Index: src/0dev.org/predictor/predictor.go
==================================================================
--- src/0dev.org/predictor/predictor.go
+++ src/0dev.org/predictor/predictor.go
@@ -1,10 +1,11 @@
 // Package predictor implements the predictor compression/decompression algorithm
 // as specified by RFC1978 - PPP Predictor Compression Protocol
 package predictor
 
 import (
+	bits "0dev.org/bits"
 	"io"
 )
 
 type context struct {
 	table [1 << 16]byte
@@ -70,11 +71,10 @@
 			// ... and stage the rest of the data in the buffer
 			ctx.input = append(ctx.input, data[blockSize-bufferLength:]...)
 			return nil
 		}
 
-		// TODO allocate this on ctx.buffer ...
 		var buf []byte = make([]byte, 1, blockSize+1)
 		for block := 0; block < len(data)/blockSize; block++ {
 			for i := 0; i < blockSize; i++ {
 				var current byte = data[(block*blockSize)+i]
 				if ctx.table[ctx.hash] == current {
@@ -117,83 +117,99 @@
 	return r(output)
 }
 
 // Returns an io.Reader implementation that wraps the provided io.Reader
 // and decompresses data according to the predictor algorithm
-func Decompressor(wrapped io.Reader) io.Reader {
+func Decompressor(reader io.Reader) io.Reader {
 	var ctx context
 	ctx.input = make([]byte, 0, 8)
 
 	return decompressor(func(output []byte) (int, error) {
 		var (
-			err       error
-			flags     byte
-			readCount int
+			err                             error
+			flags                           byte
+			rc, available, predicted, total int
 		)
 
 		// Sanity check for space to read into
 		if len(output) == 0 {
 			return 0, nil
 		}
 
 		// Check whether we have leftover data in the buffer
 		if len(ctx.input) > 0 {
-			readCount = copy(output, ctx.input)
-
-			// Check whether we still have leftover data in the buffer :)
-			if readCount < len(ctx.input) {
-				ctx.input = ctx.input[:copy(ctx.input, ctx.input[readCount:])]
-			}
-			return readCount, nil
-		}
-
-		// This is single-iteration only but it is fine according to io.Reader's contract ?!
-		// TODO - read all bytes from a block based on the hamming weight of the flag
-		// and just shuffle them for predictions instead of bite-sized reads ;)
-
-		// Read the flags
-		readCount, err = wrapped.Read(ctx.input[:1])
-		if readCount == 0 || err != nil {
-			return readCount, err
-		}
-
-		ctx.input = ctx.input[:8]
-		flags = ctx.input[0]
-
-		var i uint = 0
-		for ; i < 8; i++ {
-			if flags&(1<<i) > 0 {
-				// Guess was right
-				ctx.input[i] = ctx.table[ctx.hash]
-			} else {
-				readCount, err = wrapped.Read(ctx.input[i:(i + 1)])
-
-				if err == io.EOF {
-					break
-				}
-
-				if err != nil {
-					return readCount, err
-				}
-
-				if readCount == 0 { // treat as EoF
-					break
-				}
-
-				ctx.table[ctx.hash] = ctx.input[i]
-			}
-
+			rc = copy(output, ctx.input)
+
+			// Check whether we still have leftover data in the buffer :)
+			if rc < len(ctx.input) {
+				ctx.input = ctx.input[:copy(ctx.input, ctx.input[rc:])]
+			}
+			return rc, nil
+		}
+
+		// Read the next prediction header
+	readHeader:
+		rc, err = reader.Read(ctx.input[:1])
+		// Fail on error unless it is EOF
+		if err != nil && err != io.EOF {
+			return total, err
+		} else if rc == 0 {
+			return total, err
+		}
+
+		// Extend the buffer, copy the prediction header
+		//  and calculate the number of subsequent bytes to read
+		ctx.input = ctx.input[:8]
+		flags = ctx.input[0]
+		predicted = int(bits.Hamming(flags))
+		available = 8 - predicted
+
+		// Read the non-predicted bytes and place them in the end of the buffer
+		rc, err = reader.Read(ctx.input[predicted:])
+	retryData:
+		if rc < int(available) && err == nil {
+			// Retry the read if we have fewer bytes than what the prediction header indicates
+			var r int
+			r, err = reader.Read(ctx.input[predicted+rc:])
+			rc += r
+			goto retryData
+		} // Continue on any error, try to decompress and return it along the result
+
+		// Walk the buffer, filling in the predicted blanks,
+		// relocating read bytes and and updating the guess table
+		for i, a := uint(0), predicted; i < 8; i++ {
+			if (flags & (1 << i)) > 0 {
+				// Guess succeeded, fill in from the table
+				ctx.input[i] = ctx.table[ctx.hash]
+				rc++
+			} else {
+				// Relocate a read byte
+				ctx.input[i], a = ctx.input[a], a+1
+				// Guess failed, update the table
+				ctx.table[ctx.hash] = ctx.input[i]
+			}
+			// Update the hash
 			ctx.hash = (ctx.hash << 4) ^ uint16(ctx.input[i])
 		}
 
-		readCount = copy(output, ctx.input[:i])
+		// rc now contains the precise amount of populated data
+		ctx.input = ctx.input[:rc]
+		available = copy(output, ctx.input)
+
+		total += available
 
-		// Place any remaining bytes in the buffer
-		if uint(readCount) < i {
-			ctx.input = ctx.input[readCount:i]
+		// Check for remaining bytes that dont fit in the output buffer
+		if available < rc {
+			ctx.input = ctx.input[:copy(ctx.input, ctx.input[available:])]
 		} else {
+			// Clear the buffer
 			ctx.input = ctx.input[:0]
+
+			output = output[available:]
+			if len(output) > 0 && err == nil {
+				goto readHeader
+			}
 		}
 
-		return readCount, nil
+		return total, err
 	})
 }