extends CharsetDecoder to create Base64 Decoder
/*
* Base64Decoder.java
*
* Created on 20 December 2005, 12:14
*
*/
import java.nio.CharBuffer;
import java.nio.charset.Charset;
import java.nio.charset.CharsetDecoder;
import java.nio.charset.CoderResult;
/**
* <p>Base 64 decoder class. Converts byte array to characters.</p>
* <p>If measure is supplied this is a line width, and \n characters
* are inserted when it's exceeeded.</p>
*
* @author Malcolm McMahon
* @version $Revision: 1.1 $
*/
public class Base64Decoder extends CharsetDecoder {
Integer measure;
int charCount;
char padding;
String codeChars;
char[] excessChars;
int codeState;
int bits;
/**
* Creates a new instance of Base64Decoder
* @param cs The Charset that created this object
* @param padding The padding character, usually =
* @param measure The line width in characters
*/
public Base64Decoder(Charset cs,String codeChars, char padding, int measure) {
super(cs, 1.33f, 4f);
this.measure = measure;
this.codeChars = codeChars;
this.padding = padding;
}
/**
* @param cs The Charset that created this object
* @param padding The padding character, usually =
*/
public Base64Decoder(Charset cs,String codeChars, char padding) {
super(cs, 1.33f, 4f);
this.codeChars = codeChars;
this.padding = padding;
}
private boolean flushExcess(CharBuffer out) {
if(excessChars == null)
return true;
for(int i = 0; i < excessChars.length; i++)
if(out.remaining() > 0)
out.put(excessChars[i]);
else {
char[] nc = new char[excessChars.length - i];
System.arraycopy(excessChars, i, nc, 0, nc.length);
excessChars = nc;
return false;
}
excessChars = null;
return true;
}
/**
* Output one character. These character output methods
* park characters in excessChars if the buffer overflows.
* The also add newlines if the measure calls for it.
* @return true if space found
*/
private boolean out(CharBuffer out, char ch) {
if(measure != null && ++charCount > measure) {
charCount = -1;
return out(out, '\n', ch);
}
if(out.remaining() > 0) {
out.put(ch);
return true;
} else {
excessChars = new char[]{ch};
return false;
}
}
/**
* Output two characters
* @return true if space found
*/
private boolean out(CharBuffer out, char ch1, char ch2) {
if(measure != null && (charCount += 2) > measure) {
if(charCount - 1 != measure) {
charCount = 2;
return out(out, '\n', ch1, ch2);
} else {
charCount = 1;
return out(out, ch1, '\n', ch2);
}
}
switch(out.remaining()) {
case 0:
excessChars = new char[]{ch1, ch2};
return false;
case 1:
out.put(ch1);
excessChars = new char[]{ch2};
return false;
default:
out.put(ch1);
out.put(ch2);
return true;
}
}
private boolean out(CharBuffer out, char ch1, char ch2, char ch3) {
switch(out.remaining()) {
case 0:
excessChars = new char[]{ch1, ch2, ch3};
return false;
case 1:
out.put(ch1);
excessChars = new char[]{ch2, ch3};
return false;
case 2:
out.put(ch1);
out.put(ch2);
excessChars = new char[]{ch3};
return false;
default:
out.put(ch1);
out.put(ch2);
out.put(ch3);
return true;
}
}
/**
* Flushes this decoder.
*
* <p> The default implementation of this method does nothing, and always
* returns {@link CoderResult#UNDERFLOW}. This method should be overridden
* by decoders that may need to write final characters to the output buffer
* once the entire input sequence has been read. </p>
*
* @param out
* The output character buffer
*
* @return A coder-result object, either {@link CoderResult#UNDERFLOW} or
* {@link CoderResult#OVERFLOW}
*/
protected java.nio.charset.CoderResult implFlush(CharBuffer out) {
if(!flushExcess(out))
return CoderResult.OVERFLOW;
switch(codeState) {
case 0: // full quanta count
implReset();
return CoderResult.UNDERFLOW;
case 1:
if(!out(out, codeChars.charAt(bits), padding))
return CoderResult.OVERFLOW;
break;
case 2:
if(!out(out, codeChars.charAt(bits)))
return CoderResult.OVERFLOW;
break;
}
if(out(out, padding)) {
implReset();
return CoderResult.UNDERFLOW;
} else
return CoderResult.OVERFLOW;
}
/**
* Decodes one or more bytes into one or more characters.
*
* <p> This method encapsulates the basic decoding loop, decoding as many
* bytes as possible until it either runs out of input, runs out of room
* in the output buffer, or encounters a decoding error. This method is
* invoked by the {@link #decode decode} method, which handles result
* interpretation and error recovery.
*
* <p> The buffers are read from, and written to, starting at their current
* positions. At most {@link Buffer#remaining in.remaining()} bytes
* will be read, and at most {@link Buffer#remaining out.remaining()}
* characters will be written. The buffers' positions will be advanced to
* reflect the bytes read and the characters written, but their marks and
* limits will not be modified.
*
* <p> This method returns a {@link CoderResult} object to describe its
* reason for termination, in the same manner as the {@link #decode decode}
* method. Most implementations of this method will handle decoding errors
* by returning an appropriate result object for interpretation by the
* {@link #decode decode} method. An optimized implementation may instead
* examine the relevant error action and implement that action itself.
*
* <p> An implementation of this method may perform arbitrary lookahead by
* returning {@link CoderResult#UNDERFLOW} until it receives sufficient
* input. </p>
*
* @param in
* The input byte buffer
*
* @param out
* The output character buffer
*
* @return A coder-result object describing the reason for termination
*/
public java.nio.charset.CoderResult decodeLoop(java.nio.ByteBuffer in, java.nio.CharBuffer out) {
if(!flushExcess(out))
return CoderResult.OVERFLOW;
while(in.remaining() > 0) {
int b = in.get() & 0xff;
switch(codeState) {
case 0:
bits = (b << 4) & 0x30;
codeState = 1;
if(!out(out, codeChars.charAt((b >> 2) & 0x3f)))
return CoderResult.OVERFLOW;
break;
case 1:
int cc = bits | (b >> 4) & 0xf;
bits = (b << 2) & 0x3c;
codeState = 2;
if(!out(out, codeChars.charAt(cc)))
return CoderResult.OVERFLOW;
break;
case 2:
cc = bits | (b >> 6) & 3;
codeState = 0;
if(!out(out, codeChars.charAt(cc), codeChars.charAt(b & 0x3f)))
return CoderResult.OVERFLOW;
}
}
return CoderResult.UNDERFLOW;
}
/**
* Resets this decoder, clearing any charset-specific internal state.
*
* <p> The default implementation of this method does nothing. This method
* should be overridden by decoders that maintain internal state. </p>
*/
protected void implReset() {
bits = 0;
excessChars = null;
codeState = 0;
charCount = 0;
}
}
Related examples in the same category