ScalarEvolutionExpander.cpp revision 0b8c9a80f20772c3793201ab5b251d3520b9cea3
136f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman//===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis --*- C++ -*-===// 236f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// 336f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// The LLVM Compiler Infrastructure 436f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// 54ee451de366474b9c228b4e5fa573795a715216dChris Lattner// This file is distributed under the University of Illinois Open Source 64ee451de366474b9c228b4e5fa573795a715216dChris Lattner// License. See LICENSE.TXT for details. 736f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// 836f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman//===----------------------------------------------------------------------===// 936f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// 1036f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// This file contains the implementation of the scalar evolution expander, 1136f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// which is used to generate the code corresponding to a given scalar evolution 1236f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// expression. 1336f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman// 1436f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman//===----------------------------------------------------------------------===// 1536f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 1636f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman#include "llvm/Analysis/ScalarEvolutionExpander.h" 17d04a8d4b33ff316ca4cf961e06c9e312eff8e64fChandler Carruth#include "llvm/ADT/STLExtras.h" 18e81561909d128c6e2d8033cb5465a49b2596b26aBill Wendling#include "llvm/Analysis/LoopInfo.h" 190b8c9a80f20772c3793201ab5b251d3520b9cea3Chandler Carruth#include "llvm/IR/DataLayout.h" 200b8c9a80f20772c3793201ab5b251d3520b9cea3Chandler Carruth#include "llvm/IR/IntrinsicInst.h" 210b8c9a80f20772c3793201ab5b251d3520b9cea3Chandler Carruth#include "llvm/IR/LLVMContext.h" 22c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick#include "llvm/Support/Debug.h" 23a04a4a79ea365d1ba96ed4b5861e879b267162e2Nadav Rotem#include "llvm/Target/TargetLowering.h" 24d152d03a476b8d0d4b26577db26e2ba76034b0f3Andrew Trick 2536f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begemanusing namespace llvm; 2636f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 2719e5ada58a8a07d3d6d67312550f6d0791d84c3cGabor Greif/// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP, 28485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman/// reusing an existing cast if a suitable one exists, moving an existing 29485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman/// cast if a suitable one exists but isn't in the right place, or 3019e5ada58a8a07d3d6d67312550f6d0791d84c3cGabor Greif/// creating a new one. 31db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris LattnerValue *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty, 32485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman Instruction::CastOps Op, 33485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman BasicBlock::iterator IP) { 34919a503458c18472984c7c37b305110c9ba3a782Rafael Espindola // This function must be called with the builder having a valid insertion 35919a503458c18472984c7c37b305110c9ba3a782Rafael Espindola // point. It doesn't need to be the actual IP where the uses of the returned 36919a503458c18472984c7c37b305110c9ba3a782Rafael Espindola // cast will be added, but it must dominate such IP. 3723b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola // We use this precondition to produce a cast that will dominate all its 3823b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola // uses. In particular, this is crucial for the case where the builder's 3923b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola // insertion point *is* the point where we were asked to put the cast. 40c8e41c591741b3da1077f7000274ad040bef8002Sylvestre Ledru // Since we don't know the builder's insertion point is actually 41919a503458c18472984c7c37b305110c9ba3a782Rafael Espindola // where the uses will be added (only that it dominates it), we are 42919a503458c18472984c7c37b305110c9ba3a782Rafael Espindola // not allowed to move it. 43919a503458c18472984c7c37b305110c9ba3a782Rafael Espindola BasicBlock::iterator BIP = Builder.GetInsertPoint(); 44919a503458c18472984c7c37b305110c9ba3a782Rafael Espindola 4523b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola Instruction *Ret = NULL; 46ef4c80e07baf02dd2a8f08db49c5634a06d3ca1eRafael Espindola 47485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // Check to see if there is already a cast! 48485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); 49f64f9cf7ec39ae91bca84dc6ad3c8fc3343b358aGabor Greif UI != E; ++UI) { 50f64f9cf7ec39ae91bca84dc6ad3c8fc3343b358aGabor Greif User *U = *UI; 51f64f9cf7ec39ae91bca84dc6ad3c8fc3343b358aGabor Greif if (U->getType() == Ty) 5219e5ada58a8a07d3d6d67312550f6d0791d84c3cGabor Greif if (CastInst *CI = dyn_cast<CastInst>(U)) 53485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman if (CI->getOpcode() == Op) { 54b84d540bc9d70de00b791d0e0deab410d6ba3ad5Rafael Espindola // If the cast isn't where we want it, create a new cast at IP. 55b84d540bc9d70de00b791d0e0deab410d6ba3ad5Rafael Espindola // Likewise, do not reuse a cast at BIP because it must dominate 56b84d540bc9d70de00b791d0e0deab410d6ba3ad5Rafael Espindola // instructions that might be inserted before BIP. 57919a503458c18472984c7c37b305110c9ba3a782Rafael Espindola if (BasicBlock::iterator(CI) != IP || BIP == IP) { 58485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // Create a new cast, and leave the old cast in place in case 59485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // it is being used as an insert point. Clear its operand 60485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // so that it doesn't hold anything live. 6123b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola Ret = CastInst::Create(Op, V, Ty, "", IP); 6223b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola Ret->takeName(CI); 6323b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola CI->replaceAllUsesWith(Ret); 64485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman CI->setOperand(0, UndefValue::get(V->getType())); 6523b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola break; 66485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman } 6723b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola Ret = CI; 6823b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola break; 69485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman } 70f64f9cf7ec39ae91bca84dc6ad3c8fc3343b358aGabor Greif } 71485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman 72485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // Create a new cast. 7323b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola if (!Ret) 7423b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola Ret = CastInst::Create(Op, V, Ty, V->getName(), IP); 7523b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola 7623b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola // We assert at the end of the function since IP might point to an 7723b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola // instruction with different dominance properties than a cast 7823b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola // (an invoke for example) and not dominate BIP (but the cast does). 7923b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola assert(SE.DT->dominates(Ret, BIP)); 8023b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola 8123b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola rememberInstruction(Ret); 8223b6ec906a081259dad4672ec386ddfb52cd0d9fRafael Espindola return Ret; 83485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman} 84485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman 85267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman/// InsertNoopCastOfTo - Insert a cast of V to the specified type, 86267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman/// which must be possible with a noop cast, doing what we can to share 87267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman/// the casts. 88db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris LattnerValue *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) { 89267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false); 90267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman assert((Op == Instruction::BitCast || 91267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Op == Instruction::PtrToInt || 92267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Op == Instruction::IntToPtr) && 93267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman "InsertNoopCastOfTo cannot perform non-noop casts!"); 94267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) && 95267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman "InsertNoopCastOfTo cannot change sizes!"); 96267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman 972d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman // Short-circuit unnecessary bitcasts. 9819154f457696f286b4a597179fa07e3b38ea0310Andrew Trick if (Op == Instruction::BitCast) { 9919154f457696f286b4a597179fa07e3b38ea0310Andrew Trick if (V->getType() == Ty) 10019154f457696f286b4a597179fa07e3b38ea0310Andrew Trick return V; 10119154f457696f286b4a597179fa07e3b38ea0310Andrew Trick if (CastInst *CI = dyn_cast<CastInst>(V)) { 10219154f457696f286b4a597179fa07e3b38ea0310Andrew Trick if (CI->getOperand(0)->getType() == Ty) 10319154f457696f286b4a597179fa07e3b38ea0310Andrew Trick return CI->getOperand(0); 10419154f457696f286b4a597179fa07e3b38ea0310Andrew Trick } 10519154f457696f286b4a597179fa07e3b38ea0310Andrew Trick } 106f04fa483b83227c570bc58e1684ea096430a6697Dan Gohman // Short-circuit unnecessary inttoptr<->ptrtoint casts. 107267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) && 10880dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) { 109af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman if (CastInst *CI = dyn_cast<CastInst>(V)) 110af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman if ((CI->getOpcode() == Instruction::PtrToInt || 111af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman CI->getOpcode() == Instruction::IntToPtr) && 112af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman SE.getTypeSizeInBits(CI->getType()) == 113af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman SE.getTypeSizeInBits(CI->getOperand(0)->getType())) 114af79fb5f47b0088c6a8973a7fdbaea96973a429dDan Gohman return CI->getOperand(0); 11580dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 11680dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman if ((CE->getOpcode() == Instruction::PtrToInt || 11780dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman CE->getOpcode() == Instruction::IntToPtr) && 11880dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman SE.getTypeSizeInBits(CE->getType()) == 11980dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman SE.getTypeSizeInBits(CE->getOperand(0)->getType())) 12080dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman return CE->getOperand(0); 12180dcdee0f48820ecea86c15768324945bb0d68d1Dan Gohman } 122f04fa483b83227c570bc58e1684ea096430a6697Dan Gohman 123485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // Fold a cast of a constant. 124ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner if (Constant *C = dyn_cast<Constant>(V)) 125baf3c404409d5e47b13984a7f95bfbd6d1f2e79eOwen Anderson return ConstantExpr::getCast(Op, C, Ty); 1264c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman 127485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // Cast the argument at the beginning of the entry block, after 128485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // any bitcasts of other arguments. 129ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner if (Argument *A = dyn_cast<Argument>(V)) { 130485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin(); 131485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman while ((isa<BitCastInst>(IP) && 132485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) && 133485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman cast<BitCastInst>(IP)->getOperand(0) != A) || 134a4c86ab073d4b7a36477fc7c54c9d52499f04586Bill Wendling isa<DbgInfoIntrinsic>(IP) || 135a4c86ab073d4b7a36477fc7c54c9d52499f04586Bill Wendling isa<LandingPadInst>(IP)) 136485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman ++IP; 137485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman return ReuseOrCreateCast(A, Ty, Op, IP); 138ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner } 1393913187bf6135e6a71260c65eed664095c8f9ce9Wojciech Matyjewicz 140485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman // Cast the instruction immediately after the instruction. 141ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner Instruction *I = cast<Instruction>(V); 142ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner BasicBlock::iterator IP = I; ++IP; 143ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner if (InvokeInst *II = dyn_cast<InvokeInst>(I)) 144ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner IP = II->getNormalDest()->begin(); 145ef4c80e07baf02dd2a8f08db49c5634a06d3ca1eRafael Espindola while (isa<PHINode>(IP) || isa<LandingPadInst>(IP)) 146a4c86ab073d4b7a36477fc7c54c9d52499f04586Bill Wendling ++IP; 147485c43fc478d5e16c55e14cb2586b56cc1c4c91fDan Gohman return ReuseOrCreateCast(I, Ty, Op, IP); 148ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner} 149ca1a4bebb3928ba18fb8751e2f0c0f88fad54cfdChris Lattner 1507fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner/// InsertBinop - Insert the specified binary operator, doing a small amount 1517fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner/// of work to avoid inserting an obviously redundant operation. 152267a385342f2e7388f178b327dd87c5f29afd51bDan GohmanValue *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, 153267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Value *LHS, Value *RHS) { 1540f0eb18addc94bf3d8179fbce162c8c4622b9866Dan Gohman // Fold a binop with constant operands. 1550f0eb18addc94bf3d8179fbce162c8c4622b9866Dan Gohman if (Constant *CLHS = dyn_cast<Constant>(LHS)) 1560f0eb18addc94bf3d8179fbce162c8c4622b9866Dan Gohman if (Constant *CRHS = dyn_cast<Constant>(RHS)) 157baf3c404409d5e47b13984a7f95bfbd6d1f2e79eOwen Anderson return ConstantExpr::get(Opcode, CLHS, CRHS); 1580f0eb18addc94bf3d8179fbce162c8c4622b9866Dan Gohman 1597fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner // Do a quick scan to see if we have this binop nearby. If so, reuse it. 1607fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner unsigned ScanLimit = 6; 161267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 162267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman // Scanning starts from the last instruction before the insertion point. 163267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock::iterator IP = Builder.GetInsertPoint(); 164267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman if (IP != BlockBegin) { 1658a08769bad43a22fae2845bb0ba0fd1266cd55c8Wojciech Matyjewicz --IP; 1668a08769bad43a22fae2845bb0ba0fd1266cd55c8Wojciech Matyjewicz for (; ScanLimit; --IP, --ScanLimit) { 1678d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen // Don't count dbg.value against the ScanLimit, to avoid perturbing the 1688d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen // generated code. 1698d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen if (isa<DbgInfoIntrinsic>(IP)) 1708d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen ScanLimit++; 1715be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS && 1725be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman IP->getOperand(1) == RHS) 1735be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman return IP; 1748a08769bad43a22fae2845bb0ba0fd1266cd55c8Wojciech Matyjewicz if (IP == BlockBegin) break; 1758a08769bad43a22fae2845bb0ba0fd1266cd55c8Wojciech Matyjewicz } 1767fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner } 177267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman 178087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Save the original insertion point so we can restore it when we're done. 179087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 180087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 181087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 182087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Move the insertion point out of as many loops as we can. 183087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 184087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break; 185087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman BasicBlock *Preheader = L->getLoopPreheader(); 186087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!Preheader) break; 187087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 188087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Ok, move up a level. 189087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 190087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 191087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 1928a08769bad43a22fae2845bb0ba0fd1266cd55c8Wojciech Matyjewicz // If we haven't found this binop, insert it. 193a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS)); 194df3ad6697bb849e514c0881ca8700ea36678cbdaDevang Patel BO->setDebugLoc(SaveInsertPt->getDebugLoc()); 195a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(BO); 196087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 197087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Restore the original insert point. 198087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (SaveInsertBB) 199087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman restoreInsertPoint(SaveInsertBB, SaveInsertPt); 200087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 201cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman return BO; 2027fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner} 2037fec90ebf4ebe7aa73a6dd7d275c255587c041adChris Lattner 2044a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman/// FactorOutConstant - Test if S is divisible by Factor, using signed 205453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// division. If so, update S with Factor divided out and return true. 2063f46a3abeedba8d517b4182de34c821d752db058Dan Gohman/// S need not be evenly divisible if a reasonable remainder can be 2074a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman/// computed. 208453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made 209453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// unnecessary; in its place, just signed-divide Ops[i] by the scale and 210453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// check to see if the divide was folded. 2110bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohmanstatic bool FactorOutConstant(const SCEV *&S, 2120bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *&Remainder, 213c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *Factor, 214c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ScalarEvolution &SE, 2153574eca1b02600bac4e625297f4ecf745f4c4f32Micah Villmow const DataLayout *TD) { 216453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // Everything is divisible by one. 217c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (Factor->isOne()) 218c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman return true; 219c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 220c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // x/x == 1. 221c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (S == Factor) { 222deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman S = SE.getConstant(S->getType(), 1); 223453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman return true; 224c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 225453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 226453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // For a Constant, check for a multiple of the given factor. 2274a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 228c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // 0/x == 0. 229c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (C->isZero()) 230453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman return true; 231c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Check for divisibility. 232c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) { 233c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ConstantInt *CI = 234c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ConstantInt::get(SE.getContext(), 235c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman C->getValue()->getValue().sdiv( 236c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman FC->getValue()->getValue())); 237c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // If the quotient is zero and the remainder is non-zero, reject 238c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // the value at this scale. It will be considered for subsequent 239c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // smaller scales. 240c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (!CI->isZero()) { 241c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *Div = SE.getConstant(CI); 242c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman S = Div; 243c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman Remainder = 244c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SE.getAddExpr(Remainder, 245c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SE.getConstant(C->getValue()->getValue().srem( 246c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman FC->getValue()->getValue()))); 247c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman return true; 248c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 249453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 2504a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman } 251453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 252453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // In a Mul, check if there is a constant operand which is a multiple 253453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // of the given factor. 254c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) { 255c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (TD) { 2563574eca1b02600bac4e625297f4ecf745f4c4f32Micah Villmow // With DataLayout, the size is known. Check if there is a constant 257c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // operand which is a multiple of the given factor. If so, we can 258c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // factor it. 259c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEVConstant *FC = cast<SCEVConstant>(Factor); 260c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0))) 261c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) { 262f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 263c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman NewMulOps[0] = 264c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SE.getConstant(C->getValue()->getValue().sdiv( 265c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman FC->getValue()->getValue())); 266c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman S = SE.getMulExpr(NewMulOps); 267c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman return true; 268c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 269c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } else { 2703574eca1b02600bac4e625297f4ecf745f4c4f32Micah Villmow // Without DataLayout, check if Factor can be factored out of any of the 271c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Mul's operands. If so, we can just remove it. 272c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) { 273c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *SOp = M->getOperand(i); 274deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman const SCEV *Remainder = SE.getConstant(SOp->getType(), 0); 275c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) && 276c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman Remainder->isZero()) { 277f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 278c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman NewMulOps[i] = SOp; 279c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman S = SE.getMulExpr(NewMulOps); 280c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman return true; 281c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 282453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 283c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 284c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 285453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 286453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // In an AddRec, check if both start and step are divisible. 287453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) { 2880bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *Step = A->getStepRecurrence(SE); 289deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman const SCEV *StepRem = SE.getConstant(Step->getType(), 0); 290c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (!FactorOutConstant(Step, StepRem, Factor, SE, TD)) 2914a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman return false; 2924a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman if (!StepRem->isZero()) 2934a4f767235dac50965fc266fb822d9a2e37d5212Dan Gohman return false; 2940bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *Start = A->getStart(); 295c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (!FactorOutConstant(Start, Remainder, Factor, SE, TD)) 296453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman return false; 2973228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick // FIXME: can use A->getNoWrapFlags(FlagNW) 2983228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick S = SE.getAddRecExpr(Start, Step, A->getLoop(), SCEV::FlagAnyWrap); 299453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman return true; 300453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 301453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 302453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman return false; 303453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman} 304453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 305c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs 306c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// is the number of SCEVAddRecExprs present, which are kept at the end of 307c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// the list. 308c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// 309c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohmanstatic void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops, 310db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty, 311c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ScalarEvolution &SE) { 312c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman unsigned NumAddRecs = 0; 313c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i) 314c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ++NumAddRecs; 315c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Group Ops into non-addrecs and addrecs. 316c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs); 317c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end()); 318c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Let ScalarEvolution sort and simplify the non-addrecs list. 319c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *Sum = NoAddRecs.empty() ? 320deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman SE.getConstant(Ty, 0) : 321c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SE.getAddExpr(NoAddRecs); 322c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // If it returned an add, use the operands. Otherwise it simplified 323c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // the sum into a single value, so just use that. 324f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman Ops.clear(); 325c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum)) 326403a8cdda5e76ea689693de16474650b4b0df818Dan Gohman Ops.append(Add->op_begin(), Add->op_end()); 327f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman else if (!Sum->isZero()) 328f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman Ops.push_back(Sum); 329c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Then append the addrecs. 330403a8cdda5e76ea689693de16474650b4b0df818Dan Gohman Ops.append(AddRecs.begin(), AddRecs.end()); 331c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman} 332c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 333c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// SplitAddRecs - Flatten a list of add operands, moving addrec start values 334c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}. 335c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// This helps expose more opportunities for folding parts of the expressions 336c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// into GEP indices. 337c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman/// 338c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohmanstatic void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops, 339db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty, 340c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ScalarEvolution &SE) { 341c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Find the addrecs. 342c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SmallVector<const SCEV *, 8> AddRecs; 343c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman for (unsigned i = 0, e = Ops.size(); i != e; ++i) 344c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) { 345c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *Start = A->getStart(); 346c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (Start->isZero()) break; 347deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman const SCEV *Zero = SE.getConstant(Ty, 0); 348c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman AddRecs.push_back(SE.getAddRecExpr(Zero, 349c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman A->getStepRecurrence(SE), 3503228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick A->getLoop(), 3513228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick // FIXME: A->getNoWrapFlags(FlagNW) 3523228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick SCEV::FlagAnyWrap)); 353c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) { 354c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman Ops[i] = Zero; 355403a8cdda5e76ea689693de16474650b4b0df818Dan Gohman Ops.append(Add->op_begin(), Add->op_end()); 356c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman e += Add->getNumOperands(); 357c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } else { 358c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman Ops[i] = Start; 359c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 360c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 361c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (!AddRecs.empty()) { 362c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Add the addrecs onto the end of the list. 363403a8cdda5e76ea689693de16474650b4b0df818Dan Gohman Ops.append(AddRecs.begin(), AddRecs.end()); 364c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Resort the operand list, moving any constants to the front. 365c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SimplifyAddOperands(Ops, Ty, SE); 366c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 367c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman} 368c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 3694c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// expandAddToGEP - Expand an addition expression with a pointer type into 3704c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps 3714c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// BasicAliasAnalysis and other passes analyze the result. See the rules 3724c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// for getelementptr vs. inttoptr in 3734c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// http://llvm.org/docs/LangRef.html#pointeraliasing 3744c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// for details. 37513c5e35222afe0895f0c5e68aa9f22f134ea437aDan Gohman/// 3763abf905b50f33340ae81913da81b0eda96fa4616Dan Gohman/// Design note: The correctness of using getelementptr here depends on 3774c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as 3784c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// they may introduce pointer arithmetic which may not be safely converted 3794c0d5d5db876b0628bdf6a2174263a1c0a9130e2Dan Gohman/// into getelementptr. 380453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// 381453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// Design note: It might seem desirable for this function to be more 382453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// loop-aware. If some of the indices are loop-invariant while others 383453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// aren't, it might seem desirable to emit multiple GEPs, keeping the 384453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// loop-invariant portions of the overall computation outside the loop. 385453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// However, there are a few reasons this is not done here. Hoisting simple 386453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// arithmetic is a low-level optimization that often isn't very 387453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// important until late in the optimization process. In fact, passes 388453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// like InstructionCombining will combine GEPs, even if it means 389453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// pushing loop-invariant computation down into loops, so even if the 390453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// GEPs were split here, the work would quickly be undone. The 391453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// LoopStrengthReduction pass, which is usually run quite late (and 392453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// after the last InstructionCombining pass), takes care of hoisting 393453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// loop-invariant portions of expressions, after considering what 394453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// can be folded using target addressing modes. 395453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// 3960bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan GohmanValue *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin, 3970bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *const *op_end, 398db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner PointerType *PTy, 399db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty, 4005be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman Value *V) { 401db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *ElTy = PTy->getElementType(); 4025be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman SmallVector<Value *, 4> GepIndices; 4030bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman SmallVector<const SCEV *, 8> Ops(op_begin, op_end); 4045be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman bool AnyNonZeroIndices = false; 4055be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 406c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Split AddRecs up into parts as either of the parts may be usable 407c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // without the other. 408c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman SplitAddRecs(Ops, Ty, SE); 409c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 410eb35699e642f2c021f04154bdb4bd90a1afb3baaBob Wilson // Descend down the pointer's type and attempt to convert the other 4115be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // operands into GEP indices, at each level. The first index in a GEP 4125be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // indexes into the array implied by the pointer operand; the rest of 4135be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // the indices index into the element or field type selected by the 4145be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // preceding index. 4155be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman for (;;) { 416c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // If the scale size is not 0, attempt to factor out a scale for 417c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // array indexing. 4180bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman SmallVector<const SCEV *, 8> ScaledOps; 419150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman if (ElTy->isSized()) { 420ece6c6bb6329748b92403c06ac87f45c43485911Chandler Carruth const SCEV *ElSize = SE.getSizeOfExpr(ElTy); 421150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman if (!ElSize->isZero()) { 422150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman SmallVector<const SCEV *, 8> NewOps; 423150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 424150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman const SCEV *Op = Ops[i]; 425deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman const SCEV *Remainder = SE.getConstant(Ty, 0); 426150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) { 427150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman // Op now has ElSize factored out. 428150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman ScaledOps.push_back(Op); 429150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman if (!Remainder->isZero()) 430150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman NewOps.push_back(Remainder); 431150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman AnyNonZeroIndices = true; 432150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman } else { 433150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman // The operand was not divisible, so add it to the list of operands 434150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman // we'll scan next iteration. 435150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman NewOps.push_back(Ops[i]); 436150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman } 437150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman } 438150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman // If we made any changes, update Ops. 439150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman if (!ScaledOps.empty()) { 440150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman Ops = NewOps; 441150dfa87113e51881c75676edd8e3549eb702d99Dan Gohman SimplifyAddOperands(Ops, Ty, SE); 4425be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 443c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 4445be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 445c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 446c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Record the scaled array index for this level of the type. If 447c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // we didn't find any operands that could be factored, tentatively 448c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // assume that element zero was selected (since the zero offset 449c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // would obviously be folded away). 4505be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman Value *Scaled = ScaledOps.empty() ? 451a7235ea7245028a0723e8ab7fd011386b3900777Owen Anderson Constant::getNullValue(Ty) : 4525be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman expandCodeFor(SE.getAddExpr(ScaledOps), Ty); 4535be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman GepIndices.push_back(Scaled); 4545be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 4555be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // Collect struct field index operands. 456db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner while (StructType *STy = dyn_cast<StructType>(ElTy)) { 457c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman bool FoundFieldNo = false; 458c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // An empty struct has no fields. 459c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (STy->getNumElements() == 0) break; 460c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (SE.TD) { 4613574eca1b02600bac4e625297f4ecf745f4c4f32Micah Villmow // With DataLayout, field offsets are known. See if a constant offset 462c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // falls within any of the struct fields. 463c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (Ops.empty()) break; 4645be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0])) 4655be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (SE.getTypeSizeInBits(C->getType()) <= 64) { 4665be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman const StructLayout &SL = *SE.TD->getStructLayout(STy); 4675be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman uint64_t FullOffset = C->getValue()->getZExtValue(); 4685be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (FullOffset < SL.getSizeInBytes()) { 4695be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman unsigned ElIdx = SL.getElementContainingOffset(FullOffset); 4701d0be15f89cb5056e20e2d24faa8d6afb1573bcaOwen Anderson GepIndices.push_back( 4711d0be15f89cb5056e20e2d24faa8d6afb1573bcaOwen Anderson ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx)); 4725be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman ElTy = STy->getTypeAtIndex(ElIdx); 4735be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman Ops[0] = 4746de29f8d960505421d61c80cdb738e16720b6c0eDan Gohman SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx)); 4755be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman AnyNonZeroIndices = true; 476c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman FoundFieldNo = true; 4775be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 4785be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 479c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } else { 4803574eca1b02600bac4e625297f4ecf745f4c4f32Micah Villmow // Without DataLayout, just check for an offsetof expression of the 481c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // appropriate struct type. 482c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman for (unsigned i = 0, e = Ops.size(); i != e; ++i) 4830f5efe56258f8cd6ceff4d7955a5d80144cd9cb0Dan Gohman if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) { 484db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *CTy; 4850f5efe56258f8cd6ceff4d7955a5d80144cd9cb0Dan Gohman Constant *FieldNo; 4864f8eea82d8967cffa85b9df6c9255717b059009eDan Gohman if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) { 4870f5efe56258f8cd6ceff4d7955a5d80144cd9cb0Dan Gohman GepIndices.push_back(FieldNo); 4880f5efe56258f8cd6ceff4d7955a5d80144cd9cb0Dan Gohman ElTy = 4890f5efe56258f8cd6ceff4d7955a5d80144cd9cb0Dan Gohman STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue()); 490c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman Ops[i] = SE.getConstant(Ty, 0); 491c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman AnyNonZeroIndices = true; 492c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman FoundFieldNo = true; 493c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman break; 494c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 4950f5efe56258f8cd6ceff4d7955a5d80144cd9cb0Dan Gohman } 4965be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 497c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // If no struct field offsets were found, tentatively assume that 498c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // field zero was selected (since the zero offset would obviously 499c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // be folded away). 500c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (!FoundFieldNo) { 501c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ElTy = STy->getTypeAtIndex(0u); 502c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman GepIndices.push_back( 503c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman Constant::getNullValue(Type::getInt32Ty(Ty->getContext()))); 504c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 505c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 5065be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 507db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) 5085be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman ElTy = ATy->getElementType(); 509c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman else 510c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman break; 5115be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 5125be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 5133f46a3abeedba8d517b4182de34c821d752db058Dan Gohman // If none of the operands were convertible to proper GEP indices, cast 5145be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // the base to i8* and do an ugly getelementptr with that. It's still 5155be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // better than ptrtoint+arithmetic+inttoptr at least. 5165be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (!AnyNonZeroIndices) { 517c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Cast the base to i8*. 5185be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman V = InsertNoopCastOfTo(V, 519ac53a0b272452013124bfc70480aea5e41b60f40Duncan Sands Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace())); 520c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman 521705b48d960dff1a96ac40d0cf932eb465b9a550aRafael Espindola assert(!isa<Instruction>(V) || 522c9ae8cc24c70dda33b68cacf01d2feeeb836f6f2Rafael Espindola SE.DT->dominates(cast<Instruction>(V), Builder.GetInsertPoint())); 5234b04578d65e38cdb5077de2498889e4a174ccdfdRafael Espindola 524c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Expand the operands for a plain byte offset. 52592fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty); 5265be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 5275be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // Fold a GEP with constant operands. 5285be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (Constant *CLHS = dyn_cast<Constant>(V)) 5295be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (Constant *CRHS = dyn_cast<Constant>(Idx)) 530dab3d29605a5c83db41b28176273ef55961120c1Jay Foad return ConstantExpr::getGetElementPtr(CLHS, CRHS); 5315be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 5325be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman // Do a quick scan to see if we have this GEP nearby. If so, reuse it. 5335be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman unsigned ScanLimit = 6; 534267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 535267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman // Scanning starts from the last instruction before the insertion point. 536267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock::iterator IP = Builder.GetInsertPoint(); 537267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman if (IP != BlockBegin) { 5385be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman --IP; 5395be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman for (; ScanLimit; --IP, --ScanLimit) { 5408d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen // Don't count dbg.value against the ScanLimit, to avoid perturbing the 5418d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen // generated code. 5428d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen if (isa<DbgInfoIntrinsic>(IP)) 5438d50ea760322630dbb9ab6436ca356b279adee10Dale Johannesen ScanLimit++; 5445be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (IP->getOpcode() == Instruction::GetElementPtr && 5455be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman IP->getOperand(0) == V && IP->getOperand(1) == Idx) 5465be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman return IP; 5475be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (IP == BlockBegin) break; 5485be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 5495be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 5505be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 551087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Save the original insertion point so we can restore it when we're done. 552087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 553087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 554087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 555087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Move the insertion point out of as many loops as we can. 556087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 557087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break; 558087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman BasicBlock *Preheader = L->getLoopPreheader(); 559087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!Preheader) break; 560087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 561087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Ok, move up a level. 562087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 563087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 564087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 565c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Emit a GEP. 566c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman Value *GEP = Builder.CreateGEP(V, Idx, "uglygep"); 567a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(GEP); 568087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 569087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Restore the original insert point. 570087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (SaveInsertBB) 571087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman restoreInsertPoint(SaveInsertBB, SaveInsertPt); 572087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 5735be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman return GEP; 5745be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 5755be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 576087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Save the original insertion point so we can restore it when we're done. 577087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 578087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 579087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 580087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Move the insertion point out of as many loops as we can. 581087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 582087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!L->isLoopInvariant(V)) break; 583087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 584087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman bool AnyIndexNotLoopInvariant = false; 585087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(), 586087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman E = GepIndices.end(); I != E; ++I) 587087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!L->isLoopInvariant(*I)) { 588087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman AnyIndexNotLoopInvariant = true; 589087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman break; 590087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 591087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (AnyIndexNotLoopInvariant) 592087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman break; 593087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 594087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman BasicBlock *Preheader = L->getLoopPreheader(); 595087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!Preheader) break; 596087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 597087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Ok, move up a level. 598087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 599087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 600087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 601d6aa02de1076c801ac41295156a2379637976918Dan Gohman // Insert a pretty getelementptr. Note that this GEP is not marked inbounds, 602d6aa02de1076c801ac41295156a2379637976918Dan Gohman // because ScalarEvolution may have changed the address arithmetic to 603d6aa02de1076c801ac41295156a2379637976918Dan Gohman // compute a value which is beyond the end of the allocated object. 604a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Value *Casted = V; 605a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman if (V->getType() != PTy) 606a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Casted = InsertNoopCastOfTo(Casted, PTy); 607a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Value *GEP = Builder.CreateGEP(Casted, 6080a2a60ace9b79164b71794ce7ff981171c61e442Jay Foad GepIndices, 609267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman "scevgep"); 6105be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman Ops.push_back(SE.getUnknown(GEP)); 611a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(GEP); 612087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 613087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Restore the original insert point. 614087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (SaveInsertBB) 615087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman restoreInsertPoint(SaveInsertBB, SaveInsertPt); 616087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 6175be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman return expand(SE.getAddExpr(Ops)); 6185be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman} 6195be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 620087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman/// PickMostRelevantLoop - Given two loops pick the one that's most relevant for 621087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman/// SCEV expansion. If they are nested, this is the most nested. If they are 622087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman/// neighboring, pick the later. 623087bd1e3a12893873761736bf0f905a350e9e708Dan Gohmanstatic const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B, 624087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman DominatorTree &DT) { 625087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!A) return B; 626087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!B) return A; 627087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (A->contains(B)) return B; 628087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (B->contains(A)) return A; 629087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (DT.dominates(A->getHeader(), B->getHeader())) return B; 630087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (DT.dominates(B->getHeader(), A->getHeader())) return A; 631087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return A; // Arbitrarily break the tie. 632087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman} 633087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 6349c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman/// getRelevantLoop - Get the most relevant loop associated with the given 635087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman/// expression, according to PickMostRelevantLoop. 6369c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohmanconst Loop *SCEVExpander::getRelevantLoop(const SCEV *S) { 6379c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman // Test whether we've already computed the most relevant loop for this SCEV. 6389c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair = 6399c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman RelevantLoops.insert(std::make_pair(S, static_cast<const Loop *>(0))); 6409c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman if (!Pair.second) 6419c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman return Pair.first->second; 6429c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman 643087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (isa<SCEVConstant>(S)) 6449c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman // A constant has no relevant loops. 645087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return 0; 646087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 647087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (const Instruction *I = dyn_cast<Instruction>(U->getValue())) 6489c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman return Pair.first->second = SE.LI->getLoopFor(I->getParent()); 6499c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman // A non-instruction has no relevant loops. 650087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return 0; 651087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 652087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) { 653087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman const Loop *L = 0; 654087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 655087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman L = AR->getLoop(); 656087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end(); 657087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman I != E; ++I) 6589c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT); 6599c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman return RelevantLoops[N] = L; 6609c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman } 6619c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) { 6629c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman const Loop *Result = getRelevantLoop(C->getOperand()); 6639c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman return RelevantLoops[C] = Result; 6649c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman } 6659c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 6669c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman const Loop *Result = 6679c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman PickMostRelevantLoop(getRelevantLoop(D->getLHS()), 6689c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman getRelevantLoop(D->getRHS()), 6699c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman *SE.DT); 6709c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman return RelevantLoops[D] = Result; 671087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 672087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman llvm_unreachable("Unexpected SCEV type!"); 673087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman} 674c4f7ec85ecb760fff2b702c6deb06506b968ba4fDan Gohman 675b35798347ea87b8b6d36155b211016a7769f01abDan Gohmannamespace { 676b35798347ea87b8b6d36155b211016a7769f01abDan Gohman 677087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman/// LoopCompare - Compare loops by PickMostRelevantLoop. 678087bd1e3a12893873761736bf0f905a350e9e708Dan Gohmanclass LoopCompare { 679087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman DominatorTree &DT; 680087bd1e3a12893873761736bf0f905a350e9e708Dan Gohmanpublic: 681087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman explicit LoopCompare(DominatorTree &dt) : DT(dt) {} 682087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 683087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman bool operator()(std::pair<const Loop *, const SCEV *> LHS, 684087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman std::pair<const Loop *, const SCEV *> RHS) const { 685bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman // Keep pointer operands sorted at the end. 686bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman if (LHS.second->getType()->isPointerTy() != 687bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman RHS.second->getType()->isPointerTy()) 688bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman return LHS.second->getType()->isPointerTy(); 689bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman 690087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Compare loops with PickMostRelevantLoop. 691087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (LHS.first != RHS.first) 692087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first; 693087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 694087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // If one operand is a non-constant negative and the other is not, 695087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // put the non-constant negative on the right so that a sub can 696087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // be used instead of a negate and add. 697f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick if (LHS.second->isNonConstantNegative()) { 698f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick if (!RHS.second->isNonConstantNegative()) 699087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return false; 700f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick } else if (RHS.second->isNonConstantNegative()) 701087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return true; 702087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 703087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Otherwise they are equivalent according to this comparison. 704087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return false; 705c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman } 706087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman}; 7075be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 708b35798347ea87b8b6d36155b211016a7769f01abDan Gohman} 709b35798347ea87b8b6d36155b211016a7769f01abDan Gohman 710087bd1e3a12893873761736bf0f905a350e9e708Dan GohmanValue *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) { 711db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 712e24fa64d52330626553298f56ba5aa702624c282Dan Gohman 713087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Collect all the add operands in a loop, along with their associated loops. 714087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Iterate in reverse so that constants are emitted last, all else equal, and 715087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // so that pointer operands are inserted first, which the code below relies on 716087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // to form more involved GEPs. 717087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 718087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()), 719087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman E(S->op_begin()); I != E; ++I) 7209c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 721087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 722087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Sort by loop. Use a stable sort so that constants follow non-constants and 723087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // pointer operands precede non-pointer operands. 724087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); 725087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 726087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Emit instructions to add all the operands. Hoist as much as possible 727087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // out of loops, and form meaningful getelementptrs where possible. 728087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Value *Sum = 0; 729087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator 730087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { 731087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman const Loop *CurLoop = I->first; 732087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman const SCEV *Op = I->second; 733087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!Sum) { 734087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // This is the first operand. Just expand it. 735087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = expand(Op); 736087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman ++I; 737db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) { 738087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // The running sum expression is a pointer. Try to form a getelementptr 739087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // at this level with that as the base. 740087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman SmallVector<const SCEV *, 4> NewOps; 741bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman for (; I != E && I->first == CurLoop; ++I) { 742bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman // If the operand is SCEVUnknown and not instructions, peek through 743bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman // it, to enable more of it to be folded into the GEP. 744bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman const SCEV *X = I->second; 745bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X)) 746bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman if (!isa<Instruction>(U->getValue())) 747bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman X = SE.getSCEV(U->getValue()); 748bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman NewOps.push_back(X); 749bb5d92741b2835eff25e4524bc6a5b0fb4fda855Dan Gohman } 750087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum); 751db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) { 752087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // The running sum is an integer, and there's a pointer at this level. 753f8d0578e4cbd5922696c92f5068c5513d8e8d60eDan Gohman // Try to form a getelementptr. If the running sum is instructions, 754f8d0578e4cbd5922696c92f5068c5513d8e8d60eDan Gohman // use a SCEVUnknown to avoid re-analyzing them. 755087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman SmallVector<const SCEV *, 4> NewOps; 756f8d0578e4cbd5922696c92f5068c5513d8e8d60eDan Gohman NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) : 757f8d0578e4cbd5922696c92f5068c5513d8e8d60eDan Gohman SE.getSCEV(Sum)); 758087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman for (++I; I != E && I->first == CurLoop; ++I) 759087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman NewOps.push_back(I->second); 760087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op)); 761f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick } else if (Op->isNonConstantNegative()) { 762087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Instead of doing a negate and add, just do a subtract. 763a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty); 764087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = InsertNoopCastOfTo(Sum, Ty); 765087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = InsertBinop(Instruction::Sub, Sum, W); 766087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman ++I; 767a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } else { 768087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // A simple add. 769a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Value *W = expandCodeFor(Op, Ty); 770087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = InsertNoopCastOfTo(Sum, Ty); 771087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Canonicalize a constant to the RHS. 772087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (isa<Constant>(Sum)) std::swap(Sum, W); 773087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Sum = InsertBinop(Instruction::Add, Sum, W); 774087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman ++I; 775a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 7762d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman } 777087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 778087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return Sum; 779e24fa64d52330626553298f56ba5aa702624c282Dan Gohman} 7805be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman 781890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) { 782db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 783087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 784087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Collect all the mul operands in a loop, along with their associated loops. 785087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Iterate in reverse so that constants are emitted last, all else equal. 786087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 787087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()), 788087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman E(S->op_begin()); I != E; ++I) 7899c9fcfc719158a46cb2e41b66d7dc1a63cd48d74Dan Gohman OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 790087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 791087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Sort by loop. Use a stable sort so that constants follow non-constants. 792087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); 793087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman 794087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Emit instructions to mul all the operands. Hoist as much as possible 795087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // out of loops. 796087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Value *Prod = 0; 797087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator 798087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { 799087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman const SCEV *Op = I->second; 800087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (!Prod) { 801087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // This is the first operand. Just expand it. 802087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Prod = expand(Op); 803087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman ++I; 804087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } else if (Op->isAllOnesValue()) { 805087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Instead of doing a multiply by negative one, just do a negate. 806087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Prod = InsertNoopCastOfTo(Prod, Ty); 807087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod); 808087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman ++I; 809087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } else { 810087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // A simple mul. 811087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Value *W = expandCodeFor(Op, Ty); 812087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Prod = InsertNoopCastOfTo(Prod, Ty); 813087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman // Canonicalize a constant to the RHS. 814087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman if (isa<Constant>(Prod)) std::swap(Prod, W); 815087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman Prod = InsertBinop(Instruction::Mul, Prod, W); 816087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman ++I; 817087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman } 8182d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman } 8192d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman 820087bd1e3a12893873761736bf0f905a350e9e708Dan Gohman return Prod; 82136f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman} 82236f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 823890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) { 824db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 8252d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman 82692fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *LHS = expandCodeFor(S->getLHS(), Ty); 827890f92b744fb074465bc2b7006ee753a181f62a4Dan Gohman if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) { 8286177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky const APInt &RHS = SC->getValue()->getValue(); 8296177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky if (RHS.isPowerOf2()) 8306177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky return InsertBinop(Instruction::LShr, LHS, 831eed707b1e6097aac2bb6b3d47271f6300ace7f2eOwen Anderson ConstantInt::get(Ty, RHS.logBase2())); 8326177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky } 8336177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky 83492fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *RHS = expandCodeFor(S->getRHS(), Ty); 835267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman return InsertBinop(Instruction::UDiv, LHS, RHS); 8366177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky} 8376177fd4fcee4d82692c47e33754ffe285c38cc69Nick Lewycky 838453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// Move parts of Base into Rest to leave Base with the minimal 839453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// expression that provides a pointer operand suitable for a 840453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman/// GEP expansion. 8410bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohmanstatic void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest, 842453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman ScalarEvolution &SE) { 843453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) { 844453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman Base = A->getStart(); 845453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman Rest = SE.getAddExpr(Rest, 846deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman SE.getAddRecExpr(SE.getConstant(A->getType(), 0), 847453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman A->getStepRecurrence(SE), 8483228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick A->getLoop(), 8493228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick // FIXME: A->getNoWrapFlags(FlagNW) 8503228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick SCEV::FlagAnyWrap)); 851453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 852453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) { 853453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman Base = A->getOperand(A->getNumOperands()-1); 8540bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end()); 855453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman NewAddOps.back() = Rest; 856453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman Rest = SE.getAddExpr(NewAddOps); 857453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman ExposePointerBase(Base, Rest, SE); 858453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 859453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman} 860453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 861c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick/// Determine if this is a well-behaved chain of instructions leading back to 862c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick/// the PHI. If so, it may be reused by expanded expressions. 863c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trickbool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV, 864c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick const Loop *L) { 865c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) || 866c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV))) 867c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return false; 868c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // If any of the operands don't dominate the insert position, bail. 869c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Addrec operands are always loop-invariant, so this can only happen 870c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // if there are instructions which haven't been hoisted. 871c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (L == IVIncInsertLoop) { 872c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick for (User::op_iterator OI = IncV->op_begin()+1, 873c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick OE = IncV->op_end(); OI != OE; ++OI) 874c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (Instruction *OInst = dyn_cast<Instruction>(OI)) 875c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (!SE.DT->dominates(OInst, IVIncInsertPos)) 876c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return false; 877c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 878c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Advance to the next instruction. 879c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick IncV = dyn_cast<Instruction>(IncV->getOperand(0)); 880c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (!IncV) 881c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return false; 882c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 883c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (IncV->mayHaveSideEffects()) 884c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return false; 885c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 886c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (IncV != PN) 887c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return true; 888c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 889c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return isNormalAddRecExprPHI(PN, IncV, L); 890c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick} 891c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 892b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// getIVIncOperand returns an induction variable increment's induction 893b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// variable operand. 894b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// 895b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// If allowScale is set, any type of GEP is allowed as long as the nonIV 896b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// operands dominate InsertPos. 897b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// 898b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// If allowScale is not set, ensure that a GEP increment conforms to one of the 899b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// simple patterns generated by getAddRecExprPHILiterally and 900b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// expandAddtoGEP. If the pattern isn't recognized, return NULL. 901b5c26ef9da16052597d59a412eaae32098aa1be0Andrew TrickInstruction *SCEVExpander::getIVIncOperand(Instruction *IncV, 902b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick Instruction *InsertPos, 903b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick bool allowScale) { 904b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (IncV == InsertPos) 905b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return NULL; 90664925c55c65f9345a69fb67db07aa62cfb723577Andrew Trick 907c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick switch (IncV->getOpcode()) { 908b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick default: 909b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return NULL; 910c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Check for a simple Add/Sub or GEP of a loop invariant step. 911c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick case Instruction::Add: 912b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick case Instruction::Sub: { 913b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1)); 914c9ae8cc24c70dda33b68cacf01d2feeeb836f6f2Rafael Espindola if (!OInst || SE.DT->dominates(OInst, InsertPos)) 915b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return dyn_cast<Instruction>(IncV->getOperand(0)); 916b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return NULL; 917b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 918c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick case Instruction::BitCast: 919b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return dyn_cast<Instruction>(IncV->getOperand(0)); 920b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick case Instruction::GetElementPtr: 921c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end(); 922c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick I != E; ++I) { 923c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (isa<Constant>(*I)) 924c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick continue; 925b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (Instruction *OInst = dyn_cast<Instruction>(*I)) { 926c9ae8cc24c70dda33b68cacf01d2feeeb836f6f2Rafael Espindola if (!SE.DT->dominates(OInst, InsertPos)) 927b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return NULL; 928b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 929b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (allowScale) { 930b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // allow any kind of GEP as long as it can be hoisted. 931b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick continue; 932b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 933b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // This must be a pointer addition of constants (pretty), which is already 934b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // handled, or some number of address-size elements (ugly). Ugly geps 935b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // have 2 operands. i1* is used by the expander to represent an 936b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // address-size element. 937c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (IncV->getNumOperands() != 2) 938b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return NULL; 939365c9f1ff55bef134c6b9707f7df44d680ddabeaAndrew Trick unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace(); 940c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS) 941c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS)) 942b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return NULL; 943c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick break; 944c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 945b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return dyn_cast<Instruction>(IncV->getOperand(0)); 946c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 947b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick} 948b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick 949b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make 950b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// it available to other uses in this loop. Recursively hoist any operands, 951b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// until we reach a value that dominates InsertPos. 952b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trickbool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) { 953c9ae8cc24c70dda33b68cacf01d2feeeb836f6f2Rafael Espindola if (SE.DT->dominates(IncV, InsertPos)) 954b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return true; 955b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick 956b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // InsertPos must itself dominate IncV so that IncV's new position satisfies 957b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // its existing users. 9583de8ad8bf9081fcb7ec0600fd2a8d1ce809f556dAndrew Trick if (isa<PHINode>(InsertPos) 9593de8ad8bf9081fcb7ec0600fd2a8d1ce809f556dAndrew Trick || !SE.DT->dominates(InsertPos->getParent(), IncV->getParent())) 960c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return false; 961b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick 962b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // Check that the chain of IV operands leading back to Phi can be hoisted. 963b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick SmallVector<Instruction*, 4> IVIncs; 964b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick for(;;) { 965b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true); 966b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (!Oper) 967b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return false; 968b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // IncV is safe to hoist. 969b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick IVIncs.push_back(IncV); 970b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick IncV = Oper; 971c9ae8cc24c70dda33b68cacf01d2feeeb836f6f2Rafael Espindola if (SE.DT->dominates(IncV, InsertPos)) 972b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick break; 973b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 974b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick for (SmallVectorImpl<Instruction*>::reverse_iterator I = IVIncs.rbegin(), 975b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick E = IVIncs.rend(); I != E; ++I) { 976b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick (*I)->moveBefore(InsertPos); 977b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 978b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return true; 979b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick} 980b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick 981b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// Determine if this cyclic phi is in a form that would have been generated by 982b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// LSR. We don't care if the phi was actually expanded in this pass, as long 983b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// as it is in a low-cost form, for example, no implied multiplication. This 984b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// should match any patterns generated by getAddRecExprPHILiterally and 985b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick/// expandAddtoGEP. 986b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trickbool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV, 987b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick const Loop *L) { 988b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick for(Instruction *IVOper = IncV; 989b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(), 990b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick /*allowScale=*/false));) { 991b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (IVOper == PN) 992b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return true; 993c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 994b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick return false; 995c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick} 996c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 997553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick/// expandIVInc - Expand an IV increment at Builder's current InsertPos. 998553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick/// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may 999553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick/// need to materialize IV increments elsewhere to handle difficult situations. 1000553fe05f236f46fe27b7fcfa822b06367d50183eAndrew TrickValue *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L, 1001553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Type *ExpandTy, Type *IntTy, 1002553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick bool useSubtract) { 1003553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Value *IncV; 1004553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // If the PHI is a pointer, use a GEP, otherwise use an add or sub. 1005553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick if (ExpandTy->isPointerTy()) { 1006553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick PointerType *GEPPtrTy = cast<PointerType>(ExpandTy); 1007553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // If the step isn't constant, don't use an implicitly scaled GEP, because 1008553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // that would require a multiply inside the loop. 1009553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick if (!isa<ConstantInt>(StepV)) 1010553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()), 1011553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick GEPPtrTy->getAddressSpace()); 1012553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick const SCEV *const StepArray[1] = { SE.getSCEV(StepV) }; 1013553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN); 1014553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick if (IncV->getType() != PN->getType()) { 1015553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick IncV = Builder.CreateBitCast(IncV, PN->getType()); 1016553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick rememberInstruction(IncV); 1017553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick } 1018553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick } else { 1019553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick IncV = useSubtract ? 1020553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") : 1021553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next"); 1022553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick rememberInstruction(IncV); 1023553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick } 1024553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick return IncV; 1025553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick} 1026553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick 1027a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman/// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand 1028a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman/// the base addrec, which is the addrec without any non-loop-dominating 1029a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman/// values, and return the PHI. 1030a10756ee657a4d43a48cca5c166919093930ed6bDan GohmanPHINode * 1031a10756ee657a4d43a48cca5c166919093930ed6bDan GohmanSCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized, 1032a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const Loop *L, 1033db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *ExpandTy, 1034db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *IntTy) { 103593a896e2e36480d55de3bab53e68581e08526344Benjamin Kramer assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position"); 1036d152d03a476b8d0d4b26577db26e2ba76034b0f3Andrew Trick 1037a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Reuse a previously-inserted PHI, if present. 1038c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick BasicBlock *LatchBlock = L->getLoopLatch(); 1039c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (LatchBlock) { 1040c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick for (BasicBlock::iterator I = L->getHeader()->begin(); 1041c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1042c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (!SE.isSCEVable(PN->getType()) || 1043c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick (SE.getEffectiveSCEVType(PN->getType()) != 1044c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick SE.getEffectiveSCEVType(Normalized->getType())) || 1045c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick SE.getSCEV(PN) != Normalized) 1046c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick continue; 1047c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 1048c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick Instruction *IncV = 1049c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)); 1050c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick 1051c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (LSRMode) { 1052365c9f1ff55bef134c6b9707f7df44d680ddabeaAndrew Trick if (!isExpandedAddRecExprPHI(PN, IncV, L)) 1053c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick continue; 1054b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (L == IVIncInsertLoop && !hoistIVInc(IncV, IVIncInsertPos)) 1055b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick continue; 1056c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 1057c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick else { 1058c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick if (!isNormalAddRecExprPHI(PN, IncV, L)) 1059c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick continue; 1060b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (L == IVIncInsertLoop) 1061b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick do { 1062b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick if (SE.DT->dominates(IncV, IVIncInsertPos)) 1063b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick break; 1064b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // Make sure the increment is where we want it. But don't move it 1065b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // down past a potential existing post-inc user. 1066b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick IncV->moveBefore(IVIncInsertPos); 1067b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick IVIncInsertPos = IncV; 1068b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick IncV = cast<Instruction>(IncV->getOperand(0)); 1069b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } while (IncV != PN); 1070c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 1071c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Ok, the add recurrence looks usable. 1072c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Remember this PHI, even in post-inc mode. 1073c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick InsertedValues.insert(PN); 1074c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick // Remember the increment. 1075c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick rememberInstruction(IncV); 1076c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick return PN; 1077c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 1078c5701910604cdf65811fabd31d41e38f1d1d4eb1Andrew Trick } 1079a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1080a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Save the original insertion point so we can restore it when we're done. 1081a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 1082a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 1083a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1084ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // Another AddRec may need to be recursively expanded below. For example, if 1085ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // this AddRec is quadratic, the StepV may itself be an AddRec in this 1086ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // loop. Remove this loop from the PostIncLoops set before expanding such 1087ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // AddRecs. Otherwise, we cannot find a valid position for the step 1088ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // (i.e. StepV can never dominate its loop header). Ideally, we could do 1089ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // SavedIncLoops.swap(PostIncLoops), but we generally have a single element, 1090ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // so it's not worth implementing SmallPtrSet::swap. 1091ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick PostIncLoopSet SavedPostIncLoops = PostIncLoops; 1092ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick PostIncLoops.clear(); 1093ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick 1094a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Expand code for the start value. 1095a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy, 1096a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman L->getHeader()->begin()); 1097a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1098d152d03a476b8d0d4b26577db26e2ba76034b0f3Andrew Trick // StartV must be hoisted into L's preheader to dominate the new phi. 109993a896e2e36480d55de3bab53e68581e08526344Benjamin Kramer assert(!isa<Instruction>(StartV) || 110093a896e2e36480d55de3bab53e68581e08526344Benjamin Kramer SE.DT->properlyDominates(cast<Instruction>(StartV)->getParent(), 110193a896e2e36480d55de3bab53e68581e08526344Benjamin Kramer L->getHeader())); 1102d152d03a476b8d0d4b26577db26e2ba76034b0f3Andrew Trick 1103553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // Expand code for the step value. Do this before creating the PHI so that PHI 1104553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // reuse code doesn't see an incomplete PHI. 1105a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEV *Step = Normalized->getStepRecurrence(SE); 1106553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // If the stride is negative, insert a sub instead of an add for the increment 1107553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // (unless it's a constant, because subtracts of constants are canonicalized 1108553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // to adds). 1109f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1110553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick if (useSubtract) 1111a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Step = SE.getNegativeSCEV(Step); 1112553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // Expand the step somewhere that dominates the loop header. 1113a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin()); 1114a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1115a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Create the PHI. 1116d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad BasicBlock *Header = L->getHeader(); 1117d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad Builder.SetInsertPoint(Header, Header->begin()); 1118d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 11195e7645be4c9dd2193add44d30b5fef8036d7a3ceAndrew Trick PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE), 1120dc8e546048db2f7ff5656742b2b26975098a11a0Andrew Trick Twine(IVName) + ".iv"); 1121a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(PN); 1122a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1123a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Create the step instructions and populate the PHI. 1124d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 1125a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman BasicBlock *Pred = *HPI; 1126a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1127a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Add a start value. 1128a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman if (!L->contains(Pred)) { 1129a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman PN->addIncoming(StartV, Pred); 1130a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman continue; 1131a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1132a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1133553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // Create a step value and add it to the PHI. 1134553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the 1135553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // instructions at IVIncInsertPos. 1136a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Instruction *InsertPos = L == IVIncInsertLoop ? 1137a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman IVIncInsertPos : Pred->getTerminator(); 1138c5ecbdc1896f1cc089372feef3191ace2f840898Devang Patel Builder.SetInsertPoint(InsertPos); 1139553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1140553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick 1141a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman PN->addIncoming(IncV, Pred); 1142a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1143a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1144a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Restore the original insert point. 1145a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman if (SaveInsertBB) 1146455985501381777db03534c925a35e261e356395Dan Gohman restoreInsertPoint(SaveInsertBB, SaveInsertPt); 1147a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1148ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // After expanding subexpressions, restore the PostIncLoops set so the caller 1149ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick // can ensure that IVIncrement dominates the current uses. 1150ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick PostIncLoops = SavedPostIncLoops; 1151ba3c0bc364a31bfbcc570504aae2172914b1fe9fAndrew Trick 1152a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Remember this PHI, even in post-inc mode. 1153a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman InsertedValues.insert(PN); 1154a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1155a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman return PN; 1156a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman} 1157a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1158a10756ee657a4d43a48cca5c166919093930ed6bDan GohmanValue *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) { 1159db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *STy = S->getType(); 1160db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *IntTy = SE.getEffectiveSCEVType(STy); 1161a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const Loop *L = S->getLoop(); 1162a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1163a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Determine a normalized form of this expression, which is the expression 1164a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // before any post-inc adjustment is made. 1165a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEVAddRecExpr *Normalized = S; 1166448db1cdef5872713ef77beffacf502ae3450cd7Dan Gohman if (PostIncLoops.count(L)) { 1167448db1cdef5872713ef77beffacf502ae3450cd7Dan Gohman PostIncLoopSet Loops; 1168448db1cdef5872713ef77beffacf502ae3450cd7Dan Gohman Loops.insert(L); 1169448db1cdef5872713ef77beffacf502ae3450cd7Dan Gohman Normalized = 1170448db1cdef5872713ef77beffacf502ae3450cd7Dan Gohman cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0, 1171448db1cdef5872713ef77beffacf502ae3450cd7Dan Gohman Loops, SE, *SE.DT)); 1172a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1173a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1174a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Strip off any non-loop-dominating component from the addrec start. 1175a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEV *Start = Normalized->getStart(); 1176a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEV *PostLoopOffset = 0; 1177dc0e8fb9f9512622f55f73e1a434caa5c0915694Dan Gohman if (!SE.properlyDominates(Start, L->getHeader())) { 1178a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman PostLoopOffset = Start; 1179deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman Start = SE.getConstant(Normalized->getType(), 0); 11803228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick Normalized = cast<SCEVAddRecExpr>( 11813228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE), 11823228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick Normalized->getLoop(), 11833228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick // FIXME: Normalized->getNoWrapFlags(FlagNW) 11843228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick SCEV::FlagAnyWrap)); 1185a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1186a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1187a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Strip off any non-loop-dominating component from the addrec step. 1188a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEV *Step = Normalized->getStepRecurrence(SE); 1189a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEV *PostLoopScale = 0; 1190dc0e8fb9f9512622f55f73e1a434caa5c0915694Dan Gohman if (!SE.dominates(Step, L->getHeader())) { 1191a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman PostLoopScale = Step; 1192deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman Step = SE.getConstant(Normalized->getType(), 1); 1193a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Normalized = 1194a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step, 11953228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick Normalized->getLoop(), 11963228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick // FIXME: Normalized 11973228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick // ->getNoWrapFlags(FlagNW) 11983228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick SCEV::FlagAnyWrap)); 1199a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1200a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1201a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Expand the core addrec. If we need post-loop scaling, force it to 1202a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // expand to an integer type to avoid the need for additional casting. 1203db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *ExpandTy = PostLoopScale ? IntTy : STy; 1204a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy); 1205a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 12063f46a3abeedba8d517b4182de34c821d752db058Dan Gohman // Accommodate post-inc mode, if necessary. 1207a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Value *Result; 1208448db1cdef5872713ef77beffacf502ae3450cd7Dan Gohman if (!PostIncLoops.count(L)) 1209a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Result = PN; 1210a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman else { 1211a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // In PostInc mode, use the post-incremented value. 1212a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman BasicBlock *LatchBlock = L->getLoopLatch(); 1213a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman assert(LatchBlock && "PostInc mode requires a unique loop latch!"); 1214a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Result = PN->getIncomingValueForBlock(LatchBlock); 121548ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick 121648ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // For an expansion to use the postinc form, the client must call 121748ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // expandCodeFor with an InsertPoint that is either outside the PostIncLoop 121848ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // or dominated by IVIncInsertPos. 1219553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick if (isa<Instruction>(Result) 1220553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick && !SE.DT->dominates(cast<Instruction>(Result), 1221553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Builder.GetInsertPoint())) { 1222553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // The induction variable's postinc expansion does not dominate this use. 1223553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // IVUsers tries to prevent this case, so it is rare. However, it can 1224553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // happen when an IVUser outside the loop is not dominated by the latch 1225553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // block. Adjusting IVIncInsertPos before expansion begins cannot handle 1226553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // all cases. Consider a phi outide whose operand is replaced during 1227553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // expansion with the value of the postinc user. Without fundamentally 1228553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // changing the way postinc users are tracked, the only remedy is 1229553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // inserting an extra IV increment. StepV might fold into PostLoopOffset, 1230553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // but hopefully expandCodeFor handles that. 1231553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick bool useSubtract = 1232f8fd841a4bb7a59f81cf4642169e8251e039acfeAndrew Trick !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1233553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick if (useSubtract) 1234553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Step = SE.getNegativeSCEV(Step); 1235553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // Expand the step somewhere that dominates the loop header. 1236553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 1237553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 1238553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin()); 1239553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // Restore the insertion point to the place where the caller has 1240553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick // determined dominates all uses. 1241553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick restoreInsertPoint(SaveInsertBB, SaveInsertPt); 1242553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1243553fe05f236f46fe27b7fcfa822b06367d50183eAndrew Trick } 1244a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1245a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1246a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Re-apply any non-loop-dominating scale. 1247a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman if (PostLoopScale) { 12480a799ab15801d4ebf68eeb151d6375a799c87d9aDan Gohman Result = InsertNoopCastOfTo(Result, IntTy); 1249a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Result = Builder.CreateMul(Result, 1250a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman expandCodeFor(PostLoopScale, IntTy)); 1251a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(Result); 1252a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1253a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1254a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman // Re-apply any non-loop-dominating offset. 1255a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman if (PostLoopOffset) { 1256db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) { 1257a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman const SCEV *const OffsetArray[1] = { PostLoopOffset }; 1258a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result); 1259a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } else { 12600a799ab15801d4ebf68eeb151d6375a799c87d9aDan Gohman Result = InsertNoopCastOfTo(Result, IntTy); 1261a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman Result = Builder.CreateAdd(Result, 1262a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman expandCodeFor(PostLoopOffset, IntTy)); 1263a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(Result); 1264a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1265a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman } 1266a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1267a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman return Result; 1268a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman} 1269a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1270890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) { 1271a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman if (!CanonicalMode) return expandAddRecExprLiterally(S); 1272a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman 1273db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 127436f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman const Loop *L = S->getLoop(); 127536f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 12764d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // First check for an existing canonical IV in a suitable type. 12774d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman PHINode *CanonicalIV = 0; 12784d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman if (PHINode *PN = L->getCanonicalInductionVariable()) 1279133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty)) 12804d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman CanonicalIV = PN; 12814d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman 12824d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // Rewrite an AddRec in terms of the canonical induction variable, if 12834d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // its type is more narrow. 12844d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman if (CanonicalIV && 12854d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman SE.getTypeSizeInBits(CanonicalIV->getType()) > 12864d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman SE.getTypeSizeInBits(Ty)) { 1287f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman SmallVector<const SCEV *, 4> NewOps(S->getNumOperands()); 1288f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i) 1289f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType()); 12903228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(), 12913228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick // FIXME: S->getNoWrapFlags(FlagNW) 12923228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick SCEV::FlagAnyWrap)); 1293267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 1294267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 12954d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman BasicBlock::iterator NewInsertPt = 12967896c9f436a4eda5ec15e882a7505ba482a2fcd0Chris Lattner llvm::next(BasicBlock::iterator(cast<Instruction>(V))); 1297a4c86ab073d4b7a36477fc7c54c9d52499f04586Bill Wendling while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) || 1298a4c86ab073d4b7a36477fc7c54c9d52499f04586Bill Wendling isa<LandingPadInst>(NewInsertPt)) 129908f55d02da2cd24f3b4297259ef6429b8bf1ef97Jim Grosbach ++NewInsertPt; 13004d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0, 13014d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman NewInsertPt); 1302455985501381777db03534c925a35e261e356395Dan Gohman restoreInsertPoint(SaveInsertBB, SaveInsertPt); 13034d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman return V; 13044d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman } 13054d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman 130636f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // {X,+,F} --> X + {0,+,F} 1307cfeb6a450632f2a6cd05302633c8c2b8c90cfdfdDan Gohman if (!S->getStart()->isZero()) { 1308f9e64729afec646fe93b51417e66e552e8e630a4Dan Gohman SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end()); 1309deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman NewOps[0] = SE.getConstant(Ty, 0); 13103228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick // FIXME: can use S->getNoWrapFlags() 13113228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick const SCEV *Rest = SE.getAddRecExpr(NewOps, L, SCEV::FlagAnyWrap); 1312453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 1313453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the 1314453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman // comments on expandAddToGEP for details. 1315c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *Base = S->getStart(); 1316c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman const SCEV *RestArray[1] = { Rest }; 1317c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Dig into the expression to find the pointer base for a GEP. 1318c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman ExposePointerBase(Base, RestArray[0], SE); 1319c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // If we found a pointer, expand the AddRec with a GEP. 1320db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) { 1321c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // Make sure the Base isn't something exotic, such as a multiplied 1322c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // or divided pointer value. In those cases, the result type isn't 1323c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman // actually a pointer type. 1324c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) { 1325c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman Value *StartV = expand(Base); 1326c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!"); 1327c40f17b08774c2dcc5787fd83241e3c64ba82974Dan Gohman return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV); 1328453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 1329453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman } 1330453aa4fbf1083cc7f646a0ac21e2bcc384a91ae9Dan Gohman 133140a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // Just do a normal add. Pre-expand the operands to suppress folding. 133240a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())), 133340a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman SE.getUnknown(expand(Rest)))); 133436f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman } 133536f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 13366ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman // If we don't yet have a canonical IV, create one. 13376ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman if (!CanonicalIV) { 133836f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // Create and insert the PHI node for the induction variable in the 133936f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // specified loop. 134036f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman BasicBlock *Header = L->getHeader(); 1341d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 13423ecfc861b4365f341c5c969b40e1afccde676e6fJay Foad CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar", 13433ecfc861b4365f341c5c969b40e1afccde676e6fJay Foad Header->begin()); 13446ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman rememberInstruction(CanonicalIV); 134536f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 1346eed707b1e6097aac2bb6b3d47271f6300ace7f2eOwen Anderson Constant *One = ConstantInt::get(Ty, 1); 1347d8b4fb4aab4d6fedb2b14bed1b846451b17bde7cJay Foad for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 13487656018c2268285907cfdc106071462a01a73878Gabor Greif BasicBlock *HP = *HPI; 13497656018c2268285907cfdc106071462a01a73878Gabor Greif if (L->contains(HP)) { 13503abf905b50f33340ae81913da81b0eda96fa4616Dan Gohman // Insert a unit add instruction right before the terminator 13513abf905b50f33340ae81913da81b0eda96fa4616Dan Gohman // corresponding to the back-edge. 13526ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One, 13536ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman "indvar.next", 13546ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman HP->getTerminator()); 1355df3ad6697bb849e514c0881ca8700ea36678cbdaDevang Patel Add->setDebugLoc(HP->getTerminator()->getDebugLoc()); 1356a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(Add); 13576ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman CanonicalIV->addIncoming(Add, HP); 135883d577490b1473213b6973236110c28e59127956Dan Gohman } else { 13596ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP); 136083d577490b1473213b6973236110c28e59127956Dan Gohman } 13617656018c2268285907cfdc106071462a01a73878Gabor Greif } 136236f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman } 136336f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 13646ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman // {0,+,1} --> Insert a canonical induction variable into the loop! 13656ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman if (S->isAffine() && S->getOperand(1)->isOne()) { 13666ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) && 13676ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman "IVs with types different from the canonical IV should " 13686ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman "already have been handled!"); 13696ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman return CanonicalIV; 13706ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman } 13716ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman 13724d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // {0,+,F} --> {0,+,1} * F 137336f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 1374df14a04b5ccbbe6a46c2ccb93e27b12a36ff163eChris Lattner // If this is a simple linear addrec, emit it now as a special case. 137540a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman if (S->isAffine()) // {0,+,F} --> i*F 137640a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman return 137740a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman expand(SE.getTruncateOrNoop( 13786ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman SE.getMulExpr(SE.getUnknown(CanonicalIV), 137940a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman SE.getNoopOrAnyExtend(S->getOperand(1), 13806ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman CanonicalIV->getType())), 138140a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman Ty)); 138236f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 138336f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // If this is a chain of recurrences, turn it into a closed form, using the 138436f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // folders, then expandCodeFor the closed form. This allows the folders to 138536f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // simplify the expression without having to build a bunch of special code 138636f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman // into this folder. 13876ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV. 138836f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 13894d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // Promote S up to the canonical IV type, if the cast is foldable. 13900bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *NewS = S; 13916ebfd72f37325a1ebceb53e5ecad524d359d8d0bDan Gohman const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType()); 13924d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman if (isa<SCEVAddRecExpr>(Ext)) 13934d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman NewS = Ext; 13944d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman 13950bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE); 1396e81561909d128c6e2d8033cb5465a49b2596b26aBill Wendling //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n"; 139736f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman 13984d8414f42033a5e744e8e60d2ca188b424c76168Dan Gohman // Truncate the result down to the original type, if needed. 13990bba49cebc50c7bd4662a4807bcb3ee7f42cb470Dan Gohman const SCEV *T = SE.getTruncateOrNoop(V, Ty); 1400469f3cdc13851914f3a766cbd8f701cf8431cacaDan Gohman return expand(T); 140136f891bdf6cf38fcc655a0930ca18664e18518d4Nate Begeman} 140296fea337d27357e9b62abbf3d2d5ce29f1c8e870Anton Korobeynikov 1403890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) { 1404db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 140592fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *V = expandCodeFor(S->getOperand(), 140692fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman SE.getEffectiveSCEVType(S->getOperand()->getType())); 1407a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Value *I = Builder.CreateTrunc(V, Ty); 1408a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(I); 1409cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman return I; 141011f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman} 141111f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman 1412890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) { 1413db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 141492fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *V = expandCodeFor(S->getOperand(), 141592fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman SE.getEffectiveSCEVType(S->getOperand()->getType())); 1416a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Value *I = Builder.CreateZExt(V, Ty); 1417a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(I); 1418cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman return I; 141911f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman} 142011f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman 1421890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) { 1422db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = SE.getEffectiveSCEVType(S->getType()); 142392fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *V = expandCodeFor(S->getOperand(), 142492fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman SE.getEffectiveSCEVType(S->getOperand()->getType())); 1425a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Value *I = Builder.CreateSExt(V, Ty); 1426a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(I); 1427cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman return I; 142811f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman} 142911f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman 1430890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) { 14310196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1432db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = LHS->getType(); 14330196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman for (int i = S->getNumOperands()-2; i >= 0; --i) { 14340196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // In the case of mixed integer and pointer types, do the 14350196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // rest of the comparisons as integer. 14360196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman if (S->getOperand(i)->getType() != Ty) { 14370196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman Ty = SE.getEffectiveSCEVType(Ty); 14380196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman LHS = InsertNoopCastOfTo(LHS, Ty); 14390196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman } 144092fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1441a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Value *ICmp = Builder.CreateICmpSGT(LHS, RHS); 1442a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(ICmp); 1443267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax"); 1444a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(Sel); 1445cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman LHS = Sel; 1446c54c561c9f7270c055dd7ba75a3a003b771a42d9Nick Lewycky } 14470196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // In the case of mixed integer and pointer types, cast the 14480196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // final result back to the pointer type. 14490196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman if (LHS->getType() != S->getType()) 14500196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman LHS = InsertNoopCastOfTo(LHS, S->getType()); 1451c54c561c9f7270c055dd7ba75a3a003b771a42d9Nick Lewycky return LHS; 1452c54c561c9f7270c055dd7ba75a3a003b771a42d9Nick Lewycky} 1453c54c561c9f7270c055dd7ba75a3a003b771a42d9Nick Lewycky 1454890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) { 14550196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1456db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty = LHS->getType(); 14570196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman for (int i = S->getNumOperands()-2; i >= 0; --i) { 14580196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // In the case of mixed integer and pointer types, do the 14590196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // rest of the comparisons as integer. 14600196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman if (S->getOperand(i)->getType() != Ty) { 14610196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman Ty = SE.getEffectiveSCEVType(Ty); 14620196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman LHS = InsertNoopCastOfTo(LHS, Ty); 14630196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman } 146492fcdcac543653a62949fe9e5a7bd008500c1380Dan Gohman Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1465a9390a4d5f5d568059a80970d22194b165d097a7Benjamin Kramer Value *ICmp = Builder.CreateICmpUGT(LHS, RHS); 1466a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(ICmp); 1467267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax"); 1468a10756ee657a4d43a48cca5c166919093930ed6bDan Gohman rememberInstruction(Sel); 1469cf5ab820227dedd77fb91d0904b6dc3694a7c196Dan Gohman LHS = Sel; 14703e6307698084e7adfc10b739442ae29742beefd0Nick Lewycky } 14710196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // In the case of mixed integer and pointer types, cast the 14720196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman // final result back to the pointer type. 14730196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman if (LHS->getType() != S->getType()) 14740196dc5733bf3c6a71f7de1f631f2c43e5809fd9Dan Gohman LHS = InsertNoopCastOfTo(LHS, S->getType()); 14753e6307698084e7adfc10b739442ae29742beefd0Nick Lewycky return LHS; 14763e6307698084e7adfc10b739442ae29742beefd0Nick Lewycky} 14773e6307698084e7adfc10b739442ae29742beefd0Nick Lewycky 1478db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris LattnerValue *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty, 1479b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick Instruction *IP) { 14806c7ed6b54949949806797bafdf545fbfecb2cef5Dan Gohman Builder.SetInsertPoint(IP->getParent(), IP); 14816c7ed6b54949949806797bafdf545fbfecb2cef5Dan Gohman return expandCodeFor(SH, Ty); 14826c7ed6b54949949806797bafdf545fbfecb2cef5Dan Gohman} 14836c7ed6b54949949806797bafdf545fbfecb2cef5Dan Gohman 1484db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris LattnerValue *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) { 148511f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman // Expand the code for this SCEV. 14862d1be87ee40a4a0241d94448173879d9df2bc5b3Dan Gohman Value *V = expand(SH); 14875be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman if (Ty) { 14885be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) && 14895be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman "non-trivial casts should be done with the SCEVs directly!"); 14905be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman V = InsertNoopCastOfTo(V, Ty); 14915be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman } 14925be18e84766fb495b0bde3c8244c1df459a18683Dan Gohman return V; 149311f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman} 149411f6d3b478c4fa09d126833c57fbac1d795ead31Dan Gohman 1495890f92b744fb074465bc2b7006ee753a181f62a4Dan GohmanValue *SCEVExpander::expand(const SCEV *S) { 149640a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // Compute an insertion point for this SCEV object. Hoist the instructions 149740a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // as far out in the loop nest as possible. 1498267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Instruction *InsertPt = Builder.GetInsertPoint(); 1499267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ; 150040a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman L = L->getParentLoop()) 150117ead4ff4baceb2c5503f233d0288d363ae44165Dan Gohman if (SE.isLoopInvariant(S, L)) { 150240a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman if (!L) break; 1503e059ee832ca36d65a5fb87b0ac5bcdb0490b15cbDan Gohman if (BasicBlock *Preheader = L->getLoopPreheader()) 150440a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman InsertPt = Preheader->getTerminator(); 15050f8cd56bfdd32af4edb253654db02fb3143b25a8Andrew Trick else { 15060f8cd56bfdd32af4edb253654db02fb3143b25a8Andrew Trick // LSR sets the insertion point for AddRec start/step values to the 15070f8cd56bfdd32af4edb253654db02fb3143b25a8Andrew Trick // block start to simplify value reuse, even though it's an invalid 15080f8cd56bfdd32af4edb253654db02fb3143b25a8Andrew Trick // position. SCEVExpander must correct for this in all cases. 15090f8cd56bfdd32af4edb253654db02fb3143b25a8Andrew Trick InsertPt = L->getHeader()->getFirstInsertionPt(); 15100f8cd56bfdd32af4edb253654db02fb3143b25a8Andrew Trick } 151140a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman } else { 151240a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // If the SCEV is computable at this level, insert it into the header 151340a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // after the PHIs (and after any other instructions that we've inserted 151440a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman // there) so that it is guaranteed to dominate any user inside the loop. 15155b6f42f57e730c2d968c313a27fa505a3c3e5efaBill Wendling if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L)) 15165b6f42f57e730c2d968c313a27fa505a3c3e5efaBill Wendling InsertPt = L->getHeader()->getFirstInsertionPt(); 1517b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick while (InsertPt != Builder.GetInsertPoint() 1518b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick && (isInsertedInstruction(InsertPt) 1519b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick || isa<DbgInfoIntrinsic>(InsertPt))) { 15207896c9f436a4eda5ec15e882a7505ba482a2fcd0Chris Lattner InsertPt = llvm::next(BasicBlock::iterator(InsertPt)); 1521b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick } 152240a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman break; 152340a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman } 152440a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman 1525667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman // Check to see if we already expanded this here. 1526667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman std::map<std::pair<const SCEV *, Instruction *>, 1527667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman AssertingVH<Value> >::iterator I = 1528667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman InsertedExpressions.find(std::make_pair(S, InsertPt)); 1529267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman if (I != InsertedExpressions.end()) 1530667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman return I->second; 1531267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman 1532267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 1533267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 1534267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman Builder.SetInsertPoint(InsertPt->getParent(), InsertPt); 1535667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman 1536667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman // Expand the expression into instructions. 153796fea337d27357e9b62abbf3d2d5ce29f1c8e870Anton Korobeynikov Value *V = visit(S); 153840a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman 1539667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman // Remember the expanded value for this SCEV at this location. 154048ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // 154148ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // This is independent of PostIncLoops. The mapped value simply materializes 154248ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // the expression at this insertion point. If the mapped value happened to be 154348ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // a postinc expansion, it could be reused by a non postinc user, but only if 154448ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick // its insertion point was already at the head of the loop. 154548ba0e45ed68689ce7b384578e6272410e4e23feAndrew Trick InsertedExpressions[std::make_pair(S, InsertPt)] = V; 1546667d787c0a21cf3f5dfcde03ca471162ba35b614Dan Gohman 1547455985501381777db03534c925a35e261e356395Dan Gohman restoreInsertPoint(SaveInsertBB, SaveInsertPt); 154896fea337d27357e9b62abbf3d2d5ce29f1c8e870Anton Korobeynikov return V; 154996fea337d27357e9b62abbf3d2d5ce29f1c8e870Anton Korobeynikov} 15501d09de3eca23267855e28297fcb40de3632ea47bDan Gohman 15511d826a76f591afea445489b9a5485c345e66bf87Dan Gohmanvoid SCEVExpander::rememberInstruction(Value *I) { 155225fcaff409f5c4c6da08f148ffb9404a71e8e4a7Dan Gohman if (!PostIncLoops.empty()) 155325fcaff409f5c4c6da08f148ffb9404a71e8e4a7Dan Gohman InsertedPostIncValues.insert(I); 155425fcaff409f5c4c6da08f148ffb9404a71e8e4a7Dan Gohman else 15551d826a76f591afea445489b9a5485c345e66bf87Dan Gohman InsertedValues.insert(I); 15561d826a76f591afea445489b9a5485c345e66bf87Dan Gohman} 15571d826a76f591afea445489b9a5485c345e66bf87Dan Gohman 1558455985501381777db03534c925a35e261e356395Dan Gohmanvoid SCEVExpander::restoreInsertPoint(BasicBlock *BB, BasicBlock::iterator I) { 1559455985501381777db03534c925a35e261e356395Dan Gohman Builder.SetInsertPoint(BB, I); 1560455985501381777db03534c925a35e261e356395Dan Gohman} 1561455985501381777db03534c925a35e261e356395Dan Gohman 15621d09de3eca23267855e28297fcb40de3632ea47bDan Gohman/// getOrInsertCanonicalInductionVariable - This method returns the 15631d09de3eca23267855e28297fcb40de3632ea47bDan Gohman/// canonical induction variable of the specified type for the specified 15641d09de3eca23267855e28297fcb40de3632ea47bDan Gohman/// loop (inserting one if there is none). A canonical induction variable 15651d09de3eca23267855e28297fcb40de3632ea47bDan Gohman/// starts at zero and steps by one on each iteration. 15667c58dbd88c36c5d6c411ea6c046ddcff4c5841e9Dan GohmanPHINode * 15671d09de3eca23267855e28297fcb40de3632ea47bDan GohmanSCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L, 1568db125cfaf57cc83e7dd7453de2d509bc8efd0e5eChris Lattner Type *Ty) { 1569b0bc6c361da9009e8414efde317d9bbff755f6c0Duncan Sands assert(Ty->isIntegerTy() && "Can only insert integer induction variables!"); 1570133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman 1571133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman // Build a SCEV for {0,+,1}<L>. 15723228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick // Conservatively use FlagAnyWrap for now. 1573deff621abdd48bd70434bd4d7ef30f08ddba1cd8Dan Gohman const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0), 15743228cc259b5ca00e46af36da369a451f5736cbf4Andrew Trick SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap); 1575133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman 1576133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman // Emit code for it. 1577267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 1578267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 15797c58dbd88c36c5d6c411ea6c046ddcff4c5841e9Dan Gohman PHINode *V = cast<PHINode>(expandCodeFor(H, 0, L->getHeader()->begin())); 1580267a385342f2e7388f178b327dd87c5f29afd51bDan Gohman if (SaveInsertBB) 1581455985501381777db03534c925a35e261e356395Dan Gohman restoreInsertPoint(SaveInsertBB, SaveInsertPt); 1582133e295b363c89b6a3da7aba3ac5ae6332429e46Dan Gohman 158340a5a1b39ee1cd40ff9d04740386b667fb27b340Dan Gohman return V; 15841d09de3eca23267855e28297fcb40de3632ea47bDan Gohman} 1585204494149b6f846e8f173f525b129f5508076049Andrew Trick 1586139f333f88fb9e0c39dfa3b56fd2a26b33837b70Andrew Trick/// Sort values by integer width for replaceCongruentIVs. 1587139f333f88fb9e0c39dfa3b56fd2a26b33837b70Andrew Trickstatic bool width_descending(Value *lhs, Value *rhs) { 1588ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // Put pointers at the back and make sure pointer < pointer = false. 1589ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick if (!lhs->getType()->isIntegerTy() || !rhs->getType()->isIntegerTy()) 1590ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick return rhs->getType()->isIntegerTy() && !lhs->getType()->isIntegerTy(); 1591ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick return rhs->getType()->getPrimitiveSizeInBits() 1592ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick < lhs->getType()->getPrimitiveSizeInBits(); 1593ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick} 1594ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick 1595204494149b6f846e8f173f525b129f5508076049Andrew Trick/// replaceCongruentIVs - Check for congruent phis in this loop header and 1596204494149b6f846e8f173f525b129f5508076049Andrew Trick/// replace them with their most canonical representative. Return the number of 1597204494149b6f846e8f173f525b129f5508076049Andrew Trick/// phis eliminated. 1598204494149b6f846e8f173f525b129f5508076049Andrew Trick/// 1599204494149b6f846e8f173f525b129f5508076049Andrew Trick/// This does not depend on any SCEVExpander state but should be used in 1600204494149b6f846e8f173f525b129f5508076049Andrew Trick/// the same context that SCEVExpander is used. 1601204494149b6f846e8f173f525b129f5508076049Andrew Trickunsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT, 1602a04a4a79ea365d1ba96ed4b5861e879b267162e2Nadav Rotem SmallVectorImpl<WeakVH> &DeadInsts, 1603a04a4a79ea365d1ba96ed4b5861e879b267162e2Nadav Rotem const TargetLowering *TLI) { 1604ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // Find integer phis in order of increasing width. 1605ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick SmallVector<PHINode*, 8> Phis; 1606ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick for (BasicBlock::iterator I = L->getHeader()->begin(); 1607ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick PHINode *Phi = dyn_cast<PHINode>(I); ++I) { 1608ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick Phis.push_back(Phi); 1609ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick } 1610a04a4a79ea365d1ba96ed4b5861e879b267162e2Nadav Rotem if (TLI) 1611ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick std::sort(Phis.begin(), Phis.end(), width_descending); 1612ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick 1613204494149b6f846e8f173f525b129f5508076049Andrew Trick unsigned NumElim = 0; 1614204494149b6f846e8f173f525b129f5508076049Andrew Trick DenseMap<const SCEV *, PHINode *> ExprToIVMap; 1615ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // Process phis from wide to narrow. Mapping wide phis to the their truncation 1616ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // so narrow phis can reuse them. 1617ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick for (SmallVectorImpl<PHINode*>::const_iterator PIter = Phis.begin(), 1618ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick PEnd = Phis.end(); PIter != PEnd; ++PIter) { 1619ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick PHINode *Phi = *PIter; 1620ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick 1621239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer // Fold constant phis. They may be congruent to other constant phis and 1622239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer // would confuse the logic below that expects proper IVs. 1623239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer if (Value *V = Phi->hasConstantValue()) { 1624239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer Phi->replaceAllUsesWith(V); 1625239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer DeadInsts.push_back(Phi); 1626239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer ++NumElim; 1627239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer DEBUG_WITH_TYPE(DebugType, dbgs() 1628239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer << "INDVARS: Eliminated constant iv: " << *Phi << '\n'); 1629239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer continue; 1630239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer } 1631239fd44f7a68aae4b2e8b6f1738ef9e8fd4ddc01Benjamin Kramer 1632204494149b6f846e8f173f525b129f5508076049Andrew Trick if (!SE.isSCEVable(Phi->getType())) 1633204494149b6f846e8f173f525b129f5508076049Andrew Trick continue; 1634204494149b6f846e8f173f525b129f5508076049Andrew Trick 1635204494149b6f846e8f173f525b129f5508076049Andrew Trick PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)]; 1636204494149b6f846e8f173f525b129f5508076049Andrew Trick if (!OrigPhiRef) { 1637204494149b6f846e8f173f525b129f5508076049Andrew Trick OrigPhiRef = Phi; 1638a04a4a79ea365d1ba96ed4b5861e879b267162e2Nadav Rotem if (Phi->getType()->isIntegerTy() && TLI 1639a04a4a79ea365d1ba96ed4b5861e879b267162e2Nadav Rotem && TLI->isTruncateFree(Phi->getType(), Phis.back()->getType())) { 1640ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // This phi can be freely truncated to the narrowest phi type. Map the 1641ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // truncated expression to it so it will be reused for narrow types. 1642ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick const SCEV *TruncExpr = 1643ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType()); 1644ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick ExprToIVMap[TruncExpr] = Phi; 1645ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick } 1646204494149b6f846e8f173f525b129f5508076049Andrew Trick continue; 1647204494149b6f846e8f173f525b129f5508076049Andrew Trick } 1648204494149b6f846e8f173f525b129f5508076049Andrew Trick 1649ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // Replacing a pointer phi with an integer phi or vice-versa doesn't make 1650ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // sense. 1651ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy()) 1652204494149b6f846e8f173f525b129f5508076049Andrew Trick continue; 1653204494149b6f846e8f173f525b129f5508076049Andrew Trick 1654204494149b6f846e8f173f525b129f5508076049Andrew Trick if (BasicBlock *LatchBlock = L->getLoopLatch()) { 1655204494149b6f846e8f173f525b129f5508076049Andrew Trick Instruction *OrigInc = 1656204494149b6f846e8f173f525b129f5508076049Andrew Trick cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock)); 1657204494149b6f846e8f173f525b129f5508076049Andrew Trick Instruction *IsomorphicInc = 1658204494149b6f846e8f173f525b129f5508076049Andrew Trick cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock)); 1659204494149b6f846e8f173f525b129f5508076049Andrew Trick 1660ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick // If this phi has the same width but is more canonical, replace the 1661b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // original with it. As part of the "more canonical" determination, 1662b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick // respect a prior decision to use an IV chain. 1663ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick if (OrigPhiRef->getType() == Phi->getType() 1664b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick && !(ChainedPhis.count(Phi) 1665b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L)) 1666b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick && (ChainedPhis.count(Phi) 1667b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) { 1668204494149b6f846e8f173f525b129f5508076049Andrew Trick std::swap(OrigPhiRef, Phi); 1669204494149b6f846e8f173f525b129f5508076049Andrew Trick std::swap(OrigInc, IsomorphicInc); 1670204494149b6f846e8f173f525b129f5508076049Andrew Trick } 1671204494149b6f846e8f173f525b129f5508076049Andrew Trick // Replacing the congruent phi is sufficient because acyclic redundancy 1672204494149b6f846e8f173f525b129f5508076049Andrew Trick // elimination, CSE/GVN, should handle the rest. However, once SCEV proves 1673204494149b6f846e8f173f525b129f5508076049Andrew Trick // that a phi is congruent, it's often the head of an IV user cycle that 1674139f333f88fb9e0c39dfa3b56fd2a26b33837b70Andrew Trick // is isomorphic with the original phi. It's worth eagerly cleaning up the 1675139f333f88fb9e0c39dfa3b56fd2a26b33837b70Andrew Trick // common case of a single IV increment so that DeleteDeadPHIs can remove 1676139f333f88fb9e0c39dfa3b56fd2a26b33837b70Andrew Trick // cycles that had postinc uses. 1677ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc), 1678ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick IsomorphicInc->getType()); 1679ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick if (OrigInc != IsomorphicInc 168064925c55c65f9345a69fb67db07aa62cfb723577Andrew Trick && TruncExpr == SE.getSCEV(IsomorphicInc) 1681b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc)) 1682b5c26ef9da16052597d59a412eaae32098aa1be0Andrew Trick || hoistIVInc(OrigInc, IsomorphicInc))) { 1683204494149b6f846e8f173f525b129f5508076049Andrew Trick DEBUG_WITH_TYPE(DebugType, dbgs() 1684204494149b6f846e8f173f525b129f5508076049Andrew Trick << "INDVARS: Eliminated congruent iv.inc: " 1685204494149b6f846e8f173f525b129f5508076049Andrew Trick << *IsomorphicInc << '\n'); 1686ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick Value *NewInc = OrigInc; 1687ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick if (OrigInc->getType() != IsomorphicInc->getType()) { 1688dd1f22f25d8496b10cfddedb63d674dd39897bb0Andrew Trick Instruction *IP = isa<PHINode>(OrigInc) 1689dd1f22f25d8496b10cfddedb63d674dd39897bb0Andrew Trick ? (Instruction*)L->getHeader()->getFirstInsertionPt() 1690dd1f22f25d8496b10cfddedb63d674dd39897bb0Andrew Trick : OrigInc->getNextNode(); 1691dd1f22f25d8496b10cfddedb63d674dd39897bb0Andrew Trick IRBuilder<> Builder(IP); 1692ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc()); 1693ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick NewInc = Builder. 1694ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName); 1695ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick } 1696ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick IsomorphicInc->replaceAllUsesWith(NewInc); 1697204494149b6f846e8f173f525b129f5508076049Andrew Trick DeadInsts.push_back(IsomorphicInc); 1698204494149b6f846e8f173f525b129f5508076049Andrew Trick } 1699204494149b6f846e8f173f525b129f5508076049Andrew Trick } 1700204494149b6f846e8f173f525b129f5508076049Andrew Trick DEBUG_WITH_TYPE(DebugType, dbgs() 1701204494149b6f846e8f173f525b129f5508076049Andrew Trick << "INDVARS: Eliminated congruent iv: " << *Phi << '\n'); 1702204494149b6f846e8f173f525b129f5508076049Andrew Trick ++NumElim; 1703ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick Value *NewIV = OrigPhiRef; 1704ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick if (OrigPhiRef->getType() != Phi->getType()) { 1705ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick IRBuilder<> Builder(L->getHeader()->getFirstInsertionPt()); 1706ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick Builder.SetCurrentDebugLocation(Phi->getDebugLoc()); 1707ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName); 1708ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick } 1709ee98aa87434d9d49a8e4dab41d873888ac9c4805Andrew Trick Phi->replaceAllUsesWith(NewIV); 1710204494149b6f846e8f173f525b129f5508076049Andrew Trick DeadInsts.push_back(Phi); 1711204494149b6f846e8f173f525b129f5508076049Andrew Trick } 1712204494149b6f846e8f173f525b129f5508076049Andrew Trick return NumElim; 1713204494149b6f846e8f173f525b129f5508076049Andrew Trick} 1714e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick 1715e08c32249fca32cd7b122024a4ca252fcb235694Andrew Tricknamespace { 1716e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// Search for a SCEV subexpression that is not safe to expand. Any expression 1717e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely 1718e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// UDiv expressions. We don't know if the UDiv is derived from an IR divide 1719e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// instruction, but the important thing is that we prove the denominator is 1720e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// nonzero before expansion. 1721e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// 1722e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// IVUsers already checks that IV-derived expressions are safe. So this check is 1723e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// only needed when the expression includes some subexpression that is not IV 1724e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// derived. 1725e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// 1726e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// Currently, we only allow division by a nonzero constant here. If this is 1727e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// inadequate, we could easily allow division by SCEVUnknown by using 1728e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick// ValueTracking to check isKnownNonZero(). 1729e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trickstruct SCEVFindUnsafe { 1730e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick bool IsUnsafe; 1731e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick 1732e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick SCEVFindUnsafe(): IsUnsafe(false) {} 1733e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick 1734e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick bool follow(const SCEV *S) { 1735e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S); 1736e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick if (!D) 1737e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick return true; 1738e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS()); 1739e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick if (SC && !SC->getValue()->isZero()) 1740e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick return true; 1741e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick IsUnsafe = true; 1742e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick return false; 1743e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick } 1744e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick bool isDone() const { return IsUnsafe; } 1745e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick}; 1746e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick} 1747e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick 1748e08c32249fca32cd7b122024a4ca252fcb235694Andrew Tricknamespace llvm { 1749e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trickbool isSafeToExpand(const SCEV *S) { 1750e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick SCEVFindUnsafe Search; 1751e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick visitAll(S, Search); 1752e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick return !Search.IsUnsafe; 1753e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick} 1754e08c32249fca32cd7b122024a4ca252fcb235694Andrew Trick} 1755