1de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar//===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis ------------===// 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" 27f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama 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; 32f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama 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. 66f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama 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) 78f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama 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). 83f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar assert(SE.DT.dominates(Ret, &*BIP)); 8423b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola 8523b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola rememberInstruction(Ret); 8623b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola return Ret; 87485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman} 88485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman 89f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainarstatic BasicBlock::iterator findInsertPointAfter(Instruction *I, 90f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar BasicBlock *MustDominate) { 91f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar BasicBlock::iterator IP = ++I->getIterator(); 92f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (auto *II = dyn_cast<InvokeInst>(I)) 93f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar IP = II->getNormalDest()->begin(); 94f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 95f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar while (isa<PHINode>(IP)) 96f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar ++IP; 97f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 98de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (isa<FuncletPadInst>(IP) || isa<LandingPadInst>(IP)) { 99de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar ++IP; 100de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar } else if (isa<CatchSwitchInst>(IP)) { 101de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar IP = MustDominate->getFirstInsertionPt(); 102de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar } else { 103de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar assert(!IP->isEHPad() && "unexpected eh pad!"); 104f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar } 105f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 106f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return IP; 107f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar} 108f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 109267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman/// InsertNoopCastOfTo - Insert a cast of V to the specified type, 110267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman/// which must be possible with a noop cast, doing what we can to share 111267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman/// the casts. 112db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris LattnerValue *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) { 113267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false); 114267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman assert((Op == Instruction::BitCast || 115267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Op == Instruction::PtrToInt || 116267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Op == Instruction::IntToPtr) && 117267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman "InsertNoopCastOfTo cannot perform non-noop casts!"); 118267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) && 119267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman "InsertNoopCastOfTo cannot change sizes!"); 120267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman 1212d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman // Short-circuit unnecessary bitcasts. 12219154f457696f286b4a597179fa07e3b38ea0310Andrew Trick if (Op == Instruction::BitCast) { 12319154f457696f286b4a597179fa07e3b38ea0310Andrew Trick if (V->getType() == Ty) 12419154f457696f286b4a597179fa07e3b38ea0310Andrew Trick return V; 12519154f457696f286b4a597179fa07e3b38ea0310Andrew Trick if (CastInst *CI = dyn_cast<CastInst>(V)) { 12619154f457696f286b4a597179fa07e3b38ea0310Andrew Trick if (CI->getOperand(0)->getType() == Ty) 12719154f457696f286b4a597179fa07e3b38ea0310Andrew Trick return CI->getOperand(0); 12819154f457696f286b4a597179fa07e3b38ea0310Andrew Trick } 12919154f457696f286b4a597179fa07e3b38ea0310Andrew Trick } 130f04fa483b83227c570bc58e1684ea096430a6697Dan Gohman // Short-circuit unnecessary inttoptr<->ptrtoint casts. 131267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) && 13280dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) { 133af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman if (CastInst *CI = dyn_cast<CastInst>(V)) 134af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman if ((CI->getOpcode() == Instruction::PtrToInt || 135af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman CI->getOpcode() == Instruction::IntToPtr) && 136af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman SE.getTypeSizeInBits(CI->getType()) == 137af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman SE.getTypeSizeInBits(CI->getOperand(0)->getType())) 138af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman return CI->getOperand(0); 13980dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 14080dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman if ((CE->getOpcode() == Instruction::PtrToInt || 14180dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman CE->getOpcode() == Instruction::IntToPtr) && 14280dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman SE.getTypeSizeInBits(CE->getType()) == 14380dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman SE.getTypeSizeInBits(CE->getOperand(0)->getType())) 14480dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman return CE->getOperand(0); 14580dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman } 146f04fa483b83227c570bc58e1684ea096430a6697Dan Gohman 147485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // Fold a cast of a constant. 148ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner if (Constant *C = dyn_cast<Constant>(V)) 149baf3c404409d5e47b13984a7f95bfbd6d1f2e79eOwen Anderson return ConstantExpr::getCast(Op, C, Ty); 1504c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman 151485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // Cast the argument at the beginning of the entry block, after 152485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // any bitcasts of other arguments. 153ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner if (Argument *A = dyn_cast<Argument>(V)) { 154485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin(); 155485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman while ((isa<BitCastInst>(IP) && 156485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) && 157485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman cast<BitCastInst>(IP)->getOperand(0) != A) || 158f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar isa<DbgInfoIntrinsic>(IP)) 159485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman ++IP; 160485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman return ReuseOrCreateCast(A, Ty, Op, IP); 161ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner } 1623913187bf6135e6a71260c65eed664095c8f9ce9Wojciech Matyjewicz 163485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // Cast the instruction immediately after the instruction. 164ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner Instruction *I = cast<Instruction>(V); 165f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar BasicBlock::iterator IP = findInsertPointAfter(I, Builder.GetInsertBlock()); 166485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman return ReuseOrCreateCast(I, Ty, Op, IP); 167ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner} 168ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner 1697fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner/// InsertBinop - Insert the specified binary operator, doing a small amount 1707fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner/// of work to avoid inserting an obviously redundant operation. 171267a385342f2e7388f178b327dd87c5f29afd51bDan GohmanValue *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, 172267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Value *LHS, Value *RHS) { 1730f0eb18addc94bf3d8179fbce162c8c4622b9866Dan Gohman // Fold a binop with constant operands. 1740f0eb18addc94bf3d8179fbce162c8c4622b9866Dan Gohman if (Constant *CLHS = dyn_cast<Constant>(LHS)) 1750f0eb18addc94bf3d8179fbce162c8c4622b9866Dan Gohman if (Constant *CRHS = dyn_cast<Constant>(RHS)) 176baf3c404409d5e47b13984a7f95bfbd6d1f2e79eOwen Anderson return ConstantExpr::get(Opcode, CLHS, CRHS); 1770f0eb18addc94bf3d8179fbce162c8c4622b9866Dan Gohman 1787fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner // Do a quick scan to see if we have this binop nearby. If so, reuse it. 1797fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner unsigned ScanLimit = 6; 180267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 181267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman // Scanning starts from the last instruction before the insertion point. 182267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock::iterator IP = Builder.GetInsertPoint(); 183267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman if (IP != BlockBegin) { 1848a08769bad43a22fae2845bb0ba0fd1266cd55c8Wojciech Matyjewicz --IP; 1858a08769bad43a22fae2845bb0ba0fd1266cd55c8Wojciech Matyjewicz for (; ScanLimit; --IP, --ScanLimit) { 1868d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen // Don't count dbg.value against the ScanLimit, to avoid perturbing the 1878d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen // generated code. 1888d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen if (isa<DbgInfoIntrinsic>(IP)) 1898d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen ScanLimit++; 1905be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS && 1915be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman IP->getOperand(1) == RHS) 192f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return &*IP; 1938a08769bad43a22fae2845bb0ba0fd1266cd55c8Wojciech Matyjewicz if (IP == BlockBegin) break; 1948a08769bad43a22fae2845bb0ba0fd1266cd55c8Wojciech Matyjewicz } 1957fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner } 196267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman 197087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Save the original insertion point so we can restore it when we're done. 198d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc(); 199de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar SCEVInsertPointGuard Guard(Builder, this); 200087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 201087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Move the insertion point out of as many loops as we can. 202f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) { 203087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break; 204087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman BasicBlock *Preheader = L->getLoopPreheader(); 205087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!Preheader) break; 206087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 207087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Ok, move up a level. 208f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Builder.SetInsertPoint(Preheader->getTerminator()); 209087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 210087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 2118a08769bad43a22fae2845bb0ba0fd1266cd55c8Wojciech Matyjewicz // If we haven't found this binop, insert it. 212a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS)); 213d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer BO->setDebugLoc(Loc); 214a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(BO); 215087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 216cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman return BO; 2177fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner} 2187fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner 2194a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman/// FactorOutConstant - Test if S is divisible by Factor, using signed 220453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// division. If so, update S with Factor divided out and return true. 2213f46a3abeedba8d517b4182de34c821d752db058Dan Gohman/// S need not be evenly divisible if a reasonable remainder can be 2224a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman/// computed. 223453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made 224453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// unnecessary; in its place, just signed-divide Ops[i] by the scale and 225453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// check to see if the divide was folded. 2264c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainarstatic bool FactorOutConstant(const SCEV *&S, const SCEV *&Remainder, 2274c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar const SCEV *Factor, ScalarEvolution &SE, 2284c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar const DataLayout &DL) { 229453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // Everything is divisible by one. 230c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (Factor->isOne()) 231c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman return true; 232c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 233c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // x/x == 1. 234c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (S == Factor) { 235deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman S = SE.getConstant(S->getType(), 1); 236453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman return true; 237c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 238453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 239453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // For a Constant, check for a multiple of the given factor. 2404a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 241c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // 0/x == 0. 242c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (C->isZero()) 243453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman return true; 244c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Check for divisibility. 245c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) { 246c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ConstantInt *CI = 247f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar ConstantInt::get(SE.getContext(), C->getAPInt().sdiv(FC->getAPInt())); 248c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // If the quotient is zero and the remainder is non-zero, reject 249c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // the value at this scale. It will be considered for subsequent 250c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // smaller scales. 251c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (!CI->isZero()) { 252c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *Div = SE.getConstant(CI); 253c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman S = Div; 254f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Remainder = SE.getAddExpr( 255f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Remainder, SE.getConstant(C->getAPInt().srem(FC->getAPInt()))); 256c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman return true; 257c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 258453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 2594a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman } 260453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 261453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // In a Mul, check if there is a constant operand which is a multiple 262453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // of the given factor. 263c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) { 2644c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar // Size is known, check if there is a constant operand which is a multiple 2654c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar // of the given factor. If so, we can factor it. 2664c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar const SCEVConstant *FC = cast<SCEVConstant>(Factor); 2674c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0))) 268f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (!C->getAPInt().srem(FC->getAPInt())) { 2694c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 270f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar NewMulOps[0] = SE.getConstant(C->getAPInt().sdiv(FC->getAPInt())); 2714c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar S = SE.getMulExpr(NewMulOps); 2724c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar return true; 273453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 274c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 275453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 276453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // In an AddRec, check if both start and step are divisible. 277453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) { 2780bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *Step = A->getStepRecurrence(SE); 279deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman const SCEV *StepRem = SE.getConstant(Step->getType(), 0); 28036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (!FactorOutConstant(Step, StepRem, Factor, SE, DL)) 2814a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman return false; 2824a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman if (!StepRem->isZero()) 2834a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman return false; 2840bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *Start = A->getStart(); 28536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (!FactorOutConstant(Start, Remainder, Factor, SE, DL)) 286453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman return false; 2876f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick S = SE.getAddRecExpr(Start, Step, A->getLoop(), 2886f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick A->getNoWrapFlags(SCEV::FlagNW)); 289453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman return true; 290453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 291453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 292453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman return false; 293453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman} 294453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 295c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs 296c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// is the number of SCEVAddRecExprs present, which are kept at the end of 297c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// the list. 298c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// 299c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohmanstatic void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops, 300db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty, 301c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ScalarEvolution &SE) { 302c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman unsigned NumAddRecs = 0; 303c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i) 304c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ++NumAddRecs; 305c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Group Ops into non-addrecs and addrecs. 306c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs); 307c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end()); 308c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Let ScalarEvolution sort and simplify the non-addrecs list. 309c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *Sum = NoAddRecs.empty() ? 310deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman SE.getConstant(Ty, 0) : 311c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SE.getAddExpr(NoAddRecs); 312c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // If it returned an add, use the operands. Otherwise it simplified 313c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // the sum into a single value, so just use that. 314f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman Ops.clear(); 315c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum)) 316403a8cdda5e76ea689693de16474650b4b0df818Dan Gohman Ops.append(Add->op_begin(), Add->op_end()); 317f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman else if (!Sum->isZero()) 318f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman Ops.push_back(Sum); 319c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Then append the addrecs. 320403a8cdda5e76ea689693de16474650b4b0df818Dan Gohman Ops.append(AddRecs.begin(), AddRecs.end()); 321c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman} 322c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 323c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// SplitAddRecs - Flatten a list of add operands, moving addrec start values 324c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}. 325c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// This helps expose more opportunities for folding parts of the expressions 326c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// into GEP indices. 327c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// 328c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohmanstatic void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops, 329db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty, 330c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ScalarEvolution &SE) { 331c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Find the addrecs. 332c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SmallVector<const SCEV *, 8> AddRecs; 333c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman for (unsigned i = 0, e = Ops.size(); i != e; ++i) 334c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) { 335c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *Start = A->getStart(); 336c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (Start->isZero()) break; 337deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman const SCEV *Zero = SE.getConstant(Ty, 0); 338c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman AddRecs.push_back(SE.getAddRecExpr(Zero, 339c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman A->getStepRecurrence(SE), 3403228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick A->getLoop(), 3416f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick A->getNoWrapFlags(SCEV::FlagNW))); 342c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) { 343c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman Ops[i] = Zero; 344403a8cdda5e76ea689693de16474650b4b0df818Dan Gohman Ops.append(Add->op_begin(), Add->op_end()); 345c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman e += Add->getNumOperands(); 346c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } else { 347c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman Ops[i] = Start; 348c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 349c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 350c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (!AddRecs.empty()) { 351c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Add the addrecs onto the end of the list. 352403a8cdda5e76ea689693de16474650b4b0df818Dan Gohman Ops.append(AddRecs.begin(), AddRecs.end()); 353c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Resort the operand list, moving any constants to the front. 354c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SimplifyAddOperands(Ops, Ty, SE); 355c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 356c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman} 357c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 3584c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// expandAddToGEP - Expand an addition expression with a pointer type into 3594c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps 3604c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// BasicAliasAnalysis and other passes analyze the result. See the rules 3614c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// for getelementptr vs. inttoptr in 3624c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// http://llvm.org/docs/LangRef.html#pointeraliasing 3634c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// for details. 36413c5e35222afe0895f0c5e68aa9f22f134ea437aDan Gohman/// 3653abf905b50f33340ae81913da81b0eda96fa4616Dan Gohman/// Design note: The correctness of using getelementptr here depends on 3664c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as 3674c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// they may introduce pointer arithmetic which may not be safely converted 3684c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// into getelementptr. 369453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// 370453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// Design note: It might seem desirable for this function to be more 371453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// loop-aware. If some of the indices are loop-invariant while others 372453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// aren't, it might seem desirable to emit multiple GEPs, keeping the 373453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// loop-invariant portions of the overall computation outside the loop. 374453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// However, there are a few reasons this is not done here. Hoisting simple 375453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// arithmetic is a low-level optimization that often isn't very 376453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// important until late in the optimization process. In fact, passes 377453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// like InstructionCombining will combine GEPs, even if it means 378453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// pushing loop-invariant computation down into loops, so even if the 379453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// GEPs were split here, the work would quickly be undone. The 380453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// LoopStrengthReduction pass, which is usually run quite late (and 381453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// after the last InstructionCombining pass), takes care of hoisting 382453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// loop-invariant portions of expressions, after considering what 383453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// can be folded using target addressing modes. 384453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// 3850bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan GohmanValue *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin, 3860bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *const *op_end, 387db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner PointerType *PTy, 388db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty, 3895be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman Value *V) { 3904c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar Type *OriginalElTy = PTy->getElementType(); 3914c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar Type *ElTy = OriginalElTy; 3925be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman SmallVector<Value *, 4> GepIndices; 3930bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman SmallVector<const SCEV *, 8> Ops(op_begin, op_end); 3945be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman bool AnyNonZeroIndices = false; 3955be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 396c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Split AddRecs up into parts as either of the parts may be usable 397c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // without the other. 398c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SplitAddRecs(Ops, Ty, SE); 399c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 4004c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar Type *IntPtrTy = DL.getIntPtrType(PTy); 40114807bd8c801f976c999e5a6699f31ee9642021aMatt Arsenault 402eb35699e642f2c021f04154bdb4bd90a1afb3baaBob Wilson // Descend down the pointer's type and attempt to convert the other 4035be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // operands into GEP indices, at each level. The first index in a GEP 4045be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // indexes into the array implied by the pointer operand; the rest of 4055be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // the indices index into the element or field type selected by the 4065be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // preceding index. 4075be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman for (;;) { 408c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // If the scale size is not 0, attempt to factor out a scale for 409c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // array indexing. 4100bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman SmallVector<const SCEV *, 8> ScaledOps; 411150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman if (ElTy->isSized()) { 41214807bd8c801f976c999e5a6699f31ee9642021aMatt Arsenault const SCEV *ElSize = SE.getSizeOfExpr(IntPtrTy, ElTy); 413150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman if (!ElSize->isZero()) { 414150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman SmallVector<const SCEV *, 8> NewOps; 415f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar for (const SCEV *Op : Ops) { 416deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman const SCEV *Remainder = SE.getConstant(Ty, 0); 4174c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar if (FactorOutConstant(Op, Remainder, ElSize, SE, DL)) { 418150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman // Op now has ElSize factored out. 419150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman ScaledOps.push_back(Op); 420150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman if (!Remainder->isZero()) 421150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman NewOps.push_back(Remainder); 422150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman AnyNonZeroIndices = true; 423150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman } else { 424150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman // The operand was not divisible, so add it to the list of operands 425150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman // we'll scan next iteration. 426f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar NewOps.push_back(Op); 427150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman } 428150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman } 429150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman // If we made any changes, update Ops. 430150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman if (!ScaledOps.empty()) { 431150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman Ops = NewOps; 432150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman SimplifyAddOperands(Ops, Ty, SE); 4335be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 434c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 4355be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 436c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 437c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Record the scaled array index for this level of the type. If 438c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // we didn't find any operands that could be factored, tentatively 439c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // assume that element zero was selected (since the zero offset 440c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // would obviously be folded away). 4415be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman Value *Scaled = ScaledOps.empty() ? 442a7235ea7245028a0723e8ab7fd011386b3900777Owen Anderson Constant::getNullValue(Ty) : 4435be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman expandCodeFor(SE.getAddExpr(ScaledOps), Ty); 4445be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman GepIndices.push_back(Scaled); 4455be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 4465be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // Collect struct field index operands. 447db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner while (StructType *STy = dyn_cast<StructType>(ElTy)) { 448c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman bool FoundFieldNo = false; 449c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // An empty struct has no fields. 450c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (STy->getNumElements() == 0) break; 4514c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar // Field offsets are known. See if a constant offset falls within any of 4524c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar // the struct fields. 4534c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar if (Ops.empty()) 4544c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar break; 4554c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0])) 4564c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar if (SE.getTypeSizeInBits(C->getType()) <= 64) { 4574c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar const StructLayout &SL = *DL.getStructLayout(STy); 4584c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar uint64_t FullOffset = C->getValue()->getZExtValue(); 4594c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar if (FullOffset < SL.getSizeInBytes()) { 4604c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar unsigned ElIdx = SL.getElementContainingOffset(FullOffset); 4614c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar GepIndices.push_back( 4624c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx)); 4634c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar ElTy = STy->getTypeAtIndex(ElIdx); 4644c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar Ops[0] = 4656de29f8d960505421d61c80cdb738e16720b6c0eDan Gohman SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx)); 4664c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar AnyNonZeroIndices = true; 4674c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar FoundFieldNo = true; 4680f5efe56258f8cd6ceff4d7955a5d80144cd9cb0Dan Gohman } 4694c5e43da7792f75567b693105cc53e3f1992ad98Pirama Arumuga Nainar } 470c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // If no struct field offsets were found, tentatively assume that 471c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // field zero was selected (since the zero offset would obviously 472c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // be folded away). 473c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (!FoundFieldNo) { 474c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ElTy = STy->getTypeAtIndex(0u); 475c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman GepIndices.push_back( 476c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman Constant::getNullValue(Type::getInt32Ty(Ty->getContext()))); 477c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 478c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 4795be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 480db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) 4815be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman ElTy = ATy->getElementType(); 482c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman else 483c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman break; 4845be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 4855be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 4863f46a3abeedba8d517b4182de34c821d752db058Dan Gohman // If none of the operands were convertible to proper GEP indices, cast 4875be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // the base to i8* and do an ugly getelementptr with that. It's still 4885be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // better than ptrtoint+arithmetic+inttoptr at least. 4895be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (!AnyNonZeroIndices) { 490c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Cast the base to i8*. 4915be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman V = InsertNoopCastOfTo(V, 492ac53a0b272452013124bfc70480aea5e41b60f40Duncan Sands Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace())); 493c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 494705b48d960dff1a96ac40d0cf932eb465b9a550aRafael Espindola assert(!isa<Instruction>(V) || 495f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar SE.DT.dominates(cast<Instruction>(V), &*Builder.GetInsertPoint())); 4964b04578d65e38cdb5077de2498889e4a174ccdfdRafael Espindola 497c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Expand the operands for a plain byte offset. 49892fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty); 4995be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 5005be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // Fold a GEP with constant operands. 5015be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (Constant *CLHS = dyn_cast<Constant>(V)) 5025be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (Constant *CRHS = dyn_cast<Constant>(Idx)) 5030c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar return ConstantExpr::getGetElementPtr(Type::getInt8Ty(Ty->getContext()), 5040c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar CLHS, CRHS); 5055be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 5065be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // Do a quick scan to see if we have this GEP nearby. If so, reuse it. 5075be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman unsigned ScanLimit = 6; 508267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 509267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman // Scanning starts from the last instruction before the insertion point. 510267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock::iterator IP = Builder.GetInsertPoint(); 511267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman if (IP != BlockBegin) { 5125be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman --IP; 5135be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman for (; ScanLimit; --IP, --ScanLimit) { 5148d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen // Don't count dbg.value against the ScanLimit, to avoid perturbing the 5158d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen // generated code. 5168d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen if (isa<DbgInfoIntrinsic>(IP)) 5178d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen ScanLimit++; 5185be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (IP->getOpcode() == Instruction::GetElementPtr && 5195be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman IP->getOperand(0) == V && IP->getOperand(1) == Idx) 520f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return &*IP; 5215be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (IP == BlockBegin) break; 5225be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 5235be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 5245be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 525087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Save the original insertion point so we can restore it when we're done. 526de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar SCEVInsertPointGuard Guard(Builder, this); 527087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 528087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Move the insertion point out of as many loops as we can. 529f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) { 530087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break; 531087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman BasicBlock *Preheader = L->getLoopPreheader(); 532087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!Preheader) break; 533087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 534087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Ok, move up a level. 535f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Builder.SetInsertPoint(Preheader->getTerminator()); 536087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 537087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 538c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Emit a GEP. 5390c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar Value *GEP = Builder.CreateGEP(Builder.getInt8Ty(), V, Idx, "uglygep"); 540a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(GEP); 541087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 5425be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman return GEP; 5435be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 5445be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 545de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar { 546de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar SCEVInsertPointGuard Guard(Builder, this); 547087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 548de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Move the insertion point out of as many loops as we can. 549de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) { 550de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (!L->isLoopInvariant(V)) break; 551f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 552de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar bool AnyIndexNotLoopInvariant = 553de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar std::any_of(GepIndices.begin(), GepIndices.end(), 554de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar [L](Value *Op) { return !L->isLoopInvariant(Op); }); 555087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 556de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (AnyIndexNotLoopInvariant) 557de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar break; 558087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 559de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar BasicBlock *Preheader = L->getLoopPreheader(); 560de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (!Preheader) break; 561087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 562de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Ok, move up a level. 563de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Builder.SetInsertPoint(Preheader->getTerminator()); 564de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar } 565087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 566de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Insert a pretty getelementptr. Note that this GEP is not marked inbounds, 567de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // because ScalarEvolution may have changed the address arithmetic to 568de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // compute a value which is beyond the end of the allocated object. 569de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *Casted = V; 570de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (V->getType() != PTy) 571de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Casted = InsertNoopCastOfTo(Casted, PTy); 572de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *GEP = Builder.CreateGEP(OriginalElTy, Casted, GepIndices, "scevgep"); 573de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Ops.push_back(SE.getUnknown(GEP)); 574de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar rememberInstruction(GEP); 575de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar } 576341562b9fb6574b1c9492e52cec24106cd31ce51Benjamin Kramer 5775be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman return expand(SE.getAddExpr(Ops)); 5785be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman} 5795be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 580087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman/// PickMostRelevantLoop - Given two loops pick the one that's most relevant for 581087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman/// SCEV expansion. If they are nested, this is the most nested. If they are 582087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman/// neighboring, pick the later. 583087bd1e3a12893873761736bf0f905a350e9e708Dan Gohmanstatic const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B, 584087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman DominatorTree &DT) { 585087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!A) return B; 586087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!B) return A; 587087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (A->contains(B)) return B; 588087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (B->contains(A)) return A; 589087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (DT.dominates(A->getHeader(), B->getHeader())) return B; 590087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (DT.dominates(B->getHeader(), A->getHeader())) return A; 591087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return A; // Arbitrarily break the tie. 592087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman} 593087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 5949c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman/// getRelevantLoop - Get the most relevant loop associated with the given 595087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman/// expression, according to PickMostRelevantLoop. 5969c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohmanconst Loop *SCEVExpander::getRelevantLoop(const SCEV *S) { 5979c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman // Test whether we've already computed the most relevant loop for this SCEV. 598f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar auto Pair = RelevantLoops.insert(std::make_pair(S, nullptr)); 5999c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman if (!Pair.second) 6009c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman return Pair.first->second; 6019c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman 602087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (isa<SCEVConstant>(S)) 6039c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman // A constant has no relevant loops. 604dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines return nullptr; 605087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 606087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (const Instruction *I = dyn_cast<Instruction>(U->getValue())) 607f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return Pair.first->second = SE.LI.getLoopFor(I->getParent()); 6089c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman // A non-instruction has no relevant loops. 609dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines return nullptr; 610087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 611087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) { 612dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines const Loop *L = nullptr; 613087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 614087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman L = AR->getLoop(); 615f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar for (const SCEV *Op : N->operands()) 616f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar L = PickMostRelevantLoop(L, getRelevantLoop(Op), SE.DT); 6179c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman return RelevantLoops[N] = L; 6189c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman } 6199c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) { 6209c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman const Loop *Result = getRelevantLoop(C->getOperand()); 6219c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman return RelevantLoops[C] = Result; 6229c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman } 6239c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 624f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar const Loop *Result = PickMostRelevantLoop( 625f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar getRelevantLoop(D->getLHS()), getRelevantLoop(D->getRHS()), SE.DT); 6269c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman return RelevantLoops[D] = Result; 627087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 628087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman llvm_unreachable("Unexpected SCEV type!"); 629087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman} 630c4f7ec85ecb760fff2b702c6deb06506b968ba4fDan Gohman 631b35798347ea87b8b6d36155b211016a7769f01abDan Gohmannamespace { 632b35798347ea87b8b6d36155b211016a7769f01abDan Gohman 633087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman/// LoopCompare - Compare loops by PickMostRelevantLoop. 634087bd1e3a12893873761736bf0f905a350e9e708Dan Gohmanclass LoopCompare { 635087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman DominatorTree &DT; 636087bd1e3a12893873761736bf0f905a350e9e708Dan Gohmanpublic: 637087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman explicit LoopCompare(DominatorTree &dt) : DT(dt) {} 638087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 639087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman bool operator()(std::pair<const Loop *, const SCEV *> LHS, 640087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman std::pair<const Loop *, const SCEV *> RHS) const { 641bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman // Keep pointer operands sorted at the end. 642bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman if (LHS.second->getType()->isPointerTy() != 643bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman RHS.second->getType()->isPointerTy()) 644bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman return LHS.second->getType()->isPointerTy(); 645bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman 646087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Compare loops with PickMostRelevantLoop. 647087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (LHS.first != RHS.first) 648087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first; 649087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 650087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // If one operand is a non-constant negative and the other is not, 651087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // put the non-constant negative on the right so that a sub can 652087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // be used instead of a negate and add. 653f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick if (LHS.second->isNonConstantNegative()) { 654f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick if (!RHS.second->isNonConstantNegative()) 655087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return false; 656f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick } else if (RHS.second->isNonConstantNegative()) 657087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return true; 658087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 659087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Otherwise they are equivalent according to this comparison. 660087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return false; 661c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 662087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman}; 6635be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 664b35798347ea87b8b6d36155b211016a7769f01abDan Gohman} 665b35798347ea87b8b6d36155b211016a7769f01abDan Gohman 666087bd1e3a12893873761736bf0f905a350e9e708Dan GohmanValue *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) { 667db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 668e24fa64d52330626553298f56ba5aa702624c282Dan Gohman 669087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Collect all the add operands in a loop, along with their associated loops. 670087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Iterate in reverse so that constants are emitted last, all else equal, and 671087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // so that pointer operands are inserted first, which the code below relies on 672087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // to form more involved GEPs. 673087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 674087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()), 675087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman E(S->op_begin()); I != E; ++I) 6769c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 677087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 678087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Sort by loop. Use a stable sort so that constants follow non-constants and 679087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // pointer operands precede non-pointer operands. 680f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(SE.DT)); 681087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 682087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Emit instructions to add all the operands. Hoist as much as possible 683087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // out of loops, and form meaningful getelementptrs where possible. 684dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines Value *Sum = nullptr; 685f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar for (auto I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E;) { 686087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman const Loop *CurLoop = I->first; 687087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman const SCEV *Op = I->second; 688087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!Sum) { 689087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // This is the first operand. Just expand it. 690087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = expand(Op); 691087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman ++I; 692db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) { 693087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // The running sum expression is a pointer. Try to form a getelementptr 694087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // at this level with that as the base. 695087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman SmallVector<const SCEV *, 4> NewOps; 696bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman for (; I != E && I->first == CurLoop; ++I) { 697bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman // If the operand is SCEVUnknown and not instructions, peek through 698bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman // it, to enable more of it to be folded into the GEP. 699bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman const SCEV *X = I->second; 700bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X)) 701bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman if (!isa<Instruction>(U->getValue())) 702bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman X = SE.getSCEV(U->getValue()); 703bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman NewOps.push_back(X); 704bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman } 705087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum); 706db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) { 707087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // The running sum is an integer, and there's a pointer at this level. 708f8d0578e4cbd5922696c92f5068c5513d8e8d60eDan Gohman // Try to form a getelementptr. If the running sum is instructions, 709f8d0578e4cbd5922696c92f5068c5513d8e8d60eDan Gohman // use a SCEVUnknown to avoid re-analyzing them. 710087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman SmallVector<const SCEV *, 4> NewOps; 711f8d0578e4cbd5922696c92f5068c5513d8e8d60eDan Gohman NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) : 712f8d0578e4cbd5922696c92f5068c5513d8e8d60eDan Gohman SE.getSCEV(Sum)); 713087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman for (++I; I != E && I->first == CurLoop; ++I) 714087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman NewOps.push_back(I->second); 715087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op)); 716f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick } else if (Op->isNonConstantNegative()) { 717087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Instead of doing a negate and add, just do a subtract. 718a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty); 719087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = InsertNoopCastOfTo(Sum, Ty); 720087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = InsertBinop(Instruction::Sub, Sum, W); 721087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman ++I; 722a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } else { 723087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // A simple add. 724a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Value *W = expandCodeFor(Op, Ty); 725087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = InsertNoopCastOfTo(Sum, Ty); 726087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Canonicalize a constant to the RHS. 727087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (isa<Constant>(Sum)) std::swap(Sum, W); 728087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = InsertBinop(Instruction::Add, Sum, W); 729087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman ++I; 730a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 7312d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman } 732087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 733087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return Sum; 734e24fa64d52330626553298f56ba5aa702624c282Dan Gohman} 7355be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 736890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) { 737db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 738087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 739087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Collect all the mul operands in a loop, along with their associated loops. 740087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Iterate in reverse so that constants are emitted last, all else equal. 741087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 742087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()), 743087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman E(S->op_begin()); I != E; ++I) 7449c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 745087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 746087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Sort by loop. Use a stable sort so that constants follow non-constants. 747f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(SE.DT)); 748087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 749087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Emit instructions to mul all the operands. Hoist as much as possible 750087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // out of loops. 751dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines Value *Prod = nullptr; 752f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar for (const auto &I : OpsAndLoops) { 753f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar const SCEV *Op = I.second; 754087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!Prod) { 755087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // This is the first operand. Just expand it. 756087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Prod = expand(Op); 757087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } else if (Op->isAllOnesValue()) { 758087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Instead of doing a multiply by negative one, just do a negate. 759087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Prod = InsertNoopCastOfTo(Prod, Ty); 760087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod); 761087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } else { 762087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // A simple mul. 763087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Value *W = expandCodeFor(Op, Ty); 764087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Prod = InsertNoopCastOfTo(Prod, Ty); 765087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Canonicalize a constant to the RHS. 766087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (isa<Constant>(Prod)) std::swap(Prod, W); 767f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar const APInt *RHS; 768f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (match(W, m_Power2(RHS))) { 769f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar // Canonicalize Prod*(1<<C) to Prod<<C. 770f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar assert(!Ty->isVectorTy() && "vector types are not SCEVable"); 771f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Prod = InsertBinop(Instruction::Shl, Prod, 772f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar ConstantInt::get(Ty, RHS->logBase2())); 773f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar } else { 774f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Prod = InsertBinop(Instruction::Mul, Prod, W); 775f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar } 776087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 7772d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman } 7782d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman 779087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return Prod; 78036f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman} 78136f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 782890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) { 783db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 7842d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman 78592fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *LHS = expandCodeFor(S->getLHS(), Ty); 786890f92b744fb074465bc2b7006ee753a181f62a4Dan Gohman if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) { 787f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar const APInt &RHS = SC->getAPInt(); 7886177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky if (RHS.isPowerOf2()) 7896177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky return InsertBinop(Instruction::LShr, LHS, 790eed707b1e6097aac2bb6b3d47271f6300ace7f2eOwen Anderson ConstantInt::get(Ty, RHS.logBase2())); 7916177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky } 7926177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky 79392fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *RHS = expandCodeFor(S->getRHS(), Ty); 794267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman return InsertBinop(Instruction::UDiv, LHS, RHS); 7956177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky} 7966177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky 797453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// Move parts of Base into Rest to leave Base with the minimal 798453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// expression that provides a pointer operand suitable for a 799453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// GEP expansion. 8000bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohmanstatic void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest, 801453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman ScalarEvolution &SE) { 802453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) { 803453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman Base = A->getStart(); 804453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman Rest = SE.getAddExpr(Rest, 805deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman SE.getAddRecExpr(SE.getConstant(A->getType(), 0), 806453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman A->getStepRecurrence(SE), 8073228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick A->getLoop(), 8086f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick A->getNoWrapFlags(SCEV::FlagNW))); 809453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 810453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) { 811453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman Base = A->getOperand(A->getNumOperands()-1); 8120bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end()); 813453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman NewAddOps.back() = Rest; 814453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman Rest = SE.getAddExpr(NewAddOps); 815453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman ExposePointerBase(Base, Rest, SE); 816453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 817453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman} 818453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 819c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick/// Determine if this is a well-behaved chain of instructions leading back to 820c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick/// the PHI. If so, it may be reused by expanded expressions. 821c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trickbool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV, 822c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick const Loop *L) { 823c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) || 824c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV))) 825c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return false; 826c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // If any of the operands don't dominate the insert position, bail. 827c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Addrec operands are always loop-invariant, so this can only happen 828c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // if there are instructions which haven't been hoisted. 829c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (L == IVIncInsertLoop) { 830c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick for (User::op_iterator OI = IncV->op_begin()+1, 831c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick OE = IncV->op_end(); OI != OE; ++OI) 832c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (Instruction *OInst = dyn_cast<Instruction>(OI)) 833f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (!SE.DT.dominates(OInst, IVIncInsertPos)) 834c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return false; 835c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 836c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Advance to the next instruction. 837c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick IncV = dyn_cast<Instruction>(IncV->getOperand(0)); 838c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (!IncV) 839c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return false; 840c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 841c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (IncV->mayHaveSideEffects()) 842c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return false; 843c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 844c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (IncV != PN) 845c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return true; 846c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 847c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return isNormalAddRecExprPHI(PN, IncV, L); 848c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick} 849c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 850b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// getIVIncOperand returns an induction variable increment's induction 851b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// variable operand. 852b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// 853b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// If allowScale is set, any type of GEP is allowed as long as the nonIV 854b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// operands dominate InsertPos. 855b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// 856b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// If allowScale is not set, ensure that a GEP increment conforms to one of the 857b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// simple patterns generated by getAddRecExprPHILiterally and 858b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// expandAddtoGEP. If the pattern isn't recognized, return NULL. 859b5c26ef9da16052597d59a412eaae32098aa1be0Andrew TrickInstruction *SCEVExpander::getIVIncOperand(Instruction *IncV, 860b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick Instruction *InsertPos, 861b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick bool allowScale) { 862b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (IncV == InsertPos) 863dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines return nullptr; 86464925c55c65f9345a69fb67db07aa62cfb723577Andrew Trick 865c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick switch (IncV->getOpcode()) { 866b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick default: 867dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines return nullptr; 868c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Check for a simple Add/Sub or GEP of a loop invariant step. 869c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick case Instruction::Add: 870b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick case Instruction::Sub: { 871b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1)); 872f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (!OInst || SE.DT.dominates(OInst, InsertPos)) 873b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return dyn_cast<Instruction>(IncV->getOperand(0)); 874dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines return nullptr; 875b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 876c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick case Instruction::BitCast: 877b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return dyn_cast<Instruction>(IncV->getOperand(0)); 878b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick case Instruction::GetElementPtr: 879f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar for (auto I = IncV->op_begin() + 1, E = IncV->op_end(); I != E; ++I) { 880c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (isa<Constant>(*I)) 881c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick continue; 882b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (Instruction *OInst = dyn_cast<Instruction>(*I)) { 883f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (!SE.DT.dominates(OInst, InsertPos)) 884dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines return nullptr; 885b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 886b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (allowScale) { 887b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // allow any kind of GEP as long as it can be hoisted. 888b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick continue; 889b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 890b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // This must be a pointer addition of constants (pretty), which is already 891b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // handled, or some number of address-size elements (ugly). Ugly geps 892b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // have 2 operands. i1* is used by the expander to represent an 893b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // address-size element. 894c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (IncV->getNumOperands() != 2) 895dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines return nullptr; 896365c9f1ff55bef134c6b9707f7df44d680ddabeaAndrew Trick unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace(); 897c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS) 898c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS)) 899dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines return nullptr; 900c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick break; 901c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 902b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return dyn_cast<Instruction>(IncV->getOperand(0)); 903c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 904b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick} 905b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick 906de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar/// If the insert point of the current builder or any of the builders on the 907de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar/// stack of saved builders has 'I' as its insert point, update it to point to 908de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar/// the instruction after 'I'. This is intended to be used when the instruction 909de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar/// 'I' is being moved. If this fixup is not done and 'I' is moved to a 910de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar/// different block, the inconsistent insert point (with a mismatched 911de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar/// Instruction and Block) can lead to an instruction being inserted in a block 912de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar/// other than its parent. 913de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainarvoid SCEVExpander::fixupInsertPoints(Instruction *I) { 914de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar BasicBlock::iterator It(*I); 915de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar BasicBlock::iterator NewInsertPt = std::next(It); 916de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (Builder.GetInsertPoint() == It) 917de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Builder.SetInsertPoint(&*NewInsertPt); 918de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar for (auto *InsertPtGuard : InsertPointGuards) 919de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (InsertPtGuard->GetInsertPoint() == It) 920de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar InsertPtGuard->SetInsertPoint(NewInsertPt); 921de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar} 922de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 923b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make 924b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// it available to other uses in this loop. Recursively hoist any operands, 925b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// until we reach a value that dominates InsertPos. 926b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trickbool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) { 927f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (SE.DT.dominates(IncV, InsertPos)) 928b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return true; 929b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick 930b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // InsertPos must itself dominate IncV so that IncV's new position satisfies 931b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // its existing users. 932f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (isa<PHINode>(InsertPos) || 933f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar !SE.DT.dominates(InsertPos->getParent(), IncV->getParent())) 934f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return false; 935f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 936f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (!SE.LI.movementPreservesLCSSAForm(IncV, InsertPos)) 937c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return false; 938b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick 939b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // Check that the chain of IV operands leading back to Phi can be hoisted. 940b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick SmallVector<Instruction*, 4> IVIncs; 941b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick for(;;) { 942b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true); 943b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (!Oper) 944b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return false; 945b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // IncV is safe to hoist. 946b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick IVIncs.push_back(IncV); 947b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick IncV = Oper; 948f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (SE.DT.dominates(IncV, InsertPos)) 949b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick break; 950b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 951f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar for (auto I = IVIncs.rbegin(), E = IVIncs.rend(); I != E; ++I) { 952de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar fixupInsertPoints(*I); 953b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick (*I)->moveBefore(InsertPos); 954b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 955b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return true; 956b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick} 957b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick 958b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// Determine if this cyclic phi is in a form that would have been generated by 959b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// LSR. We don't care if the phi was actually expanded in this pass, as long 960b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// as it is in a low-cost form, for example, no implied multiplication. This 961b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// should match any patterns generated by getAddRecExprPHILiterally and 962b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// expandAddtoGEP. 963b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trickbool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV, 964b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick const Loop *L) { 965b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick for(Instruction *IVOper = IncV; 966b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(), 967b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick /*allowScale=*/false));) { 968b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (IVOper == PN) 969b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return true; 970c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 971b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return false; 972c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick} 973c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 974553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick/// expandIVInc - Expand an IV increment at Builder's current InsertPos. 975553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick/// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may 976553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick/// need to materialize IV increments elsewhere to handle difficult situations. 977553fe05f236f46fe27b7fcfa822b06367d50183eAndrew TrickValue *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L, 978553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Type *ExpandTy, Type *IntTy, 979553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick bool useSubtract) { 980553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Value *IncV; 981553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // If the PHI is a pointer, use a GEP, otherwise use an add or sub. 982553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick if (ExpandTy->isPointerTy()) { 983553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick PointerType *GEPPtrTy = cast<PointerType>(ExpandTy); 984553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // If the step isn't constant, don't use an implicitly scaled GEP, because 985553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // that would require a multiply inside the loop. 986553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick if (!isa<ConstantInt>(StepV)) 987553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()), 988553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick GEPPtrTy->getAddressSpace()); 989553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick const SCEV *const StepArray[1] = { SE.getSCEV(StepV) }; 990553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN); 991553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick if (IncV->getType() != PN->getType()) { 992553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick IncV = Builder.CreateBitCast(IncV, PN->getType()); 993553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick rememberInstruction(IncV); 994553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick } 995553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick } else { 996553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick IncV = useSubtract ? 997553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") : 998553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next"); 999553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick rememberInstruction(IncV); 1000553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick } 1001553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick return IncV; 1002553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick} 1003553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick 100436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines/// \brief Hoist the addrec instruction chain rooted in the loop phi above the 100536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines/// position. This routine assumes that this is possible (has been checked). 1006de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainarvoid SCEVExpander::hoistBeforePos(DominatorTree *DT, Instruction *InstToHoist, 1007de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Instruction *Pos, PHINode *LoopPhi) { 100836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines do { 100936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (DT->dominates(InstToHoist, Pos)) 101036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines break; 101136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Make sure the increment is where we want it. But don't move it 101236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // down past a potential existing post-inc user. 1013de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar fixupInsertPoints(InstToHoist); 101436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines InstToHoist->moveBefore(Pos); 101536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Pos = InstToHoist; 101636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines InstToHoist = cast<Instruction>(InstToHoist->getOperand(0)); 101736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } while (InstToHoist != LoopPhi); 101836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines} 101936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 102036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines/// \brief Check whether we can cheaply express the requested SCEV in terms of 1021f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar/// the available PHI SCEV by truncation and/or inversion of the step. 102236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hinesstatic bool canBeCheaplyTransformed(ScalarEvolution &SE, 102336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines const SCEVAddRecExpr *Phi, 102436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines const SCEVAddRecExpr *Requested, 102536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines bool &InvertStep) { 102636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Type *PhiTy = SE.getEffectiveSCEVType(Phi->getType()); 102736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Type *RequestedTy = SE.getEffectiveSCEVType(Requested->getType()); 102836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 102936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (RequestedTy->getIntegerBitWidth() > PhiTy->getIntegerBitWidth()) 103036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines return false; 103136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 103236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Try truncate it if necessary. 103336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Phi = dyn_cast<SCEVAddRecExpr>(SE.getTruncateOrNoop(Phi, RequestedTy)); 103436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (!Phi) 103536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines return false; 103636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 103736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Check whether truncation will help. 103836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (Phi == Requested) { 103936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines InvertStep = false; 104036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines return true; 104136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } 104236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 104336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Check whether inverting will help: {R,+,-1} == R - {0,+,1}. 104436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (SE.getAddExpr(Requested->getStart(), 104536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines SE.getNegativeSCEV(Requested)) == Phi) { 104636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines InvertStep = true; 104736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines return true; 104836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } 104936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 105036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines return false; 105136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines} 105236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 1053ebe69fe11e48d322045d5949c83283927a0d790bStephen Hinesstatic bool IsIncrementNSW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) { 1054ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines if (!isa<IntegerType>(AR->getType())) 1055ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines return false; 1056ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines 1057ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth(); 1058ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2); 1059ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines const SCEV *Step = AR->getStepRecurrence(SE); 1060ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines const SCEV *OpAfterExtend = SE.getAddExpr(SE.getSignExtendExpr(Step, WideTy), 1061ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines SE.getSignExtendExpr(AR, WideTy)); 1062ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines const SCEV *ExtendAfterOp = 1063ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines SE.getSignExtendExpr(SE.getAddExpr(AR, Step), WideTy); 1064ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines return ExtendAfterOp == OpAfterExtend; 1065ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines} 1066ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines 1067ebe69fe11e48d322045d5949c83283927a0d790bStephen Hinesstatic bool IsIncrementNUW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) { 1068ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines if (!isa<IntegerType>(AR->getType())) 1069ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines return false; 1070ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines 1071ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth(); 1072ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2); 1073ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines const SCEV *Step = AR->getStepRecurrence(SE); 1074ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines const SCEV *OpAfterExtend = SE.getAddExpr(SE.getZeroExtendExpr(Step, WideTy), 1075ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines SE.getZeroExtendExpr(AR, WideTy)); 1076ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines const SCEV *ExtendAfterOp = 1077ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines SE.getZeroExtendExpr(SE.getAddExpr(AR, Step), WideTy); 1078ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines return ExtendAfterOp == OpAfterExtend; 1079ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines} 1080ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines 1081a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman/// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand 1082a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman/// the base addrec, which is the addrec without any non-loop-dominating 1083a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman/// values, and return the PHI. 1084a10756ee657a4d43a48cca5c166919093930ed6bDan GohmanPHINode * 1085a10756ee657a4d43a48cca5c166919093930ed6bDan GohmanSCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized, 1086a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const Loop *L, 1087db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *ExpandTy, 108836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Type *IntTy, 108936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Type *&TruncTy, 109036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines bool &InvertStep) { 109193a896e2e36480d55de3bab53e68581e08526344Benjamin Kramer assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position"); 1092d152d03a476b8d0d4b26577db26e2ba76034b0f3Andrew Trick 1093a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Reuse a previously-inserted PHI, if present. 1094c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick BasicBlock *LatchBlock = L->getLoopLatch(); 1095c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (LatchBlock) { 1096dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines PHINode *AddRecPhiMatch = nullptr; 1097dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines Instruction *IncV = nullptr; 1098dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines TruncTy = nullptr; 109936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines InvertStep = false; 110036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 110136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Only try partially matching scevs that need truncation and/or 110236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // step-inversion if we know this loop is outside the current loop. 1103f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar bool TryNonMatchingSCEV = 1104f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar IVIncInsertLoop && 1105f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar SE.DT.properlyDominates(LatchBlock, IVIncInsertLoop->getHeader()); 110636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 1107f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar for (auto &I : *L->getHeader()) { 1108f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar auto *PN = dyn_cast<PHINode>(&I); 1109f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (!PN || !SE.isSCEVable(PN->getType())) 111036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines continue; 111136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 111236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines const SCEVAddRecExpr *PhiSCEV = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PN)); 111336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (!PhiSCEV) 1114c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick continue; 1115c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 111636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines bool IsMatchingSCEV = PhiSCEV == Normalized; 111736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // We only handle truncation and inversion of phi recurrences for the 111836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // expanded expression if the expanded expression's loop dominates the 111936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // loop we insert to. Check now, so we can bail out early. 112036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (!IsMatchingSCEV && !TryNonMatchingSCEV) 112136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines continue; 112236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 112336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Instruction *TempIncV = 112436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)); 1125c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 112636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Check whether we can reuse this PHI node. 1127c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (LSRMode) { 112836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (!isExpandedAddRecExprPHI(PN, TempIncV, L)) 1129c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick continue; 113036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (L == IVIncInsertLoop && !hoistIVInc(TempIncV, IVIncInsertPos)) 1131b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick continue; 113236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } else { 113336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (!isNormalAddRecExprPHI(PN, TempIncV, L)) 1134c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick continue; 1135c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 113636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 113736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Stop if we have found an exact match SCEV. 113836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (IsMatchingSCEV) { 113936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines IncV = TempIncV; 1140dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines TruncTy = nullptr; 114136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines InvertStep = false; 114236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines AddRecPhiMatch = PN; 114336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines break; 114436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } 114536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 114636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Try whether the phi can be translated into the requested form 114736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // (truncated and/or offset by a constant). 114836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if ((!TruncTy || InvertStep) && 114936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines canBeCheaplyTransformed(SE, PhiSCEV, Normalized, InvertStep)) { 115036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Record the phi node. But don't stop we might find an exact match 115136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // later. 115236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines AddRecPhiMatch = PN; 115336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines IncV = TempIncV; 115436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines TruncTy = SE.getEffectiveSCEVType(Normalized->getType()); 115536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } 115636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } 115736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 115836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (AddRecPhiMatch) { 115936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Potentially, move the increment. We have made sure in 116036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // isExpandedAddRecExprPHI or hoistIVInc that this is possible. 116136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (L == IVIncInsertLoop) 1162f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar hoistBeforePos(&SE.DT, IncV, IVIncInsertPos, AddRecPhiMatch); 116336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 1164c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Ok, the add recurrence looks usable. 1165c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Remember this PHI, even in post-inc mode. 116636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines InsertedValues.insert(AddRecPhiMatch); 1167c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Remember the increment. 1168c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick rememberInstruction(IncV); 116936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines return AddRecPhiMatch; 1170c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 1171c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 1172a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1173a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Save the original insertion point so we can restore it when we're done. 1174de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar SCEVInsertPointGuard Guard(Builder, this); 1175a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1176ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // Another AddRec may need to be recursively expanded below. For example, if 1177ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // this AddRec is quadratic, the StepV may itself be an AddRec in this 1178ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // loop. Remove this loop from the PostIncLoops set before expanding such 1179ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // AddRecs. Otherwise, we cannot find a valid position for the step 1180ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // (i.e. StepV can never dominate its loop header). Ideally, we could do 1181ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // SavedIncLoops.swap(PostIncLoops), but we generally have a single element, 1182ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // so it's not worth implementing SmallPtrSet::swap. 1183ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick PostIncLoopSet SavedPostIncLoops = PostIncLoops; 1184ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick PostIncLoops.clear(); 1185ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick 1186a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Expand code for the start value. 1187f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Value *StartV = 1188f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar expandCodeFor(Normalized->getStart(), ExpandTy, &L->getHeader()->front()); 1189a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1190d152d03a476b8d0d4b26577db26e2ba76034b0f3Andrew Trick // StartV must be hoisted into L's preheader to dominate the new phi. 119193a896e2e36480d55de3bab53e68581e08526344Benjamin Kramer assert(!isa<Instruction>(StartV) || 1192f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar SE.DT.properlyDominates(cast<Instruction>(StartV)->getParent(), 1193f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar L->getHeader())); 1194d152d03a476b8d0d4b26577db26e2ba76034b0f3Andrew Trick 1195553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // Expand code for the step value. Do this before creating the PHI so that PHI 1196553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // reuse code doesn't see an incomplete PHI. 1197a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEV *Step = Normalized->getStepRecurrence(SE); 1198553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // If the stride is negative, insert a sub instead of an add for the increment 1199553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // (unless it's a constant, because subtracts of constants are canonicalized 1200553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // to adds). 1201f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1202553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick if (useSubtract) 1203a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Step = SE.getNegativeSCEV(Step); 1204553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // Expand the step somewhere that dominates the loop header. 1205f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Value *StepV = expandCodeFor(Step, IntTy, &L->getHeader()->front()); 1206a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1207ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines // The no-wrap behavior proved by IsIncrement(NUW|NSW) is only applicable if 1208ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines // we actually do emit an addition. It does not apply if we emit a 1209ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines // subtraction. 1210ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines bool IncrementIsNUW = !useSubtract && IsIncrementNUW(SE, Normalized); 1211ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines bool IncrementIsNSW = !useSubtract && IsIncrementNSW(SE, Normalized); 1212ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines 1213a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Create the PHI. 1214d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad BasicBlock *Header = L->getHeader(); 1215d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad Builder.SetInsertPoint(Header, Header->begin()); 1216d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 12175e7645be4c9dd2193add44d30b5fef8036d7a3ceAndrew Trick PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE), 1218dc8e546048db2f7ff5656742b2b26975098a11a0Andrew Trick Twine(IVName) + ".iv"); 1219a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(PN); 1220a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1221a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Create the step instructions and populate the PHI. 1222d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 1223a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman BasicBlock *Pred = *HPI; 1224a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1225a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Add a start value. 1226a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman if (!L->contains(Pred)) { 1227a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman PN->addIncoming(StartV, Pred); 1228a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman continue; 1229a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1230a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1231553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // Create a step value and add it to the PHI. 1232553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the 1233553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // instructions at IVIncInsertPos. 1234a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Instruction *InsertPos = L == IVIncInsertLoop ? 1235a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman IVIncInsertPos : Pred->getTerminator(); 1236c5ecbdc1896f1cc089372feef3191ace2f840898Devang Patel Builder.SetInsertPoint(InsertPos); 1237553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1238ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines 1239409443b1c6415e55c2bd4f0662e14cbc52d16686Andrew Trick if (isa<OverflowingBinaryOperator>(IncV)) { 1240ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines if (IncrementIsNUW) 1241409443b1c6415e55c2bd4f0662e14cbc52d16686Andrew Trick cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap(); 1242ebe69fe11e48d322045d5949c83283927a0d790bStephen Hines if (IncrementIsNSW) 1243409443b1c6415e55c2bd4f0662e14cbc52d16686Andrew Trick cast<BinaryOperator>(IncV)->setHasNoSignedWrap(); 1244409443b1c6415e55c2bd4f0662e14cbc52d16686Andrew Trick } 1245a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman PN->addIncoming(IncV, Pred); 1246a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1247a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1248ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // After expanding subexpressions, restore the PostIncLoops set so the caller 1249ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // can ensure that IVIncrement dominates the current uses. 1250ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick PostIncLoops = SavedPostIncLoops; 1251ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick 1252a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Remember this PHI, even in post-inc mode. 1253a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman InsertedValues.insert(PN); 1254a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1255a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman return PN; 1256a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman} 1257a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1258a10756ee657a4d43a48cca5c166919093930ed6bDan GohmanValue *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) { 1259db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *STy = S->getType(); 1260db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *IntTy = SE.getEffectiveSCEVType(STy); 1261a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const Loop *L = S->getLoop(); 1262a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1263a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Determine a normalized form of this expression, which is the expression 1264a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // before any post-inc adjustment is made. 1265a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEVAddRecExpr *Normalized = S; 1266448db1cdef5872713ef77beffacf502ae3450cd7Dan Gohman if (PostIncLoops.count(L)) { 1267448db1cdef5872713ef77beffacf502ae3450cd7Dan Gohman PostIncLoopSet Loops; 1268448db1cdef5872713ef77beffacf502ae3450cd7Dan Gohman Loops.insert(L); 1269f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Normalized = cast<SCEVAddRecExpr>(TransformForPostIncUse( 1270f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Normalize, S, nullptr, nullptr, Loops, SE, SE.DT)); 1271a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1272a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1273a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Strip off any non-loop-dominating component from the addrec start. 1274a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEV *Start = Normalized->getStart(); 1275dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines const SCEV *PostLoopOffset = nullptr; 1276dc0e8fb9f9512622f55f73e1a434caa5c0915694Dan Gohman if (!SE.properlyDominates(Start, L->getHeader())) { 1277a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman PostLoopOffset = Start; 1278deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman Start = SE.getConstant(Normalized->getType(), 0); 12793228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick Normalized = cast<SCEVAddRecExpr>( 12803228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE), 12813228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick Normalized->getLoop(), 12826f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick Normalized->getNoWrapFlags(SCEV::FlagNW))); 1283a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1284a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1285a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Strip off any non-loop-dominating component from the addrec step. 1286a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEV *Step = Normalized->getStepRecurrence(SE); 1287dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines const SCEV *PostLoopScale = nullptr; 1288dc0e8fb9f9512622f55f73e1a434caa5c0915694Dan Gohman if (!SE.dominates(Step, L->getHeader())) { 1289a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman PostLoopScale = Step; 1290deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman Step = SE.getConstant(Normalized->getType(), 1); 1291de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (!Start->isZero()) { 1292de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // The normalization below assumes that Start is constant zero, so if 1293de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // it isn't re-associate Start to PostLoopOffset. 1294de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar assert(!PostLoopOffset && "Start not-null but PostLoopOffset set?"); 1295de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar PostLoopOffset = Start; 1296de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Start = SE.getConstant(Normalized->getType(), 0); 1297de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar } 1298a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Normalized = 12996f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick cast<SCEVAddRecExpr>(SE.getAddRecExpr( 13006f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick Start, Step, Normalized->getLoop(), 13016f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick Normalized->getNoWrapFlags(SCEV::FlagNW))); 1302a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1303a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1304a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Expand the core addrec. If we need post-loop scaling, force it to 1305a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // expand to an integer type to avoid the need for additional casting. 1306db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *ExpandTy = PostLoopScale ? IntTy : STy; 130736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // In some cases, we decide to reuse an existing phi node but need to truncate 130836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // it and/or invert the step. 1309dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines Type *TruncTy = nullptr; 131036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines bool InvertStep = false; 131136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy, 131236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines TruncTy, InvertStep); 1313a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 13143f46a3abeedba8d517b4182de34c821d752db058Dan Gohman // Accommodate post-inc mode, if necessary. 1315a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Value *Result; 1316448db1cdef5872713ef77beffacf502ae3450cd7Dan Gohman if (!PostIncLoops.count(L)) 1317a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Result = PN; 1318a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman else { 1319a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // In PostInc mode, use the post-incremented value. 1320a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman BasicBlock *LatchBlock = L->getLoopLatch(); 1321a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman assert(LatchBlock && "PostInc mode requires a unique loop latch!"); 1322a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Result = PN->getIncomingValueForBlock(LatchBlock); 132348ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick 132448ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // For an expansion to use the postinc form, the client must call 132548ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // expandCodeFor with an InsertPoint that is either outside the PostIncLoop 132648ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // or dominated by IVIncInsertPos. 1327f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (isa<Instruction>(Result) && 1328f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar !SE.DT.dominates(cast<Instruction>(Result), 1329f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar &*Builder.GetInsertPoint())) { 1330553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // The induction variable's postinc expansion does not dominate this use. 1331553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // IVUsers tries to prevent this case, so it is rare. However, it can 1332553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // happen when an IVUser outside the loop is not dominated by the latch 1333553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // block. Adjusting IVIncInsertPos before expansion begins cannot handle 1334553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // all cases. Consider a phi outide whose operand is replaced during 1335553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // expansion with the value of the postinc user. Without fundamentally 1336553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // changing the way postinc users are tracked, the only remedy is 1337553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // inserting an extra IV increment. StepV might fold into PostLoopOffset, 1338553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // but hopefully expandCodeFor handles that. 1339553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick bool useSubtract = 1340f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1341553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick if (useSubtract) 1342553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Step = SE.getNegativeSCEV(Step); 1343d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer Value *StepV; 1344d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer { 1345d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer // Expand the step somewhere that dominates the loop header. 1346de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar SCEVInsertPointGuard Guard(Builder, this); 1347f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar StepV = expandCodeFor(Step, IntTy, &L->getHeader()->front()); 1348d4278821665aa97f5fc0d19a32ff1fb39a22d395Benjamin Kramer } 1349553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1350553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick } 1351a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1352a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 135336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // We have decided to reuse an induction variable of a dominating loop. Apply 135436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // truncation and/or invertion of the step. 135536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (TruncTy) { 135636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Type *ResTy = Result->getType(); 135736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Normalize the result type. 135836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (ResTy != SE.getEffectiveSCEVType(ResTy)) 135936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Result = InsertNoopCastOfTo(Result, SE.getEffectiveSCEVType(ResTy)); 136036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Truncate the result. 136136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (TruncTy != Result->getType()) { 136236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Result = Builder.CreateTrunc(Result, TruncTy); 136336b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines rememberInstruction(Result); 136436b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } 136536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Invert the result. 136636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (InvertStep) { 136736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Result = Builder.CreateSub(expandCodeFor(Normalized->getStart(), TruncTy), 136836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines Result); 136936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines rememberInstruction(Result); 137036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } 137136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines } 137236b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines 1373a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Re-apply any non-loop-dominating scale. 1374a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman if (PostLoopScale) { 13754d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick assert(S->isAffine() && "Can't linearly scale non-affine recurrences."); 13760a799ab15801d4ebf68eeb151d6375a799c87d9aDan Gohman Result = InsertNoopCastOfTo(Result, IntTy); 1377a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Result = Builder.CreateMul(Result, 1378a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman expandCodeFor(PostLoopScale, IntTy)); 1379a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(Result); 1380a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1381a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1382a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Re-apply any non-loop-dominating offset. 1383a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman if (PostLoopOffset) { 1384db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) { 1385a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEV *const OffsetArray[1] = { PostLoopOffset }; 1386a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result); 1387a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } else { 13880a799ab15801d4ebf68eeb151d6375a799c87d9aDan Gohman Result = InsertNoopCastOfTo(Result, IntTy); 1389a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Result = Builder.CreateAdd(Result, 1390a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman expandCodeFor(PostLoopOffset, IntTy)); 1391a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(Result); 1392a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1393a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1394a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1395a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman return Result; 1396a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman} 1397a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1398890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) { 1399a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman if (!CanonicalMode) return expandAddRecExprLiterally(S); 1400a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1401db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 140236f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman const Loop *L = S->getLoop(); 140336f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 14044d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // First check for an existing canonical IV in a suitable type. 1405dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines PHINode *CanonicalIV = nullptr; 14064d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman if (PHINode *PN = L->getCanonicalInductionVariable()) 1407133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty)) 14084d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman CanonicalIV = PN; 14094d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman 14104d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // Rewrite an AddRec in terms of the canonical induction variable, if 14114d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // its type is more narrow. 14124d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman if (CanonicalIV && 14134d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman SE.getTypeSizeInBits(CanonicalIV->getType()) > 14144d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman SE.getTypeSizeInBits(Ty)) { 1415f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman SmallVector<const SCEV *, 4> NewOps(S->getNumOperands()); 1416f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i) 1417f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType()); 14183228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(), 14196f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick S->getNoWrapFlags(SCEV::FlagNW))); 14204d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman BasicBlock::iterator NewInsertPt = 1421f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar findInsertPointAfter(cast<Instruction>(V), Builder.GetInsertBlock()); 1422dce4a407a24b04eebc6a376f8e62b41aaa7b071fStephen Hines V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), nullptr, 1423f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar &*NewInsertPt); 14244d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman return V; 14254d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman } 14264d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman 142736f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // {X,+,F} --> X + {0,+,F} 1428cfeb6a450632f2a6cd05302633c8c2b8c90cfdfdDan Gohman if (!S->getStart()->isZero()) { 1429f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end()); 1430deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman NewOps[0] = SE.getConstant(Ty, 0); 14316f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick const SCEV *Rest = SE.getAddRecExpr(NewOps, L, 14326f71dd765ae9e1d1d0ba01d98a05627ffe3bfc8aAndrew Trick S->getNoWrapFlags(SCEV::FlagNW)); 1433453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 1434453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the 1435453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // comments on expandAddToGEP for details. 1436c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *Base = S->getStart(); 1437c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *RestArray[1] = { Rest }; 1438c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Dig into the expression to find the pointer base for a GEP. 1439c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ExposePointerBase(Base, RestArray[0], SE); 1440c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // If we found a pointer, expand the AddRec with a GEP. 1441db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) { 1442c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Make sure the Base isn't something exotic, such as a multiplied 1443c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // or divided pointer value. In those cases, the result type isn't 1444c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // actually a pointer type. 1445c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) { 1446c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman Value *StartV = expand(Base); 1447c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!"); 1448c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV); 1449453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 1450453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 1451453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 145240a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // Just do a normal add. Pre-expand the operands to suppress folding. 1453de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // 1454de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // The LHS and RHS values are factored out of the expand call to make the 1455de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // output independent of the argument evaluation order. 1456de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar const SCEV *AddExprLHS = SE.getUnknown(expand(S->getStart())); 1457de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar const SCEV *AddExprRHS = SE.getUnknown(expand(Rest)); 1458de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar return expand(SE.getAddExpr(AddExprLHS, AddExprRHS)); 145936f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman } 146036f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 14616ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman // If we don't yet have a canonical IV, create one. 14626ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman if (!CanonicalIV) { 146336f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // Create and insert the PHI node for the induction variable in the 146436f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // specified loop. 146536f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman BasicBlock *Header = L->getHeader(); 1466d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 14673ecfc861b4365f341c5c969b40e1afccde676e6fJay Foad CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar", 1468f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar &Header->front()); 14696ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman rememberInstruction(CanonicalIV); 147036f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 147119046ec19c4931f025cd4730eb43e9203ada6bb7Hal Finkel SmallSet<BasicBlock *, 4> PredSeen; 1472eed707b1e6097aac2bb6b3d47271f6300ace7f2eOwen Anderson Constant *One = ConstantInt::get(Ty, 1); 1473d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 14747656018c2268285907cfdc106071462a01a73878Gabor Greif BasicBlock *HP = *HPI; 147537ed9c199ca639565f6ce88105f9e39e898d82d0Stephen Hines if (!PredSeen.insert(HP).second) { 147637ed9c199ca639565f6ce88105f9e39e898d82d0Stephen Hines // There must be an incoming value for each predecessor, even the 147737ed9c199ca639565f6ce88105f9e39e898d82d0Stephen Hines // duplicates! 147837ed9c199ca639565f6ce88105f9e39e898d82d0Stephen Hines CanonicalIV->addIncoming(CanonicalIV->getIncomingValueForBlock(HP), HP); 147919046ec19c4931f025cd4730eb43e9203ada6bb7Hal Finkel continue; 148037ed9c199ca639565f6ce88105f9e39e898d82d0Stephen Hines } 148119046ec19c4931f025cd4730eb43e9203ada6bb7Hal Finkel 14827656018c2268285907cfdc106071462a01a73878Gabor Greif if (L->contains(HP)) { 14833abf905b50f33340ae81913da81b0eda96fa4616Dan Gohman // Insert a unit add instruction right before the terminator 14843abf905b50f33340ae81913da81b0eda96fa4616Dan Gohman // corresponding to the back-edge. 14856ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One, 14866ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman "indvar.next", 14876ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman HP->getTerminator()); 1488df3ad6697bb849e514c0881ca8700ea36678cbdaDevang Patel Add->setDebugLoc(HP->getTerminator()->getDebugLoc()); 1489a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(Add); 14906ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman CanonicalIV->addIncoming(Add, HP); 149183d577490b1473213b6973236110c28e59127956Dan Gohman } else { 14926ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP); 149383d577490b1473213b6973236110c28e59127956Dan Gohman } 14947656018c2268285907cfdc106071462a01a73878Gabor Greif } 149536f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman } 149636f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 14976ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman // {0,+,1} --> Insert a canonical induction variable into the loop! 14986ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman if (S->isAffine() && S->getOperand(1)->isOne()) { 14996ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) && 15006ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman "IVs with types different from the canonical IV should " 15016ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman "already have been handled!"); 15026ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman return CanonicalIV; 15036ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman } 15046ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman 15054d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // {0,+,F} --> {0,+,1} * F 150636f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 1507df14a04b5ccbbe6a46c2ccb93e27b12a36ff163eChris Lattner // If this is a simple linear addrec, emit it now as a special case. 150840a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman if (S->isAffine()) // {0,+,F} --> i*F 150940a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman return 151040a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman expand(SE.getTruncateOrNoop( 15116ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman SE.getMulExpr(SE.getUnknown(CanonicalIV), 151240a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman SE.getNoopOrAnyExtend(S->getOperand(1), 15136ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman CanonicalIV->getType())), 151440a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman Ty)); 151536f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 151636f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // If this is a chain of recurrences, turn it into a closed form, using the 151736f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // folders, then expandCodeFor the closed form. This allows the folders to 151836f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // simplify the expression without having to build a bunch of special code 151936f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // into this folder. 15206ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV. 152136f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 15224d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // Promote S up to the canonical IV type, if the cast is foldable. 15230bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *NewS = S; 15246ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType()); 15254d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman if (isa<SCEVAddRecExpr>(Ext)) 15264d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman NewS = Ext; 15274d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman 15280bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE); 1529e81561909d128c6e2d8033cb5465a49b2596b26aBill Wendling //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n"; 153036f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 15314d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // Truncate the result down to the original type, if needed. 15320bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *T = SE.getTruncateOrNoop(V, Ty); 1533469f3cdc13851914f3a766cbd8f701cf8431cacaDan Gohman return expand(T); 153436f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman} 153596fea337d27357e9b62abbf3d2d5ce29f1c8e870Anton Korobeynikov 1536890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) { 1537db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 153892fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *V = expandCodeFor(S->getOperand(), 153992fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman SE.getEffectiveSCEVType(S->getOperand()->getType())); 1540a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Value *I = Builder.CreateTrunc(V, Ty); 1541a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(I); 1542cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman return I; 154311f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman} 154411f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman 1545890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) { 1546db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 154792fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *V = expandCodeFor(S->getOperand(), 154892fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman SE.getEffectiveSCEVType(S->getOperand()->getType())); 1549a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Value *I = Builder.CreateZExt(V, Ty); 1550a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(I); 1551cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman return I; 155211f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman} 155311f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman 1554890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) { 1555db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 155692fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *V = expandCodeFor(S->getOperand(), 155792fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman SE.getEffectiveSCEVType(S->getOperand()->getType())); 1558a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Value *I = Builder.CreateSExt(V, Ty); 1559a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(I); 1560cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman return I; 156111f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman} 156211f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman 1563890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) { 15640196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1565db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = LHS->getType(); 15660196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman for (int i = S->getNumOperands()-2; i >= 0; --i) { 15670196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // In the case of mixed integer and pointer types, do the 15680196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // rest of the comparisons as integer. 15690196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman if (S->getOperand(i)->getType() != Ty) { 15700196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman Ty = SE.getEffectiveSCEVType(Ty); 15710196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman LHS = InsertNoopCastOfTo(LHS, Ty); 15720196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman } 157392fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1574a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Value *ICmp = Builder.CreateICmpSGT(LHS, RHS); 1575a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(ICmp); 1576267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax"); 1577a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(Sel); 1578cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman LHS = Sel; 1579c54c561c9f7270c055dd7ba75a3a003b771a42d9Nick Lewycky } 15800196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // In the case of mixed integer and pointer types, cast the 15810196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // final result back to the pointer type. 15820196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman if (LHS->getType() != S->getType()) 15830196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman LHS = InsertNoopCastOfTo(LHS, S->getType()); 1584c54c561c9f7270c055dd7ba75a3a003b771a42d9Nick Lewycky return LHS; 1585c54c561c9f7270c055dd7ba75a3a003b771a42d9Nick Lewycky} 1586c54c561c9f7270c055dd7ba75a3a003b771a42d9Nick Lewycky 1587890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) { 15880196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1589db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = LHS->getType(); 15900196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman for (int i = S->getNumOperands()-2; i >= 0; --i) { 15910196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // In the case of mixed integer and pointer types, do the 15920196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // rest of the comparisons as integer. 15930196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman if (S->getOperand(i)->getType() != Ty) { 15940196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman Ty = SE.getEffectiveSCEVType(Ty); 15950196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman LHS = InsertNoopCastOfTo(LHS, Ty); 15960196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman } 159792fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1598a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Value *ICmp = Builder.CreateICmpUGT(LHS, RHS); 1599a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(ICmp); 1600267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax"); 1601a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(Sel); 1602cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman LHS = Sel; 16033e6307698084e7adfc10b739442ae29742beefd0Nick Lewycky } 16040196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // In the case of mixed integer and pointer types, cast the 16050196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // final result back to the pointer type. 16060196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman if (LHS->getType() != S->getType()) 16070196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman LHS = InsertNoopCastOfTo(LHS, S->getType()); 16083e6307698084e7adfc10b739442ae29742beefd0Nick Lewycky return LHS; 16093e6307698084e7adfc10b739442ae29742beefd0Nick Lewycky} 16103e6307698084e7adfc10b739442ae29742beefd0Nick Lewycky 1611db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris LattnerValue *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty, 1612b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick Instruction *IP) { 1613f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar assert(IP); 1614f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Builder.SetInsertPoint(IP); 16156c7ed6b54949949806797bafdf545fbfecb2cef5Dan Gohman return expandCodeFor(SH, Ty); 16166c7ed6b54949949806797bafdf545fbfecb2cef5Dan Gohman} 16176c7ed6b54949949806797bafdf545fbfecb2cef5Dan Gohman 1618db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris LattnerValue *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) { 161911f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman // Expand the code for this SCEV. 16202d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman Value *V = expand(SH); 16215be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (Ty) { 16225be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) && 16235be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman "non-trivial casts should be done with the SCEVs directly!"); 16245be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman V = InsertNoopCastOfTo(V, Ty); 16255be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 16265be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman return V; 162711f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman} 162811f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman 1629de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga NainarValue *SCEVExpander::FindValueInExprValueMap(const SCEV *S, 1630de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar const Instruction *InsertPt) { 1631de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar SetVector<Value *> *Set = SE.getSCEVValues(S); 1632de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // If the expansion is not in CanonicalMode, and the SCEV contains any 1633de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // sub scAddRecExpr type SCEV, it is required to expand the SCEV literally. 1634de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (CanonicalMode || !SE.containsAddRecurrence(S)) { 1635de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // If S is scConstant, it may be worse to reuse an existing Value. 1636de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (S->getSCEVType() != scConstant && Set) { 1637de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Choose a Value from the set which dominates the insertPt. 1638de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // insertPt should be inside the Value's parent loop so as not to break 1639de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // the LCSSA form. 1640de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar for (auto const &Ent : *Set) { 1641de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Instruction *EntInst = nullptr; 1642de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (Ent && isa<Instruction>(Ent) && 1643de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar (EntInst = cast<Instruction>(Ent)) && 1644de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar S->getType() == Ent->getType() && 1645de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar EntInst->getFunction() == InsertPt->getFunction() && 1646de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar SE.DT.dominates(EntInst, InsertPt) && 1647de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar (SE.LI.getLoopFor(EntInst->getParent()) == nullptr || 1648de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar SE.LI.getLoopFor(EntInst->getParent())->contains(InsertPt))) { 1649de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar return Ent; 1650de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar } 1651de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar } 1652de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar } 1653de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar } 1654de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar return nullptr; 1655de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar} 1656de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 1657de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar// The expansion of SCEV will either reuse a previous Value in ExprValueMap, 1658de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar// or expand the SCEV literally. Specifically, if the expansion is in LSRMode, 1659de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar// and the SCEV contains any sub scAddRecExpr type SCEV, it will be expanded 1660de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar// literally, to prevent LSR's transformed SCEV from being reverted. Otherwise, 1661de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar// the expansion will try to reuse Value from ExprValueMap, and only when it 1662de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar// fails, expand the SCEV literally. 1663890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::expand(const SCEV *S) { 166440a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // Compute an insertion point for this SCEV object. Hoist the instructions 166540a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // as far out in the loop nest as possible. 1666f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Instruction *InsertPt = &*Builder.GetInsertPoint(); 1667f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar for (Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock());; 166840a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman L = L->getParentLoop()) 166917ead4ff4baceb2c5503f233d0288d363ae44165Dan Gohman if (SE.isLoopInvariant(S, L)) { 167040a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman if (!L) break; 1671e059ee832ca36d65a5fb87b0ac5bcdb0490b15cbDan Gohman if (BasicBlock *Preheader = L->getLoopPreheader()) 167240a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman InsertPt = Preheader->getTerminator(); 16730f8cd56bfdd32af4edb253654db02fb3143b25a8Andrew Trick else { 16740f8cd56bfdd32af4edb253654db02fb3143b25a8Andrew Trick // LSR sets the insertion point for AddRec start/step values to the 16750f8cd56bfdd32af4edb253654db02fb3143b25a8Andrew Trick // block start to simplify value reuse, even though it's an invalid 16760f8cd56bfdd32af4edb253654db02fb3143b25a8Andrew Trick // position. SCEVExpander must correct for this in all cases. 1677f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar InsertPt = &*L->getHeader()->getFirstInsertionPt(); 16780f8cd56bfdd32af4edb253654db02fb3143b25a8Andrew Trick } 167940a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman } else { 168040a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // If the SCEV is computable at this level, insert it into the header 168140a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // after the PHIs (and after any other instructions that we've inserted 168240a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // there) so that it is guaranteed to dominate any user inside the loop. 16835b6f42f57e730c2d968c313a27fa505a3c3e5efaBill Wendling if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L)) 1684f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar InsertPt = &*L->getHeader()->getFirstInsertionPt(); 1685de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar while (InsertPt->getIterator() != Builder.GetInsertPoint() && 1686de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar (isInsertedInstruction(InsertPt) || 1687de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar isa<DbgInfoIntrinsic>(InsertPt))) { 1688f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar InsertPt = &*std::next(InsertPt->getIterator()); 1689b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 169040a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman break; 169140a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman } 169240a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman 1693667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman // Check to see if we already expanded this here. 1694f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar auto I = InsertedExpressions.find(std::make_pair(S, InsertPt)); 1695267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman if (I != InsertedExpressions.end()) 1696667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman return I->second; 1697267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman 1698de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar SCEVInsertPointGuard Guard(Builder, this); 1699f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Builder.SetInsertPoint(InsertPt); 1700667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman 1701667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman // Expand the expression into instructions. 1702de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *V = FindValueInExprValueMap(S, InsertPt); 1703de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 1704de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (!V) 1705de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar V = visit(S); 170640a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman 1707667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman // Remember the expanded value for this SCEV at this location. 170848ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // 170948ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // This is independent of PostIncLoops. The mapped value simply materializes 171048ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // the expression at this insertion point. If the mapped value happened to be 171136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // a postinc expansion, it could be reused by a non-postinc user, but only if 171248ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // its insertion point was already at the head of the loop. 171348ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick InsertedExpressions[std::make_pair(S, InsertPt)] = V; 171496fea337d27357e9b62abbf3d2d5ce29f1c8e870Anton Korobeynikov return V; 171596fea337d27357e9b62abbf3d2d5ce29f1c8e870Anton Korobeynikov} 17161d09de3eca23267855e28297fcb40de3632ea47bDan Gohman 17171d826a76f591afea445489b9a5485c345e66bf87Dan Gohmanvoid SCEVExpander::rememberInstruction(Value *I) { 171825fcaff409f5c4c6da08f148ffb9404a71e8e4a7Dan Gohman if (!PostIncLoops.empty()) 171925fcaff409f5c4c6da08f148ffb9404a71e8e4a7Dan Gohman InsertedPostIncValues.insert(I); 172025fcaff409f5c4c6da08f148ffb9404a71e8e4a7Dan Gohman else 17211d826a76f591afea445489b9a5485c345e66bf87Dan Gohman InsertedValues.insert(I); 17221d826a76f591afea445489b9a5485c345e66bf87Dan Gohman} 17231d826a76f591afea445489b9a5485c345e66bf87Dan Gohman 17241d09de3eca23267855e28297fcb40de3632ea47bDan Gohman/// getOrInsertCanonicalInductionVariable - This method returns the 17251d09de3eca23267855e28297fcb40de3632ea47bDan Gohman/// canonical induction variable of the specified type for the specified 17261d09de3eca23267855e28297fcb40de3632ea47bDan Gohman/// loop (inserting one if there is none). A canonical induction variable 17271d09de3eca23267855e28297fcb40de3632ea47bDan Gohman/// starts at zero and steps by one on each iteration. 17287c58dbd88c36c5d6c411ea6c046ddcff4c5841e9Dan GohmanPHINode * 17291d09de3eca23267855e28297fcb40de3632ea47bDan GohmanSCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L, 1730db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty) { 1731b0bc6c361da9009e8414efde317d9bbff755f6c0Duncan Sands assert(Ty->isIntegerTy() && "Can only insert integer induction variables!"); 1732133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman 1733133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman // Build a SCEV for {0,+,1}<L>. 17343228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick // Conservatively use FlagAnyWrap for now. 1735deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0), 17363228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap); 1737133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman 1738133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman // Emit code for it. 1739de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar SCEVInsertPointGuard Guard(Builder, this); 1740f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar PHINode *V = 1741f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar cast<PHINode>(expandCodeFor(H, nullptr, &L->getHeader()->front())); 1742133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman 174340a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman return V; 17441d09de3eca23267855e28297fcb40de3632ea47bDan Gohman} 1745204494149b6f846e8f173f525b129f5508076049Andrew Trick 1746204494149b6f846e8f173f525b129f5508076049Andrew Trick/// replaceCongruentIVs - Check for congruent phis in this loop header and 1747204494149b6f846e8f173f525b129f5508076049Andrew Trick/// replace them with their most canonical representative. Return the number of 1748204494149b6f846e8f173f525b129f5508076049Andrew Trick/// phis eliminated. 1749204494149b6f846e8f173f525b129f5508076049Andrew Trick/// 1750204494149b6f846e8f173f525b129f5508076049Andrew Trick/// This does not depend on any SCEVExpander state but should be used in 1751204494149b6f846e8f173f525b129f5508076049Andrew Trick/// the same context that SCEVExpander is used. 1752204494149b6f846e8f173f525b129f5508076049Andrew Trickunsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT, 1753a04a4a79ea365d1ba96ed4b5861e879b267162e2Nadav Rotem SmallVectorImpl<WeakVH> &DeadInsts, 1754e4ba75f43e2ab1480d119d2d4eb878256274e0fbChandler Carruth const TargetTransformInfo *TTI) { 1755ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // Find integer phis in order of increasing width. 1756ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick SmallVector<PHINode*, 8> Phis; 1757f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar for (auto &I : *L->getHeader()) { 1758f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (auto *PN = dyn_cast<PHINode>(&I)) 1759f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Phis.push_back(PN); 1760f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar else 1761f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar break; 1762ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick } 1763f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 1764e4ba75f43e2ab1480d119d2d4eb878256274e0fbChandler Carruth if (TTI) 176536b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines std::sort(Phis.begin(), Phis.end(), [](Value *LHS, Value *RHS) { 176636b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines // Put pointers at the back and make sure pointer < pointer = false. 176736b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines if (!LHS->getType()->isIntegerTy() || !RHS->getType()->isIntegerTy()) 176836b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines return RHS->getType()->isIntegerTy() && !LHS->getType()->isIntegerTy(); 176936b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines return RHS->getType()->getPrimitiveSizeInBits() < 177036b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines LHS->getType()->getPrimitiveSizeInBits(); 177136b56886974eae4f9c5ebc96befd3e7bfe5de338Stephen Hines }); 1772ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick 1773204494149b6f846e8f173f525b129f5508076049Andrew Trick unsigned NumElim = 0; 1774204494149b6f846e8f173f525b129f5508076049Andrew Trick DenseMap<const SCEV *, PHINode *> ExprToIVMap; 1775f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar // Process phis from wide to narrow. Map wide phis to their truncation 1776ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // so narrow phis can reuse them. 1777f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar for (PHINode *Phi : Phis) { 1778f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar auto SimplifyPHINode = [&](PHINode *PN) -> Value * { 1779f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (Value *V = SimplifyInstruction(PN, DL, &SE.TLI, &SE.DT, &SE.AC)) 1780f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return V; 1781f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (!SE.isSCEVable(PN->getType())) 1782f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return nullptr; 1783f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar auto *Const = dyn_cast<SCEVConstant>(SE.getSCEV(PN)); 1784f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (!Const) 1785f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return nullptr; 1786f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return Const->getValue(); 1787f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar }; 1788ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick 1789239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer // Fold constant phis. They may be congruent to other constant phis and 1790239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer // would confuse the logic below that expects proper IVs. 1791f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (Value *V = SimplifyPHINode(Phi)) { 1792f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (V->getType() != Phi->getType()) 1793f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar continue; 1794239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer Phi->replaceAllUsesWith(V); 17956948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar DeadInsts.emplace_back(Phi); 1796239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer ++NumElim; 1797239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer DEBUG_WITH_TYPE(DebugType, dbgs() 1798239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer << "INDVARS: Eliminated constant iv: " << *Phi << '\n'); 1799239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer continue; 1800239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer } 1801239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer 1802204494149b6f846e8f173f525b129f5508076049Andrew Trick if (!SE.isSCEVable(Phi->getType())) 1803204494149b6f846e8f173f525b129f5508076049Andrew Trick continue; 1804204494149b6f846e8f173f525b129f5508076049Andrew Trick 1805204494149b6f846e8f173f525b129f5508076049Andrew Trick PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)]; 1806204494149b6f846e8f173f525b129f5508076049Andrew Trick if (!OrigPhiRef) { 1807204494149b6f846e8f173f525b129f5508076049Andrew Trick OrigPhiRef = Phi; 1808de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (Phi->getType()->isIntegerTy() && TTI && 1809de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) { 1810ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // This phi can be freely truncated to the narrowest phi type. Map the 1811ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // truncated expression to it so it will be reused for narrow types. 1812ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick const SCEV *TruncExpr = 1813ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType()); 1814ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick ExprToIVMap[TruncExpr] = Phi; 1815ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick } 1816204494149b6f846e8f173f525b129f5508076049Andrew Trick continue; 1817204494149b6f846e8f173f525b129f5508076049Andrew Trick } 1818204494149b6f846e8f173f525b129f5508076049Andrew Trick 1819ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // Replacing a pointer phi with an integer phi or vice-versa doesn't make 1820ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // sense. 1821ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy()) 1822204494149b6f846e8f173f525b129f5508076049Andrew Trick continue; 1823204494149b6f846e8f173f525b129f5508076049Andrew Trick 1824204494149b6f846e8f173f525b129f5508076049Andrew Trick if (BasicBlock *LatchBlock = L->getLoopLatch()) { 1825de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Instruction *OrigInc = dyn_cast<Instruction>( 1826de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar OrigPhiRef->getIncomingValueForBlock(LatchBlock)); 1827204494149b6f846e8f173f525b129f5508076049Andrew Trick Instruction *IsomorphicInc = 1828de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar dyn_cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock)); 1829de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 1830de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (OrigInc && IsomorphicInc) { 1831de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // If this phi has the same width but is more canonical, replace the 1832de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // original with it. As part of the "more canonical" determination, 1833de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // respect a prior decision to use an IV chain. 1834de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (OrigPhiRef->getType() == Phi->getType() && 1835de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar !(ChainedPhis.count(Phi) || 1836de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L)) && 1837de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar (ChainedPhis.count(Phi) || 1838de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) { 1839de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar std::swap(OrigPhiRef, Phi); 1840de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar std::swap(OrigInc, IsomorphicInc); 1841de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar } 1842de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Replacing the congruent phi is sufficient because acyclic 1843de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // redundancy elimination, CSE/GVN, should handle the 1844de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // rest. However, once SCEV proves that a phi is congruent, 1845de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // it's often the head of an IV user cycle that is isomorphic 1846de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // with the original phi. It's worth eagerly cleaning up the 1847de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // common case of a single IV increment so that DeleteDeadPHIs 1848de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // can remove cycles that had postinc uses. 1849de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar const SCEV *TruncExpr = 1850de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar SE.getTruncateOrNoop(SE.getSCEV(OrigInc), IsomorphicInc->getType()); 1851de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (OrigInc != IsomorphicInc && 1852de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar TruncExpr == SE.getSCEV(IsomorphicInc) && 1853de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar SE.LI.replacementPreservesLCSSAForm(IsomorphicInc, OrigInc) && 1854de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar hoistIVInc(OrigInc, IsomorphicInc)) { 1855de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar DEBUG_WITH_TYPE(DebugType, 1856de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar dbgs() << "INDVARS: Eliminated congruent iv.inc: " 1857de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar << *IsomorphicInc << '\n'); 1858de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *NewInc = OrigInc; 1859de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (OrigInc->getType() != IsomorphicInc->getType()) { 1860de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Instruction *IP = nullptr; 1861de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (PHINode *PN = dyn_cast<PHINode>(OrigInc)) 1862de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar IP = &*PN->getParent()->getFirstInsertionPt(); 1863de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar else 1864de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar IP = OrigInc->getNextNode(); 1865de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 1866de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar IRBuilder<> Builder(IP); 1867de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc()); 1868de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar NewInc = Builder.CreateTruncOrBitCast( 1869de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar OrigInc, IsomorphicInc->getType(), IVName); 1870de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar } 1871de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar IsomorphicInc->replaceAllUsesWith(NewInc); 1872de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar DeadInsts.emplace_back(IsomorphicInc); 1873ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick } 1874204494149b6f846e8f173f525b129f5508076049Andrew Trick } 1875204494149b6f846e8f173f525b129f5508076049Andrew Trick } 1876de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar DEBUG_WITH_TYPE(DebugType, dbgs() << "INDVARS: Eliminated congruent iv: " 1877de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar << *Phi << '\n'); 1878204494149b6f846e8f173f525b129f5508076049Andrew Trick ++NumElim; 1879ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick Value *NewIV = OrigPhiRef; 1880ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick if (OrigPhiRef->getType() != Phi->getType()) { 1881f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar IRBuilder<> Builder(&*L->getHeader()->getFirstInsertionPt()); 1882ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick Builder.SetCurrentDebugLocation(Phi->getDebugLoc()); 1883ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName); 1884ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick } 1885ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick Phi->replaceAllUsesWith(NewIV); 18866948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar DeadInsts.emplace_back(Phi); 1887204494149b6f846e8f173f525b129f5508076049Andrew Trick } 1888204494149b6f846e8f173f525b129f5508076049Andrew Trick return NumElim; 1889204494149b6f846e8f173f525b129f5508076049Andrew Trick} 1890e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick 1891f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga NainarValue *SCEVExpander::findExistingExpansion(const SCEV *S, 1892f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar const Instruction *At, Loop *L) { 1893f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar using namespace llvm::PatternMatch; 1894f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 1895f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar SmallVector<BasicBlock *, 4> ExitingBlocks; 1896f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar L->getExitingBlocks(ExitingBlocks); 1897f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 1898f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar // Look for suitable value in simple conditions at the loop exits. 1899f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar for (BasicBlock *BB : ExitingBlocks) { 1900f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar ICmpInst::Predicate Pred; 1901f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Instruction *LHS, *RHS; 1902f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar BasicBlock *TrueBB, *FalseBB; 1903f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 1904f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (!match(BB->getTerminator(), 1905f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar m_Br(m_ICmp(Pred, m_Instruction(LHS), m_Instruction(RHS)), 1906f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar TrueBB, FalseBB))) 1907f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar continue; 1908f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 1909f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (SE.getSCEV(LHS) == S && SE.DT.dominates(LHS, At)) 1910f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return LHS; 1911f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 1912f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (SE.getSCEV(RHS) == S && SE.DT.dominates(RHS, At)) 1913f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return RHS; 1914f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar } 1915f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 1916de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Use expand's logic which is used for reusing a previous Value in 1917de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // ExprValueMap. 1918de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (Value *Val = FindValueInExprValueMap(S, At)) 1919de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar return Val; 1920de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 1921f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar // There is potential to make this significantly smarter, but this simple 1922f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar // heuristic already gets some interesting cases. 1923f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 1924f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar // Can not find suitable value. 1925f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return nullptr; 1926f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar} 1927f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 19280c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainarbool SCEVExpander::isHighCostExpansionHelper( 1929f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar const SCEV *S, Loop *L, const Instruction *At, 1930f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar SmallPtrSetImpl<const SCEV *> &Processed) { 1931f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 1932f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar // If we can find an existing value for this scev avaliable at the point "At" 1933f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar // then consider the expression cheap. 1934f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (At && findExistingExpansion(S, At, L) != nullptr) 1935f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return false; 19366948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar 19376948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar // Zero/One operand expressions 19386948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar switch (S->getSCEVType()) { 19396948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar case scUnknown: 19406948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar case scConstant: 19416948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar return false; 19426948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar case scTruncate: 1943f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return isHighCostExpansionHelper(cast<SCEVTruncateExpr>(S)->getOperand(), 1944f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar L, At, Processed); 19456948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar case scZeroExtend: 19466948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar return isHighCostExpansionHelper(cast<SCEVZeroExtendExpr>(S)->getOperand(), 1947f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar L, At, Processed); 19486948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar case scSignExtend: 19496948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar return isHighCostExpansionHelper(cast<SCEVSignExtendExpr>(S)->getOperand(), 1950f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar L, At, Processed); 19516948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar } 19526948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar 19530c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar if (!Processed.insert(S).second) 19540c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar return false; 19550c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar 19560c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar if (auto *UDivExpr = dyn_cast<SCEVUDivExpr>(S)) { 19570c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // If the divisor is a power of two and the SCEV type fits in a native 1958f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar // integer, consider the division cheap irrespective of whether it occurs in 19590c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // the user code since it can be lowered into a right shift. 19600c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar if (auto *SC = dyn_cast<SCEVConstant>(UDivExpr->getRHS())) 1961f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (SC->getAPInt().isPowerOf2()) { 19620c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar const DataLayout &DL = 19630c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar L->getHeader()->getParent()->getParent()->getDataLayout(); 19640c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar unsigned Width = cast<IntegerType>(UDivExpr->getType())->getBitWidth(); 19650c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar return DL.isIllegalInteger(Width); 19660c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar } 19670c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar 19680c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // UDivExpr is very likely a UDiv that ScalarEvolution's HowFarToZero or 19690c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // HowManyLessThans produced to compute a precise expression, rather than a 19700c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // UDiv from the user's code. If we can't find a UDiv in the code with some 19710c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // simple searching, assume the former consider UDivExpr expensive to 19720c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // compute. 19730c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar BasicBlock *ExitingBB = L->getExitingBlock(); 19740c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar if (!ExitingBB) 19750c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar return true; 19760c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar 1977f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar // At the beginning of this function we already tried to find existing value 1978f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar // for plain 'S'. Now try to lookup 'S + 1' since it is common pattern 1979f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar // involving division. This is just a simple search heuristic. 1980f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (!At) 1981f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar At = &ExitingBB->back(); 1982f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (!findExistingExpansion( 1983f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar SE.getAddExpr(S, SE.getConstant(S->getType(), 1)), At, L)) 19840c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar return true; 19850c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar } 19860c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar 19876948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar // HowManyLessThans uses a Max expression whenever the loop is not guarded by 19886948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar // the exit condition. 19896948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S)) 19906948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar return true; 19916948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar 19926948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar // Recurse past nary expressions, which commonly occur in the 19930c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // BackedgeTakenCount. They may already exist in program code, and if not, 19940c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // they are not too expensive rematerialize. 19956948897e478cbd66626159776a8017b3c18579b9Pirama Arumuga Nainar if (const SCEVNAryExpr *NAry = dyn_cast<SCEVNAryExpr>(S)) { 1996f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar for (auto *Op : NAry->operands()) 1997f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar if (isHighCostExpansionHelper(Op, L, At, Processed)) 19980c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar return true; 19990c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar } 20000c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar 20010c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // If we haven't recognized an expensive SCEV pattern, assume it's an 20020c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar // expression produced by program code. 20030c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar return false; 20040c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar} 20050c7f116bb6950ef819323d855415b2f2b0aad987Pirama Arumuga Nainar 2006f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga NainarValue *SCEVExpander::expandCodeForPredicate(const SCEVPredicate *Pred, 2007f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Instruction *IP) { 2008f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar assert(IP); 2009f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar switch (Pred->getKind()) { 2010f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar case SCEVPredicate::P_Union: 2011f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return expandUnionPredicate(cast<SCEVUnionPredicate>(Pred), IP); 2012f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar case SCEVPredicate::P_Equal: 2013f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return expandEqualPredicate(cast<SCEVEqualPredicate>(Pred), IP); 2014de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar case SCEVPredicate::P_Wrap: { 2015de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar auto *AddRecPred = cast<SCEVWrapPredicate>(Pred); 2016de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar return expandWrapPredicate(AddRecPred, IP); 2017de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar } 2018f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar } 2019f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar llvm_unreachable("Unknown SCEV predicate type"); 2020f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar} 2021f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 2022f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga NainarValue *SCEVExpander::expandEqualPredicate(const SCEVEqualPredicate *Pred, 2023f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Instruction *IP) { 2024f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Value *Expr0 = expandCodeFor(Pred->getLHS(), Pred->getLHS()->getType(), IP); 2025f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Value *Expr1 = expandCodeFor(Pred->getRHS(), Pred->getRHS()->getType(), IP); 2026f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 2027f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Builder.SetInsertPoint(IP); 2028f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar auto *I = Builder.CreateICmpNE(Expr0, Expr1, "ident.check"); 2029f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return I; 2030f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar} 2031f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 2032de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga NainarValue *SCEVExpander::generateOverflowCheck(const SCEVAddRecExpr *AR, 2033de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Instruction *Loc, bool Signed) { 2034de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar assert(AR->isAffine() && "Cannot generate RT check for " 2035de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar "non-affine expression"); 2036de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2037de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar SCEVUnionPredicate Pred; 2038de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar const SCEV *ExitCount = 2039de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar SE.getPredicatedBackedgeTakenCount(AR->getLoop(), Pred); 2040de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2041de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar assert(ExitCount != SE.getCouldNotCompute() && "Invalid loop count"); 2042de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2043de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar const SCEV *Step = AR->getStepRecurrence(SE); 2044de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar const SCEV *Start = AR->getStart(); 2045de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2046de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar unsigned SrcBits = SE.getTypeSizeInBits(ExitCount->getType()); 2047de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar unsigned DstBits = SE.getTypeSizeInBits(AR->getType()); 2048de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2049de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // The expression {Start,+,Step} has nusw/nssw if 2050de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Step < 0, Start - |Step| * Backedge <= Start 2051de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Step >= 0, Start + |Step| * Backedge > Start 2052de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // and |Step| * Backedge doesn't unsigned overflow. 2053de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2054de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar IntegerType *CountTy = IntegerType::get(Loc->getContext(), SrcBits); 2055de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Builder.SetInsertPoint(Loc); 2056de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *TripCountVal = expandCodeFor(ExitCount, CountTy, Loc); 2057de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2058de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar IntegerType *Ty = 2059de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar IntegerType::get(Loc->getContext(), SE.getTypeSizeInBits(AR->getType())); 2060de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2061de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *StepValue = expandCodeFor(Step, Ty, Loc); 2062de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *NegStepValue = expandCodeFor(SE.getNegativeSCEV(Step), Ty, Loc); 2063de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *StartValue = expandCodeFor(Start, Ty, Loc); 2064de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2065de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar ConstantInt *Zero = 2066de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar ConstantInt::get(Loc->getContext(), APInt::getNullValue(DstBits)); 2067de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2068de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Builder.SetInsertPoint(Loc); 2069de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Compute |Step| 2070de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *StepCompare = Builder.CreateICmp(ICmpInst::ICMP_SLT, StepValue, Zero); 2071de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *AbsStep = Builder.CreateSelect(StepCompare, NegStepValue, StepValue); 2072de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2073de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Get the backedge taken count and truncate or extended to the AR type. 2074de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *TruncTripCount = Builder.CreateZExtOrTrunc(TripCountVal, Ty); 2075de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar auto *MulF = Intrinsic::getDeclaration(Loc->getModule(), 2076de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Intrinsic::umul_with_overflow, Ty); 2077de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2078de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Compute |Step| * Backedge 2079de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar CallInst *Mul = Builder.CreateCall(MulF, {AbsStep, TruncTripCount}, "mul"); 2080de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *MulV = Builder.CreateExtractValue(Mul, 0, "mul.result"); 2081de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *OfMul = Builder.CreateExtractValue(Mul, 1, "mul.overflow"); 2082de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2083de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Compute: 2084de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Start + |Step| * Backedge < Start 2085de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Start - |Step| * Backedge > Start 2086de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *Add = Builder.CreateAdd(StartValue, MulV); 2087de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *Sub = Builder.CreateSub(StartValue, MulV); 2088de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2089de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *EndCompareGT = Builder.CreateICmp( 2090de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT, Sub, StartValue); 2091de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2092de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *EndCompareLT = Builder.CreateICmp( 2093de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT, Add, StartValue); 2094de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2095de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Select the answer based on the sign of Step. 2096de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *EndCheck = 2097de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Builder.CreateSelect(StepCompare, EndCompareGT, EndCompareLT); 2098de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2099de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // If the backedge taken count type is larger than the AR type, 2100de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // check that we don't drop any bits by truncating it. If we are 2101de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // droping bits, then we have overflow (unless the step is zero). 2102de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (SE.getTypeSizeInBits(CountTy) > SE.getTypeSizeInBits(Ty)) { 2103de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar auto MaxVal = APInt::getMaxValue(DstBits).zext(SrcBits); 2104de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar auto *BackedgeCheck = 2105de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Builder.CreateICmp(ICmpInst::ICMP_UGT, TripCountVal, 2106de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar ConstantInt::get(Loc->getContext(), MaxVal)); 2107de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar BackedgeCheck = Builder.CreateAnd( 2108de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar BackedgeCheck, Builder.CreateICmp(ICmpInst::ICMP_NE, StepValue, Zero)); 2109de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2110de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar EndCheck = Builder.CreateOr(EndCheck, BackedgeCheck); 2111de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar } 2112de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2113de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar EndCheck = Builder.CreateOr(EndCheck, OfMul); 2114de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar return EndCheck; 2115de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar} 2116de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2117de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga NainarValue *SCEVExpander::expandWrapPredicate(const SCEVWrapPredicate *Pred, 2118de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Instruction *IP) { 2119de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar const auto *A = cast<SCEVAddRecExpr>(Pred->getExpr()); 2120de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar Value *NSSWCheck = nullptr, *NUSWCheck = nullptr; 2121de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2122de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Add a check for NUSW 2123de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (Pred->getFlags() & SCEVWrapPredicate::IncrementNUSW) 2124de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar NUSWCheck = generateOverflowCheck(A, IP, false); 2125de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2126de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar // Add a check for NSSW 2127de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (Pred->getFlags() & SCEVWrapPredicate::IncrementNSSW) 2128de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar NSSWCheck = generateOverflowCheck(A, IP, true); 2129de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2130de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (NUSWCheck && NSSWCheck) 2131de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar return Builder.CreateOr(NUSWCheck, NSSWCheck); 2132de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2133de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (NUSWCheck) 2134de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar return NUSWCheck; 2135de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2136de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar if (NSSWCheck) 2137de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar return NSSWCheck; 2138de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2139de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar return ConstantInt::getFalse(IP->getContext()); 2140de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar} 2141de2d8694e25a814696358e95141f4b1aa4d8847ePirama Arumuga Nainar 2142f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga NainarValue *SCEVExpander::expandUnionPredicate(const SCEVUnionPredicate *Union, 2143f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Instruction *IP) { 2144f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar auto *BoolType = IntegerType::get(IP->getContext(), 1); 2145f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Value *Check = ConstantInt::getNullValue(BoolType); 2146f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 2147f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar // Loop over all checks in this set. 2148f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar for (auto Pred : Union->getPredicates()) { 2149f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar auto *NextCheck = expandCodeForPredicate(Pred, IP); 2150f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Builder.SetInsertPoint(IP); 2151f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar Check = Builder.CreateOr(Check, NextCheck); 2152f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar } 2153f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 2154f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar return Check; 2155f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar} 2156f3ef5332fa3f4d5ec72c178a2b19dac363a19383Pirama Arumuga Nainar 2157e08c32249fca32cd7b122024a4ca252fcb235694Andrew Tricknamespace { 2158e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// Search for a SCEV subexpression that is not safe to expand. Any expression 2159e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely 2160e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// UDiv expressions. We don't know if the UDiv is derived from an IR divide 2161e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// instruction, but the important thing is that we prove the denominator is 2162e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// nonzero before expansion. 2163e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// 2164e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// IVUsers already checks that IV-derived expressions are safe. So this check is 2165e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// only needed when the expression includes some subexpression that is not IV 2166e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// derived. 2167e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// 2168e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// Currently, we only allow division by a nonzero constant here. If this is 2169e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// inadequate, we could easily allow division by SCEVUnknown by using 2170e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// ValueTracking to check isKnownNonZero(). 21714d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick// 21724d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick// We cannot generally expand recurrences unless the step dominates the loop 21734d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick// header. The expander handles the special case of affine recurrences by 21744d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick// scaling the recurrence outside the loop, but this technique isn't generally 21754d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick// applicable. Expanding a nested recurrence outside a loop requires computing 21764d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick// binomial coefficients. This could be done, but the recurrence has to be in a 21774d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick// perfectly reduced form, which can't be guaranteed. 2178e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trickstruct SCEVFindUnsafe { 21794d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick ScalarEvolution &SE; 2180e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick bool IsUnsafe; 2181e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick 21824d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick SCEVFindUnsafe(ScalarEvolution &se): SE(se), IsUnsafe(false) {} 2183e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick 2184e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick bool follow(const SCEV *S) { 21854d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 21864d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS()); 21874d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick if (!SC || SC->getValue()->isZero()) { 21884d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick IsUnsafe = true; 21894d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick return false; 21904d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick } 21914d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick } 21924d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 21934d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick const SCEV *Step = AR->getStepRecurrence(SE); 21944d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick if (!AR->isAffine() && !SE.dominates(Step, AR->getLoop()->getHeader())) { 21954d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick IsUnsafe = true; 21964d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick return false; 21974d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick } 21984d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick } 21994d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick return true; 2200e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick } 2201e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick bool isDone() const { return IsUnsafe; } 2202e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick}; 2203e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick} 2204e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick 2205e08c32249fca32cd7b122024a4ca252fcb235694Andrew Tricknamespace llvm { 22064d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trickbool isSafeToExpand(const SCEV *S, ScalarEvolution &SE) { 22074d4bbaf997c16f9e79503bd640306d784efd090eAndrew Trick SCEVFindUnsafe Search(SE); 2208e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick visitAll(S, Search); 2209e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick return !Search.IsUnsafe; 2210e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick} 2211e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick} 2212