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/DenseMap.h"
17 #include "llvm/ADT/EquivalenceClasses.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/APInt.h"
20 #include "llvm/ADT/APSInt.h"
21 #include "llvm/ADT/MapVector.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/ConstantRange.h"
24 #include "llvm/IR/InstIterator.h"
25 #include "llvm/IR/Instructions.h"
26 #include "llvm/IR/IRBuilder.h"
27 #include "llvm/IR/Module.h"
28 #include "llvm/Pass.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Transforms/Scalar.h"
31 #include <functional> // For std::function
35 // The algorithm is simple. Start at instructions that convert from the
36 // float to the int domain: fptoui, fptosi and fcmp. Walk up the def-use
37 // graph, using an equivalence datastructure to unify graphs that interfere.
39 // Mappable instructions are those with an integer corrollary that, given
40 // integer domain inputs, produce an integer output; fadd, for example.
42 // If a non-mappable instruction is seen, this entire def-use graph is marked
43 // as non-transformable. If we see an instruction that converts from the
44 // integer domain to FP domain (uitofp,sitofp), we terminate our walk.
46 /// The largest integer type worth dealing with.
47 static cl::opt<unsigned>
48 MaxIntegerBW("float2int-max-integer-bw", cl::init(64), cl::Hidden,
49 cl::desc("Max integer bitwidth to consider in float2int"
53 struct Float2Int : public FunctionPass {
54 static char ID; // Pass identification, replacement for typeid
55 Float2Int() : FunctionPass(ID) {
56 initializeFloat2IntPass(*PassRegistry::getPassRegistry());
59 bool runOnFunction(Function &F) override;
60 void getAnalysisUsage(AnalysisUsage &AU) const override {
64 void findRoots(Function &F, SmallPtrSet<Instruction*,8> &Roots);
65 ConstantRange seen(Instruction *I, ConstantRange R);
66 ConstantRange badRange();
67 ConstantRange unknownRange();
68 ConstantRange validateRange(ConstantRange R);
69 void walkBackwards(const SmallPtrSetImpl<Instruction*> &Roots);
71 bool validateAndTransform();
72 Value *convert(Instruction *I, Type *ToTy);
75 MapVector<Instruction*, ConstantRange > SeenInsts;
76 SmallPtrSet<Instruction*,8> Roots;
77 EquivalenceClasses<Instruction*> ECs;
78 MapVector<Instruction*, Value*> ConvertedInsts;
83 char Float2Int::ID = 0;
84 INITIALIZE_PASS(Float2Int, "float2int", "Float to int", false, false)
86 // Given a FCmp predicate, return a matching ICmp predicate if one
87 // exists, otherwise return BAD_ICMP_PREDICATE.
88 static CmpInst::Predicate mapFCmpPred(CmpInst::Predicate P) {
90 case CmpInst::FCMP_OEQ:
91 case CmpInst::FCMP_UEQ:
92 return CmpInst::ICMP_EQ;
93 case CmpInst::FCMP_OGT:
94 case CmpInst::FCMP_UGT:
95 return CmpInst::ICMP_SGT;
96 case CmpInst::FCMP_OGE:
97 case CmpInst::FCMP_UGE:
98 return CmpInst::ICMP_SGE;
99 case CmpInst::FCMP_OLT:
100 case CmpInst::FCMP_ULT:
101 return CmpInst::ICMP_SLT;
102 case CmpInst::FCMP_OLE:
103 case CmpInst::FCMP_ULE:
104 return CmpInst::ICMP_SLE;
105 case CmpInst::FCMP_ONE:
106 case CmpInst::FCMP_UNE:
107 return CmpInst::ICMP_NE;
109 return CmpInst::BAD_ICMP_PREDICATE;
113 // Given a floating point binary operator, return the matching
115 static Instruction::BinaryOps mapBinOpcode(unsigned Opcode) {
117 default: llvm_unreachable("Unhandled opcode!");
118 case Instruction::FAdd: return Instruction::Add;
119 case Instruction::FSub: return Instruction::Sub;
120 case Instruction::FMul: return Instruction::Mul;
124 // Find the roots - instructions that convert from the FP domain to
126 void Float2Int::findRoots(Function &F, SmallPtrSet<Instruction*,8> &Roots) {
127 for (auto &I : inst_range(F)) {
128 switch (I.getOpcode()) {
130 case Instruction::FPToUI:
131 case Instruction::FPToSI:
134 case Instruction::FCmp:
135 if (mapFCmpPred(cast<CmpInst>(&I)->getPredicate()) !=
136 CmpInst::BAD_ICMP_PREDICATE)
143 // Helper - mark I as having been traversed, having range R.
144 ConstantRange Float2Int::seen(Instruction *I, ConstantRange R) {
145 DEBUG(dbgs() << "F2I: " << *I << ":" << R << "\n");
146 if (SeenInsts.find(I) != SeenInsts.end())
147 SeenInsts.find(I)->second = R;
149 SeenInsts.insert(std::make_pair(I, R));
153 // Helper - get a range representing a poison value.
154 ConstantRange Float2Int::badRange() {
155 return ConstantRange(MaxIntegerBW + 1, true);
157 ConstantRange Float2Int::unknownRange() {
158 return ConstantRange(MaxIntegerBW + 1, false);
160 ConstantRange Float2Int::validateRange(ConstantRange R) {
161 if (R.getBitWidth() > MaxIntegerBW + 1)
166 // The most obvious way to structure the search is a depth-first, eager
167 // search from each root. However, that require direct recursion and so
168 // can only handle small instruction sequences. Instead, we split the search
169 // up into two phases:
170 // - walkBackwards: A breadth-first walk of the use-def graph starting from
171 // the roots. Populate "SeenInsts" with interesting
172 // instructions and poison values if they're obvious and
173 // cheap to compute. Calculate the equivalance set structure
174 // while we're here too.
175 // - walkForwards: Iterate over SeenInsts in reverse order, so we visit
176 // defs before their uses. Calculate the real range info.
178 // Breadth-first walk of the use-def graph; determine the set of nodes
179 // we care about and eagerly determine if some of them are poisonous.
180 void Float2Int::walkBackwards(const SmallPtrSetImpl<Instruction*> &Roots) {
181 std::deque<Instruction*> Worklist(Roots.begin(), Roots.end());
182 while (!Worklist.empty()) {
183 Instruction *I = Worklist.back();
186 if (SeenInsts.find(I) != SeenInsts.end())
190 switch (I->getOpcode()) {
191 // FIXME: Handle select and phi nodes.
193 // Path terminated uncleanly.
197 case Instruction::UIToFP: {
198 // Path terminated cleanly.
199 unsigned BW = I->getOperand(0)->getType()->getPrimitiveSizeInBits();
200 APInt Min = APInt::getMinValue(BW).zextOrSelf(MaxIntegerBW+1);
201 APInt Max = APInt::getMaxValue(BW).zextOrSelf(MaxIntegerBW+1);
202 seen(I, validateRange(ConstantRange(Min, Max)));
206 case Instruction::SIToFP: {
207 // Path terminated cleanly.
208 unsigned BW = I->getOperand(0)->getType()->getPrimitiveSizeInBits();
209 APInt SMin = APInt::getSignedMinValue(BW).sextOrSelf(MaxIntegerBW+1);
210 APInt SMax = APInt::getSignedMaxValue(BW).sextOrSelf(MaxIntegerBW+1);
211 seen(I, validateRange(ConstantRange(SMin, SMax)));
215 case Instruction::FAdd:
216 case Instruction::FSub:
217 case Instruction::FMul:
218 case Instruction::FPToUI:
219 case Instruction::FPToSI:
220 case Instruction::FCmp:
224 seen(I, unknownRange());
225 for (Value *O : I->operands()) {
226 if (Instruction *OI = dyn_cast<Instruction>(O)) {
227 // Unify def-use chains if they interfere.
228 ECs.unionSets(I, OI);
229 Worklist.push_back(OI);
230 } else if (!isa<ConstantFP>(O)) {
231 // Not an instruction or ConstantFP? we can't do anything.
239 // Walk forwards down the list of seen instructions, so we visit defs before
241 void Float2Int::walkForwards() {
242 for (auto It = SeenInsts.rbegin(), E = SeenInsts.rend(); It != E; ++It) {
243 if (It->second != unknownRange())
246 Instruction *I = It->first;
247 std::function<ConstantRange(ArrayRef<ConstantRange>)> Op;
248 switch (I->getOpcode()) {
249 // FIXME: Handle select and phi nodes.
251 case Instruction::UIToFP:
252 case Instruction::SIToFP:
253 llvm_unreachable("Should have been handled in walkForwards!");
255 case Instruction::FAdd:
256 Op = [](ArrayRef<ConstantRange> Ops) {
257 assert(Ops.size() == 2 && "FAdd is a binary operator!");
258 return Ops[0].add(Ops[1]);
262 case Instruction::FSub:
263 Op = [](ArrayRef<ConstantRange> Ops) {
264 assert(Ops.size() == 2 && "FSub is a binary operator!");
265 return Ops[0].sub(Ops[1]);
269 case Instruction::FMul:
270 Op = [](ArrayRef<ConstantRange> Ops) {
271 assert(Ops.size() == 2 && "FMul is a binary operator!");
272 return Ops[0].multiply(Ops[1]);
277 // Root-only instructions - we'll only see these if they're the
278 // first node in a walk.
280 case Instruction::FPToUI:
281 case Instruction::FPToSI:
282 Op = [](ArrayRef<ConstantRange> Ops) {
283 assert(Ops.size() == 1 && "FPTo[US]I is a unary operator!");
288 case Instruction::FCmp:
289 Op = [](ArrayRef<ConstantRange> Ops) {
290 assert(Ops.size() == 2 && "FCmp is a binary operator!");
291 return Ops[0].unionWith(Ops[1]);
297 SmallVector<ConstantRange,4> OpRanges;
298 for (Value *O : I->operands()) {
299 if (Instruction *OI = dyn_cast<Instruction>(O)) {
300 assert(SeenInsts.find(OI) != SeenInsts.end() &&
301 "def not seen before use!");
302 OpRanges.push_back(SeenInsts.find(OI)->second);
303 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(O)) {
304 // Work out if the floating point number can be losslessly represented
306 // APFloat::convertToInteger(&Exact) purports to do what we want, but
307 // the exactness can be too precise. For example, negative zero can
308 // never be exactly converted to an integer.
310 // Instead, we ask APFloat to round itself to an integral value - this
311 // preserves sign-of-zero - then compare the result with the original.
313 APFloat F = CF->getValueAPF();
315 // First, weed out obviously incorrect values. Non-finite numbers
316 // can't be represented and neither can negative zero, unless
317 // we're in fast math mode.
319 (F.isZero() && F.isNegative() && isa<FPMathOperator>(I) &&
320 !I->hasNoSignedZeros())) {
327 auto Res = NewF.roundToIntegral(APFloat::rmNearestTiesToEven);
328 if (Res != APFloat::opOK || NewF.compare(F) != APFloat::cmpEqual) {
333 // OK, it's representable. Now get it.
334 APSInt Int(MaxIntegerBW+1, false);
336 CF->getValueAPF().convertToInteger(Int,
337 APFloat::rmNearestTiesToEven,
339 OpRanges.push_back(ConstantRange(Int));
341 llvm_unreachable("Should have already marked this as badRange!");
345 // Reduce the operands' ranges to a single range and return.
347 seen(I, Op(OpRanges));
351 // If there is a valid transform to be done, do it.
352 bool Float2Int::validateAndTransform() {
353 bool MadeChange = false;
355 // Iterate over every disjoint partition of the def-use graph.
356 for (auto It = ECs.begin(), E = ECs.end(); It != E; ++It) {
357 ConstantRange R(MaxIntegerBW + 1, false);
359 Type *ConvertedToTy = nullptr;
361 // For every member of the partition, union all the ranges together.
362 for (auto MI = ECs.member_begin(It), ME = ECs.member_end();
364 Instruction *I = *MI;
365 auto SeenI = SeenInsts.find(I);
366 assert (SeenI != SeenInsts.end() && "Didn't see this instruction?");
368 R = R.unionWith(SeenI->second);
369 // We need to ensure I has no users that have not been seen.
370 // If it does, transformation would be illegal.
372 // Don't count the roots, as they terminate the graphs.
373 if (Roots.count(I) == 0) {
374 // Set the type of the conversion while we're here.
376 ConvertedToTy = I->getType();
377 for (User *U : I->users()) {
378 Instruction *UI = dyn_cast<Instruction>(U);
379 if (!UI || SeenInsts.find(UI) == SeenInsts.end()) {
380 DEBUG(dbgs() << "F2I: Failing because of " << *U << "\n");
390 // If the set was empty, or we failed, or the range is poisonous,
392 if (ECs.member_begin(It) == ECs.member_end() || Fail ||
393 R.isFullSet() || R.isSignWrappedSet())
395 assert(ConvertedToTy && "Must have set the convertedtoty by this point!");
397 // The number of bits required is the maximum of the upper and
398 // lower limits, plus one so it can be signed.
399 unsigned MinBW = std::max(R.getLower().getMinSignedBits(),
400 R.getUpper().getMinSignedBits()) + 1;
401 DEBUG(dbgs() << "F2I: MinBitwidth=" << MinBW << ", R: " << R << "\n");
403 // If we've run off the realms of the exactly representable integers,
404 // the floating point result will differ from an integer approximation.
406 // Do we need more bits than are in the mantissa of the type we converted
407 // to? semanticsPrecision returns the number of mantissa bits plus one
409 unsigned MaxRepresentableBits
410 = APFloat::semanticsPrecision(ConvertedToTy->getFltSemantics()) - 1;
411 if (MinBW > MaxRepresentableBits) {
412 DEBUG(dbgs() << "F2I: Value not guaranteed to be representable!\n");
416 DEBUG(dbgs() << "F2I: Value requires more than 64 bits to represent!\n");
420 // OK, R is known to be representable. Now pick a type for it.
421 // FIXME: Pick the smallest legal type that will fit.
422 Type *Ty = (MinBW > 32) ? Type::getInt64Ty(*Ctx) : Type::getInt32Ty(*Ctx);
424 for (auto MI = ECs.member_begin(It), ME = ECs.member_end();
433 Value *Float2Int::convert(Instruction *I, Type *ToTy) {
434 if (ConvertedInsts.find(I) != ConvertedInsts.end())
435 // Already converted this instruction.
436 return ConvertedInsts[I];
438 SmallVector<Value*,4> NewOperands;
439 for (Value *V : I->operands()) {
440 // Don't recurse if we're an instruction that terminates the path.
441 if (I->getOpcode() == Instruction::UIToFP ||
442 I->getOpcode() == Instruction::SIToFP) {
443 NewOperands.push_back(V);
444 } else if (Instruction *VI = dyn_cast<Instruction>(V)) {
445 NewOperands.push_back(convert(VI, ToTy));
446 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
447 APSInt Val(ToTy->getPrimitiveSizeInBits(), true);
449 CF->getValueAPF().convertToInteger(Val,
450 APFloat::rmNearestTiesToEven,
452 NewOperands.push_back(ConstantInt::get(ToTy, Val));
454 llvm_unreachable("Unhandled operand type?");
458 // Now create a new instruction.
460 Value *NewV = nullptr;
461 switch (I->getOpcode()) {
462 default: llvm_unreachable("Unhandled instruction!");
464 case Instruction::FPToUI:
465 NewV = IRB.CreateZExtOrTrunc(NewOperands[0], I->getType());
468 case Instruction::FPToSI:
469 NewV = IRB.CreateSExtOrTrunc(NewOperands[0], I->getType());
472 case Instruction::FCmp: {
473 CmpInst::Predicate P = mapFCmpPred(cast<CmpInst>(I)->getPredicate());
474 assert(P != CmpInst::BAD_ICMP_PREDICATE && "Unhandled predicate!");
475 NewV = IRB.CreateICmp(P, NewOperands[0], NewOperands[1], I->getName());
479 case Instruction::UIToFP:
480 NewV = IRB.CreateZExtOrTrunc(NewOperands[0], ToTy);
483 case Instruction::SIToFP:
484 NewV = IRB.CreateSExtOrTrunc(NewOperands[0], ToTy);
487 case Instruction::FAdd:
488 case Instruction::FSub:
489 case Instruction::FMul:
490 NewV = IRB.CreateBinOp(mapBinOpcode(I->getOpcode()),
491 NewOperands[0], NewOperands[1],
496 // If we're a root instruction, RAUW.
498 I->replaceAllUsesWith(NewV);
500 ConvertedInsts[I] = NewV;
504 // Perform dead code elimination on the instructions we just modified.
505 void Float2Int::cleanup() {
506 for (auto I = ConvertedInsts.rbegin(), E = ConvertedInsts.rend();
508 I->first->eraseFromParent();
511 bool Float2Int::runOnFunction(Function &F) {
512 DEBUG(dbgs() << "F2I: Looking at function " << F.getName() << "\n");
513 // Clear out all state.
514 ECs = EquivalenceClasses<Instruction*>();
516 ConvertedInsts.clear();
519 Ctx = &F.getParent()->getContext();
523 walkBackwards(Roots);
526 bool Modified = validateAndTransform();
532 FunctionPass *llvm::createFloat2IntPass() {
533 return new Float2Int();