136f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman//===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis --*- C++ -*-===// 236f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// 336f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// The LLVM Compiler Infrastructure 436f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// 54ee451de366474b9c228b4e5fa573795a715216dChris Lattner// This file is distributed under the University of Illinois Open Source 64ee451de366474b9c228b4e5fa573795a715216dChris Lattner// License. See LICENSE.TXT for details. 736f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// 836f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman//===----------------------------------------------------------------------===// 936f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// 1036f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// This file contains the implementation of the scalar evolution expander, 1136f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// which is used to generate the code corresponding to a given scalar evolution 1236f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// expression. 1336f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// 1436f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman//===----------------------------------------------------------------------===// 1536f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 1636f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman#include "llvm/Analysis/ScalarEvolutionExpander.h" 17d04a8d4b33ff316ca4cf961e06c9e312eff8e64fChandler Carruth#include "llvm/ADT/STLExtras.h" 1836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines#include "llvm/ADT/SmallSet.h" 19c6a4f5e819217e1e12c458aed8e7b122e23a3a58Stephen Hines#include "llvm/Analysis/InstructionSimplify.h" 20e81561909d128c6e2d8033cb5465a49b2596b26aBill Wendling#include "llvm/Analysis/LoopInfo.h" 21e4ba75f43e2ab1480d119d2d4eb878256274e0fbChandler Carruth#include "llvm/Analysis/TargetTransformInfo.h" 220b8c9a80f20772c3793201ab5b251d3520b9cea3Chandler Carruth#include "llvm/IR/DataLayout.h" 2336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines#include "llvm/IR/Dominators.h" 240b8c9a80f20772c3793201ab5b251d3520b9cea3Chandler Carruth#include "llvm/IR/IntrinsicInst.h" 250b8c9a80f20772c3793201ab5b251d3520b9cea3Chandler Carruth#include "llvm/IR/LLVMContext.h" 260c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar#include "llvm/IR/Module.h" 27cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar#include "llvm/IR/PatternMatch.h" 28c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick#include "llvm/Support/Debug.h" 294c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar#include "llvm/Support/raw_ostream.h" 30d152d03a476b8d0d4b26577db26e2ba76034b0f3Andrew Trick 3136f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begemanusing namespace llvm; 32cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainarusing namespace PatternMatch; 3336f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 3419e5ada58a8a07d3d6d67312550f6d0791d84c3cGabor Greif/// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP, 35485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman/// reusing an existing cast if a suitable one exists, moving an existing 36485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman/// cast if a suitable one exists but isn't in the right place, or 3719e5ada58a8a07d3d6d67312550f6d0791d84c3cGabor Greif/// creating a new one. 38db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris LattnerValue *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty, 39485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman Instruction::CastOps Op, 40485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman BasicBlock::iterator IP) { 41919a503458c18472984c7c37b305110c9ba3a782Rafael Espindola // This function must be called with the builder having a valid insertion 42919a503458c18472984c7c37b305110c9ba3a782Rafael Espindola // point. It doesn't need to be the actual IP where the uses of the returned 43919a503458c18472984c7c37b305110c9ba3a782Rafael Espindola // cast will be added, but it must dominate such IP. 4423b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola // We use this precondition to produce a cast that will dominate all its 4523b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola // uses. In particular, this is crucial for the case where the builder's 4623b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola // insertion point *is* the point where we were asked to put the cast. 47c8e41c591741b3da1077f7000274ad040bef8002Sylvestre Ledru // Since we don't know the builder's insertion point is actually 48919a503458c18472984c7c37b305110c9ba3a782Rafael Espindola // where the uses will be added (only that it dominates it), we are 49919a503458c18472984c7c37b305110c9ba3a782Rafael Espindola // not allowed to move it. 50919a503458c18472984c7c37b305110c9ba3a782Rafael Espindola BasicBlock::iterator BIP = Builder.GetInsertPoint(); 51919a503458c18472984c7c37b305110c9ba3a782Rafael Espindola 52dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines Instruction *Ret = nullptr; 53ef4c80e07baf02dd2a8f08db49c5634a06d3ca1eRafael Espindola 54485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // Check to see if there is already a cast! 5536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines for (User *U : V->users()) 56f64f9cf7ec39ae91bca84dc6ad3c8fc3343b358aGabor Greif if (U->getType() == Ty) 5719e5ada58a8a07d3d6d67312550f6d0791d84c3cGabor Greif if (CastInst *CI = dyn_cast<CastInst>(U)) 58485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman if (CI->getOpcode() == Op) { 59b84d540bc9d70de00b791d0e0deab410d6ba3ad5Rafael Espindola // If the cast isn't where we want it, create a new cast at IP. 60b84d540bc9d70de00b791d0e0deab410d6ba3ad5Rafael Espindola // Likewise, do not reuse a cast at BIP because it must dominate 61b84d540bc9d70de00b791d0e0deab410d6ba3ad5Rafael Espindola // instructions that might be inserted before BIP. 62919a503458c18472984c7c37b305110c9ba3a782Rafael Espindola if (BasicBlock::iterator(CI) != IP || BIP == IP) { 63485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // Create a new cast, and leave the old cast in place in case 64485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // it is being used as an insert point. Clear its operand 65485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // so that it doesn't hold anything live. 66cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Ret = CastInst::Create(Op, V, Ty, "", &*IP); 6723b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola Ret->takeName(CI); 6823b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola CI->replaceAllUsesWith(Ret); 69485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman CI->setOperand(0, UndefValue::get(V->getType())); 7023b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola break; 71485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman } 7223b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola Ret = CI; 7323b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola break; 74485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman } 75485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman 76485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // Create a new cast. 7723b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola if (!Ret) 78cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Ret = CastInst::Create(Op, V, Ty, V->getName(), &*IP); 7923b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola 8023b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola // We assert at the end of the function since IP might point to an 8123b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola // instruction with different dominance properties than a cast 8223b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola // (an invoke for example) and not dominate BIP (but the cast does). 83cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar assert(SE.DT.dominates(Ret, &*BIP)); 8423b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola 8523b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola rememberInstruction(Ret); 8623b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola return Ret; 87485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman} 88485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman 89cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainarstatic BasicBlock::iterator findInsertPointAfter(Instruction *I, 90cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar BasicBlock *MustDominate) { 91cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar BasicBlock::iterator IP = ++I->getIterator(); 92cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (auto *II = dyn_cast<InvokeInst>(I)) 93cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar IP = II->getNormalDest()->begin(); 94cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 95cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar while (isa<PHINode>(IP)) 96cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar ++IP; 97cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 98cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar while (IP->isEHPad()) { 99cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (isa<FuncletPadInst>(IP) || isa<LandingPadInst>(IP)) { 100cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar ++IP; 101cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar } else if (isa<CatchSwitchInst>(IP)) { 102cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar IP = MustDominate->getFirstInsertionPt(); 103cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar } else { 104cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar llvm_unreachable("unexpected eh pad!"); 105cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar } 106cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar } 107cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 108cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return IP; 109cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar} 110cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 111267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman/// InsertNoopCastOfTo - Insert a cast of V to the specified type, 112267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman/// which must be possible with a noop cast, doing what we can to share 113267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman/// the casts. 114db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris LattnerValue *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) { 115267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false); 116267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman assert((Op == Instruction::BitCast || 117267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Op == Instruction::PtrToInt || 118267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Op == Instruction::IntToPtr) && 119267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman "InsertNoopCastOfTo cannot perform non-noop casts!"); 120267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) && 121267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman "InsertNoopCastOfTo cannot change sizes!"); 122267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman 1232d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman // Short-circuit unnecessary bitcasts. 12419154f457696f286b4a597179fa07e3b38ea0310Andrew Trick if (Op == Instruction::BitCast) { 12519154f457696f286b4a597179fa07e3b38ea0310Andrew Trick if (V->getType() == Ty) 12619154f457696f286b4a597179fa07e3b38ea0310Andrew Trick return V; 12719154f457696f286b4a597179fa07e3b38ea0310Andrew Trick if (CastInst *CI = dyn_cast<CastInst>(V)) { 12819154f457696f286b4a597179fa07e3b38ea0310Andrew Trick if (CI->getOperand(0)->getType() == Ty) 12919154f457696f286b4a597179fa07e3b38ea0310Andrew Trick return CI->getOperand(0); 13019154f457696f286b4a597179fa07e3b38ea0310Andrew Trick } 13119154f457696f286b4a597179fa07e3b38ea0310Andrew Trick } 132f04fa483b83227c570bc58e1684ea096430a6697Dan Gohman // Short-circuit unnecessary inttoptr<->ptrtoint casts. 133267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) && 13480dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) { 135af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman if (CastInst *CI = dyn_cast<CastInst>(V)) 136af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman if ((CI->getOpcode() == Instruction::PtrToInt || 137af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman CI->getOpcode() == Instruction::IntToPtr) && 138af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman SE.getTypeSizeInBits(CI->getType()) == 139af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman SE.getTypeSizeInBits(CI->getOperand(0)->getType())) 140af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman return CI->getOperand(0); 14180dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 14280dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman if ((CE->getOpcode() == Instruction::PtrToInt || 14380dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman CE->getOpcode() == Instruction::IntToPtr) && 14480dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman SE.getTypeSizeInBits(CE->getType()) == 14580dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman SE.getTypeSizeInBits(CE->getOperand(0)->getType())) 14680dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman return CE->getOperand(0); 14780dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman } 148f04fa483b83227c570bc58e1684ea096430a6697Dan Gohman 149485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // Fold a cast of a constant. 150ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner if (Constant *C = dyn_cast<Constant>(V)) 151baf3c404409d5e47b13984a7f95bfbd6d1f2e79eOwen Anderson return ConstantExpr::getCast(Op, C, Ty); 1524c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman 153485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // Cast the argument at the beginning of the entry block, after 154485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // any bitcasts of other arguments. 155ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner if (Argument *A = dyn_cast<Argument>(V)) { 156485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin(); 157485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman while ((isa<BitCastInst>(IP) && 158485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) && 159485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman cast<BitCastInst>(IP)->getOperand(0) != A) || 160cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar isa<DbgInfoIntrinsic>(IP)) 161485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman ++IP; 162485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman return ReuseOrCreateCast(A, Ty, Op, IP); 163ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner } 1643913187bf6135e6a71260c65eed664095c8f9ce9Wojciech Matyjewicz 165485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // Cast the instruction immediately after the instruction. 166ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner Instruction *I = cast<Instruction>(V); 167cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar BasicBlock::iterator IP = findInsertPointAfter(I, Builder.GetInsertBlock()); 168485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman return ReuseOrCreateCast(I, Ty, Op, IP); 169ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner} 170ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner 1717fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner/// InsertBinop - Insert the specified binary operator, doing a small amount 1727fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner/// of work to avoid inserting an obviously redundant operation. 173267a385342f2e7388f178b327dd87c5f29afd51bDan GohmanValue *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, 174267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Value *LHS, Value *RHS) { 1750f0eb18addc94bf3d8179fbce162c8c4622b9866Dan Gohman // Fold a binop with constant operands. 1760f0eb18addc94bf3d8179fbce162c8c4622b9866Dan Gohman if (Constant *CLHS = dyn_cast<Constant>(LHS)) 1770f0eb18addc94bf3d8179fbce162c8c4622b9866Dan Gohman if (Constant *CRHS = dyn_cast<Constant>(RHS)) 178baf3c404409d5e47b13984a7f95bfbd6d1f2e79eOwen Anderson return ConstantExpr::get(Opcode, CLHS, CRHS); 1790f0eb18addc94bf3d8179fbce162c8c4622b9866Dan Gohman 1807fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner // Do a quick scan to see if we have this binop nearby. If so, reuse it. 1817fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner unsigned ScanLimit = 6; 182267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 183267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman // Scanning starts from the last instruction before the insertion point. 184267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock::iterator IP = Builder.GetInsertPoint(); 185267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman if (IP != BlockBegin) { 1868a08769bad43a22fae2845bb0ba0fd1266cd55c8Wojciech Matyjewicz --IP; 1878a08769bad43a22fae2845bb0ba0fd1266cd55c8Wojciech Matyjewicz for (; ScanLimit; --IP, --ScanLimit) { 1888d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen // Don't count dbg.value against the ScanLimit, to avoid perturbing the 1898d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen // generated code. 1908d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen if (isa<DbgInfoIntrinsic>(IP)) 1918d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen ScanLimit++; 1925be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS && 1935be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman IP->getOperand(1) == RHS) 194cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return &*IP; 1958a08769bad43a22fae2845bb0ba0fd1266cd55c8Wojciech Matyjewicz if (IP == BlockBegin) break; 1968a08769bad43a22fae2845bb0ba0fd1266cd55c8Wojciech Matyjewicz } 1977fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner } 198267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman 199087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Save the original insertion point so we can restore it when we're done. 200d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc(); 201d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer BuilderType::InsertPointGuard Guard(Builder); 202087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 203087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Move the insertion point out of as many loops as we can. 204cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) { 205087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break; 206087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman BasicBlock *Preheader = L->getLoopPreheader(); 207087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!Preheader) break; 208087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 209087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Ok, move up a level. 210cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Builder.SetInsertPoint(Preheader->getTerminator()); 211087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 212087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 2138a08769bad43a22fae2845bb0ba0fd1266cd55c8Wojciech Matyjewicz // If we haven't found this binop, insert it. 214a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS)); 215d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer BO->setDebugLoc(Loc); 216a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(BO); 217087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 218cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman return BO; 2197fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner} 2207fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner 2214a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman/// FactorOutConstant - Test if S is divisible by Factor, using signed 222453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// division. If so, update S with Factor divided out and return true. 2233f46a3abeedba8d517b4182de34c821d752db058Dan Gohman/// S need not be evenly divisible if a reasonable remainder can be 2244a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman/// computed. 225453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made 226453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// unnecessary; in its place, just signed-divide Ops[i] by the scale and 227453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// check to see if the divide was folded. 2284c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainarstatic bool FactorOutConstant(const SCEV *&S, const SCEV *&Remainder, 2294c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar const SCEV *Factor, ScalarEvolution &SE, 2304c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar const DataLayout &DL) { 231453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // Everything is divisible by one. 232c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (Factor->isOne()) 233c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman return true; 234c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 235c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // x/x == 1. 236c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (S == Factor) { 237deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman S = SE.getConstant(S->getType(), 1); 238453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman return true; 239c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 240453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 241453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // For a Constant, check for a multiple of the given factor. 2424a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 243c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // 0/x == 0. 244c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (C->isZero()) 245453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman return true; 246c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Check for divisibility. 247c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) { 248c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ConstantInt *CI = 249cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar ConstantInt::get(SE.getContext(), C->getAPInt().sdiv(FC->getAPInt())); 250c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // If the quotient is zero and the remainder is non-zero, reject 251c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // the value at this scale. It will be considered for subsequent 252c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // smaller scales. 253c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (!CI->isZero()) { 254c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *Div = SE.getConstant(CI); 255c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman S = Div; 256cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Remainder = SE.getAddExpr( 257cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Remainder, SE.getConstant(C->getAPInt().srem(FC->getAPInt()))); 258c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman return true; 259c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 260453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 2614a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman } 262453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 263453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // In a Mul, check if there is a constant operand which is a multiple 264453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // of the given factor. 265c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) { 2664c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar // Size is known, check if there is a constant operand which is a multiple 2674c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar // of the given factor. If so, we can factor it. 2684c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar const SCEVConstant *FC = cast<SCEVConstant>(Factor); 2694c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0))) 270cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (!C->getAPInt().srem(FC->getAPInt())) { 2714c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 272cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar NewMulOps[0] = SE.getConstant(C->getAPInt().sdiv(FC->getAPInt())); 2734c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar S = SE.getMulExpr(NewMulOps); 2744c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar return true; 275453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 276c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 277453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 278453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // In an AddRec, check if both start and step are divisible. 279453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) { 2800bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *Step = A->getStepRecurrence(SE); 281deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman const SCEV *StepRem = SE.getConstant(Step->getType(), 0); 28236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (!FactorOutConstant(Step, StepRem, Factor, SE, DL)) 2834a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman return false; 2844a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman if (!StepRem->isZero()) 2854a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman return false; 2860bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *Start = A->getStart(); 28736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (!FactorOutConstant(Start, Remainder, Factor, SE, DL)) 288453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman return false; 2896f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick S = SE.getAddRecExpr(Start, Step, A->getLoop(), 2906f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick A->getNoWrapFlags(SCEV::FlagNW)); 291453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman return true; 292453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 293453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 294453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman return false; 295453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman} 296453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 297c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs 298c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// is the number of SCEVAddRecExprs present, which are kept at the end of 299c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// the list. 300c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// 301c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohmanstatic void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops, 302db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty, 303c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ScalarEvolution &SE) { 304c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman unsigned NumAddRecs = 0; 305c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i) 306c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ++NumAddRecs; 307c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Group Ops into non-addrecs and addrecs. 308c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs); 309c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end()); 310c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Let ScalarEvolution sort and simplify the non-addrecs list. 311c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *Sum = NoAddRecs.empty() ? 312deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman SE.getConstant(Ty, 0) : 313c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SE.getAddExpr(NoAddRecs); 314c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // If it returned an add, use the operands. Otherwise it simplified 315c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // the sum into a single value, so just use that. 316f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman Ops.clear(); 317c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum)) 318403a8cdda5e76ea689693de16474650b4b0df818Dan Gohman Ops.append(Add->op_begin(), Add->op_end()); 319f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman else if (!Sum->isZero()) 320f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman Ops.push_back(Sum); 321c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Then append the addrecs. 322403a8cdda5e76ea689693de16474650b4b0df818Dan Gohman Ops.append(AddRecs.begin(), AddRecs.end()); 323c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman} 324c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 325c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// SplitAddRecs - Flatten a list of add operands, moving addrec start values 326c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}. 327c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// This helps expose more opportunities for folding parts of the expressions 328c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// into GEP indices. 329c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// 330c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohmanstatic void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops, 331db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty, 332c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ScalarEvolution &SE) { 333c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Find the addrecs. 334c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SmallVector<const SCEV *, 8> AddRecs; 335c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman for (unsigned i = 0, e = Ops.size(); i != e; ++i) 336c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) { 337c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *Start = A->getStart(); 338c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (Start->isZero()) break; 339deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman const SCEV *Zero = SE.getConstant(Ty, 0); 340c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman AddRecs.push_back(SE.getAddRecExpr(Zero, 341c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman A->getStepRecurrence(SE), 3423228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick A->getLoop(), 3436f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick A->getNoWrapFlags(SCEV::FlagNW))); 344c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) { 345c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman Ops[i] = Zero; 346403a8cdda5e76ea689693de16474650b4b0df818Dan Gohman Ops.append(Add->op_begin(), Add->op_end()); 347c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman e += Add->getNumOperands(); 348c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } else { 349c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman Ops[i] = Start; 350c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 351c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 352c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (!AddRecs.empty()) { 353c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Add the addrecs onto the end of the list. 354403a8cdda5e76ea689693de16474650b4b0df818Dan Gohman Ops.append(AddRecs.begin(), AddRecs.end()); 355c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Resort the operand list, moving any constants to the front. 356c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SimplifyAddOperands(Ops, Ty, SE); 357c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 358c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman} 359c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 3604c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// expandAddToGEP - Expand an addition expression with a pointer type into 3614c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps 3624c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// BasicAliasAnalysis and other passes analyze the result. See the rules 3634c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// for getelementptr vs. inttoptr in 3644c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// http://llvm.org/docs/LangRef.html#pointeraliasing 3654c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// for details. 36613c5e35222afe0895f0c5e68aa9f22f134ea437aDan Gohman/// 3673abf905b50f33340ae81913da81b0eda96fa4616Dan Gohman/// Design note: The correctness of using getelementptr here depends on 3684c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as 3694c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// they may introduce pointer arithmetic which may not be safely converted 3704c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// into getelementptr. 371453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// 372453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// Design note: It might seem desirable for this function to be more 373453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// loop-aware. If some of the indices are loop-invariant while others 374453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// aren't, it might seem desirable to emit multiple GEPs, keeping the 375453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// loop-invariant portions of the overall computation outside the loop. 376453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// However, there are a few reasons this is not done here. Hoisting simple 377453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// arithmetic is a low-level optimization that often isn't very 378453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// important until late in the optimization process. In fact, passes 379453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// like InstructionCombining will combine GEPs, even if it means 380453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// pushing loop-invariant computation down into loops, so even if the 381453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// GEPs were split here, the work would quickly be undone. The 382453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// LoopStrengthReduction pass, which is usually run quite late (and 383453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// after the last InstructionCombining pass), takes care of hoisting 384453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// loop-invariant portions of expressions, after considering what 385453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// can be folded using target addressing modes. 386453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// 3870bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan GohmanValue *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin, 3880bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *const *op_end, 389db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner PointerType *PTy, 390db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty, 3915be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman Value *V) { 3924c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar Type *OriginalElTy = PTy->getElementType(); 3934c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar Type *ElTy = OriginalElTy; 3945be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman SmallVector<Value *, 4> GepIndices; 3950bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman SmallVector<const SCEV *, 8> Ops(op_begin, op_end); 3965be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman bool AnyNonZeroIndices = false; 3975be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 398c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Split AddRecs up into parts as either of the parts may be usable 399c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // without the other. 400c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SplitAddRecs(Ops, Ty, SE); 401c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 4024c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar Type *IntPtrTy = DL.getIntPtrType(PTy); 40314807bd8c801f976c999e5a6699f31ee9642021aMatt Arsenault 404eb35699e642f2c021f04154bdb4bd90a1afb3baaBob Wilson // Descend down the pointer's type and attempt to convert the other 4055be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // operands into GEP indices, at each level. The first index in a GEP 4065be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // indexes into the array implied by the pointer operand; the rest of 4075be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // the indices index into the element or field type selected by the 4085be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // preceding index. 4095be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman for (;;) { 410c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // If the scale size is not 0, attempt to factor out a scale for 411c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // array indexing. 4120bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman SmallVector<const SCEV *, 8> ScaledOps; 413150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman if (ElTy->isSized()) { 41414807bd8c801f976c999e5a6699f31ee9642021aMatt Arsenault const SCEV *ElSize = SE.getSizeOfExpr(IntPtrTy, ElTy); 415150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman if (!ElSize->isZero()) { 416150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman SmallVector<const SCEV *, 8> NewOps; 417cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar for (const SCEV *Op : Ops) { 418deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman const SCEV *Remainder = SE.getConstant(Ty, 0); 4194c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar if (FactorOutConstant(Op, Remainder, ElSize, SE, DL)) { 420150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman // Op now has ElSize factored out. 421150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman ScaledOps.push_back(Op); 422150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman if (!Remainder->isZero()) 423150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman NewOps.push_back(Remainder); 424150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman AnyNonZeroIndices = true; 425150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman } else { 426150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman // The operand was not divisible, so add it to the list of operands 427150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman // we'll scan next iteration. 428cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar NewOps.push_back(Op); 429150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman } 430150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman } 431150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman // If we made any changes, update Ops. 432150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman if (!ScaledOps.empty()) { 433150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman Ops = NewOps; 434150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman SimplifyAddOperands(Ops, Ty, SE); 4355be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 436c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 4375be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 438c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 439c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Record the scaled array index for this level of the type. If 440c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // we didn't find any operands that could be factored, tentatively 441c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // assume that element zero was selected (since the zero offset 442c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // would obviously be folded away). 4435be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman Value *Scaled = ScaledOps.empty() ? 444a7235ea7245028a0723e8ab7fd011386b3900777Owen Anderson Constant::getNullValue(Ty) : 4455be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman expandCodeFor(SE.getAddExpr(ScaledOps), Ty); 4465be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman GepIndices.push_back(Scaled); 4475be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 4485be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // Collect struct field index operands. 449db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner while (StructType *STy = dyn_cast<StructType>(ElTy)) { 450c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman bool FoundFieldNo = false; 451c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // An empty struct has no fields. 452c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (STy->getNumElements() == 0) break; 4534c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar // Field offsets are known. See if a constant offset falls within any of 4544c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar // the struct fields. 4554c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar if (Ops.empty()) 4564c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar break; 4574c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0])) 4584c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar if (SE.getTypeSizeInBits(C->getType()) <= 64) { 4594c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar const StructLayout &SL = *DL.getStructLayout(STy); 4604c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar uint64_t FullOffset = C->getValue()->getZExtValue(); 4614c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar if (FullOffset < SL.getSizeInBytes()) { 4624c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar unsigned ElIdx = SL.getElementContainingOffset(FullOffset); 4634c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar GepIndices.push_back( 4644c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx)); 4654c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar ElTy = STy->getTypeAtIndex(ElIdx); 4664c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar Ops[0] = 4676de29f8d960505421d61c80cdb738e16720b6c0eDan Gohman SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx)); 4684c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar AnyNonZeroIndices = true; 4694c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar FoundFieldNo = true; 4700f5efe56258f8cd6ceff4d7955a5d80144cd9cb0Dan Gohman } 4714c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar } 472c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // If no struct field offsets were found, tentatively assume that 473c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // field zero was selected (since the zero offset would obviously 474c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // be folded away). 475c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (!FoundFieldNo) { 476c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ElTy = STy->getTypeAtIndex(0u); 477c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman GepIndices.push_back( 478c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman Constant::getNullValue(Type::getInt32Ty(Ty->getContext()))); 479c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 480c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 4815be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 482db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) 4835be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman ElTy = ATy->getElementType(); 484c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman else 485c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman break; 4865be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 4875be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 4883f46a3abeedba8d517b4182de34c821d752db058Dan Gohman // If none of the operands were convertible to proper GEP indices, cast 4895be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // the base to i8* and do an ugly getelementptr with that. It's still 4905be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // better than ptrtoint+arithmetic+inttoptr at least. 4915be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (!AnyNonZeroIndices) { 492c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Cast the base to i8*. 4935be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman V = InsertNoopCastOfTo(V, 494ac53a0b272452013124bfc70480aea5e41b60f40Duncan Sands Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace())); 495c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 496705b48d960dff1a96ac40d0cf932eb465b9a550aRafael Espindola assert(!isa<Instruction>(V) || 497cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar SE.DT.dominates(cast<Instruction>(V), &*Builder.GetInsertPoint())); 4984b04578d65e38cdb5077de2498889e4a174ccdfdRafael Espindola 499c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Expand the operands for a plain byte offset. 50092fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty); 5015be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 5025be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // Fold a GEP with constant operands. 5035be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (Constant *CLHS = dyn_cast<Constant>(V)) 5045be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (Constant *CRHS = dyn_cast<Constant>(Idx)) 5050c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar return ConstantExpr::getGetElementPtr(Type::getInt8Ty(Ty->getContext()), 5060c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar CLHS, CRHS); 5075be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 5085be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // Do a quick scan to see if we have this GEP nearby. If so, reuse it. 5095be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman unsigned ScanLimit = 6; 510267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 511267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman // Scanning starts from the last instruction before the insertion point. 512267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock::iterator IP = Builder.GetInsertPoint(); 513267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman if (IP != BlockBegin) { 5145be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman --IP; 5155be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman for (; ScanLimit; --IP, --ScanLimit) { 5168d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen // Don't count dbg.value against the ScanLimit, to avoid perturbing the 5178d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen // generated code. 5188d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen if (isa<DbgInfoIntrinsic>(IP)) 5198d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen ScanLimit++; 5205be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (IP->getOpcode() == Instruction::GetElementPtr && 5215be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman IP->getOperand(0) == V && IP->getOperand(1) == Idx) 522cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return &*IP; 5235be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (IP == BlockBegin) break; 5245be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 5255be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 5265be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 527087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Save the original insertion point so we can restore it when we're done. 528d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer BuilderType::InsertPointGuard Guard(Builder); 529087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 530087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Move the insertion point out of as many loops as we can. 531cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) { 532087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break; 533087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman BasicBlock *Preheader = L->getLoopPreheader(); 534087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!Preheader) break; 535087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 536087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Ok, move up a level. 537cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Builder.SetInsertPoint(Preheader->getTerminator()); 538087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 539087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 540c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Emit a GEP. 5410c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar Value *GEP = Builder.CreateGEP(Builder.getInt8Ty(), V, Idx, "uglygep"); 542a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(GEP); 543087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 5445be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman return GEP; 5455be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 5465be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 547087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Save the original insertion point so we can restore it when we're done. 548341562b9fb6574b1c9492e52cec24106cd31ce51Benjamin Kramer BuilderType::InsertPoint SaveInsertPt = Builder.saveIP(); 549087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 550087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Move the insertion point out of as many loops as we can. 551cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) { 552087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!L->isLoopInvariant(V)) break; 553087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 554cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar bool AnyIndexNotLoopInvariant = 555cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar std::any_of(GepIndices.begin(), GepIndices.end(), 556cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar [L](Value *Op) { return !L->isLoopInvariant(Op); }); 557cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 558087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (AnyIndexNotLoopInvariant) 559087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman break; 560087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 561087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman BasicBlock *Preheader = L->getLoopPreheader(); 562087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!Preheader) break; 563087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 564087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Ok, move up a level. 565cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Builder.SetInsertPoint(Preheader->getTerminator()); 566087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 567087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 568d6aa02de1076c801ac41295156a2379637976918Dan Gohman // Insert a pretty getelementptr. Note that this GEP is not marked inbounds, 569d6aa02de1076c801ac41295156a2379637976918Dan Gohman // because ScalarEvolution may have changed the address arithmetic to 570d6aa02de1076c801ac41295156a2379637976918Dan Gohman // compute a value which is beyond the end of the allocated object. 571a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Value *Casted = V; 572a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman if (V->getType() != PTy) 573a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Casted = InsertNoopCastOfTo(Casted, PTy); 574cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Value *GEP = Builder.CreateGEP(OriginalElTy, Casted, GepIndices, "scevgep"); 5755be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman Ops.push_back(SE.getUnknown(GEP)); 576a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(GEP); 577087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 578341562b9fb6574b1c9492e52cec24106cd31ce51Benjamin Kramer // Restore the original insert point. 579341562b9fb6574b1c9492e52cec24106cd31ce51Benjamin Kramer Builder.restoreIP(SaveInsertPt); 580341562b9fb6574b1c9492e52cec24106cd31ce51Benjamin Kramer 5815be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman return expand(SE.getAddExpr(Ops)); 5825be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman} 5835be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 584087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman/// PickMostRelevantLoop - Given two loops pick the one that's most relevant for 585087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman/// SCEV expansion. If they are nested, this is the most nested. If they are 586087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman/// neighboring, pick the later. 587087bd1e3a12893873761736bf0f905a350e9e708Dan Gohmanstatic const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B, 588087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman DominatorTree &DT) { 589087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!A) return B; 590087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!B) return A; 591087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (A->contains(B)) return B; 592087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (B->contains(A)) return A; 593087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (DT.dominates(A->getHeader(), B->getHeader())) return B; 594087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (DT.dominates(B->getHeader(), A->getHeader())) return A; 595087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return A; // Arbitrarily break the tie. 596087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman} 597087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 5989c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman/// getRelevantLoop - Get the most relevant loop associated with the given 599087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman/// expression, according to PickMostRelevantLoop. 6009c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohmanconst Loop *SCEVExpander::getRelevantLoop(const SCEV *S) { 6019c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman // Test whether we've already computed the most relevant loop for this SCEV. 602cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar auto Pair = RelevantLoops.insert(std::make_pair(S, nullptr)); 6039c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman if (!Pair.second) 6049c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman return Pair.first->second; 6059c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman 606087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (isa<SCEVConstant>(S)) 6079c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman // A constant has no relevant loops. 608dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines return nullptr; 609087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 610087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (const Instruction *I = dyn_cast<Instruction>(U->getValue())) 611cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return Pair.first->second = SE.LI.getLoopFor(I->getParent()); 6129c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman // A non-instruction has no relevant loops. 613dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines return nullptr; 614087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 615087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) { 616dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines const Loop *L = nullptr; 617087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 618087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman L = AR->getLoop(); 619cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar for (const SCEV *Op : N->operands()) 620cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar L = PickMostRelevantLoop(L, getRelevantLoop(Op), SE.DT); 6219c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman return RelevantLoops[N] = L; 6229c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman } 6239c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) { 6249c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman const Loop *Result = getRelevantLoop(C->getOperand()); 6259c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman return RelevantLoops[C] = Result; 6269c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman } 6279c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 628cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar const Loop *Result = PickMostRelevantLoop( 629cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar getRelevantLoop(D->getLHS()), getRelevantLoop(D->getRHS()), SE.DT); 6309c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman return RelevantLoops[D] = Result; 631087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 632087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman llvm_unreachable("Unexpected SCEV type!"); 633087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman} 634c4f7ec85ecb760fff2b702c6deb06506b968ba4fDan Gohman 635b35798347ea87b8b6d36155b211016a7769f01abDan Gohmannamespace { 636b35798347ea87b8b6d36155b211016a7769f01abDan Gohman 637087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman/// LoopCompare - Compare loops by PickMostRelevantLoop. 638087bd1e3a12893873761736bf0f905a350e9e708Dan Gohmanclass LoopCompare { 639087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman DominatorTree &DT; 640087bd1e3a12893873761736bf0f905a350e9e708Dan Gohmanpublic: 641087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman explicit LoopCompare(DominatorTree &dt) : DT(dt) {} 642087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 643087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman bool operator()(std::pair<const Loop *, const SCEV *> LHS, 644087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman std::pair<const Loop *, const SCEV *> RHS) const { 645bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman // Keep pointer operands sorted at the end. 646bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman if (LHS.second->getType()->isPointerTy() != 647bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman RHS.second->getType()->isPointerTy()) 648bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman return LHS.second->getType()->isPointerTy(); 649bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman 650087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Compare loops with PickMostRelevantLoop. 651087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (LHS.first != RHS.first) 652087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first; 653087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 654087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // If one operand is a non-constant negative and the other is not, 655087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // put the non-constant negative on the right so that a sub can 656087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // be used instead of a negate and add. 657f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick if (LHS.second->isNonConstantNegative()) { 658f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick if (!RHS.second->isNonConstantNegative()) 659087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return false; 660f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick } else if (RHS.second->isNonConstantNegative()) 661087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return true; 662087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 663087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Otherwise they are equivalent according to this comparison. 664087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return false; 665c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 666087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman}; 6675be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 668b35798347ea87b8b6d36155b211016a7769f01abDan Gohman} 669b35798347ea87b8b6d36155b211016a7769f01abDan Gohman 670087bd1e3a12893873761736bf0f905a350e9e708Dan GohmanValue *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) { 671db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 672e24fa64d52330626553298f56ba5aa702624c282Dan Gohman 673087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Collect all the add operands in a loop, along with their associated loops. 674087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Iterate in reverse so that constants are emitted last, all else equal, and 675087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // so that pointer operands are inserted first, which the code below relies on 676087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // to form more involved GEPs. 677087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 678087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()), 679087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman E(S->op_begin()); I != E; ++I) 6809c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 681087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 682087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Sort by loop. Use a stable sort so that constants follow non-constants and 683087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // pointer operands precede non-pointer operands. 684cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(SE.DT)); 685087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 686087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Emit instructions to add all the operands. Hoist as much as possible 687087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // out of loops, and form meaningful getelementptrs where possible. 688dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines Value *Sum = nullptr; 689cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar for (auto I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E;) { 690087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman const Loop *CurLoop = I->first; 691087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman const SCEV *Op = I->second; 692087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!Sum) { 693087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // This is the first operand. Just expand it. 694087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = expand(Op); 695087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman ++I; 696db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) { 697087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // The running sum expression is a pointer. Try to form a getelementptr 698087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // at this level with that as the base. 699087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman SmallVector<const SCEV *, 4> NewOps; 700bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman for (; I != E && I->first == CurLoop; ++I) { 701bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman // If the operand is SCEVUnknown and not instructions, peek through 702bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman // it, to enable more of it to be folded into the GEP. 703bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman const SCEV *X = I->second; 704bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X)) 705bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman if (!isa<Instruction>(U->getValue())) 706bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman X = SE.getSCEV(U->getValue()); 707bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman NewOps.push_back(X); 708bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman } 709087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum); 710db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) { 711087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // The running sum is an integer, and there's a pointer at this level. 712f8d0578e4cbd5922696c92f5068c5513d8e8d60eDan Gohman // Try to form a getelementptr. If the running sum is instructions, 713f8d0578e4cbd5922696c92f5068c5513d8e8d60eDan Gohman // use a SCEVUnknown to avoid re-analyzing them. 714087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman SmallVector<const SCEV *, 4> NewOps; 715f8d0578e4cbd5922696c92f5068c5513d8e8d60eDan Gohman NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) : 716f8d0578e4cbd5922696c92f5068c5513d8e8d60eDan Gohman SE.getSCEV(Sum)); 717087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman for (++I; I != E && I->first == CurLoop; ++I) 718087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman NewOps.push_back(I->second); 719087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op)); 720f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick } else if (Op->isNonConstantNegative()) { 721087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Instead of doing a negate and add, just do a subtract. 722a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty); 723087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = InsertNoopCastOfTo(Sum, Ty); 724087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = InsertBinop(Instruction::Sub, Sum, W); 725087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman ++I; 726a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } else { 727087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // A simple add. 728a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Value *W = expandCodeFor(Op, Ty); 729087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = InsertNoopCastOfTo(Sum, Ty); 730087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Canonicalize a constant to the RHS. 731087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (isa<Constant>(Sum)) std::swap(Sum, W); 732087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = InsertBinop(Instruction::Add, Sum, W); 733087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman ++I; 734a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 7352d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman } 736087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 737087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return Sum; 738e24fa64d52330626553298f56ba5aa702624c282Dan Gohman} 7395be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 740890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) { 741db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 742087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 743087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Collect all the mul operands in a loop, along with their associated loops. 744087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Iterate in reverse so that constants are emitted last, all else equal. 745087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 746087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()), 747087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman E(S->op_begin()); I != E; ++I) 7489c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 749087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 750087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Sort by loop. Use a stable sort so that constants follow non-constants. 751cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(SE.DT)); 752087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 753087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Emit instructions to mul all the operands. Hoist as much as possible 754087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // out of loops. 755dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines Value *Prod = nullptr; 756cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar for (const auto &I : OpsAndLoops) { 757cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar const SCEV *Op = I.second; 758087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!Prod) { 759087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // This is the first operand. Just expand it. 760087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Prod = expand(Op); 761087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } else if (Op->isAllOnesValue()) { 762087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Instead of doing a multiply by negative one, just do a negate. 763087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Prod = InsertNoopCastOfTo(Prod, Ty); 764087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod); 765087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } else { 766087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // A simple mul. 767087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Value *W = expandCodeFor(Op, Ty); 768087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Prod = InsertNoopCastOfTo(Prod, Ty); 769087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Canonicalize a constant to the RHS. 770087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (isa<Constant>(Prod)) std::swap(Prod, W); 771cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar const APInt *RHS; 772cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (match(W, m_Power2(RHS))) { 773cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar // Canonicalize Prod*(1<<C) to Prod<<C. 774cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar assert(!Ty->isVectorTy() && "vector types are not SCEVable"); 775cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Prod = InsertBinop(Instruction::Shl, Prod, 776cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar ConstantInt::get(Ty, RHS->logBase2())); 777cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar } else { 778cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Prod = InsertBinop(Instruction::Mul, Prod, W); 779cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar } 780087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 7812d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman } 7822d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman 783087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return Prod; 78436f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman} 78536f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 786890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) { 787db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 7882d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman 78992fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *LHS = expandCodeFor(S->getLHS(), Ty); 790890f92b744fb074465bc2b7006ee753a181f62a4Dan Gohman if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) { 791cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar const APInt &RHS = SC->getAPInt(); 7926177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky if (RHS.isPowerOf2()) 7936177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky return InsertBinop(Instruction::LShr, LHS, 794eed707b1e6097aac2bb6b3d47271f6300ace7f2eOwen Anderson ConstantInt::get(Ty, RHS.logBase2())); 7956177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky } 7966177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky 79792fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *RHS = expandCodeFor(S->getRHS(), Ty); 798267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman return InsertBinop(Instruction::UDiv, LHS, RHS); 7996177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky} 8006177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky 801453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// Move parts of Base into Rest to leave Base with the minimal 802453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// expression that provides a pointer operand suitable for a 803453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// GEP expansion. 8040bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohmanstatic void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest, 805453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman ScalarEvolution &SE) { 806453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) { 807453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman Base = A->getStart(); 808453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman Rest = SE.getAddExpr(Rest, 809deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman SE.getAddRecExpr(SE.getConstant(A->getType(), 0), 810453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman A->getStepRecurrence(SE), 8113228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick A->getLoop(), 8126f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick A->getNoWrapFlags(SCEV::FlagNW))); 813453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 814453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) { 815453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman Base = A->getOperand(A->getNumOperands()-1); 8160bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end()); 817453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman NewAddOps.back() = Rest; 818453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman Rest = SE.getAddExpr(NewAddOps); 819453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman ExposePointerBase(Base, Rest, SE); 820453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 821453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman} 822453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 823c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick/// Determine if this is a well-behaved chain of instructions leading back to 824c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick/// the PHI. If so, it may be reused by expanded expressions. 825c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trickbool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV, 826c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick const Loop *L) { 827c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) || 828c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV))) 829c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return false; 830c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // If any of the operands don't dominate the insert position, bail. 831c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Addrec operands are always loop-invariant, so this can only happen 832c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // if there are instructions which haven't been hoisted. 833c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (L == IVIncInsertLoop) { 834c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick for (User::op_iterator OI = IncV->op_begin()+1, 835c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick OE = IncV->op_end(); OI != OE; ++OI) 836c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (Instruction *OInst = dyn_cast<Instruction>(OI)) 837cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (!SE.DT.dominates(OInst, IVIncInsertPos)) 838c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return false; 839c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 840c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Advance to the next instruction. 841c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick IncV = dyn_cast<Instruction>(IncV->getOperand(0)); 842c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (!IncV) 843c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return false; 844c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 845c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (IncV->mayHaveSideEffects()) 846c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return false; 847c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 848c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (IncV != PN) 849c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return true; 850c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 851c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return isNormalAddRecExprPHI(PN, IncV, L); 852c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick} 853c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 854b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// getIVIncOperand returns an induction variable increment's induction 855b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// variable operand. 856b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// 857b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// If allowScale is set, any type of GEP is allowed as long as the nonIV 858b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// operands dominate InsertPos. 859b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// 860b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// If allowScale is not set, ensure that a GEP increment conforms to one of the 861b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// simple patterns generated by getAddRecExprPHILiterally and 862b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// expandAddtoGEP. If the pattern isn't recognized, return NULL. 863b5c26ef9da16052597d59a412eaae32098aa1be0Andrew TrickInstruction *SCEVExpander::getIVIncOperand(Instruction *IncV, 864b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick Instruction *InsertPos, 865b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick bool allowScale) { 866b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (IncV == InsertPos) 867dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines return nullptr; 86864925c55c65f9345a69fb67db07aa62cfb723577Andrew Trick 869c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick switch (IncV->getOpcode()) { 870b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick default: 871dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines return nullptr; 872c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Check for a simple Add/Sub or GEP of a loop invariant step. 873c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick case Instruction::Add: 874b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick case Instruction::Sub: { 875b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1)); 876cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (!OInst || SE.DT.dominates(OInst, InsertPos)) 877b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return dyn_cast<Instruction>(IncV->getOperand(0)); 878dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines return nullptr; 879b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 880c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick case Instruction::BitCast: 881b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return dyn_cast<Instruction>(IncV->getOperand(0)); 882b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick case Instruction::GetElementPtr: 883cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar for (auto I = IncV->op_begin() + 1, E = IncV->op_end(); I != E; ++I) { 884c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (isa<Constant>(*I)) 885c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick continue; 886b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (Instruction *OInst = dyn_cast<Instruction>(*I)) { 887cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (!SE.DT.dominates(OInst, InsertPos)) 888dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines return nullptr; 889b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 890b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (allowScale) { 891b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // allow any kind of GEP as long as it can be hoisted. 892b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick continue; 893b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 894b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // This must be a pointer addition of constants (pretty), which is already 895b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // handled, or some number of address-size elements (ugly). Ugly geps 896b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // have 2 operands. i1* is used by the expander to represent an 897b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // address-size element. 898c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (IncV->getNumOperands() != 2) 899dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines return nullptr; 900365c9f1ff55bef134c6b9707f7df44d680ddabeaAndrew Trick unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace(); 901c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS) 902c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS)) 903dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines return nullptr; 904c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick break; 905c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 906b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return dyn_cast<Instruction>(IncV->getOperand(0)); 907c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 908b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick} 909b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick 910b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make 911b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// it available to other uses in this loop. Recursively hoist any operands, 912b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// until we reach a value that dominates InsertPos. 913b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trickbool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) { 914cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (SE.DT.dominates(IncV, InsertPos)) 915b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return true; 916b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick 917b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // InsertPos must itself dominate IncV so that IncV's new position satisfies 918b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // its existing users. 919cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (isa<PHINode>(InsertPos) || 920cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar !SE.DT.dominates(InsertPos->getParent(), IncV->getParent())) 921cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return false; 922cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 923cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (!SE.LI.movementPreservesLCSSAForm(IncV, InsertPos)) 924c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return false; 925b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick 926b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // Check that the chain of IV operands leading back to Phi can be hoisted. 927b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick SmallVector<Instruction*, 4> IVIncs; 928b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick for(;;) { 929b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true); 930b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (!Oper) 931b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return false; 932b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // IncV is safe to hoist. 933b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick IVIncs.push_back(IncV); 934b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick IncV = Oper; 935cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (SE.DT.dominates(IncV, InsertPos)) 936b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick break; 937b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 938cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar for (auto I = IVIncs.rbegin(), E = IVIncs.rend(); I != E; ++I) { 939b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick (*I)->moveBefore(InsertPos); 940b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 941b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return true; 942b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick} 943b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick 944b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// Determine if this cyclic phi is in a form that would have been generated by 945b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// LSR. We don't care if the phi was actually expanded in this pass, as long 946b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// as it is in a low-cost form, for example, no implied multiplication. This 947b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// should match any patterns generated by getAddRecExprPHILiterally and 948b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// expandAddtoGEP. 949b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trickbool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV, 950b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick const Loop *L) { 951b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick for(Instruction *IVOper = IncV; 952b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(), 953b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick /*allowScale=*/false));) { 954b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (IVOper == PN) 955b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return true; 956c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 957b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return false; 958c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick} 959c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 960553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick/// expandIVInc - Expand an IV increment at Builder's current InsertPos. 961553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick/// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may 962553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick/// need to materialize IV increments elsewhere to handle difficult situations. 963553fe05f236f46fe27b7fcfa822b06367d50183eAndrew TrickValue *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L, 964553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Type *ExpandTy, Type *IntTy, 965553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick bool useSubtract) { 966553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Value *IncV; 967553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // If the PHI is a pointer, use a GEP, otherwise use an add or sub. 968553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick if (ExpandTy->isPointerTy()) { 969553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick PointerType *GEPPtrTy = cast<PointerType>(ExpandTy); 970553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // If the step isn't constant, don't use an implicitly scaled GEP, because 971553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // that would require a multiply inside the loop. 972553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick if (!isa<ConstantInt>(StepV)) 973553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()), 974553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick GEPPtrTy->getAddressSpace()); 975553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick const SCEV *const StepArray[1] = { SE.getSCEV(StepV) }; 976553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN); 977553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick if (IncV->getType() != PN->getType()) { 978553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick IncV = Builder.CreateBitCast(IncV, PN->getType()); 979553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick rememberInstruction(IncV); 980553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick } 981553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick } else { 982553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick IncV = useSubtract ? 983553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") : 984553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next"); 985553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick rememberInstruction(IncV); 986553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick } 987553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick return IncV; 988553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick} 989553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick 99036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines/// \brief Hoist the addrec instruction chain rooted in the loop phi above the 99136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines/// position. This routine assumes that this is possible (has been checked). 99236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hinesstatic void hoistBeforePos(DominatorTree *DT, Instruction *InstToHoist, 99336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Instruction *Pos, PHINode *LoopPhi) { 99436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines do { 99536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (DT->dominates(InstToHoist, Pos)) 99636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines break; 99736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Make sure the increment is where we want it. But don't move it 99836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // down past a potential existing post-inc user. 99936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines InstToHoist->moveBefore(Pos); 100036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Pos = InstToHoist; 100136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines InstToHoist = cast<Instruction>(InstToHoist->getOperand(0)); 100236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } while (InstToHoist != LoopPhi); 100336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines} 100436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 100536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines/// \brief Check whether we can cheaply express the requested SCEV in terms of 1006cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar/// the available PHI SCEV by truncation and/or inversion of the step. 100736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hinesstatic bool canBeCheaplyTransformed(ScalarEvolution &SE, 100836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines const SCEVAddRecExpr *Phi, 100936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines const SCEVAddRecExpr *Requested, 101036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines bool &InvertStep) { 101136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Type *PhiTy = SE.getEffectiveSCEVType(Phi->getType()); 101236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Type *RequestedTy = SE.getEffectiveSCEVType(Requested->getType()); 101336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 101436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (RequestedTy->getIntegerBitWidth() > PhiTy->getIntegerBitWidth()) 101536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines return false; 101636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 101736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Try truncate it if necessary. 101836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Phi = dyn_cast<SCEVAddRecExpr>(SE.getTruncateOrNoop(Phi, RequestedTy)); 101936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (!Phi) 102036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines return false; 102136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 102236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Check whether truncation will help. 102336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (Phi == Requested) { 102436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines InvertStep = false; 102536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines return true; 102636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } 102736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 102836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Check whether inverting will help: {R,+,-1} == R - {0,+,1}. 102936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (SE.getAddExpr(Requested->getStart(), 103036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines SE.getNegativeSCEV(Requested)) == Phi) { 103136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines InvertStep = true; 103236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines return true; 103336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } 103436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 103536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines return false; 103636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines} 103736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 1038ebe69fe11e48d322045d5949c83283927a0d790bStephen Hinesstatic bool IsIncrementNSW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) { 1039ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines if (!isa<IntegerType>(AR->getType())) 1040ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines return false; 1041ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines 1042ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth(); 1043ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2); 1044ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines const SCEV *Step = AR->getStepRecurrence(SE); 1045ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines const SCEV *OpAfterExtend = SE.getAddExpr(SE.getSignExtendExpr(Step, WideTy), 1046ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines SE.getSignExtendExpr(AR, WideTy)); 1047ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines const SCEV *ExtendAfterOp = 1048ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines SE.getSignExtendExpr(SE.getAddExpr(AR, Step), WideTy); 1049ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines return ExtendAfterOp == OpAfterExtend; 1050ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines} 1051ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines 1052ebe69fe11e48d322045d5949c83283927a0d790bStephen Hinesstatic bool IsIncrementNUW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) { 1053ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines if (!isa<IntegerType>(AR->getType())) 1054ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines return false; 1055ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines 1056ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth(); 1057ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2); 1058ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines const SCEV *Step = AR->getStepRecurrence(SE); 1059ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines const SCEV *OpAfterExtend = SE.getAddExpr(SE.getZeroExtendExpr(Step, WideTy), 1060ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines SE.getZeroExtendExpr(AR, WideTy)); 1061ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines const SCEV *ExtendAfterOp = 1062ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines SE.getZeroExtendExpr(SE.getAddExpr(AR, Step), WideTy); 1063ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines return ExtendAfterOp == OpAfterExtend; 1064ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines} 1065ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines 1066a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman/// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand 1067a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman/// the base addrec, which is the addrec without any non-loop-dominating 1068a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman/// values, and return the PHI. 1069a10756ee657a4d43a48cca5c166919093930ed6bDan GohmanPHINode * 1070a10756ee657a4d43a48cca5c166919093930ed6bDan GohmanSCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized, 1071a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const Loop *L, 1072db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *ExpandTy, 107336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Type *IntTy, 107436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Type *&TruncTy, 107536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines bool &InvertStep) { 107693a896e2e36480d55de3bab53e68581e08526344Benjamin Kramer assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position"); 1077d152d03a476b8d0d4b26577db26e2ba76034b0f3Andrew Trick 1078a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Reuse a previously-inserted PHI, if present. 1079c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick BasicBlock *LatchBlock = L->getLoopLatch(); 1080c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (LatchBlock) { 1081dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines PHINode *AddRecPhiMatch = nullptr; 1082dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines Instruction *IncV = nullptr; 1083dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines TruncTy = nullptr; 108436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines InvertStep = false; 108536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 108636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Only try partially matching scevs that need truncation and/or 108736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // step-inversion if we know this loop is outside the current loop. 1088cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar bool TryNonMatchingSCEV = 1089cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar IVIncInsertLoop && 1090cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar SE.DT.properlyDominates(LatchBlock, IVIncInsertLoop->getHeader()); 109136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 1092cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar for (auto &I : *L->getHeader()) { 1093cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar auto *PN = dyn_cast<PHINode>(&I); 1094cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (!PN || !SE.isSCEVable(PN->getType())) 109536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines continue; 109636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 109736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines const SCEVAddRecExpr *PhiSCEV = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PN)); 109836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (!PhiSCEV) 1099c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick continue; 1100c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 110136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines bool IsMatchingSCEV = PhiSCEV == Normalized; 110236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // We only handle truncation and inversion of phi recurrences for the 110336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // expanded expression if the expanded expression's loop dominates the 110436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // loop we insert to. Check now, so we can bail out early. 110536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (!IsMatchingSCEV && !TryNonMatchingSCEV) 110636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines continue; 110736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 110836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Instruction *TempIncV = 110936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)); 1110c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 111136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Check whether we can reuse this PHI node. 1112c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (LSRMode) { 111336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (!isExpandedAddRecExprPHI(PN, TempIncV, L)) 1114c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick continue; 111536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (L == IVIncInsertLoop && !hoistIVInc(TempIncV, IVIncInsertPos)) 1116b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick continue; 111736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } else { 111836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (!isNormalAddRecExprPHI(PN, TempIncV, L)) 1119c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick continue; 1120c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 112136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 112236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Stop if we have found an exact match SCEV. 112336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (IsMatchingSCEV) { 112436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines IncV = TempIncV; 1125dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines TruncTy = nullptr; 112636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines InvertStep = false; 112736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines AddRecPhiMatch = PN; 112836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines break; 112936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } 113036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 113136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Try whether the phi can be translated into the requested form 113236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // (truncated and/or offset by a constant). 113336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if ((!TruncTy || InvertStep) && 113436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines canBeCheaplyTransformed(SE, PhiSCEV, Normalized, InvertStep)) { 113536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Record the phi node. But don't stop we might find an exact match 113636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // later. 113736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines AddRecPhiMatch = PN; 113836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines IncV = TempIncV; 113936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines TruncTy = SE.getEffectiveSCEVType(Normalized->getType()); 114036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } 114136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } 114236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 114336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (AddRecPhiMatch) { 114436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Potentially, move the increment. We have made sure in 114536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // isExpandedAddRecExprPHI or hoistIVInc that this is possible. 114636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (L == IVIncInsertLoop) 1147cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar hoistBeforePos(&SE.DT, IncV, IVIncInsertPos, AddRecPhiMatch); 114836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 1149c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Ok, the add recurrence looks usable. 1150c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Remember this PHI, even in post-inc mode. 115136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines InsertedValues.insert(AddRecPhiMatch); 1152c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Remember the increment. 1153c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick rememberInstruction(IncV); 115436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines return AddRecPhiMatch; 1155c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 1156c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 1157a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1158a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Save the original insertion point so we can restore it when we're done. 1159d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer BuilderType::InsertPointGuard Guard(Builder); 1160a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1161ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // Another AddRec may need to be recursively expanded below. For example, if 1162ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // this AddRec is quadratic, the StepV may itself be an AddRec in this 1163ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // loop. Remove this loop from the PostIncLoops set before expanding such 1164ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // AddRecs. Otherwise, we cannot find a valid position for the step 1165ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // (i.e. StepV can never dominate its loop header). Ideally, we could do 1166ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // SavedIncLoops.swap(PostIncLoops), but we generally have a single element, 1167ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // so it's not worth implementing SmallPtrSet::swap. 1168ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick PostIncLoopSet SavedPostIncLoops = PostIncLoops; 1169ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick PostIncLoops.clear(); 1170ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick 1171a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Expand code for the start value. 1172cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Value *StartV = 1173cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar expandCodeFor(Normalized->getStart(), ExpandTy, &L->getHeader()->front()); 1174a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1175d152d03a476b8d0d4b26577db26e2ba76034b0f3Andrew Trick // StartV must be hoisted into L's preheader to dominate the new phi. 117693a896e2e36480d55de3bab53e68581e08526344Benjamin Kramer assert(!isa<Instruction>(StartV) || 1177cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar SE.DT.properlyDominates(cast<Instruction>(StartV)->getParent(), 1178cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar L->getHeader())); 1179d152d03a476b8d0d4b26577db26e2ba76034b0f3Andrew Trick 1180553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // Expand code for the step value. Do this before creating the PHI so that PHI 1181553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // reuse code doesn't see an incomplete PHI. 1182a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEV *Step = Normalized->getStepRecurrence(SE); 1183553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // If the stride is negative, insert a sub instead of an add for the increment 1184553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // (unless it's a constant, because subtracts of constants are canonicalized 1185553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // to adds). 1186f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1187553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick if (useSubtract) 1188a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Step = SE.getNegativeSCEV(Step); 1189553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // Expand the step somewhere that dominates the loop header. 1190cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Value *StepV = expandCodeFor(Step, IntTy, &L->getHeader()->front()); 1191a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1192ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines // The no-wrap behavior proved by IsIncrement(NUW|NSW) is only applicable if 1193ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines // we actually do emit an addition. It does not apply if we emit a 1194ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines // subtraction. 1195ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines bool IncrementIsNUW = !useSubtract && IsIncrementNUW(SE, Normalized); 1196ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines bool IncrementIsNSW = !useSubtract && IsIncrementNSW(SE, Normalized); 1197ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines 1198a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Create the PHI. 1199d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad BasicBlock *Header = L->getHeader(); 1200d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad Builder.SetInsertPoint(Header, Header->begin()); 1201d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 12025e7645be4c9dd2193add44d30b5fef8036d7a3ceAndrew Trick PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE), 1203dc8e546048db2f7ff5656742b2b26975098a11a0Andrew Trick Twine(IVName) + ".iv"); 1204a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(PN); 1205a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1206a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Create the step instructions and populate the PHI. 1207d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 1208a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman BasicBlock *Pred = *HPI; 1209a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1210a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Add a start value. 1211a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman if (!L->contains(Pred)) { 1212a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman PN->addIncoming(StartV, Pred); 1213a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman continue; 1214a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1215a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1216553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // Create a step value and add it to the PHI. 1217553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the 1218553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // instructions at IVIncInsertPos. 1219a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Instruction *InsertPos = L == IVIncInsertLoop ? 1220a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman IVIncInsertPos : Pred->getTerminator(); 1221c5ecbdc1896f1cc089372feef3191ace2f840898Devang Patel Builder.SetInsertPoint(InsertPos); 1222553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1223ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines 1224409443b1c6415e55c2bd4f0662e14cbc52d16686Andrew Trick if (isa<OverflowingBinaryOperator>(IncV)) { 1225ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines if (IncrementIsNUW) 1226409443b1c6415e55c2bd4f0662e14cbc52d16686Andrew Trick cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap(); 1227ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines if (IncrementIsNSW) 1228409443b1c6415e55c2bd4f0662e14cbc52d16686Andrew Trick cast<BinaryOperator>(IncV)->setHasNoSignedWrap(); 1229409443b1c6415e55c2bd4f0662e14cbc52d16686Andrew Trick } 1230a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman PN->addIncoming(IncV, Pred); 1231a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1232a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1233ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // After expanding subexpressions, restore the PostIncLoops set so the caller 1234ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // can ensure that IVIncrement dominates the current uses. 1235ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick PostIncLoops = SavedPostIncLoops; 1236ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick 1237a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Remember this PHI, even in post-inc mode. 1238a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman InsertedValues.insert(PN); 1239a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1240a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman return PN; 1241a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman} 1242a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1243a10756ee657a4d43a48cca5c166919093930ed6bDan GohmanValue *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) { 1244db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *STy = S->getType(); 1245db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *IntTy = SE.getEffectiveSCEVType(STy); 1246a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const Loop *L = S->getLoop(); 1247a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1248a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Determine a normalized form of this expression, which is the expression 1249a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // before any post-inc adjustment is made. 1250a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEVAddRecExpr *Normalized = S; 1251448db1cdef5872713ef77beffacf502ae3450cd7Dan Gohman if (PostIncLoops.count(L)) { 1252448db1cdef5872713ef77beffacf502ae3450cd7Dan Gohman PostIncLoopSet Loops; 1253448db1cdef5872713ef77beffacf502ae3450cd7Dan Gohman Loops.insert(L); 1254cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Normalized = cast<SCEVAddRecExpr>(TransformForPostIncUse( 1255cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Normalize, S, nullptr, nullptr, Loops, SE, SE.DT)); 1256a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1257a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1258a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Strip off any non-loop-dominating component from the addrec start. 1259a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEV *Start = Normalized->getStart(); 1260dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines const SCEV *PostLoopOffset = nullptr; 1261dc0e8fb9f9512622f55f73e1a434caa5c0915694Dan Gohman if (!SE.properlyDominates(Start, L->getHeader())) { 1262a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman PostLoopOffset = Start; 1263deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman Start = SE.getConstant(Normalized->getType(), 0); 12643228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick Normalized = cast<SCEVAddRecExpr>( 12653228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE), 12663228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick Normalized->getLoop(), 12676f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick Normalized->getNoWrapFlags(SCEV::FlagNW))); 1268a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1269a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1270a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Strip off any non-loop-dominating component from the addrec step. 1271a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEV *Step = Normalized->getStepRecurrence(SE); 1272dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines const SCEV *PostLoopScale = nullptr; 1273dc0e8fb9f9512622f55f73e1a434caa5c0915694Dan Gohman if (!SE.dominates(Step, L->getHeader())) { 1274a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman PostLoopScale = Step; 1275deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman Step = SE.getConstant(Normalized->getType(), 1); 1276a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Normalized = 12776f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick cast<SCEVAddRecExpr>(SE.getAddRecExpr( 12786f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick Start, Step, Normalized->getLoop(), 12796f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick Normalized->getNoWrapFlags(SCEV::FlagNW))); 1280a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1281a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1282a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Expand the core addrec. If we need post-loop scaling, force it to 1283a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // expand to an integer type to avoid the need for additional casting. 1284db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *ExpandTy = PostLoopScale ? IntTy : STy; 128536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // In some cases, we decide to reuse an existing phi node but need to truncate 128636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // it and/or invert the step. 1287dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines Type *TruncTy = nullptr; 128836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines bool InvertStep = false; 128936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy, 129036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines TruncTy, InvertStep); 1291a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 12923f46a3abeedba8d517b4182de34c821d752db058Dan Gohman // Accommodate post-inc mode, if necessary. 1293a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Value *Result; 1294448db1cdef5872713ef77beffacf502ae3450cd7Dan Gohman if (!PostIncLoops.count(L)) 1295a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Result = PN; 1296a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman else { 1297a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // In PostInc mode, use the post-incremented value. 1298a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman BasicBlock *LatchBlock = L->getLoopLatch(); 1299a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman assert(LatchBlock && "PostInc mode requires a unique loop latch!"); 1300a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Result = PN->getIncomingValueForBlock(LatchBlock); 130148ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick 130248ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // For an expansion to use the postinc form, the client must call 130348ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // expandCodeFor with an InsertPoint that is either outside the PostIncLoop 130448ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // or dominated by IVIncInsertPos. 1305cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (isa<Instruction>(Result) && 1306cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar !SE.DT.dominates(cast<Instruction>(Result), 1307cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar &*Builder.GetInsertPoint())) { 1308553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // The induction variable's postinc expansion does not dominate this use. 1309553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // IVUsers tries to prevent this case, so it is rare. However, it can 1310553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // happen when an IVUser outside the loop is not dominated by the latch 1311553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // block. Adjusting IVIncInsertPos before expansion begins cannot handle 1312553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // all cases. Consider a phi outide whose operand is replaced during 1313553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // expansion with the value of the postinc user. Without fundamentally 1314553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // changing the way postinc users are tracked, the only remedy is 1315553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // inserting an extra IV increment. StepV might fold into PostLoopOffset, 1316553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // but hopefully expandCodeFor handles that. 1317553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick bool useSubtract = 1318f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1319553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick if (useSubtract) 1320553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Step = SE.getNegativeSCEV(Step); 1321d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer Value *StepV; 1322d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer { 1323d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer // Expand the step somewhere that dominates the loop header. 1324d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer BuilderType::InsertPointGuard Guard(Builder); 1325cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar StepV = expandCodeFor(Step, IntTy, &L->getHeader()->front()); 1326d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer } 1327553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1328553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick } 1329a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1330a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 133136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // We have decided to reuse an induction variable of a dominating loop. Apply 133236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // truncation and/or invertion of the step. 133336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (TruncTy) { 133436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Type *ResTy = Result->getType(); 133536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Normalize the result type. 133636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (ResTy != SE.getEffectiveSCEVType(ResTy)) 133736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Result = InsertNoopCastOfTo(Result, SE.getEffectiveSCEVType(ResTy)); 133836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Truncate the result. 133936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (TruncTy != Result->getType()) { 134036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Result = Builder.CreateTrunc(Result, TruncTy); 134136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines rememberInstruction(Result); 134236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } 134336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Invert the result. 134436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (InvertStep) { 134536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Result = Builder.CreateSub(expandCodeFor(Normalized->getStart(), TruncTy), 134636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Result); 134736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines rememberInstruction(Result); 134836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } 134936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } 135036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 1351a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Re-apply any non-loop-dominating scale. 1352a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman if (PostLoopScale) { 13534d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick assert(S->isAffine() && "Can't linearly scale non-affine recurrences."); 13540a799ab15801d4ebf68eeb151d6375a799c87d9aDan Gohman Result = InsertNoopCastOfTo(Result, IntTy); 1355a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Result = Builder.CreateMul(Result, 1356a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman expandCodeFor(PostLoopScale, IntTy)); 1357a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(Result); 1358a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1359a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1360a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Re-apply any non-loop-dominating offset. 1361a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman if (PostLoopOffset) { 1362db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) { 1363a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEV *const OffsetArray[1] = { PostLoopOffset }; 1364a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result); 1365a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } else { 13660a799ab15801d4ebf68eeb151d6375a799c87d9aDan Gohman Result = InsertNoopCastOfTo(Result, IntTy); 1367a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Result = Builder.CreateAdd(Result, 1368a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman expandCodeFor(PostLoopOffset, IntTy)); 1369a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(Result); 1370a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1371a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1372a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1373a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman return Result; 1374a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman} 1375a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1376890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) { 1377a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman if (!CanonicalMode) return expandAddRecExprLiterally(S); 1378a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1379db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 138036f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman const Loop *L = S->getLoop(); 138136f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 13824d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // First check for an existing canonical IV in a suitable type. 1383dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines PHINode *CanonicalIV = nullptr; 13844d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman if (PHINode *PN = L->getCanonicalInductionVariable()) 1385133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty)) 13864d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman CanonicalIV = PN; 13874d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman 13884d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // Rewrite an AddRec in terms of the canonical induction variable, if 13894d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // its type is more narrow. 13904d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman if (CanonicalIV && 13914d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman SE.getTypeSizeInBits(CanonicalIV->getType()) > 13924d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman SE.getTypeSizeInBits(Ty)) { 1393f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman SmallVector<const SCEV *, 4> NewOps(S->getNumOperands()); 1394f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i) 1395f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType()); 13963228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(), 13976f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick S->getNoWrapFlags(SCEV::FlagNW))); 13984d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman BasicBlock::iterator NewInsertPt = 1399cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar findInsertPointAfter(cast<Instruction>(V), Builder.GetInsertBlock()); 1400dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), nullptr, 1401cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar &*NewInsertPt); 14024d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman return V; 14034d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman } 14044d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman 140536f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // {X,+,F} --> X + {0,+,F} 1406cfeb6a450632f2a6cd05302633c8c2b8c90cfdfdDan Gohman if (!S->getStart()->isZero()) { 1407f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end()); 1408deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman NewOps[0] = SE.getConstant(Ty, 0); 14096f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick const SCEV *Rest = SE.getAddRecExpr(NewOps, L, 14106f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick S->getNoWrapFlags(SCEV::FlagNW)); 1411453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 1412453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the 1413453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // comments on expandAddToGEP for details. 1414c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *Base = S->getStart(); 1415c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *RestArray[1] = { Rest }; 1416c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Dig into the expression to find the pointer base for a GEP. 1417c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ExposePointerBase(Base, RestArray[0], SE); 1418c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // If we found a pointer, expand the AddRec with a GEP. 1419db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) { 1420c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Make sure the Base isn't something exotic, such as a multiplied 1421c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // or divided pointer value. In those cases, the result type isn't 1422c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // actually a pointer type. 1423c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) { 1424c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman Value *StartV = expand(Base); 1425c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!"); 1426c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV); 1427453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 1428453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 1429453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 143040a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // Just do a normal add. Pre-expand the operands to suppress folding. 143140a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())), 143240a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman SE.getUnknown(expand(Rest)))); 143336f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman } 143436f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 14356ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman // If we don't yet have a canonical IV, create one. 14366ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman if (!CanonicalIV) { 143736f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // Create and insert the PHI node for the induction variable in the 143836f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // specified loop. 143936f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman BasicBlock *Header = L->getHeader(); 1440d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 14413ecfc861b4365f341c5c969b40e1afccde676e6fJay Foad CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar", 1442cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar &Header->front()); 14436ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman rememberInstruction(CanonicalIV); 144436f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 144519046ec19c4931f025cd4730eb43e9203ada6bb7Hal Finkel SmallSet<BasicBlock *, 4> PredSeen; 1446eed707b1e6097aac2bb6b3d47271f6300ace7f2eOwen Anderson Constant *One = ConstantInt::get(Ty, 1); 1447d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 14487656018c2268285907cfdc106071462a01a73878Gabor Greif BasicBlock *HP = *HPI; 144937ed9c199ca639565f6ce88105f9e39e898d82d0Stephen Hines if (!PredSeen.insert(HP).second) { 145037ed9c199ca639565f6ce88105f9e39e898d82d0Stephen Hines // There must be an incoming value for each predecessor, even the 145137ed9c199ca639565f6ce88105f9e39e898d82d0Stephen Hines // duplicates! 145237ed9c199ca639565f6ce88105f9e39e898d82d0Stephen Hines CanonicalIV->addIncoming(CanonicalIV->getIncomingValueForBlock(HP), HP); 145319046ec19c4931f025cd4730eb43e9203ada6bb7Hal Finkel continue; 145437ed9c199ca639565f6ce88105f9e39e898d82d0Stephen Hines } 145519046ec19c4931f025cd4730eb43e9203ada6bb7Hal Finkel 14567656018c2268285907cfdc106071462a01a73878Gabor Greif if (L->contains(HP)) { 14573abf905b50f33340ae81913da81b0eda96fa4616Dan Gohman // Insert a unit add instruction right before the terminator 14583abf905b50f33340ae81913da81b0eda96fa4616Dan Gohman // corresponding to the back-edge. 14596ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One, 14606ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman "indvar.next", 14616ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman HP->getTerminator()); 1462df3ad6697bb849e514c0881ca8700ea36678cbdaDevang Patel Add->setDebugLoc(HP->getTerminator()->getDebugLoc()); 1463a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(Add); 14646ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman CanonicalIV->addIncoming(Add, HP); 146583d577490b1473213b6973236110c28e59127956Dan Gohman } else { 14666ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP); 146783d577490b1473213b6973236110c28e59127956Dan Gohman } 14687656018c2268285907cfdc106071462a01a73878Gabor Greif } 146936f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman } 147036f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 14716ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman // {0,+,1} --> Insert a canonical induction variable into the loop! 14726ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman if (S->isAffine() && S->getOperand(1)->isOne()) { 14736ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) && 14746ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman "IVs with types different from the canonical IV should " 14756ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman "already have been handled!"); 14766ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman return CanonicalIV; 14776ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman } 14786ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman 14794d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // {0,+,F} --> {0,+,1} * F 148036f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 1481df14a04b5ccbbe6a46c2ccb93e27b12a36ff163eChris Lattner // If this is a simple linear addrec, emit it now as a special case. 148240a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman if (S->isAffine()) // {0,+,F} --> i*F 148340a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman return 148440a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman expand(SE.getTruncateOrNoop( 14856ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman SE.getMulExpr(SE.getUnknown(CanonicalIV), 148640a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman SE.getNoopOrAnyExtend(S->getOperand(1), 14876ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman CanonicalIV->getType())), 148840a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman Ty)); 148936f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 149036f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // If this is a chain of recurrences, turn it into a closed form, using the 149136f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // folders, then expandCodeFor the closed form. This allows the folders to 149236f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // simplify the expression without having to build a bunch of special code 149336f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // into this folder. 14946ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV. 149536f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 14964d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // Promote S up to the canonical IV type, if the cast is foldable. 14970bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *NewS = S; 14986ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType()); 14994d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman if (isa<SCEVAddRecExpr>(Ext)) 15004d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman NewS = Ext; 15014d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman 15020bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE); 1503e81561909d128c6e2d8033cb5465a49b2596b26aBill Wendling //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n"; 150436f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 15054d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // Truncate the result down to the original type, if needed. 15060bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *T = SE.getTruncateOrNoop(V, Ty); 1507469f3cdc13851914f3a766cbd8f701cf8431cacaDan Gohman return expand(T); 150836f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman} 150996fea337d27357e9b62abbf3d2d5ce29f1c8e870Anton Korobeynikov 1510890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) { 1511db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 151292fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *V = expandCodeFor(S->getOperand(), 151392fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman SE.getEffectiveSCEVType(S->getOperand()->getType())); 1514a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Value *I = Builder.CreateTrunc(V, Ty); 1515a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(I); 1516cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman return I; 151711f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman} 151811f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman 1519890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) { 1520db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 152192fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *V = expandCodeFor(S->getOperand(), 152292fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman SE.getEffectiveSCEVType(S->getOperand()->getType())); 1523a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Value *I = Builder.CreateZExt(V, Ty); 1524a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(I); 1525cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman return I; 152611f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman} 152711f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman 1528890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) { 1529db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 153092fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *V = expandCodeFor(S->getOperand(), 153192fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman SE.getEffectiveSCEVType(S->getOperand()->getType())); 1532a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Value *I = Builder.CreateSExt(V, Ty); 1533a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(I); 1534cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman return I; 153511f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman} 153611f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman 1537890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) { 15380196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1539db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = LHS->getType(); 15400196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman for (int i = S->getNumOperands()-2; i >= 0; --i) { 15410196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // In the case of mixed integer and pointer types, do the 15420196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // rest of the comparisons as integer. 15430196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman if (S->getOperand(i)->getType() != Ty) { 15440196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman Ty = SE.getEffectiveSCEVType(Ty); 15450196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman LHS = InsertNoopCastOfTo(LHS, Ty); 15460196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman } 154792fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1548a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Value *ICmp = Builder.CreateICmpSGT(LHS, RHS); 1549a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(ICmp); 1550267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax"); 1551a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(Sel); 1552cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman LHS = Sel; 1553c54c561c9f7270c055dd7ba75a3a003b771a42d9Nick Lewycky } 15540196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // In the case of mixed integer and pointer types, cast the 15550196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // final result back to the pointer type. 15560196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman if (LHS->getType() != S->getType()) 15570196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman LHS = InsertNoopCastOfTo(LHS, S->getType()); 1558c54c561c9f7270c055dd7ba75a3a003b771a42d9Nick Lewycky return LHS; 1559c54c561c9f7270c055dd7ba75a3a003b771a42d9Nick Lewycky} 1560c54c561c9f7270c055dd7ba75a3a003b771a42d9Nick Lewycky 1561890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) { 15620196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1563db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = LHS->getType(); 15640196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman for (int i = S->getNumOperands()-2; i >= 0; --i) { 15650196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // In the case of mixed integer and pointer types, do the 15660196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // rest of the comparisons as integer. 15670196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman if (S->getOperand(i)->getType() != Ty) { 15680196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman Ty = SE.getEffectiveSCEVType(Ty); 15690196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman LHS = InsertNoopCastOfTo(LHS, Ty); 15700196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman } 157192fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1572a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Value *ICmp = Builder.CreateICmpUGT(LHS, RHS); 1573a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(ICmp); 1574267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax"); 1575a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(Sel); 1576cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman LHS = Sel; 15773e6307698084e7adfc10b739442ae29742beefd0Nick Lewycky } 15780196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // In the case of mixed integer and pointer types, cast the 15790196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // final result back to the pointer type. 15800196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman if (LHS->getType() != S->getType()) 15810196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman LHS = InsertNoopCastOfTo(LHS, S->getType()); 15823e6307698084e7adfc10b739442ae29742beefd0Nick Lewycky return LHS; 15833e6307698084e7adfc10b739442ae29742beefd0Nick Lewycky} 15843e6307698084e7adfc10b739442ae29742beefd0Nick Lewycky 1585db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris LattnerValue *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty, 1586b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick Instruction *IP) { 1587cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar assert(IP); 1588cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Builder.SetInsertPoint(IP); 15896c7ed6b54949949806797bafdf545fbfecb2cef5Dan Gohman return expandCodeFor(SH, Ty); 15906c7ed6b54949949806797bafdf545fbfecb2cef5Dan Gohman} 15916c7ed6b54949949806797bafdf545fbfecb2cef5Dan Gohman 1592db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris LattnerValue *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) { 159311f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman // Expand the code for this SCEV. 15942d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman Value *V = expand(SH); 15955be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (Ty) { 15965be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) && 15975be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman "non-trivial casts should be done with the SCEVs directly!"); 15985be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman V = InsertNoopCastOfTo(V, Ty); 15995be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 16005be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman return V; 160111f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman} 160211f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman 1603890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::expand(const SCEV *S) { 160440a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // Compute an insertion point for this SCEV object. Hoist the instructions 160540a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // as far out in the loop nest as possible. 1606cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Instruction *InsertPt = &*Builder.GetInsertPoint(); 1607cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar for (Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock());; 160840a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman L = L->getParentLoop()) 160917ead4ff4baceb2c5503f233d0288d363ae44165Dan Gohman if (SE.isLoopInvariant(S, L)) { 161040a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman if (!L) break; 1611e059ee832ca36d65a5fb87b0ac5bcdb0490b15cbDan Gohman if (BasicBlock *Preheader = L->getLoopPreheader()) 161240a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman InsertPt = Preheader->getTerminator(); 16130f8cd56bfdd32af4edb253654db02fb3143b25a8Andrew Trick else { 16140f8cd56bfdd32af4edb253654db02fb3143b25a8Andrew Trick // LSR sets the insertion point for AddRec start/step values to the 16150f8cd56bfdd32af4edb253654db02fb3143b25a8Andrew Trick // block start to simplify value reuse, even though it's an invalid 16160f8cd56bfdd32af4edb253654db02fb3143b25a8Andrew Trick // position. SCEVExpander must correct for this in all cases. 1617cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar InsertPt = &*L->getHeader()->getFirstInsertionPt(); 16180f8cd56bfdd32af4edb253654db02fb3143b25a8Andrew Trick } 161940a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman } else { 162040a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // If the SCEV is computable at this level, insert it into the header 162140a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // after the PHIs (and after any other instructions that we've inserted 162240a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // there) so that it is guaranteed to dominate any user inside the loop. 16235b6f42f57e730c2d968c313a27fa505a3c3e5efaBill Wendling if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L)) 1624cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar InsertPt = &*L->getHeader()->getFirstInsertionPt(); 1625b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick while (InsertPt != Builder.GetInsertPoint() 1626b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick && (isInsertedInstruction(InsertPt) 1627b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick || isa<DbgInfoIntrinsic>(InsertPt))) { 1628cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar InsertPt = &*std::next(InsertPt->getIterator()); 1629b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 163040a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman break; 163140a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman } 163240a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman 1633667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman // Check to see if we already expanded this here. 1634cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar auto I = InsertedExpressions.find(std::make_pair(S, InsertPt)); 1635267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman if (I != InsertedExpressions.end()) 1636667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman return I->second; 1637267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman 1638d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer BuilderType::InsertPointGuard Guard(Builder); 1639cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Builder.SetInsertPoint(InsertPt); 1640667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman 1641667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman // Expand the expression into instructions. 164296fea337d27357e9b62abbf3d2d5ce29f1c8e870Anton Korobeynikov Value *V = visit(S); 164340a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman 1644667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman // Remember the expanded value for this SCEV at this location. 164548ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // 164648ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // This is independent of PostIncLoops. The mapped value simply materializes 164748ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // the expression at this insertion point. If the mapped value happened to be 164836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // a postinc expansion, it could be reused by a non-postinc user, but only if 164948ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // its insertion point was already at the head of the loop. 165048ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick InsertedExpressions[std::make_pair(S, InsertPt)] = V; 165196fea337d27357e9b62abbf3d2d5ce29f1c8e870Anton Korobeynikov return V; 165296fea337d27357e9b62abbf3d2d5ce29f1c8e870Anton Korobeynikov} 16531d09de3eca23267855e28297fcb40de3632ea47bDan Gohman 16541d826a76f591afea445489b9a5485c345e66bf87Dan Gohmanvoid SCEVExpander::rememberInstruction(Value *I) { 165525fcaff409f5c4c6da08f148ffb9404a71e8e4a7Dan Gohman if (!PostIncLoops.empty()) 165625fcaff409f5c4c6da08f148ffb9404a71e8e4a7Dan Gohman InsertedPostIncValues.insert(I); 165725fcaff409f5c4c6da08f148ffb9404a71e8e4a7Dan Gohman else 16581d826a76f591afea445489b9a5485c345e66bf87Dan Gohman InsertedValues.insert(I); 16591d826a76f591afea445489b9a5485c345e66bf87Dan Gohman} 16601d826a76f591afea445489b9a5485c345e66bf87Dan Gohman 16611d09de3eca23267855e28297fcb40de3632ea47bDan Gohman/// getOrInsertCanonicalInductionVariable - This method returns the 16621d09de3eca23267855e28297fcb40de3632ea47bDan Gohman/// canonical induction variable of the specified type for the specified 16631d09de3eca23267855e28297fcb40de3632ea47bDan Gohman/// loop (inserting one if there is none). A canonical induction variable 16641d09de3eca23267855e28297fcb40de3632ea47bDan Gohman/// starts at zero and steps by one on each iteration. 16657c58dbd88c36c5d6c411ea6c046ddcff4c5841e9Dan GohmanPHINode * 16661d09de3eca23267855e28297fcb40de3632ea47bDan GohmanSCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L, 1667db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty) { 1668b0bc6c361da9009e8414efde317d9bbff755f6c0Duncan Sands assert(Ty->isIntegerTy() && "Can only insert integer induction variables!"); 1669133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman 1670133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman // Build a SCEV for {0,+,1}<L>. 16713228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick // Conservatively use FlagAnyWrap for now. 1672deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0), 16733228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap); 1674133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman 1675133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman // Emit code for it. 1676d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer BuilderType::InsertPointGuard Guard(Builder); 1677cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar PHINode *V = 1678cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar cast<PHINode>(expandCodeFor(H, nullptr, &L->getHeader()->front())); 1679133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman 168040a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman return V; 16811d09de3eca23267855e28297fcb40de3632ea47bDan Gohman} 1682204494149b6f846e8f173f525b129f5508076049Andrew Trick 1683204494149b6f846e8f173f525b129f5508076049Andrew Trick/// replaceCongruentIVs - Check for congruent phis in this loop header and 1684204494149b6f846e8f173f525b129f5508076049Andrew Trick/// replace them with their most canonical representative. Return the number of 1685204494149b6f846e8f173f525b129f5508076049Andrew Trick/// phis eliminated. 1686204494149b6f846e8f173f525b129f5508076049Andrew Trick/// 1687204494149b6f846e8f173f525b129f5508076049Andrew Trick/// This does not depend on any SCEVExpander state but should be used in 1688204494149b6f846e8f173f525b129f5508076049Andrew Trick/// the same context that SCEVExpander is used. 1689204494149b6f846e8f173f525b129f5508076049Andrew Trickunsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT, 1690a04a4a79ea365d1ba96ed4b5861e879b267162e2Nadav Rotem SmallVectorImpl<WeakVH> &DeadInsts, 1691e4ba75f43e2ab1480d119d2d4eb878256274e0fbChandler Carruth const TargetTransformInfo *TTI) { 1692ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // Find integer phis in order of increasing width. 1693ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick SmallVector<PHINode*, 8> Phis; 1694cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar for (auto &I : *L->getHeader()) { 1695cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (auto *PN = dyn_cast<PHINode>(&I)) 1696cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Phis.push_back(PN); 1697cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar else 1698cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar break; 1699ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick } 1700cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 1701e4ba75f43e2ab1480d119d2d4eb878256274e0fbChandler Carruth if (TTI) 170236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines std::sort(Phis.begin(), Phis.end(), [](Value *LHS, Value *RHS) { 170336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Put pointers at the back and make sure pointer < pointer = false. 170436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (!LHS->getType()->isIntegerTy() || !RHS->getType()->isIntegerTy()) 170536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines return RHS->getType()->isIntegerTy() && !LHS->getType()->isIntegerTy(); 170636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines return RHS->getType()->getPrimitiveSizeInBits() < 170736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines LHS->getType()->getPrimitiveSizeInBits(); 170836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines }); 1709ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick 1710204494149b6f846e8f173f525b129f5508076049Andrew Trick unsigned NumElim = 0; 1711204494149b6f846e8f173f525b129f5508076049Andrew Trick DenseMap<const SCEV *, PHINode *> ExprToIVMap; 1712cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar // Process phis from wide to narrow. Map wide phis to their truncation 1713ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // so narrow phis can reuse them. 1714cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar for (PHINode *Phi : Phis) { 1715cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar auto SimplifyPHINode = [&](PHINode *PN) -> Value * { 1716cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (Value *V = SimplifyInstruction(PN, DL, &SE.TLI, &SE.DT, &SE.AC)) 1717cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return V; 1718cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (!SE.isSCEVable(PN->getType())) 1719cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return nullptr; 1720cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar auto *Const = dyn_cast<SCEVConstant>(SE.getSCEV(PN)); 1721cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (!Const) 1722cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return nullptr; 1723cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return Const->getValue(); 1724cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar }; 1725ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick 1726239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer // Fold constant phis. They may be congruent to other constant phis and 1727239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer // would confuse the logic below that expects proper IVs. 1728cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (Value *V = SimplifyPHINode(Phi)) { 1729cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (V->getType() != Phi->getType()) 1730cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar continue; 1731239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer Phi->replaceAllUsesWith(V); 17326948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar DeadInsts.emplace_back(Phi); 1733239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer ++NumElim; 1734239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer DEBUG_WITH_TYPE(DebugType, dbgs() 1735239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer << "INDVARS: Eliminated constant iv: " << *Phi << '\n'); 1736239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer continue; 1737239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer } 1738239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer 1739204494149b6f846e8f173f525b129f5508076049Andrew Trick if (!SE.isSCEVable(Phi->getType())) 1740204494149b6f846e8f173f525b129f5508076049Andrew Trick continue; 1741204494149b6f846e8f173f525b129f5508076049Andrew Trick 1742204494149b6f846e8f173f525b129f5508076049Andrew Trick PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)]; 1743204494149b6f846e8f173f525b129f5508076049Andrew Trick if (!OrigPhiRef) { 1744204494149b6f846e8f173f525b129f5508076049Andrew Trick OrigPhiRef = Phi; 1745e4ba75f43e2ab1480d119d2d4eb878256274e0fbChandler Carruth if (Phi->getType()->isIntegerTy() && TTI 1746e4ba75f43e2ab1480d119d2d4eb878256274e0fbChandler Carruth && TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) { 1747ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // This phi can be freely truncated to the narrowest phi type. Map the 1748ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // truncated expression to it so it will be reused for narrow types. 1749ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick const SCEV *TruncExpr = 1750ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType()); 1751ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick ExprToIVMap[TruncExpr] = Phi; 1752ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick } 1753204494149b6f846e8f173f525b129f5508076049Andrew Trick continue; 1754204494149b6f846e8f173f525b129f5508076049Andrew Trick } 1755204494149b6f846e8f173f525b129f5508076049Andrew Trick 1756ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // Replacing a pointer phi with an integer phi or vice-versa doesn't make 1757ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // sense. 1758ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy()) 1759204494149b6f846e8f173f525b129f5508076049Andrew Trick continue; 1760204494149b6f846e8f173f525b129f5508076049Andrew Trick 1761204494149b6f846e8f173f525b129f5508076049Andrew Trick if (BasicBlock *LatchBlock = L->getLoopLatch()) { 1762204494149b6f846e8f173f525b129f5508076049Andrew Trick Instruction *OrigInc = 1763204494149b6f846e8f173f525b129f5508076049Andrew Trick cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock)); 1764204494149b6f846e8f173f525b129f5508076049Andrew Trick Instruction *IsomorphicInc = 1765204494149b6f846e8f173f525b129f5508076049Andrew Trick cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock)); 1766204494149b6f846e8f173f525b129f5508076049Andrew Trick 1767ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // If this phi has the same width but is more canonical, replace the 1768b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // original with it. As part of the "more canonical" determination, 1769b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // respect a prior decision to use an IV chain. 1770ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick if (OrigPhiRef->getType() == Phi->getType() 1771b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick && !(ChainedPhis.count(Phi) 1772b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L)) 1773b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick && (ChainedPhis.count(Phi) 1774b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) { 1775204494149b6f846e8f173f525b129f5508076049Andrew Trick std::swap(OrigPhiRef, Phi); 1776204494149b6f846e8f173f525b129f5508076049Andrew Trick std::swap(OrigInc, IsomorphicInc); 1777204494149b6f846e8f173f525b129f5508076049Andrew Trick } 1778204494149b6f846e8f173f525b129f5508076049Andrew Trick // Replacing the congruent phi is sufficient because acyclic redundancy 1779204494149b6f846e8f173f525b129f5508076049Andrew Trick // elimination, CSE/GVN, should handle the rest. However, once SCEV proves 1780204494149b6f846e8f173f525b129f5508076049Andrew Trick // that a phi is congruent, it's often the head of an IV user cycle that 1781139f333f88fb9e0c39dfa3b56fd2a26b33837b70Andrew Trick // is isomorphic with the original phi. It's worth eagerly cleaning up the 1782139f333f88fb9e0c39dfa3b56fd2a26b33837b70Andrew Trick // common case of a single IV increment so that DeleteDeadPHIs can remove 1783139f333f88fb9e0c39dfa3b56fd2a26b33837b70Andrew Trick // cycles that had postinc uses. 1784ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc), 1785ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick IsomorphicInc->getType()); 1786ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick if (OrigInc != IsomorphicInc 178764925c55c65f9345a69fb67db07aa62cfb723577Andrew Trick && TruncExpr == SE.getSCEV(IsomorphicInc) 1788b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc)) 1789b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick || hoistIVInc(OrigInc, IsomorphicInc))) { 1790204494149b6f846e8f173f525b129f5508076049Andrew Trick DEBUG_WITH_TYPE(DebugType, dbgs() 1791204494149b6f846e8f173f525b129f5508076049Andrew Trick << "INDVARS: Eliminated congruent iv.inc: " 1792204494149b6f846e8f173f525b129f5508076049Andrew Trick << *IsomorphicInc << '\n'); 1793ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick Value *NewInc = OrigInc; 1794ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick if (OrigInc->getType() != IsomorphicInc->getType()) { 17954c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar Instruction *IP = nullptr; 17964c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar if (PHINode *PN = dyn_cast<PHINode>(OrigInc)) 1797cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar IP = &*PN->getParent()->getFirstInsertionPt(); 17984c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar else 17994c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar IP = OrigInc->getNextNode(); 18004c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar 1801dd1f22f25d8496b10cfddedb63d674dd39897bb0Andrew Trick IRBuilder<> Builder(IP); 1802ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc()); 1803ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick NewInc = Builder. 1804ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName); 1805ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick } 1806ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick IsomorphicInc->replaceAllUsesWith(NewInc); 18076948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar DeadInsts.emplace_back(IsomorphicInc); 1808204494149b6f846e8f173f525b129f5508076049Andrew Trick } 1809204494149b6f846e8f173f525b129f5508076049Andrew Trick } 1810204494149b6f846e8f173f525b129f5508076049Andrew Trick DEBUG_WITH_TYPE(DebugType, dbgs() 1811204494149b6f846e8f173f525b129f5508076049Andrew Trick << "INDVARS: Eliminated congruent iv: " << *Phi << '\n'); 1812204494149b6f846e8f173f525b129f5508076049Andrew Trick ++NumElim; 1813ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick Value *NewIV = OrigPhiRef; 1814ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick if (OrigPhiRef->getType() != Phi->getType()) { 1815cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar IRBuilder<> Builder(&*L->getHeader()->getFirstInsertionPt()); 1816ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick Builder.SetCurrentDebugLocation(Phi->getDebugLoc()); 1817ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName); 1818ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick } 1819ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick Phi->replaceAllUsesWith(NewIV); 18206948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar DeadInsts.emplace_back(Phi); 1821204494149b6f846e8f173f525b129f5508076049Andrew Trick } 1822204494149b6f846e8f173f525b129f5508076049Andrew Trick return NumElim; 1823204494149b6f846e8f173f525b129f5508076049Andrew Trick} 1824e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick 1825cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga NainarValue *SCEVExpander::findExistingExpansion(const SCEV *S, 1826cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar const Instruction *At, Loop *L) { 1827cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar using namespace llvm::PatternMatch; 1828cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 1829cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar SmallVector<BasicBlock *, 4> ExitingBlocks; 1830cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar L->getExitingBlocks(ExitingBlocks); 1831cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 1832cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar // Look for suitable value in simple conditions at the loop exits. 1833cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar for (BasicBlock *BB : ExitingBlocks) { 1834cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar ICmpInst::Predicate Pred; 1835cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Instruction *LHS, *RHS; 1836cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar BasicBlock *TrueBB, *FalseBB; 1837cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 1838cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (!match(BB->getTerminator(), 1839cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar m_Br(m_ICmp(Pred, m_Instruction(LHS), m_Instruction(RHS)), 1840cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar TrueBB, FalseBB))) 1841cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar continue; 1842cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 1843cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (SE.getSCEV(LHS) == S && SE.DT.dominates(LHS, At)) 1844cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return LHS; 1845cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 1846cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (SE.getSCEV(RHS) == S && SE.DT.dominates(RHS, At)) 1847cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return RHS; 1848cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar } 1849cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 1850cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar // There is potential to make this significantly smarter, but this simple 1851cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar // heuristic already gets some interesting cases. 1852cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 1853cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar // Can not find suitable value. 1854cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return nullptr; 1855cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar} 1856cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 18570c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainarbool SCEVExpander::isHighCostExpansionHelper( 1858cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar const SCEV *S, Loop *L, const Instruction *At, 1859cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar SmallPtrSetImpl<const SCEV *> &Processed) { 1860cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 1861cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar // If we can find an existing value for this scev avaliable at the point "At" 1862cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar // then consider the expression cheap. 1863cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (At && findExistingExpansion(S, At, L) != nullptr) 1864cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return false; 18656948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar 18666948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar // Zero/One operand expressions 18676948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar switch (S->getSCEVType()) { 18686948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar case scUnknown: 18696948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar case scConstant: 18706948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar return false; 18716948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar case scTruncate: 1872cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return isHighCostExpansionHelper(cast<SCEVTruncateExpr>(S)->getOperand(), 1873cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar L, At, Processed); 18746948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar case scZeroExtend: 18756948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar return isHighCostExpansionHelper(cast<SCEVZeroExtendExpr>(S)->getOperand(), 1876cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar L, At, Processed); 18776948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar case scSignExtend: 18786948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar return isHighCostExpansionHelper(cast<SCEVSignExtendExpr>(S)->getOperand(), 1879cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar L, At, Processed); 18806948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar } 18816948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar 18820c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar if (!Processed.insert(S).second) 18830c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar return false; 18840c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar 18850c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar if (auto *UDivExpr = dyn_cast<SCEVUDivExpr>(S)) { 18860c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // If the divisor is a power of two and the SCEV type fits in a native 1887cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar // integer, consider the division cheap irrespective of whether it occurs in 18880c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // the user code since it can be lowered into a right shift. 18890c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar if (auto *SC = dyn_cast<SCEVConstant>(UDivExpr->getRHS())) 1890cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (SC->getAPInt().isPowerOf2()) { 18910c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar const DataLayout &DL = 18920c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar L->getHeader()->getParent()->getParent()->getDataLayout(); 18930c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar unsigned Width = cast<IntegerType>(UDivExpr->getType())->getBitWidth(); 18940c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar return DL.isIllegalInteger(Width); 18950c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar } 18960c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar 18970c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // UDivExpr is very likely a UDiv that ScalarEvolution's HowFarToZero or 18980c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // HowManyLessThans produced to compute a precise expression, rather than a 18990c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // UDiv from the user's code. If we can't find a UDiv in the code with some 19000c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // simple searching, assume the former consider UDivExpr expensive to 19010c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // compute. 19020c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar BasicBlock *ExitingBB = L->getExitingBlock(); 19030c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar if (!ExitingBB) 19040c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar return true; 19050c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar 1906cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar // At the beginning of this function we already tried to find existing value 1907cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar // for plain 'S'. Now try to lookup 'S + 1' since it is common pattern 1908cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar // involving division. This is just a simple search heuristic. 1909cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (!At) 1910cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar At = &ExitingBB->back(); 1911cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (!findExistingExpansion( 1912cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar SE.getAddExpr(S, SE.getConstant(S->getType(), 1)), At, L)) 19130c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar return true; 19140c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar } 19150c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar 19166948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar // HowManyLessThans uses a Max expression whenever the loop is not guarded by 19176948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar // the exit condition. 19186948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S)) 19196948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar return true; 19206948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar 19216948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar // Recurse past nary expressions, which commonly occur in the 19220c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // BackedgeTakenCount. They may already exist in program code, and if not, 19230c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // they are not too expensive rematerialize. 19246948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar if (const SCEVNAryExpr *NAry = dyn_cast<SCEVNAryExpr>(S)) { 1925cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar for (auto *Op : NAry->operands()) 1926cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar if (isHighCostExpansionHelper(Op, L, At, Processed)) 19270c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar return true; 19280c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar } 19290c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar 19300c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // If we haven't recognized an expensive SCEV pattern, assume it's an 19310c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // expression produced by program code. 19320c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar return false; 19330c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar} 19340c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar 1935cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga NainarValue *SCEVExpander::expandCodeForPredicate(const SCEVPredicate *Pred, 1936cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Instruction *IP) { 1937cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar assert(IP); 1938cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar switch (Pred->getKind()) { 1939cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar case SCEVPredicate::P_Union: 1940cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return expandUnionPredicate(cast<SCEVUnionPredicate>(Pred), IP); 1941cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar case SCEVPredicate::P_Equal: 1942cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return expandEqualPredicate(cast<SCEVEqualPredicate>(Pred), IP); 1943cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar } 1944cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar llvm_unreachable("Unknown SCEV predicate type"); 1945cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar} 1946cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 1947cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga NainarValue *SCEVExpander::expandEqualPredicate(const SCEVEqualPredicate *Pred, 1948cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Instruction *IP) { 1949cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Value *Expr0 = expandCodeFor(Pred->getLHS(), Pred->getLHS()->getType(), IP); 1950cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Value *Expr1 = expandCodeFor(Pred->getRHS(), Pred->getRHS()->getType(), IP); 1951cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 1952cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Builder.SetInsertPoint(IP); 1953cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar auto *I = Builder.CreateICmpNE(Expr0, Expr1, "ident.check"); 1954cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return I; 1955cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar} 1956cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 1957cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga NainarValue *SCEVExpander::expandUnionPredicate(const SCEVUnionPredicate *Union, 1958cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Instruction *IP) { 1959cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar auto *BoolType = IntegerType::get(IP->getContext(), 1); 1960cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Value *Check = ConstantInt::getNullValue(BoolType); 1961cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 1962cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar // Loop over all checks in this set. 1963cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar for (auto Pred : Union->getPredicates()) { 1964cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar auto *NextCheck = expandCodeForPredicate(Pred, IP); 1965cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Builder.SetInsertPoint(IP); 1966cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar Check = Builder.CreateOr(Check, NextCheck); 1967cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar } 1968cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 1969cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar return Check; 1970cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar} 1971cddc3e03e4ec99c0268c03a126195173e519ed58Pirama Arumuga Nainar 1972e08c32249fca32cd7b122024a4ca252fcb235694Andrew Tricknamespace { 1973e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// Search for a SCEV subexpression that is not safe to expand. Any expression 1974e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely 1975e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// UDiv expressions. We don't know if the UDiv is derived from an IR divide 1976e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// instruction, but the important thing is that we prove the denominator is 1977e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// nonzero before expansion. 1978e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// 1979e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// IVUsers already checks that IV-derived expressions are safe. So this check is 1980e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// only needed when the expression includes some subexpression that is not IV 1981e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// derived. 1982e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// 1983e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// Currently, we only allow division by a nonzero constant here. If this is 1984e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// inadequate, we could easily allow division by SCEVUnknown by using 1985e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// ValueTracking to check isKnownNonZero(). 19864d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick// 19874d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick// We cannot generally expand recurrences unless the step dominates the loop 19884d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick// header. The expander handles the special case of affine recurrences by 19894d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick// scaling the recurrence outside the loop, but this technique isn't generally 19904d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick// applicable. Expanding a nested recurrence outside a loop requires computing 19914d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick// binomial coefficients. This could be done, but the recurrence has to be in a 19924d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick// perfectly reduced form, which can't be guaranteed. 1993e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trickstruct SCEVFindUnsafe { 19944d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick ScalarEvolution &SE; 1995e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick bool IsUnsafe; 1996e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick 19974d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick SCEVFindUnsafe(ScalarEvolution &se): SE(se), IsUnsafe(false) {} 1998e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick 1999e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick bool follow(const SCEV *S) { 20004d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 20014d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS()); 20024d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick if (!SC || SC->getValue()->isZero()) { 20034d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick IsUnsafe = true; 20044d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick return false; 20054d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick } 20064d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick } 20074d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 20084d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick const SCEV *Step = AR->getStepRecurrence(SE); 20094d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick if (!AR->isAffine() && !SE.dominates(Step, AR->getLoop()->getHeader())) { 20104d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick IsUnsafe = true; 20114d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick return false; 20124d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick } 20134d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick } 20144d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick return true; 2015e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick } 2016e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick bool isDone() const { return IsUnsafe; } 2017e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick}; 2018e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick} 2019e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick 2020e08c32249fca32cd7b122024a4ca252fcb235694Andrew Tricknamespace llvm { 20214d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trickbool isSafeToExpand(const SCEV *S, ScalarEvolution &SE) { 20224d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick SCEVFindUnsafe Search(SE); 2023e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick visitAll(S, Search); 2024e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick return !Search.IsUnsafe; 2025e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick} 2026e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick} 2027