Change the algorithm and generated code for parallel moves on IA32.

src/ia32/lithium-gap-resolver-ia32.cc

Index: src/ia32/lithium-gap-resolver-ia32.cc
diff --git a/src/ia32/lithium-gap-resolver-ia32.cc b/src/ia32/lithium-gap-resolver-ia32.cc
new file mode 100644
index 0000000000000000000000000000000000000000..3bef55135fb7c61f98d7e4781fdd1159c408c229
--- /dev/null
+++ b/src/ia32/lithium-gap-resolver-ia32.cc
@@ -0,0 +1,446 @@
+// Copyright 2011 the V8 project authors. All rights reserved.
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+//       notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+//       copyright notice, this list of conditions and the following
+//       disclaimer in the documentation and/or other materials provided
+//       with the distribution.
+//     * Neither the name of Google Inc. nor the names of its
+//       contributors may be used to endorse or promote products derived
+//       from this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+#include "ia32/lithium-gap-resolver-ia32.h"
+#include "ia32/lithium-codegen-ia32.h"
+
+namespace v8 {
+namespace internal {
+
+LGapResolver::LGapResolver(LCodeGen* owner)
+    : cgen_(owner), moves_(32), spilled_register_(-1) {

[William Hesse, 2011/01/17 10:43:43]

Can we use no_reg, or no_reg.code(), for no spilled register?

[Kevin Millikin (Chromium), 2011/01/17 11:18:17]

That seems like a subtle type error.  This is an allocation index, not a
register code.

+  for (int i = 0; i < Register::kNumAllocatableRegisters; ++i) {
+    source_uses_[i] = 0;
+    destination_uses_[i] = 0;
+  }
+}
+
+
+void LGapResolver::Resolve(LParallelMove* parallel_move) {
+  ASSERT(HasBeenReset());
+  // Build up a worklist of moves.
+  BuildInitialMoveList(parallel_move);
+
+  for (int i = moves_.length() - 1; i >= 0; --i) {
+    LMoveOperands move = moves_[i];
+    // Skip constants to perform them last.  They don't block other moves
+    // and skipping such moves with register destinations keeps those
+    // registers free for the whole algorithm.
+    if (!move.IsEliminated() && !move.source()->IsConstantOperand()) {
+      PerformMove(i);
+    }
+  }
+
+  // Perform the moves with constant sources.
+  for (int i = moves_.length() - 1; i >= 0; --i) {
+    if (!moves_[i].IsEliminated()) {
+      ASSERT(moves_[i].source()->IsConstantOperand());
+      EmitMove(i);
+    }
+  }
+
+  Finish();
+  ASSERT(HasBeenReset());
+}
+
+
+void LGapResolver::BuildInitialMoveList(LParallelMove* parallel_move) {
+  // Perform a linear sweep of the moves to add them to the initial list of
+  // moves to perform, ignoring any move that is redundant (the source is
+  // the same as the destination, the destination is ignored and
+  // unallocated, or the move was already eliminated).
+  const ZoneList<LMoveOperands>* moves = parallel_move->move_operands();
+  for (int i = moves->length() - 1; i >= 0; --i) {
+    LMoveOperands move = moves->at(i);
+    if (!move.IsRedundant()) AddMove(move);
+  }
+}
+
+
+void LGapResolver::PerformMove(int index) {

[fschneider, 2011/01/17 09:44:16]

Would it make the code more concise by passing a LMoveOperands* to
PerformMove/RemoveMove/...?

[Kevin Millikin (Chromium), 2011/01/17 10:13:09]

A little, but I'm nervous about passing addresses of elements in a List.  Adding
to the list can reallocate the backing store.

Of course, here the algorithm doesn't necessarily work if we add to the list
after we start performing moves, but I'd still like to be careful.

If nothing else, the index makes it clear that the list is holding the pair and
passing a pointer obscures that a bit.

+  // Each call to this function performs a move and deletes it from the move
+  // graph.  We first recursively perform any move blocking this one.  We
+  // mark a move as "pending" on entry to PerformMove in order to detect
+  // cycles in the move graph.  We use operand swaps to resolve cycles,
+  // which means that a call to PerformMove could change any source operand
+  // in the move graph.
+
+  ASSERT(!moves_[index].IsRedundant());

[fschneider, 2011/01/17 09:44:16]

Can we ASSSERT(!moves_[index].IsPending()) here? 

I think this is already covered by the !IsRedundant(), but
the predicates are a little confusing.

[Kevin Millikin (Chromium), 2011/01/17 10:13:09]

I agree the predicates are confusing (one has to read their implementation). 
That's probably unavoidable while keeping their names relatively short.

I will assert !IsPending() && !IsRedundant() (the assert will hit a null
dereference otherwise, which still catches the case but is a little less direct.

+
+  // Clear this move's destination to indicate a pending move.  The actual
+  // destination is saved on the side.
+  ASSERT(moves_[index].source() != NULL);  // Or else it will look eliminated.
+  LOperand* destination = moves_[index].destination();
+  moves_[index].set_destination(NULL);
+
+  // Perform a depth-first traversal of the move graph to resolve
+  // dependencies.  Any unperformed, unpending move with a source the same
+  // as this one's destination blocks this one so recursively perform all
+  // such moves.
+  for (int i = moves_.length() - 1; i >= 0; --i) {

[fschneider, 2011/01/17 09:44:16]

Reason for iterating in reverse order? Maybe add a comment about it.

[Kevin Millikin (Chromium), 2011/01/17 10:13:09]

No reason, order is abitrary.  I'll just stick with forward so nobody has to
wonder about it.

[William Hesse, 2011/01/17 10:43:43]

Do we have the source use counts at this point?  I guess only for registers. 
Can we skip this loop if the destination is a register, and its use count is 0?

[Kevin Millikin (Chromium), 2011/01/17 11:18:17]

We have source use counts for the general purpose registers.  I wanted to keep
things simple: most gaps have 0, 1, or two moves.

+    LMoveOperands other_move = moves_[i];
+    if (other_move.Blocks(destination) && !other_move.IsPending()) {
+      // Though PerformMove can change any source operand in the move graph,
+      // this call cannot create a blocking move via a swap (this loop does
+      // not miss any).  Assume there is a non-blocking move with source A
+      // and this move is blocked on source B and there is a swap of A and
+      // B.  Then A and B must be involved in the same cycle (or they would
+      // not be swapped).  Since this move's destination is B and there is
+      // only a single incoming edge to an operand, this move must also be
+      // involved in the same cycle.  In that case, the blocking move will
+      // be created but will be "pending" when we return from PerformMove.
+      PerformMove(i);
+    }
+  }
+
+  // We are about to resolve this move and don't need it marked as
+  // pending, so restore its destination.
+  moves_[index].set_destination(destination);
+
+  // This move's source may have changed due to swaps to resolve cycles and
+  // so it may now be the last move in the cycle.  If so remove it.
+  if (moves_[index].source()->Equals(destination)) {
+    RemoveMove(index);
+    return;
+  }
+
+  // The move may be blocked on a (at most one) pending move, in which case
+  // we have a cycle.  Search for such a blocking move and perform a swap to
+  // resolve it.
+  for (int i = moves_.length() - 1; i >= 0; --i) {
+    LMoveOperands other_move = moves_[i];
+    if (other_move.Blocks(destination)) {
+      ASSERT(other_move.IsPending());
+      EmitSwap(index);
+      return;
+    }
+  }
+
+  // This move is not blocked.
+  EmitMove(index);
+}
+
+
+void LGapResolver::AddMove(LMoveOperands move) {

[fschneider, 2011/01/17 09:44:16]

Maybe pass a LMoveOperands* here.

[Kevin Millikin (Chromium), 2011/01/17 10:13:09]

I'd rather not.  We don't need to mutate it and it's another pointer into a
List.

+  LOperand* source = move.source();
+  if (source->IsRegister()) ++source_uses_[source->index()];
+
+  LOperand* destination = move.destination();
+  if (destination->IsRegister()) ++destination_uses_[destination->index()];
+
+  moves_.Add(move);
+}
+
+
+void LGapResolver::RemoveMove(int index) {
+  LOperand* source = moves_[index].source();
+  if (source->IsRegister()) {
+    --source_uses_[source->index()];
+    ASSERT(source_uses_[source->index()] >= 0);
+  }
+
+  LOperand* destination = moves_[index].destination();
+  if (destination->IsRegister()) {
+    --destination_uses_[destination->index()];
+    ASSERT(destination_uses_[destination->index()] >= 0);
+  }
+
+  moves_[index].Eliminate();
+}
+
+
+int LGapResolver::CountSourceUses(LOperand* operand) {
+  int count = 0;
+  for (int i = moves_.length() - 1; i >= 0; --i) {
+    if (!moves_[i].IsEliminated() && moves_[i].source()->Equals(operand)) {
+      ++count;
+    }
+  }
+  return count;
+}
+
+
+Register LGapResolver::GetFreeRegisterNot(Register reg) {
+  int skip_index = reg.is(no_reg) ? -1 : Register::ToAllocationIndex(reg);
+  for (int i = 0; i < Register::kNumAllocatableRegisters; ++i) {
+    if (source_uses_[i] == 0 && destination_uses_[i] > 0 && i != skip_index) {
+      return Register::FromAllocationIndex(i);
+    }
+  }
+  return no_reg;
+}
+
+
+bool LGapResolver::HasBeenReset() {
+  if (!moves_.is_empty()) return false;
+  if (spilled_register_ >= 0) return false;
+
+  for (int i = 0; i < Register::kNumAllocatableRegisters; ++i) {
+    if (source_uses_[i] != 0) return false;
+    if (destination_uses_[i] != 0) return false;
+  }
+  return true;
+}
+
+
+#define __ ACCESS_MASM(cgen_->masm())
+
+void LGapResolver::Finish() {
+  if (spilled_register_ >= 0) {
+    __ pop(Register::FromAllocationIndex(spilled_register_));
+    spilled_register_ = -1;
+  }
+  moves_.Rewind(0);

[William Hesse, 2011/01/17 10:43:43]

Shouldn't we have a List::Clear for Rewind(0), and also shouldn't Rewind()
assert that the new index < length_?

Obviously, this could go in a different change.

[Kevin Millikin (Chromium), 2011/01/17 11:18:17]

We have List::Clear but it does something different (deallocates the backing
store).  At this point List::Rewind(0) is a V8 idiom.

+}
+
+
+void LGapResolver::EnsureRestored(LOperand* operand) {
+  if (operand->IsRegister() && operand->index() == spilled_register_) {
+    __ pop(Register::FromAllocationIndex(spilled_register_));
+    spilled_register_ = -1;
+  }
+}
+
+
+Register LGapResolver::EnsureTempRegister() {
+  // 1. We may have already spilled to create a temp register.
+  if (spilled_register_ >= 0) {
+    return Register::FromAllocationIndex(spilled_register_);
+  }
+
+  // 2. We may have a free register that we can use without spilling.
+  Register free = GetFreeRegisterNot(no_reg);
+  if (!free.is(no_reg)) return free;
+
+  // 3. Prefer to spill a register that is not used in any remaining move
+  // because it will not need to be restored until the end.
+  for (int i = 0; i < Register::kNumAllocatableRegisters; ++i) {
+    if (source_uses_[i] == 0 && destination_uses_[i] == 0) {
+      Register scratch = Register::FromAllocationIndex(i);
+      __ push(scratch);
+      spilled_register_ = i;
+      return scratch;
+    }
+  }
+
+  // 4. Use an arbitrary register.  Register 0 is as arbitrary as any other.
+  Register scratch = Register::FromAllocationIndex(0);
+  __ push(scratch);
+  spilled_register_ = 0;
+  return scratch;
+}
+
+
+void LGapResolver::EmitMove(int index) {
+  LOperand* source = moves_[index].source();
+  LOperand* destination = moves_[index].destination();
+  EnsureRestored(source);
+  EnsureRestored(destination);
+
+  // Dispatch on the source and destination operand kinds.  Not all
+  // combinations are possible.

[William Hesse, 2011/01/17 10:43:43]

Are only combinations prohibited by the double vs word size difference, and
constants as destinations, prohibited, or are there other, hidden, non-supported
cases?

[Kevin Millikin (Chromium), 2011/01/17 11:18:17]

The first.

+  if (source->IsRegister()) {
+    ASSERT(destination->IsRegister() || destination->IsStackSlot());
+    Register src = cgen_->ToRegister(source);
+    Operand dst = cgen_->ToOperand(destination);
+    __ mov(dst, src);
+
+  } else if (source->IsStackSlot()) {
+    ASSERT(destination->IsRegister() || destination->IsStackSlot());
+    Operand src = cgen_->ToOperand(source);
+    if (destination->IsRegister()) {
+      Register dst = cgen_->ToRegister(destination);
+      __ mov(dst, src);
+    } else {
+      // Spill on demand to use a temporary register for memory-to-memory
+      // moves.
+      Register tmp = EnsureTempRegister();
+      Operand dst = cgen_->ToOperand(destination);
+      __ mov(tmp, src);
+      __ mov(dst, tmp);
+    }
+
+  } else if (source->IsConstantOperand()) {
+    ASSERT(destination->IsRegister() || destination->IsStackSlot());
+    Immediate src = cgen_->ToImmediate(source);
+    Operand dst = cgen_->ToOperand(destination);
+    __ mov(dst, src);
+
+  } else if (source->IsDoubleRegister()) {
+    ASSERT(destination->IsDoubleRegister() ||
+           destination->IsDoubleStackSlot());
+    XMMRegister src = cgen_->ToDoubleRegister(source);
+    Operand dst = cgen_->ToOperand(destination);
+    __ movdbl(dst, src);
+
+  } else if (source->IsDoubleStackSlot()) {
+    ASSERT(destination->IsDoubleRegister() ||
+           destination->IsDoubleStackSlot());
+    Operand src = cgen_->ToOperand(source);
+    if (destination->IsDoubleRegister()) {
+      XMMRegister dst = cgen_->ToDoubleRegister(destination);
+      __ movdbl(dst, src);
+    } else {
+      // We rely on having xmm0 available as a fixed scratch register.
+      Operand dst = cgen_->ToOperand(destination);
+      __ movdbl(xmm0, src);
+      __ movdbl(dst, xmm0);
+    }
+
+  } else {
+    UNREACHABLE();
+  }
+
+  RemoveMove(index);
+}
+
+
+void LGapResolver::EmitSwap(int index) {
+  LOperand* source = moves_[index].source();
+  LOperand* destination = moves_[index].destination();
+  EnsureRestored(source);
+  EnsureRestored(destination);
+
+  // Dispatch on the source and destination operand kinds.  Not all
+  // combinations are possible.
+  if (source->IsRegister() && destination->IsRegister()) {
+    // Register-register.
+    Register src = cgen_->ToRegister(source);
+    Register dst = cgen_->ToRegister(destination);
+    __ xchg(dst, src);
+
+  } else if ((source->IsRegister() && destination->IsStackSlot()) ||
+             (source->IsStackSlot() && destination->IsRegister())) {
+    // Register-memory.  Use a free register as a temp if possible.  Do not
+    // spill on demand because the simple spill implementation cannot avoid
+    // spilling src at this point.  Memory-to-memory swaps are a rare case.

[William Hesse, 2011/01/17 10:43:43]

Why is the memory-to-memory comment here?

+    Register tmp = GetFreeRegisterNot(no_reg);
+    Register reg =
+        cgen_->ToRegister(source->IsRegister() ? source : destination);
+    Operand mem =
+        cgen_->ToOperand(source->IsRegister() ? destination : source);
+    if (tmp.is(no_reg)) {
+      __ xor_(reg, mem);
+      __ xor_(mem, reg);
+      __ xor_(reg, mem);
+    } else {
+      __ mov(tmp, mem);
+      __ mov(mem, reg);
+      __ mov(reg, tmp);
+    }
+
+  } else if (source->IsStackSlot() && destination->IsStackSlot()) {
+    // Memory-memory.  Spill on demand to use a temporary.  If there is a
+    // free register after that, use it as a second temporary.
+    Register tmp0 = EnsureTempRegister();
+    Register tmp1 = GetFreeRegisterNot(tmp0);
+    Operand src = cgen_->ToOperand(source);
+    Operand dst = cgen_->ToOperand(destination);
+    if (tmp1.is(no_reg)) {
+      // Only one temp register available to us.
+      __ mov(tmp0, dst);
+      __ xor_(tmp0, src);
+      __ xor_(src, tmp0);
+      __ xor_(tmp0, src);
+      __ mov(dst, tmp0);
+    } else {
+      __ mov(tmp0, dst);
+      __ mov(tmp1, src);
+      __ mov(dst, tmp1);
+      __ mov(src, tmp0);
+    }
+
+  } else if (source->IsDoubleRegister() || destination->IsDoubleRegister()) {
+    // XMM register-register or register-memory.  We rely on having xmm0
+    // available as a fixed scratch register.
+    ASSERT(source->IsDoubleRegister() || source->IsDoubleStackSlot());
+    ASSERT(destination->IsDoubleRegister() ||
+           destination->IsDoubleStackSlot());
+    XMMRegister reg = cgen_->ToDoubleRegister(source->IsDoubleRegister()
+                                                  ? source
+                                                  : destination);
+    Operand other =
+        cgen_->ToOperand(source->IsDoubleRegister() ? destination : source);
+    __ movdbl(xmm0, other);
+    __ movdbl(other, reg);
+    __ movdbl(reg, Operand(xmm0));
+
+  } else if (source->IsDoubleStackSlot() && destination->IsDoubleStackSlot()) {
+    // Double-width memory-to-memory.  Spill on demand to use a general
+    // purpose temporary register and also rely on having xmm0 available as
+    // a fixed scratch register.
+    Register tmp = EnsureTempRegister();
+    Operand src0 = cgen_->ToOperand(source);
+    Operand src1 = cgen_->HighOperand(source);
+    Operand dst0 = cgen_->ToOperand(destination);
+    Operand dst1 = cgen_->HighOperand(destination);
+    __ movdbl(xmm0, dst0);  // Save destination in xmm0.
+    __ mov(tmp, src0);  // Then use tmp to copy source to destination.
+    __ mov(dst0, tmp);
+    __ mov(tmp, src1);
+    __ mov(dst1, tmp);
+    __ movdbl(src0, xmm0);
+
+  } else {
+    // No other combinations are possible.
+    UNREACHABLE();
+  }
+
+  // The swap of source and destination has executed a move from source to
+  // destination.
+  RemoveMove(index);
+
+  // Any unperformed (including pending) move with a source of either
+  // this move's source or destination needs to have their source
+  // changed to reflect the state of affairs after the swap.
+  for (int j = moves_.length() - 1; j >= 0; --j) {
+    LMoveOperands other_move = moves_[j];
+    if (other_move.Blocks(source)) {
+      moves_[j].set_source(destination);
+    } else if (other_move.Blocks(destination)) {
+      moves_[j].set_source(source);
+    }
+  }
+
+  // In addition to swapping the actual uses as sources, we need to update
+  // the use counts.
+  if (source->IsRegister() && destination->IsRegister()) {
+    int temp = source_uses_[source->index()];
+    source_uses_[source->index()] = source_uses_[destination->index()];
+    source_uses_[destination->index()] = temp;
+  } else if (source->IsRegister()) {
+    // We don't have use counts for non-register operands like destination.
+    // Compute those counts now.
+    source_uses_[source->index()] = CountSourceUses(source);
+  } else if (destination->IsRegister()) {
+    source_uses_[destination->index()] = CountSourceUses(destination);
+  }
+}
+
+#undef __
+
+} }  // namespace v8::internal