeError(off, "raw literal size too large")
+		}
+		outbuf = append(outbuf, data[off:off+regeneratedSize]...)
+		off += regeneratedSize
+	} else {
+		// RFC 3.1.1.3.1.3.
+		if off >= len(data) {
+			return 0, nil, r.makeError(off, "RLE literal missing")
+		}
+		rle := data[off]
+		off++
+		for i := 0; i < regeneratedSize; i++ {
+			outbuf = append(outbuf, rle)
+		}
+	}
+
+	return off, outbuf, nil
+}
+
+// readHuffLiterals reads and decompresses a Compressed_Literals_Block or
+// a Treeless_Literals_Block. RFC 3.1.1.3.1.4.
+func (r *Reader) readHuffLiterals(data block, off int, hdr byte, outbuf []byte) (int, []byte, error) {
+	var (
+		regeneratedSize int
+		compressedSize  int
+		streams         int
+	)
+	switch (hdr >> 2) & 3 {
+	case 0, 1:
+		if off+1 >= len(data) {
+			return 0, nil, r.makeEOFError(off)
+		}
+		regeneratedSize = (int(hdr) >> 4) | ((int(data[off]) & 0x3f) << 4)
+		compressedSize = (int(data[off]) >> 6) | (int(data[off+1]) << 2)
+		off += 2
+		if ((hdr >> 2) & 3) == 0 {
+			streams = 1
+		} else {
+			streams = 4
+		}
+	case 2:
+		if off+2 >= len(data) {
+			return 0, nil, r.makeEOFError(off)
+		}
+		regeneratedSize = (int(hdr) >> 4) | (int(data[off]) << 4) | ((int(data[off+1]) & 3) << 12)
+		compressedSize = (int(data[off+1]) >> 2) | (int(data[off+2]) << 6)
+		off += 3
+		streams = 4
+	case 3:
+		if off+3 >= len(data) {
+			return 0, nil, r.makeEOFError(off)
+		}
+		regeneratedSize = (int(hdr) >> 4) | (int(data[off]) << 4) | ((int(data[off+1]) & 0x3f) << 12)
+		compressedSize = (int(data[off+1]) >> 6) | (int(data[off+2]) << 2) | (int(data[off+3]) << 10)
+		off += 4
+		streams = 4
+	}
+
+	// We are going to use the entire literal block in the output.
+	// The maximum size of one decompressed block is 128K,
+	// so we can't have more literals than that.
+	if regeneratedSize > 128<<10 {
+		return 0, nil, r.makeError(off, "literal size too large")
+	}
+
+	roff := off + compressedSize
+	if roff > len(data) || roff < 0 {
+		return 0, nil, r.makeEOFError(off)
+	}
+
+	totalStreamsSize := compressedSize
+	if (hdr & 3) == 2 {
+		// Compressed_Literals_Block.
+		// Read new huffman tree.
+
+		if len(r.huffmanTable) < 1<<maxHuffmanBits {
+			r.huffmanTable = make([]uint16, 1<<maxHuffmanBits)
+		}
+
+		huffmanTableBits, hoff, err := r.readHuff(data, off, r.huffmanTable)
+		if err != nil {
+			return 0, nil, err
+		}
+		r.huffmanTableBits = huffmanTableBits
+
+		if totalStreamsSize < hoff-off {
+			return 0, nil, r.makeError(off, "Huffman table too big")
+		}
+		totalStreamsSize -= hoff - off
+		off = hoff
+	} else {
+		// Treeless_Literals_Block
+		// Reuse previous Huffman tree.
+		if r.huffmanTableBits == 0 {
+			return 0, nil, r.makeError(off, "missing literals Huffman tree")
+		}
+	}
+
+	// Decompress compressedSize bytes of data at off using the
+	// Huffman tree.
+
+	var err error
+	if streams == 1 {
+		outbuf, err = r.readLiteralsOneStream(data, off, totalStreamsSize, regeneratedSize, outbuf)
+	} else {
+		outbuf, err = r.readLiteralsFourStreams(data, off, totalStreamsSize, regeneratedSize, outbuf)
+	}
+
+	if err != nil {
+		return 0, nil, err
+	}
+
+	return roff, outbuf, nil
+}
+
+// readLiteralsOneStream reads a single stream of compressed literals.
+func (r *Reader) readLiteralsOneStream(data block, off, compressedSize, regeneratedSize int, outbuf []byte) ([]byte, error) {
+	// We let the reverse bit reader read earlier bytes,
+	// because the Huffman table ignores bits that it doesn't need.
+	rbr, err := r.makeReverseBitReader(data, off+compressedSize-1, off-2)
+	if err != nil {
+		return nil, err
+	}
+
+	huffTable := r.huffmanTable
+	huffBits := uint32(r.huffmanTableBits)
+	huffMask := (uint32(1) << huffBits) - 1
+
+	for i := 0; i < regeneratedSize; i++ {
+		if !rbr.fetch(uint8(huffBits)) {
+			return nil, rbr.makeError("literals Huffman stream out of bits")
+		}
+
+		var t uint16
+		idx := (rbr.bits >> (rbr.cnt - huffBits)) & huffMask
+		t = huffTable[idx]
+		outbuf = append(outbuf, byte(t>>8))
+		rbr.cnt -= uint32(t & 0xff)
+	}
+
+	return outbuf, nil
+}
+
+// readLiteralsFourStreams reads four interleaved streams of
+// compressed literals.
+func (r *Reader) readLiteralsFourStreams(data block, off, totalStreamsSize, regeneratedSize int, outbuf []byte) ([]byte, error) {
+	// Read the jump table to find out where the streams are.
+	// RFC 3.1.1.3.1.6.
+	if off+5 >= len(data) {
+		return nil, r.makeEOFError(off)
+	}
+	if totalStreamsSize < 6 {
+		return nil, r.makeError(off, "total streams size too small for jump table")
+	}
+
+	streamSize1 := binary.LittleEndian.Uint16(data[off:])
+	streamSize2 := binary.LittleEndian.Uint16(data[off+2:])
+	streamSize3 := binary.LittleEndian.Uint16(data[off+4:])
+	off += 6
+
+	tot := uint64(streamSize1) + uint64(streamSize2) + uint64(streamSize3)
+	if tot > uint64(totalStreamsSize)-6 {
+		return nil, r.makeEOFError(off)
+	}
+	streamSize4 := uint32(totalStreamsSize) - 6 - uint32(tot)
+
+	off--
+	off1 := off + int(streamSize1)
+	start1 := off + 1
+
+	off2 := off1 + int(streamSize2)
+	start2 := off1 + 1
+
+	off3 := off2 + int(streamSize3)
+	start3 := off2 + 1
+
+	off4 := off3 + int(streamSize4)
+	start4 := off3 + 1
+
+	// We let the reverse bit readers read earlier bytes,
+	// because the Huffman tables ignore bits that they don't need.
+
+	rbr1, err := r.makeReverseBitReader(data, off1, start1-2)
+	if err != nil {
+		return nil, err
+	}
+
+	rbr2, err := r.makeReverseBitReader(data, off2, start2-2)
+	if err != nil {
+		return nil, err
+	}
+
+	rbr3, err := r.makeReverseBitReader(data, off3, start3-2)
+	if err != nil {
+		return nil, err
+	}
+
+	rbr4, err := r.makeReverseBitReader(data, off4, start4-2)
+	if err != nil {
+		return nil, err
+	}
+
+	regeneratedStreamSize := (regeneratedSize + 3) / 4
+
+	out1 := len(outbuf)
+	out2 := out1 + regeneratedStreamSize
+	out3 := out2 + regeneratedStreamSize
+	out4 := out3 + regeneratedStreamSize
+
+	regeneratedStreamSize4 := regeneratedSize - regeneratedStreamSize*3
+
+	outbuf = append(outbuf, make([]byte, regeneratedSize)...)
+
+	huffTable := r.huffmanTable
+	huffBits := uint32(r.huffmanTableBits)
+	huffMask := (uint32(1) << huffBits) - 1
+
+	for i := 0; i < regeneratedStreamSize; i++ {
+		use4 := i < regeneratedStreamSize4
+
+		fetchHuff := func(rbr *reverseBitReader) (uint16, error) {
+			if !rbr.fetch(uint8(huffBits)) {
+				return 0, rbr.makeError("literals Huffman stream out of bits")
+			}
+			idx := (rbr.bits >> (rbr.cnt - huffBits)) & huffMask
+			return huffTable[idx], nil
+		}
+
+		t1, err := fetchHuff(&rbr1)
+		if err != nil {
+			return nil, err
+		}
+
+		t2, err := fetchHuff(&rbr2)
+		if err != nil {
+			return nil, err
+		}
+
+		t3, err := fetchHuff(&rbr3)
+		if err != nil {
+			return nil, err
+		}
+
+		if use4 {
+			t4, err := fetchHuff(&rbr4)
+			if err != nil {
+				return nil, err
+			}
+			outbuf[out4] = byte(t4 >> 8)
+			out4++
+			rbr4.cnt -= uint32(t4 & 0xff)
+		}
+
+		outbuf[out1] = byte(t1 >> 8)
+		out1++
+		rbr1.cnt -= uint32(t1 & 0xff)
+
+		outbuf[out2] = byte(t2 >> 8)
+		out2++
+		rbr2.cnt -= uint32(t2 & 0xff)
+
+		outbuf[out3] = byte(t3 >> 8)
+		out3++
+		rbr3.cnt -= uint32(t3 & 0xff)
+	}
+
+	return outbuf, nil
+}
diff --git a/src/internal/zstd/xxhash.go b/src/internal/zstd/xxhash.go
new file mode 100644
index 0000000..4d579ee
--- /dev/null
+++ b/src/internal/zstd/xxhash.go
@@ -0,0 +1,148 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package zstd
+
+import (
+	"encoding/binary"
+	"math/bits"
+)
+
+const (
+	xxhPrime64c1 = 0x9e3779b185ebca87
+	xxhPrime64c2 = 0xc2b2ae3d27d4eb4f
+	xxhPrime64c3 = 0x165667b19e3779f9
+	xxhPrime64c4 = 0x85ebca77c2b2ae63
+	xxhPrime64c5 = 0x27d4eb2f165667c5
+)
+
+// xxhash64 is the state of a xxHash-64 checksum.
+type xxhash64 struct {
+	len uint64    // total length hashed
+	v   [4]uint64 // accumulators
+	buf [32]byte  // buffer
+	cnt int       // number of bytes in buffer
+}
+
+// reset discards the current state and prepares to compute a new hash.
+// We assume a seed of 0 since that is what zstd uses.
+func (xh *xxhash64) reset() {
+	xh.len = 0
+
+	// Separate addition for awkward constant overflow.
+	xh.v[0] = xxhPrime64c1
+	xh.v[0] += xxhPrime64c2
+
+	xh.v[1] = xxhPrime64c2
+	xh.v[2] = 0
+
+	// Separate negation for awkward constant overflow.
+	xh.v[3] = xxhPrime64c1
+	xh.v[3] = -xh.v[3]
+
+	for i := range xh.buf {
+		xh.buf[i] = 0
+	}
+	xh.cnt = 0
+}
+
+// update adds a buffer to the has.
+func (xh *xxhash64) update(b []byte) {
+	xh.len += uint64(len(b))
+
+	if xh.cnt+len(b) < len(xh.buf) {
+		copy(xh.buf[xh.cnt:], b)
+		xh.cnt += len(b)
+		return
+	}
+
+	if xh.cnt > 0 {
+		n := copy(xh.buf[xh.cnt:], b)
+		b = b[n:]
+		xh.v[0] = xh.round(xh.v[0], binary.LittleEndian.Uint64(xh.buf[:]))
+		xh.v[1] = xh.round(xh.v[1], binary.LittleEndian.Uint64(xh.buf[8:]))
+		xh.v[2] = xh.round(xh.v[2], binary.LittleEndian.Uint64(xh.buf[16:]))
+		xh.v[3] = xh.round(xh.v[3], binary.LittleEndian.Uint64(xh.buf[24:]))
+		xh.cnt = 0
+	}
+
+	for len(b) >= 32 {
+		xh.v[0] = xh.round(xh.v[0], binary.LittleEndian.Uint64(b))
+		xh.v[1] = xh.round(xh.v[1], binary.LittleEndian.Uint64(b[8:]))
+		xh.v[2] = xh.round(xh.v[2], binary.LittleEndian.Uint64(b[16:]))
+		xh.v[3] = xh.round(xh.v[3], binary.LittleEndian.Uint64(b[24:]))
+		b = b[32:]
+	}
+
+	if len(b) > 0 {
+		copy(xh.buf[:], b)
+		xh.cnt = len(b)
+	}
+}
+
+// digest returns the final hash value.
+func (xh *xxhash64) digest() uint64 {
+	var h64 uint64
+	if xh.len < 32 {
+		h64 = xh.v[2] + xxhPrime64c5
+	} else {
+		h64 = bits.RotateLeft64(xh.v[0], 1) +
+			bits.RotateLeft64(xh.v[1], 7) +
+			bits.RotateLeft64(xh.v[2], 12) +
+			bits.RotateLeft64(xh.v[3], 18)
+		h64 = xh.mergeRound(h64, xh.v[0])
+		h64 = xh.mergeRound(h64, xh.v[1])
+		h64 = xh.mergeRound(h64, xh.v[2])
+		h64 = xh.mergeRound(h64, xh.v[3])
+	}
+
+	h64 += xh.len
+
+	len := xh.len
+	len &= 31
+	buf := xh.buf[:]
+	for len >= 8 {
+		k1 := xh.round(0, binary.LittleEndian.Uint64(buf))
+		buf = buf[8:]
+		h64 ^= k1
+		h64 = bits.RotateLeft64(h64, 27)*xxhPrime64c1 + xxhPrime64c4
+		len -= 8
+	}
+	if len >= 4 {
+		h64 ^= uint64(binary.LittleEndian.Uint32(buf)) * xxhPrime64c1
+		buf = buf[4:]
+		h64 = bits.RotateLeft64(h64, 23)*xxhPrime64c2 + xxhPrime64c3
+		len -= 4
+	}
+	for len > 0 {
+		h64 ^= uint64(buf[0]) * xxhPrime64c5
+		buf = buf[1:]
+		h64 = bits.RotateLeft64(h64, 11) * xxhPrime64c1
+		len--
+	}
+
+	h64 ^= h64 >> 33
+	h64 *= xxhPrime64c2
+	h64 ^= h64 >> 29
+	h64 *= xxhPrime64c3
+	h64 ^= h64 >> 32
+
+	return h64
+}
+
+// round updates a value.
+func (xh *xxhash64) round(v, n uint64) uint64 {
+	v += n * xxhPrime64c2
+	v = bits.RotateLeft64(v, 31)
+	v *= xxhPrime64c1
+	return v
+}
+
+// mergeRound updates a value in the final round.
+func (xh *xxhash64) mergeRound(v, n uint64) uint64 {
+	n = xh.round(0, n)
+	v ^= n
+	v = v*xxhPrime64c1 + xxhPrime64c4
+	return v
+}
diff --git a/src/internal/zstd/xxhash_test.go b/src/internal/zstd/xxhash_test.go
new file mode 100644
index 0000000..646cee8
--- /dev/null
+++ b/src/internal/zstd/xxhash_test.go
@@ -0,0 +1,105 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package zstd
+
+import (
+	"bytes"
+	"os"
+	"os/exec"
+	"strconv"
+	"testing"
+)
+
+var xxHashTests = []struct {
+	data string
+	hash uint64
+}{
+	{
+		"hello, world",
+		0xb33a384e6d1b1242,
+	},
+	{
+		"abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789$",
+		0x1032d841e824f998,
+	},
+}
+
+func TestXXHash(t *testing.T) {
+	var xh xxhash64
+	for i, test := range xxHashTests {
+		xh.reset()
+		xh.update([]byte(test.data))
+		if got := xh.digest(); got != test.hash {
+			t.Errorf("#%d: got %#x want %#x", i, got, test.hash)
+		}
+	}
+}
+
+func TestLargeXXHash(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping expensive test in short mode")
+	}
+
+	data := bigData(t)
+	var xh xxhash64
+	xh.reset()
+	i := 0
+	for i < len(data) {
+		// Write varying amounts to test buffering.
+		c := i%4094 + 1
+		if i+c > len(data) {
+			c = len(data) - i
+		}
+		xh.update(data[i : i+c])
+		i += c
+	}
+
+	got := xh.digest()
+	want := uint64(0xf0dd39fd7e063f82)
+	if got != want {
+		t.Errorf("got %#x want %#x", got, want)
+	}
+}
+
+func FuzzXXHash(f *testing.F) {
+	if _, err := os.Stat("/usr/bin/xxhsum"); err != nil {
+		f.Skip("skipping because /usr/bin/xxhsum does not exist")
+	}
+
+	for _, test := range xxHashTests {
+		f.Add([]byte(test.data))
+	}
+	f.Add(bytes.Repeat([]byte("abcdefghijklmnop"), 256))
+	var buf bytes.Buffer
+	for i := 0; i < 256; i++ {
+		buf.WriteByte(byte(i))
+	}
+	f.Add(bytes.Repeat(buf.Bytes(), 64))
+	f.Add(bigData(f))
+
+	f.Fuzz(func(t *testing.T, b []byte) {
+		cmd := exec.Command("/usr/bin/xxhsum", "-H64")
+		cmd.Stdin = bytes.NewReader(b)
+		var hhsumHash bytes.Buffer
+		cmd.Stdout = &hhsumHash
+		if err := cmd.Run(); err != nil {
+			t.Fatalf("running hhsum failed: %v", err)
+		}
+		hhHashBytes := bytes.Fields(bytes.TrimSpace(hhsumHash.Bytes()))[0]
+		hhHash, err := strconv.ParseUint(string(hhHashBytes), 16, 64)
+		if err != nil {
+			t.Fatalf("could not parse hash %q: %v", hhHashBytes, err)
+		}
+
+		var xh xxhash64
+		xh.reset()
+		xh.update(b)
+		goHash := xh.digest()
+
+		if goHash != hhHash {
+			t.Errorf("Go hash %#x != xxhsum hash %#x", goHash, hhHash)
+		}
+	})
+}
diff --git a/src/internal/zstd/zstd.go b/src/internal/zstd/zstd.go
new file mode 100644
index 0000000..a860789
--- /dev/null
+++ b/src/internal/zstd/zstd.go
@@ -0,0 +1,508 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package zstd provides a decompressor for zstd streams,
+// described in RFC 8878. It does not support dictionaries.
+package zstd
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+)
+
+// fuzzing is a fuzzer hook set to true when fuzzing.
+// This is used to reject cases where we don't match zstd.
+var fuzzing = false
+
+// Reader implements [io.Reader] to read a zstd compressed stream.
+type Reader struct {
+	// The underlying Reader.
+	r io.Reader
+
+	// Whether we have read the frame header.
+	// This is of interest when buffer is empty.
+	// If true we expect to see a new block.
+	sawFrameHeader bool
+
+	// Whether the current frame expects a checksum.
+	hasChecksum bool
+
+	// Whether we have read at least one frame.
+	readOneFrame bool
+
+	// True if the frame size is not known.
+	frameSizeUnknown bool
+
+	// The number of uncompressed bytes remaining in the current frame.
+	// If frameSizeUnknown is true, this is not valid.
+	remainingFrameSize uint64
+
+	// The number of bytes read from r up to the start of the current
+	// block, for error reporting.
+	blockOffset int64
+
+	// Buffered decompressed data.
+	buffer []byte
+	// Current read offset in buffer.
+	off int
+
+	// The current repeated offsets.
+	repeatedOffset1 uint32
+	repeatedOffset2 uint32
+	repeatedOffset3 uint32
+
+	// The current Huffman tree used for compressing literals.
+	huffmanTable     []uint16
+	huffmanTableBits int
+
+	// The window for back references.
+	windowSize int    // maximum required window size
+	window     []byte // window data
+
+	// A buffer available to hold a compressed block.
+	compressedBuf []byte
+
+	// A buffer for literals.
+	literals []byte
+
+	// Sequence decode FSE tables.
+	seqTables    [3][]fseBaselineEntry
+	seqTableBits [3]uint8
+
+	// Buffers for sequence decode FSE tables.
+	seqTableBuffers [3][]fseBaselineEntry
+
+	// Scratch space used for small reads, to avoid allocation.
+	scratch [16]byte
+
+	// A scratch table for reading an FSE. Only temporarily valid.
+	fseScratch []fseEntry
+
+	// For checksum computation.
+	checksum xxhash64
+}
+
+// NewReader creates a new Reader that decompresses data from the given reader.
+func NewReader(input io.Reader) *Reader {
+	r := new(Reader)
+	r.Reset(input)
+	return r
+}
+
+// Reset discards the current state and starts reading a new stream from r.
+// This permits reusing a Reader rather than allocating a new one.
+func (r *Reader) Reset(input io.Reader) {
+	r.r = input
+
+	// Several fields are preserved to avoid allocation.
+	// Others are always set before they are used.
+	r.sawFrameHeader = false
+	r.hasChecksum = false
+	r.readOneFrame = false
+	r.frameSizeUnknown = false
+	r.remainingFrameSize = 0
+	r.blockOffset = 0
+	// buffer
+	r.off = 0
+	// repeatedOffset1
+	// repeatedOffset2
+	// repeatedOffset3
+	// huffmanTable
+	// huffmanTableBits
+	// windowSize
+	// window
+	// compressedBuf
+	// literals
+	// seqTables
+	// seqTableBits
+	// seqTableBuffers
+	// scratch
+	// fseScratch
+}
+
+// Read implements [io.Reader].
+func (r *Reader) Read(p []byte) (int, error) {
+	if err := r.refillIfNeeded(); err != nil {
+		return 0, err
+	}
+	n := copy(p, r.buffer[r.off:])
+	r.off += n
+	return n, nil
+}
+
+// ReadByte implements [io.ByteReader].
+func (r *Reader) ReadByte() (byte, error) {
+	if err := r.refillIfNeeded(); err != nil {
+		return 0, err
+	}
+	ret := r.buffer[r.off]
+	r.off++
+	return ret, nil
+}
+
+// refillIfNeeded reads the next block if necessary.
+func (r *Reader) refillIfNeeded() error {
+	for r.off >= len(r.buffer) {
+		if err := r.refill(); err != nil {
+			return err
+		}
+		r.off = 0
+	}
+	return nil
+}
+
+// refill reads and decompresses the next block.
+func (r *Reader) refill() error {
+	if !r.sawFrameHeader {
+		if err := r.readFrameHeader(); err != nil {
+			return err
+		}
+	}
+	return r.readBlock()
+}
+
+// readFrameHeader reads the frame header and prepares to read a block.
+func (r *Reader) readFrameHeader() error {
+retry:
+	relativeOffset := 0
+
+	// Read magic number. RFC 3.1.1.
+	if _, err := io.ReadFull(r.r, r.scratch[:4]); err != nil {
+		// We require that the stream contain at least one frame.
+		if err == io.EOF && !r.readOneFrame {
+			err = io.ErrUnexpectedEOF
+		}
+		return r.wrapError(relativeOffset, err)
+	}
+
+	if magic := binary.LittleEndian.Uint32(r.scratch[:4]); magic != 0xfd2fb528 {
+		if magic >= 0x184d2a50 && magic <= 0x184d2a5f {
+			// This is a skippable frame.
+			r.blockOffset += int64(relativeOffset) + 4
+			if err := r.skipFrame(); err != nil {
+				return err
+			}
+			goto retry
+		}
+
+		return r.makeError(relativeOffset, "invalid magic number")
+	}
+
+	relativeOffset += 4
+
+	// Read Frame_Header_Descriptor. RFC 3.1.1.1.1.
+	if _, err := io.ReadFull(r.r, r.scratch[:1]); err != nil {
+		return r.wrapNonEOFError(relativeOffset, err)
+	}
+	descriptor := r.scratch[0]
+
+	singleSegment := descriptor&(1<<5) != 0
+
+	fcsFieldSize := 1 << (descriptor >> 6)
+	if fcsFieldSize == 1 && !singleSegment {
+		fcsFieldSize = 0
+	}
+
+	var windowDescriptorSize int
+	if singleSegment {
+		windowDescriptorSize = 0
+	} else {
+		windowDescriptorSize = 1
+	}
+
+	if descriptor&(1<<3) != 0 {
+		return r.makeError(relativeOffset, "reserved bit set in frame header descriptor")
+	}
+
+	r.hasChecksum = descriptor&(1<<2) != 0
+	if r.hasChecksum {
+		r.checksum.reset()
+	}
+
+	if descriptor&3 != 0 {
+		return r.makeError(relativeOffset, "dictionaries are not supported")
+	}
+
+	relativeOffset++
+
+	headerSize := windowDescriptorSize + fcsFieldSize
+
+	if _, err := io.ReadFull(r.r, r.scratch[:headerSize]); err != nil {
+		return r.wrapNonEOFError(relativeOffset, err)
+	}
+
+	// Figure out the maximum amount of data we need to retain
+	// for backreferences.
+
+	if singleSegment {
+		// No window required, as all the data is in a single buffer.
+		r.windowSize = 0
+	} else {
+		// Window descriptor. RFC 3.1.1.1.2.
+		windowDescriptor := r.scratch[0]
+		exponent := uint64(windowDescriptor >> 3)
+		mantissa := uint64(windowDescriptor & 7)
+		windowLog := exponent + 10
+		windowBase := uint64(1) << windowLog
+		windowAdd := (windowBase / 8) * mantissa
+		windowSize := windowBase + windowAdd
+
+		// Default zstd sets limits on the window size.
+		if fuzzing && (windowLog > 31 || windowSize > 1<<27) {
+			return r.makeError(relativeOffset, "windowSize too large")
+		}
+
+		// RFC 8878 permits us to set an 8M max on window size.
+		if windowSize > 8<<20 {
+			windowSize = 8 << 20
+		}
+
+		r.windowSize = int(windowSize)
+	}
+
+	// Frame_Content_Size. RFC 3.1.1.4.
+	r.frameSizeUnknown = false
+	r.remainingFrameSize = 0
+	fb := r.scratch[windowDescriptorSize:]
+	switch fcsFieldSize {
+	case 0:
+		r.frameSizeUnknown = true
+	case 1:
+		r.remainingFrameSize = uint64(fb[0])
+	case 2:
+		r.remainingFrameSize = 256 + uint64(binary.LittleEndian.Uint16(fb))
+	case 4:
+		r.remainingFrameSize = uint64(binary.LittleEndian.Uint32(fb))
+	case 8:
+		r.remainingFrameSize = binary.LittleEndian.Uint64(fb)
+	default:
+		panic("unreachable")
+	}
+
+	relativeOffset += headerSize
+
+	r.sawFrameHeader = true
+	r.readOneFrame = true
+	r.blockOffset += int64(relativeOffset)
+
+	// Prepare to read blocks from the frame.
+	r.repeatedOffset1 = 1
+	r.repeatedOffset2 = 4
+	r.repeatedOffset3 = 8
+	r.huffmanTableBits = 0
+	r.window = r.window[:0]
+	r.seqTables[0] = nil
+	r.seqTables[1] = nil
+	r.seqTables[2] = nil
+
+	return nil
+}
+
+// skipFrame skips a skippable frame. RFC 3.1.2.
+func (r *Reader) skipFrame() error {
+	relativeOffset := 0
+
+	if _, err := io.ReadFull(r.r, r.scratch[:4]); err != nil {
+		return r.wrapNonEOFError(relativeOffset, err)
+	}
+
+	relativeOffset += 4
+
+	size := binary.LittleEndian.Uint32(r.scratch[:4])
+
+	if seeker, ok := r.r.(io.Seeker); ok {
+		if _, err := seeker.Seek(int64(size), io.SeekCurrent); err != nil {
+			return err
+		}
+		r.blockOffset += int64(relativeOffset) + int64(size)
+		return nil
+	}
+
+	var skip []byte
+	const chunk = 1 << 20 // 1M
+	for size >= chunk {
+		if len(skip) == 0 {
+			skip = make([]byte, chunk)
+		}
+		if _, err := io.ReadFull(r.r, skip); err != nil {
+			return r.wrapNonEOFError(relativeOffset, err)
+		}
+		relativeOffset += chunk
+		size -= chunk
+	}
+	if size > 0 {
+		if len(skip) == 0 {
+			skip = make([]byte, size)
+		}
+		if _, err := io.ReadFull(r.r, skip); err != nil {
+			return r.wrapNonEOFError(relativeOffset, err)
+		}
+		relativeOffset += int(size)
+	}
+
+	r.blockOffset += int64(relativeOffset)
+
+	return nil
+}
+
+// readBlock reads the next block from a frame.
+func (r *Reader) readBlock() error {
+	relativeOffset := 0
+
+	// Read Block_Header. RFC 3.1.1.2.
+	if _, err := io.ReadFull(r.r, r.scratch[:3]); err != nil {
+		return r.wrapNonEOFError(relativeOffset, err)
+	}
+
+	relativeOffset += 3
+
+	header := uint32(r.scratch[0]) | (uint32(r.scratch[1]) << 8) | (uint32(r.scratch[2]) << 16)
+
+	lastBlock := header&1 != 0
+	blockType := (header >> 1) & 3
+	blockSize := int(header >> 3)
+
+	// Maximum block size is smaller of window size and 128K.
+	// We don't record the window size for a single segment frame,
+	// so just use 128K. RFC 3.1.1.2.3, 3.1.1.2.4.
+	if blockSize > 128<<10 || (r.windowSize > 0 && blockSize > r.windowSize) {
+		return r.makeError(relativeOffset, "block size too large")
+	}
+
+	// Handle different block types. RFC 3.1.1.2.2.
+	switch blockType {
+	case 0:
+		r.setBufferSize(blockSize)
+		if _, err := io.ReadFull(r.r, r.buffer); err != nil {
+			return r.wrapNonEOFError(relativeOffset, err)
+		}
+		relativeOffset += blockSize
+		r.blockOffset += int64(relativeOffset)
+	case 1:
+		r.setBufferSize(blockSize)
+		if _, err := io.ReadFull(r.r, r.scratch[:1]); err != nil {
+			return r.wrapNonEOFError(relativeOffset, err)
+		}
+		relativeOffset++
+		v := r.scratch[0]
+		for i := range r.buffer {
+			r.buffer[i] = v
+		}
+		r.blockOffset += int64(relativeOffset)
+	case 2:
+		r.blockOffset += int64(relativeOffset)
+		if err := r.compressedBlock(blockSize); err != nil {
+			return err
+		}
+		r.blockOffset += int64(blockSize)
+	case 3:
+		return r.makeError(relativeOffset, "invalid block type")
+	}
+
+	if !r.frameSizeUnknown {
+		if uint64(len(r.buffer)) > r.remainingFrameSize {
+			return r.makeError(relativeOffset, "too many uncompressed bytes in frame")
+		}
+		r.remainingFrameSize -= uint64(len(r.buffer))
+	}
+
+	if r.hasChecksum {
+		r.checksum.update(r.buffer)
+	}
+
+	if !lastBlock {
+		r.saveWindow(r.buffer)
+	} else {
+		if !r.frameSizeUnknown && r.remainingFrameSize != 0 {
+			return r.makeError(relativeOffset, "not enough uncompressed bytes for frame")
+		}
+		// Check for checksum at end of frame. RFC 3.1.1.
+		if r.hasChecksum {
+			if _, err := io.ReadFull(r.r, r.scratch[:4]); err != nil {
+				return r.wrapNonEOFError(0, err)
+			}
+
+			inputChecksum := binary.LittleEndian.Uint32(r.scratch[:4])
+			dataChecksum := uint32(r.checksum.digest())
+			if inputChecksum != dataChecksum {
+				return r.wrapError(0, fmt.Errorf("invalid checksum: got %#x want %#x", dataChecksum, inputChecksum))
+			}
+
+			r.blockOffset += 4
+		}
+		r.sawFrameHeader = false
+	}
+
+	return nil
+}
+
+// setBufferSize sets the decompressed buffer size.
+// When this is called the buffer is empty.
+func (r *Reader) setBufferSize(size int) {
+	if cap(r.buffer) < size {
+		need := size - cap(r.buffer)
+		r.buffer = append(r.buffer[:cap(r.buffer)], make([]byte, need)...)
+	}
+	r.buffer = r.buffer[:size]
+}
+
+// saveWindow saves bytes in the backreference window.
+// TODO: use a circular buffer for less data movement.
+func (r *Reader) saveWindow(buf []byte) {
+	if r.windowSize == 0 {
+		return
+	}
+
+	if len(buf) >= r.windowSize {
+		from := len(buf) - r.windowSize
+		r.window = append(r.window[:0], buf[from:]...)
+		return
+	}
+
+	keep := r.windowSize - len(buf) // must be positive
+	if keep < len(r.window) {
+		remove := len(r.window) - keep
+		copy(r.window[:], r.window[remove:])
+	}
+
+	r.window = append(r.window, buf...)
+}
+
+// zstdError is an error while decompressing.
+type zstdError struct {
+	offset int64
+	err    error
+}
+
+func (ze *zstdError) Error() string {
+	return fmt.Sprintf("zstd decompression error at %d: %v", ze.offset, ze.err)
+}
+
+func (ze *zstdError) Unwrap() error {
+	return ze.err
+}
+
+func (r *Reader) makeEOFError(off int) error {
+	return r.wrapError(off, io.ErrUnexpectedEOF)
+}
+
+func (r *Reader) wrapNonEOFError(off int, err error) error {
+	if err == io.EOF {
+		err = io.ErrUnexpectedEOF
+	}
+	return r.wrapError(off, err)
+}
+
+func (r *Reader) makeError(off int, msg string) error {
+	return r.wrapError(off, errors.New(msg))
+}
+
+func (r *Reader) wrapError(off int, err error) error {
+	if err == io.EOF {
+		return err
+	}
+	return &zstdError{r.blockOffset + int64(off), err}
+}
diff --git a/src/internal/zstd/zstd_test.go b/src/internal/zstd/zstd_test.go
new file mode 100644
index 0000000..bc75e0f
--- /dev/null
+++ b/src/internal/zstd/zstd_test.go
@@ -0,0 +1,249 @@
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package zstd
+
+import (
+	"bytes"
+	"fmt"
+	"internal/race"
+	"internal/testenv"
+	"io"
+	"os"
+	"os/exec"
+	"strings"
+	"sync"
+	"testing"
+)
+
+// tests holds some simple test cases, including some found by fuzzing.
+var tests = []struct {
+	name, uncompressed, compressed string
+}{
+	{
+		"hello",
+		"hello, world\n",
+		"\x28\xb5\x2f\xfd\x24\x0d\x69\x00\x00\x68\x65\x6c\x6c\x6f\x2c\x20\x77\x6f\x72\x6c\x64\x0a\x4c\x1f\xf9\xf1",
+	},
+	{
+		// a small compressed .debug_ranges section.
+		"ranges",
+		"\xcc\x11\x00\x00\x00\x00\x00\x00\xd5\x13\x00\x00\x00\x00\x00\x00" +
+			"\x1c\x14\x00\x00\x00\x00\x00\x00\x72\x14\x00\x00\x00\x00\x00\x00" +
+			"\x9d\x14\x00\x00\x00\x00\x00\x00\xd5\x14\x00\x00\x00\x00\x00\x00" +
+			"\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00" +
+			"\xfb\x12\x00\x00\x00\x00\x00\x00\x09\x13\x00\x00\x00\x00\x00\x00" +
+			"\x0c\x13\x00\x00\x00\x00\x00\x00\xcb\x13\x00\x00\x00\x00\x00\x00" +
+			"\x29\x14\x00\x00\x00\x00\x00\x00\x4e\x14\x00\x00\x00\x00\x00\x00" +
+			"\x9d\x14\x00\x00\x00\x00\x00\x00\xd5\x14\x00\x00\x00\x00\x00\x00" +
+			"\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00" +
+			"\xfb\x12\x00\x00\x00\x00\x00\x00\x09\x13\x00\x00\x00\x00\x00\x00" +
+			"\x67\x13\x00\x00\x00\x00\x00\x00\xcb\x13\x00\x00\x00\x00\x00\x00" +
+			"\x9d\x14\x00\x00\x00\x00\x00\x00\xd5\x14\x00\x00\x00\x00\x00\x00" +
+			"\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00" +
+			"\x5f\x0b\x00\x00\x00\x00\x00\x00\x6c\x0b\x00\x00\x00\x00\x00\x00" +
+			"\x7d\x0b\x00\x00\x00\x00\x00\x00\x7e\x0c\x00\x00\x00\x00\x00\x00" +
+			"\x38\x0f\x00\x00\x00\x00\x00\x00\x5c\x0f\x00\x00\x00\x00\x00\x00" +
+			"\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00" +
+			"\x83\x0c\x00\x00\x00\x00\x00\x00\xfa\x0c\x00\x00\x00\x00\x00\x00" +
+			"\xfd\x0d\x00\x00\x00\x00\x00\x00\xef\x0e\x00\x00\x00\x00\x00\x00" +
+			"\x14\x0f\x00\x00\x00\x00\x00\x00\x38\x0f\x00\x00\x00\x00\x00\x00" +
+			"\x9f\x0f\x00\x00\x00\x00\x00\x00\xac\x0f\x00\x00\x00\x00\x00\x00" +
+			"\xdb\x0f\x00\x00\x00\x00\x00\x00\xff\x0f\x00\x00\x00\x00\x00\x00" +
+			"\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00" +
+			"\xfd\x0d\x00\x00\x00\x00\x00\x00\xd8\x0e\x00\x00\x00\x00\x00\x00" +
+			"\x9f\x0f\x00\x00\x00\x00\x00\x00\xac\x0f\x00\x00\x00\x00\x00\x00" +
+			"\xdb\x0f\x00\x00\x00\x00\x00\x00\xff\x0f\x00\x00\x00\x00\x00\x00" +
+			"\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00" +
+			"\xfa\x0c\x00\x00\x00\x00\x00\x00\xea\x0d\x00\x00\x00\x00\x00\x00" +
+			"\xef\x0e\x00\x00\x00\x00\x00\x00\x14\x0f\x00\x00\x00\x00\x00\x00" +
+			"\x5c\x0f\x00\x00\x00\x00\x00\x00\x9f\x0f\x00\x00\x00\x00\x00\x00" +
+			"\xac\x0f\x00\x00\x00\x00\x00\x00\xdb\x0f\x00\x00\x00\x00\x00\x00" +
+			"\xff\x0f\x00\x00\x00\x00\x00\x00\x2c\x10\x00\x00\x00\x00\x00\x00" +
+			"\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00" +
+			"\x60\x11\x00\x00\x00\x00\x00\x00\xd1\x16\x00\x00\x00\x00\x00\x00" +
+			"\x40\x0b\x00\x00\x00\x00\x00\x00\x2c\x10\x00\x00\x00\x00\x00\x00" +
+			"\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00" +
+			"\x7a\x00\x00\x00\x00\x00\x00\x00\xb6\x00\x00\x00\x00\x00\x00\x00" +
+			"\x9f\x01\x00\x00\x00\x00\x00\x00\xa7\x01\x00\x00\x00\x00\x00\x00" +
+			"\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00" +
+			"\x7a\x00\x00\x00\x00\x00\x00\x00\xa9\x00\x00\x00\x00\x00\x00\x00" +
+			"\x9f\x01\x00\x00\x00\x00\x00\x00\xa7\x01\x00\x00\x00\x00\x00\x00" +
+			"\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00",
+
+		"\x28\xb5\x2f\xfd\x64\xa0\x01\x2d\x05\x00\xc4\x04\xcc\x11\x00\xd5" +
+			"\x13\x00\x1c\x14\x00\x72\x9d\xd5\xfb\x12\x00\x09\x0c\x13\xcb\x13" +
+			"\x29\x4e\x67\x5f\x0b\x6c\x0b\x7d\x0b\x7e\x0c\x38\x0f\x5c\x0f\x83" +
+			"\x0c\xfa\x0c\xfd\x0d\xef\x0e\x14\x38\x9f\x0f\xac\x0f\xdb\x0f\xff" +
+			"\x0f\xd8\x9f\xac\xdb\xff\xea\x5c\x2c\x10\x60\xd1\x16\x40\x0b\x7a" +
+			"\x00\xb6\x00\x9f\x01\xa7\x01\xa9\x36\x20\xa0\x83\x14\x34\x63\x4a" +
+			"\x21\x70\x8c\x07\x46\x03\x4e\x10\x62\x3c\x06\x4e\xc8\x8c\xb0\x32" +
+			"\x2a\x59\xad\xb2\xf1\x02\x82\x7c\x33\xcb\x92\x6f\x32\x4f\x9b\xb0" +
+			"\xa2\x30\xf0\xc0\x06\x1e\x98\x99\x2c\x06\x1e\xd8\xc0\x03\x56\xd8" +
+			"\xc0\x03\x0f\x6c\xe0\x01\xf1\xf0\xee\x9a\xc6\xc8\x97\x99\xd1\x6c" +
+			"\xb4\x21\x45\x3b\x10\xe4\x7b\x99\x4d\x8a\x36\x64\x5c\x77\x08\x02" +
+			"\xcb\xe0\xce",
+	},
+	{
+		"fuzz1",
+		"0\x00\x00\x00\x00\x000\x00\x00\x00\x00\x001\x00\x00\x00\x00\x000000",
+		"(\xb5/\xfd\x04X\x8d\x00\x00P0\x000\x001\x000000\x03T\x02\x00\x01\x01m\xf9\xb7G",
+	},
+}
+
+func TestSamples(t *testing.T) {
+	for _, test := range tests {
+		test := test
+		t.Run(test.name, func(t *testing.T) {
+			r := NewReader(strings.NewReader(test.compressed))
+			got, err := io.ReadAll(r)
+			if err != nil {
+				t.Fatal(err)
+			}
+			gotstr := string(got)
+			if gotstr != test.uncompressed {
+				t.Errorf("got %q want %q", gotstr, test.uncompressed)
+			}
+		})
+	}
+}
+
+var (
+	bigDataOnce  sync.Once
+	bigDataBytes []byte
+	bigDataErr   error
+)
+
+// bigData returns the contents of our large test file.
+func bigData(t testing.TB) []byte {
+	bigDataOnce.Do(func() {
+		bigDataBytes, bigDataErr = os.ReadFile("../../testdata/Isaac.Newton-Opticks.txt")
+	})
+	if bigDataErr != nil {
+		t.Fatal(bigDataErr)
+	}
+	return bigDataBytes
+}
+
+var (
+	zstdBigOnce  sync.Once
+	zstdBigBytes []byte
+	zstdBigSkip  bool
+	zstdBigErr   error
+)
+
+// zstdBigData returns the compressed contents of our large test file.
+// This will only run on Unix systems with zstd installed.
+// That's OK as the package is GOOS-independent.
+func zstdBigData(t testing.TB) []byte {
+	input := bigData(t)
+
+	zstdBigOnce.Do(func() {
+		if _, err := os.Stat("/usr/bin/zstd"); err != nil {
+			zstdBigSkip = true
+			return
+		}
+
+		cmd := exec.Command("/usr/bin/zstd", "-z")
+		cmd.Stdin = bytes.NewReader(input)
+		var compressed bytes.Buffer
+		cmd.Stdout = &compressed
+		cmd.Stderr = os.Stderr
+		if err := cmd.Run(); err != nil {
+			zstdBigErr = fmt.Errorf("running zstd failed: %v", err)
+			return
+		}
+
+		zstdBigBytes = compressed.Bytes()
+	})
+	if zstdBigSkip {
+		t.Skip("skipping because /usr/bin/zstd does not exist")
+	}
+	if zstdBigErr != nil {
+		t.Fatal(zstdBigErr)
+	}
+	return zstdBigBytes
+}
+
+// Test decompressing a large file. We don't have a compressor,
+// so this test only runs on systems with zstd installed.
+func TestLarge(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping expensive test in short mode")
+	}
+
+	data := bigData(t)
+	compressed := zstdBigData(t)
+
+	t.Logf("/usr/bin/zstd compressed %d bytes to %d", len(data), len(compressed))
+
+	r := NewReader(bytes.NewReader(compressed))
+	got, err := io.ReadAll(r)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if !bytes.Equal(got, data) {
+		showDiffs(t, got, data)
+	}
+}
+
+// showDiffs reports the first few differences in two []byte.
+func showDiffs(t *testing.T, got, want []byte) {
+	t.Error("data mismatch")
+	if len(got) != len(want) {
+		t.Errorf("got data length %d, want %d", len(got), len(want))
+	}
+	diffs := 0
+	for i, b := range got {
+		if i >= len(want) {
+			break
+		}
+		if b != want[i] {
+			diffs++
+			if diffs > 20 {
+				break
+			}
+			t.Logf("%d: %#x != %#x", i, b, want[i])
+		}
+	}
+}
+
+func TestAlloc(t *testing.T) {
+	testenv.SkipIfOptimizationOff(t)
+	if race.Enabled {
+		t.Skip("skipping allocation test under race detector")
+	}
+
+	compressed := zstdBigData(t)
+	input := bytes.NewReader(compressed)
+	r := NewReader(input)
+	c := testing.AllocsPerRun(10, func() {
+		input.Reset(compressed)
+		r.Reset(input)
+		io.Copy(io.Discard, r)
+	})
+	if c != 0 {
+		t.Errorf("got %v allocs, want 0", c)
+	}
+}
+
+func BenchmarkLarge(b *testing.B) {
+	b.StopTimer()
+	b.ReportAllocs()
+
+	compressed := zstdBigData(b)
+
+	b.SetBytes(int64(len(compressed)))
+
+	input := bytes.NewReader(compressed)
+	r := NewReader(input)
+
+	b.StartTimer()
+	for i := 0; i < b.N; i++ {
+		input.Reset(compressed)
+		r.Reset(input)
+		io.Copy(io.Discard, r)
+	}
+}