1 //===- Float2Int.cpp - Demote floating point ops to work on integers ------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the Float2Int pass, which aims to demote floating
11 // point operations to work on integers, where that is losslessly possible.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "float2int"
16 #include "llvm/ADT/APInt.h"
17 #include "llvm/ADT/APSInt.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/EquivalenceClasses.h"
20 #include "llvm/ADT/MapVector.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/IR/ConstantRange.h"
23 #include "llvm/IR/Constants.h"
24 #include "llvm/IR/IRBuilder.h"
25 #include "llvm/IR/InstIterator.h"
26 #include "llvm/IR/Instructions.h"
27 #include "llvm/IR/Module.h"
28 #include "llvm/Pass.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include "llvm/Transforms/Scalar.h"
33 #include <functional> // For std::function
36 // The algorithm is simple. Start at instructions that convert from the
37 // float to the int domain: fptoui, fptosi and fcmp. Walk up the def-use
38 // graph, using an equivalence datastructure to unify graphs that interfere.
40 // Mappable instructions are those with an integer corrollary that, given
41 // integer domain inputs, produce an integer output; fadd, for example.
43 // If a non-mappable instruction is seen, this entire def-use graph is marked
44 // as non-transformable. If we see an instruction that converts from the
45 // integer domain to FP domain (uitofp,sitofp), we terminate our walk.
47 /// The largest integer type worth dealing with.
48 static cl::opt<unsigned>
49 MaxIntegerBW("float2int-max-integer-bw", cl::init(64), cl::Hidden,
50 cl::desc("Max integer bitwidth to consider in float2int"
54 struct Float2Int : public FunctionPass {
55 static char ID; // Pass identification, replacement for typeid
56 Float2Int() : FunctionPass(ID) {
57 initializeFloat2IntPass(*PassRegistry::getPassRegistry());
60 bool runOnFunction(Function &F) override;
61 void getAnalysisUsage(AnalysisUsage &AU) const override {
65 void findRoots(Function &F, SmallPtrSet<Instruction*,8> &Roots);
66 ConstantRange seen(Instruction *I, ConstantRange R);
67 ConstantRange badRange();
68 ConstantRange unknownRange();
69 ConstantRange validateRange(ConstantRange R);
70 void walkBackwards(const SmallPtrSetImpl<Instruction*> &Roots);
72 bool validateAndTransform();
73 Value *convert(Instruction *I, Type *ToTy);
76 MapVector<Instruction*, ConstantRange > SeenInsts;
77 SmallPtrSet<Instruction*,8> Roots;
78 EquivalenceClasses<Instruction*> ECs;
79 MapVector<Instruction*, Value*> ConvertedInsts;
84 char Float2Int::ID = 0;
85 INITIALIZE_PASS(Float2Int, "float2int", "Float to int", false, false)
87 // Given a FCmp predicate, return a matching ICmp predicate if one
88 // exists, otherwise return BAD_ICMP_PREDICATE.
89 static CmpInst::Predicate mapFCmpPred(CmpInst::Predicate P) {
91 case CmpInst::FCMP_OEQ:
92 case CmpInst::FCMP_UEQ:
93 return CmpInst::ICMP_EQ;
94 case CmpInst::FCMP_OGT:
95 case CmpInst::FCMP_UGT:
96 return CmpInst::ICMP_SGT;
97 case CmpInst::FCMP_OGE:
98 case CmpInst::FCMP_UGE:
99 return CmpInst::ICMP_SGE;
100 case CmpInst::FCMP_OLT:
101 case CmpInst::FCMP_ULT:
102 return CmpInst::ICMP_SLT;
103 case CmpInst::FCMP_OLE:
104 case CmpInst::FCMP_ULE:
105 return CmpInst::ICMP_SLE;
106 case CmpInst::FCMP_ONE:
107 case CmpInst::FCMP_UNE:
108 return CmpInst::ICMP_NE;
110 return CmpInst::BAD_ICMP_PREDICATE;
114 // Given a floating point binary operator, return the matching
116 static Instruction::BinaryOps mapBinOpcode(unsigned Opcode) {
118 default: llvm_unreachable("Unhandled opcode!");
119 case Instruction::FAdd: return Instruction::Add;
120 case Instruction::FSub: return Instruction::Sub;
121 case Instruction::FMul: return Instruction::Mul;
125 // Find the roots - instructions that convert from the FP domain to
127 void Float2Int::findRoots(Function &F, SmallPtrSet<Instruction*,8> &Roots) {
128 for (auto &I : inst_range(F)) {
129 switch (I.getOpcode()) {
131 case Instruction::FPToUI:
132 case Instruction::FPToSI:
135 case Instruction::FCmp:
136 if (mapFCmpPred(cast<CmpInst>(&I)->getPredicate()) !=
137 CmpInst::BAD_ICMP_PREDICATE)
144 // Helper - mark I as having been traversed, having range R.
145 ConstantRange Float2Int::seen(Instruction *I, ConstantRange R) {
146 DEBUG(dbgs() << "F2I: " << *I << ":" << R << "\n");
147 if (SeenInsts.find(I) != SeenInsts.end())
148 SeenInsts.find(I)->second = R;
150 SeenInsts.insert(std::make_pair(I, R));
154 // Helper - get a range representing a poison value.
155 ConstantRange Float2Int::badRange() {
156 return ConstantRange(MaxIntegerBW + 1, true);
158 ConstantRange Float2Int::unknownRange() {
159 return ConstantRange(MaxIntegerBW + 1, false);
161 ConstantRange Float2Int::validateRange(ConstantRange R) {
162 if (R.getBitWidth() > MaxIntegerBW + 1)
167 // The most obvious way to structure the search is a depth-first, eager
168 // search from each root. However, that require direct recursion and so
169 // can only handle small instruction sequences. Instead, we split the search
170 // up into two phases:
171 // - walkBackwards: A breadth-first walk of the use-def graph starting from
172 // the roots. Populate "SeenInsts" with interesting
173 // instructions and poison values if they're obvious and
174 // cheap to compute. Calculate the equivalance set structure
175 // while we're here too.
176 // - walkForwards: Iterate over SeenInsts in reverse order, so we visit
177 // defs before their uses. Calculate the real range info.
179 // Breadth-first walk of the use-def graph; determine the set of nodes
180 // we care about and eagerly determine if some of them are poisonous.
181 void Float2Int::walkBackwards(const SmallPtrSetImpl<Instruction*> &Roots) {
182 std::deque<Instruction*> Worklist(Roots.begin(), Roots.end());
183 while (!Worklist.empty()) {
184 Instruction *I = Worklist.back();
187 if (SeenInsts.find(I) != SeenInsts.end())
191 switch (I->getOpcode()) {
192 // FIXME: Handle select and phi nodes.
194 // Path terminated uncleanly.
198 case Instruction::UIToFP: {
199 // Path terminated cleanly.
200 unsigned BW = I->getOperand(0)->getType()->getPrimitiveSizeInBits();
201 APInt Min = APInt::getMinValue(BW).zextOrSelf(MaxIntegerBW+1);
202 APInt Max = APInt::getMaxValue(BW).zextOrSelf(MaxIntegerBW+1);
203 seen(I, validateRange(ConstantRange(Min, Max)));
207 case Instruction::SIToFP: {
208 // Path terminated cleanly.
209 unsigned BW = I->getOperand(0)->getType()->getPrimitiveSizeInBits();
210 APInt SMin = APInt::getSignedMinValue(BW).sextOrSelf(MaxIntegerBW+1);
211 APInt SMax = APInt::getSignedMaxValue(BW).sextOrSelf(MaxIntegerBW+1);
212 seen(I, validateRange(ConstantRange(SMin, SMax)));
216 case Instruction::FAdd:
217 case Instruction::FSub:
218 case Instruction::FMul:
219 case Instruction::FPToUI:
220 case Instruction::FPToSI:
221 case Instruction::FCmp:
225 seen(I, unknownRange());
226 for (Value *O : I->operands()) {
227 if (Instruction *OI = dyn_cast<Instruction>(O)) {
228 // Unify def-use chains if they interfere.
229 ECs.unionSets(I, OI);
230 Worklist.push_back(OI);
231 } else if (!isa<ConstantFP>(O)) {
232 // Not an instruction or ConstantFP? we can't do anything.
240 // Walk forwards down the list of seen instructions, so we visit defs before
242 void Float2Int::walkForwards() {
243 for (auto It = SeenInsts.rbegin(), E = SeenInsts.rend(); It != E; ++It) {
244 if (It->second != unknownRange())
247 Instruction *I = It->first;
248 std::function<ConstantRange(ArrayRef<ConstantRange>)> Op;
249 switch (I->getOpcode()) {
250 // FIXME: Handle select and phi nodes.
252 case Instruction::UIToFP:
253 case Instruction::SIToFP:
254 llvm_unreachable("Should have been handled in walkForwards!");
256 case Instruction::FAdd:
257 Op = [](ArrayRef<ConstantRange> Ops) {
258 assert(Ops.size() == 2 && "FAdd is a binary operator!");
259 return Ops[0].add(Ops[1]);
263 case Instruction::FSub:
264 Op = [](ArrayRef<ConstantRange> Ops) {
265 assert(Ops.size() == 2 && "FSub is a binary operator!");
266 return Ops[0].sub(Ops[1]);
270 case Instruction::FMul:
271 Op = [](ArrayRef<ConstantRange> Ops) {
272 assert(Ops.size() == 2 && "FMul is a binary operator!");
273 return Ops[0].multiply(Ops[1]);
278 // Root-only instructions - we'll only see these if they're the
279 // first node in a walk.
281 case Instruction::FPToUI:
282 case Instruction::FPToSI:
283 Op = [](ArrayRef<ConstantRange> Ops) {
284 assert(Ops.size() == 1 && "FPTo[US]I is a unary operator!");
289 case Instruction::FCmp:
290 Op = [](ArrayRef<ConstantRange> Ops) {
291 assert(Ops.size() == 2 && "FCmp is a binary operator!");
292 return Ops[0].unionWith(Ops[1]);
298 SmallVector<ConstantRange,4> OpRanges;
299 for (Value *O : I->operands()) {
300 if (Instruction *OI = dyn_cast<Instruction>(O)) {
301 assert(SeenInsts.find(OI) != SeenInsts.end() &&
302 "def not seen before use!");
303 OpRanges.push_back(SeenInsts.find(OI)->second);
304 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(O)) {
305 // Work out if the floating point number can be losslessly represented
307 // APFloat::convertToInteger(&Exact) purports to do what we want, but
308 // the exactness can be too precise. For example, negative zero can
309 // never be exactly converted to an integer.
311 // Instead, we ask APFloat to round itself to an integral value - this
312 // preserves sign-of-zero - then compare the result with the original.
314 APFloat F = CF->getValueAPF();
316 // First, weed out obviously incorrect values. Non-finite numbers
317 // can't be represented and neither can negative zero, unless
318 // we're in fast math mode.
320 (F.isZero() && F.isNegative() && isa<FPMathOperator>(I) &&
321 !I->hasNoSignedZeros())) {
328 auto Res = NewF.roundToIntegral(APFloat::rmNearestTiesToEven);
329 if (Res != APFloat::opOK || NewF.compare(F) != APFloat::cmpEqual) {
334 // OK, it's representable. Now get it.
335 APSInt Int(MaxIntegerBW+1, false);
337 CF->getValueAPF().convertToInteger(Int,
338 APFloat::rmNearestTiesToEven,
340 OpRanges.push_back(ConstantRange(Int));
342 llvm_unreachable("Should have already marked this as badRange!");
346 // Reduce the operands' ranges to a single range and return.
348 seen(I, Op(OpRanges));
352 // If there is a valid transform to be done, do it.
353 bool Float2Int::validateAndTransform() {
354 bool MadeChange = false;
356 // Iterate over every disjoint partition of the def-use graph.
357 for (auto It = ECs.begin(), E = ECs.end(); It != E; ++It) {
358 ConstantRange R(MaxIntegerBW + 1, false);
360 Type *ConvertedToTy = nullptr;
362 // For every member of the partition, union all the ranges together.
363 for (auto MI = ECs.member_begin(It), ME = ECs.member_end();
365 Instruction *I = *MI;
366 auto SeenI = SeenInsts.find(I);
367 assert (SeenI != SeenInsts.end() && "Didn't see this instruction?");
369 R = R.unionWith(SeenI->second);
370 // We need to ensure I has no users that have not been seen.
371 // If it does, transformation would be illegal.
373 // Don't count the roots, as they terminate the graphs.
374 if (Roots.count(I) == 0) {
375 // Set the type of the conversion while we're here.
377 ConvertedToTy = I->getType();
378 for (User *U : I->users()) {
379 Instruction *UI = dyn_cast<Instruction>(U);
380 if (!UI || SeenInsts.find(UI) == SeenInsts.end()) {
381 DEBUG(dbgs() << "F2I: Failing because of " << *U << "\n");
391 // If the set was empty, or we failed, or the range is poisonous,
393 if (ECs.member_begin(It) == ECs.member_end() || Fail ||
394 R.isFullSet() || R.isSignWrappedSet())
396 assert(ConvertedToTy && "Must have set the convertedtoty by this point!");
398 // The number of bits required is the maximum of the upper and
399 // lower limits, plus one so it can be signed.
400 unsigned MinBW = std::max(R.getLower().getMinSignedBits(),
401 R.getUpper().getMinSignedBits()) + 1;
402 DEBUG(dbgs() << "F2I: MinBitwidth=" << MinBW << ", R: " << R << "\n");
404 // If we've run off the realms of the exactly representable integers,
405 // the floating point result will differ from an integer approximation.
407 // Do we need more bits than are in the mantissa of the type we converted
408 // to? semanticsPrecision returns the number of mantissa bits plus one
410 unsigned MaxRepresentableBits
411 = APFloat::semanticsPrecision(ConvertedToTy->getFltSemantics()) - 1;
412 if (MinBW > MaxRepresentableBits) {
413 DEBUG(dbgs() << "F2I: Value not guaranteed to be representable!\n");
417 DEBUG(dbgs() << "F2I: Value requires more than 64 bits to represent!\n");
421 // OK, R is known to be representable. Now pick a type for it.
422 // FIXME: Pick the smallest legal type that will fit.
423 Type *Ty = (MinBW > 32) ? Type::getInt64Ty(*Ctx) : Type::getInt32Ty(*Ctx);
425 for (auto MI = ECs.member_begin(It), ME = ECs.member_end();
434 Value *Float2Int::convert(Instruction *I, Type *ToTy) {
435 if (ConvertedInsts.find(I) != ConvertedInsts.end())
436 // Already converted this instruction.
437 return ConvertedInsts[I];
439 SmallVector<Value*,4> NewOperands;
440 for (Value *V : I->operands()) {
441 // Don't recurse if we're an instruction that terminates the path.
442 if (I->getOpcode() == Instruction::UIToFP ||
443 I->getOpcode() == Instruction::SIToFP) {
444 NewOperands.push_back(V);
445 } else if (Instruction *VI = dyn_cast<Instruction>(V)) {
446 NewOperands.push_back(convert(VI, ToTy));
447 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
448 APSInt Val(ToTy->getPrimitiveSizeInBits(), true);
450 CF->getValueAPF().convertToInteger(Val,
451 APFloat::rmNearestTiesToEven,
453 NewOperands.push_back(ConstantInt::get(ToTy, Val));
455 llvm_unreachable("Unhandled operand type?");
459 // Now create a new instruction.
461 Value *NewV = nullptr;
462 switch (I->getOpcode()) {
463 default: llvm_unreachable("Unhandled instruction!");
465 case Instruction::FPToUI:
466 NewV = IRB.CreateZExtOrTrunc(NewOperands[0], I->getType());
469 case Instruction::FPToSI:
470 NewV = IRB.CreateSExtOrTrunc(NewOperands[0], I->getType());
473 case Instruction::FCmp: {
474 CmpInst::Predicate P = mapFCmpPred(cast<CmpInst>(I)->getPredicate());
475 assert(P != CmpInst::BAD_ICMP_PREDICATE && "Unhandled predicate!");
476 NewV = IRB.CreateICmp(P, NewOperands[0], NewOperands[1], I->getName());
480 case Instruction::UIToFP:
481 NewV = IRB.CreateZExtOrTrunc(NewOperands[0], ToTy);
484 case Instruction::SIToFP:
485 NewV = IRB.CreateSExtOrTrunc(NewOperands[0], ToTy);
488 case Instruction::FAdd:
489 case Instruction::FSub:
490 case Instruction::FMul:
491 NewV = IRB.CreateBinOp(mapBinOpcode(I->getOpcode()),
492 NewOperands[0], NewOperands[1],
497 // If we're a root instruction, RAUW.
499 I->replaceAllUsesWith(NewV);
501 ConvertedInsts[I] = NewV;
505 // Perform dead code elimination on the instructions we just modified.
506 void Float2Int::cleanup() {
507 for (auto I = ConvertedInsts.rbegin(), E = ConvertedInsts.rend();
509 I->first->eraseFromParent();
512 bool Float2Int::runOnFunction(Function &F) {
513 DEBUG(dbgs() << "F2I: Looking at function " << F.getName() << "\n");
514 // Clear out all state.
515 ECs = EquivalenceClasses<Instruction*>();
517 ConvertedInsts.clear();
520 Ctx = &F.getParent()->getContext();
524 walkBackwards(Roots);
527 bool Modified = validateAndTransform();
533 FunctionPass *llvm::createFloat2IntPass() {
534 return new Float2Int();