// The Store can be captured only if nothing after the allocation
// and before the Store is using the memory location that the store
// overwrites.
bool failed = false;
// If is_complete_with_arraycopy() is true the shape of the graph is
// well defined and is safe so no need for extra checks.
if (!is_complete_with_arraycopy()) {
// We are going to look at each use of the memory state following
// the allocation to make sure nothing reads the memory that the
// Store writes.
const TypePtr* t_adr = phase->type(adr)->isa_ptr();
int alias_idx = phase->C->get_alias_index(t_adr);
ResourceMark rm;
Unique_Node_List mems;
mems.push(mem);
Node* unique_merge = NULL;
for (uint next = 0; next < mems.size(); ++next) {
Node *m = mems.at(next);
for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
Node *n = m->fast_out(j);
if (n->outcnt() == 0) {
continue;
}
if (n == st) {
continue;
} else if (n->in(0) != NULL && n->in(0) != ctl) {
// If the control of this use is different from the control
// of the Store which is right after the InitializeNode then
// this node cannot be between the InitializeNode and the
// Store.
continue;
} else if (n->is_MergeMem()) {
if (n->as_MergeMem()->memory_at(alias_idx) == m) {
// We can hit a MergeMemNode (that will likely go away
// later) that is a direct use of the memory state
// following the InitializeNode on the same slice as the
// store node that we'd like to capture. We need to check
// the uses of the MergeMemNode.
mems.push(n);
}
} else if (n->is_Mem()) {
Node* other_adr = n->in(MemNode::Address);
if (other_adr == adr) {
failed = true;
break;
} else {
const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
if (other_t_adr != NULL) {
int other_alias_idx = phase->C->get_alias_index(other_t_adr);
if (other_alias_idx == alias_idx) {
// A load from the same memory slice as the store right
// after the InitializeNode. We check the control of the
// object/array that is loaded from. If it's the same as
// the store control then we cannot capture the store.
assert(!n->is_Store(), "2 stores to same slice on same control?");
Node* base = other_adr;
assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name()));
base = base->in(AddPNode::Base);
if (base != NULL) {
base = base->uncast();
if (base->is_Proj() && base->in(0) == alloc) {
failed = true;
break;
}
}
}
}
}
} else {
failed = true;
break;
}
}
}
}
if (failed) {
if (!can_reshape) {
// We decided we couldn't capture the store during parsing. We
// should try again during the next IGVN once the graph is
// cleaner.
phase->C->record_for_igvn(st);
}
return FAIL;
}
return offset; // success
}
// Find the captured store in(i) which corresponds to the range
// [start..start+size) in the initialized object.
// If there is one, return its index i. If there isn't, return the
// negative of the index where it should be inserted.
// Return 0 if the queried range overlaps an initialization boundary
// or if dead code is encountered.
// If size_in_bytes is zero, do not bother with overlap checks.
int InitializeNode::captured_store_insertion_point(intptr_t start,
int size_in_bytes,
PhaseTransform* phase) {
const int FAIL = 0, MAX_STORE = BytesPerLong;
if (is_complete())
return FAIL; // arraycopy got here first; punt
assert(allocation() != NULL, "must be present");
// no negatives, no header fields:
if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
// after a certain size, we bail out on tracking all the stores:
intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
if (start >= ti_limit) return FAIL;
for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
if (i >= limit) return -(int)i; // not found; here is where to put it
Node* st = in(i);
intptr_t st_off = get_store_offset(st, phase);
if (st_off < 0) {
if (st != zero_memory()) {
return FAIL; // bail out if there is dead garbage
}
} else if (st_off > start) {
// ...we are done, since stores are ordered
if (st_off < start + size_in_bytes) {
return FAIL; // the next store overlaps
}
return -(int)i; // not found; here is where to put it
} else if (st_off < start) {
if (size_in_bytes != 0 &&
start < st_off + MAX_STORE &&
start < st_off + st->as_Store()->memory_size()) {
return FAIL; // the previous store overlaps
}
} else {
if (size_in_bytes != 0 &&
st->as_Store()->memory_size() != size_in_bytes) {
return FAIL; // mismatched store size
}
return i;
}
++i;
}
}
// Look for a captured store which initializes at the offset 'start'
// with the given size. If there is no such store, and no other
// initialization interferes, then return zero_memory (the memory
// projection of the AllocateNode).
Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
PhaseTransform* phase) {
assert(stores_are_sane(phase), "");
int i = captured_store_insertion_point(start, size_in_bytes, phase);
if (i == 0) {
return NULL; // something is dead
} else if (i < 0) {
return zero_memory(); // just primordial zero bits here
} else {
Node* st = in(i); // here is the store at this position
assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
return st;
}
}
// Create, as a raw pointer, an address within my new object at 'offset'.
Node* InitializeNode::make_raw_address(intptr_t offset,
PhaseTransform* phase) {
Node* addr = in(RawAddress);
if (offset != 0) {
Compile* C = phase->C;
addr = phase->transform( new (C) AddPNode(C->top(), addr,
phase->MakeConX(offset)) );
}
return addr;
}
// Clone the given store, converting it into a raw store
// initializing a field or element of my new object.
// Caller is responsible for retiring the original store,
// with subsume_node or the like.
//
// From the example above InitializeNode::InitializeNode,
// here are the old stores to be captured:
// store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
// store2 = (StoreC init.Control store1 (+ oop 14) 2)
//
// Here is the changed code; note the extra edges on init:
// alloc = (Allocate ...)
// rawoop = alloc.RawAddress
// rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
// rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
// init = (Initialize alloc.Control alloc.Memory rawoop
// rawstore1 rawstore2)
//
Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
PhaseTransform* phase, bool can_reshape) {
assert(stores_are_sane(phase), "");
if (start < 0) return NULL;
assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
Compile* C = phase->C;
int size_in_bytes = st->memory_size();
int i = captured_store_insertion_point(start, size_in_bytes, phase);
if (i == 0) return NULL; // bail out
Node* prev_mem = NULL; // raw memory for the captured store
if (i > 0) {
prev_mem = in(i); // there is a pre-existing store under this one
set_req(i, C->top()); // temporarily disconnect it
// See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
} else {
i = -i; // no pre-existing store
prev_mem = zero_memory(); // a slice of the newly allocated object
if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
set_req(--i, C->top()); // reuse this edge; it has been folded away
else
ins_req(i, C->top()); // build a new edge
}
Node* new_st = st->clone();
new_st->set_req(MemNode::Control, in(Control));
new_st->set_req(MemNode::Memory, prev_mem);
new_st->set_req(MemNode::Address, make_raw_address(start, phase));
new_st = phase->transform(new_st);
// At this point, new_st might have swallowed a pre-existing store
// at the same offset, or perhaps new_st might have disappeared,
// if it redundantly stored the same value (or zero to fresh memory).
// In any case, wire it in:
set_req(i, new_st);
// The caller may now kill the old guy.
DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
assert(check_st == new_st || check_st == NULL, "must be findable");
assert(!is_complete(), "");
return new_st;
}
static bool store_constant(jlong* tiles, int num_tiles,
intptr_t st_off, int st_size,
jlong con) {
if ((st_off & (st_size-1)) != 0)
return false; // strange store offset (assume size==2**N)
address addr = (address)tiles + st_off;
assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
switch (st_size) {
case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
case sizeof(jint): *(jint*) addr = (jint) con; break;
case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
default: return false; // strange store size (detect size!=2**N here)
}
return true; // return success to caller
}
// Coalesce subword constants into int constants and possibly
// into long constants. The goal, if the CPU permits,
// is to initialize the object with a small number of 64-bit tiles.
// Also, convert floating-point constants to bit patterns.
// Non-constants are not relevant to this pass.
//
// In terms of the running example on InitializeNode::InitializeNode
// and InitializeNode::capture_store, here is the transformation
// of rawstore1 and rawstore2 into rawstore12:
// alloc = (Allocate ...)
// rawoop = alloc.RawAddress
// tile12 = 0x00010002
// rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
// init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
//
void
InitializeNode::coalesce_subword_stores(intptr_t header_size,
Node* size_in_bytes,
PhaseGVN* phase) {
Compile* C = phase->C;
assert(stores_are_sane(phase), "");
// Note: After this pass, they are not completely sane,
// since there may be some overlaps.
int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
size_limit = MIN2(size_limit, ti_limit);
size_limit = align_size_up(size_limit, BytesPerLong);
int num_tiles = size_limit / BytesPerLong;
// allocate space for the tile map:
const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
jlong tiles_buf[small_len];
Node* nodes_buf[small_len];
jlong inits_buf[small_len];
jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
: NEW_RESOURCE_ARRAY(jlong, num_tiles));
Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
: NEW_RESOURCE_ARRAY(Node*, num_tiles));
jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
: NEW_RESOURCE_ARRAY(jlong, num_tiles));
// tiles: exact bitwise model of all primitive constants
// nodes: last constant-storing node subsumed into the tiles model
// inits: which bytes (in each tile) are touched by any initializations
Pass A: Fill in the tile model with any relevant stores.
Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
Node* zmem = zero_memory(); // initially zero memory state
for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
Node* st = in(i);
intptr_t st_off = get_store_offset(st, phase);
// Figure out the store's offset and constant value:
if (st_off < header_size) continue; //skip (ignore header)
if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
int st_size = st->as_Store()->memory_size();
if (st_off + st_size > size_limit) break;
// Record which bytes are touched, whether by constant or not.
if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
continue; // skip (strange store size)
const Type* val = phase->type(st->in(MemNode::ValueIn));
if (!val->singleton()) continue; //skip (non-con store)
BasicType type = val->basic_type();
jlong con = 0;
switch (type) {
case T_INT: con = val->is_int()->get_con(); break;
case T_LONG: con = val->is_long()->get_con(); break;
case T_FLOAT: con = jint_cast(val->getf()); break;
case T_DOUBLE: con = jlong_cast(val->getd()); break;
default: continue; //skip (odd store type)
}
if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
st->Opcode() == Op_StoreL) {
continue; // This StoreL is already optimal.
}
// Store down the constant.
store_constant(tiles, num_tiles, st_off, st_size, con);
intptr_t j = st_off >> LogBytesPerLong;
if (type == T_INT && st_size == BytesPerInt
&& (st_off & BytesPerInt) == BytesPerInt) {
jlong lcon = tiles[j];
if (!Matcher::isSimpleConstant64(lcon) &&
st->Opcode() == Op_StoreI) {
// This StoreI is already optimal by itself.
jint* intcon = (jint*) &tiles[j];
intcon[1] = 0; // undo the store_constant()
// If the previous store is also optimal by itself, back up and
// undo the action of the previous loop iteration... if we can.
// But if we can't, just let the previous half take care of itself.
st = nodes[j];
st_off -= BytesPerInt;
con = intcon[0];
if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
assert(st_off >= header_size, "still ignoring header");
assert(get_store_offset(st, phase) == st_off, "must be");
assert(in(i-1) == zmem, "must be");
DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
assert(con == tcon->is_int()->get_con(), "must be");
// Undo the effects of the previous loop trip, which swallowed st:
intcon[0] = 0; // undo store_constant()
set_req(i-1, st); // undo set_req(i, zmem)
nodes[j] = NULL; // undo nodes[j] = st
--old_subword; // undo ++old_subword
}
continue; // This StoreI is already optimal.
}
}
// This store is not needed.
set_req(i, zmem);
nodes[j] = st; // record for the moment
if (st_size < BytesPerLong) // something has changed
++old_subword; // includes int/float, but who's counting...
else ++old_long;
}
if ((old_subword + old_long) == 0)
return; // nothing more to do
Pass B: Convert any non-zero tiles into optimal constant stores.
// Be sure to insert them before overlapping non-constant stores.
// (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
for (int j = 0; j < num_tiles; j++) {
jlong con = tiles[j];
jlong init = inits[j];
if (con == 0) continue;
jint con0, con1; // split the constant, address-wise
jint init0, init1; // split the init map, address-wise
{ union { jlong con; jint intcon[2]; } u;
u.con = con;
con0 = u.intcon[0];
con1 = u.intcon[1];
u.con = init;
init0 = u.intcon[0];
init1 = u.intcon[1];
}
Node* old = nodes[j];
assert(old != NULL, "need the prior store");
intptr_t offset = (j * BytesPerLong);
bool split = !Matcher::isSimpleConstant64(con);
if (offset < header_size) {
assert(offset + BytesPerInt >= header_size, "second int counts");
assert(*(jint*)&tiles[j] == 0, "junk in header");
split = true; // only the second word counts
// Example: int a[] = { 42 ... }
} else if (con0 == 0 && init0 == -1) {
split = true; // first word is covered by full inits
// Example: int a[] = { ... foo(), 42 ... }
} else if (con1 == 0 && init1 == -1) {
split = true; // second word is covered by full inits
// Example: int a[] = { ... 42, foo() ... }
}
// Here's a case where init0 is neither 0 nor -1:
// byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
// Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
// In this case the tile is not split; it is (jlong)42.
// The big tile is stored down, and then the foo() value is inserted.
// (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
Node* ctl = old->in(MemNode::Control);
Node* adr = make_raw_address(offset, phase);
const TypePtr* atp = TypeRawPtr::BOTTOM;
// One or two coalesced stores to plop down.
Node* st[2];
intptr_t off[2];
int nst = 0;
if (!split) {
++new_long;
off[nst] = offset;
st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
phase->longcon(con), T_LONG, MemNode::unordered);
} else {
// Omit either if it is a zero.
if (con0 != 0) {
++new_int;
off[nst] = offset;
st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
phase->intcon(con0), T_INT, MemNode::unordered);
}
if (con1 != 0) {
++new_int;
offset += BytesPerInt;
adr = make_raw_address(offset, phase);
off[nst] = offset;
st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
phase->intcon(con1), T_INT, MemNode::unordered);
}
}
// Insert second store first, then the first before the second.
// Insert each one just before any overlapping non-constant stores.
while (nst > 0) {
Node* st1 = st[--nst];
C->copy_node_notes_to(st1, old);
st1 = phase->transform(st1);
offset = off[nst];
assert(offset >= header_size, "do not smash header");
int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
guarantee(ins_idx != 0, "must re-insert constant store");
if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
set_req(--ins_idx, st1);
else
ins_req(ins_idx, st1);
}
}
if (PrintCompilation && WizardMode)
tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
old_subword, old_long, new_int, new_long);
if (C->log() != NULL)
C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
old_subword, old_long, new_int, new_long);
// Clean up any remaining occurrences of zmem:
remove_extra_zeroes();
}
// Explore forward from in(start) to find the first fully initialized
// word, and return its offset. Skip groups of subword stores which
// together initialize full words. If in(start) is itself part of a
// fully initialized word, return the offset of in(start). If there
// are no following full-word stores, or if something is fishy, return
// a negative value.
intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
int int_map = 0;
intptr_t int_map_off = 0;
const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
for (uint i = start, limit = req(); i < limit; i++) {
Node* st = in(i);
intptr_t st_off = get_store_offset(st, phase);
if (st_off < 0) break; // return conservative answer
int st_size = st->as_Store()->memory_size();
if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
return st_off; // we found a complete word init
}
// update the map:
intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
if (this_int_off != int_map_off) {
// reset the map:
int_map = 0;
int_map_off = this_int_off;
}
int subword_off = st_off - this_int_off;
int_map |= right_n_bits(st_size) << subword_off;
if ((int_map & FULL_MAP) == FULL_MAP) {
return this_int_off; // we found a complete word init
}
// Did this store hit or cross the word boundary?
intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
if (next_int_off == this_int_off + BytesPerInt) {
// We passed the current int, without fully initializing it.
int_map_off = next_int_off;
int_map >>= BytesPerInt;
} else if (next_int_off > this_int_off + BytesPerInt) {
// We passed the current and next int.
return this_int_off + BytesPerInt;
}
}
return -1;
}
// Called when the associated AllocateNode is expanded into CFG.
// At this point, we may perform additional optimizations.
// Linearize the stores by ascending offset, to make memory
// activity as coherent as possible.
Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
intptr_t header_size,
Node* size_in_bytes,
PhaseGVN* phase) {
assert(!is_complete(), "not already complete");
assert(stores_are_sane(phase), "");
assert(allocation() != NULL, "must be present");
remove_extra_zeroes();
if (ReduceFieldZeroing || ReduceBulkZeroing)
// reduce instruction count for common initialization patterns
coalesce_subword_stores(header_size, size_in_bytes, phase);
Node* zmem = zero_memory(); // initially zero memory state
Node* inits = zmem; // accumulating a linearized chain of inits
#ifdef ASSERT
intptr_t first_offset = allocation()->minimum_header_size();
intptr_t last_init_off = first_offset; // previous init offset
intptr_t last_init_end = first_offset; // previous init offset+size
intptr_t last_tile_end = first_offset; // previous tile offset+size
#endif
intptr_t zeroes_done = header_size;
bool do_zeroing = true; // we might give up if inits are very sparse
int big_init_gaps = 0; // how many large gaps have we seen?
if (ZeroTLAB) do_zeroing = false;
if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
Node* st = in(i);
intptr_t st_off = get_store_offset(st, phase);
if (st_off < 0)
break; // unknown junk in the inits
if (st->in(MemNode::Memory) != zmem)
break; // complicated store chains somehow in list
int st_size = st->as_Store()->memory_size();
intptr_t next_init_off = st_off + st_size;
if (do_zeroing && zeroes_done < next_init_off) {
// See if this store needs a zero before it or under it.
intptr_t zeroes_needed = st_off;
if (st_size < BytesPerInt) {
// Look for subword stores which only partially initialize words.
// If we find some, we must lay down some word-level zeroes first,
// underneath the subword stores.
//
// Examples:
// byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
// byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
// byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
//
// Note: coalesce_subword_stores may have already done this,
// if it was prompted by constant non-zero subword initializers.
// But this case can still arise with non-constant stores.
intptr_t next_full_store = find_next_fullword_store(i, phase);
// In the examples above:
// in(i) p q r s x y z
// st_off 12 13 14 15 12 13 14
// st_size 1 1 1 1 1 1 1
// next_full_s. 12 16 16 16 16 16 16
// z's_done 12 16 16 16 12 16 12
// z's_needed 12 16 16 16 16 16 16
// zsize 0 0 0 0 4 0 4
if (next_full_store < 0) {
// Conservative tack: Zero to end of current word.
zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
} else {
// Zero to beginning of next fully initialized word.
// Or, don't zero at all, if we are already in that word.
assert(next_full_store >= zeroes_needed, "must go forward");
assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
zeroes_needed = next_full_store;
}
}
if (zeroes_needed > zeroes_done) {
intptr_t zsize = zeroes_needed - zeroes_done;
// Do some incremental zeroing on rawmem, in parallel with inits.
zeroes_done = align_size_down(zeroes_done, BytesPerInt);
rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
zeroes_done, zeroes_needed,
phase);
zeroes_done = zeroes_needed;
if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
do_zeroing = false; // leave the hole, next time
}
}
// Collect the store and move on:
st->set_req(MemNode::Memory, inits);
inits = st; // put it on the linearized chain
set_req(i, zmem); // unhook from previous position
if (zeroes_done == st_off)
zeroes_done = next_init_off;
assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
#ifdef ASSERT
// Various order invariants. Weaker than stores_are_sane because
// a large constant tile can be filled in by smaller non-constant stores.
assert(st_off >= last_init_off, "inits do not reverse");
last_init_off = st_off;
const Type* val = NULL;
if (st_size >= BytesPerInt &&
(val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
(int)val->basic_type() < (int)T_OBJECT) {
assert(st_off >= last_tile_end, "tiles do not overlap");
assert(st_off >= last_init_end, "tiles do not overwrite inits");
last_tile_end = MAX2(last_tile_end, next_init_off);
} else {
intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
assert(st_tile_end >= last_tile_end, "inits stay with tiles");
assert(st_off >= last_init_end, "inits do not overlap");
last_init_end = next_init_off; // it's a non-tile
}
#endif //ASSERT
}
remove_extra_zeroes(); // clear out all the zmems left over
add_req(inits);
if (!ZeroTLAB) {
// If anything remains to be zeroed, zero it all now.
zeroes_done = align_size_down(zeroes_done, BytesPerInt);
// if it is the last unused 4 bytes of an instance, forget about it
intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
if (zeroes_done + BytesPerLong >= size_limit) {
AllocateNode* alloc = allocation();
assert(alloc != NULL, "must be present");
if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
Node* klass_node = alloc->in(AllocateNode::KlassNode);
ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
if (zeroes_done == k->layout_helper())
zeroes_done = size_limit;
}
}
if (zeroes_done < size_limit) {
rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
zeroes_done, size_in_bytes, phase);
}
}
set_complete(phase);
return rawmem;
}
#ifdef ASSERT
bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
if (is_complete())
return true; // stores could be anything at this point
assert(allocation() != NULL, "must be present");
intptr_t last_off = allocation()->minimum_header_size();
for (uint i = InitializeNode::RawStores; i < req(); i++) {
Node* st = in(i);
intptr_t st_off = get_store_offset(st, phase);
if (st_off < 0) continue; // ignore dead garbage
if (last_off > st_off) {
tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
this->dump(2);
assert(false, "ascending store offsets");
return false;
}
last_off = st_off + st->as_Store()->memory_size();
}
return true;
}
#endif //ASSERT
//============================MergeMemNode=====================================
//
// SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
// contributing store or call operations. Each contributor provides the memory
// state for a particular "alias type" (see Compile::alias_type). For example,
// if a MergeMem has an input X for alias category #6, then any memory reference
// to alias category #6 may use X as its memory state input, as an exact equivalent
// to using the MergeMem as a whole.
// Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
//
// (Here, the <N> notation gives the index of the relevant adr_type.)
//
// In one special case (and more cases in the future), alias categories overlap.
// The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
// states. Therefore, if a MergeMem has only one contributing input W for Bot,
// it is exactly equivalent to that state W:
// MergeMem(<Bot>: W) <==> W
//
// Usually, the merge has more than one input. In that case, where inputs
// overlap (i.e., one is Bot), the narrower alias type determines the memory
// state for that type, and the wider alias type (Bot) fills in everywhere else:
// Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
// Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
//
// A merge can take a "wide" memory state as one of its narrow inputs.
// This simply means that the merge observes out only the relevant parts of
// the wide input. That is, wide memory states arriving at narrow merge inputs
// are implicitly "filtered" or "sliced" as necessary. (This is rare.)
//
// These rules imply that MergeMem nodes may cascade (via their <Bot> links),
// and that memory slices "leak through":
// MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
//
// But, in such a cascade, repeated memory slices can "block the leak":
// MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
//
// In the last example, Y is not part of the combined memory state of the
// outermost MergeMem. The system must, of course, prevent unschedulable
// memory states from arising, so you can be sure that the state Y is somehow
// a precursor to state Y'.
//
//
// REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
// of each MergeMemNode array are exactly the numerical alias indexes, including
// but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
// Compile::alias_type (and kin) produce and manage these indexes.
//
// By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
// (Note that this provides quick access to the top node inside MergeMem methods,
// without the need to reach out via TLS to Compile::current.)
//
// As a consequence of what was just described, a MergeMem that represents a full
// memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
// containing all alias categories.
//
// MergeMem nodes never (?) have control inputs, so in(0) is NULL.
//
// All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
// a memory state for the alias type <N>, or else the top node, meaning that
// there is no particular input for that alias type. Note that the length of
// a MergeMem is variable, and may be extended at any time to accommodate new
// memory states at larger alias indexes. When merges grow, they are of course
// filled with "top" in the unused in() positions.
//
// This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
// (Top was chosen because it works smoothly with passes like GCM.)
//
// For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
// the type of random VM bits like TLS references.) Since it is always the
// first non-Bot memory slice, some low-level loops use it to initialize an
// index variable: for (i = AliasIdxRaw; i < req(); i++).
//
//
// ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
// the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
// the memory state for alias type <N>, or (if there is no particular slice at <N>,
// it returns the base memory. To prevent bugs, memory_at does not accept <Top>
// or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
// MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
//
// %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
// really that different from the other memory inputs. An abbreviation called
// "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
//
//
// PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
// partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
// that "emerges though" the base memory will be marked as excluding the alias types
// of the other (narrow-memory) copies which "emerged through" the narrow edges:
//
// Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
// ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
//
// This strange "subtraction" effect is necessary to ensure IGVN convergence.
// (It is currently unimplemented.) As you can see, the resulting merge is
// actually a disjoint union of memory states, rather than an overlay.
//
//------------------------------MergeMemNode-----------------------------------
Node* MergeMemNode::make_empty_memory() {
Node* empty_memory = (Node*) Compile::current()->top();
assert(empty_memory->is_top(), "correct sentinel identity");
return empty_memory;
}
MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
init_class_id(Class_MergeMem);
// all inputs are nullified in Node::Node(int)
// set_input(0, NULL); // no control input
// Initialize the edges uniformly to top, for starters.
Node* empty_mem = make_empty_memory();
for (uint i = Compile::AliasIdxTop; i < req(); i++) {
init_req(i,empty_mem);
}
assert(empty_memory() == empty_mem, "");
if( new_base != NULL && new_base->is_MergeMem() ) {
MergeMemNode* mdef = new_base->as_MergeMem();
assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
mms.set_memory(mms.memory2());
}
assert(base_memory() == mdef->base_memory(), "");
} else {
set_base_memory(new_base);
}
}
// Make a new, untransformed MergeMem with the same base as 'mem'.
// If mem is itself a MergeMem, populate the result with the same edges.
MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
return new(C) MergeMemNode(mem);
}
//------------------------------cmp--------------------------------------------
uint MergeMemNode::hash() const { return NO_HASH; }
uint MergeMemNode::cmp( const Node &n ) const {
return (&n == this); // Always fail except on self
}
//------------------------------Identity---------------------------------------
Node* MergeMemNode::Identity(PhaseTransform *phase) {
// Identity if this merge point does not record any interesting memory
// disambiguations.
Node* base_mem = base_memory();
Node* empty_mem = empty_memory();
if (base_mem != empty_mem) { // Memory path is not dead?
for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
Node* mem = in(i);
if (mem != empty_mem && mem != base_mem) {
return this; // Many memory splits; no change
}
}
}
return base_mem; // No memory splits; ID on the one true input
}
//------------------------------Ideal------------------------------------------
// This method is invoked recursively on chains of MergeMem nodes
Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
// Remove chain'd MergeMems
//
// This is delicate, because the each "in(i)" (i >= Raw) is interpreted
// relative to the "in(Bot)". Since we are patching both at the same time,
// we have to be careful to read each "in(i)" relative to the old "in(Bot)",
// but rewrite each "in(i)" relative to the new "in(Bot)".
Node *progress = NULL;
Node* old_base = base_memory();
Node* empty_mem = empty_memory();
if (old_base == empty_mem)
return NULL; // Dead memory path.
MergeMemNode* old_mbase;
if (old_base != NULL && old_base->is_MergeMem())
old_mbase = old_base->as_MergeMem();
else
old_mbase = NULL;
Node* new_base = old_base;
// simplify stacked MergeMems in base memory
if (old_mbase) new_base = old_mbase->base_memory();
// the base memory might contribute new slices beyond my req()
if (old_mbase) grow_to_match(old_mbase);
// Look carefully at the base node if it is a phi.
PhiNode* phi_base;
if (new_base != NULL && new_base->is_Phi())
phi_base = new_base->as_Phi();
else
phi_base = NULL;
Node* phi_reg = NULL;
uint phi_len = (uint)-1;
if (phi_base != NULL && !phi_base->is_copy()) {
// do not examine phi if degraded to a copy
phi_reg = phi_base->region();
phi_len = phi_base->req();
// see if the phi is unfinished
for (uint i = 1; i < phi_len; i++) {
if (phi_base->in(i) == NULL) {
// incomplete phi; do not look at it yet!
phi_reg = NULL;
phi_len = (uint)-1;
break;
}
}
}
// Note: We do not call verify_sparse on entry, because inputs
// can normalize to the base_memory via subsume_node or similar
// mechanisms. This method repairs that damage.
assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
// Look at each slice.
for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
Node* old_in = in(i);
// calculate the old memory value
Node* old_mem = old_in;
if (old_mem == empty_mem) old_mem = old_base;
assert(old_mem == memory_at(i), "");
// maybe update (reslice) the old memory value
// simplify stacked MergeMems
Node* new_mem = old_mem;
MergeMemNode* old_mmem;
if (old_mem != NULL && old_mem->is_MergeMem())
old_mmem = old_mem->as_MergeMem();
else
old_mmem = NULL;
if (old_mmem == this) {
// This can happen if loops break up and safepoints disappear.
// A merge of BotPtr (default) with a RawPtr memory derived from a
// safepoint can be rewritten to a merge of the same BotPtr with
// the BotPtr phi coming into the loop. If that phi disappears
// also, we can end up with a self-loop of the mergemem.
// In general, if loops degenerate and memory effects disappear,
// a mergemem can be left looking at itself. This simply means
// that the mergemem's default should be used, since there is
// no longer any apparent effect on this slice.
// Note: If a memory slice is a MergeMem cycle, it is unreachable
// from start. Update the input to TOP.
new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
}
else if (old_mmem != NULL) {
new_mem = old_mmem->memory_at(i);
}
// else preceding memory was not a MergeMem
// replace equivalent phis (unfortunately, they do not GVN together)
if (new_mem != NULL && new_mem != new_base &&
new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
if (new_mem->is_Phi()) {
PhiNode* phi_mem = new_mem->as_Phi();
for (uint i = 1; i < phi_len; i++) {
if (phi_base->in(i) != phi_mem->in(i)) {
phi_mem = NULL;
break;
}
}
if (phi_mem != NULL) {
// equivalent phi nodes; revert to the def
new_mem = new_base;
}
}
}
// maybe store down a new value
Node* new_in = new_mem;
if (new_in == new_base) new_in = empty_mem;
if (new_in != old_in) {
// Warning: Do not combine this "if" with the previous "if"
// A memory slice might have be be rewritten even if it is semantically
// unchanged, if the base_memory value has changed.
set_req(i, new_in);
progress = this; // Report progress
}
}
if (new_base != old_base) {
set_req(Compile::AliasIdxBot, new_base);
// Don't use set_base_memory(new_base), because we need to update du.
assert(base_memory() == new_base, "");
progress = this;
}
if( base_memory() == this ) {
// a self cycle indicates this memory path is dead
set_req(Compile::AliasIdxBot, empty_mem);
}
// Resolve external cycles by calling Ideal on a MergeMem base_memory
// Recursion must occur after the self cycle check above
if( base_memory()->is_MergeMem() ) {
MergeMemNode *new_mbase = base_memory()->as_MergeMem();
Node *m = phase->transform(new_mbase); // Rollup any cycles
if( m != NULL && (m->is_top() ||
m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
// propagate rollup of dead cycle to self
set_req(Compile::AliasIdxBot, empty_mem);
}
}
if( base_memory() == empty_mem ) {
progress = this;
// Cut inputs during Parse phase only.
// During Optimize phase a dead MergeMem node will be subsumed by Top.
if( !can_reshape ) {
for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
if( in(i) != empty_mem ) { set_req(i, empty_mem); }
}
}
}
if( !progress && base_memory()->is_Phi() && can_reshape ) {
// Check if PhiNode::Ideal's "Split phis through memory merges"
// transform should be attempted. Look for this->phi->this cycle.
uint merge_width = req();
if (merge_width > Compile::AliasIdxRaw) {
PhiNode* phi = base_memory()->as_Phi();
for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
if (phi->in(i) == this) {
phase->is_IterGVN()->_worklist.push(phi);
break;
}
}
}
}
assert(progress || verify_sparse(), "please, no dups of base");
return progress;
}
//-------------------------set_base_memory-------------------------------------
void MergeMemNode::set_base_memory(Node *new_base) {
Node* empty_mem = empty_memory();
set_req(Compile::AliasIdxBot, new_base);
assert(memory_at(req()) == new_base, "must set default memory");
// Clear out other occurrences of new_base:
if (new_base != empty_mem) {
for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
if (in(i) == new_base) set_req(i, empty_mem);
}
}
}
//------------------------------out_RegMask------------------------------------
const RegMask &MergeMemNode::out_RegMask() const {
return RegMask::Empty;
}
//------------------------------dump_spec--------------------------------------
#ifndef PRODUCT
void MergeMemNode::dump_spec(outputStream *st) const {
st->print(" {");
Node* base_mem = base_memory();
for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
Node* mem = memory_at(i);
if (mem == base_mem) { st->print(" -"); continue; }
st->print( " N%d:", mem->_idx );
Compile::current()->get_adr_type(i)->dump_on(st);
}
st->print(" }");
}
#endif // !PRODUCT
#ifdef ASSERT
static bool might_be_same(Node* a, Node* b) {
if (a == b) return true;
if (!(a->is_Phi() || b->is_Phi())) return false;
// phis shift around during optimization
return true; // pretty stupid...
}
// verify a narrow slice (either incoming or outgoing)
static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
if (!VerifyAliases) return; // don't bother to verify unless requested
if (is_error_reported()) return; // muzzle asserts when debugging an error
if (Node::in_dump()) return; // muzzle asserts when printing
assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
assert(n != NULL, "");
// Elide intervening MergeMem's
while (n->is_MergeMem()) {
n = n->as_MergeMem()->memory_at(alias_idx);
}
Compile* C = Compile::current();
const TypePtr* n_adr_type = n->adr_type();
if (n == m->empty_memory()) {
// Implicit copy of base_memory()
} else if (n_adr_type != TypePtr::BOTTOM) {
assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
} else {
// A few places like make_runtime_call "know" that VM calls are narrow,
// and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
bool expected_wide_mem = false;
if (n == m->base_memory()) {
expected_wide_mem = true;
} else if (alias_idx == Compile::AliasIdxRaw ||
n == m->memory_at(Compile::AliasIdxRaw)) {
expected_wide_mem = true;
} else if (!C->alias_type(alias_idx)->is_rewritable()) {
// memory can "leak through" calls on channels that
// are write-once. Allow this also.
expected_wide_mem = true;
}
assert(expected_wide_mem, "expected narrow slice replacement");
}
}
#else // !ASSERT
#define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
#endif
//-----------------------------memory_at---------------------------------------
Node* MergeMemNode::memory_at(uint alias_idx) const {
assert(alias_idx >= Compile::AliasIdxRaw ||
alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
"must avoid base_memory and AliasIdxTop");
// Otherwise, it is a narrow slice.
Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
Compile *C = Compile::current();
if (is_empty_memory(n)) {
// the array is sparse; empty slots are the "top" node
n = base_memory();
assert(Node::in_dump()
|| n == NULL || n->bottom_type() == Type::TOP
|| n->adr_type() == NULL // address is TOP
|| n->adr_type() == TypePtr::BOTTOM
|| n->adr_type() == TypeRawPtr::BOTTOM
|| Compile::current()->AliasLevel() == 0,
"must be a wide memory");
// AliasLevel == 0 if we are organizing the memory states manually.
// See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
} else {
// make sure the stored slice is sane
#ifdef ASSERT
if (is_error_reported() || Node::in_dump()) {
} else if (might_be_same(n, base_memory())) {
// Give it a pass: It is a mostly harmless repetition of the base.
// This can arise normally from node subsumption during optimization.
} else {
verify_memory_slice(this, alias_idx, n);
}
#endif
}
return n;
}
//---------------------------set_memory_at-------------------------------------
void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
verify_memory_slice(this, alias_idx, n);
Node* empty_mem = empty_memory();
if (n == base_memory()) n = empty_mem; // collapse default
uint need_req = alias_idx+1;
if (req() < need_req) {
if (n == empty_mem) return; // already the default, so do not grow me
// grow the sparse array
do {
add_req(empty_mem);
} while (req() < need_req);
}
set_req( alias_idx, n );
}
//--------------------------iteration_setup------------------------------------
void MergeMemNode::iteration_setup(const MergeMemNode* other) {
if (other != NULL) {
grow_to_match(other);
// invariant: the finite support of mm2 is within mm->req()
#ifdef ASSERT
for (uint i = req(); i < other->req(); i++) {
assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
}
#endif
}
// Replace spurious copies of base_memory by top.
Node* base_mem = base_memory();
if (base_mem != NULL && !base_mem->is_top()) {
for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
if (in(i) == base_mem)
set_req(i, empty_memory());
}
}
}
//---------------------------grow_to_match-------------------------------------
void MergeMemNode::grow_to_match(const MergeMemNode* other) {
Node* empty_mem = empty_memory();
assert(other->is_empty_memory(empty_mem), "consistent sentinels");
// look for the finite support of the other memory
for (uint i = other->req(); --i >= req(); ) {
if (other->in(i) != empty_mem) {
uint new_len = i+1;
while (req() < new_len) add_req(empty_mem);
break;
}
}
}
//---------------------------verify_sparse-------------------------------------
#ifndef PRODUCT
bool MergeMemNode::verify_sparse() const {
assert(is_empty_memory(make_empty_memory()), "sane sentinel");
Node* base_mem = base_memory();
// The following can happen in degenerate cases, since empty==top.
if (is_empty_memory(base_mem)) return true;
for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
assert(in(i) != NULL, "sane slice");
if (in(i) == base_mem) return false; // should have been the sentinel value!
}
return true;
}
bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
Node* n;
n = mm->in(idx);
if (mem == n) return true; // might be empty_memory()
n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
if (mem == n) return true;
while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
if (mem == n) return true;
if (n == NULL) break;
}
return false;
}
#endif // !PRODUCT
C:\hotspot-69087d08d473\src\share\vm/opto/memnode.hpp
/*
* Copyright (c) 1997, 2017, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#ifndef SHARE_VM_OPTO_MEMNODE_HPP
#define SHARE_VM_OPTO_MEMNODE_HPP
#include "opto/multnode.hpp"
#include "opto/node.hpp"
#include "opto/opcodes.hpp"
#include "opto/type.hpp"
// Portions of code courtesy of Clifford Click
class MultiNode;
class PhaseCCP;
class PhaseTransform;
//------------------------------MemNode----------------------------------------
// Load or Store, possibly throwing a NULL pointer exception
class MemNode : public Node {
private:
bool _unaligned_access; // Unaligned access from unsafe
bool _mismatched_access; // Mismatched access from unsafe: byte read in integer array for instance
protected:
#ifdef ASSERT
const TypePtr* _adr_type; // What kind of memory is being addressed?
#endif
virtual uint size_of() const;
public:
enum { Control, // When is it safe to do this load?
Memory, // Chunk of memory is being loaded from
Address, // Actually address, derived from base
ValueIn, // Value to store
OopStore // Preceeding oop store, only in StoreCM
};
typedef enum { unordered = 0,
acquire, // Load has to acquire or be succeeded by MemBarAcquire.
release // Store has to release or be preceded by MemBarRelease.
} MemOrd;
protected:
MemNode( Node *c0, Node *c1, Node *c2, const TypePtr* at )
: Node(c0,c1,c2 ), _unaligned_access(false), _mismatched_access(false) {
init_class_id(Class_Mem);
debug_only(_adr_type=at; adr_type();)
}
MemNode( Node *c0, Node *c1, Node *c2, const TypePtr* at, Node *c3 )
: Node(c0,c1,c2,c3), _unaligned_access(false), _mismatched_access(false) {
init_class_id(Class_Mem);
debug_only(_adr_type=at; adr_type();)
}
MemNode( Node *c0, Node *c1, Node *c2, const TypePtr* at, Node *c3, Node *c4)
: Node(c0,c1,c2,c3,c4), _unaligned_access(false), _mismatched_access(false) {
init_class_id(Class_Mem);
debug_only(_adr_type=at; adr_type();)
}
static bool check_if_adr_maybe_raw(Node* adr);
public:
// Helpers for the optimizer. Documented in memnode.cpp.
static bool detect_ptr_independence(Node* p1, AllocateNode* a1,
Node* p2, AllocateNode* a2,
PhaseTransform* phase);
static bool adr_phi_is_loop_invariant(Node* adr_phi, Node* cast);
static Node *optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase);
static Node *optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase);
// This one should probably be a phase-specific function:
static bool all_controls_dominate(Node* dom, Node* sub);
// Find any cast-away of null-ness and keep its control.
static Node *Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr );
virtual Node *Ideal_DU_postCCP( PhaseCCP *ccp );
virtual const class TypePtr *adr_type() const; // returns bottom_type of address
// Shared code for Ideal methods:
Node *Ideal_common(PhaseGVN *phase, bool can_reshape); // Return -1 for short-circuit NULL.
// Helper function for adr_type() implementations.
static const TypePtr* calculate_adr_type(const Type* t, const TypePtr* cross_check = NULL);
// Raw access function, to allow copying of adr_type efficiently in
// product builds and retain the debug info for debug builds.
const TypePtr *raw_adr_type() const {
#ifdef ASSERT
return _adr_type;
#else
return 0;
#endif
}
// Map a load or store opcode to its corresponding store opcode.
// (Return -1 if unknown.)
virtual int store_Opcode() const { return -1; }
// What is the type of the value in memory? (T_VOID mean "unspecified".)
virtual BasicType memory_type() const = 0;
virtual int memory_size() const {
#ifdef ASSERT
return type2aelembytes(memory_type(), true);
#else
return type2aelembytes(memory_type());
#endif
}
// Search through memory states which precede this node (load or store).
// Look for an exact match for the address, with no intervening
// aliased stores.
Node* find_previous_store(PhaseTransform* phase);
// Can this node (load or store) accurately see a stored value in
// the given memory state? (The state may or may not be in(Memory).)
Node* can_see_stored_value(Node* st, PhaseTransform* phase) const;
void set_unaligned_access() { _unaligned_access = true; }
bool is_unaligned_access() const { return _unaligned_access; }
void set_mismatched_access() { _mismatched_access = true; }
bool is_mismatched_access() const { return _mismatched_access; }
#ifndef PRODUCT
static void dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st);
virtual void dump_spec(outputStream *st) const;
#endif
};
//------------------------------LoadNode---------------------------------------
// Load value; requires Memory and Address
class LoadNode : public MemNode {
public:
// Some loads (from unsafe) should be pinned: they don't depend only
// on the dominating test. The boolean field _depends_only_on_test
// below records whether that node depends only on the dominating
// test.
// Methods used to build LoadNodes pass an argument of type enum
// ControlDependency instead of a boolean because those methods
// typically have multiple boolean parameters with default values:
// passing the wrong boolean to one of these parameters by mistake
// goes easily unnoticed. Using an enum, the compiler can check that
// the type of a value and the type of the parameter match.
enum ControlDependency {
Pinned,
DependsOnlyOnTest
};
private:
// LoadNode::hash() doesn't take the _depends_only_on_test field
// into account: If the graph already has a non-pinned LoadNode and
// we add a pinned LoadNode with the same inputs, it's safe for GVN
// to replace the pinned LoadNode with the non-pinned LoadNode,
// otherwise it wouldn't be safe to have a non pinned LoadNode with
// those inputs in the first place. If the graph already has a
// pinned LoadNode and we add a non pinned LoadNode with the same
// inputs, it's safe (but suboptimal) for GVN to replace the
// non-pinned LoadNode by the pinned LoadNode.
bool _depends_only_on_test;
// On platforms with weak memory ordering (e.g., PPC, Ia64) we distinguish
// loads that can be reordered, and such requiring acquire semantics to
// adhere to the Java specification. The required behaviour is stored in
// this field.
const MemOrd _mo;
protected:
virtual uint cmp(const Node &n) const;
virtual uint size_of() const; // Size is bigger
// Should LoadNode::Ideal() attempt to remove control edges?
virtual bool can_remove_control() const;
const Type* const _type; // What kind of value is loaded?
public:
LoadNode(Node *c, Node *mem, Node *adr, const TypePtr* at, const Type *rt, MemOrd mo, ControlDependency control_dependency)
: MemNode(c,mem,adr,at), _type(rt), _mo(mo), _depends_only_on_test(control_dependency == DependsOnlyOnTest) {
init_class_id(Class_Load);
}
inline bool is_unordered() const { return !is_acquire(); }
inline bool is_acquire() const {
assert(_mo == unordered || _mo == acquire, "unexpected");
return _mo == acquire;
}
// Polymorphic factory method:
static Node* make(PhaseGVN& gvn, Node *c, Node *mem, Node *adr,
const TypePtr* at, const Type *rt, BasicType bt,
MemOrd mo, ControlDependency control_dependency = DependsOnlyOnTest);
virtual uint hash() const; // Check the type
// Handle algebraic identities here. If we have an identity, return the Node
// we are equivalent to. We look for Load of a Store.
virtual Node *Identity( PhaseTransform *phase );
// If the load is from Field memory and the pointer is non-null, it might be possible to
// zero out the control input.
// If the offset is constant and the base is an object allocation,
// try to hook me up to the exact initializing store.
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
// Split instance field load through Phi.
Node* split_through_phi(PhaseGVN *phase);
// Recover original value from boxed values
Node *eliminate_autobox(PhaseGVN *phase);
// Compute a new Type for this node. Basically we just do the pre-check,
// then call the virtual add() to set the type.
virtual const Type *Value( PhaseTransform *phase ) const;
// Common methods for LoadKlass and LoadNKlass nodes.
const Type *klass_value_common( PhaseTransform *phase ) const;
Node *klass_identity_common( PhaseTransform *phase );
virtual uint ideal_reg() const;
virtual const Type *bottom_type() const;
// Following method is copied from TypeNode:
void set_type(const Type* t) {
assert(t != NULL, "sanity");
debug_only(uint check_hash = (VerifyHashTableKeys && _hash_lock) ? hash() : NO_HASH);
*(const Type**)&_type = t; // cast away const-ness
// If this node is in the hash table, make sure it doesn't need a rehash.
assert(check_hash == NO_HASH || check_hash == hash(), "type change must preserve hash code");
}
const Type* type() const { assert(_type != NULL, "sanity"); return _type; };
// Do not match memory edge
virtual uint match_edge(uint idx) const;
// Map a load opcode to its corresponding store opcode.
virtual int store_Opcode() const = 0;
// Check if the load's memory input is a Phi node with the same control.
bool is_instance_field_load_with_local_phi(Node* ctrl);
#ifndef PRODUCT
virtual void dump_spec(outputStream *st) const;
#endif
#ifdef ASSERT
// Helper function to allow a raw load without control edge for some cases
static bool is_immutable_value(Node* adr);
#endif
protected:
const Type* load_array_final_field(const TypeKlassPtr *tkls,
ciKlass* klass) const;
// depends_only_on_test is almost always true, and needs to be almost always
// true to enable key hoisting & commoning optimizations. However, for the
// special case of RawPtr loads from TLS top & end, and other loads performed by
// GC barriers, the control edge carries the dependence preventing hoisting past
// a Safepoint instead of the memory edge. (An unfortunate consequence of having
// Safepoints not set Raw Memory; itself an unfortunate consequence of having Nodes
// which produce results (new raw memory state) inside of loops preventing all
// manner of other optimizations). Basically, it's ugly but so is the alternative.
// See comment in macro.cpp, around line 125 expand_allocate_common().
virtual bool depends_only_on_test() const { return adr_type() != TypeRawPtr::BOTTOM && _depends_only_on_test; }
};
//------------------------------LoadBNode--------------------------------------
// Load a byte (8bits signed) from memory
class LoadBNode : public LoadNode {
public:
LoadBNode(Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeInt *ti, MemOrd mo, ControlDependency control_dependency = DependsOnlyOnTest)
: LoadNode(c, mem, adr, at, ti, mo, control_dependency) {}
virtual int Opcode() const;
virtual uint ideal_reg() const { return Op_RegI; }
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value(PhaseTransform *phase) const;
virtual int store_Opcode() const { return Op_StoreB; }
virtual BasicType memory_type() const { return T_BYTE; }
};
//------------------------------LoadUBNode-------------------------------------
// Load a unsigned byte (8bits unsigned) from memory
class LoadUBNode : public LoadNode {
public:
LoadUBNode(Node* c, Node* mem, Node* adr, const TypePtr* at, const TypeInt* ti, MemOrd mo, ControlDependency control_dependency = DependsOnlyOnTest)
: LoadNode(c, mem, adr, at, ti, mo, control_dependency) {}
virtual int Opcode() const;
virtual uint ideal_reg() const { return Op_RegI; }
virtual Node* Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value(PhaseTransform *phase) const;
virtual int store_Opcode() const { return Op_StoreB; }
virtual BasicType memory_type() const { return T_BYTE; }
};
//------------------------------LoadUSNode-------------------------------------
// Load an unsigned short/char (16bits unsigned) from memory
class LoadUSNode : public LoadNode {
public:
LoadUSNode(Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeInt *ti, MemOrd mo, ControlDependency control_dependency = DependsOnlyOnTest)
: LoadNode(c, mem, adr, at, ti, mo, control_dependency) {}
virtual int Opcode() const;
virtual uint ideal_reg() const { return Op_RegI; }
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value(PhaseTransform *phase) const;
virtual int store_Opcode() const { return Op_StoreC; }
virtual BasicType memory_type() const { return T_CHAR; }
};
//------------------------------LoadSNode--------------------------------------
// Load a short (16bits signed) from memory
class LoadSNode : public LoadNode {
public:
LoadSNode(Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeInt *ti, MemOrd mo, ControlDependency control_dependency = DependsOnlyOnTest)
: LoadNode(c, mem, adr, at, ti, mo, control_dependency) {}
virtual int Opcode() const;
virtual uint ideal_reg() const { return Op_RegI; }
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value(PhaseTransform *phase) const;
virtual int store_Opcode() const { return Op_StoreC; }
virtual BasicType memory_type() const { return T_SHORT; }
};
//------------------------------LoadINode--------------------------------------
// Load an integer from memory
class LoadINode : public LoadNode {
public:
LoadINode(Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeInt *ti, MemOrd mo, ControlDependency control_dependency = DependsOnlyOnTest)
: LoadNode(c, mem, adr, at, ti, mo, control_dependency) {}
virtual int Opcode() const;
virtual uint ideal_reg() const { return Op_RegI; }
virtual int store_Opcode() const { return Op_StoreI; }
virtual BasicType memory_type() const { return T_INT; }
};
//------------------------------LoadRangeNode----------------------------------
// Load an array length from the array
class LoadRangeNode : public LoadINode {
public:
LoadRangeNode(Node *c, Node *mem, Node *adr, const TypeInt *ti = TypeInt::POS)
: LoadINode(c, mem, adr, TypeAryPtr::RANGE, ti, MemNode::unordered) {}
virtual int Opcode() const;
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
};
//------------------------------LoadLNode--------------------------------------
// Load a long from memory
class LoadLNode : public LoadNode {
virtual uint hash() const { return LoadNode::hash() + _require_atomic_access; }
virtual uint cmp( const Node &n ) const {
return _require_atomic_access == ((LoadLNode&)n)._require_atomic_access
&& LoadNode::cmp(n);
}
virtual uint size_of() const { return sizeof(*this); }
const bool _require_atomic_access; // is piecewise load forbidden?
public:
LoadLNode(Node *c, Node *mem, Node *adr, const TypePtr* at, const TypeLong *tl,
MemOrd mo, ControlDependency control_dependency = DependsOnlyOnTest, bool require_atomic_access = false)
: LoadNode(c, mem, adr, at, tl, mo, control_dependency), _require_atomic_access(require_atomic_access) {}
virtual int Opcode() const;
virtual uint ideal_reg() const { return Op_RegL; }
virtual int store_Opcode() const { return Op_StoreL; }
virtual BasicType memory_type() const { return T_LONG; }
bool require_atomic_access() const { return _require_atomic_access; }
static LoadLNode* make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type,
const Type* rt, MemOrd mo, ControlDependency control_dependency = DependsOnlyOnTest);
#ifndef PRODUCT
virtual void dump_spec(outputStream *st) const {
LoadNode::dump_spec(st);
if (_require_atomic_access) st->print(" Atomic!");
}
#endif
};
//------------------------------LoadL_unalignedNode----------------------------
// Load a long from unaligned memory
class LoadL_unalignedNode : public LoadLNode {
public:
LoadL_unalignedNode(Node *c, Node *mem, Node *adr, const TypePtr* at, MemOrd mo, ControlDependency control_dependency = DependsOnlyOnTest)
: LoadLNode(c, mem, adr, at, TypeLong::LONG, mo, control_dependency) {}
virtual int Opcode() const;
};
//------------------------------LoadFNode--------------------------------------
// Load a float (64 bits) from memory
class LoadFNode : public LoadNode {
public:
LoadFNode(Node *c, Node *mem, Node *adr, const TypePtr* at, const Type *t, MemOrd mo, ControlDependency control_dependency = DependsOnlyOnTest)
: LoadNode(c, mem, adr, at, t, mo, control_dependency) {}
virtual int Opcode() const;
virtual uint ideal_reg() const { return Op_RegF; }
virtual int store_Opcode() const { return Op_StoreF; }
virtual BasicType memory_type() const { return T_FLOAT; }
};
//------------------------------LoadDNode--------------------------------------
// Load a double (64 bits) from memory
class LoadDNode : public LoadNode {
virtual uint hash() const { return LoadNode::hash() + _require_atomic_access; }
virtual uint cmp( const Node &n ) const {
return _require_atomic_access == ((LoadDNode&)n)._require_atomic_access
&& LoadNode::cmp(n);
}
virtual uint size_of() const { return sizeof(*this); }
const bool _require_atomic_access; // is piecewise load forbidden?
public:
LoadDNode(Node *c, Node *mem, Node *adr, const TypePtr* at, const Type *t,
MemOrd mo, ControlDependency control_dependency = DependsOnlyOnTest, bool require_atomic_access = false)
: LoadNode(c, mem, adr, at, t, mo, control_dependency), _require_atomic_access(require_atomic_access) {}
virtual int Opcode() const;
virtual uint ideal_reg() const { return Op_RegD; }
virtual int store_Opcode() const { return Op_StoreD; }
virtual BasicType memory_type() const { return T_DOUBLE; }
bool require_atomic_access() const { return _require_atomic_access; }
static LoadDNode* make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type,
const Type* rt, MemOrd mo, ControlDependency control_dependency = DependsOnlyOnTest);
#ifndef PRODUCT
virtual void dump_spec(outputStream *st) const {
LoadNode::dump_spec(st);
if (_require_atomic_access) st->print(" Atomic!");
}
#endif
};
//------------------------------LoadD_unalignedNode----------------------------
// Load a double from unaligned memory
class LoadD_unalignedNode : public LoadDNode {
public:
LoadD_unalignedNode(Node *c, Node *mem, Node *adr, const TypePtr* at, MemOrd mo, ControlDependency control_dependency = DependsOnlyOnTest)
: LoadDNode(c, mem, adr, at, Type::DOUBLE, mo, control_dependency) {}
virtual int Opcode() const;
};
//------------------------------LoadPNode--------------------------------------
// Load a pointer from memory (either object or array)
class LoadPNode : public LoadNode {
public:
LoadPNode(Node *c, Node *mem, Node *adr, const TypePtr *at, const TypePtr* t, MemOrd mo, ControlDependency control_dependency = DependsOnlyOnTest)
: LoadNode(c, mem, adr, at, t, mo, control_dependency) {}
virtual int Opcode() const;
virtual uint ideal_reg() const { return Op_RegP; }
virtual int store_Opcode() const { return Op_StoreP; }
virtual BasicType memory_type() const { return T_ADDRESS; }
};
//------------------------------LoadNNode--------------------------------------
// Load a narrow oop from memory (either object or array)
class LoadNNode : public LoadNode {
public:
LoadNNode(Node *c, Node *mem, Node *adr, const TypePtr *at, const Type* t, MemOrd mo, ControlDependency control_dependency = DependsOnlyOnTest)
: LoadNode(c, mem, adr, at, t, mo, control_dependency) {}
virtual int Opcode() const;
virtual uint ideal_reg() const { return Op_RegN; }
virtual int store_Opcode() const { return Op_StoreN; }
virtual BasicType memory_type() const { return T_NARROWOOP; }
};
//------------------------------LoadKlassNode----------------------------------
// Load a Klass from an object
class LoadKlassNode : public LoadPNode {
protected:
// In most cases, LoadKlassNode does not have the control input set. If the control
// input is set, it must not be removed (by LoadNode::Ideal()).
virtual bool can_remove_control() const;
public:
LoadKlassNode(Node *c, Node *mem, Node *adr, const TypePtr *at, const TypeKlassPtr *tk, MemOrd mo)
: LoadPNode(c, mem, adr, at, tk, mo) {}
virtual int Opcode() const;
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Identity( PhaseTransform *phase );
virtual bool depends_only_on_test() const { return true; }
// Polymorphic factory method:
static Node* make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at,
const TypeKlassPtr* tk = TypeKlassPtr::OBJECT);
};
//------------------------------LoadNKlassNode---------------------------------
// Load a narrow Klass from an object.
class LoadNKlassNode : public LoadNNode {
public:
LoadNKlassNode(Node *c, Node *mem, Node *adr, const TypePtr *at, const TypeNarrowKlass *tk, MemOrd mo)
: LoadNNode(c, mem, adr, at, tk, mo) {}
virtual int Opcode() const;
virtual uint ideal_reg() const { return Op_RegN; }
virtual int store_Opcode() const { return Op_StoreNKlass; }
virtual BasicType memory_type() const { return T_NARROWKLASS; }
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Identity( PhaseTransform *phase );
virtual bool depends_only_on_test() const { return true; }
};
//------------------------------StoreNode--------------------------------------
// Store value; requires Store, Address and Value
class StoreNode : public MemNode {
private:
// On platforms with weak memory ordering (e.g., PPC, Ia64) we distinguish
// stores that can be reordered, and such requiring release semantics to
// adhere to the Java specification. The required behaviour is stored in
// this field.
const MemOrd _mo;
// Needed for proper cloning.
virtual uint size_of() const { return sizeof(*this); }
protected:
virtual uint cmp( const Node &n ) const;
virtual bool depends_only_on_test() const { return false; }
Node *Ideal_masked_input (PhaseGVN *phase, uint mask);
Node *Ideal_sign_extended_input(PhaseGVN *phase, int num_bits);
public:
// We must ensure that stores of object references will be visible
// only after the object's initialization. So the callers of this
// procedure must indicate that the store requires `release'
// semantics, if the stored value is an object reference that might
// point to a new object and may become externally visible.
StoreNode(Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, MemOrd mo)
: MemNode(c, mem, adr, at, val), _mo(mo) {
init_class_id(Class_Store);
}
StoreNode(Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, Node *oop_store, MemOrd mo)
: MemNode(c, mem, adr, at, val, oop_store), _mo(mo) {
init_class_id(Class_Store);
}
inline bool is_unordered() const { return !is_release(); }
inline bool is_release() const {
assert((_mo == unordered || _mo == release), "unexpected");
return _mo == release;
}
// Conservatively release stores of object references in order to
// ensure visibility of object initialization.
static inline MemOrd release_if_reference(const BasicType t) {
const MemOrd mo = (t == T_ARRAY ||
t == T_ADDRESS || // Might be the address of an object reference (`boxing').
t == T_OBJECT) ? release : unordered;
return mo;
}
// Polymorphic factory method
//
// We must ensure that stores of object references will be visible
// only after the object's initialization. So the callers of this
// procedure must indicate that the store requires `release'
// semantics, if the stored value is an object reference that might
// point to a new object and may become externally visible.
static StoreNode* make(PhaseGVN& gvn, Node *c, Node *mem, Node *adr,
const TypePtr* at, Node *val, BasicType bt, MemOrd mo);
virtual uint hash() const; // Check the type
// If the store is to Field memory and the pointer is non-null, we can
// zero out the control input.
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
// Compute a new Type for this node. Basically we just do the pre-check,
// then call the virtual add() to set the type.
virtual const Type *Value( PhaseTransform *phase ) const;
// Check for identity function on memory (Load then Store at same address)
virtual Node *Identity( PhaseTransform *phase );
// Do not match memory edge
virtual uint match_edge(uint idx) const;
virtual const Type *bottom_type() const; // returns Type::MEMORY
// Map a store opcode to its corresponding own opcode, trivially.
virtual int store_Opcode() const { return Opcode(); }
// have all possible loads of the value stored been optimized away?
bool value_never_loaded(PhaseTransform *phase) const;
MemBarNode* trailing_membar() const;
};
//------------------------------StoreBNode-------------------------------------
// Store byte to memory
class StoreBNode : public StoreNode {
public:
StoreBNode(Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, MemOrd mo)
: StoreNode(c, mem, adr, at, val, mo) {}
virtual int Opcode() const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual BasicType memory_type() const { return T_BYTE; }
};
//------------------------------StoreCNode-------------------------------------
// Store char/short to memory
class StoreCNode : public StoreNode {
public:
StoreCNode(Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, MemOrd mo)
: StoreNode(c, mem, adr, at, val, mo) {}
virtual int Opcode() const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual BasicType memory_type() const { return T_CHAR; }
};
//------------------------------StoreINode-------------------------------------
// Store int to memory
class StoreINode : public StoreNode {
public:
StoreINode(Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, MemOrd mo)
: StoreNode(c, mem, adr, at, val, mo) {}
virtual int Opcode() const;
virtual BasicType memory_type() const { return T_INT; }
};
//------------------------------StoreLNode-------------------------------------
// Store long to memory
class StoreLNode : public StoreNode {
virtual uint hash() const { return StoreNode::hash() + _require_atomic_access; }
virtual uint cmp( const Node &n ) const {
return _require_atomic_access == ((StoreLNode&)n)._require_atomic_access
&& StoreNode::cmp(n);
}
virtual uint size_of() const { return sizeof(*this); }
const bool _require_atomic_access; // is piecewise store forbidden?
public:
StoreLNode(Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, MemOrd mo, bool require_atomic_access = false)
: StoreNode(c, mem, adr, at, val, mo), _require_atomic_access(require_atomic_access) {}
virtual int Opcode() const;
virtual BasicType memory_type() const { return T_LONG; }
bool require_atomic_access() const { return _require_atomic_access; }
static StoreLNode* make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo);
#ifndef PRODUCT
virtual void dump_spec(outputStream *st) const {
StoreNode::dump_spec(st);
if (_require_atomic_access) st->print(" Atomic!");
}
#endif
};
//------------------------------StoreFNode-------------------------------------
// Store float to memory
class StoreFNode : public StoreNode {
public:
StoreFNode(Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, MemOrd mo)
: StoreNode(c, mem, adr, at, val, mo) {}
virtual int Opcode() const;
virtual BasicType memory_type() const { return T_FLOAT; }
};
//------------------------------StoreDNode-------------------------------------
// Store double to memory
class StoreDNode : public StoreNode {
virtual uint hash() const { return StoreNode::hash() + _require_atomic_access; }
virtual uint cmp( const Node &n ) const {
return _require_atomic_access == ((StoreDNode&)n)._require_atomic_access
&& StoreNode::cmp(n);
}
virtual uint size_of() const { return sizeof(*this); }
const bool _require_atomic_access; // is piecewise store forbidden?
public:
StoreDNode(Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val,
MemOrd mo, bool require_atomic_access = false)
: StoreNode(c, mem, adr, at, val, mo), _require_atomic_access(require_atomic_access) {}
virtual int Opcode() const;
virtual BasicType memory_type() const { return T_DOUBLE; }
bool require_atomic_access() const { return _require_atomic_access; }
static StoreDNode* make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo);
#ifndef PRODUCT
virtual void dump_spec(outputStream *st) const {
StoreNode::dump_spec(st);
if (_require_atomic_access) st->print(" Atomic!");
}
#endif
};
//------------------------------StorePNode-------------------------------------
// Store pointer to memory
class StorePNode : public StoreNode {
public:
StorePNode(Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, MemOrd mo)
: StoreNode(c, mem, adr, at, val, mo) {}
virtual int Opcode() const;
virtual BasicType memory_type() const { return T_ADDRESS; }
};
//------------------------------StoreNNode-------------------------------------
// Store narrow oop to memory
class StoreNNode : public StoreNode {
public:
StoreNNode(Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, MemOrd mo)
: StoreNode(c, mem, adr, at, val, mo) {}
virtual int Opcode() const;
virtual BasicType memory_type() const { return T_NARROWOOP; }
};
//------------------------------StoreNKlassNode--------------------------------------
// Store narrow klass to memory
class StoreNKlassNode : public StoreNNode {
public:
StoreNKlassNode(Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, MemOrd mo)
: StoreNNode(c, mem, adr, at, val, mo) {}
virtual int Opcode() const;
virtual BasicType memory_type() const { return T_NARROWKLASS; }
};
//------------------------------StoreCMNode-----------------------------------
// Store card-mark byte to memory for CM
// The last StoreCM before a SafePoint must be preserved and occur after its "oop" store
// Preceeding equivalent StoreCMs may be eliminated.
class StoreCMNode : public StoreNode {
private:
virtual uint hash() const { return StoreNode::hash() + _oop_alias_idx; }
virtual uint cmp( const Node &n ) const {
return _oop_alias_idx == ((StoreCMNode&)n)._oop_alias_idx
&& StoreNode::cmp(n);
}
virtual uint size_of() const { return sizeof(*this); }
int _oop_alias_idx; // The alias_idx of OopStore
public:
StoreCMNode( Node *c, Node *mem, Node *adr, const TypePtr* at, Node *val, Node *oop_store, int oop_alias_idx ) :
StoreNode(c, mem, adr, at, val, oop_store, MemNode::release),
_oop_alias_idx(oop_alias_idx) {
assert(_oop_alias_idx >= Compile::AliasIdxRaw ||
_oop_alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
"bad oop alias idx");
}
virtual int Opcode() const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value( PhaseTransform *phase ) const;
virtual BasicType memory_type() const { return T_VOID; } // unspecific
int oop_alias_idx() const { return _oop_alias_idx; }
};
//------------------------------LoadPLockedNode---------------------------------
// Load-locked a pointer from memory (either object or array).
// On Sparc & Intel this is implemented as a normal pointer load.
// On PowerPC and friends it's a real load-locked.
class LoadPLockedNode : public LoadPNode {
public:
LoadPLockedNode(Node *c, Node *mem, Node *adr, MemOrd mo)
: LoadPNode(c, mem, adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, mo) {}
virtual int Opcode() const;
virtual int store_Opcode() const { return Op_StorePConditional; }
virtual bool depends_only_on_test() const { return true; }
};
//------------------------------SCMemProjNode---------------------------------------
// This class defines a projection of the memory state of a store conditional node.
// These nodes return a value, but also update memory.
class SCMemProjNode : public ProjNode {
public:
enum {SCMEMPROJCON = (uint)-2};
SCMemProjNode( Node *src) : ProjNode( src, SCMEMPROJCON) { }
virtual int Opcode() const;
virtual bool is_CFG() const { return false; }
virtual const Type *bottom_type() const {return Type::MEMORY;}
virtual const TypePtr *adr_type() const { return in(0)->in(MemNode::Memory)->adr_type();}
virtual uint ideal_reg() const { return 0;} // memory projections don't have a register
virtual const Type *Value( PhaseTransform *phase ) const;
#ifndef PRODUCT
virtual void dump_spec(outputStream *st) const {};
#endif
};
//------------------------------LoadStoreNode---------------------------
// Note: is_Mem() method returns 'true' for this class.
class LoadStoreNode : public Node {
private:
const Type* const _type; // What kind of value is loaded?
const TypePtr* _adr_type; // What kind of memory is being addressed?
virtual uint size_of() const; // Size is bigger
public:
LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required );
virtual bool depends_only_on_test() const { return false; }
virtual uint match_edge(uint idx) const { return idx == MemNode::Address || idx == MemNode::ValueIn; }
virtual const Type *bottom_type() const { return _type; }
virtual uint ideal_reg() const;
virtual const class TypePtr *adr_type() const { return _adr_type; } // returns bottom_type of address
bool result_not_used() const;
MemBarNode* trailing_membar() const;
};
class LoadStoreConditionalNode : public LoadStoreNode {
public:
enum {
ExpectedIn = MemNode::ValueIn+1 // One more input than MemNode
};
LoadStoreConditionalNode(Node *c, Node *mem, Node *adr, Node *val, Node *ex);
};
//------------------------------StorePConditionalNode---------------------------
// Conditionally store pointer to memory, if no change since prior
// load-locked. Sets flags for success or failure of the store.
class StorePConditionalNode : public LoadStoreConditionalNode {
public:
StorePConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ll ) : LoadStoreConditionalNode(c, mem, adr, val, ll) { }
virtual int Opcode() const;
// Produces flags
virtual uint ideal_reg() const { return Op_RegFlags; }
};
//------------------------------StoreIConditionalNode---------------------------
// Conditionally store int to memory, if no change since prior
// load-locked. Sets flags for success or failure of the store.
class StoreIConditionalNode : public LoadStoreConditionalNode {
public:
StoreIConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ii ) : LoadStoreConditionalNode(c, mem, adr, val, ii) { }
virtual int Opcode() const;
// Produces flags
virtual uint ideal_reg() const { return Op_RegFlags; }
};
//------------------------------StoreLConditionalNode---------------------------
// Conditionally store long to memory, if no change since prior
// load-locked. Sets flags for success or failure of the store.
class StoreLConditionalNode : public LoadStoreConditionalNode {
public:
StoreLConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ll ) : LoadStoreConditionalNode(c, mem, adr, val, ll) { }
virtual int Opcode() const;
// Produces flags
virtual uint ideal_reg() const { return Op_RegFlags; }
};
//------------------------------CompareAndSwapLNode---------------------------
class CompareAndSwapLNode : public LoadStoreConditionalNode {
public:
CompareAndSwapLNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex) : LoadStoreConditionalNode(c, mem, adr, val, ex) { }
virtual int Opcode() const;
};
//------------------------------CompareAndSwapINode---------------------------
class CompareAndSwapINode : public LoadStoreConditionalNode {
public:
CompareAndSwapINode( Node *c, Node *mem, Node *adr, Node *val, Node *ex) : LoadStoreConditionalNode(c, mem, adr, val, ex) { }
virtual int Opcode() const;
};
//------------------------------CompareAndSwapPNode---------------------------
class CompareAndSwapPNode : public LoadStoreConditionalNode {
public:
CompareAndSwapPNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex) : LoadStoreConditionalNode(c, mem, adr, val, ex) { }
virtual int Opcode() const;
};
//------------------------------CompareAndSwapNNode---------------------------
class CompareAndSwapNNode : public LoadStoreConditionalNode {
public:
CompareAndSwapNNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex) : LoadStoreConditionalNode(c, mem, adr, val, ex) { }
virtual int Opcode() const;
};
//------------------------------GetAndAddINode---------------------------
class GetAndAddINode : public LoadStoreNode {
public:
GetAndAddINode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at ) : LoadStoreNode(c, mem, adr, val, at, TypeInt::INT, 4) { }
virtual int Opcode() const;
};
//------------------------------GetAndAddLNode---------------------------
class GetAndAddLNode : public LoadStoreNode {
public:
GetAndAddLNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at ) : LoadStoreNode(c, mem, adr, val, at, TypeLong::LONG, 4) { }
virtual int Opcode() const;
};
//------------------------------GetAndSetINode---------------------------
class GetAndSetINode : public LoadStoreNode {
public:
GetAndSetINode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at ) : LoadStoreNode(c, mem, adr, val, at, TypeInt::INT, 4) { }
virtual int Opcode() const;
};
//------------------------------GetAndSetINode---------------------------
class GetAndSetLNode : public LoadStoreNode {
public:
GetAndSetLNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at ) : LoadStoreNode(c, mem, adr, val, at, TypeLong::LONG, 4) { }
virtual int Opcode() const;
};
//------------------------------GetAndSetPNode---------------------------
class GetAndSetPNode : public LoadStoreNode {
public:
GetAndSetPNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* t ) : LoadStoreNode(c, mem, adr, val, at, t, 4) { }
virtual int Opcode() const;
};
//------------------------------GetAndSetNNode---------------------------
class GetAndSetNNode : public LoadStoreNode {
public:
GetAndSetNNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* t ) : LoadStoreNode(c, mem, adr, val, at, t, 4) { }
virtual int Opcode() const;
};
//------------------------------ClearArray-------------------------------------
class ClearArrayNode: public Node {
public:
ClearArrayNode( Node *ctrl, Node *arymem, Node *word_cnt, Node *base )
: Node(ctrl,arymem,word_cnt,base) {
init_class_id(Class_ClearArray);
}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return Type::MEMORY; }
// ClearArray modifies array elements, and so affects only the
// array memory addressed by the bottom_type of its base address.
virtual const class TypePtr *adr_type() const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual uint match_edge(uint idx) const;
// Clear the given area of an object or array.
// The start offset must always be aligned mod BytesPerInt.
// The end offset must always be aligned mod BytesPerLong.
// Return the new memory.
static Node* clear_memory(Node* control, Node* mem, Node* dest,
intptr_t start_offset,
intptr_t end_offset,
PhaseGVN* phase);
static Node* clear_memory(Node* control, Node* mem, Node* dest,
intptr_t start_offset,
Node* end_offset,
PhaseGVN* phase);
static Node* clear_memory(Node* control, Node* mem, Node* dest,
Node* start_offset,
Node* end_offset,
PhaseGVN* phase);
// Return allocation input memory edge if it is different instance
// or itself if it is the one we are looking for.
static bool step_through(Node** np, uint instance_id, PhaseTransform* phase);
};
//------------------------------StrIntrinsic-------------------------------
// Base class for Ideal nodes used in String instrinsic code.
class StrIntrinsicNode: public Node {
public:
StrIntrinsicNode(Node* control, Node* char_array_mem,
Node* s1, Node* c1, Node* s2, Node* c2):
Node(control, char_array_mem, s1, c1, s2, c2) {
}
StrIntrinsicNode(Node* control, Node* char_array_mem,
Node* s1, Node* s2, Node* c):
Node(control, char_array_mem, s1, s2, c) {
}
StrIntrinsicNode(Node* control, Node* char_array_mem,
Node* s1, Node* s2):
Node(control, char_array_mem, s1, s2) {
}
virtual bool depends_only_on_test() const { return false; }
virtual const TypePtr* adr_type() const { return TypeAryPtr::CHARS; }
virtual uint match_edge(uint idx) const;
virtual uint ideal_reg() const { return Op_RegI; }
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value(PhaseTransform *phase) const;
};
//------------------------------StrComp-------------------------------------
class StrCompNode: public StrIntrinsicNode {
public:
StrCompNode(Node* control, Node* char_array_mem,
Node* s1, Node* c1, Node* s2, Node* c2):
StrIntrinsicNode(control, char_array_mem, s1, c1, s2, c2) {};
virtual int Opcode() const;
virtual const Type* bottom_type() const { return TypeInt::INT; }
};
//------------------------------StrEquals-------------------------------------
class StrEqualsNode: public StrIntrinsicNode {
public:
StrEqualsNode(Node* control, Node* char_array_mem,
Node* s1, Node* s2, Node* c):
StrIntrinsicNode(control, char_array_mem, s1, s2, c) {};
virtual int Opcode() const;
virtual const Type* bottom_type() const { return TypeInt::BOOL; }
};
//------------------------------StrIndexOf-------------------------------------
class StrIndexOfNode: public StrIntrinsicNode {
public:
StrIndexOfNode(Node* control, Node* char_array_mem,
Node* s1, Node* c1, Node* s2, Node* c2):
StrIntrinsicNode(control, char_array_mem, s1, c1, s2, c2) {};
virtual int Opcode() const;
virtual const Type* bottom_type() const { return TypeInt::INT; }
};
//------------------------------AryEq---------------------------------------
class AryEqNode: public StrIntrinsicNode {
public:
AryEqNode(Node* control, Node* char_array_mem, Node* s1, Node* s2):
StrIntrinsicNode(control, char_array_mem, s1, s2) {};
virtual int Opcode() const;
virtual const Type* bottom_type() const { return TypeInt::BOOL; }
};
//------------------------------EncodeISOArray--------------------------------
// encode char[] to byte[] in ISO_8859_1
class EncodeISOArrayNode: public Node {
public:
EncodeISOArrayNode(Node *control, Node* arymem, Node* s1, Node* s2, Node* c): Node(control, arymem, s1, s2, c) {};
virtual int Opcode() const;
virtual bool depends_only_on_test() const { return false; }
virtual const Type* bottom_type() const { return TypeInt::INT; }
virtual const TypePtr* adr_type() const { return TypePtr::BOTTOM; }
virtual uint match_edge(uint idx) const;
virtual uint ideal_reg() const { return Op_RegI; }
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value(PhaseTransform *phase) const;
};
//------------------------------MemBar-----------------------------------------
// There are different flavors of Memory Barriers to match the Java Memory
// Model. Monitor-enter and volatile-load act as Aquires: no following ref
// can be moved to before them. We insert a MemBar-Acquire after a FastLock or
// volatile-load. Monitor-exit and volatile-store act as Release: no
// preceding ref can be moved to after them. We insert a MemBar-Release
// before a FastUnlock or volatile-store. All volatiles need to be
// serialized, so we follow all volatile-stores with a MemBar-Volatile to
// separate it from any following volatile-load.
class MemBarNode: public MultiNode {
virtual uint hash() const ; // { return NO_HASH; }
virtual uint cmp( const Node &n ) const ; // Always fail, except on self
virtual uint size_of() const { return sizeof(*this); }
// Memory type this node is serializing. Usually either rawptr or bottom.
const TypePtr* _adr_type;
// How is this membar related to a nearby memory access?
enum {
Standalone,
TrailingLoad,
TrailingStore,
LeadingStore,
TrailingLoadStore,
LeadingLoadStore
} _kind;
#ifdef ASSERT
uint _pair_idx;
#endif
public:
enum {
Precedent = TypeFunc::Parms // optional edge to force precedence
};
MemBarNode(Compile* C, int alias_idx, Node* precedent);
virtual int Opcode() const = 0;
virtual const class TypePtr *adr_type() const { return _adr_type; }
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual uint match_edge(uint idx) const { return 0; }
virtual const Type *bottom_type() const { return TypeTuple::MEMBAR; }
virtual Node *match( const ProjNode *proj, const Matcher *m );
// Factory method. Builds a wide or narrow membar.
// Optional 'precedent' becomes an extra edge if not null.
static MemBarNode* make(Compile* C, int opcode,
int alias_idx = Compile::AliasIdxBot,
Node* precedent = NULL);
MemBarNode* trailing_membar() const;
MemBarNode* leading_membar() const;
void set_trailing_load() { _kind = TrailingLoad; }
bool trailing_load() const { return _kind == TrailingLoad; }
bool trailing_store() const { return _kind == TrailingStore; }
bool leading_store() const { return _kind == LeadingStore; }
bool trailing_load_store() const { return _kind == TrailingLoadStore; }
bool leading_load_store() const { return _kind == LeadingLoadStore; }
bool trailing() const { return _kind == TrailingLoad || _kind == TrailingStore || _kind == TrailingLoadStore; }
bool leading() const { return _kind == LeadingStore || _kind == LeadingLoadStore; }
bool standalone() const { return _kind == Standalone; }
static void set_store_pair(MemBarNode* leading, MemBarNode* trailing);
static void set_load_store_pair(MemBarNode* leading, MemBarNode* trailing);
void remove(PhaseIterGVN *igvn);
};
// "Acquire" - no following ref can move before (but earlier refs can
// follow, like an early Load stalled in cache). Requires multi-cpu
// visibility. Inserted after a volatile load.
class MemBarAcquireNode: public MemBarNode {
public:
MemBarAcquireNode(Compile* C, int alias_idx, Node* precedent)
: MemBarNode(C, alias_idx, precedent) {}
virtual int Opcode() const;
};
// "Acquire" - no following ref can move before (but earlier refs can
// follow, like an early Load stalled in cache). Requires multi-cpu
// visibility. Inserted independ of any load, as required
// for intrinsic sun.misc.Unsafe.loadFence().
class LoadFenceNode: public MemBarNode {
public:
LoadFenceNode(Compile* C, int alias_idx, Node* precedent)
: MemBarNode(C, alias_idx, precedent) {}
virtual int Opcode() const;
};
// "Release" - no earlier ref can move after (but later refs can move
// up, like a speculative pipelined cache-hitting Load). Requires
// multi-cpu visibility. Inserted before a volatile store.
class MemBarReleaseNode: public MemBarNode {
public:
MemBarReleaseNode(Compile* C, int alias_idx, Node* precedent)
: MemBarNode(C, alias_idx, precedent) {}
virtual int Opcode() const;
};
// "Release" - no earlier ref can move after (but later refs can move
// up, like a speculative pipelined cache-hitting Load). Requires
// multi-cpu visibility. Inserted independent of any store, as required
// for intrinsic sun.misc.Unsafe.storeFence().
class StoreFenceNode: public MemBarNode {
public:
StoreFenceNode(Compile* C, int alias_idx, Node* precedent)
: MemBarNode(C, alias_idx, precedent) {}
virtual int Opcode() const;
};
// "Acquire" - no following ref can move before (but earlier refs can
// follow, like an early Load stalled in cache). Requires multi-cpu
// visibility. Inserted after a FastLock.
class MemBarAcquireLockNode: public MemBarNode {
public:
MemBarAcquireLockNode(Compile* C, int alias_idx, Node* precedent)
: MemBarNode(C, alias_idx, precedent) {}
virtual int Opcode() const;
};
// "Release" - no earlier ref can move after (but later refs can move
// up, like a speculative pipelined cache-hitting Load). Requires
// multi-cpu visibility. Inserted before a FastUnLock.
class MemBarReleaseLockNode: public MemBarNode {
public:
MemBarReleaseLockNode(Compile* C, int alias_idx, Node* precedent)
: MemBarNode(C, alias_idx, precedent) {}
virtual int Opcode() const;
};
class MemBarStoreStoreNode: public MemBarNode {
public:
MemBarStoreStoreNode(Compile* C, int alias_idx, Node* precedent)
: MemBarNode(C, alias_idx, precedent) {
init_class_id(Class_MemBarStoreStore);
}
virtual int Opcode() const;
};
// Ordering between a volatile store and a following volatile load.
// Requires multi-CPU visibility?
class MemBarVolatileNode: public MemBarNode {
public:
MemBarVolatileNode(Compile* C, int alias_idx, Node* precedent)
: MemBarNode(C, alias_idx, precedent) {}
virtual int Opcode() const;
};
// Ordering within the same CPU. Used to order unsafe memory references
// inside the compiler when we lack alias info. Not needed "outside" the
// compiler because the CPU does all the ordering for us.
class MemBarCPUOrderNode: public MemBarNode {
public:
MemBarCPUOrderNode(Compile* C, int alias_idx, Node* precedent)
: MemBarNode(C, alias_idx, precedent) {}
virtual int Opcode() const;
virtual uint ideal_reg() const { return 0; } // not matched in the AD file
};
// Isolation of object setup after an AllocateNode and before next safepoint.
// (See comment in memnode.cpp near InitializeNode::InitializeNode for semantics.)
class InitializeNode: public MemBarNode {
friend class AllocateNode;
enum {
Incomplete = 0,
Complete = 1,
WithArraycopy = 2
};
int _is_complete;
bool _does_not_escape;
public:
enum {
Control = TypeFunc::Control,
Memory = TypeFunc::Memory, // MergeMem for states affected by this op
RawAddress = TypeFunc::Parms+0, // the newly-allocated raw address
RawStores = TypeFunc::Parms+1 // zero or more stores (or TOP)
};
InitializeNode(Compile* C, int adr_type, Node* rawoop);
virtual int Opcode() const;
virtual uint size_of() const { return sizeof(*this); }
virtual uint ideal_reg() const { return 0; } // not matched in the AD file
virtual const RegMask &in_RegMask(uint) const; // mask for RawAddress
// Manage incoming memory edges via a MergeMem on in(Memory):
Node* memory(uint alias_idx);
// The raw memory edge coming directly from the Allocation.
// The contents of this memory are *always* all-zero-bits.
Node* zero_memory() { return memory(Compile::AliasIdxRaw); }
// Return the corresponding allocation for this initialization (or null if none).
// (Note: Both InitializeNode::allocation and AllocateNode::initialization
// are defined in graphKit.cpp, which sets up the bidirectional relation.)
AllocateNode* allocation();
// Anything other than zeroing in this init?
bool is_non_zero();
// An InitializeNode must completed before macro expansion is done.
// Completion requires that the AllocateNode must be followed by
// initialization of the new memory to zero, then to any initializers.
bool is_complete() { return _is_complete != Incomplete; }
bool is_complete_with_arraycopy() { return (_is_complete & WithArraycopy) != 0; }
// Mark complete. (Must not yet be complete.)
void set_complete(PhaseGVN* phase);
void set_complete_with_arraycopy() { _is_complete = Complete | WithArraycopy; }
bool does_not_escape() { return _does_not_escape; }
void set_does_not_escape() { _does_not_escape = true; }
#ifdef ASSERT
// ensure all non-degenerate stores are ordered and non-overlapping
bool stores_are_sane(PhaseTransform* phase);
#endif //ASSERT
// See if this store can be captured; return offset where it initializes.
// Return 0 if the store cannot be moved (any sort of problem).
intptr_t can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape);
// Capture another store; reformat it to write my internal raw memory.
// Return the captured copy, else NULL if there is some sort of problem.
Node* capture_store(StoreNode* st, intptr_t start, PhaseTransform* phase, bool can_reshape);
// Find captured store which corresponds to the range [start..start+size).
// Return my own memory projection (meaning the initial zero bits)
// if there is no such store. Return NULL if there is a problem.
Node* find_captured_store(intptr_t start, int size_in_bytes, PhaseTransform* phase);
// Called when the associated AllocateNode is expanded into CFG.
Node* complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
intptr_t header_size, Node* size_in_bytes,
PhaseGVN* phase);
private:
void remove_extra_zeroes();
// Find out where a captured store should be placed (or already is placed).
int captured_store_insertion_point(intptr_t start, int size_in_bytes,
PhaseTransform* phase);
static intptr_t get_store_offset(Node* st, PhaseTransform* phase);
Node* make_raw_address(intptr_t offset, PhaseTransform* phase);
bool detect_init_independence(Node* n, int& count);
void coalesce_subword_stores(intptr_t header_size, Node* size_in_bytes,
PhaseGVN* phase);
intptr_t find_next_fullword_store(uint i, PhaseGVN* phase);
};
//------------------------------MergeMem---------------------------------------
// (See comment in memnode.cpp near MergeMemNode::MergeMemNode for semantics.)
class MergeMemNode: public Node {
virtual uint hash() const ; // { return NO_HASH; }
virtual uint cmp( const Node &n ) const ; // Always fail, except on self
friend class MergeMemStream;
MergeMemNode(Node* def); // clients use MergeMemNode::make
public:
// If the input is a whole memory state, clone it with all its slices intact.
// Otherwise, make a new memory state with just that base memory input.
// In either case, the result is a newly created MergeMem.
static MergeMemNode* make(Compile* C, Node* base_memory);
virtual int Opcode() const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual uint ideal_reg() const { return NotAMachineReg; }
virtual uint match_edge(uint idx) const { return 0; }
virtual const RegMask &out_RegMask() const;
virtual const Type *bottom_type() const { return Type::MEMORY; }
virtual const TypePtr *adr_type() const { return TypePtr::BOTTOM; }
// sparse accessors
// Fetch the previously stored "set_memory_at", or else the base memory.
// (Caller should clone it if it is a phi-nest.)
Node* memory_at(uint alias_idx) const;
// set the memory, regardless of its previous value
void set_memory_at(uint alias_idx, Node* n);
// the "base" is the memory that provides the non-finite support
Node* base_memory() const { return in(Compile::AliasIdxBot); }
// warning: setting the base can implicitly set any of the other slices too
void set_base_memory(Node* def);
// sentinel value which denotes a copy of the base memory:
Node* empty_memory() const { return in(Compile::AliasIdxTop); }
static Node* make_empty_memory(); // where the sentinel comes from
bool is_empty_memory(Node* n) const { assert((n == empty_memory()) == n->is_top(), "sanity"); return n->is_top(); }
// hook for the iterator, to perform any necessary setup
void iteration_setup(const MergeMemNode* other = NULL);
// push sentinels until I am at least as long as the other (semantic no-op)
void grow_to_match(const MergeMemNode* other);
bool verify_sparse() const PRODUCT_RETURN0;
#ifndef PRODUCT
virtual void dump_spec(outputStream *st) const;
#endif
};
class MergeMemStream : public StackObj {
private:
MergeMemNode* _mm;
const MergeMemNode* _mm2; // optional second guy, contributes non-empty iterations
Node* _mm_base; // loop-invariant base memory of _mm
int _idx;
int _cnt;
Node* _mem;
Node* _mem2;
int _cnt2;
void init(MergeMemNode* mm, const MergeMemNode* mm2 = NULL) {
// subsume_node will break sparseness at times, whenever a memory slice
// folds down to a copy of the base ("fat") memory. In such a case,
// the raw edge will update to base, although it should be top.
// This iterator will recognize either top or base_memory as an
// "empty" slice. See is_empty, is_empty2, and next below.
//
// The sparseness property is repaired in MergeMemNode::Ideal.
// As long as access to a MergeMem goes through this iterator
// or the memory_at accessor, flaws in the sparseness will
// never be observed.
//
// Also, iteration_setup repairs sparseness.
assert(mm->verify_sparse(), "please, no dups of base");
assert(mm2==NULL || mm2->verify_sparse(), "please, no dups of base");
_mm = mm;
_mm_base = mm->base_memory();
_mm2 = mm2;
_cnt = mm->req();
_idx = Compile::AliasIdxBot-1; // start at the base memory
_mem = NULL;
_mem2 = NULL;
}
#ifdef ASSERT
Node* check_memory() const {
if (at_base_memory())
return _mm->base_memory();
else if ((uint)_idx < _mm->req() && !_mm->in(_idx)->is_top())
return _mm->memory_at(_idx);
else
return _mm_base;
}
Node* check_memory2() const {
return at_base_memory()? _mm2->base_memory(): _mm2->memory_at(_idx);
}
#endif
static bool match_memory(Node* mem, const MergeMemNode* mm, int idx) PRODUCT_RETURN0;
void assert_synch() const {
assert(!_mem || _idx >= _cnt || match_memory(_mem, _mm, _idx),
"no side-effects except through the stream");
}
public:
// expected usages:
// for (MergeMemStream mms(mem->is_MergeMem()); next_non_empty(); ) { ... }
// for (MergeMemStream mms(mem1, mem2); next_non_empty2(); ) { ... }
// iterate over one merge
MergeMemStream(MergeMemNode* mm) {
mm->iteration_setup();
init(mm);
debug_only(_cnt2 = 999);
}
// iterate in parallel over two merges
// only iterates through non-empty elements of mm2
MergeMemStream(MergeMemNode* mm, const MergeMemNode* mm2) {
assert(mm2, "second argument must be a MergeMem also");
((MergeMemNode*)mm2)->iteration_setup(); // update hidden state
mm->iteration_setup(mm2);
init(mm, mm2);
_cnt2 = mm2->req();
}
#ifdef ASSERT
~MergeMemStream() {
assert_synch();
}
#endif
MergeMemNode* all_memory() const {
return _mm;
}
Node* base_memory() const {
assert(_mm_base == _mm->base_memory(), "no update to base memory, please");
return _mm_base;
}
const MergeMemNode* all_memory2() const {
assert(_mm2 != NULL, "");
return _mm2;
}
bool at_base_memory() const {
return _idx == Compile::AliasIdxBot;
}
int alias_idx() const {
assert(_mem, "must call next 1st");
return _idx;
}
const TypePtr* adr_type() const {
return Compile::current()->get_adr_type(alias_idx());
}
const TypePtr* adr_type(Compile* C) const {
return C->get_adr_type(alias_idx());
}
bool is_empty() const {
assert(_mem, "must call next 1st");
assert(_mem->is_top() == (_mem==_mm->empty_memory()), "correct sentinel");
return _mem->is_top();
}
bool is_empty2() const {
assert(_mem2, "must call next 1st");
assert(_mem2->is_top() == (_mem2==_mm2->empty_memory()), "correct sentinel");
return _mem2->is_top();
}
Node* memory() const {
assert(!is_empty(), "must not be empty");
assert_synch();
return _mem;
}
// get the current memory, regardless of empty or non-empty status
Node* force_memory() const {
assert(!is_empty() || !at_base_memory(), "");
// Use _mm_base to defend against updates to _mem->base_memory().
Node *mem = _mem->is_top() ? _mm_base : _mem;
assert(mem == check_memory(), "");
return mem;
}
Node* memory2() const {
assert(_mem2 == check_memory2(), "");
return _mem2;
}
void set_memory(Node* mem) {
if (at_base_memory()) {
// Note that this does not change the invariant _mm_base.
_mm->set_base_memory(mem);
} else {
_mm->set_memory_at(_idx, mem);
}
_mem = mem;
assert_synch();
}
// Recover from a side effect to the MergeMemNode.
void set_memory() {
_mem = _mm->in(_idx);
}
bool next() { return next(false); }
bool next2() { return next(true); }
bool next_non_empty() { return next_non_empty(false); }
bool next_non_empty2() { return next_non_empty(true); }
// next_non_empty2 can yield states where is_empty() is true
private:
// find the next item, which might be empty
bool next(bool have_mm2) {
assert((_mm2 != NULL) == have_mm2, "use other next");
assert_synch();
if (++_idx < _cnt) {
// Note: This iterator allows _mm to be non-sparse.
// It behaves the same whether _mem is top or base_memory.
_mem = _mm->in(_idx);
if (have_mm2)
_mem2 = _mm2->in((_idx < _cnt2) ? _idx : Compile::AliasIdxTop);
return true;
}
return false;
}
// find the next non-empty item
bool next_non_empty(bool have_mm2) {
while (next(have_mm2)) {
if (!is_empty()) {
// make sure _mem2 is filled in sensibly
if (have_mm2 && _mem2->is_top()) _mem2 = _mm2->base_memory();
return true;
} else if (have_mm2 && !is_empty2()) {
return true; // is_empty() == true
}
}
return false;
}
};
//------------------------------Prefetch---------------------------------------
// Non-faulting prefetch load. Prefetch for many reads.
class PrefetchReadNode : public Node {
public:
PrefetchReadNode(Node *abio, Node *adr) : Node(0,abio,adr) {}
virtual int Opcode() const;
virtual uint ideal_reg() const { return NotAMachineReg; }
virtual uint match_edge(uint idx) const { return idx==2; }
virtual const Type *bottom_type() const { return Type::ABIO; }
};
// Non-faulting prefetch load. Prefetch for many reads & many writes.
class PrefetchWriteNode : public Node {
public:
PrefetchWriteNode(Node *abio, Node *adr) : Node(0,abio,adr) {}
virtual int Opcode() const;
virtual uint ideal_reg() const { return NotAMachineReg; }
virtual uint match_edge(uint idx) const { return idx==2; }
virtual const Type *bottom_type() const { return Type::ABIO; }
};
// Allocation prefetch which may fault, TLAB size have to be adjusted.
class PrefetchAllocationNode : public Node {
public:
PrefetchAllocationNode(Node *mem, Node *adr) : Node(0,mem,adr) {}
virtual int Opcode() const;
virtual uint ideal_reg() const { return NotAMachineReg; }
virtual uint match_edge(uint idx) const { return idx==2; }
virtual const Type *bottom_type() const { return ( AllocatePrefetchStyle == 3 ) ? Type::MEMORY : Type::ABIO; }
};
#endif // SHARE_VM_OPTO_MEMNODE_HPP
C:\hotspot-69087d08d473\src\share\vm/opto/mulnode.cpp
/*
* Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#include "precompiled.hpp"
#include "memory/allocation.inline.hpp"
#include "opto/addnode.hpp"
#include "opto/connode.hpp"
#include "opto/memnode.hpp"
#include "opto/mulnode.hpp"
#include "opto/phaseX.hpp"
#include "opto/subnode.hpp"
// Portions of code courtesy of Clifford Click
//=============================================================================
//------------------------------hash-------------------------------------------
// Hash function over MulNodes. Needs to be commutative; i.e., I swap
// (commute) inputs to MulNodes willy-nilly so the hash function must return
// the same value in the presence of edge swapping.
uint MulNode::hash() const {
return (uintptr_t)in(1) + (uintptr_t)in(2) + Opcode();
}
//------------------------------Identity---------------------------------------
// Multiplying a one preserves the other argument
Node *MulNode::Identity( PhaseTransform *phase ) {
register const Type *one = mul_id(); // The multiplicative identity
if( phase->type( in(1) )->higher_equal( one ) ) return in(2);
if( phase->type( in(2) )->higher_equal( one ) ) return in(1);
return this;
}
//------------------------------Ideal------------------------------------------
// We also canonicalize the Node, moving constants to the right input,
// and flatten expressions (so that 1+x+2 becomes x+3).
Node *MulNode::Ideal(PhaseGVN *phase, bool can_reshape) {
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
Node *progress = NULL; // Progress flag
// We are OK if right is a constant, or right is a load and
// left is a non-constant.
if( !(t2->singleton() ||
(in(2)->is_Load() && !(t1->singleton() || in(1)->is_Load())) ) ) {
if( t1->singleton() || // Left input is a constant?
// Otherwise, sort inputs (commutativity) to help value numbering.
(in(1)->_idx > in(2)->_idx) ) {
swap_edges(1, 2);
const Type *t = t1;
t1 = t2;
t2 = t;
progress = this; // Made progress
}
}
// If the right input is a constant, and the left input is a product of a
// constant, flatten the expression tree.
uint op = Opcode();
if( t2->singleton() && // Right input is a constant?
op != Op_MulF && // Float & double cannot reassociate
op != Op_MulD ) {
if( t2 == Type::TOP ) return NULL;
Node *mul1 = in(1);
#ifdef ASSERT
// Check for dead loop
int op1 = mul1->Opcode();
if( phase->eqv( mul1, this ) || phase->eqv( in(2), this ) ||
( op1 == mul_opcode() || op1 == add_opcode() ) &&
( phase->eqv( mul1->in(1), this ) || phase->eqv( mul1->in(2), this ) ||
phase->eqv( mul1->in(1), mul1 ) || phase->eqv( mul1->in(2), mul1 ) ) )
assert(false, "dead loop in MulNode::Ideal");
#endif
if( mul1->Opcode() == mul_opcode() ) { // Left input is a multiply?
// Mul of a constant?
const Type *t12 = phase->type( mul1->in(2) );
if( t12->singleton() && t12 != Type::TOP) { // Left input is an add of a constant?
// Compute new constant; check for overflow
const Type *tcon01 = ((MulNode*)mul1)->mul_ring(t2,t12);
if( tcon01->singleton() ) {
// The Mul of the flattened expression
set_req(1, mul1->in(1));
set_req(2, phase->makecon( tcon01 ));
t2 = tcon01;
progress = this; // Made progress
}
}
}
// If the right input is a constant, and the left input is an add of a
// constant, flatten the tree: (X+con1)*con0 ==> X*con0 + con1*con0
const Node *add1 = in(1);
if( add1->Opcode() == add_opcode() ) { // Left input is an add?
// Add of a constant?
const Type *t12 = phase->type( add1->in(2) );
if( t12->singleton() && t12 != Type::TOP ) { // Left input is an add of a constant?
assert( add1->in(1) != add1, "dead loop in MulNode::Ideal" );
// Compute new constant; check for overflow
const Type *tcon01 = mul_ring(t2,t12);
if( tcon01->singleton() ) {
// Convert (X+con1)*con0 into X*con0
Node *mul = clone(); // mul = ()*con0
mul->set_req(1,add1->in(1)); // mul = X*con0
mul = phase->transform(mul);
Node *add2 = add1->clone();
add2->set_req(1, mul); // X*con0 + con0*con1
add2->set_req(2, phase->makecon(tcon01) );
progress = add2;
}
}
} // End of is left input an add
} // End of is right input a Mul
return progress;
}
//------------------------------Value-----------------------------------------
const Type *MulNode::Value( PhaseTransform *phase ) const {
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
// Either input is TOP ==> the result is TOP
if( t1 == Type::TOP ) return Type::TOP;
if( t2 == Type::TOP ) return Type::TOP;
// Either input is ZERO ==> the result is ZERO.
// Not valid for floats or doubles since +0.0 * -0.0 --> +0.0
int op = Opcode();
if( op == Op_MulI || op == Op_AndI || op == Op_MulL || op == Op_AndL ) {
const Type *zero = add_id(); // The multiplicative zero
if( t1->higher_equal( zero ) ) return zero;
if( t2->higher_equal( zero ) ) return zero;
}
// Either input is BOTTOM ==> the result is the local BOTTOM
if( t1 == Type::BOTTOM || t2 == Type::BOTTOM )
return bottom_type();
#if defined(IA32)
// Can't trust native compilers to properly fold strict double
// multiplication with round-to-zero on this platform.
if (op == Op_MulD && phase->C->method()->is_strict()) {
return TypeD::DOUBLE;
}
#endif
return mul_ring(t1,t2); // Local flavor of type multiplication
}
//=============================================================================
//------------------------------Ideal------------------------------------------
// Check for power-of-2 multiply, then try the regular MulNode::Ideal
Node *MulINode::Ideal(PhaseGVN *phase, bool can_reshape) {
// Swap constant to right
jint con;
if ((con = in(1)->find_int_con(0)) != 0) {
swap_edges(1, 2);
// Finish rest of method to use info in 'con'
} else if ((con = in(2)->find_int_con(0)) == 0) {
return MulNode::Ideal(phase, can_reshape);
}
// Now we have a constant Node on the right and the constant in con
if (con == 0) return NULL; // By zero is handled by Value call
if (con == 1) return NULL; // By one is handled by Identity call
// Check for negative constant; if so negate the final result
bool sign_flip = false;
unsigned int abs_con = uabs(con);
if (abs_con != (unsigned int)con) {
sign_flip = true;
}
// Get low bit; check for being the only bit
Node *res = NULL;
unsigned int bit1 = abs_con & (0-abs_con); // Extract low bit
if (bit1 == abs_con) { // Found a power of 2?
res = new (phase->C) LShiftINode(in(1), phase->intcon(log2_uint(bit1)));
} else {
// Check for constant with 2 bits set
unsigned int bit2 = abs_con-bit1;
bit2 = bit2 & (0-bit2); // Extract 2nd bit
if (bit2 + bit1 == abs_con) { // Found all bits in con?
Node *n1 = phase->transform( new (phase->C) LShiftINode(in(1), phase->intcon(log2_uint(bit1))));
Node *n2 = phase->transform( new (phase->C) LShiftINode(in(1), phase->intcon(log2_uint(bit2))));
res = new (phase->C) AddINode(n2, n1);
} else if (is_power_of_2(abs_con+1)) {
// Sleezy: power-of-2 -1. Next time be generic.
unsigned int temp = abs_con + 1;
Node *n1 = phase->transform(new (phase->C) LShiftINode(in(1), phase->intcon(log2_uint(temp))));
res = new (phase->C) SubINode(n1, in(1));
} else {
return MulNode::Ideal(phase, can_reshape);
}
}
if (sign_flip) { // Need to negate result?
res = phase->transform(res);// Transform, before making the zero con
res = new (phase->C) SubINode(phase->intcon(0),res);
}
return res; // Return final result
}
//------------------------------mul_ring---------------------------------------
// Compute the product type of two integer ranges into this node.
const Type *MulINode::mul_ring(const Type *t0, const Type *t1) const {
const TypeInt *r0 = t0->is_int(); // Handy access
const TypeInt *r1 = t1->is_int();
// Fetch endpoints of all ranges
int32 lo0 = r0->_lo;
double a = (double)lo0;
int32 hi0 = r0->_hi;
double b = (double)hi0;
int32 lo1 = r1->_lo;
double c = (double)lo1;
int32 hi1 = r1->_hi;
double d = (double)hi1;
// Compute all endpoints & check for overflow
int32 A = java_multiply(lo0, lo1);
if( (double)A != a*c ) return TypeInt::INT; // Overflow?
int32 B = java_multiply(lo0, hi1);
if( (double)B != a*d ) return TypeInt::INT; // Overflow?
int32 C = java_multiply(hi0, lo1);
if( (double)C != b*c ) return TypeInt::INT; // Overflow?
int32 D = java_multiply(hi0, hi1);
if( (double)D != b*d ) return TypeInt::INT; // Overflow?
if( A < B ) { lo0 = A; hi0 = B; } // Sort range endpoints
else { lo0 = B; hi0 = A; }
if( C < D ) {
if( C < lo0 ) lo0 = C;
if( D > hi0 ) hi0 = D;
} else {
if( D < lo0 ) lo0 = D;
if( C > hi0 ) hi0 = C;
}
return TypeInt::make(lo0, hi0, MAX2(r0->_widen,r1->_widen));
}
//=============================================================================
//------------------------------Ideal------------------------------------------
// Check for power-of-2 multiply, then try the regular MulNode::Ideal
Node *MulLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
// Swap constant to right
jlong con;
if ((con = in(1)->find_long_con(0)) != 0) {
swap_edges(1, 2);
// Finish rest of method to use info in 'con'
} else if ((con = in(2)->find_long_con(0)) == 0) {
return MulNode::Ideal(phase, can_reshape);
}
// Now we have a constant Node on the right and the constant in con
if (con == CONST64(0)) return NULL; // By zero is handled by Value call
if (con == CONST64(1)) return NULL; // By one is handled by Identity call
// Check for negative constant; if so negate the final result
bool sign_flip = false;
julong abs_con = uabs(con);
if (abs_con != (julong)con) {
sign_flip = true;
}
// Get low bit; check for being the only bit
Node *res = NULL;
julong bit1 = abs_con & (0-abs_con); // Extract low bit
if (bit1 == abs_con) { // Found a power of 2?
res = new (phase->C) LShiftLNode(in(1), phase->intcon(log2_long(bit1)));
} else {
// Check for constant with 2 bits set
julong bit2 = abs_con-bit1;
bit2 = bit2 & (0-bit2); // Extract 2nd bit
if (bit2 + bit1 == abs_con) { // Found all bits in con?
Node *n1 = phase->transform(new (phase->C) LShiftLNode(in(1), phase->intcon(log2_long(bit1))));
Node *n2 = phase->transform(new (phase->C) LShiftLNode(in(1), phase->intcon(log2_long(bit2))));
res = new (phase->C) AddLNode(n2, n1);
} else if (is_power_of_2_long(abs_con+1)) {
// Sleezy: power-of-2 -1. Next time be generic.
julong temp = abs_con + 1;
Node *n1 = phase->transform( new (phase->C) LShiftLNode(in(1), phase->intcon(log2_long(temp))));
res = new (phase->C) SubLNode(n1, in(1));
} else {
return MulNode::Ideal(phase, can_reshape);
}
}
if (sign_flip) { // Need to negate result?
res = phase->transform(res);// Transform, before making the zero con
res = new (phase->C) SubLNode(phase->longcon(0),res);
}
return res; // Return final result
}
//------------------------------mul_ring---------------------------------------
// Compute the product type of two integer ranges into this node.
const Type *MulLNode::mul_ring(const Type *t0, const Type *t1) const {
const TypeLong *r0 = t0->is_long(); // Handy access
const TypeLong *r1 = t1->is_long();
// Fetch endpoints of all ranges
jlong lo0 = r0->_lo;
double a = (double)lo0;
jlong hi0 = r0->_hi;
double b = (double)hi0;
jlong lo1 = r1->_lo;
double c = (double)lo1;
jlong hi1 = r1->_hi;
double d = (double)hi1;
// Compute all endpoints & check for overflow
jlong A = java_multiply(lo0, lo1);
if( (double)A != a*c ) return TypeLong::LONG; // Overflow?
jlong B = java_multiply(lo0, hi1);
if( (double)B != a*d ) return TypeLong::LONG; // Overflow?
jlong C = java_multiply(hi0, lo1);
if( (double)C != b*c ) return TypeLong::LONG; // Overflow?
jlong D = java_multiply(hi0, hi1);
if( (double)D != b*d ) return TypeLong::LONG; // Overflow?
if( A < B ) { lo0 = A; hi0 = B; } // Sort range endpoints
else { lo0 = B; hi0 = A; }
if( C < D ) {
if( C < lo0 ) lo0 = C;
if( D > hi0 ) hi0 = D;
} else {
if( D < lo0 ) lo0 = D;
if( C > hi0 ) hi0 = C;
}
return TypeLong::make(lo0, hi0, MAX2(r0->_widen,r1->_widen));
}
//=============================================================================
//------------------------------mul_ring---------------------------------------
// Compute the product type of two double ranges into this node.
const Type *MulFNode::mul_ring(const Type *t0, const Type *t1) const {
if( t0 == Type::FLOAT || t1 == Type::FLOAT ) return Type::FLOAT;
return TypeF::make( t0->getf() * t1->getf() );
}
//=============================================================================
//------------------------------mul_ring---------------------------------------
// Compute the product type of two double ranges into this node.
const Type *MulDNode::mul_ring(const Type *t0, const Type *t1) const {
if( t0 == Type::DOUBLE || t1 == Type::DOUBLE ) return Type::DOUBLE;
// We must be multiplying 2 double constants.
return TypeD::make( t0->getd() * t1->getd() );
}
//=============================================================================
//------------------------------Value------------------------------------------
const Type *MulHiLNode::Value( PhaseTransform *phase ) const {
// Either input is TOP ==> the result is TOP
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
if( t1 == Type::TOP ) return Type::TOP;
if( t2 == Type::TOP ) return Type::TOP;
// Either input is BOTTOM ==> the result is the local BOTTOM
const Type *bot = bottom_type();
if( (t1 == bot) || (t2 == bot) ||
(t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
return bot;
// It is not worth trying to constant fold this stuff!
return TypeLong::LONG;
}
//=============================================================================
//------------------------------mul_ring---------------------------------------
// Supplied function returns the product of the inputs IN THE CURRENT RING.
// For the logical operations the ring's MUL is really a logical AND function.
// This also type-checks the inputs for sanity. Guaranteed never to
// be passed a TOP or BOTTOM type, these are filtered out by pre-check.
const Type *AndINode::mul_ring( const Type *t0, const Type *t1 ) const {
const TypeInt *r0 = t0->is_int(); // Handy access
const TypeInt *r1 = t1->is_int();
int widen = MAX2(r0->_widen,r1->_widen);
// If either input is a constant, might be able to trim cases
if( !r0->is_con() && !r1->is_con() )
return TypeInt::INT; // No constants to be had
// Both constants? Return bits
if( r0->is_con() && r1->is_con() )
return TypeInt::make( r0->get_con() & r1->get_con() );
if( r0->is_con() && r0->get_con() > 0 )
return TypeInt::make(0, r0->get_con(), widen);
if( r1->is_con() && r1->get_con() > 0 )
return TypeInt::make(0, r1->get_con(), widen);
if( r0 == TypeInt::BOOL || r1 == TypeInt::BOOL ) {
return TypeInt::BOOL;
}
return TypeInt::INT; // No constants to be had
}
//------------------------------Identity---------------------------------------
// Masking off the high bits of an unsigned load is not required
Node *AndINode::Identity( PhaseTransform *phase ) {
// x & x => x
if (phase->eqv(in(1), in(2))) return in(1);
Node* in1 = in(1);
uint op = in1->Opcode();
const TypeInt* t2 = phase->type(in(2))->isa_int();
if (t2 && t2->is_con()) {
int con = t2->get_con();
// Masking off high bits which are always zero is useless.
const TypeInt* t1 = phase->type( in(1) )->isa_int();
if (t1 != NULL && t1->_lo >= 0) {
jint t1_support = right_n_bits(1 + log2_jint(t1->_hi));
if ((t1_support & con) == t1_support)
return in1;
}
// Masking off the high bits of a unsigned-shift-right is not
// needed either.
if (op == Op_URShiftI) {
const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
if (t12 && t12->is_con()) { // Shift is by a constant
int shift = t12->get_con();
shift &= BitsPerJavaInteger - 1; // semantics of Java shifts
int mask = max_juint >> shift;
if ((mask & con) == mask) // If AND is useless, skip it
return in1;
}
}
}
return MulNode::Identity(phase);
}
//------------------------------Ideal------------------------------------------
Node *AndINode::Ideal(PhaseGVN *phase, bool can_reshape) {
// Special case constant AND mask
const TypeInt *t2 = phase->type( in(2) )->isa_int();
if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
const int mask = t2->get_con();
Node *load = in(1);
uint lop = load->Opcode();
// Masking bits off of a Character? Hi bits are already zero.
if( lop == Op_LoadUS &&
(mask & 0xFFFF0000) ) // Can we make a smaller mask?
return new (phase->C) AndINode(load,phase->intcon(mask&0xFFFF));
// Masking bits off of a Short? Loading a Character does some masking
if (can_reshape &&
load->outcnt() == 1 && load->unique_out() == this) {
if (lop == Op_LoadS && (mask & 0xFFFF0000) == 0 ) {
Node *ldus = new (phase->C) LoadUSNode(load->in(MemNode::Control),
load->in(MemNode::Memory),
load->in(MemNode::Address),
load->adr_type(),
TypeInt::CHAR, MemNode::unordered);
ldus = phase->transform(ldus);
return new (phase->C) AndINode(ldus, phase->intcon(mask & 0xFFFF));
}
// Masking sign bits off of a Byte? Do an unsigned byte load plus
// an and.
if (lop == Op_LoadB && (mask & 0xFFFFFF00) == 0) {
Node* ldub = new (phase->C) LoadUBNode(load->in(MemNode::Control),
load->in(MemNode::Memory),
load->in(MemNode::Address),
load->adr_type(),
TypeInt::UBYTE, MemNode::unordered);
ldub = phase->transform(ldub);
return new (phase->C) AndINode(ldub, phase->intcon(mask));
}
}
// Masking off sign bits? Dont make them!
if( lop == Op_RShiftI ) {
const TypeInt *t12 = phase->type(load->in(2))->isa_int();
if( t12 && t12->is_con() ) { // Shift is by a constant
int shift = t12->get_con();
shift &= BitsPerJavaInteger-1; // semantics of Java shifts
const int sign_bits_mask = ~right_n_bits(BitsPerJavaInteger - shift);
// If the AND'ing of the 2 masks has no bits, then only original shifted
// bits survive. NO sign-extension bits survive the maskings.
if( (sign_bits_mask & mask) == 0 ) {
// Use zero-fill shift instead
Node *zshift = phase->transform(new (phase->C) URShiftINode(load->in(1),load->in(2)));
return new (phase->C) AndINode( zshift, in(2) );
}
}
}
// Check for 'negate/and-1', a pattern emitted when someone asks for
// 'mod 2'. Negate leaves the low order bit unchanged (think: complement
// plus 1) and the mask is of the low order bit. Skip the negate.
if( lop == Op_SubI && mask == 1 && load->in(1) &&
phase->type(load->in(1)) == TypeInt::ZERO )
return new (phase->C) AndINode( load->in(2), in(2) );
return MulNode::Ideal(phase, can_reshape);
}
//=============================================================================
//------------------------------mul_ring---------------------------------------
// Supplied function returns the product of the inputs IN THE CURRENT RING.
// For the logical operations the ring's MUL is really a logical AND function.
// This also type-checks the inputs for sanity. Guaranteed never to
// be passed a TOP or BOTTOM type, these are filtered out by pre-check.
const Type *AndLNode::mul_ring( const Type *t0, const Type *t1 ) const {
const TypeLong *r0 = t0->is_long(); // Handy access
const TypeLong *r1 = t1->is_long();
int widen = MAX2(r0->_widen,r1->_widen);
// If either input is a constant, might be able to trim cases
if( !r0->is_con() && !r1->is_con() )
return TypeLong::LONG; // No constants to be had
// Both constants? Return bits
if( r0->is_con() && r1->is_con() )
return TypeLong::make( r0->get_con() & r1->get_con() );
if( r0->is_con() && r0->get_con() > 0 )
return TypeLong::make(CONST64(0), r0->get_con(), widen);
if( r1->is_con() && r1->get_con() > 0 )
return TypeLong::make(CONST64(0), r1->get_con(), widen);
return TypeLong::LONG; // No constants to be had
}
//------------------------------Identity---------------------------------------
// Masking off the high bits of an unsigned load is not required
Node *AndLNode::Identity( PhaseTransform *phase ) {
// x & x => x
if (phase->eqv(in(1), in(2))) return in(1);
Node *usr = in(1);
const TypeLong *t2 = phase->type( in(2) )->isa_long();
if( t2 && t2->is_con() ) {
jlong con = t2->get_con();
// Masking off high bits which are always zero is useless.
const TypeLong* t1 = phase->type( in(1) )->isa_long();
if (t1 != NULL && t1->_lo >= 0) {
int bit_count = log2_long(t1->_hi) + 1;
jlong t1_support = jlong(max_julong >> (BitsPerJavaLong - bit_count));
if ((t1_support & con) == t1_support)
return usr;
}
uint lop = usr->Opcode();
// Masking off the high bits of a unsigned-shift-right is not
// needed either.
if( lop == Op_URShiftL ) {
const TypeInt *t12 = phase->type( usr->in(2) )->isa_int();
if( t12 && t12->is_con() ) { // Shift is by a constant
int shift = t12->get_con();
shift &= BitsPerJavaLong - 1; // semantics of Java shifts
jlong mask = max_julong >> shift;
if( (mask&con) == mask ) // If AND is useless, skip it
return usr;
}
}
}
return MulNode::Identity(phase);
}
//------------------------------Ideal------------------------------------------
Node *AndLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
// Special case constant AND mask
const TypeLong *t2 = phase->type( in(2) )->isa_long();
if( !t2 || !t2->is_con() ) return MulNode::Ideal(phase, can_reshape);
const jlong mask = t2->get_con();
Node* in1 = in(1);
uint op = in1->Opcode();
// Are we masking a long that was converted from an int with a mask
// that fits in 32-bits? Commute them and use an AndINode. Don't
// convert masks which would cause a sign extension of the integer
// value. This check includes UI2L masks (0x00000000FFFFFFFF) which
// would be optimized away later in Identity.
if (op == Op_ConvI2L && (mask & CONST64(0xFFFFFFFF80000000)) == 0) {
Node* andi = new (phase->C) AndINode(in1->in(1), phase->intcon(mask));
andi = phase->transform(andi);
return new (phase->C) ConvI2LNode(andi);
}
// Masking off sign bits? Dont make them!
if (op == Op_RShiftL) {
const TypeInt* t12 = phase->type(in1->in(2))->isa_int();
if( t12 && t12->is_con() ) { // Shift is by a constant
int shift = t12->get_con();
shift &= BitsPerJavaLong - 1; // semantics of Java shifts
const jlong sign_bits_mask = ~(((jlong)CONST64(1) << (jlong)(BitsPerJavaLong - shift)) -1);
// If the AND'ing of the 2 masks has no bits, then only original shifted
// bits survive. NO sign-extension bits survive the maskings.
if( (sign_bits_mask & mask) == 0 ) {
// Use zero-fill shift instead
Node *zshift = phase->transform(new (phase->C) URShiftLNode(in1->in(1), in1->in(2)));
return new (phase->C) AndLNode(zshift, in(2));
}
}
}
return MulNode::Ideal(phase, can_reshape);
}
//=============================================================================
//------------------------------Identity---------------------------------------
Node *LShiftINode::Identity( PhaseTransform *phase ) {
const TypeInt *ti = phase->type( in(2) )->isa_int(); // shift count is an int
return ( ti && ti->is_con() && ( ti->get_con() & ( BitsPerInt - 1 ) ) == 0 ) ? in(1) : this;
}
//------------------------------Ideal------------------------------------------
// If the right input is a constant, and the left input is an add of a
// constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0
Node *LShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
const Type *t = phase->type( in(2) );
if( t == Type::TOP ) return NULL; // Right input is dead
const TypeInt *t2 = t->isa_int();
if( !t2 || !t2->is_con() ) return NULL; // Right input is a constant
const int con = t2->get_con() & ( BitsPerInt - 1 ); // masked shift count
if ( con == 0 ) return NULL; // let Identity() handle 0 shift count
// Left input is an add of a constant?
Node *add1 = in(1);
int add1_op = add1->Opcode();
if( add1_op == Op_AddI ) { // Left input is an add?
assert( add1 != add1->in(1), "dead loop in LShiftINode::Ideal" );
const TypeInt *t12 = phase->type(add1->in(2))->isa_int();
if( t12 && t12->is_con() ){ // Left input is an add of a con?
// Transform is legal, but check for profit. Avoid breaking 'i2s'
// and 'i2b' patterns which typically fold into 'StoreC/StoreB'.
if( con < 16 ) {
// Compute X << con0
Node *lsh = phase->transform( new (phase->C) LShiftINode( add1->in(1), in(2) ) );
// Compute X<<con0 + (con1<<con0)
return new (phase->C) AddINode( lsh, phase->intcon(t12->get_con() << con));
}
}
}
// Check for "(x>>c0)<<c0" which just masks off low bits
if( (add1_op == Op_RShiftI || add1_op == Op_URShiftI ) &&
add1->in(2) == in(2) )
// Convert to "(x & -(1<<c0))"
return new (phase->C) AndINode(add1->in(1),phase->intcon( -(1<<con)));
// Check for "((x>>c0) & Y)<<c0" which just masks off more low bits
if( add1_op == Op_AndI ) {
Node *add2 = add1->in(1);
int add2_op = add2->Opcode();
if( (add2_op == Op_RShiftI || add2_op == Op_URShiftI ) &&
add2->in(2) == in(2) ) {
// Convert to "(x & (Y<<c0))"
Node *y_sh = phase->transform( new (phase->C) LShiftINode( add1->in(2), in(2) ) );
return new (phase->C) AndINode( add2->in(1), y_sh );
}
}
// Check for ((x & ((1<<(32-c0))-1)) << c0) which ANDs off high bits
// before shifting them away.
const jint bits_mask = right_n_bits(BitsPerJavaInteger-con);
if( add1_op == Op_AndI &&
phase->type(add1->in(2)) == TypeInt::make( bits_mask ) )
return new (phase->C) LShiftINode( add1->in(1), in(2) );
return NULL;
}
//------------------------------Value------------------------------------------
// A LShiftINode shifts its input2 left by input1 amount.
const Type *LShiftINode::Value( PhaseTransform *phase ) const {
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
// Either input is TOP ==> the result is TOP
if( t1 == Type::TOP ) return Type::TOP;
if( t2 == Type::TOP ) return Type::TOP;
// Left input is ZERO ==> the result is ZERO.
if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
// Shift by zero does nothing
if( t2 == TypeInt::ZERO ) return t1;
// Either input is BOTTOM ==> the result is BOTTOM
if( (t1 == TypeInt::INT) || (t2 == TypeInt::INT) ||
(t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
return TypeInt::INT;
const TypeInt *r1 = t1->is_int(); // Handy access
const TypeInt *r2 = t2->is_int(); // Handy access
if (!r2->is_con())
return TypeInt::INT;
uint shift = r2->get_con();
shift &= BitsPerJavaInteger-1; // semantics of Java shifts
// Shift by a multiple of 32 does nothing:
if (shift == 0) return t1;
// If the shift is a constant, shift the bounds of the type,
// unless this could lead to an overflow.
if (!r1->is_con()) {
jint lo = r1->_lo, hi = r1->_hi;
if (((lo << shift) >> shift) == lo &&
((hi << shift) >> shift) == hi) {
// No overflow. The range shifts up cleanly.
return TypeInt::make((jint)lo << (jint)shift,
(jint)hi << (jint)shift,
MAX2(r1->_widen,r2->_widen));
}
return TypeInt::INT;
}
return TypeInt::make( (jint)r1->get_con() << (jint)shift );
}
//=============================================================================
//------------------------------Identity---------------------------------------
Node *LShiftLNode::Identity( PhaseTransform *phase ) {
const TypeInt *ti = phase->type( in(2) )->isa_int(); // shift count is an int
return ( ti && ti->is_con() && ( ti->get_con() & ( BitsPerLong - 1 ) ) == 0 ) ? in(1) : this;
}
//------------------------------Ideal------------------------------------------
// If the right input is a constant, and the left input is an add of a
// constant, flatten the tree: (X+con1)<<con0 ==> X<<con0 + con1<<con0
Node *LShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
const Type *t = phase->type( in(2) );
if( t == Type::TOP ) return NULL; // Right input is dead
const TypeInt *t2 = t->isa_int();
if( !t2 || !t2->is_con() ) return NULL; // Right input is a constant
const int con = t2->get_con() & ( BitsPerLong - 1 ); // masked shift count
if ( con == 0 ) return NULL; // let Identity() handle 0 shift count
// Left input is an add of a constant?
Node *add1 = in(1);
int add1_op = add1->Opcode();
if( add1_op == Op_AddL ) { // Left input is an add?
// Avoid dead data cycles from dead loops
assert( add1 != add1->in(1), "dead loop in LShiftLNode::Ideal" );
const TypeLong *t12 = phase->type(add1->in(2))->isa_long();
if( t12 && t12->is_con() ){ // Left input is an add of a con?
// Compute X << con0
Node *lsh = phase->transform( new (phase->C) LShiftLNode( add1->in(1), in(2) ) );
// Compute X<<con0 + (con1<<con0)
return new (phase->C) AddLNode( lsh, phase->longcon(t12->get_con() << con));
}
}
// Check for "(x>>c0)<<c0" which just masks off low bits
if( (add1_op == Op_RShiftL || add1_op == Op_URShiftL ) &&
add1->in(2) == in(2) )
// Convert to "(x & -(1<<c0))"
return new (phase->C) AndLNode(add1->in(1),phase->longcon( -(CONST64(1)<<con)));
// Check for "((x>>c0) & Y)<<c0" which just masks off more low bits
if( add1_op == Op_AndL ) {
Node *add2 = add1->in(1);
int add2_op = add2->Opcode();
if( (add2_op == Op_RShiftL || add2_op == Op_URShiftL ) &&
add2->in(2) == in(2) ) {
// Convert to "(x & (Y<<c0))"
Node *y_sh = phase->transform( new (phase->C) LShiftLNode( add1->in(2), in(2) ) );
return new (phase->C) AndLNode( add2->in(1), y_sh );
}
}
// Check for ((x & ((CONST64(1)<<(64-c0))-1)) << c0) which ANDs off high bits
// before shifting them away.
const jlong bits_mask = jlong(max_julong >> con);
if( add1_op == Op_AndL &&
phase->type(add1->in(2)) == TypeLong::make( bits_mask ) )
return new (phase->C) LShiftLNode( add1->in(1), in(2) );
return NULL;
}
//------------------------------Value------------------------------------------
// A LShiftLNode shifts its input2 left by input1 amount.
const Type *LShiftLNode::Value( PhaseTransform *phase ) const {
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
// Either input is TOP ==> the result is TOP
if( t1 == Type::TOP ) return Type::TOP;
if( t2 == Type::TOP ) return Type::TOP;
// Left input is ZERO ==> the result is ZERO.
if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
// Shift by zero does nothing
if( t2 == TypeInt::ZERO ) return t1;
// Either input is BOTTOM ==> the result is BOTTOM
if( (t1 == TypeLong::LONG) || (t2 == TypeInt::INT) ||
(t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
return TypeLong::LONG;
const TypeLong *r1 = t1->is_long(); // Handy access
const TypeInt *r2 = t2->is_int(); // Handy access
if (!r2->is_con())
return TypeLong::LONG;
uint shift = r2->get_con();
shift &= BitsPerJavaLong - 1; // semantics of Java shifts
// Shift by a multiple of 64 does nothing:
if (shift == 0) return t1;
// If the shift is a constant, shift the bounds of the type,
// unless this could lead to an overflow.
if (!r1->is_con()) {
jlong lo = r1->_lo, hi = r1->_hi;
if (((lo << shift) >> shift) == lo &&
((hi << shift) >> shift) == hi) {
// No overflow. The range shifts up cleanly.
return TypeLong::make((jlong)lo << (jint)shift,
(jlong)hi << (jint)shift,
MAX2(r1->_widen,r2->_widen));
}
return TypeLong::LONG;
}
return TypeLong::make( (jlong)r1->get_con() << (jint)shift );
}
//=============================================================================
//------------------------------Identity---------------------------------------
Node *RShiftINode::Identity( PhaseTransform *phase ) {
const TypeInt *t2 = phase->type(in(2))->isa_int();
if( !t2 ) return this;
if ( t2->is_con() && ( t2->get_con() & ( BitsPerInt - 1 ) ) == 0 )
return in(1);
// Check for useless sign-masking
if( in(1)->Opcode() == Op_LShiftI &&
in(1)->req() == 3 &&
in(1)->in(2) == in(2) &&
t2->is_con() ) {
uint shift = t2->get_con();
shift &= BitsPerJavaInteger-1; // semantics of Java shifts
// Compute masks for which this shifting doesn't change
int lo = (-1 << (BitsPerJavaInteger - shift-1)); // FFFF8000
int hi = ~lo; // 00007FFF
const TypeInt *t11 = phase->type(in(1)->in(1))->isa_int();
if( !t11 ) return this;
// Does actual value fit inside of mask?
if( lo <= t11->_lo && t11->_hi <= hi )
return in(1)->in(1); // Then shifting is a nop
}
return this;
}
//------------------------------Ideal------------------------------------------
Node *RShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
// Inputs may be TOP if they are dead.
const TypeInt *t1 = phase->type( in(1) )->isa_int();
if( !t1 ) return NULL; // Left input is an integer
const TypeInt *t2 = phase->type( in(2) )->isa_int();
if( !t2 || !t2->is_con() ) return NULL; // Right input is a constant
const TypeInt *t3; // type of in(1).in(2)
int shift = t2->get_con();
shift &= BitsPerJavaInteger-1; // semantics of Java shifts
if ( shift == 0 ) return NULL; // let Identity() handle 0 shift count
// Check for (x & 0xFF000000) >> 24, whose mask can be made smaller.
// Such expressions arise normally from shift chains like (byte)(x >> 24).
const Node *mask = in(1);
if( mask->Opcode() == Op_AndI &&
(t3 = phase->type(mask->in(2))->isa_int()) &&
t3->is_con() ) {
Node *x = mask->in(1);
jint maskbits = t3->get_con();
// Convert to "(x >> shift) & (mask >> shift)"
Node *shr_nomask = phase->transform( new (phase->C) RShiftINode(mask->in(1), in(2)) );
return new (phase->C) AndINode(shr_nomask, phase->intcon( maskbits >> shift));
}
// Check for "(short[i] <<16)>>16" which simply sign-extends
const Node *shl = in(1);
if( shl->Opcode() != Op_LShiftI ) return NULL;
if( shift == 16 &&
(t3 = phase->type(shl->in(2))->isa_int()) &&
t3->is_con(16) ) {
Node *ld = shl->in(1);
if( ld->Opcode() == Op_LoadS ) {
// Sign extension is just useless here. Return a RShiftI of zero instead
// returning 'ld' directly. We cannot return an old Node directly as
// that is the job of 'Identity' calls and Identity calls only work on
// direct inputs ('ld' is an extra Node removed from 'this'). The
// combined optimization requires Identity only return direct inputs.
set_req(1, ld);
set_req(2, phase->intcon(0));
return this;
}
else if( can_reshape &&
ld->Opcode() == Op_LoadUS &&
ld->outcnt() == 1 && ld->unique_out() == shl)
// Replace zero-extension-load with sign-extension-load
return new (phase->C) LoadSNode( ld->in(MemNode::Control),
ld->in(MemNode::Memory),
ld->in(MemNode::Address),
ld->adr_type(), TypeInt::SHORT,
MemNode::unordered);
}
// Check for "(byte[i] <<24)>>24" which simply sign-extends
if( shift == 24 &&
(t3 = phase->type(shl->in(2))->isa_int()) &&
t3->is_con(24) ) {
Node *ld = shl->in(1);
if( ld->Opcode() == Op_LoadB ) {
// Sign extension is just useless here
set_req(1, ld);
set_req(2, phase->intcon(0));
return this;
}
}
return NULL;
}
//------------------------------Value------------------------------------------
// A RShiftINode shifts its input2 right by input1 amount.
const Type *RShiftINode::Value( PhaseTransform *phase ) const {
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
// Either input is TOP ==> the result is TOP
if( t1 == Type::TOP ) return Type::TOP;
if( t2 == Type::TOP ) return Type::TOP;
// Left input is ZERO ==> the result is ZERO.
if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
// Shift by zero does nothing
if( t2 == TypeInt::ZERO ) return t1;
// Either input is BOTTOM ==> the result is BOTTOM
if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
return TypeInt::INT;
if (t2 == TypeInt::INT)
return TypeInt::INT;
const TypeInt *r1 = t1->is_int(); // Handy access
const TypeInt *r2 = t2->is_int(); // Handy access
// If the shift is a constant, just shift the bounds of the type.
// For example, if the shift is 31, we just propagate sign bits.
if (r2->is_con()) {
uint shift = r2->get_con();
shift &= BitsPerJavaInteger-1; // semantics of Java shifts
// Shift by a multiple of 32 does nothing:
if (shift == 0) return t1;
// Calculate reasonably aggressive bounds for the result.
// This is necessary if we are to correctly type things
// like (x<<24>>24) == ((byte)x).
jint lo = (jint)r1->_lo >> (jint)shift;
jint hi = (jint)r1->_hi >> (jint)shift;
assert(lo <= hi, "must have valid bounds");
const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen));
#ifdef ASSERT
// Make sure we get the sign-capture idiom correct.
if (shift == BitsPerJavaInteger-1) {
if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>31 of + is 0");
if (r1->_hi < 0) assert(ti == TypeInt::MINUS_1, ">>31 of - is -1");
}
#endif
return ti;
}
if( !r1->is_con() || !r2->is_con() )
return TypeInt::INT;
// Signed shift right
return TypeInt::make( r1->get_con() >> (r2->get_con()&31) );
}
//=============================================================================
//------------------------------Identity---------------------------------------
Node *RShiftLNode::Identity( PhaseTransform *phase ) {
const TypeInt *ti = phase->type( in(2) )->isa_int(); // shift count is an int
return ( ti && ti->is_con() && ( ti->get_con() & ( BitsPerLong - 1 ) ) == 0 ) ? in(1) : this;
}
//------------------------------Value------------------------------------------
// A RShiftLNode shifts its input2 right by input1 amount.
const Type *RShiftLNode::Value( PhaseTransform *phase ) const {
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
// Either input is TOP ==> the result is TOP
if( t1 == Type::TOP ) return Type::TOP;
if( t2 == Type::TOP ) return Type::TOP;
// Left input is ZERO ==> the result is ZERO.
if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
// Shift by zero does nothing
if( t2 == TypeInt::ZERO ) return t1;
// Either input is BOTTOM ==> the result is BOTTOM
if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
return TypeLong::LONG;
if (t2 == TypeInt::INT)
return TypeLong::LONG;
const TypeLong *r1 = t1->is_long(); // Handy access
const TypeInt *r2 = t2->is_int (); // Handy access
// If the shift is a constant, just shift the bounds of the type.
// For example, if the shift is 63, we just propagate sign bits.
if (r2->is_con()) {
uint shift = r2->get_con();
shift &= (2*BitsPerJavaInteger)-1; // semantics of Java shifts
// Shift by a multiple of 64 does nothing:
if (shift == 0) return t1;
// Calculate reasonably aggressive bounds for the result.
// This is necessary if we are to correctly type things
// like (x<<24>>24) == ((byte)x).
jlong lo = (jlong)r1->_lo >> (jlong)shift;
jlong hi = (jlong)r1->_hi >> (jlong)shift;
assert(lo <= hi, "must have valid bounds");
const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen));
#ifdef ASSERT
// Make sure we get the sign-capture idiom correct.
if (shift == (2*BitsPerJavaInteger)-1) {
if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>63 of + is 0");
if (r1->_hi < 0) assert(tl == TypeLong::MINUS_1, ">>63 of - is -1");
}
#endif
return tl;
}
return TypeLong::LONG; // Give up
}
//=============================================================================
//------------------------------Identity---------------------------------------
Node *URShiftINode::Identity( PhaseTransform *phase ) {
const TypeInt *ti = phase->type( in(2) )->isa_int();
if ( ti && ti->is_con() && ( ti->get_con() & ( BitsPerInt - 1 ) ) == 0 ) return in(1);
// Check for "((x << LogBytesPerWord) + (wordSize-1)) >> LogBytesPerWord" which is just "x".
// Happens during new-array length computation.
// Safe if 'x' is in the range [0..(max_int>>LogBytesPerWord)]
Node *add = in(1);
if( add->Opcode() == Op_AddI ) {
const TypeInt *t2 = phase->type(add->in(2))->isa_int();
if( t2 && t2->is_con(wordSize - 1) &&
add->in(1)->Opcode() == Op_LShiftI ) {
// Check that shift_counts are LogBytesPerWord
Node *lshift_count = add->in(1)->in(2);
const TypeInt *t_lshift_count = phase->type(lshift_count)->isa_int();
if( t_lshift_count && t_lshift_count->is_con(LogBytesPerWord) &&
t_lshift_count == phase->type(in(2)) ) {
Node *x = add->in(1)->in(1);
const TypeInt *t_x = phase->type(x)->isa_int();
if( t_x != NULL && 0 <= t_x->_lo && t_x->_hi <= (max_jint>>LogBytesPerWord) ) {
return x;
}
}
}
}
return (phase->type(in(2))->higher_equal(TypeInt::ZERO)) ? in(1) : this;
}
//------------------------------Ideal------------------------------------------
Node *URShiftINode::Ideal(PhaseGVN *phase, bool can_reshape) {
const TypeInt *t2 = phase->type( in(2) )->isa_int();
if( !t2 || !t2->is_con() ) return NULL; // Right input is a constant
const int con = t2->get_con() & 31; // Shift count is always masked
if ( con == 0 ) return NULL; // let Identity() handle a 0 shift count
// We'll be wanting the right-shift amount as a mask of that many bits
const int mask = right_n_bits(BitsPerJavaInteger - con);
int in1_op = in(1)->Opcode();
// Check for ((x>>>a)>>>b) and replace with (x>>>(a+b)) when a+b < 32
if( in1_op == Op_URShiftI ) {
const TypeInt *t12 = phase->type( in(1)->in(2) )->isa_int();
if( t12 && t12->is_con() ) { // Right input is a constant
assert( in(1) != in(1)->in(1), "dead loop in URShiftINode::Ideal" );
const int con2 = t12->get_con() & 31; // Shift count is always masked
const int con3 = con+con2;
if( con3 < 32 ) // Only merge shifts if total is < 32
return new (phase->C) URShiftINode( in(1)->in(1), phase->intcon(con3) );
}
}
// Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
// The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
// If Q is "X << z" the rounding is useless. Look for patterns like
// ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
Node *add = in(1);
if( in1_op == Op_AddI ) {
Node *lshl = add->in(1);
if( lshl->Opcode() == Op_LShiftI &&
phase->type(lshl->in(2)) == t2 ) {
Node *y_z = phase->transform( new (phase->C) URShiftINode(add->in(2),in(2)) );
Node *sum = phase->transform( new (phase->C) AddINode( lshl->in(1), y_z ) );
return new (phase->C) AndINode( sum, phase->intcon(mask) );
}
}
// Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
// This shortens the mask. Also, if we are extracting a high byte and
// storing it to a buffer, the mask will be removed completely.
Node *andi = in(1);
if( in1_op == Op_AndI ) {
const TypeInt *t3 = phase->type( andi->in(2) )->isa_int();
if( t3 && t3->is_con() ) { // Right input is a constant
jint mask2 = t3->get_con();
mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
Node *newshr = phase->transform( new (phase->C) URShiftINode(andi->in(1), in(2)) );
return new (phase->C) AndINode(newshr, phase->intcon(mask2));
// The negative values are easier to materialize than positive ones.
// A typical case from address arithmetic is ((x & ~15) >> 4).
// It's better to change that to ((x >> 4) & ~0) versus
// ((x >> 4) & 0x0FFFFFFF). The difference is greatest in LP64.
}
}
// Check for "(X << z ) >>> z" which simply zero-extends
Node *shl = in(1);
if( in1_op == Op_LShiftI &&
phase->type(shl->in(2)) == t2 )
return new (phase->C) AndINode( shl->in(1), phase->intcon(mask) );
return NULL;
}
//------------------------------Value------------------------------------------
// A URShiftINode shifts its input2 right by input1 amount.
const Type *URShiftINode::Value( PhaseTransform *phase ) const {
// (This is a near clone of RShiftINode::Value.)
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
// Either input is TOP ==> the result is TOP
if( t1 == Type::TOP ) return Type::TOP;
if( t2 == Type::TOP ) return Type::TOP;
// Left input is ZERO ==> the result is ZERO.
if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
// Shift by zero does nothing
if( t2 == TypeInt::ZERO ) return t1;
// Either input is BOTTOM ==> the result is BOTTOM
if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
return TypeInt::INT;
if (t2 == TypeInt::INT)
return TypeInt::INT;
const TypeInt *r1 = t1->is_int(); // Handy access
const TypeInt *r2 = t2->is_int(); // Handy access
if (r2->is_con()) {
uint shift = r2->get_con();
shift &= BitsPerJavaInteger-1; // semantics of Java shifts
// Shift by a multiple of 32 does nothing:
if (shift == 0) return t1;
// Calculate reasonably aggressive bounds for the result.
jint lo = (juint)r1->_lo >> (juint)shift;
jint hi = (juint)r1->_hi >> (juint)shift;
if (r1->_hi >= 0 && r1->_lo < 0) {
// If the type has both negative and positive values,
// there are two separate sub-domains to worry about:
// The positive half and the negative half.
jint neg_lo = lo;
jint neg_hi = (juint)-1 >> (juint)shift;
jint pos_lo = (juint) 0 >> (juint)shift;
jint pos_hi = hi;
lo = MIN2(neg_lo, pos_lo); // == 0
hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
}
assert(lo <= hi, "must have valid bounds");
const TypeInt* ti = TypeInt::make(lo, hi, MAX2(r1->_widen,r2->_widen));
#ifdef ASSERT
// Make sure we get the sign-capture idiom correct.
if (shift == BitsPerJavaInteger-1) {
if (r1->_lo >= 0) assert(ti == TypeInt::ZERO, ">>>31 of + is 0");
if (r1->_hi < 0) assert(ti == TypeInt::ONE, ">>>31 of - is +1");
}
#endif
return ti;
}
//
// Do not support shifted oops in info for GC
//
// else if( t1->base() == Type::InstPtr ) {
//
// const TypeInstPtr *o = t1->is_instptr();
// if( t1->singleton() )
// return TypeInt::make( ((uint32)o->const_oop() + o->_offset) >> shift );
// }
// else if( t1->base() == Type::KlassPtr ) {
// const TypeKlassPtr *o = t1->is_klassptr();
// if( t1->singleton() )
// return TypeInt::make( ((uint32)o->const_oop() + o->_offset) >> shift );
// }
return TypeInt::INT;
}
//=============================================================================
//------------------------------Identity---------------------------------------
Node *URShiftLNode::Identity( PhaseTransform *phase ) {
const TypeInt *ti = phase->type( in(2) )->isa_int(); // shift count is an int
return ( ti && ti->is_con() && ( ti->get_con() & ( BitsPerLong - 1 ) ) == 0 ) ? in(1) : this;
}
//------------------------------Ideal------------------------------------------
Node *URShiftLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
const TypeInt *t2 = phase->type( in(2) )->isa_int();
if( !t2 || !t2->is_con() ) return NULL; // Right input is a constant
const int con = t2->get_con() & ( BitsPerLong - 1 ); // Shift count is always masked
if ( con == 0 ) return NULL; // let Identity() handle a 0 shift count
// note: mask computation below does not work for 0 shift count
// We'll be wanting the right-shift amount as a mask of that many bits
const jlong mask = jlong(max_julong >> con);
// Check for ((x << z) + Y) >>> z. Replace with x + con>>>z
// The idiom for rounding to a power of 2 is "(Q+(2^z-1)) >>> z".
// If Q is "X << z" the rounding is useless. Look for patterns like
// ((X<<Z) + Y) >>> Z and replace with (X + Y>>>Z) & Z-mask.
Node *add = in(1);
if( add->Opcode() == Op_AddL ) {
Node *lshl = add->in(1);
if( lshl->Opcode() == Op_LShiftL &&
phase->type(lshl->in(2)) == t2 ) {
Node *y_z = phase->transform( new (phase->C) URShiftLNode(add->in(2),in(2)) );
Node *sum = phase->transform( new (phase->C) AddLNode( lshl->in(1), y_z ) );
return new (phase->C) AndLNode( sum, phase->longcon(mask) );
}
}
// Check for (x & mask) >>> z. Replace with (x >>> z) & (mask >>> z)
// This shortens the mask. Also, if we are extracting a high byte and
// storing it to a buffer, the mask will be removed completely.
Node *andi = in(1);
if( andi->Opcode() == Op_AndL ) {
const TypeLong *t3 = phase->type( andi->in(2) )->isa_long();
if( t3 && t3->is_con() ) { // Right input is a constant
jlong mask2 = t3->get_con();
mask2 >>= con; // *signed* shift downward (high-order zeroes do not help)
Node *newshr = phase->transform( new (phase->C) URShiftLNode(andi->in(1), in(2)) );
return new (phase->C) AndLNode(newshr, phase->longcon(mask2));
}
}
// Check for "(X << z ) >>> z" which simply zero-extends
Node *shl = in(1);
if( shl->Opcode() == Op_LShiftL &&
phase->type(shl->in(2)) == t2 )
return new (phase->C) AndLNode( shl->in(1), phase->longcon(mask) );
return NULL;
}
//------------------------------Value------------------------------------------
// A URShiftINode shifts its input2 right by input1 amount.
const Type *URShiftLNode::Value( PhaseTransform *phase ) const {
// (This is a near clone of RShiftLNode::Value.)
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
// Either input is TOP ==> the result is TOP
if( t1 == Type::TOP ) return Type::TOP;
if( t2 == Type::TOP ) return Type::TOP;
// Left input is ZERO ==> the result is ZERO.
if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
// Shift by zero does nothing
if( t2 == TypeInt::ZERO ) return t1;
// Either input is BOTTOM ==> the result is BOTTOM
if (t1 == Type::BOTTOM || t2 == Type::BOTTOM)
return TypeLong::LONG;
if (t2 == TypeInt::INT)
return TypeLong::LONG;
const TypeLong *r1 = t1->is_long(); // Handy access
const TypeInt *r2 = t2->is_int (); // Handy access
if (r2->is_con()) {
uint shift = r2->get_con();
shift &= BitsPerJavaLong - 1; // semantics of Java shifts
// Shift by a multiple of 64 does nothing:
if (shift == 0) return t1;
// Calculate reasonably aggressive bounds for the result.
jlong lo = (julong)r1->_lo >> (juint)shift;
jlong hi = (julong)r1->_hi >> (juint)shift;
if (r1->_hi >= 0 && r1->_lo < 0) {
// If the type has both negative and positive values,
// there are two separate sub-domains to worry about:
// The positive half and the negative half.
jlong neg_lo = lo;
jlong neg_hi = (julong)-1 >> (juint)shift;
jlong pos_lo = (julong) 0 >> (juint)shift;
jlong pos_hi = hi;
//lo = MIN2(neg_lo, pos_lo); // == 0
lo = neg_lo < pos_lo ? neg_lo : pos_lo;
//hi = MAX2(neg_hi, pos_hi); // == -1 >>> shift;
hi = neg_hi > pos_hi ? neg_hi : pos_hi;
}
assert(lo <= hi, "must have valid bounds");
const TypeLong* tl = TypeLong::make(lo, hi, MAX2(r1->_widen,r2->_widen));
#ifdef ASSERT
// Make sure we get the sign-capture idiom correct.
if (shift == BitsPerJavaLong - 1) {
if (r1->_lo >= 0) assert(tl == TypeLong::ZERO, ">>>63 of + is 0");
if (r1->_hi < 0) assert(tl == TypeLong::ONE, ">>>63 of - is +1");
}
#endif
return tl;
}
return TypeLong::LONG; // Give up
}
C:\hotspot-69087d08d473\src\share\vm/opto/mulnode.hpp
/*
* Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#ifndef SHARE_VM_OPTO_MULNODE_HPP
#define SHARE_VM_OPTO_MULNODE_HPP
#include "opto/node.hpp"
#include "opto/opcodes.hpp"
#include "opto/type.hpp"
// Portions of code courtesy of Clifford Click
class PhaseTransform;
//------------------------------MulNode----------------------------------------
// Classic MULTIPLY functionality. This covers all the usual 'multiply'
// behaviors for an algebraic ring. Multiply-integer, multiply-float,
// multiply-double, and binary-and are all inherited from this class. The
// various identity values are supplied by virtual functions.
class MulNode : public Node {
virtual uint hash() const;
public:
MulNode( Node *in1, Node *in2 ): Node(0,in1,in2) {
init_class_id(Class_Mul);
}
// Handle algebraic identities here. If we have an identity, return the Node
// we are equivalent to. We look for "add of zero" as an identity.
virtual Node *Identity( PhaseTransform *phase );
// We also canonicalize the Node, moving constants to the right input,
// and flatten expressions (so that 1+x+2 becomes x+3).
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
// Compute a new Type for this node. Basically we just do the pre-check,
// then call the virtual add() to set the type.
virtual const Type *Value( PhaseTransform *phase ) const;
// Supplied function returns the product of the inputs.
// This also type-checks the inputs for sanity. Guaranteed never to
// be passed a TOP or BOTTOM type, these are filtered out by a pre-check.
// This call recognizes the multiplicative zero type.
virtual const Type *mul_ring( const Type *, const Type * ) const = 0;
// Supplied function to return the multiplicative identity type
virtual const Type *mul_id() const = 0;
// Supplied function to return the additive identity type
virtual const Type *add_id() const = 0;
// Supplied function to return the additive opcode
virtual int add_opcode() const = 0;
// Supplied function to return the multiplicative opcode
virtual int mul_opcode() const = 0;
};
//------------------------------MulINode---------------------------------------
// Multiply 2 integers
class MulINode : public MulNode {
public:
MulINode( Node *in1, Node *in2 ) : MulNode(in1,in2) {}
virtual int Opcode() const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *mul_ring( const Type *, const Type * ) const;
const Type *mul_id() const { return TypeInt::ONE; }
const Type *add_id() const { return TypeInt::ZERO; }
int add_opcode() const { return Op_AddI; }
int mul_opcode() const { return Op_MulI; }
const Type *bottom_type() const { return TypeInt::INT; }
virtual uint ideal_reg() const { return Op_RegI; }
};
//------------------------------MulLNode---------------------------------------
// Multiply 2 longs
class MulLNode : public MulNode {
public:
MulLNode( Node *in1, Node *in2 ) : MulNode(in1,in2) {}
virtual int Opcode() const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *mul_ring( const Type *, const Type * ) const;
const Type *mul_id() const { return TypeLong::ONE; }
const Type *add_id() const { return TypeLong::ZERO; }
int add_opcode() const { return Op_AddL; }
int mul_opcode() const { return Op_MulL; }
const Type *bottom_type() const { return TypeLong::LONG; }
virtual uint ideal_reg() const { return Op_RegL; }
};
//------------------------------MulFNode---------------------------------------
// Multiply 2 floats
class MulFNode : public MulNode {
public:
MulFNode( Node *in1, Node *in2 ) : MulNode(in1,in2) {}
virtual int Opcode() const;
virtual const Type *mul_ring( const Type *, const Type * ) const;
const Type *mul_id() const { return TypeF::ONE; }
const Type *add_id() const { return TypeF::ZERO; }
int add_opcode() const { return Op_AddF; }
int mul_opcode() const { return Op_MulF; }
const Type *bottom_type() const { return Type::FLOAT; }
virtual uint ideal_reg() const { return Op_RegF; }
};
//------------------------------MulDNode---------------------------------------
// Multiply 2 doubles
class MulDNode : public MulNode {
public:
MulDNode( Node *in1, Node *in2 ) : MulNode(in1,in2) {}
virtual int Opcode() const;
virtual const Type *mul_ring( const Type *, const Type * ) const;
const Type *mul_id() const { return TypeD::ONE; }
const Type *add_id() const { return TypeD::ZERO; }
int add_opcode() const { return Op_AddD; }
int mul_opcode() const { return Op_MulD; }
const Type *bottom_type() const { return Type::DOUBLE; }
virtual uint ideal_reg() const { return Op_RegD; }
};
//-------------------------------MulHiLNode------------------------------------
// Upper 64 bits of a 64 bit by 64 bit multiply
class MulHiLNode : public Node {
public:
MulHiLNode( Node *in1, Node *in2 ) : Node(0,in1,in2) {}
virtual int Opcode() const;
virtual const Type *Value( PhaseTransform *phase ) const;
const Type *bottom_type() const { return TypeLong::LONG; }
virtual uint ideal_reg() const { return Op_RegL; }
};
//------------------------------AndINode---------------------------------------
// Logically AND 2 integers. Included with the MUL nodes because it inherits
// all the behavior of multiplication on a ring.
class AndINode : public MulINode {
public:
AndINode( Node *in1, Node *in2 ) : MulINode(in1,in2) {}
virtual int Opcode() const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual Node *Identity( PhaseTransform *phase );
virtual const Type *mul_ring( const Type *, const Type * ) const;
const Type *mul_id() const { return TypeInt::MINUS_1; }
const Type *add_id() const { return TypeInt::ZERO; }
int add_opcode() const { return Op_OrI; }
int mul_opcode() const { return Op_AndI; }
virtual uint ideal_reg() const { return Op_RegI; }
};
//------------------------------AndINode---------------------------------------
// Logically AND 2 longs. Included with the MUL nodes because it inherits
// all the behavior of multiplication on a ring.
class AndLNode : public MulLNode {
public:
AndLNode( Node *in1, Node *in2 ) : MulLNode(in1,in2) {}
virtual int Opcode() const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual Node *Identity( PhaseTransform *phase );
virtual const Type *mul_ring( const Type *, const Type * ) const;
const Type *mul_id() const { return TypeLong::MINUS_1; }
const Type *add_id() const { return TypeLong::ZERO; }
int add_opcode() const { return Op_OrL; }
int mul_opcode() const { return Op_AndL; }
virtual uint ideal_reg() const { return Op_RegL; }
};
//------------------------------LShiftINode------------------------------------
// Logical shift left
class LShiftINode : public Node {
public:
LShiftINode( Node *in1, Node *in2 ) : Node(0,in1,in2) {}
virtual int Opcode() const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value( PhaseTransform *phase ) const;
const Type *bottom_type() const { return TypeInt::INT; }
virtual uint ideal_reg() const { return Op_RegI; }
};
//------------------------------LShiftLNode------------------------------------
// Logical shift left
class LShiftLNode : public Node {
public:
LShiftLNode( Node *in1, Node *in2 ) : Node(0,in1,in2) {}
virtual int Opcode() const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value( PhaseTransform *phase ) const;
const Type *bottom_type() const { return TypeLong::LONG; }
virtual uint ideal_reg() const { return Op_RegL; }
};
//------------------------------RShiftINode------------------------------------
// Signed shift right
class RShiftINode : public Node {
public:
RShiftINode( Node *in1, Node *in2 ) : Node(0,in1,in2) {}
virtual int Opcode() const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value( PhaseTransform *phase ) const;
const Type *bottom_type() const { return TypeInt::INT; }
virtual uint ideal_reg() const { return Op_RegI; }
};
//------------------------------RShiftLNode------------------------------------
// Signed shift right
class RShiftLNode : public Node {
public:
RShiftLNode( Node *in1, Node *in2 ) : Node(0,in1,in2) {}
virtual int Opcode() const;
virtual Node *Identity( PhaseTransform *phase );
virtual const Type *Value( PhaseTransform *phase ) const;
const Type *bottom_type() const { return TypeLong::LONG; }
virtual uint ideal_reg() const { return Op_RegL; }
};
//------------------------------URShiftINode-----------------------------------
// Logical shift right
class URShiftINode : public Node {
public:
URShiftINode( Node *in1, Node *in2 ) : Node(0,in1,in2) {}
virtual int Opcode() const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value( PhaseTransform *phase ) const;
const Type *bottom_type() const { return TypeInt::INT; }
virtual uint ideal_reg() const { return Op_RegI; }
};
//------------------------------URShiftLNode-----------------------------------
// Logical shift right
class URShiftLNode : public Node {
public:
URShiftLNode( Node *in1, Node *in2 ) : Node(0,in1,in2) {}
virtual int Opcode() const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value( PhaseTransform *phase ) const;
const Type *bottom_type() const { return TypeLong::LONG; }
virtual uint ideal_reg() const { return Op_RegL; }
};
#endif // SHARE_VM_OPTO_MULNODE_HPP
C:\hotspot-69087d08d473\src\share\vm/opto/multnode.cpp
/*
* Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#include "precompiled.hpp"
#include "opto/callnode.hpp"
#include "opto/cfgnode.hpp"
#include "opto/matcher.hpp"
#include "opto/mathexactnode.hpp"
#include "opto/multnode.hpp"
#include "opto/opcodes.hpp"
#include "opto/phaseX.hpp"
#include "opto/regmask.hpp"
#include "opto/type.hpp"
//=============================================================================
//------------------------------MultiNode--------------------------------------
const RegMask &MultiNode::out_RegMask() const {
return RegMask::Empty;
}
Node *MultiNode::match( const ProjNode *proj, const Matcher *m ) { return proj->clone(); }
//------------------------------proj_out---------------------------------------
// Get a named projection
ProjNode* MultiNode::proj_out(uint which_proj) const {
assert(Opcode() != Op_If || which_proj == (uint)true || which_proj == (uint)false, "must be 1 or 0");
assert(Opcode() != Op_If || outcnt() == 2, "bad if #1");
for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
Node *p = fast_out(i);
if (p->is_Proj()) {
ProjNode *proj = p->as_Proj();
if (proj->_con == which_proj) {
assert(Opcode() != Op_If || proj->Opcode() == (which_proj?Op_IfTrue:Op_IfFalse), "bad if #2");
return proj;
}
} else {
assert(p == this && this->is_Start(), "else must be proj");
continue;
}
}
return NULL;
}
//=============================================================================
//------------------------------ProjNode---------------------------------------
uint ProjNode::hash() const {
// only one input
return (uintptr_t)in(TypeFunc::Control) + (_con << 1) + (_is_io_use ? 1 : 0);
}
uint ProjNode::cmp( const Node &n ) const { return _con == ((ProjNode&)n)._con && ((ProjNode&)n)._is_io_use == _is_io_use; }
uint ProjNode::size_of() const { return sizeof(ProjNode); }
// Test if we propagate interesting control along this projection
bool ProjNode::is_CFG() const {
Node *def = in(0);
return (_con == TypeFunc::Control && def->is_CFG());
}
const Type* ProjNode::proj_type(const Type* t) const {
if (t == Type::TOP) {
return Type::TOP;
}
if (t == Type::BOTTOM) {
return Type::BOTTOM;
}
t = t->is_tuple()->field_at(_con);
Node* n = in(0);
if ((_con == TypeFunc::Parms) &&
n->is_CallStaticJava() && n->as_CallStaticJava()->is_boxing_method()) {
// The result of autoboxing is always non-null on normal path.
t = t->join_speculative(TypePtr::NOTNULL);
}
return t;
}
const Type *ProjNode::bottom_type() const {
if (in(0) == NULL) return Type::TOP;
return proj_type(in(0)->bottom_type());
}
const TypePtr *ProjNode::adr_type() const {
if (bottom_type() == Type::MEMORY) {
// in(0) might be a narrow MemBar; otherwise we will report TypePtr::BOTTOM
const TypePtr* adr_type = in(0)->adr_type();
#ifdef ASSERT
if (!is_error_reported() && !Node::in_dump())
assert(adr_type != NULL, "source must have adr_type");
#endif
return adr_type;
}
assert(bottom_type()->base() != Type::Memory, "no other memories?");
return NULL;
}
bool ProjNode::pinned() const { return in(0)->pinned(); }
#ifndef PRODUCT
void ProjNode::dump_spec(outputStream *st) const { st->print("#%d",_con); if(_is_io_use) st->print(" (i_o_use)");}
#endif
//----------------------------check_con----------------------------------------
void ProjNode::check_con() const {
Node* n = in(0);
if (n == NULL) return; // should be assert, but NodeHash makes bogons
if (n->is_Mach()) return; // mach. projs. are not type-safe
if (n->is_Start()) return; // alas, starts can have mach. projs. also
if (_con == SCMemProjNode::SCMEMPROJCON ) return;
const Type* t = n->bottom_type();
if (t == Type::TOP) return; // multi is dead
assert(_con < t->is_tuple()->cnt(), "ProjNode::_con must be in range");
}
//------------------------------Value------------------------------------------
const Type *ProjNode::Value( PhaseTransform *phase ) const {
if (in(0) == NULL) return Type::TOP;
return proj_type(phase->type(in(0)));
}
//------------------------------out_RegMask------------------------------------
// Pass the buck uphill
const RegMask &ProjNode::out_RegMask() const {
return RegMask::Empty;
}
//------------------------------ideal_reg--------------------------------------
uint ProjNode::ideal_reg() const {
return bottom_type()->ideal_reg();
}
//-------------------------------is_uncommon_trap_proj----------------------------
// Return true if proj is the form of "proj->[region->..]call_uct"
bool ProjNode::is_uncommon_trap_proj(Deoptimization::DeoptReason reason) {
int path_limit = 10;
Node* out = this;
for (int ct = 0; ct < path_limit; ct++) {
out = out->unique_ctrl_out();
if (out == NULL)
return false;
if (out->is_CallStaticJava()) {
int req = out->as_CallStaticJava()->uncommon_trap_request();
if (req != 0) {
Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
if (trap_reason == reason || reason == Deoptimization::Reason_none) {
return true;
}
}
return false; // don't do further after call
}
if (out->Opcode() != Op_Region)
return false;
}
return false;
}
//-------------------------------is_uncommon_trap_if_pattern-------------------------
// Return true for "if(test)-> proj -> ...
// |
// V
// other_proj->[region->..]call_uct"
//
// "must_reason_predicate" means the uct reason must be Reason_predicate
bool ProjNode::is_uncommon_trap_if_pattern(Deoptimization::DeoptReason reason) {
Node *in0 = in(0);
if (!in0->is_If()) return false;
// Variation of a dead If node.
if (in0->outcnt() < 2) return false;
IfNode* iff = in0->as_If();
// we need "If(Conv2B(Opaque1(...)))" pattern for reason_predicate
if (reason != Deoptimization::Reason_none) {
if (iff->in(1)->Opcode() != Op_Conv2B ||
iff->in(1)->in(1)->Opcode() != Op_Opaque1) {
return false;
}
}
ProjNode* other_proj = iff->proj_out(1-_con);
if (other_proj == NULL) // Should never happen, but make Parfait happy.
return false;
if (other_proj->is_uncommon_trap_proj(reason)) {
assert(reason == Deoptimization::Reason_none ||
Compile::current()->is_predicate_opaq(iff->in(1)->in(1)), "should be on the list");
return true;
}
return false;
}
C:\hotspot-69087d08d473\src\share\vm/opto/multnode.hpp
/*
* Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#ifndef SHARE_VM_OPTO_MULTNODE_HPP
#define SHARE_VM_OPTO_MULTNODE_HPP
#include "opto/node.hpp"
class Matcher;
class ProjNode;
//------------------------------MultiNode--------------------------------------
// This class defines a MultiNode, a Node which produces many values. The
// values are wrapped up in a tuple Type, i.e. a TypeTuple.
class MultiNode : public Node {
public:
MultiNode( uint required ) : Node(required) {
init_class_id(Class_Multi);
}
virtual int Opcode() const;
virtual const Type *bottom_type() const = 0;
virtual bool is_CFG() const { return true; }
virtual uint hash() const { return NO_HASH; } // CFG nodes do not hash
virtual bool depends_only_on_test() const { return false; }
virtual const RegMask &out_RegMask() const;
virtual Node *match( const ProjNode *proj, const Matcher *m );
virtual uint ideal_reg() const { return NotAMachineReg; }
ProjNode* proj_out(uint which_proj) const; // Get a named projection
};
//------------------------------ProjNode---------------------------------------
// This class defines a Projection node. Projections project a single element
// out of a tuple (or Signature) type. Only MultiNodes produce TypeTuple
// results.
class ProjNode : public Node {
protected:
virtual uint hash() const;
virtual uint cmp( const Node &n ) const;
virtual uint size_of() const;
void check_con() const; // Called from constructor.
const Type* proj_type(const Type* t) const;
public:
ProjNode( Node *src, uint con, bool io_use = false )
: Node( src ), _con(con), _is_io_use(io_use)
{
init_class_id(Class_Proj);
// Optimistic setting. Need additional checks in Node::is_dead_loop_safe().
if (con != TypeFunc::Memory || src->is_Start())
init_flags(Flag_is_dead_loop_safe);
debug_only(check_con());
}
const uint _con; // The field in the tuple we are projecting
const bool _is_io_use; // Used to distinguish between the projections
// used on the control and io paths from a macro node
virtual int Opcode() const;
virtual bool is_CFG() const;
virtual bool depends_only_on_test() const { return false; }
virtual const Type *bottom_type() const;
virtual const TypePtr *adr_type() const;
virtual bool pinned() const;
virtual const Type *Value( PhaseTransform *phase ) const;
virtual uint ideal_reg() const;
virtual const RegMask &out_RegMask() const;
#ifndef PRODUCT
virtual void dump_spec(outputStream *st) const;
#endif
// Return true if proj is for "proj->[region->..]call_uct"
bool is_uncommon_trap_proj(Deoptimization::DeoptReason reason);
// Return true for "if(test)-> proj -> ...
// |
// V
// other_proj->[region->..]call_uct"
bool is_uncommon_trap_if_pattern(Deoptimization::DeoptReason reason);
};
#endif // SHARE_VM_OPTO_MULTNODE_HPP
C:\hotspot-69087d08d473\src\share\vm/opto/node.cpp
/*
* Copyright (c) 1997, 2016, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#include "precompiled.hpp"
#include "libadt/vectset.hpp"
#include "memory/allocation.inline.hpp"
#include "opto/cfgnode.hpp"
#include "opto/connode.hpp"
#include "opto/loopnode.hpp"
#include "opto/machnode.hpp"
#include "opto/matcher.hpp"
#include "opto/node.hpp"
#include "opto/opcodes.hpp"
#include "opto/regmask.hpp"
#include "opto/type.hpp"
#include "utilities/copy.hpp"
class RegMask;
// #include "phase.hpp"
class PhaseTransform;
class PhaseGVN;
// Arena we are currently building Nodes in
const uint Node::NotAMachineReg = 0xffff0000;
#ifndef PRODUCT
extern int nodes_created;
#endif
#ifdef ASSERT
//-------------------------- construct_node------------------------------------
// Set a breakpoint here to identify where a particular node index is built.
void Node::verify_construction() {
_debug_orig = NULL;
int old_debug_idx = Compile::debug_idx();
int new_debug_idx = old_debug_idx+1;
if (new_debug_idx > 0) {
// Arrange that the lowest five decimal digits of _debug_idx
// will repeat those of _idx. In case this is somehow pathological,
// we continue to assign negative numbers (!) consecutively.
const int mod = 100000;
int bump = (int)(_idx - new_debug_idx) % mod;
if (bump < 0) bump += mod;
assert(bump >= 0 && bump < mod, "");
new_debug_idx += bump;
}
Compile::set_debug_idx(new_debug_idx);
set_debug_idx( new_debug_idx );
assert(Compile::current()->unique() < (INT_MAX - 1), "Node limit exceeded INT_MAX");
assert(Compile::current()->live_nodes() < Compile::current()->max_node_limit(), "Live Node limit exceeded limit");
if (BreakAtNode != 0 && (_debug_idx == BreakAtNode || (int)_idx == BreakAtNode)) {
tty->print_cr("BreakAtNode: _idx=%d _debug_idx=%d", _idx, _debug_idx);
BREAKPOINT;
}
#if OPTO_DU_ITERATOR_ASSERT
_last_del = NULL;
_del_tick = 0;
#endif
_hash_lock = 0;
}
// #ifdef ASSERT ...
#if OPTO_DU_ITERATOR_ASSERT
void DUIterator_Common::sample(const Node* node) {
_vdui = VerifyDUIterators;
_node = node;
_outcnt = node->_outcnt;
_del_tick = node->_del_tick;
_last = NULL;
}
void DUIterator_Common::verify(const Node* node, bool at_end_ok) {
assert(_node == node, "consistent iterator source");
assert(_del_tick == node->_del_tick, "no unexpected deletions allowed");
}
void DUIterator_Common::verify_resync() {
// Ensure that the loop body has just deleted the last guy produced.
const Node* node = _node;
// Ensure that at least one copy of the last-seen edge was deleted.
// Note: It is OK to delete multiple copies of the last-seen edge.
// Unfortunately, we have no way to verify that all the deletions delete
// that same edge. On this point we must use the Honor System.
assert(node->_del_tick >= _del_tick+1, "must have deleted an edge");
assert(node->_last_del == _last, "must have deleted the edge just produced");
// We liked this deletion, so accept the resulting outcnt and tick.
_outcnt = node->_outcnt;
_del_tick = node->_del_tick;
}
void DUIterator_Common::reset(const DUIterator_Common& that) {
if (this == &that) return; // ignore assignment to self
if (!_vdui) {
// We need to initialize everything, overwriting garbage values.
_last = that._last;
_vdui = that._vdui;
}
// Note: It is legal (though odd) for an iterator over some node x
// to be reassigned to iterate over another node y. Some doubly-nested
// progress loops depend on being able to do this.
const Node* node = that._node;
// Re-initialize everything, except _last.
_node = node;
_outcnt = node->_outcnt;
_del_tick = node->_del_tick;
}
void DUIterator::sample(const Node* node) {
DUIterator_Common::sample(node); // Initialize the assertion data.
_refresh_tick = 0; // No refreshes have happened, as yet.
}
void DUIterator::verify(const Node* node, bool at_end_ok) {
DUIterator_Common::verify(node, at_end_ok);
assert(_idx < node->_outcnt + (uint)at_end_ok, "idx in range");
}
void DUIterator::verify_increment() {
if (_refresh_tick & 1) {
// We have refreshed the index during this loop.
// Fix up _idx to meet asserts.
if (_idx > _outcnt) _idx = _outcnt;
}
verify(_node, true);
}
void DUIterator::verify_resync() {
// Note: We do not assert on _outcnt, because insertions are OK here.
DUIterator_Common::verify_resync();
// Make sure we are still in sync, possibly with no more out-edges:
verify(_node, true);
}
void DUIterator::reset(const DUIterator& that) {
if (this == &that) return; // self assignment is always a no-op
assert(that._refresh_tick == 0, "assign only the result of Node::outs()");
assert(that._idx == 0, "assign only the result of Node::outs()");
assert(_idx == that._idx, "already assigned _idx");
if (!_vdui) {
// We need to initialize everything, overwriting garbage values.
sample(that._node);
} else {
DUIterator_Common::reset(that);
if (_refresh_tick & 1) {
_refresh_tick++; // Clear the "was refreshed" flag.
}
assert(_refresh_tick < 2*100000, "DU iteration must converge quickly");
}
}
void DUIterator::refresh() {
DUIterator_Common::sample(_node); // Re-fetch assertion data.
_refresh_tick |= 1; // Set the "was refreshed" flag.
}
void DUIterator::verify_finish() {
// If the loop has killed the node, do not require it to re-run.
if (_node->_outcnt == 0) _refresh_tick &= ~1;
// If this assert triggers, it means that a loop used refresh_out_pos
// to re-synch an iteration index, but the loop did not correctly
// re-run itself, using a "while (progress)" construct.
// This iterator enforces the rule that you must keep trying the loop
// until it "runs clean" without any need for refreshing.
assert(!(_refresh_tick & 1), "the loop must run once with no refreshing");
}
void DUIterator_Fast::verify(const Node* node, bool at_end_ok) {
DUIterator_Common::verify(node, at_end_ok);
Node** out = node->_out;
uint cnt = node->_outcnt;
assert(cnt == _outcnt, "no insertions allowed");
assert(_outp >= out && _outp <= out + cnt - !at_end_ok, "outp in range");
// This last check is carefully designed to work for NO_OUT_ARRAY.
}
void DUIterator_Fast::verify_limit() {
const Node* node = _node;
verify(node, true);
assert(_outp == node->_out + node->_outcnt, "limit still correct");
}
void DUIterator_Fast::verify_resync() {
const Node* node = _node;
if (_outp == node->_out + _outcnt) {
// Note that the limit imax, not the pointer i, gets updated with the
// exact count of deletions. (For the pointer it's always "--i".)
assert(node->_outcnt+node->_del_tick == _outcnt+_del_tick, "no insertions allowed with deletion(s)");
// This is a limit pointer, with a name like "imax".
// Fudge the _last field so that the common assert will be happy.
_last = (Node*) node->_last_del;
DUIterator_Common::verify_resync();
} else {
assert(node->_outcnt < _outcnt, "no insertions allowed with deletion(s)");
// A normal internal pointer.
DUIterator_Common::verify_resync();
// Make sure we are still in sync, possibly with no more out-edges:
verify(node, true);
}
}
void DUIterator_Fast::verify_relimit(uint n) {
const Node* node = _node;
assert((int)n > 0, "use imax -= n only with a positive count");
// This must be a limit pointer, with a name like "imax".
assert(_outp == node->_out + node->_outcnt, "apply -= only to a limit (imax)");
// The reported number of deletions must match what the node saw.
assert(node->_del_tick == _del_tick + n, "must have deleted n edges");
// Fudge the _last field so that the common assert will be happy.
_last = (Node*) node->_last_del;
DUIterator_Common::verify_resync();
}
void DUIterator_Fast::reset(const DUIterator_Fast& that) {
assert(_outp == that._outp, "already assigned _outp");
DUIterator_Common::reset(that);
}
void DUIterator_Last::verify(const Node* node, bool at_end_ok) {
// at_end_ok means the _outp is allowed to underflow by 1
_outp += at_end_ok;
DUIterator_Fast::verify(node, at_end_ok); // check _del_tick, etc.
_outp -= at_end_ok;
assert(_outp == (node->_out + node->_outcnt) - 1, "pointer must point to end of nodes");
}
void DUIterator_Last::verify_limit() {
// Do not require the limit address to be resynched.
//verify(node, true);
assert(_outp == _node->_out, "limit still correct");
}
void DUIterator_Last::verify_step(uint num_edges) {
assert((int)num_edges > 0, "need non-zero edge count for loop progress");
_outcnt -= num_edges;
_del_tick += num_edges;
// Make sure we are still in sync, possibly with no more out-edges:
const Node* node = _node;
verify(node, true);
assert(node->_last_del == _last, "must have deleted the edge just produced");
}
#endif //OPTO_DU_ITERATOR_ASSERT
#endif //ASSERT
// This constant used to initialize _out may be any non-null value.
// The value NULL is reserved for the top node only.
#define NO_OUT_ARRAY ((Node**)-1)
// This funny expression handshakes with Node::operator new
// to pull Compile::current out of the new node's _out field,
// and then calls a subroutine which manages most field
// initializations. The only one which is tricky is the
// _idx field, which is const, and so must be initialized
// by a return value, not an assignment.
//
// (Aren't you thankful that Java finals don't require so many tricks?)
#define IDX_INIT(req) this->Init((req), (Compile*) this->_out)
#ifdef _MSC_VER // the IDX_INIT hack falls foul of warning C4355
#pragma warning( disable:4355 ) // 'this' : used in base member initializer list
#endif
#ifdef __clang__
#pragma clang diagnostic push
#pragma GCC diagnostic ignored "-Wuninitialized"
#endif
// Out-of-line code from node constructors.
// Executed only when extra debug info. is being passed around.
static void init_node_notes(Compile* C, int idx, Node_Notes* nn) {
C->set_node_notes_at(idx, nn);
}
// Shared initialization code.
inline int Node::Init(int req, Compile* C) {
assert(Compile::current() == C, "must use operator new(Compile*)");
int idx = C->next_unique();
// Allocate memory for the necessary number of edges.
if (req > 0) {
// Allocate space for _in array to have double alignment.
_in = (Node **) ((char *) (C->node_arena()->Amalloc_D(req * sizeof(void*))));
#ifdef ASSERT
_in[req-1] = this; // magic cookie for assertion check
#endif
}
// If there are default notes floating around, capture them:
Node_Notes* nn = C->default_node_notes();
if (nn != NULL) init_node_notes(C, idx, nn);
// Note: At this point, C is dead,
// and we begin to initialize the new Node.
_cnt = _max = req;
_outcnt = _outmax = 0;
_class_id = Class_Node;
_flags = 0;
_out = NO_OUT_ARRAY;
return idx;
}
//------------------------------Node-------------------------------------------
// Create a Node, with a given number of required edges.
Node::Node(uint req)
: _idx(IDX_INIT(req))
#ifdef ASSERT
, _parse_idx(_idx)
#endif
{
assert( req < Compile::current()->max_node_limit() - NodeLimitFudgeFactor, "Input limit exceeded" );
debug_only( verify_construction() );
NOT_PRODUCT(nodes_created++);
if (req == 0) {
assert( _in == (Node**)this, "Must not pass arg count to 'new'" );
_in = NULL;
} else {
assert( _in[req-1] == this, "Must pass arg count to 'new'" );
Node** to = _in;
for(uint i = 0; i < req; i++) {
to[i] = NULL;
}
}
}
//------------------------------Node-------------------------------------------
Node::Node(Node *n0)
: _idx(IDX_INIT(1))
#ifdef ASSERT
, _parse_idx(_idx)
#endif
{
debug_only( verify_construction() );
NOT_PRODUCT(nodes_created++);
// Assert we allocated space for input array already
assert( _in[0] == this, "Must pass arg count to 'new'" );
assert( is_not_dead(n0), "can not use dead node");
_in[0] = n0; if (n0 != NULL) n0->add_out((Node *)this);
}
//------------------------------Node-------------------------------------------
Node::Node(Node *n0, Node *n1)
: _idx(IDX_INIT(2))
#ifdef ASSERT
, _parse_idx(_idx)
#endif
{
debug_only( verify_construction() );
NOT_PRODUCT(nodes_created++);
// Assert we allocated space for input array already
assert( _in[1] == this, "Must pass arg count to 'new'" );
assert( is_not_dead(n0), "can not use dead node");
assert( is_not_dead(n1), "can not use dead node");
_in[0] = n0; if (n0 != NULL) n0->add_out((Node *)this);
_in[1] = n1; if (n1 != NULL) n1->add_out((Node *)this);
}
//------------------------------Node-------------------------------------------
Node::Node(Node *n0, Node *n1, Node *n2)
: _idx(IDX_INIT(3))
#ifdef ASSERT
, _parse_idx(_idx)
#endif
{
debug_only( verify_construction() );
NOT_PRODUCT(nodes_created++);
// Assert we allocated space for input array already
assert( _in[2] == this, "Must pass arg count to 'new'" );
assert( is_not_dead(n0), "can not use dead node");
assert( is_not_dead(n1), "can not use dead node");
assert( is_not_dead(n2), "can not use dead node");
_in[0] = n0; if (n0 != NULL) n0->add_out((Node *)this);
_in[1] = n1; if (n1 != NULL) n1->add_out((Node *)this);
_in[2] = n2; if (n2 != NULL) n2->add_out((Node *)this);
}
//------------------------------Node-------------------------------------------
Node::Node(Node *n0, Node *n1, Node *n2, Node *n3)
: _idx(IDX_INIT(4))
#ifdef ASSERT
, _parse_idx(_idx)
#endif
{
debug_only( verify_construction() );
NOT_PRODUCT(nodes_created++);
// Assert we allocated space for input array already
assert( _in[3] == this, "Must pass arg count to 'new'" );
assert( is_not_dead(n0), "can not use dead node");
assert( is_not_dead(n1), "can not use dead node");
assert( is_not_dead(n2), "can not use dead node");
assert( is_not_dead(n3), "can not use dead node");
_in[0] = n0; if (n0 != NULL) n0->add_out((Node *)this);
_in[1] = n1; if (n1 != NULL) n1->add_out((Node *)this);
_in[2] = n2; if (n2 != NULL) n2->add_out((Node *)this);
_in[3] = n3; if (n3 != NULL) n3->add_out((Node *)this);
}
//------------------------------Node-------------------------------------------
Node::Node(Node *n0, Node *n1, Node *n2, Node *n3, Node *n4)
: _idx(IDX_INIT(5))
#ifdef ASSERT
, _parse_idx(_idx)
#endif
{
debug_only( verify_construction() );
NOT_PRODUCT(nodes_created++);
// Assert we allocated space for input array already
assert( _in[4] == this, "Must pass arg count to 'new'" );
assert( is_not_dead(n0), "can not use dead node");
assert( is_not_dead(n1), "can not use dead node");
assert( is_not_dead(n2), "can not use dead node");
assert( is_not_dead(n3), "can not use dead node");
assert( is_not_dead(n4), "can not use dead node");
_in[0] = n0; if (n0 != NULL) n0->add_out((Node *)this);
_in[1] = n1; if (n1 != NULL) n1->add_out((Node *)this);
_in[2] = n2; if (n2 != NULL) n2->add_out((Node *)this);
_in[3] = n3; if (n3 != NULL) n3->add_out((Node *)this);
_in[4] = n4; if (n4 != NULL) n4->add_out((Node *)this);
}
//------------------------------Node-------------------------------------------
Node::Node(Node *n0, Node *n1, Node *n2, Node *n3,
Node *n4, Node *n5)
: _idx(IDX_INIT(6))
#ifdef ASSERT
, _parse_idx(_idx)
#endif
{
debug_only( verify_construction() );
NOT_PRODUCT(nodes_created++);
// Assert we allocated space for input array already
assert( _in[5] == this, "Must pass arg count to 'new'" );
assert( is_not_dead(n0), "can not use dead node");
assert( is_not_dead(n1), "can not use dead node");
assert( is_not_dead(n2), "can not use dead node");
assert( is_not_dead(n3), "can not use dead node");
assert( is_not_dead(n4), "can not use dead node");
assert( is_not_dead(n5), "can not use dead node");
_in[0] = n0; if (n0 != NULL) n0->add_out((Node *)this);
_in[1] = n1; if (n1 != NULL) n1->add_out((Node *)this);
_in[2] = n2; if (n2 != NULL) n2->add_out((Node *)this);
_in[3] = n3; if (n3 != NULL) n3->add_out((Node *)this);
_in[4] = n4; if (n4 != NULL) n4->add_out((Node *)this);
_in[5] = n5; if (n5 != NULL) n5->add_out((Node *)this);
}
//------------------------------Node-------------------------------------------
Node::Node(Node *n0, Node *n1, Node *n2, Node *n3,
Node *n4, Node *n5, Node *n6)
: _idx(IDX_INIT(7))
#ifdef ASSERT
, _parse_idx(_idx)
#endif
{
debug_only( verify_construction() );
NOT_PRODUCT(nodes_created++);
// Assert we allocated space for input array already
assert( _in[6] == this, "Must pass arg count to 'new'" );
assert( is_not_dead(n0), "can not use dead node");
assert( is_not_dead(n1), "can not use dead node");
assert( is_not_dead(n2), "can not use dead node");
assert( is_not_dead(n3), "can not use dead node");
assert( is_not_dead(n4), "can not use dead node");
assert( is_not_dead(n5), "can not use dead node");
assert( is_not_dead(n6), "can not use dead node");
_in[0] = n0; if (n0 != NULL) n0->add_out((Node *)this);
_in[1] = n1; if (n1 != NULL) n1->add_out((Node *)this);
_in[2] = n2; if (n2 != NULL) n2->add_out((Node *)this);
_in[3] = n3; if (n3 != NULL) n3->add_out((Node *)this);
_in[4] = n4; if (n4 != NULL) n4->add_out((Node *)this);
_in[5] = n5; if (n5 != NULL) n5->add_out((Node *)this);
_in[6] = n6; if (n6 != NULL) n6->add_out((Node *)this);
}
#ifdef __clang__
#pragma clang diagnostic pop
#endif
//------------------------------clone------------------------------------------
// Clone a Node.
Node *Node::clone() const {
Compile* C = Compile::current();
uint s = size_of(); // Size of inherited Node
Node *n = (Node*)C->node_arena()->Amalloc_D(size_of() + _max*sizeof(Node*));
Copy::conjoint_words_to_lower((HeapWord*)this, (HeapWord*)n, s);
// Set the new input pointer array
n->_in = (Node**)(((char*)n)+s);
// Cannot share the old output pointer array, so kill it
n->_out = NO_OUT_ARRAY;
// And reset the counters to 0
n->_outcnt = 0;
n->_outmax = 0;
// Unlock this guy, since he is not in any hash table.
debug_only(n->_hash_lock = 0);
// Walk the old node's input list to duplicate its edges
uint i;
for( i = 0; i < len(); i++ ) {
Node *x = in(i);
n->_in[i] = x;
if (x != NULL) x->add_out(n);
}
if (is_macro())
C->add_macro_node(n);
if (is_expensive())
C->add_expensive_node(n);
// If the cloned node is a range check dependent CastII, add it to the list.
CastIINode* cast = n->isa_CastII();
if (cast != NULL && cast->has_range_check()) {
C->add_range_check_cast(cast);
}
n->set_idx(C->next_unique()); // Get new unique index as well
debug_only( n->verify_construction() );
NOT_PRODUCT(nodes_created++);
// Do not patch over the debug_idx of a clone, because it makes it
// impossible to break on the clone's moment of creation.
//debug_only( n->set_debug_idx( debug_idx() ) );
C->copy_node_notes_to(n, (Node*) this);
// MachNode clone
uint nopnds;
if (this->is_Mach() && (nopnds = this->as_Mach()->num_opnds()) > 0) {
MachNode *mach = n->as_Mach();
MachNode *mthis = this->as_Mach();
// Get address of _opnd_array.
// It should be the same offset since it is the clone of this node.
MachOper **from = mthis->_opnds;
MachOper **to = (MachOper **)((size_t)(&mach->_opnds) +
pointer_delta((const void*)from,
(const void*)(&mthis->_opnds), 1));
mach->_opnds = to;
for ( uint i = 0; i < nopnds; ++i ) {
to[i] = from[i]->clone(C);
}
}
// cloning CallNode may need to clone JVMState
if (n->is_Call()) {
n->as_Call()->clone_jvms(C);
}
if (n->is_SafePoint()) {
n->as_SafePoint()->clone_replaced_nodes();
}
return n; // Return the clone
}
//---------------------------setup_is_top--------------------------------------
// Call this when changing the top node, to reassert the invariants
// required by Node::is_top. See Compile::set_cached_top_node.
void Node::setup_is_top() {
if (this == (Node*)Compile::current()->top()) {
// This node has just become top. Kill its out array.
_outcnt = _outmax = 0;
_out = NULL; // marker value for top
assert(is_top(), "must be top");
} else {
if (_out == NULL) _out = NO_OUT_ARRAY;
assert(!is_top(), "must not be top");
}
}
//------------------------------~Node------------------------------------------
// Fancy destructor; eagerly attempt to reclaim Node numberings and storage
extern int reclaim_idx ;
extern int reclaim_in ;
extern int reclaim_node;
void Node::destruct() {
// Eagerly reclaim unique Node numberings
Compile* compile = Compile::current();
if ((uint)_idx+1 == compile->unique()) {
compile->set_unique(compile->unique()-1);
#ifdef ASSERT
reclaim_idx++;
#endif
}
// Clear debug info:
Node_Notes* nn = compile->node_notes_at(_idx);
if (nn != NULL) nn->clear();
// Walk the input array, freeing the corresponding output edges
_cnt = _max; // forget req/prec distinction
uint i;
for( i = 0; i < _max; i++ ) {
set_req(i, NULL);
//assert(def->out(def->outcnt()-1) == (Node *)this,"bad def-use hacking in reclaim");
}
assert(outcnt() == 0, "deleting a node must not leave a dangling use");
// See if the input array was allocated just prior to the object
int edge_size = _max*sizeof(void*);
int out_edge_size = _outmax*sizeof(void*);
char *edge_end = ((char*)_in) + edge_size;
char *out_array = (char*)(_out == NO_OUT_ARRAY? NULL: _out);
char *out_edge_end = out_array + out_edge_size;
int node_size = size_of();
// Free the output edge array
if (out_edge_size > 0) {
#ifdef ASSERT
if( out_edge_end == compile->node_arena()->hwm() )
reclaim_in += out_edge_size; // count reclaimed out edges with in edges
#endif
compile->node_arena()->Afree(out_array, out_edge_size);
}
// Free the input edge array and the node itself
if( edge_end == (char*)this ) {
#ifdef ASSERT
if( edge_end+node_size == compile->node_arena()->hwm() ) {
reclaim_in += edge_size;
reclaim_node+= node_size;
}
#else
// It was; free the input array and object all in one hit
compile->node_arena()->Afree(_in,edge_size+node_size);
#endif
} else {
// Free just the input array
#ifdef ASSERT
if( edge_end == compile->node_arena()->hwm() )
reclaim_in += edge_size;
#endif
compile->node_arena()->Afree(_in,edge_size);
// Free just the object
#ifdef ASSERT
if( ((char*)this) + node_size == compile->node_arena()->hwm() )
reclaim_node+= node_size;
#else
compile->node_arena()->Afree(this,node_size);
#endif
}
if (is_macro()) {
compile->remove_macro_node(this);
}
if (is_expensive()) {
compile->remove_expensive_node(this);
}
CastIINode* cast = isa_CastII();
if (cast != NULL && cast->has_range_check()) {
compile->remove_range_check_cast(cast);
}
if (is_SafePoint()) {
as_SafePoint()->delete_replaced_nodes();
}
#ifdef ASSERT
// We will not actually delete the storage, but we'll make the node unusable.
*(address*)this = badAddress; // smash the C++ vtbl, probably
_in = _out = (Node**) badAddress;
_max = _cnt = _outmax = _outcnt = 0;
#endif
}
//------------------------------grow-------------------------------------------
// Grow the input array, making space for more edges
void Node::grow( uint len ) {
Arena* arena = Compile::current()->node_arena();
uint new_max = _max;
if( new_max == 0 ) {
_max = 4;
_in = (Node**)arena->Amalloc(4*sizeof(Node*));
Node** to = _in;
to[0] = NULL;
to[1] = NULL;
to[2] = NULL;
to[3] = NULL;
return;
}
while( new_max <= len ) new_max <<= 1; // Find next power-of-2
// Trimming to limit allows a uint8 to handle up to 255 edges.
// Previously I was using only powers-of-2 which peaked at 128 edges.
//if( new_max >= limit ) new_max = limit-1;
_in = (Node**)arena->Arealloc(_in, _max*sizeof(Node*), new_max*sizeof(Node*));
Copy::zero_to_bytes(&_in[_max], (new_max-_max)*sizeof(Node*)); // NULL all new space
_max = new_max; // Record new max length
// This assertion makes sure that Node::_max is wide enough to
// represent the numerical value of new_max.
assert(_max == new_max && _max > len, "int width of _max is too small");
}
//-----------------------------out_grow----------------------------------------
// Grow the input array, making space for more edges
void Node::out_grow( uint len ) {
assert(!is_top(), "cannot grow a top node's out array");
Arena* arena = Compile::current()->node_arena();
uint new_max = _outmax;
if( new_max == 0 ) {
_outmax = 4;
_out = (Node **)arena->Amalloc(4*sizeof(Node*));
return;
}
while( new_max <= len ) new_max <<= 1; // Find next power-of-2
// Trimming to limit allows a uint8 to handle up to 255 edges.
// Previously I was using only powers-of-2 which peaked at 128 edges.
//if( new_max >= limit ) new_max = limit-1;
assert(_out != NULL && _out != NO_OUT_ARRAY, "out must have sensible value");
_out = (Node**)arena->Arealloc(_out,_outmax*sizeof(Node*),new_max*sizeof(Node*));
//Copy::zero_to_bytes(&_out[_outmax], (new_max-_outmax)*sizeof(Node*)); // NULL all new space
_outmax = new_max; // Record new max length
// This assertion makes sure that Node::_max is wide enough to
// represent the numerical value of new_max.
assert(_outmax == new_max && _outmax > len, "int width of _outmax is too small");
}
#ifdef ASSERT
//------------------------------is_dead----------------------------------------
bool Node::is_dead() const {
// Mach and pinch point nodes may look like dead.
if( is_top() || is_Mach() || (Opcode() == Op_Node && _outcnt > 0) )
return false;
for( uint i = 0; i < _max; i++ )
if( _in[i] != NULL )
return false;
dump();
return true;
}
#endif
//------------------------------is_unreachable---------------------------------
bool Node::is_unreachable(PhaseIterGVN &igvn) const {
assert(!is_Mach(), "doesn't work with MachNodes");
return outcnt() == 0 || igvn.type(this) == Type::TOP || in(0)->is_top();
}
//------------------------------add_req----------------------------------------
// Add a new required input at the end
void Node::add_req( Node *n ) {
assert( is_not_dead(n), "can not use dead node");
// Look to see if I can move precedence down one without reallocating
if( (_cnt >= _max) || (in(_max-1) != NULL) )
grow( _max+1 );
// Find a precedence edge to move
if( in(_cnt) != NULL ) { // Next precedence edge is busy?
uint i;
for( i=_cnt; i<_max; i++ )
if( in(i) == NULL ) // Find the NULL at end of prec edge list
break; // There must be one, since we grew the array
_in[i] = in(_cnt); // Move prec over, making space for req edge
}
_in[_cnt++] = n; // Stuff over old prec edge
if (n != NULL) n->add_out((Node *)this);
}
//---------------------------add_req_batch-------------------------------------
// Add a new required input at the end
void Node::add_req_batch( Node *n, uint m ) {
assert( is_not_dead(n), "can not use dead node");
// check various edge cases
if ((int)m <= 1) {
assert((int)m >= 0, "oob");
if (m != 0) add_req(n);
return;
}
// Look to see if I can move precedence down one without reallocating
if( (_cnt+m) > _max || _in[_max-m] )
grow( _max+m );
// Find a precedence edge to move
if( _in[_cnt] != NULL ) { // Next precedence edge is busy?
uint i;
for( i=_cnt; i<_max; i++ )
if( _in[i] == NULL ) // Find the NULL at end of prec edge list
break; // There must be one, since we grew the array
// Slide all the precs over by m positions (assume #prec << m).
Copy::conjoint_words_to_higher((HeapWord*)&_in[_cnt], (HeapWord*)&_in[_cnt+m], ((i-_cnt)*sizeof(Node*)));
}
// Stuff over the old prec edges
for(uint i=0; i<m; i++ ) {
_in[_cnt++] = n;
}
// Insert multiple out edges on the node.
if (n != NULL && !n->is_top()) {
for(uint i=0; i<m; i++ ) {
n->add_out((Node *)this);
}
}
}
//------------------------------del_req----------------------------------------
// Delete the required edge and compact the edge array
void Node::del_req( uint idx ) {
assert( idx < _cnt, "oob");
assert( !VerifyHashTableKeys || _hash_lock == 0,
"remove node from hash table before modifying it");
// First remove corresponding def-use edge
Node *n = in(idx);
if (n != NULL) n->del_out((Node *)this);
_in[idx] = in(--_cnt); // Compact the array
// Avoid spec violation: Gap in prec edges.
close_prec_gap_at(_cnt);
}
//------------------------------del_req_ordered--------------------------------
// Delete the required edge and compact the edge array with preserved order
void Node::del_req_ordered( uint idx ) {
assert( idx < _cnt, "oob");
assert( !VerifyHashTableKeys || _hash_lock == 0,
"remove node from hash table before modifying it");
// First remove corresponding def-use edge
Node *n = in(idx);
if (n != NULL) n->del_out((Node *)this);
if (idx < --_cnt) { // Not last edge ?
Copy::conjoint_words_to_lower((HeapWord*)&_in[idx+1], (HeapWord*)&_in[idx], ((_cnt-idx)*sizeof(Node*)));
}
// Avoid spec violation: Gap in prec edges.
close_prec_gap_at(_cnt);
}
//------------------------------ins_req----------------------------------------
// Insert a new required input at the end
void Node::ins_req( uint idx, Node *n ) {
assert( is_not_dead(n), "can not use dead node");
add_req(NULL); // Make space
assert( idx < _max, "Must have allocated enough space");
// Slide over
if(_cnt-idx-1 > 0) {
Copy::conjoint_words_to_higher((HeapWord*)&_in[idx], (HeapWord*)&_in[idx+1], ((_cnt-idx-1)*sizeof(Node*)));
}
_in[idx] = n; // Stuff over old required edge
if (n != NULL) n->add_out((Node *)this); // Add reciprocal def-use edge
}
//-----------------------------find_edge---------------------------------------
int Node::find_edge(Node* n) {
for (uint i = 0; i < len(); i++) {
if (_in[i] == n) return i;
}
return -1;
}
//----------------------------replace_edge-------------------------------------
int Node::replace_edge(Node* old, Node* neww) {
if (old == neww) return 0; // nothing to do
uint nrep = 0;
for (uint i = 0; i < len(); i++) {
if (in(i) == old) {
if (i < req()) {
set_req(i, neww);
} else {
assert(find_prec_edge(neww) == -1, err_msg("spec violation: duplicated prec edge (node %d -> %d)", _idx, neww->_idx));
set_prec(i, neww);
}
nrep++;
}
}
return nrep;
}
/**
* Replace input edges in the range pointing to 'old' node.
*/
int Node::replace_edges_in_range(Node* old, Node* neww, int start, int end) {
if (old == neww) return 0; // nothing to do
uint nrep = 0;
for (int i = start; i < end; i++) {
if (in(i) == old) {
set_req(i, neww);
nrep++;
}
}
return nrep;
}
//-------------------------disconnect_inputs-----------------------------------
// NULL out all inputs to eliminate incoming Def-Use edges.
// Return the number of edges between 'n' and 'this'
int Node::disconnect_inputs(Node *n, Compile* C) {
int edges_to_n = 0;
uint cnt = req();
for( uint i = 0; i < cnt; ++i ) {
if( in(i) == 0 ) continue;
if( in(i) == n ) ++edges_to_n;
set_req(i, NULL);
}
// Remove precedence edges if any exist
// Note: Safepoints may have precedence edges, even during parsing
if( (req() != len()) && (in(req()) != NULL) ) {
uint max = len();
for( uint i = 0; i < max; ++i ) {
if( in(i) == 0 ) continue;
if( in(i) == n ) ++edges_to_n;
set_prec(i, NULL);
}
}
// Node::destruct requires all out edges be deleted first
// debug_only(destruct();) // no reuse benefit expected
if (edges_to_n == 0) {
C->record_dead_node(_idx);
}
return edges_to_n;
}
//-----------------------------uncast---------------------------------------
// %%% Temporary, until we sort out CheckCastPP vs. CastPP.
// Strip away casting. (It is depth-limited.)
Node* Node::uncast() const {
// Should be inline:
//return is_ConstraintCast() ? uncast_helper(this) : (Node*) this;
if (is_ConstraintCast() || is_CheckCastPP())
return uncast_helper(this);
else
return (Node*) this;
}
// Find out of current node that matches opcode.
Node* Node::find_out_with(int opcode) {
for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
Node* use = fast_out(i);
if (use->Opcode() == opcode) {
return use;
}
}
return NULL;
}
//---------------------------uncast_helper-------------------------------------
Node* Node::uncast_helper(const Node* p) {
#ifdef ASSERT
uint depth_count = 0;
const Node* orig_p = p;
#endif
while (true) {
#ifdef ASSERT
if (depth_count >= K) {
orig_p->dump(4);
if (p != orig_p)
p->dump(1);
}
assert(depth_count++ < K, "infinite loop in Node::uncast_helper");
#endif
if (p == NULL || p->req() != 2) {
break;
} else if (p->is_ConstraintCast()) {
p = p->in(1);
} else if (p->is_CheckCastPP()) {
p = p->in(1);
} else {
break;
}
}
return (Node*) p;
}
//------------------------------add_prec---------------------------------------
// Add a new precedence input. Precedence inputs are unordered, with
// duplicates removed and NULLs packed down at the end.
void Node::add_prec( Node *n ) {
assert( is_not_dead(n), "can not use dead node");
// Check for NULL at end
if( _cnt >= _max || in(_max-1) )
grow( _max+1 );
// Find a precedence edge to move
uint i = _cnt;
while( in(i) != NULL ) {
if (in(i) == n) return; // Avoid spec violation: duplicated prec edge.
i++;
}
_in[i] = n; // Stuff prec edge over NULL
if ( n != NULL) n->add_out((Node *)this); // Add mirror edge
#ifdef ASSERT
while ((++i)<_max) { assert(_in[i] == NULL, err_msg("spec violation: Gap in prec edges (node %d)", _idx)); }
#endif
}
//------------------------------rm_prec----------------------------------------
// Remove a precedence input. Precedence inputs are unordered, with
// duplicates removed and NULLs packed down at the end.
void Node::rm_prec( uint j ) {
assert(j < _max, err_msg("oob: i=%d, _max=%d", j, _max));
assert(j >= _cnt, "not a precedence edge");
if (_in[j] == NULL) return; // Avoid spec violation: Gap in prec edges.
_in[j]->del_out((Node *)this);
close_prec_gap_at(j);
}
//------------------------------size_of----------------------------------------
uint Node::size_of() const { return sizeof(*this); }
//------------------------------ideal_reg--------------------------------------
uint Node::ideal_reg() const { return 0; }
//------------------------------jvms-------------------------------------------
JVMState* Node::jvms() const { return NULL; }
#ifdef ASSERT
//------------------------------jvms-------------------------------------------
bool Node::verify_jvms(const JVMState* using_jvms) const {
for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
if (jvms == using_jvms) return true;
}
return false;
}
//------------------------------init_NodeProperty------------------------------
void Node::init_NodeProperty() {
assert(_max_classes <= max_jushort, "too many NodeProperty classes");
assert(_max_flags <= max_jushort, "too many NodeProperty flags");
}
#endif
//------------------------------format-----------------------------------------
// Print as assembly
void Node::format( PhaseRegAlloc *, outputStream *st ) const {}
//------------------------------emit-------------------------------------------
// Emit bytes starting at parameter 'ptr'.
void Node::emit(CodeBuffer &cbuf, PhaseRegAlloc *ra_) const {}
//------------------------------size-------------------------------------------
// Size of instruction in bytes
uint Node::size(PhaseRegAlloc *ra_) const { return 0; }
//------------------------------CFG Construction-------------------------------
// Nodes that end basic blocks, e.g. IfTrue/IfFalse, JumpProjNode, Root,
// Goto and Return.
const Node *Node::is_block_proj() const { return 0; }
// Minimum guaranteed type
const Type *Node::bottom_type() const { return Type::BOTTOM; }
//------------------------------raise_bottom_type------------------------------
// Get the worst-case Type output for this Node.
void Node::raise_bottom_type(const Type* new_type) {
if (is_Type()) {
TypeNode *n = this->as_Type();
if (VerifyAliases) {
assert(new_type->higher_equal_speculative(n->type()), "new type must refine old type");
}
n->set_type(new_type);
} else if (is_Load()) {
LoadNode *n = this->as_Load();
if (VerifyAliases) {
assert(new_type->higher_equal_speculative(n->type()), "new type must refine old type");
}
n->set_type(new_type);
}
}
//------------------------------Identity---------------------------------------
// Return a node that the given node is equivalent to.
Node *Node::Identity( PhaseTransform * ) {
return this; // Default to no identities
}
//------------------------------Value------------------------------------------
// Compute a new Type for a node using the Type of the inputs.
const Type *Node::Value( PhaseTransform * ) const {
return bottom_type(); // Default to worst-case Type
}
//------------------------------Ideal------------------------------------------
//
// 'Idealize' the graph rooted at this Node.
//
// In order to be efficient and flexible there are some subtle invariants
// these Ideal calls need to hold. Running with '+VerifyIterativeGVN' checks
// these invariants, although its too slow to have on by default. If you are
// hacking an Ideal call, be sure to test with +VerifyIterativeGVN!
//
// The Ideal call almost arbitrarily reshape the graph rooted at the 'this'
// pointer. If ANY change is made, it must return the root of the reshaped
// graph - even if the root is the same Node. Example: swapping the inputs
// to an AddINode gives the same answer and same root, but you still have to
// return the 'this' pointer instead of NULL.
//
// You cannot return an OLD Node, except for the 'this' pointer. Use the
// Identity call to return an old Node; basically if Identity can find
// another Node have the Ideal call make no change and return NULL.
// Example: AddINode::Ideal must check for add of zero; in this case it
// returns NULL instead of doing any graph reshaping.
//
// You cannot modify any old Nodes except for the 'this' pointer. Due to
// sharing there may be other users of the old Nodes relying on their current
// semantics. Modifying them will break the other users.
// Example: when reshape "(X+3)+4" into "X+7" you must leave the Node for
// "X+3" unchanged in case it is shared.
//
// If you modify the 'this' pointer's inputs, you should use
// 'set_req'. If you are making a new Node (either as the new root or
// some new internal piece) you may use 'init_req' to set the initial
// value. You can make a new Node with either 'new' or 'clone'. In
// either case, def-use info is correctly maintained.
//
// Example: reshape "(X+3)+4" into "X+7":
// set_req(1, in(1)->in(1));
// set_req(2, phase->intcon(7));
// return this;
// Example: reshape "X*4" into "X<<2"
// return new (C) LShiftINode(in(1), phase->intcon(2));
//
// You must call 'phase->transform(X)' on any new Nodes X you make, except
// for the returned root node. Example: reshape "X*31" with "(X<<5)-X".
// Node *shift=phase->transform(new(C)LShiftINode(in(1),phase->intcon(5)));
// return new (C) AddINode(shift, in(1));
//
// When making a Node for a constant use 'phase->makecon' or 'phase->intcon'.
// These forms are faster than 'phase->transform(new (C) ConNode())' and Do
// The Right Thing with def-use info.
//
// You cannot bury the 'this' Node inside of a graph reshape. If the reshaped
// graph uses the 'this' Node it must be the root. If you want a Node with
// the same Opcode as the 'this' pointer use 'clone'.
//
Node *Node::Ideal(PhaseGVN *phase, bool can_reshape) {
return NULL; // Default to being Ideal already
}
// Some nodes have specific Ideal subgraph transformations only if they are
// unique users of specific nodes. Such nodes should be put on IGVN worklist
// for the transformations to happen.
bool Node::has_special_unique_user() const {
assert(outcnt() == 1, "match only for unique out");
Node* n = unique_out();
int op = Opcode();
if( this->is_Store() ) {
// Condition for back-to-back stores folding.
return n->Opcode() == op && n->in(MemNode::Memory) == this;
} else if (this->is_Load() || this->is_DecodeN()) {
// Condition for removing an unused LoadNode or DecodeNNode from the MemBarAcquire precedence input
return n->Opcode() == Op_MemBarAcquire;
} else if( op == Op_AddL ) {
// Condition for convL2I(addL(x,y)) ==> addI(convL2I(x),convL2I(y))
return n->Opcode() == Op_ConvL2I && n->in(1) == this;
} else if( op == Op_SubI || op == Op_SubL ) {
// Condition for subI(x,subI(y,z)) ==> subI(addI(x,z),y)
return n->Opcode() == op && n->in(2) == this;
}
return false;
};
//--------------------------find_exact_control---------------------------------
// Skip Proj and CatchProj nodes chains. Check for Null and Top.
Node* Node::find_exact_control(Node* ctrl) {
if (ctrl == NULL && this->is_Region())
ctrl = this->as_Region()->is_copy();
if (ctrl != NULL && ctrl->is_CatchProj()) {
if (ctrl->as_CatchProj()->_con == CatchProjNode::fall_through_index)
ctrl = ctrl->in(0);
if (ctrl != NULL && !ctrl->is_top())
ctrl = ctrl->in(0);
}
if (ctrl != NULL && ctrl->is_Proj())
ctrl = ctrl->in(0);
return ctrl;
}
//--------------------------dominates------------------------------------------
// Helper function for MemNode::all_controls_dominate().
// Check if 'this' control node dominates or equal to 'sub' control node.
// We already know that if any path back to Root or Start reaches 'this',
// then all paths so, so this is a simple search for one example,
// not an exhaustive search for a counterexample.
bool Node::dominates(Node* sub, Node_List &nlist) {
assert(this->is_CFG(), "expecting control");
assert(sub != NULL && sub->is_CFG(), "expecting control");
// detect dead cycle without regions
int iterations_without_region_limit = DominatorSearchLimit;
Node* orig_sub = sub;
Node* dom = this;
bool met_dom = false;
nlist.clear();
// Walk 'sub' backward up the chain to 'dom', watching for regions.
// After seeing 'dom', continue up to Root or Start.
// If we hit a region (backward split point), it may be a loop head.
// Keep going through one of the region's inputs. If we reach the
// same region again, go through a different input. Eventually we
// will either exit through the loop head, or give up.
// (If we get confused, break out and return a conservative 'false'.)
while (sub != NULL) {
if (sub->is_top()) break; // Conservative answer for dead code.
if (sub == dom) {
if (nlist.size() == 0) {
// No Region nodes except loops were visited before and the EntryControl
// path was taken for loops: it did not walk in a cycle.
return true;
} else if (met_dom) {
break; // already met before: walk in a cycle
} else {
// Region nodes were visited. Continue walk up to Start or Root
// to make sure that it did not walk in a cycle.
met_dom = true; // first time meet
iterations_without_region_limit = DominatorSearchLimit; // Reset
}
}
if (sub->is_Start() || sub->is_Root()) {
// Success if we met 'dom' along a path to Start or Root.
// We assume there are no alternative paths that avoid 'dom'.
// (This assumption is up to the caller to ensure!)
return met_dom;
}
Node* up = sub->in(0);
// Normalize simple pass-through regions and projections:
up = sub->find_exact_control(up);
// If sub == up, we found a self-loop. Try to push past it.
if (sub == up && sub->is_Loop()) {
// Take loop entry path on the way up to 'dom'.
up = sub->in(1); // in(LoopNode::EntryControl);
} else if (sub == up && sub->is_Region() && sub->req() != 3) {
// Always take in(1) path on the way up to 'dom' for clone regions
// (with only one input) or regions which merge > 2 paths
// (usually used to merge fast/slow paths).
up = sub->in(1);
} else if (sub == up && sub->is_Region()) {
// Try both paths for Regions with 2 input paths (it may be a loop head).
// It could give conservative 'false' answer without information
// which region's input is the entry path.
iterations_without_region_limit = DominatorSearchLimit; // Reset
bool region_was_visited_before = false;
// Was this Region node visited before?
// If so, we have reached it because we accidentally took a
// loop-back edge from 'sub' back into the body of the loop,
// and worked our way up again to the loop header 'sub'.
// So, take the first unexplored path on the way up to 'dom'.
for (int j = nlist.size() - 1; j >= 0; j--) {
intptr_t ni = (intptr_t)nlist.at(j);
Node* visited = (Node*)(ni & ~1);
bool visited_twice_already = ((ni & 1) != 0);
if (visited == sub) {
if (visited_twice_already) {
// Visited 2 paths, but still stuck in loop body. Give up.
return false;
}
// The Region node was visited before only once.
// (We will repush with the low bit set, below.)
nlist.remove(j);
// We will find a new edge and re-insert.
region_was_visited_before = true;
break;
}
}
// Find an incoming edge which has not been seen yet; walk through it.
assert(up == sub, "");
uint skip = region_was_visited_before ? 1 : 0;
for (uint i = 1; i < sub->req(); i++) {
Node* in = sub->in(i);
if (in != NULL && !in->is_top() && in != sub) {
if (skip == 0) {
up = in;
break;
}
--skip; // skip this nontrivial input
}
}
// Set 0 bit to indicate that both paths were taken.
nlist.push((Node*)((intptr_t)sub + (region_was_visited_before ? 1 : 0)));
}
if (up == sub) {
break; // some kind of tight cycle
}
if (up == orig_sub && met_dom) {
// returned back after visiting 'dom'
break; // some kind of cycle
}
if (--iterations_without_region_limit < 0) {
break; // dead cycle
}
sub = up;
}
// Did not meet Root or Start node in pred. chain.
// Conservative answer for dead code.
return false;
}
//------------------------------remove_dead_region-----------------------------
// This control node is dead. Follow the subgraph below it making everything
// using it dead as well. This will happen normally via the usual IterGVN
// worklist but this call is more efficient. Do not update use-def info
// inside the dead region, just at the borders.
static void kill_dead_code( Node *dead, PhaseIterGVN *igvn ) {
// Con's are a popular node to re-hit in the hash table again.
if( dead->is_Con() ) return;
// Can't put ResourceMark here since igvn->_worklist uses the same arena
// for verify pass with +VerifyOpto and we add/remove elements in it here.
Node_List nstack(Thread::current()->resource_area());
Node *top = igvn->C->top();
nstack.push(dead);
bool has_irreducible_loop = igvn->C->has_irreducible_loop();
while (nstack.size() > 0) {
dead = nstack.pop();
if (dead->Opcode() == Op_SafePoint) {
dead->as_SafePoint()->disconnect_from_root(igvn);
}
if (dead->outcnt() > 0) {
// Keep dead node on stack until all uses are processed.
nstack.push(dead);
// For all Users of the Dead... ;-)
for (DUIterator_Last kmin, k = dead->last_outs(kmin); k >= kmin; ) {
Node* use = dead->last_out(k);
igvn->hash_delete(use); // Yank from hash table prior to mod
if (use->in(0) == dead) { // Found another dead node
assert (!use->is_Con(), "Control for Con node should be Root node.");
use->set_req(0, top); // Cut dead edge to prevent processing
nstack.push(use); // the dead node again.
} else if (!has_irreducible_loop && // Backedge could be alive in irreducible loop
use->is_Loop() && !use->is_Root() && // Don't kill Root (RootNode extends LoopNode)
use->in(LoopNode::EntryControl) == dead) { // Dead loop if its entry is dead
use->set_req(LoopNode::EntryControl, top); // Cut dead edge to prevent processing
use->set_req(0, top); // Cut self edge
nstack.push(use);
} else { // Else found a not-dead user
// Dead if all inputs are top or null
bool dead_use = !use->is_Root(); // Keep empty graph alive
for (uint j = 1; j < use->req(); j++) {
Node* in = use->in(j);
if (in == dead) { // Turn all dead inputs into TOP
use->set_req(j, top);
} else if (in != NULL && !in->is_top()) {
dead_use = false;
}
}
if (dead_use) {
if (use->is_Region()) {
use->set_req(0, top); // Cut self edge
}
nstack.push(use);
} else {
igvn->_worklist.push(use);
}
}
// Refresh the iterator, since any number of kills might have happened.
k = dead->last_outs(kmin);
}
} else { // (dead->outcnt() == 0)
// Done with outputs.
igvn->hash_delete(dead);
igvn->_worklist.remove(dead);
igvn->set_type(dead, Type::TOP);
if (dead->is_macro()) {
igvn->C->remove_macro_node(dead);
}
if (dead->is_expensive()) {
igvn->C->remove_expensive_node(dead);
}
CastIINode* cast = dead->isa_CastII();
if (cast != NULL && cast->has_range_check()) {
igvn->C->remove_range_check_cast(cast);
}
igvn->C->record_dead_node(dead->_idx);
// Kill all inputs to the dead guy
for (uint i=0; i < dead->req(); i++) {
Node *n = dead->in(i); // Get input to dead guy
if (n != NULL && !n->is_top()) { // Input is valid?
dead->set_req(i, top); // Smash input away
if (n->outcnt() == 0) { // Input also goes dead?
if (!n->is_Con())
nstack.push(n); // Clear it out as well
} else if (n->outcnt() == 1 &&
n->has_special_unique_user()) {
igvn->add_users_to_worklist( n );
} else if (n->outcnt() <= 2 && n->is_Store()) {
// Push store's uses on worklist to enable folding optimization for
// store/store and store/load to the same address.
// The restriction (outcnt() <= 2) is the same as in set_req_X()
// and remove_globally_dead_node().
igvn->add_users_to_worklist( n );
}
}
}
} // (dead->outcnt() == 0)
} // while (nstack.size() > 0) for outputs
return;
}
//------------------------------remove_dead_region-----------------------------
bool Node::remove_dead_region(PhaseGVN *phase, bool can_reshape) {
Node *n = in(0);
if( !n ) return false;
// Lost control into this guy? I.e., it became unreachable?
// Aggressively kill all unreachable code.
if (can_reshape && n->is_top()) {
kill_dead_code(this, phase->is_IterGVN());
return false; // Node is dead.
}
if( n->is_Region() && n->as_Region()->is_copy() ) {
Node *m = n->nonnull_req();
set_req(0, m);
return true;
}
return false;
}
//------------------------------Ideal_DU_postCCP-------------------------------
// Idealize graph, using DU info. Must clone result into new-space
Node *Node::Ideal_DU_postCCP( PhaseCCP * ) {
return NULL; // Default to no change
}
//------------------------------hash-------------------------------------------
// Hash function over Nodes.
uint Node::hash() const {
uint sum = 0;
for( uint i=0; i<_cnt; i++ ) // Add in all inputs
sum = (sum<<1)-(uintptr_t)in(i); // Ignore embedded NULLs
return (sum>>2) + _cnt + Opcode();
}
//------------------------------cmp--------------------------------------------
// Compare special parts of simple Nodes
uint Node::cmp( const Node &n ) const {
return 1; // Must be same
}
//------------------------------rematerialize-----------------------------------
// Should we clone rather than spill this instruction?
bool Node::rematerialize() const {
if ( is_Mach() )
return this->as_Mach()->rematerialize();
else
return (_flags & Flag_rematerialize) != 0;
}
//------------------------------needs_anti_dependence_check---------------------
// Nodes which use memory without consuming it, hence need antidependences.
bool Node::needs_anti_dependence_check() const {
if( req() < 2 || (_flags & Flag_needs_anti_dependence_check) == 0 )
return false;
else
return in(1)->bottom_type()->has_memory();
}
// Get an integer constant from a ConNode (or CastIINode).
// Return a default value if there is no apparent constant here.
const TypeInt* Node::find_int_type() const {
if (this->is_Type()) {
return this->as_Type()->type()->isa_int();
} else if (this->is_Con()) {
assert(is_Mach(), "should be ConNode(TypeNode) or else a MachNode");
return this->bottom_type()->isa_int();
}
return NULL;
}
// Get a pointer constant from a ConstNode.
// Returns the constant if it is a pointer ConstNode
intptr_t Node::get_ptr() const {
assert( Opcode() == Op_ConP, "" );
return ((ConPNode*)this)->type()->is_ptr()->get_con();
}
// Get a narrow oop constant from a ConNNode.
intptr_t Node::get_narrowcon() const {
assert( Opcode() == Op_ConN, "" );
return ((ConNNode*)this)->type()->is_narrowoop()->get_con();
}
// Get a long constant from a ConNode.
// Return a default value if there is no apparent constant here.
const TypeLong* Node::find_long_type() const {
if (this->is_Type()) {
return this->as_Type()->type()->isa_long();
} else if (this->is_Con()) {
assert(is_Mach(), "should be ConNode(TypeNode) or else a MachNode");
return this->bottom_type()->isa_long();
}
return NULL;
}
/**
* Return a ptr type for nodes which should have it.
*/
const TypePtr* Node::get_ptr_type() const {
const TypePtr* tp = this->bottom_type()->make_ptr();
#ifdef ASSERT
if (tp == NULL) {
this->dump(1);
assert((tp != NULL), "unexpected node type");
}
#endif
return tp;
}
// Get a double constant from a ConstNode.
// Returns the constant if it is a double ConstNode
jdouble Node::getd() const {
assert( Opcode() == Op_ConD, "" );
return ((ConDNode*)this)->type()->is_double_constant()->getd();
}
// Get a float constant from a ConstNode.
// Returns the constant if it is a float ConstNode
jfloat Node::getf() const {
assert( Opcode() == Op_ConF, "" );
return ((ConFNode*)this)->type()->is_float_constant()->getf();
}
#ifndef PRODUCT
//----------------------------NotANode----------------------------------------
// Used in debugging code to avoid walking across dead or uninitialized edges.
static inline bool NotANode(const Node* n) {
if (n == NULL) return true;
if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
if (*(address*)n == badAddress) return true; // kill by Node::destruct
return false;
}
//------------------------------find------------------------------------------
// Find a neighbor of this Node with the given _idx
// If idx is negative, find its absolute value, following both _in and _out.
static void find_recur(Compile* C, Node* &result, Node *n, int idx, bool only_ctrl,
VectorSet* old_space, VectorSet* new_space ) {
int node_idx = (idx >= 0) ? idx : -idx;
if (NotANode(n)) return; // Gracefully handle NULL, -1, 0xabababab, etc.
// Contained in new_space or old_space? Check old_arena first since it's mostly empty.
VectorSet *v = C->old_arena()->contains(n) ? old_space : new_space;
if( v->test(n->_idx) ) return;
if( (int)n->_idx == node_idx
debug_only(|| n->debug_idx() == node_idx) ) {
if (result != NULL)
tty->print("find: " INTPTR_FORMAT " and " INTPTR_FORMAT " both have idx==%d\n",
(uintptr_t)result, (uintptr_t)n, node_idx);
result = n;
}
v->set(n->_idx);
for( uint i=0; i<n->len(); i++ ) {
if( only_ctrl && !(n->is_Region()) && (n->Opcode() != Op_Root) && (i != TypeFunc::Control) ) continue;
find_recur(C, result, n->in(i), idx, only_ctrl, old_space, new_space );
}
// Search along forward edges also:
if (idx < 0 && !only_ctrl) {
for( uint j=0; j<n->outcnt(); j++ ) {
find_recur(C, result, n->raw_out(j), idx, only_ctrl, old_space, new_space );
}
}
#ifdef ASSERT
// Search along debug_orig edges last, checking for cycles
Node* orig = n->debug_orig();
if (orig != NULL) {
do {
if (NotANode(orig)) break;
find_recur(C, result, orig, idx, only_ctrl, old_space, new_space );
orig = orig->debug_orig();
} while (orig != NULL && orig != n->debug_orig());
}
#endif //ASSERT
}
// call this from debugger:
Node* find_node(Node* n, int idx) {
return n->find(idx);
}
//------------------------------find-------------------------------------------
Node* Node::find(int idx) const {
ResourceArea *area = Thread::current()->resource_area();
VectorSet old_space(area), new_space(area);
Node* result = NULL;
find_recur(Compile::current(), result, (Node*) this, idx, false, &old_space, &new_space );
return result;
}
//------------------------------find_ctrl--------------------------------------
// Find an ancestor to this node in the control history with given _idx
Node* Node::find_ctrl(int idx) const {
ResourceArea *area = Thread::current()->resource_area();
VectorSet old_space(area), new_space(area);
Node* result = NULL;
find_recur(Compile::current(), result, (Node*) this, idx, true, &old_space, &new_space );
return result;
}
#endif
#ifndef PRODUCT
// -----------------------------Name-------------------------------------------
extern const char *NodeClassNames[];
const char *Node::Name() const { return NodeClassNames[Opcode()]; }
static bool is_disconnected(const Node* n) {
for (uint i = 0; i < n->req(); i++) {
if (n->in(i) != NULL) return false;
}
return true;
}
#ifdef ASSERT
static void dump_orig(Node* orig, outputStream *st) {
Compile* C = Compile::current();
if (NotANode(orig)) orig = NULL;
if (orig != NULL && !C->node_arena()->contains(orig)) orig = NULL;
if (orig == NULL) return;
st->print(" !orig=");
Node* fast = orig->debug_orig(); // tortoise & hare algorithm to detect loops
if (NotANode(fast)) fast = NULL;
while (orig != NULL) {
bool discon = is_disconnected(orig); // if discon, print [123] else 123
if (discon) st->print("[");
if (!Compile::current()->node_arena()->contains(orig))
st->print("o");
st->print("%d", orig->_idx);
if (discon) st->print("]");
orig = orig->debug_orig();
if (NotANode(orig)) orig = NULL;
if (orig != NULL && !C->node_arena()->contains(orig)) orig = NULL;
if (orig != NULL) st->print(",");
if (fast != NULL) {
// Step fast twice for each single step of orig:
fast = fast->debug_orig();
if (NotANode(fast)) fast = NULL;
if (fast != NULL && fast != orig) {
fast = fast->debug_orig();
if (NotANode(fast)) fast = NULL;
}
if (fast == orig) {
st->print("...");
break;
}
}
}
}
void Node::set_debug_orig(Node* orig) {
_debug_orig = orig;
if (BreakAtNode == 0) return;
if (NotANode(orig)) orig = NULL;
int trip = 10;
while (orig != NULL) {
if (orig->debug_idx() == BreakAtNode || (int)orig->_idx == BreakAtNode) {
tty->print_cr("BreakAtNode: _idx=%d _debug_idx=%d orig._idx=%d orig._debug_idx=%d",
this->_idx, this->debug_idx(), orig->_idx, orig->debug_idx());
BREAKPOINT;
}
orig = orig->debug_orig();
if (NotANode(orig)) orig = NULL;
if (trip-- <= 0) break;
}
}
#endif //ASSERT
//------------------------------dump------------------------------------------
// Dump a Node
void Node::dump(const char* suffix, outputStream *st) const {
Compile* C = Compile::current();
bool is_new = C->node_arena()->contains(this);
C->_in_dump_cnt++;
st->print("%c%d\t%s\t=== ", is_new ? ' ' : 'o', _idx, Name());
// Dump the required and precedence inputs
dump_req(st);
dump_prec(st);
// Dump the outputs
dump_out(st);
if (is_disconnected(this)) {
#ifdef ASSERT
st->print(" [%d]",debug_idx());
dump_orig(debug_orig(), st);
#endif
st->cr();
C->_in_dump_cnt--;
return; // don't process dead nodes
}
// Dump node-specific info
dump_spec(st);
#ifdef ASSERT
// Dump the non-reset _debug_idx
if (Verbose && WizardMode) {
st->print(" [%d]",debug_idx());
}
#endif
const Type *t = bottom_type();
if (t != NULL && (t->isa_instptr() || t->isa_klassptr())) {
const TypeInstPtr *toop = t->isa_instptr();
const TypeKlassPtr *tkls = t->isa_klassptr();
ciKlass* klass = toop ? toop->klass() : (tkls ? tkls->klass() : NULL );
if (klass && klass->is_loaded() && klass->is_interface()) {
st->print(" Interface:");
} else if (toop) {
st->print(" Oop:");
} else if (tkls) {
st->print(" Klass:");
}
t->dump_on(st);
} else if (t == Type::MEMORY) {
st->print(" Memory:");
MemNode::dump_adr_type(this, adr_type(), st);
} else if (Verbose || WizardMode) {
st->print(" Type:");
if (t) {
t->dump_on(st);
} else {
st->print("no type");
}
} else if (t->isa_vect() && this->is_MachSpillCopy()) {
// Dump MachSpillcopy vector type.
t->dump_on(st);
}
if (is_new) {
debug_only(dump_orig(debug_orig(), st));
Node_Notes* nn = C->node_notes_at(_idx);
if (nn != NULL && !nn->is_clear()) {
if (nn->jvms() != NULL) {
st->print(" !jvms:");
nn->jvms()->dump_spec(st);
}
}
}
if (suffix) st->print("%s", suffix);
C->_in_dump_cnt--;
}
//------------------------------dump_req--------------------------------------
void Node::dump_req(outputStream *st) const {
// Dump the required input edges
for (uint i = 0; i < req(); i++) { // For all required inputs
Node* d = in(i);
if (d == NULL) {
st->print("_ ");
} else if (NotANode(d)) {
st->print("NotANode "); // uninitialized, sentinel, garbage, etc.
} else {
st->print("%c%d ", Compile::current()->node_arena()->contains(d) ? ' ' : 'o', d->_idx);
}
}
}
//------------------------------dump_prec-------------------------------------
void Node::dump_prec(outputStream *st) const {
// Dump the precedence edges
int any_prec = 0;
for (uint i = req(); i < len(); i++) { // For all precedence inputs
Node* p = in(i);
if (p != NULL) {
if (!any_prec++) st->print(" |");
if (NotANode(p)) { st->print("NotANode "); continue; }
st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
}
}
}
//------------------------------dump_out--------------------------------------
void Node::dump_out(outputStream *st) const {
// Delimit the output edges
st->print(" [[");
// Dump the output edges
for (uint i = 0; i < _outcnt; i++) { // For all outputs
Node* u = _out[i];
if (u == NULL) {
st->print("_ ");
} else if (NotANode(u)) {
st->print("NotANode ");
} else {
st->print("%c%d ", Compile::current()->node_arena()->contains(u) ? ' ' : 'o', u->_idx);
}
}
st->print("]] ");
}
//------------------------------dump_nodes-------------------------------------
static void dump_nodes(const Node* start, int d, bool only_ctrl) {
Node* s = (Node*)start; // remove const
if (NotANode(s)) return;
uint depth = (uint)ABS(d);
int direction = d;
Compile* C = Compile::current();
GrowableArray <Node *> nstack(C->live_nodes());
nstack.append(s);
int begin = 0;
int end = 0;
for(uint i = 0; i < depth; i++) {
end = nstack.length();
for(int j = begin; j < end; j++) {
Node* tp = nstack.at(j);
uint limit = direction > 0 ? tp->len() : tp->outcnt();
for(uint k = 0; k < limit; k++) {
Node* n = direction > 0 ? tp->in(k) : tp->raw_out(k);
if (NotANode(n)) continue;
// do not recurse through top or the root (would reach unrelated stuff)
if (n->is_Root() || n->is_top()) continue;
if (only_ctrl && !n->is_CFG()) continue;
bool on_stack = nstack.contains(n);
if (!on_stack) {
nstack.append(n);
}
}
}
begin = end;
}
end = nstack.length();
if (direction > 0) {
for(int j = end-1; j >= 0; j--) {
nstack.at(j)->dump();
}
} else {
for(int j = 0; j < end; j++) {
nstack.at(j)->dump();
}
}
}
//------------------------------dump-------------------------------------------
void Node::dump(int d) const {
dump_nodes(this, d, false);
}
//------------------------------dump_ctrl--------------------------------------
// Dump a Node's control history to depth
void Node::dump_ctrl(int d) const {
dump_nodes(this, d, true);
}
// VERIFICATION CODE
// For each input edge to a node (ie - for each Use-Def edge), verify that
// there is a corresponding Def-Use edge.
//------------------------------verify_edges-----------------------------------
void Node::verify_edges(Unique_Node_List &visited) {
uint i, j, idx;
int cnt;
Node *n;
// Recursive termination test
if (visited.member(this)) return;
visited.push(this);
// Walk over all input edges, checking for correspondence
for( i = 0; i < len(); i++ ) {
n = in(i);
if (n != NULL && !n->is_top()) {
// Count instances of (Node *)this
cnt = 0;
for (idx = 0; idx < n->_outcnt; idx++ ) {
if (n->_out[idx] == (Node *)this) cnt++;
}
assert( cnt > 0,"Failed to find Def-Use edge." );
// Check for duplicate edges
// walk the input array downcounting the input edges to n
for( j = 0; j < len(); j++ ) {
if( in(j) == n ) cnt--;
}
assert( cnt == 0,"Mismatched edge count.");
} else if (n == NULL) {
assert(i >= req() || i == 0 || is_Region() || is_Phi(), "only regions or phis have null data edges");
} else {
assert(n->is_top(), "sanity");
// Nothing to check.
}
}
// Recursive walk over all input edges
for( i = 0; i < len(); i++ ) {
n = in(i);
if( n != NULL )
in(i)->verify_edges(visited);
}
}
//------------------------------verify_recur-----------------------------------
static const Node *unique_top = NULL;
void Node::verify_recur(const Node *n, int verify_depth,
VectorSet &old_space, VectorSet &new_space) {
if ( verify_depth == 0 ) return;
if (verify_depth > 0) --verify_depth;
Compile* C = Compile::current();
// Contained in new_space or old_space?
VectorSet *v = C->node_arena()->contains(n) ? &new_space : &old_space;
// Check for visited in the proper space. Numberings are not unique
// across spaces so we need a separate VectorSet for each space.
if( v->test_set(n->_idx) ) return;
if (n->is_Con() && n->bottom_type() == Type::TOP) {
if (C->cached_top_node() == NULL)
C->set_cached_top_node((Node*)n);
assert(C->cached_top_node() == n, "TOP node must be unique");
}
for( uint i = 0; i < n->len(); i++ ) {
Node *x = n->in(i);
if (!x || x->is_top()) continue;
// Verify my input has a def-use edge to me
if (true /*VerifyDefUse*/) {
// Count use-def edges from n to x
int cnt = 0;
for( uint j = 0; j < n->len(); j++ )
if( n->in(j) == x )
cnt++;
// Count def-use edges from x to n
uint max = x->_outcnt;
for( uint k = 0; k < max; k++ )
if (x->_out[k] == n)
cnt--;
assert( cnt == 0, "mismatched def-use edge counts" );
}
verify_recur(x, verify_depth, old_space, new_space);
}
}
//------------------------------verify-----------------------------------------
// Check Def-Use info for my subgraph
void Node::verify() const {
Compile* C = Compile::current();
Node* old_top = C->cached_top_node();
ResourceMark rm;
ResourceArea *area = Thread::current()->resource_area();
VectorSet old_space(area), new_space(area);
verify_recur(this, -1, old_space, new_space);
C->set_cached_top_node(old_top);
}
#endif
//------------------------------walk-------------------------------------------
// Graph walk, with both pre-order and post-order functions
void Node::walk(NFunc pre, NFunc post, void *env) {
VectorSet visited(Thread::current()->resource_area()); // Setup for local walk
walk_(pre, post, env, visited);
}
void Node::walk_(NFunc pre, NFunc post, void *env, VectorSet &visited) {
if( visited.test_set(_idx) ) return;
pre(*this,env); // Call the pre-order walk function
for( uint i=0; i<_max; i++ )
if( in(i) ) // Input exists and is not walked?
in(i)->walk_(pre,post,env,visited); // Walk it with pre & post functions
post(*this,env); // Call the post-order walk function
}
void Node::nop(Node &, void*) {}
//------------------------------Registers--------------------------------------
// Do we Match on this edge index or not? Generally false for Control
// and true for everything else. Weird for calls & returns.
uint Node::match_edge(uint idx) const {
return idx; // True for other than index 0 (control)
}
static RegMask _not_used_at_all;
// Register classes are defined for specific machines
const RegMask &Node::out_RegMask() const {
ShouldNotCallThis();
return _not_used_at_all;
}
const RegMask &Node::in_RegMask(uint) const {
ShouldNotCallThis();
return _not_used_at_all;
}
//=============================================================================
//-----------------------------------------------------------------------------
void Node_Array::reset( Arena *new_arena ) {
_a->Afree(_nodes,_max*sizeof(Node*));
_max = 0;
_nodes = NULL;
_a = new_arena;
}
//------------------------------clear------------------------------------------
// Clear all entries in _nodes to NULL but keep storage
void Node_Array::clear() {
Copy::zero_to_bytes( _nodes, _max*sizeof(Node*) );
}
//-----------------------------------------------------------------------------
void Node_Array::grow( uint i ) {
if( !_max ) {
_max = 1;
_nodes = (Node**)_a->Amalloc( _max * sizeof(Node*) );
_nodes[0] = NULL;
}
uint old = _max;
while( i >= _max ) _max <<= 1; // Double to fit
_nodes = (Node**)_a->Arealloc( _nodes, old*sizeof(Node*),_max*sizeof(Node*));
Copy::zero_to_bytes( &_nodes[old], (_max-old)*sizeof(Node*) );
}
//-----------------------------------------------------------------------------
void Node_Array::insert( uint i, Node *n ) {
if( _nodes[_max-1] ) grow(_max); // Get more space if full
Copy::conjoint_words_to_higher((HeapWord*)&_nodes[i], (HeapWord*)&_nodes[i+1], ((_max-i-1)*sizeof(Node*)));
_nodes[i] = n;
}
//-----------------------------------------------------------------------------
void Node_Array::remove( uint i ) {
Copy::conjoint_words_to_lower((HeapWord*)&_nodes[i+1], (HeapWord*)&_nodes[i], ((_max-i-1)*sizeof(Node*)));
_nodes[_max-1] = NULL;
}
//-----------------------------------------------------------------------------
void Node_Array::sort( C_sort_func_t func) {
qsort( _nodes, _max, sizeof( Node* ), func );
}
//-----------------------------------------------------------------------------
void Node_Array::dump() const {
#ifndef PRODUCT
for( uint i = 0; i < _max; i++ ) {
Node *nn = _nodes[i];
if( nn != NULL ) {
tty->print("%5d--> ",i); nn->dump();
}
}
#endif
}
//--------------------------is_iteratively_computed------------------------------
// Operation appears to be iteratively computed (such as an induction variable)
// It is possible for this operation to return false for a loop-varying
// value, if it appears (by local graph inspection) to be computed by a simple conditional.
bool Node::is_iteratively_computed() {
if (ideal_reg()) { // does operation have a result register?
for (uint i = 1; i < req(); i++) {
Node* n = in(i);
if (n != NULL && n->is_Phi()) {
for (uint j = 1; j < n->req(); j++) {
if (n->in(j) == this) {
return true;
}
}
}
}
}
return false;
}
//--------------------------find_similar------------------------------
// Return a node with opcode "opc" and same inputs as "this" if one can
// be found; Otherwise return NULL;
Node* Node::find_similar(int opc) {
if (req() >= 2) {
Node* def = in(1);
if (def && def->outcnt() >= 2) {
for (DUIterator_Fast dmax, i = def->fast_outs(dmax); i < dmax; i++) {
Node* use = def->fast_out(i);
if (use->Opcode() == opc &&
use->req() == req()) {
uint j;
for (j = 0; j < use->req(); j++) {
if (use->in(j) != in(j)) {
break;
}
}
if (j == use->req()) {
return use;
}
}
}
}
}
return NULL;
}
//--------------------------unique_ctrl_out------------------------------
// Return the unique control out if only one. Null if none or more than one.
Node* Node::unique_ctrl_out() {
Node* found = NULL;
for (uint i = 0; i < outcnt(); i++) {
Node* use = raw_out(i);
if (use->is_CFG() && use != this) {
if (found != NULL) return NULL;
found = use;
}
}
return found;
}
//=============================================================================
//------------------------------yank-------------------------------------------
// Find and remove
void Node_List::yank( Node *n ) {
uint i;
for( i = 0; i < _cnt; i++ )
if( _nodes[i] == n )
break;
if( i < _cnt )
_nodes[i] = _nodes[--_cnt];
}
//------------------------------dump-------------------------------------------
void Node_List::dump() const {
#ifndef PRODUCT
for( uint i = 0; i < _cnt; i++ )
if( _nodes[i] ) {
tty->print("%5d--> ",i);
_nodes[i]->dump();
}
#endif
}
void Node_List::dump_simple() const {
#ifndef PRODUCT
for( uint i = 0; i < _cnt; i++ )
if( _nodes[i] ) {
tty->print(" %d", _nodes[i]->_idx);
} else {
tty->print(" NULL");
}
#endif
}
//=============================================================================
//------------------------------remove-----------------------------------------
void Unique_Node_List::remove( Node *n ) {
if( _in_worklist[n->_idx] ) {
for( uint i = 0; i < size(); i++ )
if( _nodes[i] == n ) {
map(i,Node_List::pop());
_in_worklist >>= n->_idx;
return;
}
ShouldNotReachHere();
}
}
//-----------------------remove_useless_nodes----------------------------------
// Remove useless nodes from worklist
void Unique_Node_List::remove_useless_nodes(VectorSet &useful) {
for( uint i = 0; i < size(); ++i ) {
Node *n = at(i);
assert( n != NULL, "Did not expect null entries in worklist");
if( ! useful.test(n->_idx) ) {
_in_worklist >>= n->_idx;
map(i,Node_List::pop());
// Node *replacement = Node_List::pop();
// if( i != size() ) { // Check if removing last entry
// _nodes[i] = replacement;
// }
--i; // Visit popped node
// If it was last entry, loop terminates since size() was also reduced
}
}
}
//=============================================================================
void Node_Stack::grow() {
size_t old_top = pointer_delta(_inode_top,_inodes,sizeof(INode)); // save _top
size_t old_max = pointer_delta(_inode_max,_inodes,sizeof(INode));
size_t max = old_max << 1; // max * 2
_inodes = REALLOC_ARENA_ARRAY(_a, INode, _inodes, old_max, max);
_inode_max = _inodes + max;
_inode_top = _inodes + old_top; // restore _top
}
// Node_Stack is used to map nodes.
Node* Node_Stack::find(uint idx) const {
uint sz = size();
for (uint i=0; i < sz; i++) {
if (idx == index_at(i) )
return node_at(i);
}
return NULL;
}
sssssssss71
最新推荐文章于 2024-05-13 09:43:22 发布