com.ctc.wstx.io.UTF8Writer Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of woodstox-core-asl Show documentation
Show all versions of woodstox-core-asl Show documentation
Woodstox is a high-performance XML processor that
implements Stax (JSR-173) and SAX2 APIs
The newest version!
package com.ctc.wstx.io;
import java.io.*;
import com.ctc.wstx.api.WriterConfig;
/**
* Specialized buffering UTF-8 writer used by
* {@link com.ctc.wstx.sw.XmlWriter}.
* The main reason for custom version is to allow for efficient
* buffer recycling; the second benefit is that encoder has less
* overhead for short content encoding (compared to JDK default
* codecs).
*/
public final class UTF8Writer
extends Writer
implements CompletelyCloseable
{
private final static int DEFAULT_BUF_LEN = 4000;
final static int SURR1_FIRST = 0xD800;
final static int SURR1_LAST = 0xDBFF;
final static int SURR2_FIRST = 0xDC00;
final static int SURR2_LAST = 0xDFFF;
final WriterConfig mConfig;
final boolean mAutoCloseOutput;
final OutputStream mOut;
byte[] mOutBuffer;
final int mOutBufferLast;
int mOutPtr;
/**
* When outputting chars from BMP, surrogate pairs need to be coalesced.
* To do this, both pairs must be known first; and since it is possible
* pairs may be split, we need temporary storage for the first half
*/
int mSurrogate = 0;
public UTF8Writer(WriterConfig cfg, OutputStream out, boolean autoclose)
{
mConfig = cfg;
mAutoCloseOutput = autoclose;
mOut = out;
mOutBuffer = (mConfig == null) ? new byte[DEFAULT_BUF_LEN] : cfg.allocFullBBuffer(DEFAULT_BUF_LEN);
/* Max. expansion for a single char (in unmodified UTF-8) is
* 4 bytes (or 3 depending on how you view it -- 4 when recombining
* surrogate pairs)
*/
mOutBufferLast = mOutBuffer.length - 4;
mOutPtr = 0;
}
/*
////////////////////////////////////////////////////////
// CompletelyCloseable impl
////////////////////////////////////////////////////////
*/
public void closeCompletely() throws IOException
{
_close(true);
}
/*
////////////////////////////////////////////////////////
// java.io.Writer implementation
////////////////////////////////////////////////////////
*/
/* !!! 30-Nov-2006, TSa: Due to co-variance between Appendable and
* Writer, this would not compile with javac 1.5, in 1.4 mode
* (source and target set to "1.4". Not a huge deal, but since
* the base impl is just fine, no point in overriding it.
*/
/*
public Writer append(char c)
throws IOException
// note: this is a JDK 1.5 method
{
write(c);
return this;
}
*/
public void close() throws IOException
{
_close(mAutoCloseOutput);
}
public void flush()
throws IOException
{
if (mOutPtr > 0 && mOutBuffer != null) {
mOut.write(mOutBuffer, 0, mOutPtr);
mOutPtr = 0;
}
mOut.flush();
}
public void write(char[] cbuf)
throws IOException
{
write(cbuf, 0, cbuf.length);
}
public void write(char[] cbuf, int off, int len)
throws IOException
{
if (len < 2) {
if (len == 1) {
write(cbuf[off]);
}
return;
}
// First: do we have a leftover surrogate to deal with?
if (mSurrogate > 0) {
char second = cbuf[off++];
--len;
write(_convertSurrogate(second));
// will have at least one more char
}
int outPtr = mOutPtr;
byte[] outBuf = mOutBuffer;
int outBufLast = mOutBufferLast; // has 4 'spare' bytes
// All right; can just loop it nice and easy now:
len += off; // len will now be the end of input buffer
output_loop:
for (; off < len; ) {
/* First, let's ensure we can output at least 4 bytes
* (longest UTF-8 encoded codepoint):
*/
if (outPtr >= outBufLast) {
mOut.write(outBuf, 0, outPtr);
outPtr = 0;
}
int c = cbuf[off++];
// And then see if we have an Ascii char:
if (c < 0x80) { // If so, can do a tight inner loop:
outBuf[outPtr++] = (byte)c;
// Let's calc how many ascii chars we can copy at most:
int maxInCount = (len - off);
int maxOutCount = (outBufLast - outPtr);
if (maxInCount > maxOutCount) {
maxInCount = maxOutCount;
}
maxInCount += off;
ascii_loop:
while (true) {
if (off >= maxInCount) { // done with max. ascii seq
continue output_loop;
}
c = cbuf[off++];
if (c >= 0x80) {
break ascii_loop;
}
outBuf[outPtr++] = (byte) c;
}
}
// Nope, multi-byte:
if (c < 0x800) { // 2-byte
outBuf[outPtr++] = (byte) (0xc0 | (c >> 6));
outBuf[outPtr++] = (byte) (0x80 | (c & 0x3f));
} else { // 3 or 4 bytes
// Surrogates?
if (c < SURR1_FIRST || c > SURR2_LAST) {
outBuf[outPtr++] = (byte) (0xe0 | (c >> 12));
outBuf[outPtr++] = (byte) (0x80 | ((c >> 6) & 0x3f));
outBuf[outPtr++] = (byte) (0x80 | (c & 0x3f));
continue;
}
// Yup, a surrogate:
if (c > SURR1_LAST) { // must be from first range
mOutPtr = outPtr;
throwIllegal(c);
}
mSurrogate = c;
// and if so, followed by another from next range
if (off >= len) { // unless we hit the end?
break;
}
c = _convertSurrogate(cbuf[off++]);
if (c > 0x10FFFF) { // illegal, as per RFC 3629
mOutPtr = outPtr;
throwIllegal(c);
}
outBuf[outPtr++] = (byte) (0xf0 | (c >> 18));
outBuf[outPtr++] = (byte) (0x80 | ((c >> 12) & 0x3f));
outBuf[outPtr++] = (byte) (0x80 | ((c >> 6) & 0x3f));
outBuf[outPtr++] = (byte) (0x80 | (c & 0x3f));
}
}
mOutPtr = outPtr;
}
public void write(int c)
throws IOException
{
// First; do we have a left over surrogate?
if (mSurrogate > 0) {
c = _convertSurrogate(c);
// If not, do we start with a surrogate?
} else if (c >= SURR1_FIRST && c <= SURR2_LAST) {
// Illegal to get second part without first:
if (c > SURR1_LAST) {
throwIllegal(c);
}
// First part just needs to be held for now
mSurrogate = c;
return;
}
if (mOutPtr >= mOutBufferLast) { // let's require enough room, first
mOut.write(mOutBuffer, 0, mOutPtr);
mOutPtr = 0;
}
if (c < 0x80) { // ascii
mOutBuffer[mOutPtr++] = (byte) c;
} else {
int ptr = mOutPtr;
if (c < 0x800) { // 2-byte
mOutBuffer[ptr++] = (byte) (0xc0 | (c >> 6));
mOutBuffer[ptr++] = (byte) (0x80 | (c & 0x3f));
} else if (c <= 0xFFFF) { // 3 bytes
mOutBuffer[ptr++] = (byte) (0xe0 | (c >> 12));
mOutBuffer[ptr++] = (byte) (0x80 | ((c >> 6) & 0x3f));
mOutBuffer[ptr++] = (byte) (0x80 | (c & 0x3f));
} else { // 4 bytes
if (c > 0x10FFFF) { // illegal, as per RFC 3629
throwIllegal(c);
}
mOutBuffer[ptr++] = (byte) (0xf0 | (c >> 18));
mOutBuffer[ptr++] = (byte) (0x80 | ((c >> 12) & 0x3f));
mOutBuffer[ptr++] = (byte) (0x80 | ((c >> 6) & 0x3f));
mOutBuffer[ptr++] = (byte) (0x80 | (c & 0x3f));
}
mOutPtr = ptr;
}
}
public void write(String str)
throws IOException
{
write(str, 0, str.length());
}
public void write(String str, int off, int len)
throws IOException
{
if (len < 2) {
if (len == 1) {
write(str.charAt(off));
}
return;
}
// First: do we have a leftover surrogate to deal with?
if (mSurrogate > 0) {
char second = str.charAt(off++);
--len;
write(_convertSurrogate(second));
// will have at least one more char (case of 1 char was checked earlier on)
}
int outPtr = mOutPtr;
byte[] outBuf = mOutBuffer;
int outBufLast = mOutBufferLast; // has 4 'spare' bytes
// All right; can just loop it nice and easy now:
len += off; // len will now be the end of input buffer
output_loop:
for (; off < len; ) {
/* First, let's ensure we can output at least 4 bytes
* (longest UTF-8 encoded codepoint):
*/
if (outPtr >= outBufLast) {
mOut.write(outBuf, 0, outPtr);
outPtr = 0;
}
int c = str.charAt(off++);
// And then see if we have an Ascii char:
if (c < 0x80) { // If so, can do a tight inner loop:
outBuf[outPtr++] = (byte)c;
// Let's calc how many ascii chars we can copy at most:
int maxInCount = (len - off);
int maxOutCount = (outBufLast - outPtr);
if (maxInCount > maxOutCount) {
maxInCount = maxOutCount;
}
maxInCount += off;
ascii_loop:
while (true) {
if (off >= maxInCount) { // done with max. ascii seq
continue output_loop;
}
c = str.charAt(off++);
if (c >= 0x80) {
break ascii_loop;
}
outBuf[outPtr++] = (byte) c;
}
}
// Nope, multi-byte:
if (c < 0x800) { // 2-byte
outBuf[outPtr++] = (byte) (0xc0 | (c >> 6));
outBuf[outPtr++] = (byte) (0x80 | (c & 0x3f));
} else { // 3 or 4 bytes
// Surrogates?
if (c < SURR1_FIRST || c > SURR2_LAST) {
outBuf[outPtr++] = (byte) (0xe0 | (c >> 12));
outBuf[outPtr++] = (byte) (0x80 | ((c >> 6) & 0x3f));
outBuf[outPtr++] = (byte) (0x80 | (c & 0x3f));
continue;
}
// Yup, a surrogate:
if (c > SURR1_LAST) { // must be from first range
mOutPtr = outPtr;
throwIllegal(c);
}
mSurrogate = c;
// and if so, followed by another from next range
if (off >= len) { // unless we hit the end?
break;
}
c = _convertSurrogate(str.charAt(off++));
if (c > 0x10FFFF) { // illegal, as per RFC 3629
mOutPtr = outPtr;
throwIllegal(c);
}
outBuf[outPtr++] = (byte) (0xf0 | (c >> 18));
outBuf[outPtr++] = (byte) (0x80 | ((c >> 12) & 0x3f));
outBuf[outPtr++] = (byte) (0x80 | ((c >> 6) & 0x3f));
outBuf[outPtr++] = (byte) (0x80 | (c & 0x3f));
}
}
mOutPtr = outPtr;
}
/*
////////////////////////////////////////////////////////////
// Internal methods
////////////////////////////////////////////////////////////
*/
private final void _close(boolean forceClosing)
throws IOException
{
byte[] buf = mOutBuffer;
if (buf != null) {
mOutBuffer = null;
if (mOutPtr > 0) {
mOut.write(buf, 0, mOutPtr);
mOutPtr = 0;
}
if (mConfig != null) {
mConfig.freeFullBBuffer(buf);
}
}
if (forceClosing) {
mOut.close();
}
/* Let's 'flush' orphan surrogate, no matter what; but only
* after cleanly closing everything else.
*/
int code = mSurrogate;
if (code > 0) {
mSurrogate = 0;
throwIllegal(code);
}
}
/**
* Method called to calculate UTF codepoint, from a surrogate pair.
*/
private final int _convertSurrogate(int secondPart)
throws IOException
{
int firstPart = mSurrogate;
mSurrogate = 0;
// Ok, then, is the second part valid?
if (secondPart < SURR2_FIRST || secondPart > SURR2_LAST) {
throw new IOException("Broken surrogate pair: first char 0x"+Integer.toHexString(firstPart)+", second 0x"+Integer.toHexString(secondPart)+"; illegal combination");
}
return 0x10000 + ((firstPart - SURR1_FIRST) << 10) + (secondPart - SURR2_FIRST);
}
private void throwIllegal(int code)
throws IOException
{
if (code > 0x10FFFF) { // over max?
throw new IOException("Illegal character point (0x"+Integer.toHexString(code)+") to output; max is 0x10FFFF as per RFC 3629");
}
if (code >= SURR1_FIRST) {
if (code <= SURR1_LAST) { // Unmatched first part (closing without second part?)
throw new IOException("Unmatched first part of surrogate pair (0x"+Integer.toHexString(code)+")");
}
throw new IOException("Unmatched second part of surrogate pair (0x"+Integer.toHexString(code)+")");
}
// should we ever get this?
throw new IOException("Illegal character point (0x"+Integer.toHexString(code)+") to output");
}
}