xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/Float2Int.cpp (revision dd41de95a84d979615a2ef11df6850622bf6184e)
1 //===- Float2Int.cpp - Demote floating point ops to work on integers ------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the Float2Int pass, which aims to demote floating
10 // point operations to work on integers, where that is losslessly possible.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/InitializePasses.h"
15 #include "llvm/Support/CommandLine.h"
16 #define DEBUG_TYPE "float2int"
17 
18 #include "llvm/Transforms/Scalar/Float2Int.h"
19 #include "llvm/ADT/APInt.h"
20 #include "llvm/ADT/APSInt.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/Analysis/AliasAnalysis.h"
23 #include "llvm/Analysis/GlobalsModRef.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/IRBuilder.h"
26 #include "llvm/IR/InstIterator.h"
27 #include "llvm/IR/Instructions.h"
28 #include "llvm/IR/Module.h"
29 #include "llvm/Pass.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/raw_ostream.h"
32 #include "llvm/Transforms/Scalar.h"
33 #include <deque>
34 #include <functional> // For std::function
35 using namespace llvm;
36 
37 // The algorithm is simple. Start at instructions that convert from the
38 // float to the int domain: fptoui, fptosi and fcmp. Walk up the def-use
39 // graph, using an equivalence datastructure to unify graphs that interfere.
40 //
41 // Mappable instructions are those with an integer corrollary that, given
42 // integer domain inputs, produce an integer output; fadd, for example.
43 //
44 // If a non-mappable instruction is seen, this entire def-use graph is marked
45 // as non-transformable. If we see an instruction that converts from the
46 // integer domain to FP domain (uitofp,sitofp), we terminate our walk.
47 
48 /// The largest integer type worth dealing with.
49 static cl::opt<unsigned>
50 MaxIntegerBW("float2int-max-integer-bw", cl::init(64), cl::Hidden,
51              cl::desc("Max integer bitwidth to consider in float2int"
52                       "(default=64)"));
53 
54 namespace {
55   struct Float2IntLegacyPass : public FunctionPass {
56     static char ID; // Pass identification, replacement for typeid
57     Float2IntLegacyPass() : FunctionPass(ID) {
58       initializeFloat2IntLegacyPassPass(*PassRegistry::getPassRegistry());
59     }
60 
61     bool runOnFunction(Function &F) override {
62       if (skipFunction(F))
63         return false;
64 
65       const DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
66       return Impl.runImpl(F, DT);
67     }
68 
69     void getAnalysisUsage(AnalysisUsage &AU) const override {
70       AU.setPreservesCFG();
71       AU.addRequired<DominatorTreeWrapperPass>();
72       AU.addPreserved<GlobalsAAWrapperPass>();
73     }
74 
75   private:
76     Float2IntPass Impl;
77   };
78 }
79 
80 char Float2IntLegacyPass::ID = 0;
81 INITIALIZE_PASS(Float2IntLegacyPass, "float2int", "Float to int", false, false)
82 
83 // Given a FCmp predicate, return a matching ICmp predicate if one
84 // exists, otherwise return BAD_ICMP_PREDICATE.
85 static CmpInst::Predicate mapFCmpPred(CmpInst::Predicate P) {
86   switch (P) {
87   case CmpInst::FCMP_OEQ:
88   case CmpInst::FCMP_UEQ:
89     return CmpInst::ICMP_EQ;
90   case CmpInst::FCMP_OGT:
91   case CmpInst::FCMP_UGT:
92     return CmpInst::ICMP_SGT;
93   case CmpInst::FCMP_OGE:
94   case CmpInst::FCMP_UGE:
95     return CmpInst::ICMP_SGE;
96   case CmpInst::FCMP_OLT:
97   case CmpInst::FCMP_ULT:
98     return CmpInst::ICMP_SLT;
99   case CmpInst::FCMP_OLE:
100   case CmpInst::FCMP_ULE:
101     return CmpInst::ICMP_SLE;
102   case CmpInst::FCMP_ONE:
103   case CmpInst::FCMP_UNE:
104     return CmpInst::ICMP_NE;
105   default:
106     return CmpInst::BAD_ICMP_PREDICATE;
107   }
108 }
109 
110 // Given a floating point binary operator, return the matching
111 // integer version.
112 static Instruction::BinaryOps mapBinOpcode(unsigned Opcode) {
113   switch (Opcode) {
114   default: llvm_unreachable("Unhandled opcode!");
115   case Instruction::FAdd: return Instruction::Add;
116   case Instruction::FSub: return Instruction::Sub;
117   case Instruction::FMul: return Instruction::Mul;
118   }
119 }
120 
121 // Find the roots - instructions that convert from the FP domain to
122 // integer domain.
123 void Float2IntPass::findRoots(Function &F, const DominatorTree &DT) {
124   for (BasicBlock &BB : F) {
125     // Unreachable code can take on strange forms that we are not prepared to
126     // handle. For example, an instruction may have itself as an operand.
127     if (!DT.isReachableFromEntry(&BB))
128       continue;
129 
130     for (Instruction &I : BB) {
131       if (isa<VectorType>(I.getType()))
132         continue;
133       switch (I.getOpcode()) {
134       default: break;
135       case Instruction::FPToUI:
136       case Instruction::FPToSI:
137         Roots.insert(&I);
138         break;
139       case Instruction::FCmp:
140         if (mapFCmpPred(cast<CmpInst>(&I)->getPredicate()) !=
141             CmpInst::BAD_ICMP_PREDICATE)
142           Roots.insert(&I);
143         break;
144       }
145     }
146   }
147 }
148 
149 // Helper - mark I as having been traversed, having range R.
150 void Float2IntPass::seen(Instruction *I, ConstantRange R) {
151   LLVM_DEBUG(dbgs() << "F2I: " << *I << ":" << R << "\n");
152   auto IT = SeenInsts.find(I);
153   if (IT != SeenInsts.end())
154     IT->second = std::move(R);
155   else
156     SeenInsts.insert(std::make_pair(I, std::move(R)));
157 }
158 
159 // Helper - get a range representing a poison value.
160 ConstantRange Float2IntPass::badRange() {
161   return ConstantRange::getFull(MaxIntegerBW + 1);
162 }
163 ConstantRange Float2IntPass::unknownRange() {
164   return ConstantRange::getEmpty(MaxIntegerBW + 1);
165 }
166 ConstantRange Float2IntPass::validateRange(ConstantRange R) {
167   if (R.getBitWidth() > MaxIntegerBW + 1)
168     return badRange();
169   return R;
170 }
171 
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.
183 
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 Float2IntPass::walkBackwards() {
187   std::deque<Instruction*> Worklist(Roots.begin(), Roots.end());
188   while (!Worklist.empty()) {
189     Instruction *I = Worklist.back();
190     Worklist.pop_back();
191 
192     if (SeenInsts.find(I) != SeenInsts.end())
193       // Seen already.
194       continue;
195 
196     switch (I->getOpcode()) {
197       // FIXME: Handle select and phi nodes.
198     default:
199       // Path terminated uncleanly.
200       seen(I, badRange());
201       break;
202 
203     case Instruction::UIToFP:
204     case Instruction::SIToFP: {
205       // Path terminated cleanly - use the type of the integer input to seed
206       // the analysis.
207       unsigned BW = I->getOperand(0)->getType()->getPrimitiveSizeInBits();
208       auto Input = ConstantRange::getFull(BW);
209       auto CastOp = (Instruction::CastOps)I->getOpcode();
210       seen(I, validateRange(Input.castOp(CastOp, MaxIntegerBW+1)));
211       continue;
212     }
213 
214     case Instruction::FNeg:
215     case Instruction::FAdd:
216     case Instruction::FSub:
217     case Instruction::FMul:
218     case Instruction::FPToUI:
219     case Instruction::FPToSI:
220     case Instruction::FCmp:
221       seen(I, unknownRange());
222       break;
223     }
224 
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         if (SeenInsts.find(I)->second != badRange())
230           Worklist.push_back(OI);
231       } else if (!isa<ConstantFP>(O)) {
232         // Not an instruction or ConstantFP? we can't do anything.
233         seen(I, badRange());
234       }
235     }
236   }
237 }
238 
239 // Walk forwards down the list of seen instructions, so we visit defs before
240 // uses.
241 void Float2IntPass::walkForwards() {
242   for (auto &It : reverse(SeenInsts)) {
243     if (It.second != unknownRange())
244       continue;
245 
246     Instruction *I = It.first;
247     std::function<ConstantRange(ArrayRef<ConstantRange>)> Op;
248     switch (I->getOpcode()) {
249       // FIXME: Handle select and phi nodes.
250     default:
251     case Instruction::UIToFP:
252     case Instruction::SIToFP:
253       llvm_unreachable("Should have been handled in walkForwards!");
254 
255     case Instruction::FNeg:
256       Op = [](ArrayRef<ConstantRange> Ops) {
257         assert(Ops.size() == 1 && "FNeg is a unary operator!");
258         unsigned Size = Ops[0].getBitWidth();
259         auto Zero = ConstantRange(APInt::getNullValue(Size));
260         return Zero.sub(Ops[0]);
261       };
262       break;
263 
264     case Instruction::FAdd:
265     case Instruction::FSub:
266     case Instruction::FMul:
267       Op = [I](ArrayRef<ConstantRange> Ops) {
268         assert(Ops.size() == 2 && "its a binary operator!");
269         auto BinOp = (Instruction::BinaryOps) I->getOpcode();
270         return Ops[0].binaryOp(BinOp, Ops[1]);
271       };
272       break;
273 
274     //
275     // Root-only instructions - we'll only see these if they're the
276     //                          first node in a walk.
277     //
278     case Instruction::FPToUI:
279     case Instruction::FPToSI:
280       Op = [I](ArrayRef<ConstantRange> Ops) {
281         assert(Ops.size() == 1 && "FPTo[US]I is a unary operator!");
282         // Note: We're ignoring the casts output size here as that's what the
283         // caller expects.
284         auto CastOp = (Instruction::CastOps)I->getOpcode();
285         return Ops[0].castOp(CastOp, MaxIntegerBW+1);
286       };
287       break;
288 
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]);
293       };
294       break;
295     }
296 
297     bool Abort = false;
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
306         // as an integer.
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.
310         //
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.
313         //
314         const APFloat &F = CF->getValueAPF();
315 
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.
319         if (!F.isFinite() ||
320             (F.isZero() && F.isNegative() && isa<FPMathOperator>(I) &&
321              !I->hasNoSignedZeros())) {
322           seen(I, badRange());
323           Abort = true;
324           break;
325         }
326 
327         APFloat NewF = F;
328         auto Res = NewF.roundToIntegral(APFloat::rmNearestTiesToEven);
329         if (Res != APFloat::opOK || NewF != F) {
330           seen(I, badRange());
331           Abort = true;
332           break;
333         }
334         // OK, it's representable. Now get it.
335         APSInt Int(MaxIntegerBW+1, false);
336         bool Exact;
337         CF->getValueAPF().convertToInteger(Int,
338                                            APFloat::rmNearestTiesToEven,
339                                            &Exact);
340         OpRanges.push_back(ConstantRange(Int));
341       } else {
342         llvm_unreachable("Should have already marked this as badRange!");
343       }
344     }
345 
346     // Reduce the operands' ranges to a single range and return.
347     if (!Abort)
348       seen(I, Op(OpRanges));
349   }
350 }
351 
352 // If there is a valid transform to be done, do it.
353 bool Float2IntPass::validateAndTransform() {
354   bool MadeChange = false;
355 
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);
359     bool Fail = false;
360     Type *ConvertedToTy = nullptr;
361 
362     // For every member of the partition, union all the ranges together.
363     for (auto MI = ECs.member_begin(It), ME = ECs.member_end();
364          MI != ME; ++MI) {
365       Instruction *I = *MI;
366       auto SeenI = SeenInsts.find(I);
367       if (SeenI == SeenInsts.end())
368         continue;
369 
370       R = R.unionWith(SeenI->second);
371       // We need to ensure I has no users that have not been seen.
372       // If it does, transformation would be illegal.
373       //
374       // Don't count the roots, as they terminate the graphs.
375       if (Roots.count(I) == 0) {
376         // Set the type of the conversion while we're here.
377         if (!ConvertedToTy)
378           ConvertedToTy = I->getType();
379         for (User *U : I->users()) {
380           Instruction *UI = dyn_cast<Instruction>(U);
381           if (!UI || SeenInsts.find(UI) == SeenInsts.end()) {
382             LLVM_DEBUG(dbgs() << "F2I: Failing because of " << *U << "\n");
383             Fail = true;
384             break;
385           }
386         }
387       }
388       if (Fail)
389         break;
390     }
391 
392     // If the set was empty, or we failed, or the range is poisonous,
393     // bail out.
394     if (ECs.member_begin(It) == ECs.member_end() || Fail ||
395         R.isFullSet() || R.isSignWrappedSet())
396       continue;
397     assert(ConvertedToTy && "Must have set the convertedtoty by this point!");
398 
399     // The number of bits required is the maximum of the upper and
400     // lower limits, plus one so it can be signed.
401     unsigned MinBW = std::max(R.getLower().getMinSignedBits(),
402                               R.getUpper().getMinSignedBits()) + 1;
403     LLVM_DEBUG(dbgs() << "F2I: MinBitwidth=" << MinBW << ", R: " << R << "\n");
404 
405     // If we've run off the realms of the exactly representable integers,
406     // the floating point result will differ from an integer approximation.
407 
408     // Do we need more bits than are in the mantissa of the type we converted
409     // to? semanticsPrecision returns the number of mantissa bits plus one
410     // for the sign bit.
411     unsigned MaxRepresentableBits
412       = APFloat::semanticsPrecision(ConvertedToTy->getFltSemantics()) - 1;
413     if (MinBW > MaxRepresentableBits) {
414       LLVM_DEBUG(dbgs() << "F2I: Value not guaranteed to be representable!\n");
415       continue;
416     }
417     if (MinBW > 64) {
418       LLVM_DEBUG(
419           dbgs() << "F2I: Value requires more than 64 bits to represent!\n");
420       continue;
421     }
422 
423     // OK, R is known to be representable. Now pick a type for it.
424     // FIXME: Pick the smallest legal type that will fit.
425     Type *Ty = (MinBW > 32) ? Type::getInt64Ty(*Ctx) : Type::getInt32Ty(*Ctx);
426 
427     for (auto MI = ECs.member_begin(It), ME = ECs.member_end();
428          MI != ME; ++MI)
429       convert(*MI, Ty);
430     MadeChange = true;
431   }
432 
433   return MadeChange;
434 }
435 
436 Value *Float2IntPass::convert(Instruction *I, Type *ToTy) {
437   if (ConvertedInsts.find(I) != ConvertedInsts.end())
438     // Already converted this instruction.
439     return ConvertedInsts[I];
440 
441   SmallVector<Value*,4> NewOperands;
442   for (Value *V : I->operands()) {
443     // Don't recurse if we're an instruction that terminates the path.
444     if (I->getOpcode() == Instruction::UIToFP ||
445         I->getOpcode() == Instruction::SIToFP) {
446       NewOperands.push_back(V);
447     } else if (Instruction *VI = dyn_cast<Instruction>(V)) {
448       NewOperands.push_back(convert(VI, ToTy));
449     } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
450       APSInt Val(ToTy->getPrimitiveSizeInBits(), /*isUnsigned=*/false);
451       bool Exact;
452       CF->getValueAPF().convertToInteger(Val,
453                                          APFloat::rmNearestTiesToEven,
454                                          &Exact);
455       NewOperands.push_back(ConstantInt::get(ToTy, Val));
456     } else {
457       llvm_unreachable("Unhandled operand type?");
458     }
459   }
460 
461   // Now create a new instruction.
462   IRBuilder<> IRB(I);
463   Value *NewV = nullptr;
464   switch (I->getOpcode()) {
465   default: llvm_unreachable("Unhandled instruction!");
466 
467   case Instruction::FPToUI:
468     NewV = IRB.CreateZExtOrTrunc(NewOperands[0], I->getType());
469     break;
470 
471   case Instruction::FPToSI:
472     NewV = IRB.CreateSExtOrTrunc(NewOperands[0], I->getType());
473     break;
474 
475   case Instruction::FCmp: {
476     CmpInst::Predicate P = mapFCmpPred(cast<CmpInst>(I)->getPredicate());
477     assert(P != CmpInst::BAD_ICMP_PREDICATE && "Unhandled predicate!");
478     NewV = IRB.CreateICmp(P, NewOperands[0], NewOperands[1], I->getName());
479     break;
480   }
481 
482   case Instruction::UIToFP:
483     NewV = IRB.CreateZExtOrTrunc(NewOperands[0], ToTy);
484     break;
485 
486   case Instruction::SIToFP:
487     NewV = IRB.CreateSExtOrTrunc(NewOperands[0], ToTy);
488     break;
489 
490   case Instruction::FNeg:
491     NewV = IRB.CreateNeg(NewOperands[0], I->getName());
492     break;
493 
494   case Instruction::FAdd:
495   case Instruction::FSub:
496   case Instruction::FMul:
497     NewV = IRB.CreateBinOp(mapBinOpcode(I->getOpcode()),
498                            NewOperands[0], NewOperands[1],
499                            I->getName());
500     break;
501   }
502 
503   // If we're a root instruction, RAUW.
504   if (Roots.count(I))
505     I->replaceAllUsesWith(NewV);
506 
507   ConvertedInsts[I] = NewV;
508   return NewV;
509 }
510 
511 // Perform dead code elimination on the instructions we just modified.
512 void Float2IntPass::cleanup() {
513   for (auto &I : reverse(ConvertedInsts))
514     I.first->eraseFromParent();
515 }
516 
517 bool Float2IntPass::runImpl(Function &F, const DominatorTree &DT) {
518   LLVM_DEBUG(dbgs() << "F2I: Looking at function " << F.getName() << "\n");
519   // Clear out all state.
520   ECs = EquivalenceClasses<Instruction*>();
521   SeenInsts.clear();
522   ConvertedInsts.clear();
523   Roots.clear();
524 
525   Ctx = &F.getParent()->getContext();
526 
527   findRoots(F, DT);
528 
529   walkBackwards();
530   walkForwards();
531 
532   bool Modified = validateAndTransform();
533   if (Modified)
534     cleanup();
535   return Modified;
536 }
537 
538 namespace llvm {
539 FunctionPass *createFloat2IntPass() { return new Float2IntLegacyPass(); }
540 
541 PreservedAnalyses Float2IntPass::run(Function &F, FunctionAnalysisManager &AM) {
542   const DominatorTree &DT = AM.getResult<DominatorTreeAnalysis>(F);
543   if (!runImpl(F, DT))
544     return PreservedAnalyses::all();
545 
546   PreservedAnalyses PA;
547   PA.preserveSet<CFGAnalyses>();
548   PA.preserve<GlobalsAA>();
549   return PA;
550 }
551 } // End namespace llvm
552