in org.eclipse.sisu.inject/src/org/eclipse/sisu/space/asm/MethodWriter.java [2372:2787]
private void resizeInstructions() {
byte[] b = code.data; // bytecode of the method
int u, v, label; // indexes in b
int i, j; // loop indexes
/*
* 1st step: As explained above, resizing an instruction may require to
* resize another one, which may require to resize yet another one, and
* so on. The first step of the algorithm consists in finding all the
* instructions that need to be resized, without modifying the code.
* This is done by the following "fix point" algorithm:
*
* Parse the code to find the jump instructions whose offset will need
* more than 2 bytes to be stored (the future offset is computed from
* the current offset and from the number of bytes that will be inserted
* or removed between the source and target instructions). For each such
* instruction, adds an entry in (a copy of) the indexes and sizes
* arrays (if this has not already been done in a previous iteration!).
*
* If at least one entry has been added during the previous step, go
* back to the beginning, otherwise stop.
*
* In fact the real algorithm is complicated by the fact that the size
* of TABLESWITCH and LOOKUPSWITCH instructions depends on their
* position in the bytecode (because of padding). In order to ensure the
* convergence of the algorithm, the number of bytes to be added or
* removed from these instructions is over estimated during the previous
* loop, and computed exactly only after the loop is finished (this
* requires another pass to parse the bytecode of the method).
*/
int[] allIndexes = new int[0]; // copy of indexes
int[] allSizes = new int[0]; // copy of sizes
boolean[] resize; // instructions to be resized
int newOffset; // future offset of a jump instruction
resize = new boolean[code.length];
// 3 = loop again, 2 = loop ended, 1 = last pass, 0 = done
int state = 3;
do {
if (state == 3) {
state = 2;
}
u = 0;
while (u < b.length) {
int opcode = b[u] & 0xFF; // opcode of current instruction
int insert = 0; // bytes to be added after this instruction
switch (ClassWriter.TYPE[opcode]) {
case ClassWriter.NOARG_INSN:
case ClassWriter.IMPLVAR_INSN:
u += 1;
break;
case ClassWriter.LABEL_INSN:
if (opcode > 201) {
// converts temporary opcodes 202 to 217, 218 and
// 219 to IFEQ ... JSR (inclusive), IFNULL and
// IFNONNULL
opcode = opcode < 218 ? opcode - 49 : opcode - 20;
label = u + readUnsignedShort(b, u + 1);
} else {
label = u + readShort(b, u + 1);
}
newOffset = getNewOffset(allIndexes, allSizes, u, label);
if (newOffset < Short.MIN_VALUE
|| newOffset > Short.MAX_VALUE) {
if (!resize[u]) {
if (opcode == Opcodes.GOTO || opcode == Opcodes.JSR) {
// two additional bytes will be required to
// replace this GOTO or JSR instruction with
// a GOTO_W or a JSR_W
insert = 2;
} else {
// five additional bytes will be required to
// replace this IFxxx <l> instruction with
// IFNOTxxx <l'> GOTO_W <l>, where IFNOTxxx
// is the "opposite" opcode of IFxxx (i.e.,
// IFNE for IFEQ) and where <l'> designates
// the instruction just after the GOTO_W.
insert = 5;
}
resize[u] = true;
}
}
u += 3;
break;
case ClassWriter.LABELW_INSN:
u += 5;
break;
case ClassWriter.TABL_INSN:
if (state == 1) {
// true number of bytes to be added (or removed)
// from this instruction = (future number of padding
// bytes - current number of padding byte) -
// previously over estimated variation =
// = ((3 - newOffset%4) - (3 - u%4)) - u%4
// = (-newOffset%4 + u%4) - u%4
// = -(newOffset & 3)
newOffset = getNewOffset(allIndexes, allSizes, 0, u);
insert = -(newOffset & 3);
} else if (!resize[u]) {
// over estimation of the number of bytes to be
// added to this instruction = 3 - current number
// of padding bytes = 3 - (3 - u%4) = u%4 = u & 3
insert = u & 3;
resize[u] = true;
}
// skips instruction
u = u + 4 - (u & 3);
u += 4 * (readInt(b, u + 8) - readInt(b, u + 4) + 1) + 12;
break;
case ClassWriter.LOOK_INSN:
if (state == 1) {
// like TABL_INSN
newOffset = getNewOffset(allIndexes, allSizes, 0, u);
insert = -(newOffset & 3);
} else if (!resize[u]) {
// like TABL_INSN
insert = u & 3;
resize[u] = true;
}
// skips instruction
u = u + 4 - (u & 3);
u += 8 * readInt(b, u + 4) + 8;
break;
case ClassWriter.WIDE_INSN:
opcode = b[u + 1] & 0xFF;
if (opcode == Opcodes.IINC) {
u += 6;
} else {
u += 4;
}
break;
case ClassWriter.VAR_INSN:
case ClassWriter.SBYTE_INSN:
case ClassWriter.LDC_INSN:
u += 2;
break;
case ClassWriter.SHORT_INSN:
case ClassWriter.LDCW_INSN:
case ClassWriter.FIELDORMETH_INSN:
case ClassWriter.TYPE_INSN:
case ClassWriter.IINC_INSN:
u += 3;
break;
case ClassWriter.ITFMETH_INSN:
case ClassWriter.INDYMETH_INSN:
u += 5;
break;
// case ClassWriter.MANA_INSN:
default:
u += 4;
break;
}
if (insert != 0) {
// adds a new (u, insert) entry in the allIndexes and
// allSizes arrays
int[] newIndexes = new int[allIndexes.length + 1];
int[] newSizes = new int[allSizes.length + 1];
System.arraycopy(allIndexes, 0, newIndexes, 0,
allIndexes.length);
System.arraycopy(allSizes, 0, newSizes, 0, allSizes.length);
newIndexes[allIndexes.length] = u;
newSizes[allSizes.length] = insert;
allIndexes = newIndexes;
allSizes = newSizes;
if (insert > 0) {
state = 3;
}
}
}
if (state < 3) {
--state;
}
} while (state != 0);
// 2nd step:
// copies the bytecode of the method into a new bytevector, updates the
// offsets, and inserts (or removes) bytes as requested.
ByteVector newCode = new ByteVector(code.length);
u = 0;
while (u < code.length) {
int opcode = b[u] & 0xFF;
switch (ClassWriter.TYPE[opcode]) {
case ClassWriter.NOARG_INSN:
case ClassWriter.IMPLVAR_INSN:
newCode.putByte(opcode);
u += 1;
break;
case ClassWriter.LABEL_INSN:
if (opcode > 201) {
// changes temporary opcodes 202 to 217 (inclusive), 218
// and 219 to IFEQ ... JSR (inclusive), IFNULL and
// IFNONNULL
opcode = opcode < 218 ? opcode - 49 : opcode - 20;
label = u + readUnsignedShort(b, u + 1);
} else {
label = u + readShort(b, u + 1);
}
newOffset = getNewOffset(allIndexes, allSizes, u, label);
if (resize[u]) {
// replaces GOTO with GOTO_W, JSR with JSR_W and IFxxx
// <l> with IFNOTxxx <l'> GOTO_W <l>, where IFNOTxxx is
// the "opposite" opcode of IFxxx (i.e., IFNE for IFEQ)
// and where <l'> designates the instruction just after
// the GOTO_W.
if (opcode == Opcodes.GOTO) {
newCode.putByte(200); // GOTO_W
} else if (opcode == Opcodes.JSR) {
newCode.putByte(201); // JSR_W
} else {
newCode.putByte(opcode <= 166 ? ((opcode + 1) ^ 1) - 1
: opcode ^ 1);
newCode.putShort(8); // jump offset
newCode.putByte(200); // GOTO_W
// newOffset now computed from start of GOTO_W
newOffset -= 3;
}
newCode.putInt(newOffset);
} else {
newCode.putByte(opcode);
newCode.putShort(newOffset);
}
u += 3;
break;
case ClassWriter.LABELW_INSN:
label = u + readInt(b, u + 1);
newOffset = getNewOffset(allIndexes, allSizes, u, label);
newCode.putByte(opcode);
newCode.putInt(newOffset);
u += 5;
break;
case ClassWriter.TABL_INSN:
// skips 0 to 3 padding bytes
v = u;
u = u + 4 - (v & 3);
// reads and copies instruction
newCode.putByte(Opcodes.TABLESWITCH);
newCode.putByteArray(null, 0, (4 - newCode.length % 4) % 4);
label = v + readInt(b, u);
u += 4;
newOffset = getNewOffset(allIndexes, allSizes, v, label);
newCode.putInt(newOffset);
j = readInt(b, u);
u += 4;
newCode.putInt(j);
j = readInt(b, u) - j + 1;
u += 4;
newCode.putInt(readInt(b, u - 4));
for (; j > 0; --j) {
label = v + readInt(b, u);
u += 4;
newOffset = getNewOffset(allIndexes, allSizes, v, label);
newCode.putInt(newOffset);
}
break;
case ClassWriter.LOOK_INSN:
// skips 0 to 3 padding bytes
v = u;
u = u + 4 - (v & 3);
// reads and copies instruction
newCode.putByte(Opcodes.LOOKUPSWITCH);
newCode.putByteArray(null, 0, (4 - newCode.length % 4) % 4);
label = v + readInt(b, u);
u += 4;
newOffset = getNewOffset(allIndexes, allSizes, v, label);
newCode.putInt(newOffset);
j = readInt(b, u);
u += 4;
newCode.putInt(j);
for (; j > 0; --j) {
newCode.putInt(readInt(b, u));
u += 4;
label = v + readInt(b, u);
u += 4;
newOffset = getNewOffset(allIndexes, allSizes, v, label);
newCode.putInt(newOffset);
}
break;
case ClassWriter.WIDE_INSN:
opcode = b[u + 1] & 0xFF;
if (opcode == Opcodes.IINC) {
newCode.putByteArray(b, u, 6);
u += 6;
} else {
newCode.putByteArray(b, u, 4);
u += 4;
}
break;
case ClassWriter.VAR_INSN:
case ClassWriter.SBYTE_INSN:
case ClassWriter.LDC_INSN:
newCode.putByteArray(b, u, 2);
u += 2;
break;
case ClassWriter.SHORT_INSN:
case ClassWriter.LDCW_INSN:
case ClassWriter.FIELDORMETH_INSN:
case ClassWriter.TYPE_INSN:
case ClassWriter.IINC_INSN:
newCode.putByteArray(b, u, 3);
u += 3;
break;
case ClassWriter.ITFMETH_INSN:
case ClassWriter.INDYMETH_INSN:
newCode.putByteArray(b, u, 5);
u += 5;
break;
// case MANA_INSN:
default:
newCode.putByteArray(b, u, 4);
u += 4;
break;
}
}
// updates the stack map frame labels
if (compute == FRAMES) {
Label l = labels;
while (l != null) {
/*
* Detects the labels that are just after an IF instruction that
* has been resized with the IFNOT GOTO_W pattern. These labels
* are now the target of a jump instruction (the IFNOT
* instruction). Note that we need the original label position
* here. getNewOffset must therefore never have been called for
* this label.
*/
u = l.position - 3;
if (u >= 0 && resize[u]) {
l.status |= Label.TARGET;
}
getNewOffset(allIndexes, allSizes, l);
l = l.successor;
}
// Update the offsets in the uninitialized types
for (i = 0; i < cw.typeTable.length; ++i) {
Item item = cw.typeTable[i];
if (item != null && item.type == ClassWriter.TYPE_UNINIT) {
item.intVal = getNewOffset(allIndexes, allSizes, 0,
item.intVal);
}
}
// The stack map frames are not serialized yet, so we don't need
// to update them. They will be serialized in visitMaxs.
} else if (frameCount > 0) {
/*
* Resizing an existing stack map frame table is really hard. Not
* only the table must be parsed to update the offets, but new
* frames may be needed for jump instructions that were inserted by
* this method. And updating the offsets or inserting frames can
* change the format of the following frames, in case of packed
* frames. In practice the whole table must be recomputed. For this
* the frames are marked as potentially invalid. This will cause the
* whole class to be reread and rewritten with the COMPUTE_FRAMES
* option (see the ClassWriter.toByteArray method). This is not very
* efficient but is much easier and requires much less code than any
* other method I can think of.
*/
cw.invalidFrames = true;
}
// updates the exception handler block labels
Handler h = firstHandler;
while (h != null) {
getNewOffset(allIndexes, allSizes, h.start);
getNewOffset(allIndexes, allSizes, h.end);
getNewOffset(allIndexes, allSizes, h.handler);
h = h.next;
}
// updates the instructions addresses in the
// local var and line number tables
for (i = 0; i < 2; ++i) {
ByteVector bv = i == 0 ? localVar : localVarType;
if (bv != null) {
b = bv.data;
u = 0;
while (u < bv.length) {
label = readUnsignedShort(b, u);
newOffset = getNewOffset(allIndexes, allSizes, 0, label);
writeShort(b, u, newOffset);
label += readUnsignedShort(b, u + 2);
newOffset = getNewOffset(allIndexes, allSizes, 0, label)
- newOffset;
writeShort(b, u + 2, newOffset);
u += 10;
}
}
}
if (lineNumber != null) {
b = lineNumber.data;
u = 0;
while (u < lineNumber.length) {
writeShort(
b,
u,
getNewOffset(allIndexes, allSizes, 0,
readUnsignedShort(b, u)));
u += 4;
}
}
// updates the labels of the other attributes
Attribute attr = cattrs;
while (attr != null) {
Label[] labels = attr.getLabels();
if (labels != null) {
for (i = labels.length - 1; i >= 0; --i) {
getNewOffset(allIndexes, allSizes, labels[i]);
}
}
attr = attr.next;
}
// replaces old bytecodes with new ones
code = newCode;
}