JSFlatString *
JSRope::flattenInternal(ExclusiveContext *maybecx)
{
/*
* Perform a depth-first dag traversal, splatting each node's characters
* into a contiguous buffer. Visit each rope node three times:
* 1. record position in the buffer and recurse into left child;
* 2. recurse into the right child;
* 3. transform the node into a dependent string.
* To avoid maintaining a stack, tree nodes are mutated to indicate how many
* times they have been visited. Since ropes can be dags, a node may be
* encountered multiple times during traversal. However, step 3 above leaves
* a valid dependent string, so everything works out.
*
* While ropes avoid all sorts of quadratic cases with string
* concatenation, they can't help when ropes are immediately flattened.
* One idiomatic case that we'd like to keep linear (and has traditionally
* been linear in SM and other JS engines) is:
*
* while (...) {
* s += ...
* s.flatten
* }
*
* To do this, when the buffer for a to-be-flattened rope is allocated, the
* allocation size is rounded up. Then, if the resulting flat string is the
* left-hand side of a new rope that gets flattened and there is enough
* capacity, the rope is flattened into the same buffer, thereby avoiding
* copying the left-hand side. Clearing the 'extensible' bit turns off this
* optimization. This is necessary, e.g., when the JSAPI hands out the raw
* null-terminated char array of a flat string.
*
* N.B. This optimization can create chains of dependent strings.
*/
const size_t wholeLength = length();
size_t wholeCapacity;
CharT *wholeChars;
JSString *str = this;
CharT *pos;
/*
* JSString::flattenData is a tagged pointer to the parent node.
* The tag indicates what to do when we return to the parent.
*/
static const uintptr_t Tag_Mask = 0x3;
static const uintptr_t Tag_FinishNode = 0x0;
static const uintptr_t Tag_VisitRightChild = 0x1;
AutoCheckCannotGC nogc;
/* Find the left most string, containing the first string. */
JSRope *leftMostRope = this;
while (leftMostRope->leftChild()->isRope())
leftMostRope = &leftMostRope->leftChild()->asRope();
if (leftMostRope->leftChild()->isExtensible()) {
JSExtensibleString &left = leftMostRope->leftChild()->asExtensible();
size_t capacity = left.capacity();
if (capacity >= wholeLength && left.hasTwoByteChars() == IsSame<CharT, jschar>::value) {
/*
* Simulate a left-most traversal from the root to leftMost->leftChild()
* via first_visit_node
*/
JS_ASSERT(str->isRope());
while (str != leftMostRope) {
if (b == WithIncrementalBarrier) {
JSString::writeBarrierPre(str->d.s.u2.left);
JSString::writeBarrierPre(str->d.s.u3.right);
}
JSString *child = str->d.s.u2.left;
JS_ASSERT(child->isRope());
str->setNonInlineChars(left.nonInlineChars<CharT>(nogc));
child->d.u1.flattenData = uintptr_t(str) | Tag_VisitRightChild;
str = child;
}
if (b == WithIncrementalBarrier) {
JSString::writeBarrierPre(str->d.s.u2.left);
JSString::writeBarrierPre(str->d.s.u3.right);
}
str->setNonInlineChars(left.nonInlineChars<CharT>(nogc));
wholeCapacity = capacity;
wholeChars = const_cast<CharT *>(left.nonInlineChars<CharT>(nogc));
pos = wholeChars + left.d.u1.length;
JS_STATIC_ASSERT(!(EXTENSIBLE_FLAGS & DEPENDENT_FLAGS));
left.d.u1.flags ^= (EXTENSIBLE_FLAGS | DEPENDENT_FLAGS);
left.d.s.u3.base = (JSLinearString *)this; /* will be true on exit */
StringWriteBarrierPostRemove(maybecx, &left.d.s.u2.left);
StringWriteBarrierPost(maybecx, (JSString **)&left.d.s.u3.base);
goto visit_right_child;
}
}
if (!AllocChars(maybecx, wholeLength, &wholeChars, &wholeCapacity))
return nullptr;
pos = wholeChars;
first_visit_node: {
if (b == WithIncrementalBarrier) {
JSString::writeBarrierPre(str->d.s.u2.left);
JSString::writeBarrierPre(str->d.s.u3.right);
}
JSString &left = *str->d.s.u2.left;
str->setNonInlineChars(pos);
StringWriteBarrierPostRemove(maybecx, &str->d.s.u2.left);
if (left.isRope()) {
/* Return to this node when 'left' done, then goto visit_right_child. */
left.d.u1.flattenData = uintptr_t(str) | Tag_VisitRightChild;
str = &left;
goto first_visit_node;
}
CopyChars(pos, left.asLinear());
pos += left.length();
}
visit_right_child: {
JSString &right = *str->d.s.u3.right;
if (right.isRope()) {
/* Return to this node when 'right' done, then goto finish_node. */
right.d.u1.flattenData = uintptr_t(str) | Tag_FinishNode;
str = &right;
goto first_visit_node;
}
CopyChars(pos, right.asLinear());
pos += right.length();
}
finish_node: {
if (str == this) {
JS_ASSERT(pos == wholeChars + wholeLength);
*pos = '\0';
str->d.u1.length = wholeLength;
if (IsSame<CharT, jschar>::value)
str->d.u1.flags = EXTENSIBLE_FLAGS;
else
str->d.u1.flags = EXTENSIBLE_FLAGS | LATIN1_CHARS_BIT;
str->setNonInlineChars(wholeChars);
str->d.s.u3.capacity = wholeCapacity;
StringWriteBarrierPostRemove(maybecx, &str->d.s.u2.left);
StringWriteBarrierPostRemove(maybecx, &str->d.s.u3.right);
return &this->asFlat();
}
uintptr_t flattenData = str->d.u1.flattenData;
if (IsSame<CharT, jschar>::value)
str->d.u1.flags = DEPENDENT_FLAGS;
else
str->d.u1.flags = DEPENDENT_FLAGS | LATIN1_CHARS_BIT;
str->d.u1.length = pos - str->asLinear().nonInlineChars<CharT>(nogc);
str->d.s.u3.base = (JSLinearString *)this; /* will be true on exit */
StringWriteBarrierPost(maybecx, (JSString **)&str->d.s.u3.base);
str = (JSString *)(flattenData & ~Tag_Mask);
if ((flattenData & Tag_Mask) == Tag_VisitRightChild)
goto visit_right_child;
JS_ASSERT((flattenData & Tag_Mask) == Tag_FinishNode);
goto finish_node;
}
}
JSFlatString*
JSRope::flattenInternal(ExclusiveContext* maybecx)
{
/*
* Consider the DAG of JSRopes rooted at this JSRope, with non-JSRopes as
* its leaves. Mutate the root JSRope into a JSExtensibleString containing
* the full flattened text that the root represents, and mutate all other
* JSRopes in the interior of the DAG into JSDependentStrings that refer to
* this new JSExtensibleString.
*
* If the leftmost leaf of our DAG is a JSExtensibleString, consider
* stealing its buffer for use in our new root, and transforming it into a
* JSDependentString too. Do not mutate any of the other leaves.
*
* Perform a depth-first dag traversal, splatting each node's characters
* into a contiguous buffer. Visit each rope node three times:
* 1. record position in the buffer and recurse into left child;
* 2. recurse into the right child;
* 3. transform the node into a dependent string.
* To avoid maintaining a stack, tree nodes are mutated to indicate how many
* times they have been visited. Since ropes can be dags, a node may be
* encountered multiple times during traversal. However, step 3 above leaves
* a valid dependent string, so everything works out.
*
* While ropes avoid all sorts of quadratic cases with string concatenation,
* they can't help when ropes are immediately flattened. One idiomatic case
* that we'd like to keep linear (and has traditionally been linear in SM
* and other JS engines) is:
*
* while (...) {
* s += ...
* s.flatten
* }
*
* Two behaviors accomplish this:
*
* - When the leftmost non-rope in the DAG we're flattening is a
* JSExtensibleString with sufficient capacity to hold the entire
* flattened string, we just flatten the DAG into its buffer. Then, when
* we transform the root of the DAG from a JSRope into a
* JSExtensibleString, we steal that buffer, and change the victim from a
* JSExtensibleString to a JSDependentString. In this case, the left-hand
* side of the string never needs to be copied.
*
* - Otherwise, we round up the total flattened size and create a fresh
* JSExtensibleString with that much capacity. If this in turn becomes the
* leftmost leaf of a subsequent flatten, we will hopefully be able to
* fill it, as in the case above.
*
* Note that, even though the code for creating JSDependentStrings avoids
* creating dependents of dependents, we can create that situation here: the
* JSExtensibleStrings we transform into JSDependentStrings might have
* JSDependentStrings pointing to them already. Stealing the buffer doesn't
* change its address, only its owning JSExtensibleString, so all chars()
* pointers in the JSDependentStrings are still valid.
*/
const size_t wholeLength = length();
size_t wholeCapacity;
CharT* wholeChars;
JSString* str = this;
CharT* pos;
/*
* JSString::flattenData is a tagged pointer to the parent node.
* The tag indicates what to do when we return to the parent.
*/
static const uintptr_t Tag_Mask = 0x3;
static const uintptr_t Tag_FinishNode = 0x0;
static const uintptr_t Tag_VisitRightChild = 0x1;
AutoCheckCannotGC nogc;
/* Find the left most string, containing the first string. */
JSRope* leftMostRope = this;
while (leftMostRope->leftChild()->isRope())
leftMostRope = &leftMostRope->leftChild()->asRope();
if (leftMostRope->leftChild()->isExtensible()) {
JSExtensibleString& left = leftMostRope->leftChild()->asExtensible();
size_t capacity = left.capacity();
if (capacity >= wholeLength && left.hasTwoByteChars() == IsSame<CharT, char16_t>::value) {
/*
* Simulate a left-most traversal from the root to leftMost->leftChild()
* via first_visit_node
*/
MOZ_ASSERT(str->isRope());
while (str != leftMostRope) {
if (b == WithIncrementalBarrier) {
JSString::writeBarrierPre(str->d.s.u2.left);
JSString::writeBarrierPre(str->d.s.u3.right);
}
JSString* child = str->d.s.u2.left;
MOZ_ASSERT(child->isRope());
str->setNonInlineChars(left.nonInlineChars<CharT>(nogc));
child->d.u1.flattenData = uintptr_t(str) | Tag_VisitRightChild;
str = child;
}
if (b == WithIncrementalBarrier) {
JSString::writeBarrierPre(str->d.s.u2.left);
JSString::writeBarrierPre(str->d.s.u3.right);
}
str->setNonInlineChars(left.nonInlineChars<CharT>(nogc));
wholeCapacity = capacity;
wholeChars = const_cast<CharT*>(left.nonInlineChars<CharT>(nogc));
pos = wholeChars + left.d.u1.length;
JS_STATIC_ASSERT(!(EXTENSIBLE_FLAGS & DEPENDENT_FLAGS));
left.d.u1.flags ^= (EXTENSIBLE_FLAGS | DEPENDENT_FLAGS);
left.d.s.u3.base = (JSLinearString*)this; /* will be true on exit */
StringWriteBarrierPostRemove(maybecx, &left.d.s.u2.left);
StringWriteBarrierPost(maybecx, (JSString**)&left.d.s.u3.base);
goto visit_right_child;
}
}
if (!AllocChars(this, wholeLength, &wholeChars, &wholeCapacity))
return nullptr;
pos = wholeChars;
first_visit_node: {
if (b == WithIncrementalBarrier) {
JSString::writeBarrierPre(str->d.s.u2.left);
JSString::writeBarrierPre(str->d.s.u3.right);
}
JSString& left = *str->d.s.u2.left;
str->setNonInlineChars(pos);
StringWriteBarrierPostRemove(maybecx, &str->d.s.u2.left);
if (left.isRope()) {
/* Return to this node when 'left' done, then goto visit_right_child. */
left.d.u1.flattenData = uintptr_t(str) | Tag_VisitRightChild;
str = &left;
goto first_visit_node;
}
CopyChars(pos, left.asLinear());
pos += left.length();
}
visit_right_child: {
JSString& right = *str->d.s.u3.right;
if (right.isRope()) {
/* Return to this node when 'right' done, then goto finish_node. */
right.d.u1.flattenData = uintptr_t(str) | Tag_FinishNode;
str = &right;
goto first_visit_node;
}
CopyChars(pos, right.asLinear());
pos += right.length();
}
finish_node: {
if (str == this) {
MOZ_ASSERT(pos == wholeChars + wholeLength);
*pos = '\0';
str->d.u1.length = wholeLength;
if (IsSame<CharT, char16_t>::value)
str->d.u1.flags = EXTENSIBLE_FLAGS;
else
str->d.u1.flags = EXTENSIBLE_FLAGS | LATIN1_CHARS_BIT;
str->setNonInlineChars(wholeChars);
str->d.s.u3.capacity = wholeCapacity;
StringWriteBarrierPostRemove(maybecx, &str->d.s.u2.left);
StringWriteBarrierPostRemove(maybecx, &str->d.s.u3.right);
return &this->asFlat();
}
uintptr_t flattenData = str->d.u1.flattenData;
if (IsSame<CharT, char16_t>::value)
str->d.u1.flags = DEPENDENT_FLAGS;
else
str->d.u1.flags = DEPENDENT_FLAGS | LATIN1_CHARS_BIT;
str->d.u1.length = pos - str->asLinear().nonInlineChars<CharT>(nogc);
str->d.s.u3.base = (JSLinearString*)this; /* will be true on exit */
StringWriteBarrierPost(maybecx, (JSString**)&str->d.s.u3.base);
str = (JSString*)(flattenData & ~Tag_Mask);
if ((flattenData & Tag_Mask) == Tag_VisitRightChild)
goto visit_right_child;
MOZ_ASSERT((flattenData & Tag_Mask) == Tag_FinishNode);
goto finish_node;
}
}
