// Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'.
// Note that if D is also part of the expression tree that we recurse to
// linearize it as well. Besides that case, this does not recurse into A,B, or
// C.
void Reassociate::LinearizeExpr(BinaryOperator *I) {
BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1));
assert(isReassociableOp(LHS, I->getOpcode()) &&
isReassociableOp(RHS, I->getOpcode()) &&
"Not an expression that needs linearization?");
DEBUG(dbgs() << "Linear" << *LHS << '\n' << *RHS << '\n' << *I << '\n');
// Move the RHS instruction to live immediately before I, avoiding breaking
// dominator properties.
RHS->moveBefore(I);
// Move operands around to do the linearization.
I->setOperand(1, RHS->getOperand(0));
RHS->setOperand(0, LHS);
I->setOperand(0, RHS);
// Conservatively clear all the optional flags, which may not hold
// after the reassociation.
I->clearSubclassOptionalData();
LHS->clearSubclassOptionalData();
RHS->clearSubclassOptionalData();
++NumLinear;
MadeChange = true;
DEBUG(dbgs() << "Linearized: " << *I << '\n');
// If D is part of this expression tree, tail recurse.
if (isReassociableOp(I->getOperand(1), I->getOpcode()))
LinearizeExpr(I);
}
Value *ConstantOffsetExtractor::removeConstOffset(unsigned ChainIndex) {
if (ChainIndex == 0) {
assert(isa<ConstantInt>(UserChain[ChainIndex]));
return ConstantInt::getNullValue(UserChain[ChainIndex]->getType());
}
BinaryOperator *BO = cast<BinaryOperator>(UserChain[ChainIndex]);
unsigned OpNo = (BO->getOperand(0) == UserChain[ChainIndex - 1] ? 0 : 1);
assert(BO->getOperand(OpNo) == UserChain[ChainIndex - 1]);
Value *NextInChain = removeConstOffset(ChainIndex - 1);
Value *TheOther = BO->getOperand(1 - OpNo);
// If NextInChain is 0 and not the LHS of a sub, we can simplify the
// sub-expression to be just TheOther.
if (ConstantInt *CI = dyn_cast<ConstantInt>(NextInChain)) {
if (CI->isZero() && !(BO->getOpcode() == Instruction::Sub && OpNo == 0))
return TheOther;
}
if (BO->getOpcode() == Instruction::Or) {
// Rebuild "or" as "add", because "or" may be invalid for the new
// epxression.
//
// For instance, given
// a | (b + 5) where a and b + 5 have no common bits,
// we can extract 5 as the constant offset.
//
// However, reusing the "or" in the new index would give us
// (a | b) + 5
// which does not equal a | (b + 5).
//
// Replacing the "or" with "add" is fine, because
// a | (b + 5) = a + (b + 5) = (a + b) + 5
if (OpNo == 0) {
return BinaryOperator::CreateAdd(NextInChain, TheOther, BO->getName(),
IP);
} else {
return BinaryOperator::CreateAdd(TheOther, NextInChain, BO->getName(),
IP);
}
}
// We can reuse BO in this case, because the new expression shares the same
// instruction type and BO is used at most once.
assert(BO->getNumUses() <= 1 &&
"distributeExtsAndCloneChain clones each BinaryOperator in "
"UserChain, so no one should be used more than "
"once");
BO->setOperand(OpNo, NextInChain);
BO->setHasNoSignedWrap(false);
BO->setHasNoUnsignedWrap(false);
// Make sure it appears after all instructions we've inserted so far.
BO->moveBefore(IP);
return BO;
}
// RewriteExprTree - Now that the operands for this expression tree are
// linearized and optimized, emit them in-order. This function is written to be
// tail recursive.
void Reassociate::RewriteExprTree(BinaryOperator *I,
SmallVectorImpl<ValueEntry> &Ops,
unsigned i) {
if (i+2 == Ops.size()) {
if (I->getOperand(0) != Ops[i].Op ||
I->getOperand(1) != Ops[i+1].Op) {
Value *OldLHS = I->getOperand(0);
DEBUG(dbgs() << "RA: " << *I << '\n');
I->setOperand(0, Ops[i].Op);
I->setOperand(1, Ops[i+1].Op);
// Clear all the optional flags, which may not hold after the
// reassociation if the expression involved more than just this operation.
if (Ops.size() != 2)
I->clearSubclassOptionalData();
DEBUG(dbgs() << "TO: " << *I << '\n');
MadeChange = true;
++NumChanged;
// If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3)
// delete the extra, now dead, nodes.
RemoveDeadBinaryOp(OldLHS);
}
return;
}
assert(i+2 < Ops.size() && "Ops index out of range!");
if (I->getOperand(1) != Ops[i].Op) {
DEBUG(dbgs() << "RA: " << *I << '\n');
I->setOperand(1, Ops[i].Op);
// Conservatively clear all the optional flags, which may not hold
// after the reassociation.
I->clearSubclassOptionalData();
DEBUG(dbgs() << "TO: " << *I << '\n');
MadeChange = true;
++NumChanged;
}
BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0));
assert(LHS->getOpcode() == I->getOpcode() &&
"Improper expression tree!");
// Compactify the tree instructions together with each other to guarantee
// that the expression tree is dominated by all of Ops.
LHS->moveBefore(I);
RewriteExprTree(LHS, Ops, i+1);
}
// NegateValue - Insert instructions before the instruction pointed to by BI,
// that computes the negative version of the value specified. The negative
// version of the value is returned, and BI is left pointing at the instruction
// that should be processed next by the reassociation pass.
//
static Value *NegateValue(Value *V, Instruction *BI) {
if (Constant *C = dyn_cast<Constant>(V))
return ConstantExpr::getNeg(C);
// We are trying to expose opportunity for reassociation. One of the things
// that we want to do to achieve this is to push a negation as deep into an
// expression chain as possible, to expose the add instructions. In practice,
// this means that we turn this:
// X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D
// so that later, a: Y = 12+X could get reassociated with the -12 to eliminate
// the constants. We assume that instcombine will clean up the mess later if
// we introduce tons of unnecessary negation instructions.
//
if (Instruction *I = dyn_cast<Instruction>(V))
if (I->getOpcode() == Instruction::Add && I->hasOneUse()) {
// Push the negates through the add.
I->setOperand(0, NegateValue(I->getOperand(0), BI));
I->setOperand(1, NegateValue(I->getOperand(1), BI));
// We must move the add instruction here, because the neg instructions do
// not dominate the old add instruction in general. By moving it, we are
// assured that the neg instructions we just inserted dominate the
// instruction we are about to insert after them.
//
I->moveBefore(BI);
I->setName(I->getName()+".neg");
return I;
}
// Okay, we need to materialize a negated version of V with an instruction.
// Scan the use lists of V to see if we have one already.
for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
User *U = *UI;
if (!BinaryOperator::isNeg(U)) continue;
// We found one! Now we have to make sure that the definition dominates
// this use. We do this by moving it to the entry block (if it is a
// non-instruction value) or right after the definition. These negates will
// be zapped by reassociate later, so we don't need much finesse here.
BinaryOperator *TheNeg = cast<BinaryOperator>(U);
// Verify that the negate is in this function, V might be a constant expr.
if (TheNeg->getParent()->getParent() != BI->getParent()->getParent())
continue;
BasicBlock::iterator InsertPt;
if (Instruction *InstInput = dyn_cast<Instruction>(V)) {
if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) {
InsertPt = II->getNormalDest()->begin();
} else {
InsertPt = InstInput;
++InsertPt;
}
while (isa<PHINode>(InsertPt)) ++InsertPt;
} else {
InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin();
}
TheNeg->moveBefore(InsertPt);
return TheNeg;
}
// Insert a 'neg' instruction that subtracts the value from zero to get the
// negation.
return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI);
}
/// LinearizeExprTree - Given an associative binary expression tree, traverse
/// all of the uses putting it into canonical form. This forces a left-linear
/// form of the expression (((a+b)+c)+d), and collects information about the
/// rank of the non-tree operands.
///
/// NOTE: These intentionally destroys the expression tree operands (turning
/// them into undef values) to reduce #uses of the values. This means that the
/// caller MUST use something like RewriteExprTree to put the values back in.
///
void Reassociate::LinearizeExprTree(BinaryOperator *I,
SmallVectorImpl<ValueEntry> &Ops) {
Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
unsigned Opcode = I->getOpcode();
// First step, linearize the expression if it is in ((A+B)+(C+D)) form.
BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode);
BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode);
// If this is a multiply expression tree and it contains internal negations,
// transform them into multiplies by -1 so they can be reassociated.
if (I->getOpcode() == Instruction::Mul) {
if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) {
LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap);
LHSBO = isReassociableOp(LHS, Opcode);
}
if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) {
RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap);
RHSBO = isReassociableOp(RHS, Opcode);
}
}
if (!LHSBO) {
if (!RHSBO) {
// Neither the LHS or RHS as part of the tree, thus this is a leaf. As
// such, just remember these operands and their rank.
Ops.push_back(ValueEntry(getRank(LHS), LHS));
Ops.push_back(ValueEntry(getRank(RHS), RHS));
// Clear the leaves out.
I->setOperand(0, UndefValue::get(I->getType()));
I->setOperand(1, UndefValue::get(I->getType()));
return;
}
// Turn X+(Y+Z) -> (Y+Z)+X
std::swap(LHSBO, RHSBO);
std::swap(LHS, RHS);
bool Success = !I->swapOperands();
assert(Success && "swapOperands failed");
(void)Success;
MadeChange = true;
} else if (RHSBO) {
// Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the RHS is not
// part of the expression tree.
LinearizeExpr(I);
LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0));
RHS = I->getOperand(1);
RHSBO = 0;
}
// Okay, now we know that the LHS is a nested expression and that the RHS is
// not. Perform reassociation.
assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!");
// Move LHS right before I to make sure that the tree expression dominates all
// values.
LHSBO->moveBefore(I);
// Linearize the expression tree on the LHS.
LinearizeExprTree(LHSBO, Ops);
// Remember the RHS operand and its rank.
Ops.push_back(ValueEntry(getRank(RHS), RHS));
// Clear the RHS leaf out.
I->setOperand(1, UndefValue::get(I->getType()));
}