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/Analysis/AliasAnalysis.h"
23 #include "llvm/Analysis/GlobalsModRef.h"
24 #include "llvm/IR/ConstantRange.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/IRBuilder.h"
27 #include "llvm/IR/InstIterator.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/Module.h"
30 #include "llvm/Pass.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/raw_ostream.h"
33 #include "llvm/Transforms/Scalar.h"
35 #include <functional> // For std::function
38 // The algorithm is simple. Start at instructions that convert from the
39 // float to the int domain: fptoui, fptosi and fcmp. Walk up the def-use
40 // graph, using an equivalence datastructure to unify graphs that interfere.
42 // Mappable instructions are those with an integer corrollary that, given
43 // integer domain inputs, produce an integer output; fadd, for example.
45 // If a non-mappable instruction is seen, this entire def-use graph is marked
46 // as non-transformable. If we see an instruction that converts from the
47 // integer domain to FP domain (uitofp,sitofp), we terminate our walk.
49 /// The largest integer type worth dealing with.
50 static cl::opt<unsigned>
51 MaxIntegerBW("float2int-max-integer-bw", cl::init(64), cl::Hidden,
52 cl::desc("Max integer bitwidth to consider in float2int"
56 struct Float2Int : public FunctionPass {
57 static char ID; // Pass identification, replacement for typeid
58 Float2Int() : FunctionPass(ID) {
59 initializeFloat2IntPass(*PassRegistry::getPassRegistry());
62 bool runOnFunction(Function &F) override;
63 void getAnalysisUsage(AnalysisUsage &AU) const override {
65 AU.addPreserved<GlobalsAAWrapperPass>();
68 void findRoots(Function &F, SmallPtrSet<Instruction*,8> &Roots);
69 ConstantRange seen(Instruction *I, ConstantRange R);
70 ConstantRange badRange();
71 ConstantRange unknownRange();
72 ConstantRange validateRange(ConstantRange R);
73 void walkBackwards(const SmallPtrSetImpl<Instruction*> &Roots);
75 bool validateAndTransform();
76 Value *convert(Instruction *I, Type *ToTy);
79 MapVector<Instruction*, ConstantRange > SeenInsts;
80 SmallPtrSet<Instruction*,8> Roots;
81 EquivalenceClasses<Instruction*> ECs;
82 MapVector<Instruction*, Value*> ConvertedInsts;
87 char Float2Int::ID = 0;
88 INITIALIZE_PASS_BEGIN(Float2Int, "float2int", "Float to int", false, false)
89 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
90 INITIALIZE_PASS_END(Float2Int, "float2int", "Float to int", false, false)
92 // Given a FCmp predicate, return a matching ICmp predicate if one
93 // exists, otherwise return BAD_ICMP_PREDICATE.
94 static CmpInst::Predicate mapFCmpPred(CmpInst::Predicate P) {
96 case CmpInst::FCMP_OEQ:
97 case CmpInst::FCMP_UEQ:
98 return CmpInst::ICMP_EQ;
99 case CmpInst::FCMP_OGT:
100 case CmpInst::FCMP_UGT:
101 return CmpInst::ICMP_SGT;
102 case CmpInst::FCMP_OGE:
103 case CmpInst::FCMP_UGE:
104 return CmpInst::ICMP_SGE;
105 case CmpInst::FCMP_OLT:
106 case CmpInst::FCMP_ULT:
107 return CmpInst::ICMP_SLT;
108 case CmpInst::FCMP_OLE:
109 case CmpInst::FCMP_ULE:
110 return CmpInst::ICMP_SLE;
111 case CmpInst::FCMP_ONE:
112 case CmpInst::FCMP_UNE:
113 return CmpInst::ICMP_NE;
115 return CmpInst::BAD_ICMP_PREDICATE;
119 // Given a floating point binary operator, return the matching
121 static Instruction::BinaryOps mapBinOpcode(unsigned Opcode) {
123 default: llvm_unreachable("Unhandled opcode!");
124 case Instruction::FAdd: return Instruction::Add;
125 case Instruction::FSub: return Instruction::Sub;
126 case Instruction::FMul: return Instruction::Mul;
130 // Find the roots - instructions that convert from the FP domain to
132 void Float2Int::findRoots(Function &F, SmallPtrSet<Instruction*,8> &Roots) {
133 for (auto &I : instructions(F)) {
134 switch (I.getOpcode()) {
136 case Instruction::FPToUI:
137 case Instruction::FPToSI:
140 case Instruction::FCmp:
141 if (mapFCmpPred(cast<CmpInst>(&I)->getPredicate()) !=
142 CmpInst::BAD_ICMP_PREDICATE)
149 // Helper - mark I as having been traversed, having range R.
150 ConstantRange Float2Int::seen(Instruction *I, ConstantRange R) {
151 DEBUG(dbgs() << "F2I: " << *I << ":" << R << "\n");
152 if (SeenInsts.find(I) != SeenInsts.end())
153 SeenInsts.find(I)->second = R;
155 SeenInsts.insert(std::make_pair(I, R));
159 // Helper - get a range representing a poison value.
160 ConstantRange Float2Int::badRange() {
161 return ConstantRange(MaxIntegerBW + 1, true);
163 ConstantRange Float2Int::unknownRange() {
164 return ConstantRange(MaxIntegerBW + 1, false);
166 ConstantRange Float2Int::validateRange(ConstantRange R) {
167 if (R.getBitWidth() > MaxIntegerBW + 1)
172 // The most obvious way to structure the search is a depth-first, eager
173 // search from each root. However, that require direct recursion and so
174 // can only handle small instruction sequences. Instead, we split the search
175 // up into two phases:
176 // - walkBackwards: A breadth-first walk of the use-def graph starting from
177 // the roots. Populate "SeenInsts" with interesting
178 // instructions and poison values if they're obvious and
179 // cheap to compute. Calculate the equivalance set structure
180 // while we're here too.
181 // - walkForwards: Iterate over SeenInsts in reverse order, so we visit
182 // defs before their uses. Calculate the real range info.
184 // Breadth-first walk of the use-def graph; determine the set of nodes
185 // we care about and eagerly determine if some of them are poisonous.
186 void Float2Int::walkBackwards(const SmallPtrSetImpl<Instruction*> &Roots) {
187 std::deque<Instruction*> Worklist(Roots.begin(), Roots.end());
188 while (!Worklist.empty()) {
189 Instruction *I = Worklist.back();
192 if (SeenInsts.find(I) != SeenInsts.end())
196 switch (I->getOpcode()) {
197 // FIXME: Handle select and phi nodes.
199 // Path terminated uncleanly.
203 case Instruction::UIToFP: {
204 // Path terminated cleanly.
205 unsigned BW = I->getOperand(0)->getType()->getPrimitiveSizeInBits();
206 APInt Min = APInt::getMinValue(BW).zextOrSelf(MaxIntegerBW+1);
207 APInt Max = APInt::getMaxValue(BW).zextOrSelf(MaxIntegerBW+1);
208 seen(I, validateRange(ConstantRange(Min, Max)));
212 case Instruction::SIToFP: {
213 // Path terminated cleanly.
214 unsigned BW = I->getOperand(0)->getType()->getPrimitiveSizeInBits();
215 APInt SMin = APInt::getSignedMinValue(BW).sextOrSelf(MaxIntegerBW+1);
216 APInt SMax = APInt::getSignedMaxValue(BW).sextOrSelf(MaxIntegerBW+1);
217 seen(I, validateRange(ConstantRange(SMin, SMax)));
221 case Instruction::FAdd:
222 case Instruction::FSub:
223 case Instruction::FMul:
224 case Instruction::FPToUI:
225 case Instruction::FPToSI:
226 case Instruction::FCmp:
227 seen(I, unknownRange());
231 for (Value *O : I->operands()) {
232 if (Instruction *OI = dyn_cast<Instruction>(O)) {
233 // Unify def-use chains if they interfere.
234 ECs.unionSets(I, OI);
235 if (SeenInsts.find(I)->second != badRange())
236 Worklist.push_back(OI);
237 } else if (!isa<ConstantFP>(O)) {
238 // Not an instruction or ConstantFP? we can't do anything.
245 // Walk forwards down the list of seen instructions, so we visit defs before
247 void Float2Int::walkForwards() {
248 for (auto &It : make_range(SeenInsts.rbegin(), SeenInsts.rend())) {
249 if (It.second != unknownRange())
252 Instruction *I = It.first;
253 std::function<ConstantRange(ArrayRef<ConstantRange>)> Op;
254 switch (I->getOpcode()) {
255 // FIXME: Handle select and phi nodes.
257 case Instruction::UIToFP:
258 case Instruction::SIToFP:
259 llvm_unreachable("Should have been handled in walkForwards!");
261 case Instruction::FAdd:
262 Op = [](ArrayRef<ConstantRange> Ops) {
263 assert(Ops.size() == 2 && "FAdd is a binary operator!");
264 return Ops[0].add(Ops[1]);
268 case Instruction::FSub:
269 Op = [](ArrayRef<ConstantRange> Ops) {
270 assert(Ops.size() == 2 && "FSub is a binary operator!");
271 return Ops[0].sub(Ops[1]);
275 case Instruction::FMul:
276 Op = [](ArrayRef<ConstantRange> Ops) {
277 assert(Ops.size() == 2 && "FMul is a binary operator!");
278 return Ops[0].multiply(Ops[1]);
283 // Root-only instructions - we'll only see these if they're the
284 // first node in a walk.
286 case Instruction::FPToUI:
287 case Instruction::FPToSI:
288 Op = [](ArrayRef<ConstantRange> Ops) {
289 assert(Ops.size() == 1 && "FPTo[US]I is a unary operator!");
294 case Instruction::FCmp:
295 Op = [](ArrayRef<ConstantRange> Ops) {
296 assert(Ops.size() == 2 && "FCmp is a binary operator!");
297 return Ops[0].unionWith(Ops[1]);
303 SmallVector<ConstantRange,4> OpRanges;
304 for (Value *O : I->operands()) {
305 if (Instruction *OI = dyn_cast<Instruction>(O)) {
306 assert(SeenInsts.find(OI) != SeenInsts.end() &&
307 "def not seen before use!");
308 OpRanges.push_back(SeenInsts.find(OI)->second);
309 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(O)) {
310 // Work out if the floating point number can be losslessly represented
312 // APFloat::convertToInteger(&Exact) purports to do what we want, but
313 // the exactness can be too precise. For example, negative zero can
314 // never be exactly converted to an integer.
316 // Instead, we ask APFloat to round itself to an integral value - this
317 // preserves sign-of-zero - then compare the result with the original.
319 APFloat F = CF->getValueAPF();
321 // First, weed out obviously incorrect values. Non-finite numbers
322 // can't be represented and neither can negative zero, unless
323 // we're in fast math mode.
325 (F.isZero() && F.isNegative() && isa<FPMathOperator>(I) &&
326 !I->hasNoSignedZeros())) {
333 auto Res = NewF.roundToIntegral(APFloat::rmNearestTiesToEven);
334 if (Res != APFloat::opOK || NewF.compare(F) != APFloat::cmpEqual) {
339 // OK, it's representable. Now get it.
340 APSInt Int(MaxIntegerBW+1, false);
342 CF->getValueAPF().convertToInteger(Int,
343 APFloat::rmNearestTiesToEven,
345 OpRanges.push_back(ConstantRange(Int));
347 llvm_unreachable("Should have already marked this as badRange!");
351 // Reduce the operands' ranges to a single range and return.
353 seen(I, Op(OpRanges));
357 // If there is a valid transform to be done, do it.
358 bool Float2Int::validateAndTransform() {
359 bool MadeChange = false;
361 // Iterate over every disjoint partition of the def-use graph.
362 for (auto It = ECs.begin(), E = ECs.end(); It != E; ++It) {
363 ConstantRange R(MaxIntegerBW + 1, false);
365 Type *ConvertedToTy = nullptr;
367 // For every member of the partition, union all the ranges together.
368 for (auto MI = ECs.member_begin(It), ME = ECs.member_end();
370 Instruction *I = *MI;
371 auto SeenI = SeenInsts.find(I);
372 if (SeenI == SeenInsts.end())
375 R = R.unionWith(SeenI->second);
376 // We need to ensure I has no users that have not been seen.
377 // If it does, transformation would be illegal.
379 // Don't count the roots, as they terminate the graphs.
380 if (Roots.count(I) == 0) {
381 // Set the type of the conversion while we're here.
383 ConvertedToTy = I->getType();
384 for (User *U : I->users()) {
385 Instruction *UI = dyn_cast<Instruction>(U);
386 if (!UI || SeenInsts.find(UI) == SeenInsts.end()) {
387 DEBUG(dbgs() << "F2I: Failing because of " << *U << "\n");
397 // If the set was empty, or we failed, or the range is poisonous,
399 if (ECs.member_begin(It) == ECs.member_end() || Fail ||
400 R.isFullSet() || R.isSignWrappedSet())
402 assert(ConvertedToTy && "Must have set the convertedtoty by this point!");
404 // The number of bits required is the maximum of the upper and
405 // lower limits, plus one so it can be signed.
406 unsigned MinBW = std::max(R.getLower().getMinSignedBits(),
407 R.getUpper().getMinSignedBits()) + 1;
408 DEBUG(dbgs() << "F2I: MinBitwidth=" << MinBW << ", R: " << R << "\n");
410 // If we've run off the realms of the exactly representable integers,
411 // the floating point result will differ from an integer approximation.
413 // Do we need more bits than are in the mantissa of the type we converted
414 // to? semanticsPrecision returns the number of mantissa bits plus one
416 unsigned MaxRepresentableBits
417 = APFloat::semanticsPrecision(ConvertedToTy->getFltSemantics()) - 1;
418 if (MinBW > MaxRepresentableBits) {
419 DEBUG(dbgs() << "F2I: Value not guaranteed to be representable!\n");
423 DEBUG(dbgs() << "F2I: Value requires more than 64 bits to represent!\n");
427 // OK, R is known to be representable. Now pick a type for it.
428 // FIXME: Pick the smallest legal type that will fit.
429 Type *Ty = (MinBW > 32) ? Type::getInt64Ty(*Ctx) : Type::getInt32Ty(*Ctx);
431 for (auto MI = ECs.member_begin(It), ME = ECs.member_end();
440 Value *Float2Int::convert(Instruction *I, Type *ToTy) {
441 if (ConvertedInsts.find(I) != ConvertedInsts.end())
442 // Already converted this instruction.
443 return ConvertedInsts[I];
445 SmallVector<Value*,4> NewOperands;
446 for (Value *V : I->operands()) {
447 // Don't recurse if we're an instruction that terminates the path.
448 if (I->getOpcode() == Instruction::UIToFP ||
449 I->getOpcode() == Instruction::SIToFP) {
450 NewOperands.push_back(V);
451 } else if (Instruction *VI = dyn_cast<Instruction>(V)) {
452 NewOperands.push_back(convert(VI, ToTy));
453 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
454 APSInt Val(ToTy->getPrimitiveSizeInBits(), /*IsUnsigned=*/false);
456 CF->getValueAPF().convertToInteger(Val,
457 APFloat::rmNearestTiesToEven,
459 NewOperands.push_back(ConstantInt::get(ToTy, Val));
461 llvm_unreachable("Unhandled operand type?");
465 // Now create a new instruction.
467 Value *NewV = nullptr;
468 switch (I->getOpcode()) {
469 default: llvm_unreachable("Unhandled instruction!");
471 case Instruction::FPToUI:
472 NewV = IRB.CreateZExtOrTrunc(NewOperands[0], I->getType());
475 case Instruction::FPToSI:
476 NewV = IRB.CreateSExtOrTrunc(NewOperands[0], I->getType());
479 case Instruction::FCmp: {
480 CmpInst::Predicate P = mapFCmpPred(cast<CmpInst>(I)->getPredicate());
481 assert(P != CmpInst::BAD_ICMP_PREDICATE && "Unhandled predicate!");
482 NewV = IRB.CreateICmp(P, NewOperands[0], NewOperands[1], I->getName());
486 case Instruction::UIToFP:
487 NewV = IRB.CreateZExtOrTrunc(NewOperands[0], ToTy);
490 case Instruction::SIToFP:
491 NewV = IRB.CreateSExtOrTrunc(NewOperands[0], ToTy);
494 case Instruction::FAdd:
495 case Instruction::FSub:
496 case Instruction::FMul:
497 NewV = IRB.CreateBinOp(mapBinOpcode(I->getOpcode()),
498 NewOperands[0], NewOperands[1],
503 // If we're a root instruction, RAUW.
505 I->replaceAllUsesWith(NewV);
507 ConvertedInsts[I] = NewV;
511 // Perform dead code elimination on the instructions we just modified.
512 void Float2Int::cleanup() {
513 for (auto &I : make_range(ConvertedInsts.rbegin(), ConvertedInsts.rend()))
514 I.first->eraseFromParent();
517 bool Float2Int::runOnFunction(Function &F) {
518 if (skipOptnoneFunction(F))
521 DEBUG(dbgs() << "F2I: Looking at function " << F.getName() << "\n");
522 // Clear out all state.
523 ECs = EquivalenceClasses<Instruction*>();
525 ConvertedInsts.clear();
528 Ctx = &F.getParent()->getContext();
532 walkBackwards(Roots);
535 bool Modified = validateAndTransform();
541 FunctionPass *llvm::createFloat2IntPass() { return new Float2Int(); }