compress/flate: optimize huffSym

By using local variables and assigning them back to decompressor
at the end of huffSym, we allow compiler to keep them in registers
and avoid reloading/storing them repeatedly. To archive this,
moreBits was inlined and specialized to work with local variables.
Also move EOF error conversion to helper function, to make inlined
part of moreBits more readable. Together this results in nice speed-up:

name                             old time/op    new time/op    delta
Decode/Digits/Huffman/1e4-6         278µs ± 1%     240µs ± 2%  -13.72%  (p=0.000 n=10+10)
Decode/Digits/Huffman/1e5-6        2.38ms ± 1%    2.05ms ± 1%  -14.12%  (p=0.000 n=10+10)
Decode/Digits/Huffman/1e6-6        23.4ms ± 1%    19.9ms ± 0%  -14.69%  (p=0.000 n=9+9)
Decode/Digits/Speed/1e4-6           280µs ± 2%     254µs ± 1%   -9.28%  (p=0.000 n=10+9)
Decode/Digits/Speed/1e5-6          2.53ms ± 1%    2.35ms ± 1%   -7.17%  (p=0.000 n=10+10)
Decode/Digits/Speed/1e6-6          24.8ms ± 1%    23.0ms ± 1%   -7.22%  (p=0.000 n=10+10)
Decode/Digits/Default/1e4-6         281µs ± 2%     259µs ± 3%   -8.03%  (p=0.000 n=10+10)
Decode/Digits/Default/1e5-6        2.45ms ± 1%    2.30ms ± 1%   -6.15%  (p=0.000 n=10+10)
Decode/Digits/Default/1e6-6        24.1ms ± 1%    22.6ms ± 0%   -6.31%  (p=0.000 n=9+9)
Decode/Digits/Compression/1e4-6     279µs ± 2%     261µs ± 2%   -6.53%  (p=0.000 n=8+9)
Decode/Digits/Compression/1e5-6    2.44ms ± 1%    2.30ms ± 1%   -5.72%  (p=0.000 n=10+9)
Decode/Digits/Compression/1e6-6    24.0ms ± 1%    22.6ms ± 0%   -6.10%  (p=0.000 n=10+9)
Decode/Twain/Huffman/1e4-6          316µs ± 2%     267µs ± 3%  -15.30%  (p=0.000 n=9+10)
Decode/Twain/Huffman/1e5-6         2.62ms ± 0%    2.22ms ± 0%  -15.24%  (p=0.000 n=10+10)
Decode/Twain/Huffman/1e6-6         25.7ms ± 1%    21.8ms ± 0%  -15.19%  (p=0.000 n=10+10)
Decode/Twain/Speed/1e4-6            290µs ± 1%     264µs ± 2%   -9.17%  (p=0.000 n=9+10)
Decode/Twain/Speed/1e5-6           2.35ms ± 1%    2.13ms ± 1%   -9.74%  (p=0.000 n=9+10)
Decode/Twain/Speed/1e6-6           22.9ms ± 0%    20.7ms ± 0%   -9.68%  (p=0.000 n=10+9)
Decode/Twain/Default/1e4-6          270µs ± 2%     252µs ± 2%   -6.67%  (p=0.000 n=9+10)
Decode/Twain/Default/1e5-6         2.02ms ± 1%    1.84ms ± 1%   -8.85%  (p=0.000 n=10+10)
Decode/Twain/Default/1e6-6         19.1ms ± 0%    17.5ms ± 1%   -8.73%  (p=0.000 n=9+10)
Decode/Twain/Compression/1e4-6      272µs ± 1%     250µs ± 4%   -8.20%  (p=0.000 n=9+10)
Decode/Twain/Compression/1e5-6     2.01ms ± 0%    1.84ms ± 1%   -8.57%  (p=0.000 n=9+10)
Decode/Twain/Compression/1e6-6     19.1ms ± 0%    17.4ms ± 1%   -8.75%  (p=0.000 n=9+10)

name                             old speed      new speed      delta
Decode/Digits/Huffman/1e4-6      35.9MB/s ± 1%  41.7MB/s ± 2%  +15.91%  (p=0.000 n=10+10)
Decode/Digits/Huffman/1e5-6      41.9MB/s ± 1%  48.8MB/s ± 1%  +16.44%  (p=0.000 n=10+10)
Decode/Digits/Huffman/1e6-6      42.8MB/s ± 1%  50.2MB/s ± 0%  +17.22%  (p=0.000 n=9+9)
Decode/Digits/Speed/1e4-6        35.7MB/s ± 2%  39.4MB/s ± 1%  +10.22%  (p=0.000 n=10+9)
Decode/Digits/Speed/1e5-6        39.6MB/s ± 1%  42.6MB/s ± 1%   +7.73%  (p=0.000 n=10+10)
Decode/Digits/Speed/1e6-6        40.3MB/s ± 1%  43.4MB/s ± 1%   +7.78%  (p=0.000 n=10+10)
Decode/Digits/Default/1e4-6      35.6MB/s ± 2%  38.7MB/s ± 2%   +8.74%  (p=0.000 n=10+10)
Decode/Digits/Default/1e5-6      40.9MB/s ± 1%  43.6MB/s ± 1%   +6.55%  (p=0.000 n=10+10)
Decode/Digits/Default/1e6-6      41.5MB/s ± 1%  44.3MB/s ± 0%   +6.73%  (p=0.000 n=9+9)
Decode/Digits/Compression/1e4-6  35.8MB/s ± 2%  38.3MB/s ± 2%   +6.99%  (p=0.000 n=8+9)
Decode/Digits/Compression/1e5-6  40.9MB/s ± 1%  43.4MB/s ± 1%   +6.07%  (p=0.000 n=10+9)
Decode/Digits/Compression/1e6-6  41.6MB/s ± 1%  44.3MB/s ± 0%   +6.49%  (p=0.000 n=10+9)
Decode/Twain/Huffman/1e4-6       31.7MB/s ± 2%  37.4MB/s ± 3%  +18.08%  (p=0.000 n=9+10)
Decode/Twain/Huffman/1e5-6       38.2MB/s ± 0%  45.0MB/s ± 0%  +17.97%  (p=0.000 n=10+10)
Decode/Twain/Huffman/1e6-6       38.9MB/s ± 1%  45.9MB/s ± 0%  +17.90%  (p=0.000 n=10+10)
Decode/Twain/Speed/1e4-6         34.5MB/s ± 1%  38.0MB/s ± 2%  +10.11%  (p=0.000 n=9+10)
Decode/Twain/Speed/1e5-6         42.5MB/s ± 1%  47.0MB/s ± 1%  +10.79%  (p=0.000 n=9+10)
Decode/Twain/Speed/1e6-6         43.7MB/s ± 0%  48.3MB/s ± 0%  +10.72%  (p=0.000 n=10+9)
Decode/Twain/Default/1e4-6       37.1MB/s ± 2%  39.8MB/s ± 2%   +7.15%  (p=0.000 n=9+10)
Decode/Twain/Default/1e5-6       49.5MB/s ± 1%  54.3MB/s ± 1%   +9.71%  (p=0.000 n=10+10)
Decode/Twain/Default/1e6-6       52.3MB/s ± 0%  57.3MB/s ± 1%   +9.57%  (p=0.000 n=9+10)
Decode/Twain/Compression/1e4-6   36.7MB/s ± 1%  40.0MB/s ± 4%   +8.96%  (p=0.000 n=9+10)
Decode/Twain/Compression/1e5-6   49.8MB/s ± 0%  54.5MB/s ± 1%   +9.38%  (p=0.000 n=9+10)
Decode/Twain/Compression/1e6-6   52.3MB/s ± 0%  57.3MB/s ± 1%   +9.58%  (p=0.000 n=9+10)

Change-Id: Iabfd285535ddb210f7f48f33317c6463b5532400
Reviewed-on: https://go-review.googlesource.com/102235
Run-TryBot: Ilya Tocar <ilya.tocar@intel.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>

src/compress/flate/inflate.go

@@ -629,10 +629,7 @@ func (f *decompressor) dataBlock() {
 	nr, err := io.ReadFull(f.r, f.buf[0:4])
 	f.roffset += int64(nr)
 	if err != nil {
-		if err == io.EOF {
-			err = io.ErrUnexpectedEOF
-		}
-		f.err = err
+		f.err = noEOF(err)
 		return
 	}
 	n := int(f.buf[0]) | int(f.buf[1])<<8
@@ -665,10 +662,7 @@ func (f *decompressor) copyData() {
 	f.copyLen -= cnt
 	f.dict.writeMark(cnt)
 	if err != nil {
-		if err == io.EOF {
-			err = io.ErrUnexpectedEOF
-		}
-		f.err = err
+		f.err = noEOF(err)
 		return
 	}
 
@@ -690,13 +684,18 @@ func (f *decompressor) finishBlock() {
 	f.step = (*decompressor).nextBlock
 }
 
+// noEOF returns err, unless err == io.EOF, in which case it returns io.ErrUnexpectedEOF.
+func noEOF(e error) error {
+	if e == io.EOF {
+		return io.ErrUnexpectedEOF
+	}
+	return e
+}
+
 func (f *decompressor) moreBits() error {
 	c, err := f.r.ReadByte()
 	if err != nil {
-		if err == io.EOF {
-			err = io.ErrUnexpectedEOF
-		}
-		return err
+		return noEOF(err)
 	}
 	f.roffset++
 	f.b |= uint32(c) << f.nb
@@ -711,25 +710,37 @@ func (f *decompressor) huffSym(h *huffmanDecoder) (int, error) {
 	// cases, the chunks slice will be 0 for the invalid sequence, leading it
 	// satisfy the n == 0 check below.
 	n := uint(h.min)
+	// Optimization. Compiler isn't smart enough to keep f.b,f.nb in registers,
+	// but is smart enough to keep local variables in registers, so use nb and b,
+	// inline call to moreBits and reassign b,nb back to f on return.
+	nb, b := f.nb, f.b
 	for {
-		for f.nb < n {
-			if err := f.moreBits(); err != nil {
-				return 0, err
+		for nb < n {
+			c, err := f.r.ReadByte()
+			if err != nil {
+				f.b = b
+				f.nb = nb
+				return 0, noEOF(err)
 			}
+			f.roffset++
+			b |= uint32(c) << (nb & 31)
+			nb += 8
 		}
-		chunk := h.chunks[f.b&(huffmanNumChunks-1)]
+		chunk := h.chunks[b&(huffmanNumChunks-1)]
 		n = uint(chunk & huffmanCountMask)
 		if n > huffmanChunkBits {
-			chunk = h.links[chunk>>huffmanValueShift][(f.b>>huffmanChunkBits)&h.linkMask]
+			chunk = h.links[chunk>>huffmanValueShift][(b>>huffmanChunkBits)&h.linkMask]
 			n = uint(chunk & huffmanCountMask)
 		}
-		if n <= f.nb {
+		if n <= nb {
 			if n == 0 {
+				f.b = b
+				f.nb = nb
 				f.err = CorruptInputError(f.roffset)
 				return 0, f.err
 			}
-			f.b >>= n
-			f.nb -= n
+			f.b = b >> (n & 31)
+			f.nb = nb - n
 			return int(chunk >> huffmanValueShift), nil
 		}
 	}