cmd/compile/internal/syntax: faster and simpler source reader

This is one of several changes that were part of a larger rewrite
which I made in early 2019 after switching to the new number literal
syntax implementation. The purpose of the rewrite was to simplify
reading of source code (Unicode character by character) and speed up
the scanner but was never submitted for review due to other priorities.

Part 3 of 3:

This change contains a complete rewrite of source.go, the file that
implements reading individual Unicode characters from the source.
The new implementation is easier to use and has simpler literal
buffer management, resulting in faster scanner and thus parser
performance.

Thew new source.go (internal) API is centered around nextch() which
advances the scanner by one character. The scanner has been adjusted
around nextch() and now consistently does one character look-ahead
(there's no need for complicated ungetr-ing anymore). Only in one
case backtrack is needed (when finding '..' rather than '...') and
that case is now more cleanly solved with the new reset() function.

Measuring line/s parsing peformance by running

go test -run StdLib -fast -skip "syntax/(scanner|source)\.go"

(best of 5 runs on "quiet" MacBook Pro, 3.3GHz Dual-Core i7, 16GB RAM,
OS X 10.15.3) before and after shows consistently 3-5% improvement of
line parsing speed:

old: parsed 1788155 lines (3969 files) in 1.255520307s (1424234 lines/s)
new: parsed 1788155 lines (3969 files) in 1.213197037s (1473919 lines/s)

(scanner.go and parser.go are skipped because this CL changed those files.)

Change-Id: Ida947f4b538d42eb2d2349062c69edb6c9e5ca66
Reviewed-on: https://go-review.googlesource.com/c/go/+/221603
Reviewed-by: Matthew Dempsky <mdempsky@google.com>

1 files changed, 143 insertions, 144 deletions

diff --git a/src/cmd/compile/internal/syntax/source.go b/src/cmd/compile/internal/syntax/source.go
index c671e3c11e..01b592152b 100644
--- a/src/cmd/compile/internal/syntax/source.go
+++ b/src/cmd/compile/internal/syntax/source.go
@@ -3,11 +3,10 @@
 // license that can be found in the LICENSE file.
 
 // This file implements source, a buffered rune reader
-// which is specialized for the needs of the Go scanner:
-// Contiguous sequences of runes (literals) are extracted
-// directly as []byte without the need to re-encode the
-// runes in UTF-8 (as would be necessary with bufio.Reader).
-//
+// specialized for scanning Go code: Reading
+// ASCII characters, maintaining current (line, col)
+// position information, and recording of the most
+// recently read source segment are highly optimized.
 // This file is self-contained (go tool compile source.go
 // compiles) and thus could be made into its own package.
 
@@ -18,202 +17,202 @@ import (
 	"unicode/utf8"
 )
 
-// starting points for line and column numbers
-const linebase = 1
-const colbase = 1
-
-// max. number of bytes to unread
-const maxunread = 10
-
-// buf [...read...|...|...unread...|s|...free...]
-//         ^      ^   ^            ^
-//         |      |   |            |
-//        suf     r0  r            w
+// The source buffer is accessed using three indices b (begin),
+// r (read), and e (end):
+//
+// - If b >= 0, it points to the beginning of a segment of most
+//   recently read characters (typically a Go literal).
+//
+// - r points to the byte immediately following the most recently
+//   read character ch, which starts at r-chw.
+//
+// - e points to the byte immediately following the last byte that
+//   was read into the buffer.
+//
+// The buffer content is terminated at buf[e] with the sentinel
+// character utf8.RuneSelf. This makes it possible to test for
+// the common case of ASCII characters with a single 'if' (see
+// nextch method).
+//
+//                +------ content in use -------+
+//                v                             v
+// buf [...read...|...segment...|ch|...unread...|s|...free...]
+//                ^             ^  ^            ^
+//                |             |  |            |
+//                b         r-chw  r            e
+//
+// Invariant: -1 <= b < r <= e < len(buf) && buf[e] == sentinel
 
 type source struct {
-	src  io.Reader
-	errh func(line, pos uint, msg string)
-
-	// source buffer
-	buf         [4 << 10]byte
-	r0, r, w    int   // previous/current read and write buf positions, excluding sentinel
-	line0, line uint  // previous/current line
-	col0, col   uint  // previous/current column (byte offsets from line start)
-	ioerr       error // pending io error
-
-	// literal buffer
-	lit []byte // literal prefix
-	suf int    // literal suffix; suf >= 0 means we are scanning a literal
+	in   io.Reader
+	errh func(line, col uint, msg string)
+
+	buf       []byte // source buffer
+	ioerr     error  // pending I/O error, or nil
+	b, r, e   int    // buffer indices (see comment above)
+	line, col uint   // source position of ch (0-based)
+	ch        rune   // most recently read character
+	chw       int    // width of ch
 }
 
-// init initializes source to read from src and to report errors via errh.
-// errh must not be nil.
-func (s *source) init(src io.Reader, errh func(line, pos uint, msg string)) {
-	s.src = src
+const sentinel = utf8.RuneSelf
+
+func (s *source) init(in io.Reader, errh func(line, col uint, msg string)) {
+	s.in = in
 	s.errh = errh
 
-	s.buf[0] = utf8.RuneSelf // terminate with sentinel
-	s.r0, s.r, s.w = 0, 0, 0
-	s.line0, s.line = 0, linebase
-	s.col0, s.col = 0, colbase
+	if s.buf == nil {
+		s.buf = make([]byte, nextSize(0))
+	}
+	s.buf[0] = sentinel
 	s.ioerr = nil
-
-	s.lit = s.lit[:0]
-	s.suf = -1
+	s.b, s.r, s.e = -1, 0, 0
+	s.line, s.col = 0, 0
+	s.ch = ' '
+	s.chw = 0
 }
 
-// ungetr sets the reading position to a previous reading
-// position, usually the one of the most recently read
-// rune, but possibly earlier (see unread below).
-func (s *source) ungetr() {
-	s.r, s.line, s.col = s.r0, s.line0, s.col0
-}
+// starting points for line and column numbers
+const linebase = 1
+const colbase = 1
 
-// unread moves the previous reading position to a position
-// that is n bytes earlier in the source. The next ungetr
-// call will set the reading position to that moved position.
-// The "unread" runes must be single byte and not contain any
-// newlines; and 0 <= n <= maxunread must hold.
-func (s *source) unread(n int) {
-	s.r0 -= n
-	s.col0 -= uint(n)
+// pos returns the (line, col) source position of s.ch.
+func (s *source) pos() (line, col uint) {
+	return linebase + s.line, colbase + s.col
 }
 
+// error reports the error msg at source position s.pos().
 func (s *source) error(msg string) {
-	s.errh(s.line0, s.col0, msg)
+	line, col := s.pos()
+	s.errh(line, col, msg)
 }
 
-// getr reads and returns the next rune.
-//
-// If a read or source encoding error occurs, getr
-// calls the error handler installed with init.
-// The handler must exist.
-//
-// The (line, col) position passed to the error handler
-// is always at the current source reading position.
-func (s *source) getr() rune {
-redo:
-	s.r0, s.line0, s.col0 = s.r, s.line, s.col
-
-	// We could avoid at least one test that is always taken in the
-	// for loop below by duplicating the common case code (ASCII)
-	// here since we always have at least the sentinel (utf8.RuneSelf)
-	// in the buffer. Measure and optimize if necessary.
+// start starts a new active source segment (including s.ch).
+// As long as stop has not been called, the active segment's
+// bytes (excluding s.ch) may be retrieved by calling segment.
+func (s *source) start()          { s.b = s.r - s.chw }
+func (s *source) stop()           { s.b = -1 }
+func (s *source) segment() []byte { return s.buf[s.b : s.r-s.chw] }
+
+// rewind rewinds the scanner's read position and character s.ch
+// to the start of the currently active segment, which must not
+// contain any newlines (otherwise position information will be
+// incorrect). Currently, rewind is only needed for handling the
+// source sequence ".."; it must not be called outside an active
+// segment.
+func (s *source) rewind() {
+	// ok to verify precondition - rewind is rarely called
+	if s.b < 0 {
+		panic("no active segment")
+	}
+	s.col -= uint(s.r - s.b)
+	s.r = s.b
+	s.nextch()
+}
 
-	// make sure we have at least one rune in buffer, or we are at EOF
-	for s.r+utf8.UTFMax > s.w && !utf8.FullRune(s.buf[s.r:s.w]) && s.ioerr == nil && s.w-s.r < len(s.buf) {
-		s.fill() // s.w-s.r < len(s.buf) => buffer is not full
+func (s *source) nextch() {
+redo:
+	s.col += uint(s.chw)
+	if s.ch == '\n' {
+		s.line++
+		s.col = 0
 	}
 
-	// common case: ASCII and enough bytes
-	// (invariant: s.buf[s.w] == utf8.RuneSelf)
-	if b := s.buf[s.r]; b < utf8.RuneSelf {
+	// fast common case: at least one ASCII character
+	if s.ch = rune(s.buf[s.r]); s.ch < sentinel {
 		s.r++
-		// TODO(gri) Optimization: Instead of adjusting s.col for each character,
-		// remember the line offset instead and then compute the offset as needed
-		// (which is less often).
-		s.col++
-		if b == 0 {
+		s.chw = 1
+		if s.ch == 0 {
 			s.error("invalid NUL character")
 			goto redo
 		}
-		if b == '\n' {
-			s.line++
-			s.col = colbase
-		}
-		return rune(b)
+		return
+	}
+
+	// slower general case: add more bytes to buffer if we don't have a full rune
+	for s.e-s.r < utf8.UTFMax && !utf8.FullRune(s.buf[s.r:s.e]) && s.ioerr == nil {
+		s.fill()
 	}
 
 	// EOF
-	if s.r == s.w {
+	if s.r == s.e {
 		if s.ioerr != io.EOF {
 			// ensure we never start with a '/' (e.g., rooted path) in the error message
 			s.error("I/O error: " + s.ioerr.Error())
+			s.ioerr = nil
 		}
-		return -1
+		s.ch = -1
+		s.chw = 0
+		return
 	}
 
-	// uncommon case: not ASCII
-	r, w := utf8.DecodeRune(s.buf[s.r:s.w])
-	s.r += w
-	s.col += uint(w)
+	s.ch, s.chw = utf8.DecodeRune(s.buf[s.r:s.e])
+	s.r += s.chw
 
-	if r == utf8.RuneError && w == 1 {
+	if s.ch == utf8.RuneError && s.chw == 1 {
 		s.error("invalid UTF-8 encoding")
 		goto redo
 	}
 
 	// BOM's are only allowed as the first character in a file
 	const BOM = 0xfeff
-	if r == BOM {
-		if s.r0 > 0 { // s.r0 is always > 0 after 1st character (fill will set it to maxunread)
+	if s.ch == BOM {
+		if s.line > 0 || s.col > 0 {
 			s.error("invalid BOM in the middle of the file")
 		}
 		goto redo
 	}
-
-	return r
 }
 
+// fill reads more source bytes into s.buf.
+// It returns with at least one more byte in the buffer, or with s.ioerr != nil.
 func (s *source) fill() {
-	// Slide unread bytes to beginning but preserve last read char
-	// (for one ungetr call) plus maxunread extra bytes (for one
-	// unread call).
-	if s.r0 > maxunread {
-		n := s.r0 - maxunread // number of bytes to slide down
-		// save literal prefix, if any
-		// (make sure we keep maxunread bytes and the last
-		// read char in the buffer)
-		if s.suf >= 0 {
-			// we have a literal
-			if s.suf < n {
-				// save literal prefix
-				s.lit = append(s.lit, s.buf[s.suf:n]...)
-				s.suf = 0
-			} else {
-				s.suf -= n
-			}
-		}
-		copy(s.buf[:], s.buf[n:s.w])
-		s.r0 = maxunread // eqv: s.r0 -= n
-		s.r -= n
-		s.w -= n
+	// determine content to preserve
+	b := s.r
+	if s.b >= 0 {
+		b = s.b
+		s.b = 0 // after buffer has grown or content has been moved down
 	}
+	content := s.buf[b:s.e]
+
+	// grow buffer or move content down
+	if len(content)*2 > len(s.buf) {
+		s.buf = make([]byte, nextSize(len(s.buf)))
+		copy(s.buf, content)
+	} else if b > 0 {
+		copy(s.buf, content)
+	}
+	s.r -= b
+	s.e -= b
 
 	// read more data: try a limited number of times
-	for i := 100; i > 0; i-- {
-		n, err := s.src.Read(s.buf[s.w : len(s.buf)-1]) // -1 to leave space for sentinel
+	for i := 0; i < 10; i++ {
+		var n int
+		n, s.ioerr = s.in.Read(s.buf[s.e : len(s.buf)-1]) // -1 to leave space for sentinel
 		if n < 0 {
 			panic("negative read") // incorrect underlying io.Reader implementation
 		}
-		s.w += n
-		if n > 0 || err != nil {
-			s.buf[s.w] = utf8.RuneSelf // sentinel
-			if err != nil {
-				s.ioerr = err
-			}
+		if n > 0 || s.ioerr != nil {
+			s.e += n
+			s.buf[s.e] = sentinel
 			return
 		}
+		// n == 0
 	}
 
-	s.buf[s.w] = utf8.RuneSelf // sentinel
+	s.buf[s.e] = sentinel
 	s.ioerr = io.ErrNoProgress
 }
 
-func (s *source) startLit() {
-	s.suf = s.r0
-	s.lit = s.lit[:0] // reuse lit
-}
-
-func (s *source) stopLit() []byte {
-	lit := s.buf[s.suf:s.r]
-	if len(s.lit) > 0 {
-		lit = append(s.lit, lit...)
+// nextSize returns the next bigger size for a buffer of a given size.
+func nextSize(size int) int {
+	const min = 4 << 10 // 4K: minimum buffer size
+	const max = 1 << 20 // 1M: maximum buffer size which is still doubled
+	if size < min {
+		return min
 	}
-	s.killLit()
-	return lit
-}
-
-func (s *source) killLit() {
-	s.suf = -1 // no pending literal
+	if size <= max {
+		return size << 1
+	}
+	return size + max
 }