1 //===- InductionVars.cpp - Induction Variable Cannonicalization code --------=//
3 // This file implements induction variable cannonicalization of loops.
5 // Specifically, after this executes, the following is true:
6 // - There is a single induction variable for each loop (at least loops that
7 // used to contain at least one induction variable)
8 // * This induction variable starts at 0 and steps by 1 per iteration
9 // * This induction variable is represented by the first PHI node in the
10 // Header block, allowing it to be found easily.
11 // - All other preexisting induction variables are adjusted to operate in
12 // terms of this primary induction variable
13 // - Induction variables with a step size of 0 have been eliminated.
15 // This code assumes the following is true to perform its full job:
16 // - The CFG has been simplified to not have multiple entrances into an
17 // interval header. Interval headers should only have two predecessors,
18 // one from inside of the loop and one from outside of the loop.
20 //===----------------------------------------------------------------------===//
22 #include "llvm/Optimizations/InductionVars.h"
23 #include "llvm/ConstPoolVals.h"
24 #include "llvm/Analysis/IntervalPartition.h"
25 #include "llvm/Assembly/Writer.h"
26 #include "llvm/SymbolTable.h"
27 #include "llvm/iOther.h"
28 #include "Support/STLExtras.h"
31 #include "llvm/Analysis/LoopDepth.h"
35 // isLoopInvariant - Return true if the specified value/basic block source is
36 // an interval invariant computation.
38 static bool isLoopInvariant(cfg::Interval *Int, Value *V) {
39 assert(isa<ConstPoolVal>(V) || isa<Instruction>(V) || isa<MethodArgument>(V));
41 if (!isa<Instruction>(V))
42 return true; // Constants and arguments are always loop invariant
44 BasicBlock *ValueBlock = cast<Instruction>(V)->getParent();
45 assert(ValueBlock && "Instruction not embedded in basic block!");
47 // For now, only consider values from outside of the interval, regardless of
48 // whether the expression could be lifted out of the loop by some LICM.
50 // TODO: invoke LICM library if we find out it would be useful.
52 return !Int->contains(ValueBlock);
56 // isLinearInductionVariableH - Return isLIV if the expression V is a linear
57 // expression defined in terms of loop invariant computations, and a single
58 // instance of the PHI node PN. Return isLIC if the expression V is a loop
59 // invariant computation. Return isNLIV if the expression is a negated linear
60 // induction variable. Return isOther if it is neither.
62 // Currently allowed operators are: ADD, SUB, NEG
63 // TODO: This should allow casts!
65 enum LIVType { isLIV, isLIC, isNLIV, isOther };
67 // neg - Negate the sign of a LIV expression.
68 inline LIVType neg(LIVType T) {
69 assert(T == isLIV || T == isNLIV && "Negate Only works on LIV expressions");
70 return T == isLIV ? isNLIV : isLIV;
73 static LIVType isLinearInductionVariableH(cfg::Interval *Int, Value *V,
75 if (V == PN) { return isLIV; } // PHI node references are (0+PHI)
76 if (isLoopInvariant(Int, V)) return isLIC;
78 // loop variant computations must be instructions!
79 Instruction *I = cast<Instruction>(V);
80 switch (I->getOpcode()) { // Handle each instruction seperately
81 case Instruction::Add:
82 case Instruction::Sub: {
83 Value *SubV1 = cast<BinaryOperator>(I)->getOperand(0);
84 Value *SubV2 = cast<BinaryOperator>(I)->getOperand(1);
85 LIVType SubLIVType1 = isLinearInductionVariableH(Int, SubV1, PN);
86 if (SubLIVType1 == isOther) return isOther; // Early bailout
87 LIVType SubLIVType2 = isLinearInductionVariableH(Int, SubV2, PN);
89 switch (SubLIVType2) {
90 case isOther: return isOther; // Unknown subexpression type
91 case isLIC: return SubLIVType1; // Constant offset, return type #1
94 // So now we know that we have a linear induction variable on the RHS of
95 // the ADD or SUB instruction. SubLIVType1 cannot be isOther, so it is
96 // either a Loop Invariant computation, or a LIV type.
97 if (SubLIVType1 == isLIC) {
98 // Loop invariant computation, we know this is a LIV then.
99 return (I->getOpcode() == Instruction::Add) ?
100 SubLIVType2 : neg(SubLIVType2);
103 // If the LHS is also a LIV Expression, we cannot add two LIVs together
104 if (I->getOpcode() == Instruction::Add) return isOther;
106 // We can only subtract two LIVs if they are the same type, which yields
107 // a LIC, because the LIVs cancel each other out.
108 return (SubLIVType1 == SubLIVType2) ? isLIC : isOther;
113 default: // Any other instruction is not a LINEAR induction var
118 // isLinearInductionVariable - Return true if the specified expression is a
119 // "linear induction variable", which is an expression involving a single
120 // instance of the PHI node and a loop invariant value that is added or
121 // subtracted to the PHI node. This is calculated by walking the SSA graph
123 static inline bool isLinearInductionVariable(cfg::Interval *Int, Value *V,
125 return isLinearInductionVariableH(Int, V, PN) == isLIV;
129 // isSimpleInductionVar - Return true iff the cannonical induction variable PN
130 // has an initializer of the constant value 0, and has a step size of constant
132 static inline bool isSimpleInductionVar(PHINode *PN) {
133 assert(PN->getNumIncomingValues() == 2 && "Must have cannonical PHI node!");
134 Value *Initializer = PN->getIncomingValue(0);
135 if (!isa<ConstPoolVal>(Initializer)) return false;
137 if (Initializer->getType()->isSigned()) { // Signed constant value...
138 if (((ConstPoolSInt*)Initializer)->getValue() != 0) return false;
139 } else if (Initializer->getType()->isUnsigned()) { // Unsigned constant value
140 if (((ConstPoolUInt*)Initializer)->getValue() != 0) return false;
142 return false; // Not signed or unsigned? Must be FP type or something
145 Value *StepExpr = PN->getIncomingValue(1);
146 if (!isa<Instruction>(StepExpr) ||
147 cast<Instruction>(StepExpr)->getOpcode() != Instruction::Add)
150 BinaryOperator *I = cast<BinaryOperator>(StepExpr);
151 assert(isa<PHINode>(I->getOperand(0)) &&
152 "PHI node should be first operand of ADD instruction!");
154 // Get the right hand side of the ADD node. See if it is a constant 1.
155 Value *StepSize = I->getOperand(1);
156 if (!isa<ConstPoolVal>(StepSize)) return false;
158 if (StepSize->getType()->isSigned()) { // Signed constant value...
159 if (((ConstPoolSInt*)StepSize)->getValue() != 1) return false;
160 } else if (StepSize->getType()->isUnsigned()) { // Unsigned constant value
161 if (((ConstPoolUInt*)StepSize)->getValue() != 1) return false;
163 return false; // Not signed or unsigned? Must be FP type or something
166 // At this point, we know the initializer is a constant value 0 and the step
167 // size is a constant value 1. This is our simple induction variable!
171 // InjectSimpleInductionVariable - Insert a cannonical induction variable into
172 // the interval header Header. This assumes that the flow graph is in
173 // simplified form (so we know that the header block has exactly 2 predecessors)
175 // TODO: This should inherit the largest type that is being used by the already
176 // present induction variables (instead of always using uint)
178 static PHINode *InjectSimpleInductionVariable(cfg::Interval *Int) {
179 string PHIName, AddName;
181 BasicBlock *Header = Int->getHeaderNode();
182 Method *M = Header->getParent();
184 if (M->hasSymbolTable()) {
185 // Only name the induction variable if the method isn't stripped.
186 PHIName = M->getSymbolTable()->getUniqueName(Type::UIntTy, "ind_var");
187 AddName = M->getSymbolTable()->getUniqueName(Type::UIntTy, "ind_var_next");
190 // Create the neccesary instructions...
191 PHINode *PN = new PHINode(Type::UIntTy, PHIName);
192 ConstPoolVal *One = ConstPoolUInt::get(Type::UIntTy, 1);
193 ConstPoolVal *Zero = ConstPoolUInt::get(Type::UIntTy, 0);
194 BinaryOperator *AddNode = BinaryOperator::create(Instruction::Add,
197 // Figure out which predecessors I have to play with... there should be
198 // exactly two... one of which is a loop predecessor, and one of which is not.
200 BasicBlock::pred_iterator PI = Header->pred_begin();
201 assert(PI != Header->pred_end() && "Header node should have 2 preds!");
202 BasicBlock *Pred1 = *PI; ++PI;
203 assert(PI != Header->pred_end() && "Header node should have 2 preds!");
204 BasicBlock *Pred2 = *PI;
205 assert(++PI == Header->pred_end() && "Header node should have 2 preds!");
207 // Make Pred1 be the loop entrance predecessor, Pred2 be the Loop predecessor
208 if (Int->contains(Pred1)) swap(Pred1, Pred2);
210 assert(!Int->contains(Pred1) && "Pred1 should be loop entrance!");
211 assert( Int->contains(Pred2) && "Pred2 should be looping edge!");
213 // Link the instructions into the PHI node...
214 PN->addIncoming(Zero, Pred1); // The initializer is first argument
215 PN->addIncoming(AddNode, Pred2); // The step size is second PHI argument
217 // Insert the PHI node into the Header of the loop. It shall be the first
218 // instruction, because the "Simple" Induction Variable must be first in the
221 BasicBlock::InstListType &IL = Header->getInstList();
224 // Insert the Add instruction as the first (non-phi) instruction in the
225 // header node's basic block.
226 BasicBlock::iterator I = IL.begin();
227 while (isa<PHINode>(*I)) ++I;
228 IL.insert(I, AddNode);
232 // ProcessInterval - This function is invoked once for each interval in the
233 // IntervalPartition of the program. It looks for auxilliary induction
234 // variables in loops. If it finds one, it:
235 // * Cannonicalizes the induction variable. This consists of:
236 // A. Making the first element of the PHI node be the loop invariant
237 // computation, and the second element be the linear induction portion.
238 // B. Changing the first element of the linear induction portion of the PHI
239 // node to be of the form ADD(PHI, <loop invariant expr>).
240 // * Add the induction variable PHI to a list of induction variables found.
242 // After this, a list of cannonical induction variables is known. This list
243 // is searched to see if there is an induction variable that counts from
244 // constant 0 with a step size of constant 1. If there is not one, one is
245 // injected into the loop. Thus a "simple" induction variable is always known
247 // One a simple induction variable is known, all other induction variables are
248 // modified to refer to the "simple" induction variable.
250 static bool ProcessInterval(cfg::Interval *Int) {
251 if (!Int->isLoop()) return false; // Not a loop? Ignore it!
253 vector<PHINode *> InductionVars;
255 BasicBlock *Header = Int->getHeaderNode();
256 // Loop over all of the PHI nodes in the interval header...
257 for (BasicBlock::iterator I = Header->begin(), E = Header->end();
258 I != E && isa<PHINode>(*I); ++I) {
259 PHINode *PN = cast<PHINode>(*I);
260 if (PN->getNumIncomingValues() != 2) { // These should be eliminated by now.
261 cerr << "Found interval header with more than 2 predecessors! Ignoring\n";
262 return false; // Todo, make an assertion.
265 // For this to be an induction variable, one of the arguments must be a
266 // loop invariant expression, and the other must be an expression involving
267 // the PHI node, along with possible additions and subtractions of loop
270 BasicBlock *BB1 = PN->getIncomingBlock(0);
271 Value *V1 = PN->getIncomingValue(0);
272 BasicBlock *BB2 = PN->getIncomingBlock(1);
273 Value *V2 = PN->getIncomingValue(1);
275 // Figure out which computation is loop invariant...
276 if (!isLoopInvariant(Int, V1)) {
277 // V1 is *not* loop invariant. Check to see if V2 is:
278 if (isLoopInvariant(Int, V2)) {
279 // They *are* loop invariant. Exchange BB1/BB2 and V1/V2 so that
280 // V1 is always the loop invariant computation.
281 swap(V1, V2); swap(BB1, BB2);
283 // Neither value is loop invariant. Must not be an induction variable.
284 // This case can happen if there is an unreachable loop in the CFG that
285 // has two tail loops in it that was not split by the cleanup phase
291 // At this point, we know that BB1/V1 are loop invariant. We don't know
292 // anything about BB2/V2. Check now to see if V2 is a linear induction
295 cerr << "Found loop invariant computation: " << V1 << endl;
297 if (!isLinearInductionVariable(Int, V2, PN))
298 continue; // No, it is not a linear ind var, ignore the PHI node.
299 cerr << "Found linear induction variable: " << V2;
301 // TODO: Cannonicalize V2
303 // Add this PHI node to the list of induction variables found...
304 InductionVars.push_back(PN);
307 // No induction variables found?
308 if (InductionVars.empty()) return false;
310 // Search to see if there is already a "simple" induction variable.
311 vector<PHINode*>::iterator It =
312 find_if(InductionVars.begin(), InductionVars.end(), isSimpleInductionVar);
314 PHINode *PrimaryIndVar;
316 // A simple induction variable was not found, inject one now...
317 if (It == InductionVars.end()) {
318 PrimaryIndVar = InjectSimpleInductionVariable(Int);
320 // Move the PHI node for this induction variable to the start of the PHI
321 // list in HeaderNode... we do not need to do this for the inserted case
322 // because the inserted node will always be placed at the beginning of
326 BasicBlock::iterator i =
327 find(Header->begin(), Header->end(), PrimaryIndVar);
328 assert(i != Header->end() &&
329 "How could Primary IndVar not be in the header!?!!?");
331 if (i != Header->begin())
332 iter_swap(i, Header->begin());
335 // Now we know that there is a simple induction variable PrimaryIndVar.
336 // Simplify all of the other induction variables to use this induction
337 // variable as their counter, and destroy the PHI nodes that correspond to
343 cerr << "Found Interval Header with indvars (primary indvar should be first "
344 << "phi): \n" << Header << "\nPrimaryIndVar: " << PrimaryIndVar;
346 return false; // TODO: true;
350 // ProcessIntervalPartition - This function loops over the interval partition
351 // processing each interval with ProcessInterval
353 static bool ProcessIntervalPartition(cfg::IntervalPartition &IP) {
354 // This currently just prints out information about the interval structure
357 static unsigned N = 0;
358 cerr << "\n***********Interval Partition #" << (++N) << "************\n\n";
359 copy(IP.begin(), IP.end(), ostream_iterator<cfg::Interval*>(cerr, "\n"));
361 cerr << "\n*********** PERFORMING WORK ************\n\n";
363 // Loop over all of the intervals in the partition and look for induction
364 // variables in intervals that represent loops.
366 return reduce_apply(IP.begin(), IP.end(), bitwise_or<bool>(), false,
367 ptr_fun(ProcessInterval));
370 // DoInductionVariableCannonicalize - Simplify induction variables in loops.
371 // This function loops over an interval partition of a program, reducing it
372 // until the graph is gone.
374 bool opt::InductionVariableCannonicalize::doIt(Method *M) {
376 if (0) { // Print basic blocks with their depth
377 LoopDepthCalculator LDC(M);
378 for (Method::iterator I = M->begin(); I != M->end(); ++I) {
379 cerr << "Basic Block Depth: " << LDC.getLoopDepth(*I) << *I;
384 cfg::IntervalPartition *IP = new cfg::IntervalPartition(M);
385 bool Changed = false;
387 while (!IP->isDegeneratePartition()) {
388 Changed |= ProcessIntervalPartition(*IP);
390 // Calculate the reduced version of this graph until we get to an
391 // irreducible graph or a degenerate graph...
393 cfg::IntervalPartition *NewIP = new cfg::IntervalPartition(*IP, false);
394 if (NewIP->size() == IP->size()) {
395 cerr << "IRREDUCIBLE GRAPH FOUND!!!\n";