// Licensed to the .NET Foundation under one or more agreements. // The .NET Foundation licenses this file to you under the MIT license. // See the LICENSE file in the project root for more information. /*XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX XX XX XX OptCSE XX XX XX XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX */ #include "jitpch.h" #ifdef _MSC_VER #pragma hdrstop #endif /*****************************************************************************/ #if FEATURE_ANYCSE /*****************************************************************************/ /* static */ const size_t Compiler::s_optCSEhashSize = EXPSET_SZ * 2; /***************************************************************************** * * We've found all the candidates, build the index for easy access. */ void Compiler::optCSEstop() { if (optCSECandidateCount == 0) { return; } CSEdsc* dsc; CSEdsc** ptr; unsigned cnt; optCSEtab = new (this, CMK_CSE) CSEdsc*[optCSECandidateCount](); for (cnt = s_optCSEhashSize, ptr = optCSEhash; cnt; cnt--, ptr++) { for (dsc = *ptr; dsc; dsc = dsc->csdNextInBucket) { if (dsc->csdIndex) { noway_assert((unsigned)dsc->csdIndex <= optCSECandidateCount); if (optCSEtab[dsc->csdIndex - 1] == nullptr) { optCSEtab[dsc->csdIndex - 1] = dsc; } } } } #ifdef DEBUG for (cnt = 0; cnt < optCSECandidateCount; cnt++) { noway_assert(optCSEtab[cnt] != nullptr); } #endif } /***************************************************************************** * * Return the descriptor for the CSE with the given index. */ inline Compiler::CSEdsc* Compiler::optCSEfindDsc(unsigned index) { noway_assert(index); noway_assert(index <= optCSECandidateCount); noway_assert(optCSEtab[index - 1]); return optCSEtab[index - 1]; } //------------------------------------------------------------------------ // Compiler::optUnmarkCSE // // Arguments: // tree - A sub tree that originally was part of a CSE use // that we are currently in the process of removing. // // Return Value: // Returns true if we can safely remove the 'tree' node. // Returns false if the node is a CSE def that the caller // needs to extract and preserve. // // Notes: // If 'tree' is a CSE use then we perform an unmark CSE operation // so that the CSE used counts and weight are updated properly. // The only caller for this method is optUnmarkCSEs which is a // tree walker vistor function. When we return false this method // returns WALK_SKIP_SUBTREES so that we don't visit the remaining // nodes of the CSE def. // bool Compiler::optUnmarkCSE(GenTree* tree) { if (!IS_CSE_INDEX(tree->gtCSEnum)) { // If this node isn't a CSE use or def we can safely remove this node. // return true; } // make sure it's been initialized noway_assert(optCSEweight <= BB_MAX_WEIGHT); // Is this a CSE use? if (IS_CSE_USE(tree->gtCSEnum)) { unsigned CSEnum = GET_CSE_INDEX(tree->gtCSEnum); CSEdsc* desc = optCSEfindDsc(CSEnum); #ifdef DEBUG if (verbose) { printf("Unmark CSE use #%02d at ", CSEnum); printTreeID(tree); printf(": %3d -> %3d\n", desc->csdUseCount, desc->csdUseCount - 1); } #endif // DEBUG // Perform an unmark CSE operation // 1. Reduce the nested CSE's 'use' count noway_assert(desc->csdUseCount > 0); if (desc->csdUseCount > 0) { desc->csdUseCount -= 1; if (desc->csdUseWtCnt < optCSEweight) { desc->csdUseWtCnt = 0; } else { desc->csdUseWtCnt -= optCSEweight; } } // 2. Unmark the CSE infomation in the node tree->gtCSEnum = NO_CSE; return true; } else { // It is not safe to remove this node, so we will return false // and the caller must add this node to the side effect list // return false; } } Compiler::fgWalkResult Compiler::optCSE_MaskHelper(GenTree** pTree, fgWalkData* walkData) { GenTree* tree = *pTree; Compiler* comp = walkData->compiler; optCSE_MaskData* pUserData = (optCSE_MaskData*)(walkData->pCallbackData); if (IS_CSE_INDEX(tree->gtCSEnum)) { unsigned cseIndex = GET_CSE_INDEX(tree->gtCSEnum); unsigned cseBit = genCSEnum2bit(cseIndex); if (IS_CSE_DEF(tree->gtCSEnum)) { BitVecOps::AddElemD(comp->cseTraits, pUserData->CSE_defMask, cseBit); } else { BitVecOps::AddElemD(comp->cseTraits, pUserData->CSE_useMask, cseBit); } } return WALK_CONTINUE; } // This functions walks all the node for an given tree // and return the mask of CSE defs and uses for the tree // void Compiler::optCSE_GetMaskData(GenTree* tree, optCSE_MaskData* pMaskData) { pMaskData->CSE_defMask = BitVecOps::MakeEmpty(cseTraits); pMaskData->CSE_useMask = BitVecOps::MakeEmpty(cseTraits); fgWalkTreePre(&tree, optCSE_MaskHelper, (void*)pMaskData); } //------------------------------------------------------------------------ // optCSE_canSwap: Determine if the execution order of two nodes can be swapped. // // Arguments: // op1 - The first node // op2 - The second node // // Return Value: // Return true iff it safe to swap the execution order of 'op1' and 'op2', // considering only the locations of the CSE defs and uses. // // Assumptions: // 'op1' currently occurse before 'op2' in the execution order. // bool Compiler::optCSE_canSwap(GenTree* op1, GenTree* op2) { // op1 and op2 must be non-null. assert(op1 != nullptr); assert(op2 != nullptr); bool canSwap = true; // the default result unless proven otherwise. optCSE_MaskData op1MaskData; optCSE_MaskData op2MaskData; optCSE_GetMaskData(op1, &op1MaskData); optCSE_GetMaskData(op2, &op2MaskData); // We cannot swap if op1 contains a CSE def that is used by op2 if (!BitVecOps::IsEmptyIntersection(cseTraits, op1MaskData.CSE_defMask, op2MaskData.CSE_useMask)) { canSwap = false; } else { // We also cannot swap if op2 contains a CSE def that is used by op1. if (!BitVecOps::IsEmptyIntersection(cseTraits, op2MaskData.CSE_defMask, op1MaskData.CSE_useMask)) { canSwap = false; } } return canSwap; } //------------------------------------------------------------------------ // optCSE_canSwap: Determine if the execution order of a node's operands can be swapped. // // Arguments: // tree - The node of interest // // Return Value: // Return true iff it safe to swap the execution order of the operands of 'tree', // considering only the locations of the CSE defs and uses. // bool Compiler::optCSE_canSwap(GenTree* tree) { // We must have a binary treenode with non-null op1 and op2 assert((tree->OperKind() & GTK_SMPOP) != 0); GenTree* op1 = tree->gtOp.gtOp1; GenTree* op2 = tree->gtGetOp2(); return optCSE_canSwap(op1, op2); } /***************************************************************************** * * Compare function passed to qsort() by CSE_Heuristic::SortCandidates * when (CodeOptKind() != Compiler::SMALL_CODE) */ /* static */ int __cdecl Compiler::optCSEcostCmpEx(const void* op1, const void* op2) { CSEdsc* dsc1 = *(CSEdsc**)op1; CSEdsc* dsc2 = *(CSEdsc**)op2; GenTree* exp1 = dsc1->csdTree; GenTree* exp2 = dsc2->csdTree; int diff; diff = (int)(exp2->gtCostEx - exp1->gtCostEx); if (diff != 0) { return diff; } // Sort the higher Use Counts toward the top diff = (int)(dsc2->csdUseWtCnt - dsc1->csdUseWtCnt); if (diff != 0) { return diff; } // With the same use count, Sort the lower Def Counts toward the top diff = (int)(dsc1->csdDefWtCnt - dsc2->csdDefWtCnt); if (diff != 0) { return diff; } // In order to ensure that we have a stable sort, we break ties using the csdIndex return (int)(dsc1->csdIndex - dsc2->csdIndex); } /***************************************************************************** * * Compare function passed to qsort() by CSE_Heuristic::SortCandidates * when (CodeOptKind() == Compiler::SMALL_CODE) */ /* static */ int __cdecl Compiler::optCSEcostCmpSz(const void* op1, const void* op2) { CSEdsc* dsc1 = *(CSEdsc**)op1; CSEdsc* dsc2 = *(CSEdsc**)op2; GenTree* exp1 = dsc1->csdTree; GenTree* exp2 = dsc2->csdTree; int diff; diff = (int)(exp2->gtCostSz - exp1->gtCostSz); if (diff != 0) { return diff; } // Sort the higher Use Counts toward the top diff = (int)(dsc2->csdUseCount - dsc1->csdUseCount); if (diff != 0) { return diff; } // With the same use count, Sort the lower Def Counts toward the top diff = (int)(dsc1->csdDefCount - dsc2->csdDefCount); if (diff != 0) { return diff; } // In order to ensure that we have a stable sort, we break ties using the csdIndex return (int)(dsc1->csdIndex - dsc2->csdIndex); } /*****************************************************************************/ #if FEATURE_VALNUM_CSE /*****************************************************************************/ /***************************************************************************** * * Initialize the Value Number CSE tracking logic. */ void Compiler::optValnumCSE_Init() { #ifdef DEBUG optCSEtab = nullptr; #endif // Init traits and full/empty bitvectors. This will be used to track the // individual cse indexes. cseTraits = new (getAllocator()) BitVecTraits(EXPSET_SZ, this); cseFull = BitVecOps::MakeFull(cseTraits); /* Allocate and clear the hash bucket table */ optCSEhash = new (this, CMK_CSE) CSEdsc*[s_optCSEhashSize](); optCSECandidateCount = 0; optDoCSE = false; // Stays false until we find duplicate CSE tree // optCseCheckedBoundMap is unused in most functions, allocated only when used optCseCheckedBoundMap = nullptr; } /***************************************************************************** * * Assign an index to the given expression (adding it to the lookup table, * if necessary). Returns the index or 0 if the expression can not be a CSE. */ unsigned Compiler::optValnumCSE_Index(GenTree* tree, GenTree* stmt) { unsigned key; unsigned hash; unsigned hval; CSEdsc* hashDsc; ValueNum vnlib = tree->GetVN(VNK_Liberal); /* Compute the hash value for the expression */ key = (unsigned)vnlib; hash = key; hash *= (unsigned)(s_optCSEhashSize + 1); hash >>= 7; hval = hash % s_optCSEhashSize; /* Look for a matching index in the hash table */ bool newCSE = false; for (hashDsc = optCSEhash[hval]; hashDsc; hashDsc = hashDsc->csdNextInBucket) { if (hashDsc->csdHashValue == key) { treeStmtLst* newElem; /* Have we started the list of matching nodes? */ if (hashDsc->csdTreeList == nullptr) { // Create the new element based upon the matching hashDsc element. newElem = new (this, CMK_TreeStatementList) treeStmtLst; newElem->tslTree = hashDsc->csdTree; newElem->tslStmt = hashDsc->csdStmt; newElem->tslBlock = hashDsc->csdBlock; newElem->tslNext = nullptr; /* Start the list with the first CSE candidate recorded */ hashDsc->csdTreeList = newElem; hashDsc->csdTreeLast = newElem; } noway_assert(hashDsc->csdTreeList); /* Append this expression to the end of the list */ newElem = new (this, CMK_TreeStatementList) treeStmtLst; newElem->tslTree = tree; newElem->tslStmt = stmt; newElem->tslBlock = compCurBB; newElem->tslNext = nullptr; hashDsc->csdTreeLast->tslNext = newElem; hashDsc->csdTreeLast = newElem; optDoCSE = true; // Found a duplicate CSE tree /* Have we assigned a CSE index? */ if (hashDsc->csdIndex == 0) { newCSE = true; break; } #if 0 // Use this to see if this Value Number base CSE is also a lexical CSE bool treeMatch = GenTree::Compare(hashDsc->csdTree, tree, true); #endif assert(FitsIn(hashDsc->csdIndex)); tree->gtCSEnum = ((signed char)hashDsc->csdIndex); return hashDsc->csdIndex; } } if (!newCSE) { /* Not found, create a new entry (unless we have too many already) */ if (optCSECandidateCount < MAX_CSE_CNT) { hashDsc = new (this, CMK_CSE) CSEdsc; hashDsc->csdHashValue = key; hashDsc->csdIndex = 0; hashDsc->csdLiveAcrossCall = 0; hashDsc->csdDefCount = 0; hashDsc->csdUseCount = 0; hashDsc->csdDefWtCnt = 0; hashDsc->csdUseWtCnt = 0; hashDsc->csdTree = tree; hashDsc->csdStmt = stmt; hashDsc->csdBlock = compCurBB; hashDsc->csdTreeList = nullptr; /* Append the entry to the hash bucket */ hashDsc->csdNextInBucket = optCSEhash[hval]; optCSEhash[hval] = hashDsc; } return 0; } else // newCSE is true { /* We get here only after finding a matching CSE */ /* Create a new CSE (unless we have the maximum already) */ if (optCSECandidateCount == MAX_CSE_CNT) { return 0; } C_ASSERT((signed char)MAX_CSE_CNT == MAX_CSE_CNT); unsigned CSEindex = ++optCSECandidateCount; // EXPSET_TP CSEmask = genCSEnum2bit(CSEindex); /* Record the new CSE index in the hashDsc */ hashDsc->csdIndex = CSEindex; /* Update the gtCSEnum field in the original tree */ noway_assert(hashDsc->csdTreeList->tslTree->gtCSEnum == 0); assert(FitsIn(CSEindex)); hashDsc->csdTreeList->tslTree->gtCSEnum = ((signed char)CSEindex); noway_assert(((unsigned)hashDsc->csdTreeList->tslTree->gtCSEnum) == CSEindex); tree->gtCSEnum = ((signed char)CSEindex); #ifdef DEBUG if (verbose) { EXPSET_TP tempMask = BitVecOps::MakeSingleton(cseTraits, genCSEnum2bit(CSEindex)); printf("\nCSE candidate #%02u, vn=", CSEindex); vnPrint(vnlib, 0); printf(" cseMask=%s in " FMT_BB ", [cost=%2u, size=%2u]: \n", genES2str(cseTraits, tempMask), compCurBB->bbNum, tree->gtCostEx, tree->gtCostSz); gtDispTree(tree); } #endif // DEBUG return CSEindex; } } /***************************************************************************** * * Locate CSE candidates and assign indices to them * return 0 if no CSE candidates were found * Also initialize bbCseIn, bbCseout and bbCseGen sets for all blocks */ unsigned Compiler::optValnumCSE_Locate() { // Locate CSE candidates and assign them indices for (BasicBlock* block = fgFirstBB; block; block = block->bbNext) { GenTree* stmt; GenTree* tree; /* Make the block publicly available */ compCurBB = block; /* Ensure that the BBF_VISITED and BBF_MARKED flag are clear */ /* Everyone who uses these flags are required to clear afterwards */ noway_assert((block->bbFlags & (BBF_VISITED | BBF_MARKED)) == 0); /* Walk the statement trees in this basic block */ for (stmt = block->FirstNonPhiDef(); stmt; stmt = stmt->gtNext) { noway_assert(stmt->gtOper == GT_STMT); /* We walk the tree in the forwards direction (bottom up) */ bool stmtHasArrLenCandidate = false; for (tree = stmt->gtStmt.gtStmtList; tree; tree = tree->gtNext) { if (tree->OperIsCompare() && stmtHasArrLenCandidate) { // Check if this compare is a function of (one of) the checked // bound candidate(s); we may want to update its value number. // if the array length gets CSEd optCseUpdateCheckedBoundMap(tree); } if (!optIsCSEcandidate(tree)) { continue; } ValueNum vnlib = tree->GetVN(VNK_Liberal); if (ValueNumStore::isReservedVN(vnlib)) { continue; } // Don't CSE constant values, instead let the Value Number // based Assertion Prop phase handle them. Here, unlike // the rest of optCSE, we use the conservative value number // rather than the liberal one, since the conservative one // is what the Value Number based Assertion Prop will use // and the point is to avoid optimizing cases that it will // handle. // if (vnStore->IsVNConstant(tree->GetVN(VNK_Conservative))) { continue; } /* Assign an index to this expression */ unsigned CSEindex = optValnumCSE_Index(tree, stmt); if (CSEindex != 0) { noway_assert(((unsigned)tree->gtCSEnum) == CSEindex); } if (IS_CSE_INDEX(CSEindex) && (tree->OperGet() == GT_ARR_LENGTH)) { stmtHasArrLenCandidate = true; } } } } /* We're done if there were no interesting expressions */ if (!optDoCSE) { return 0; } /* We're finished building the expression lookup table */ optCSEstop(); return 1; } //------------------------------------------------------------------------ // optCseUpdateCheckedBoundMap: Check if this compare is a tractable function of // a checked bound that is a CSE candidate, and insert // an entry in the optCseCheckedBoundMap if so. This facilitates // subsequently updating the compare's value number if // the bound gets CSEd. // // Arguments: // compare - The compare node to check void Compiler::optCseUpdateCheckedBoundMap(GenTree* compare) { assert(compare->OperIsCompare()); ValueNum compareVN = compare->gtVNPair.GetConservative(); VNFuncApp cmpVNFuncApp; if (!vnStore->GetVNFunc(compareVN, &cmpVNFuncApp) || (cmpVNFuncApp.m_func != GetVNFuncForNode(compare))) { // Value numbering inferred this compare as something other // than its own operator; leave its value number alone. return; } // Now look for a checked bound feeding the compare ValueNumStore::CompareCheckedBoundArithInfo info; GenTree* boundParent = nullptr; if (vnStore->IsVNCompareCheckedBound(compareVN)) { // Simple compare of an bound against something else. vnStore->GetCompareCheckedBound(compareVN, &info); boundParent = compare; } else if (vnStore->IsVNCompareCheckedBoundArith(compareVN)) { // Compare of a bound +/- some offset to something else. GenTree* op1 = compare->gtGetOp1(); GenTree* op2 = compare->gtGetOp2(); vnStore->GetCompareCheckedBoundArithInfo(compareVN, &info); if (GetVNFuncForNode(op1) == (VNFunc)info.arrOper) { // The arithmetic node is the bound's parent. boundParent = op1; } else if (GetVNFuncForNode(op2) == (VNFunc)info.arrOper) { // The arithmetic node is the bound's parent. boundParent = op2; } } if (boundParent != nullptr) { GenTree* bound = nullptr; // Find which child of boundParent is the bound. Abort if neither // conservative value number matches the one from the compare VN. GenTree* child1 = boundParent->gtGetOp1(); if ((info.vnBound == child1->gtVNPair.GetConservative()) && IS_CSE_INDEX(child1->gtCSEnum)) { bound = child1; } else { GenTree* child2 = boundParent->gtGetOp2(); if ((info.vnBound == child2->gtVNPair.GetConservative()) && IS_CSE_INDEX(child2->gtCSEnum)) { bound = child2; } } if (bound != nullptr) { // Found a checked bound feeding a compare that is a tractable function of it; // record this in the map so we can update the compare VN if the bound // node gets CSEd. if (optCseCheckedBoundMap == nullptr) { // Allocate map on first use. optCseCheckedBoundMap = new (getAllocator()) NodeToNodeMap(getAllocator()); } optCseCheckedBoundMap->Set(bound, compare); } } } /***************************************************************************** * * Compute each blocks bbCseGen * This is the bitset that represents the CSEs that are generated within the block */ void Compiler::optValnumCSE_InitDataFlow() { for (BasicBlock* block = fgFirstBB; block; block = block->bbNext) { /* Initialize the blocks's bbCseIn set */ bool init_to_zero = false; if (block == fgFirstBB) { /* Clear bbCseIn for the entry block */ init_to_zero = true; } #if !CSE_INTO_HANDLERS else { if (bbIsHandlerBeg(block)) { /* Clear everything on entry to filters or handlers */ init_to_zero = true; } } #endif if (init_to_zero) { /* Initialize to {ZERO} prior to dataflow */ block->bbCseIn = BitVecOps::MakeEmpty(cseTraits); } else { /* Initialize to {ALL} prior to dataflow */ block->bbCseIn = BitVecOps::MakeCopy(cseTraits, cseFull); } block->bbCseOut = BitVecOps::MakeCopy(cseTraits, cseFull); /* Initialize to {ZERO} prior to locating the CSE candidates */ block->bbCseGen = BitVecOps::MakeEmpty(cseTraits); } // We walk the set of CSE candidates and set the bit corresponsing to the CSEindex // in the block's bbCseGen bitset // for (unsigned cnt = 0; cnt < optCSECandidateCount; cnt++) { CSEdsc* dsc = optCSEtab[cnt]; unsigned CSEindex = dsc->csdIndex; treeStmtLst* lst = dsc->csdTreeList; noway_assert(lst); while (lst != nullptr) { BasicBlock* block = lst->tslBlock; BitVecOps::AddElemD(cseTraits, block->bbCseGen, genCSEnum2bit(CSEindex)); lst = lst->tslNext; } } #ifdef DEBUG // Dump out the bbCseGen information that we just created // if (verbose) { bool headerPrinted = false; for (BasicBlock* block = fgFirstBB; block; block = block->bbNext) { if (block->bbCseGen != nullptr) { if (!headerPrinted) { printf("\nBlocks that generate CSE def/uses\n"); headerPrinted = true; } printf(FMT_BB, block->bbNum); printf(" cseGen = %s\n", genES2str(cseTraits, block->bbCseGen)); } } } fgDebugCheckLinks(); #endif // DEBUG } /***************************************************************************** * * CSE Dataflow, so that all helper methods for dataflow are in a single place * */ class CSE_DataFlow { BitVecTraits* m_pBitVecTraits; EXPSET_TP m_preMergeOut; public: CSE_DataFlow(Compiler* pCompiler) : m_pBitVecTraits(pCompiler->cseTraits), m_preMergeOut(BitVecOps::UninitVal()) { } // At the start of the merge function of the dataflow equations, initialize premerge state (to detect changes.) void StartMerge(BasicBlock* block) { BitVecOps::Assign(m_pBitVecTraits, m_preMergeOut, block->bbCseOut); } // During merge, perform the actual merging of the predecessor's (since this is a forward analysis) dataflow flags. void Merge(BasicBlock* block, BasicBlock* predBlock, flowList* preds) { BitVecOps::IntersectionD(m_pBitVecTraits, block->bbCseIn, predBlock->bbCseOut); } // At the end of the merge store results of the dataflow equations, in a postmerge state. bool EndMerge(BasicBlock* block) { BitVecOps::DataFlowD(m_pBitVecTraits, block->bbCseOut, block->bbCseGen, block->bbCseIn); return !BitVecOps::Equal(m_pBitVecTraits, block->bbCseOut, m_preMergeOut); } }; /***************************************************************************** * * Perform a DataFlow forward analysis using the block CSE bitsets: * Inputs: * bbCseGen - Exact CSEs that are become available within the block * bbCseIn - Maximal estimate of CSEs that are/could be available at input to the block * bbCseOut - Maximal estimate of CSEs that are/could be available at exit to the block * * Outputs: * bbCseIn - Computed CSEs that are available at input to the block * bbCseOut - Computed CSEs that are available at exit to the block */ void Compiler::optValnumCSE_DataFlow() { CSE_DataFlow cse(this); // Modified dataflow algorithm for available expressions. DataFlow cse_flow(this); cse_flow.ForwardAnalysis(cse); #ifdef DEBUG if (verbose) { printf("\nAfter performing DataFlow for ValnumCSE's\n"); for (BasicBlock* block = fgFirstBB; block; block = block->bbNext) { printf(FMT_BB, block->bbNum); printf(" cseIn = %s", genES2str(cseTraits, block->bbCseIn)); printf(" cseOut = %s", genES2str(cseTraits, block->bbCseOut)); printf("\n"); } printf("\n"); } #endif // DEBUG } /***************************************************************************** * * Using the information computed by CSE_DataFlow determine for each * CSE whether the CSE is a definition (if the CSE was not available) * or if the CSE is a use (if the CSE was previously made available) * The implementation iterates of all blocks setting 'available_cses' * to the CSEs that are available at input to the block. * When a CSE expression is encountered it is classified as either * as a definition (if the CSE is not in the 'available_cses' set) or * as a use (if the CSE is in the 'available_cses' set). If the CSE * is a definition then it is added to the 'available_cses' set. * In the Value Number based CSEs we do not need to have kill sets */ void Compiler::optValnumCSE_Availablity() { #ifdef DEBUG if (verbose) { printf("Labeling the CSEs with Use/Def information\n"); } #endif EXPSET_TP available_cses = BitVecOps::MakeEmpty(cseTraits); for (BasicBlock* block = fgFirstBB; block; block = block->bbNext) { GenTree* stmt; GenTree* tree; /* Make the block publicly available */ compCurBB = block; BitVecOps::Assign(cseTraits, available_cses, block->bbCseIn); optCSEweight = block->getBBWeight(this); /* Walk the statement trees in this basic block */ for (stmt = block->FirstNonPhiDef(); stmt; stmt = stmt->gtNext) { noway_assert(stmt->gtOper == GT_STMT); /* We walk the tree in the forwards direction (bottom up) */ for (tree = stmt->gtStmt.gtStmtList; tree; tree = tree->gtNext) { if (IS_CSE_INDEX(tree->gtCSEnum)) { unsigned int cseBit = genCSEnum2bit(tree->gtCSEnum); CSEdsc* desc = optCSEfindDsc(tree->gtCSEnum); unsigned stmw = block->getBBWeight(this); /* Is this expression available here? */ if (BitVecOps::IsMember(cseTraits, available_cses, cseBit)) { /* This is a CSE use */ desc->csdUseCount += 1; desc->csdUseWtCnt += stmw; } else { if (tree->gtFlags & GTF_COLON_COND) { // We can't create CSE definitions inside QMARK-COLON trees tree->gtCSEnum = NO_CSE; continue; } /* This is a CSE def */ if (desc->csdDefCount == 0) { // This is the first def visited, so copy its conservative VN desc->defConservativeVN = tree->gtVNPair.GetConservative(); } else if (tree->gtVNPair.GetConservative() != desc->defConservativeVN) { // This candidate has defs with differing conservative VNs desc->defConservativeVN = ValueNumStore::NoVN; } desc->csdDefCount += 1; desc->csdDefWtCnt += stmw; /* Mark the node as a CSE definition */ tree->gtCSEnum = TO_CSE_DEF(tree->gtCSEnum); /* This CSE will be available after this def */ BitVecOps::AddElemD(cseTraits, available_cses, cseBit); } #ifdef DEBUG if (verbose && IS_CSE_INDEX(tree->gtCSEnum)) { printf(FMT_BB " ", block->bbNum); printTreeID(tree); printf(" %s of CSE #%02u [weight=%s]\n", IS_CSE_USE(tree->gtCSEnum) ? "Use" : "Def", GET_CSE_INDEX(tree->gtCSEnum), refCntWtd2str(stmw)); } #endif } } } } } // The following class handles the CSE heuristics // we use a complex set of heuristic rules // to determine if it is likely to be profitable to perform this CSE // class CSE_Heuristic { Compiler* m_pCompiler; unsigned m_addCSEcount; unsigned aggressiveRefCnt; unsigned moderateRefCnt; unsigned enregCount; // count of the number of enregisterable variables bool largeFrame; bool hugeFrame; Compiler::codeOptimize codeOptKind; Compiler::CSEdsc** sortTab; size_t sortSiz; #ifdef DEBUG CLRRandom m_cseRNG; unsigned m_bias; #endif public: CSE_Heuristic(Compiler* pCompiler) : m_pCompiler(pCompiler) { codeOptKind = m_pCompiler->compCodeOpt(); } Compiler::codeOptimize CodeOptKind() { return codeOptKind; } // Perform the Initialization step for our CSE Heuristics // determine the various cut off values to use for // the aggressive, moderate and conservative CSE promotions // count the number of enregisterable variables // determine if the method has a large or huge stack frame. // void Initialize() { m_addCSEcount = 0; /* Count of the number of LclVars for CSEs that we added */ // Record the weighted ref count of the last "for sure" callee saved LclVar aggressiveRefCnt = 0; moderateRefCnt = 0; enregCount = 0; largeFrame = false; hugeFrame = false; sortTab = nullptr; sortSiz = 0; #ifdef _TARGET_XARCH_ if (m_pCompiler->compLongUsed) { enregCount++; } #endif unsigned frameSize = 0; unsigned regAvailEstimate = ((CNT_CALLEE_ENREG * 3) + (CNT_CALLEE_TRASH * 2) + 1); unsigned lclNum; LclVarDsc* varDsc; for (lclNum = 0, varDsc = m_pCompiler->lvaTable; lclNum < m_pCompiler->lvaCount; lclNum++, varDsc++) { if (varDsc->lvRefCnt() == 0) { continue; } #if FEATURE_FIXED_OUT_ARGS // Skip the OutgoingArgArea in computing frame size, since // its size is not yet known and it doesn't affect local // offsets from the frame pointer (though it may affect // them from the stack pointer). noway_assert(m_pCompiler->lvaOutgoingArgSpaceVar != BAD_VAR_NUM); if (lclNum == m_pCompiler->lvaOutgoingArgSpaceVar) { continue; } #endif // FEATURE_FIXED_OUT_ARGS bool onStack = (regAvailEstimate == 0); // true when it is likely that this LclVar will have a stack home // Some LclVars always have stack homes if ((varDsc->lvDoNotEnregister) || (varDsc->lvType == TYP_LCLBLK)) { onStack = true; } #ifdef _TARGET_X86_ // Treat floating point and 64 bit integers as always on the stack if (varTypeIsFloating(varDsc->TypeGet()) || varTypeIsLong(varDsc->TypeGet())) onStack = true; #endif if (onStack) { frameSize += m_pCompiler->lvaLclSize(lclNum); } else { // For the purposes of estimating the frameSize we // will consider this LclVar as being enregistered. // Now we reduce the remaining regAvailEstimate by // an appropriate amount. if (varDsc->lvRefCnt() <= 2) { // a single use single def LclVar only uses 1 regAvailEstimate -= 1; } else { // a LclVar with multiple uses and defs uses 2 if (regAvailEstimate >= 2) { regAvailEstimate -= 2; } else { // Don't try to subtract when regAvailEstimate is 1 regAvailEstimate = 0; } } } #ifdef _TARGET_XARCH_ if (frameSize > 0x080) { // We likely have a large stack frame. // Thus we might need to use large displacements when loading or storing // to CSE LclVars that are not enregistered largeFrame = true; break; // early out, we don't need to keep increasing frameSize } #else // _TARGET_ARM_ if (frameSize > 0x0400) { largeFrame = true; } if (frameSize > 0x10000) { hugeFrame = true; break; } #endif } unsigned sortNum = 0; while (sortNum < m_pCompiler->lvaTrackedCount) { LclVarDsc* varDsc = m_pCompiler->lvaRefSorted[sortNum++]; var_types varTyp = varDsc->TypeGet(); if (varDsc->lvDoNotEnregister) { continue; } if (!varTypeIsFloating(varTyp)) { // TODO-1stClassStructs: Remove this; it is here to duplicate previous behavior. // Note that this makes genTypeStSz return 1. if (varTypeIsStruct(varTyp)) { varTyp = TYP_STRUCT; } enregCount += genTypeStSz(varTyp); } if ((aggressiveRefCnt == 0) && (enregCount > (CNT_CALLEE_ENREG * 3 / 2))) { if (CodeOptKind() == Compiler::SMALL_CODE) { aggressiveRefCnt = varDsc->lvRefCnt() + BB_UNITY_WEIGHT; } else { aggressiveRefCnt = varDsc->lvRefCntWtd() + BB_UNITY_WEIGHT; } } if ((moderateRefCnt == 0) && (enregCount > ((CNT_CALLEE_ENREG * 3) + (CNT_CALLEE_TRASH * 2)))) { if (CodeOptKind() == Compiler::SMALL_CODE) { moderateRefCnt = varDsc->lvRefCnt(); } else { moderateRefCnt = varDsc->lvRefCntWtd(); } } } unsigned mult = 3; // use smaller value for mult when enregCount is in [0..4] if (enregCount <= 4) { mult = (enregCount <= 2) ? 1 : 2; } aggressiveRefCnt = max(BB_UNITY_WEIGHT * mult, aggressiveRefCnt); moderateRefCnt = max((BB_UNITY_WEIGHT * mult) / 2, moderateRefCnt); #ifdef DEBUG if (m_pCompiler->verbose) { printf("\n"); printf("Aggressive CSE Promotion cutoff is %u\n", aggressiveRefCnt); printf("Moderate CSE Promotion cutoff is %u\n", moderateRefCnt); printf("Framesize estimate is 0x%04X\n", frameSize); printf("We have a %s frame\n", hugeFrame ? "huge" : (largeFrame ? "large" : "small")); } #endif } void SortCandidates() { /* Create an expression table sorted by decreasing cost */ sortTab = new (m_pCompiler, CMK_CSE) Compiler::CSEdsc*[m_pCompiler->optCSECandidateCount]; sortSiz = m_pCompiler->optCSECandidateCount * sizeof(*sortTab); memcpy(sortTab, m_pCompiler->optCSEtab, sortSiz); if (CodeOptKind() == Compiler::SMALL_CODE) { qsort(sortTab, m_pCompiler->optCSECandidateCount, sizeof(*sortTab), m_pCompiler->optCSEcostCmpSz); } else { qsort(sortTab, m_pCompiler->optCSECandidateCount, sizeof(*sortTab), m_pCompiler->optCSEcostCmpEx); } #ifdef DEBUG if (m_pCompiler->verbose) { printf("\nSorted CSE candidates:\n"); /* Print out the CSE candidates */ EXPSET_TP tempMask; for (unsigned cnt = 0; cnt < m_pCompiler->optCSECandidateCount; cnt++) { Compiler::CSEdsc* dsc = sortTab[cnt]; GenTree* expr = dsc->csdTree; unsigned def; unsigned use; if (CodeOptKind() == Compiler::SMALL_CODE) { def = dsc->csdDefCount; // def count use = dsc->csdUseCount; // use count (excluding the implicit uses at defs) } else { def = dsc->csdDefWtCnt; // weighted def count use = dsc->csdUseWtCnt; // weighted use count (excluding the implicit uses at defs) } tempMask = BitVecOps::MakeSingleton(m_pCompiler->cseTraits, genCSEnum2bit(dsc->csdIndex)); printf("CSE #%02u,cseMask=%s,useCnt=%d: [def=%3u, use=%3u", dsc->csdIndex, genES2str(m_pCompiler->cseTraits, tempMask), dsc->csdUseCount, def, use); printf("] :: "); m_pCompiler->gtDispTree(expr, nullptr, nullptr, true); } printf("\n"); } #endif // DEBUG } // The following class nested within CSE_Heuristic encapsulates the information // about the current CSE candidate that is under consideration // // TODO-Cleanup: This is still very much based upon the old Lexical CSE implementation // and needs to be reworked for the Value Number based implementation // class CSE_Candidate { CSE_Heuristic* m_context; Compiler::CSEdsc* m_CseDsc; unsigned m_cseIndex; unsigned m_defCount; unsigned m_useCount; unsigned m_Cost; unsigned m_Size; public: CSE_Candidate(CSE_Heuristic* context, Compiler::CSEdsc* cseDsc) : m_context(context), m_CseDsc(cseDsc) { m_cseIndex = m_CseDsc->csdIndex; } Compiler::CSEdsc* CseDsc() { return m_CseDsc; } unsigned CseIndex() { return m_cseIndex; } unsigned DefCount() { return m_defCount; } unsigned UseCount() { return m_useCount; } // TODO-CQ: With ValNum CSE's the Expr and its cost can vary. GenTree* Expr() { return m_CseDsc->csdTree; } unsigned Cost() { return m_Cost; } unsigned Size() { return m_Size; } bool LiveAcrossCall() { return (m_CseDsc->csdLiveAcrossCall != 0); } void InitializeCounts() { if (m_context->CodeOptKind() == Compiler::SMALL_CODE) { m_Cost = Expr()->gtCostSz; // the estimated code size m_Size = Expr()->gtCostSz; // always the gtCostSz m_defCount = m_CseDsc->csdDefCount; // def count m_useCount = m_CseDsc->csdUseCount; // use count (excluding the implicit uses at defs) } else { m_Cost = Expr()->gtCostEx; // the estimated execution cost m_Size = Expr()->gtCostSz; // always the gtCostSz m_defCount = m_CseDsc->csdDefWtCnt; // weighted def count m_useCount = m_CseDsc->csdUseWtCnt; // weighted use count (excluding the implicit uses at defs) } } }; #ifdef DEBUG //------------------------------------------------------------------------ // optConfigBiasedCSE: // Stress mode to shuffle the decision to CSE or not using environment // variable COMPlus_JitStressBiasedCSE (= 0 to 100%). When the bias value // is not specified but COMPlus_JitStress is ON, generate a random bias. // // Return Value: // 0 -- This method is indifferent about this CSE (no bias specified and no stress) // 1 -- This CSE must be performed to maintain specified/generated bias. // -1 -- This CSE mustn't be performed to maintain specified/generated bias. // // Operation: // A debug stress only method that returns "1" with probability (P) // defined by: // // P = (COMPlus_JitStressBiasedCSE / 100) (or) // P = (random(100) / 100) when COMPlus_JitStress is specified and // COMPlus_JitStressBiasedCSE is unspecified. // // When specified, the bias is reinterpreted as a decimal number between 0 // to 100. // When bias is not specified, a bias is randomly generated if COMPlus_JitStress // is non-zero. // // Callers are supposed to call this method for each CSE promotion decision // and ignore the call if return value is 0 and honor the 1 with a CSE and // -1 with a no-CSE to maintain the specified/generated bias. // int optConfigBiasedCSE() { // Seed the PRNG, if never done before. if (!m_cseRNG.IsInitialized()) { m_cseRNG.Init(m_pCompiler->info.compMethodHash()); m_bias = m_cseRNG.Next(100); } // Obtain the bias value and reinterpret as decimal. unsigned bias = ReinterpretHexAsDecimal(JitConfig.JitStressBiasedCSE()); // Invalid value, check if JitStress is ON. if (bias > 100) { if (!m_pCompiler->compStressCompile(Compiler::STRESS_MAKE_CSE, MAX_STRESS_WEIGHT)) { // JitStress is OFF for CSE, nothing to do. return 0; } bias = m_bias; JITDUMP("JitStressBiasedCSE is OFF, but JitStress is ON: generated bias=%d.\n", bias); } // Generate a number between (0, 99) and if the generated // number is smaller than bias, then perform CSE. unsigned gen = m_cseRNG.Next(100); int ret = (gen < bias) ? 1 : -1; if (m_pCompiler->verbose) { if (ret < 0) { printf("No CSE because gen=%d >= bias=%d\n", gen, bias); } else { printf("Promoting CSE because gen=%d < bias=%d\n", gen, bias); } } // Indicate whether to perform CSE or not. return ret; } #endif // Given a CSE candidate decide whether it passes or fails the profitability heuristic // return true if we believe that it is profitable to promote this candidate to a CSE // bool PromotionCheck(CSE_Candidate* candidate) { bool result = false; #ifdef DEBUG int stressResult = optConfigBiasedCSE(); if (stressResult != 0) { // Stress is enabled. Check whether to perform CSE or not. return (stressResult > 0); } if (m_pCompiler->optConfigDisableCSE2()) { return false; // skip this CSE } #endif /* Our calculation is based on the following cost estimate formula Existing costs are: (def + use) * cost If we introduce a CSE temp are each definition and replace the use with a CSE temp then our cost is: (def * (cost + cse-def-cost)) + (use * cse-use-cost) We must estimate the values to use for cse-def-cost and cse-use-cost If we are able to enregister the CSE then the cse-use-cost is one and cse-def-cost is either zero or one. Zero in the case where we needed to evaluate the def into a register and we can use that register as the CSE temp as well. If we are unable to enregister the CSE then the cse-use-cost is IND_COST and the cse-def-cost is also IND_COST. If we want to be conservative we use IND_COST as the the value for both cse-def-cost and cse-use-cost and then we never introduce a CSE that could pessimize the execution time of the method. If we want to be more moderate we use (IND_COST_EX + 1) / 2 as the values for both cse-def-cost and cse-use-cost. If we want to be aggressive we use 1 as the values for both cse-def-cost and cse-use-cost. If we believe that the CSE very valuable in terms of weighted ref counts such that it would always be enregistered by the register allocator we choose the aggressive use def costs. If we believe that the CSE is somewhat valuable in terms of weighted ref counts such that it could be likely be enregistered by the register allocator we choose the moderate use def costs. otherwise we choose the conservative use def costs. */ unsigned cse_def_cost; unsigned cse_use_cost; unsigned no_cse_cost = 0; unsigned yes_cse_cost = 0; unsigned extra_yes_cost = 0; unsigned extra_no_cost = 0; // The 'cseRefCnt' is the RefCnt that we will have if we promote this CSE into a new LclVar // Each CSE Def will contain two Refs and each CSE Use will have one Ref of this new LclVar unsigned cseRefCnt = (candidate->DefCount() * 2) + candidate->UseCount(); if (CodeOptKind() == Compiler::SMALL_CODE) { if (cseRefCnt >= aggressiveRefCnt) { #ifdef DEBUG if (m_pCompiler->verbose) { printf("Aggressive CSE Promotion (%u >= %u)\n", cseRefCnt, aggressiveRefCnt); } #endif cse_def_cost = 1; cse_use_cost = 1; if (candidate->LiveAcrossCall() != 0) { if (largeFrame) { cse_def_cost++; cse_use_cost++; } if (hugeFrame) { cse_def_cost++; cse_use_cost++; } } } else if (largeFrame) { #ifdef DEBUG if (m_pCompiler->verbose) { printf("Codesize CSE Promotion (large frame)\n"); } #endif #ifdef _TARGET_XARCH_ /* The following formula is good choice when optimizing CSE for SMALL_CODE */ cse_def_cost = 6; // mov [EBP-0x00001FC],reg cse_use_cost = 5; // [EBP-0x00001FC] #else // _TARGET_ARM_ if (hugeFrame) { cse_def_cost = 12; // movw/movt r10 and str reg,[sp+r10] cse_use_cost = 12; } else { cse_def_cost = 8; // movw r10 and str reg,[sp+r10] cse_use_cost = 8; } #endif } else // small frame { #ifdef DEBUG if (m_pCompiler->verbose) { printf("Codesize CSE Promotion (small frame)\n"); } #endif #ifdef _TARGET_XARCH_ /* The following formula is good choice when optimizing CSE for SMALL_CODE */ cse_def_cost = 3; // mov [EBP-1C],reg cse_use_cost = 2; // [EBP-1C] #else // _TARGET_ARM_ cse_def_cost = 2; // str reg,[sp+0x9c] cse_use_cost = 2; // ldr reg,[sp+0x9c] #endif } } else // not SMALL_CODE ... { if (cseRefCnt >= aggressiveRefCnt) { #ifdef DEBUG if (m_pCompiler->verbose) { printf("Aggressive CSE Promotion (%u >= %u)\n", cseRefCnt, aggressiveRefCnt); } #endif cse_def_cost = 1; cse_use_cost = 1; } else if (cseRefCnt >= moderateRefCnt) { if (candidate->LiveAcrossCall() == 0) { #ifdef DEBUG if (m_pCompiler->verbose) { printf("Moderate CSE Promotion (CSE never live at call) (%u >= %u)\n", cseRefCnt, moderateRefCnt); } #endif cse_def_cost = 2; cse_use_cost = 1; } else // candidate is live across call { #ifdef DEBUG if (m_pCompiler->verbose) { printf("Moderate CSE Promotion (%u >= %u)\n", cseRefCnt, moderateRefCnt); } #endif cse_def_cost = 2; cse_use_cost = 2; extra_yes_cost = BB_UNITY_WEIGHT * 2; // Extra cost in case we have to spill/restore a caller // saved register } } else // Conservative CSE promotion { if (candidate->LiveAcrossCall() == 0) { #ifdef DEBUG if (m_pCompiler->verbose) { printf("Conservative CSE Promotion (CSE never live at call) (%u < %u)\n", cseRefCnt, moderateRefCnt); } #endif cse_def_cost = 2; cse_use_cost = 2; } else // candidate is live across call { #ifdef DEBUG if (m_pCompiler->verbose) { printf("Conservative CSE Promotion (%u < %u)\n", cseRefCnt, moderateRefCnt); } #endif cse_def_cost = 3; cse_use_cost = 3; extra_yes_cost = BB_UNITY_WEIGHT * 4; // Extra cost in case we have to spill/restore a caller // saved register } // If we have maxed out lvaTrackedCount then this CSE may end up as an untracked variable if (m_pCompiler->lvaTrackedCount == lclMAX_TRACKED) { cse_def_cost++; cse_use_cost++; } } if (largeFrame) { cse_def_cost++; cse_use_cost++; } if (hugeFrame) { cse_def_cost++; cse_use_cost++; } } // estimate the cost from lost codesize reduction if we do not perform the CSE if (candidate->Size() > cse_use_cost) { Compiler::CSEdsc* dsc = candidate->CseDsc(); // We need to retrieve the actual use count, not the // weighted count extra_no_cost = candidate->Size() - cse_use_cost; extra_no_cost = extra_no_cost * dsc->csdUseCount * 2; } /* no_cse_cost is the cost estimate when we decide not to make a CSE */ /* yes_cse_cost is the cost estimate when we decide to make a CSE */ no_cse_cost = candidate->UseCount() * candidate->Cost(); yes_cse_cost = (candidate->DefCount() * cse_def_cost) + (candidate->UseCount() * cse_use_cost); no_cse_cost += extra_no_cost; yes_cse_cost += extra_yes_cost; #ifdef DEBUG if (m_pCompiler->verbose) { printf("cseRefCnt=%d, aggressiveRefCnt=%d, moderateRefCnt=%d\n", cseRefCnt, aggressiveRefCnt, moderateRefCnt); printf("defCnt=%d, useCnt=%d, cost=%d, size=%d\n", candidate->DefCount(), candidate->UseCount(), candidate->Cost(), candidate->Size()); printf("def_cost=%d, use_cost=%d, extra_no_cost=%d, extra_yes_cost=%d\n", cse_def_cost, cse_use_cost, extra_no_cost, extra_yes_cost); printf("CSE cost savings check (%u >= %u) %s\n", no_cse_cost, yes_cse_cost, (no_cse_cost >= yes_cse_cost) ? "passes" : "fails"); } #endif // Should we make this candidate into a CSE? // Is the yes cost less than the no cost // if (yes_cse_cost <= no_cse_cost) { result = true; // Yes make this a CSE } else { /* In stress mode we will make some extra CSEs */ if (no_cse_cost > 0) { int percentage = (no_cse_cost * 100) / yes_cse_cost; if (m_pCompiler->compStressCompile(Compiler::STRESS_MAKE_CSE, percentage)) { result = true; // Yes make this a CSE } } } return result; } // PerformCSE() takes a successful candidate and performs the appropriate replacements: // // It will replace all of the CSE defs with assignments to a new "cse0" LclVar // and will replace all of the CSE uses with reads of the "cse0" LclVar // void PerformCSE(CSE_Candidate* successfulCandidate) { unsigned cseRefCnt = (successfulCandidate->DefCount() * 2) + successfulCandidate->UseCount(); if (successfulCandidate->LiveAcrossCall() != 0) { // As we introduce new LclVars for these CSE we slightly // increase the cutoffs for aggressive and moderate CSE's // int incr = BB_UNITY_WEIGHT; if (cseRefCnt > aggressiveRefCnt) { aggressiveRefCnt += incr; } if (cseRefCnt > moderateRefCnt) { moderateRefCnt += (incr / 2); } } /* Introduce a new temp for the CSE */ // we will create a long lifetime temp for the new cse LclVar unsigned cseLclVarNum = m_pCompiler->lvaGrabTemp(false DEBUGARG("ValNumCSE")); var_types cseLclVarTyp = genActualType(successfulCandidate->Expr()->TypeGet()); if (varTypeIsStruct(cseLclVarTyp)) { m_pCompiler->lvaSetStruct(cseLclVarNum, m_pCompiler->gtGetStructHandle(successfulCandidate->Expr()), false); } m_pCompiler->lvaTable[cseLclVarNum].lvType = cseLclVarTyp; m_pCompiler->lvaTable[cseLclVarNum].lvIsCSE = true; m_addCSEcount++; // Record that we created a new LclVar for use as a CSE temp m_pCompiler->optCSEcount++; ValueNum defConservativeVN = successfulCandidate->CseDsc()->defConservativeVN; /* Walk all references to this CSE, adding an assignment to the CSE temp to all defs and changing all refs to a simple use of the CSE temp. We also unmark nested CSE's for all uses. */ Compiler::treeStmtLst* lst; lst = successfulCandidate->CseDsc()->csdTreeList; noway_assert(lst); #define QQQ_CHECK_CSE_VNS 0 #if QQQ_CHECK_CSE_VNS assert(lst != NULL); ValueNum firstVN = lst->tslTree->gtVN; lst = lst->tslNext; bool allSame = true; while (lst != NULL) { if (IS_CSE_INDEX(lst->tslTree->gtCSEnum)) { if (lst->tslTree->gtVN != firstVN) { allSame = false; break; } } lst = lst->tslNext; } if (!allSame) { lst = dsc->csdTreeList; GenTree* firstTree = lst->tslTree; printf("In %s, CSE (oper = %s, type = %s) has differing VNs: ", info.compFullName, GenTree::OpName(firstTree->OperGet()), varTypeName(firstTree->TypeGet())); while (lst != NULL) { if (IS_CSE_INDEX(lst->tslTree->gtCSEnum)) { printf("0x%x(%s,%d) ", lst->tslTree, IS_CSE_USE(lst->tslTree->gtCSEnum) ? "u" : "d", lst->tslTree->gtVN); } lst = lst->tslNext; } printf("\n"); } lst = dsc->csdTreeList; #endif do { /* Process the next node in the list */ GenTree* exp = lst->tslTree; GenTree* stm = lst->tslStmt; noway_assert(stm->gtOper == GT_STMT); BasicBlock* blk = lst->tslBlock; /* Advance to the next node in the list */ lst = lst->tslNext; // Assert if we used DEBUG_DESTROY_NODE on this CSE exp assert(exp->gtOper != GT_COUNT); /* Ignore the node if it's not been marked as a CSE */ if (!IS_CSE_INDEX(exp->gtCSEnum)) { continue; } /* Make sure we update the weighted ref count correctly */ m_pCompiler->optCSEweight = blk->getBBWeight(m_pCompiler); /* Figure out the actual type of the value */ var_types expTyp = genActualType(exp->TypeGet()); noway_assert(expTyp == cseLclVarTyp); // This will contain the replacement tree for exp // It will either be the CSE def or CSE ref // GenTree* cse = nullptr; bool isDef; FieldSeqNode* fldSeq = nullptr; bool hasZeroMapAnnotation = m_pCompiler->GetZeroOffsetFieldMap()->Lookup(exp, &fldSeq); if (IS_CSE_USE(exp->gtCSEnum)) { /* This is a use of the CSE */ isDef = false; #ifdef DEBUG if (m_pCompiler->verbose) { printf("\nWorking on the replacement of the CSE #%02u use at ", exp->gtCSEnum); Compiler::printTreeID(exp); printf(" in " FMT_BB "\n", blk->bbNum); } #endif // DEBUG // We will replace the CSE ref with a new tree // this is typically just a simple use of the new CSE LclVar // cse = m_pCompiler->gtNewLclvNode(cseLclVarNum, cseLclVarTyp); cse->gtVNPair = exp->gtVNPair; // assign the proper Value Numbers if (defConservativeVN != ValueNumStore::NoVN) { // All defs of this CSE share the same conservative VN, and we are rewriting this // use to fetch the same value with no reload, so we can safely propagate that // conservative VN to this use. This can help range check elimination later on. cse->gtVNPair.SetConservative(defConservativeVN); // If the old VN was flagged as a checked bound, propagate that to the new VN // to make sure assertion prop will pay attention to this VN. ValueNumStore* vnStore = m_pCompiler->vnStore; ValueNum oldVN = exp->gtVNPair.GetConservative(); if (!vnStore->IsVNConstant(defConservativeVN) && vnStore->IsVNCheckedBound(oldVN)) { vnStore->SetVNIsCheckedBound(defConservativeVN); } GenTree* cmp; if ((m_pCompiler->optCseCheckedBoundMap != nullptr) && (m_pCompiler->optCseCheckedBoundMap->Lookup(exp, &cmp))) { // Propagate the new value number to this compare node as well, since // subsequent range check elimination will try to correlate it with // the other appearances that are getting CSEd. ValueNum oldCmpVN = cmp->gtVNPair.GetConservative(); ValueNum newCmpArgVN; ValueNumStore::CompareCheckedBoundArithInfo info; if (vnStore->IsVNCompareCheckedBound(oldCmpVN)) { // Comparison is against the bound directly. newCmpArgVN = defConservativeVN; vnStore->GetCompareCheckedBound(oldCmpVN, &info); } else { // Comparison is against the bound +/- some offset. assert(vnStore->IsVNCompareCheckedBoundArith(oldCmpVN)); vnStore->GetCompareCheckedBoundArithInfo(oldCmpVN, &info); newCmpArgVN = vnStore->VNForFunc(vnStore->TypeOfVN(info.arrOp), (VNFunc)info.arrOper, info.arrOp, defConservativeVN); } ValueNum newCmpVN = vnStore->VNForFunc(vnStore->TypeOfVN(oldCmpVN), (VNFunc)info.cmpOper, info.cmpOp, newCmpArgVN); cmp->gtVNPair.SetConservative(newCmpVN); } } #ifdef DEBUG cse->gtDebugFlags |= GTF_DEBUG_VAR_CSE_REF; #endif // DEBUG // Now we need to unmark any nested CSE's uses that are found in 'exp' // As well we extract any nested CSE defs that are found in 'exp' and // these are appended to the sideEffList // Afterwards the set of nodes in the 'sideEffectList' are preserved and // all other nodes are removed and have their ref counts decremented // exp->gtCSEnum = NO_CSE; // clear the gtCSEnum field GenTree* sideEffList = nullptr; m_pCompiler->gtExtractSideEffList(exp, &sideEffList, GTF_PERSISTENT_SIDE_EFFECTS | GTF_IS_IN_CSE); // If we have any side effects or extracted CSE defs then we need to create a GT_COMMA tree instead // if (sideEffList != nullptr) { #ifdef DEBUG if (m_pCompiler->verbose) { printf("\nThis CSE use has side effects and/or nested CSE defs. The sideEffectList:\n"); m_pCompiler->gtDispTree(sideEffList); printf("\n"); } #endif GenTree* cseVal = cse; GenTree* curSideEff = sideEffList; ValueNumStore* vnStore = m_pCompiler->vnStore; ValueNumPair exceptions_vnp = ValueNumStore::VNPForEmptyExcSet(); while ((curSideEff->OperGet() == GT_COMMA) || (curSideEff->OperGet() == GT_ASG)) { GenTree* op1 = curSideEff->gtOp.gtOp1; GenTree* op2 = curSideEff->gtOp.gtOp2; ValueNumPair op1vnp; ValueNumPair op1Xvnp = ValueNumStore::VNPForEmptyExcSet(); vnStore->VNPUnpackExc(op1->gtVNPair, &op1vnp, &op1Xvnp); exceptions_vnp = vnStore->VNPExcSetUnion(exceptions_vnp, op1Xvnp); curSideEff = op2; } // We may have inserted a narrowing cast during a previous remorph // and it will not have a value number. if ((curSideEff->OperGet() == GT_CAST) && !curSideEff->gtVNPair.BothDefined()) { // The inserted cast will have no exceptional effects assert(curSideEff->gtOverflow() == false); // Process the exception effects from the cast's operand. curSideEff = curSideEff->gtOp.gtOp1; } ValueNumPair op2vnp; ValueNumPair op2Xvnp = ValueNumStore::VNPForEmptyExcSet(); vnStore->VNPUnpackExc(curSideEff->gtVNPair, &op2vnp, &op2Xvnp); exceptions_vnp = vnStore->VNPExcSetUnion(exceptions_vnp, op2Xvnp); op2Xvnp = ValueNumStore::VNPForEmptyExcSet(); vnStore->VNPUnpackExc(cseVal->gtVNPair, &op2vnp, &op2Xvnp); exceptions_vnp = vnStore->VNPExcSetUnion(exceptions_vnp, op2Xvnp); // Create a comma node with the sideEffList as op1 cse = m_pCompiler->gtNewOperNode(GT_COMMA, expTyp, sideEffList, cseVal); cse->gtVNPair = vnStore->VNPWithExc(op2vnp, exceptions_vnp); } } else { /* This is a def of the CSE */ isDef = true; #ifdef DEBUG if (m_pCompiler->verbose) { printf("\nCSE #%02u def at ", GET_CSE_INDEX(exp->gtCSEnum)); Compiler::printTreeID(exp); printf(" replaced in " FMT_BB " with def of V%02u\n", blk->bbNum, cseLclVarNum); } #endif // DEBUG exp->gtCSEnum = NO_CSE; // clear the gtCSEnum field GenTree* val = exp; /* Create an assignment of the value to the temp */ GenTree* asg = m_pCompiler->gtNewTempAssign(cseLclVarNum, val); // assign the proper Value Numbers asg->gtVNPair.SetBoth(ValueNumStore::VNForVoid()); // The GT_ASG node itself is $VN.Void asg->gtOp.gtOp1->gtVNPair = val->gtVNPair; // The dest op is the same as 'val' noway_assert(asg->gtOp.gtOp1->gtOper == GT_LCL_VAR); noway_assert(asg->gtOp.gtOp2 == val); /* Create a reference to the CSE temp */ GenTree* ref = m_pCompiler->gtNewLclvNode(cseLclVarNum, cseLclVarTyp); ref->gtVNPair = val->gtVNPair; // The new 'ref' is the same as 'val' // If it has a zero-offset field seq, copy annotation to the ref if (hasZeroMapAnnotation) { m_pCompiler->GetZeroOffsetFieldMap()->Set(ref, fldSeq); } /* Create a comma node for the CSE assignment */ cse = m_pCompiler->gtNewOperNode(GT_COMMA, expTyp, asg, ref); cse->gtVNPair = ref->gtVNPair; // The comma's value is the same as 'val' // as the assignment to the CSE LclVar // cannot add any new exceptions } // Walk the statement 'stm' and find the pointer // in the tree is pointing to 'exp' // GenTree** link = m_pCompiler->gtFindLink(stm, exp); #ifdef DEBUG if (link == nullptr) { printf("\ngtFindLink failed: stm="); Compiler::printTreeID(stm); printf(", exp="); Compiler::printTreeID(exp); printf("\n"); printf("stm ="); m_pCompiler->gtDispTree(stm); printf("\n"); printf("exp ="); m_pCompiler->gtDispTree(exp); printf("\n"); } #endif // DEBUG noway_assert(link); // Mutate this link, thus replacing the old exp with the new cse representation // *link = cse; // If it has a zero-offset field seq, copy annotation. if (hasZeroMapAnnotation) { m_pCompiler->GetZeroOffsetFieldMap()->Set(cse, fldSeq); } assert(m_pCompiler->fgRemoveRestOfBlock == false); /* re-morph the statement */ m_pCompiler->fgMorphBlockStmt(blk, stm->AsStmt() DEBUGARG("optValnumCSE")); } while (lst != nullptr); } // Consider each of the CSE candidates and if the CSE passes // the PromotionCheck then transform the CSE by calling PerformCSE // void ConsiderCandidates() { /* Consider each CSE candidate, in order of decreasing cost */ unsigned cnt = m_pCompiler->optCSECandidateCount; Compiler::CSEdsc** ptr = sortTab; for (; (cnt > 0); cnt--, ptr++) { Compiler::CSEdsc* dsc = *ptr; CSE_Candidate candidate(this, dsc); candidate.InitializeCounts(); if (candidate.UseCount() == 0) { #ifdef DEBUG if (m_pCompiler->verbose) { printf("Skipped CSE #%02u because use count is 0\n", candidate.CseIndex()); } #endif continue; } #ifdef DEBUG if (m_pCompiler->verbose) { printf("\nConsidering CSE #%02u [def=%2u, use=%2u, cost=%2u] CSE Expression:\n", candidate.CseIndex(), candidate.DefCount(), candidate.UseCount(), candidate.Cost()); m_pCompiler->gtDispTree(candidate.Expr()); printf("\n"); } #endif if ((dsc->csdDefCount <= 0) || (dsc->csdUseCount == 0)) { // If we reach this point, then the CSE def was incorrectly marked or the // block with this use is unreachable. So skip and go to the next CSE. // Without the "continue", we'd generate bad code in retail. // Commented out a noway_assert(false) here due to bug: 3290124. // The problem is if there is sub-graph that is not reachable from the // entry point, the CSE flags propagated, would be incorrect for it. continue; } bool doCSE = PromotionCheck(&candidate); #ifdef DEBUG if (m_pCompiler->verbose) { if (doCSE) { printf("\nPromoting CSE:\n"); } else { printf("Did Not promote this CSE\n"); } } #endif // DEBUG if (doCSE) { PerformCSE(&candidate); } } } // Perform the necessary cleanup after our CSE heuristics have run // void Cleanup() { // Nothing to do, currently. } }; /***************************************************************************** * * Routine for performing the Value Number based CSE using our heuristics */ void Compiler::optValnumCSE_Heuristic() { #ifdef DEBUG if (verbose) { printf("\n************ Trees at start of optValnumCSE_Heuristic()\n"); fgDumpTrees(fgFirstBB, nullptr); printf("\n"); } #endif // DEBUG CSE_Heuristic cse_heuristic(this); cse_heuristic.Initialize(); cse_heuristic.SortCandidates(); cse_heuristic.ConsiderCandidates(); cse_heuristic.Cleanup(); } /***************************************************************************** * * Perform common sub-expression elimination. */ void Compiler::optOptimizeValnumCSEs() { #ifdef DEBUG if (verbose) { printf("\n*************** In optOptimizeValnumCSEs()\n"); } if (optConfigDisableCSE()) { return; // Disabled by JitNoCSE } #endif optValnumCSE_phase = true; /* Initialize the expression tracking logic */ optValnumCSE_Init(); /* Locate interesting expressions and assign indices to them */ if (optValnumCSE_Locate() > 0) { optCSECandidateTotal += optCSECandidateCount; optValnumCSE_InitDataFlow(); optValnumCSE_DataFlow(); optValnumCSE_Availablity(); optValnumCSE_Heuristic(); } optValnumCSE_phase = false; } #endif // FEATURE_VALNUM_CSE /***************************************************************************** * * The following determines whether the given expression is a worthy CSE * candidate. */ bool Compiler::optIsCSEcandidate(GenTree* tree) { /* No good if the expression contains side effects or if it was marked as DONT CSE */ if (tree->gtFlags & (GTF_ASG | GTF_DONT_CSE)) { return false; } /* The only reason a TYP_STRUCT tree might occur is as an argument to GT_ADDR. It will never be actually materialized. So ignore them. Also TYP_VOIDs */ var_types type = tree->TypeGet(); genTreeOps oper = tree->OperGet(); // TODO-1stClassStructs: Enable CSE for struct types (depends on either transforming // to use regular assignments, or handling copyObj. if (varTypeIsStruct(type) || type == TYP_VOID) { return false; } #ifdef _TARGET_X86_ if (type == TYP_FLOAT) { // TODO-X86-CQ: Revisit this // Don't CSE a TYP_FLOAT on x86 as we currently can only enregister doubles return false; } #else if (oper == GT_CNS_DBL) { // TODO-CQ: Revisit this // Don't try to CSE a GT_CNS_DBL as they can represent both float and doubles return false; } #endif unsigned cost; if (compCodeOpt() == SMALL_CODE) { cost = tree->gtCostSz; } else { cost = tree->gtCostEx; } /* Don't bother if the potential savings are very low */ if (cost < MIN_CSE_COST) { return false; } #if !CSE_CONSTS /* Don't bother with constants */ if (tree->OperKind() & GTK_CONST) return false; #endif /* Check for some special cases */ switch (oper) { case GT_CALL: // If we have a simple helper call with no other persistent side-effects // then we allow this tree to be a CSE candidate // if (gtTreeHasSideEffects(tree, GTF_PERSISTENT_SIDE_EFFECTS | GTF_IS_IN_CSE) == false) { return true; } else { // Calls generally cannot be CSE-ed return false; } case GT_IND: // TODO-CQ: Review this... /* We try to cse GT_ARR_ELEM nodes instead of GT_IND(GT_ARR_ELEM). Doing the first allows cse to also kick in for code like "GT_IND(GT_ARR_ELEM) = GT_IND(GT_ARR_ELEM) + xyz", whereas doing the second would not allow it */ return (tree->gtOp.gtOp1->gtOper != GT_ARR_ELEM); case GT_CNS_INT: case GT_CNS_LNG: case GT_CNS_DBL: case GT_CNS_STR: return true; // We reach here only when CSE_CONSTS is enabled case GT_ARR_ELEM: case GT_ARR_LENGTH: case GT_CLS_VAR: case GT_LCL_FLD: return true; case GT_LCL_VAR: return false; // Can't CSE a volatile LCL_VAR case GT_NEG: case GT_NOT: case GT_CAST: return true; // CSE these Unary Operators case GT_SUB: case GT_DIV: case GT_MOD: case GT_UDIV: case GT_UMOD: case GT_OR: case GT_AND: case GT_XOR: case GT_RSH: case GT_RSZ: case GT_ROL: case GT_ROR: return true; // CSE these Binary Operators case GT_ADD: // Check for ADDRMODE flag on these Binary Operators case GT_MUL: case GT_LSH: if ((tree->gtFlags & GTF_ADDRMODE_NO_CSE) != 0) { return false; } case GT_EQ: case GT_NE: case GT_LT: case GT_LE: case GT_GE: case GT_GT: return true; // Also CSE these Comparison Operators case GT_INTRINSIC: return true; // Intrinsics case GT_COMMA: return true; // Allow GT_COMMA nodes to be CSE-ed. case GT_COLON: case GT_QMARK: case GT_NOP: case GT_RETURN: return false; // Currently the only special nodes that we hit // that we know that we don't want to CSE default: break; // Any new nodes that we might add later... } return false; } #ifdef DEBUG // // A Debug only method that allows you to control whether the CSE logic is enabled for this method. // // If this method returns false then the CSE phase should be performed. // If the method returns true then the CSE phase should be skipped. // bool Compiler::optConfigDisableCSE() { // Next check if COMPlus_JitNoCSE is set and applies to this method // unsigned jitNoCSE = JitConfig.JitNoCSE(); if (jitNoCSE > 0) { unsigned methodCount = Compiler::jitTotalMethodCompiled; if ((jitNoCSE & 0xF000000) == 0xF000000) { unsigned methodCountMask = methodCount & 0xFFF; unsigned bitsZero = (jitNoCSE >> 12) & 0xFFF; unsigned bitsOne = (jitNoCSE >> 0) & 0xFFF; if (((methodCountMask & bitsOne) == bitsOne) && ((~methodCountMask & bitsZero) == bitsZero)) { if (verbose) { printf(" Disabled by JitNoCSE methodCountMask\n"); } return true; // The CSE phase for this method is disabled } } else if (jitNoCSE <= (methodCount + 1)) { if (verbose) { printf(" Disabled by JitNoCSE > methodCount\n"); } return true; // The CSE phase for this method is disabled } } return false; } // // A Debug only method that allows you to control whether the CSE logic is enabled for // a particular CSE in a method // // If this method returns false then the CSE should be performed. // If the method returns true then the CSE should be skipped. // bool Compiler::optConfigDisableCSE2() { static unsigned totalCSEcount = 0; unsigned jitNoCSE2 = JitConfig.JitNoCSE2(); totalCSEcount++; if (jitNoCSE2 > 0) { if ((jitNoCSE2 & 0xF000000) == 0xF000000) { unsigned totalCSEMask = totalCSEcount & 0xFFF; unsigned bitsZero = (jitNoCSE2 >> 12) & 0xFFF; unsigned bitsOne = (jitNoCSE2 >> 0) & 0xFFF; if (((totalCSEMask & bitsOne) == bitsOne) && ((~totalCSEMask & bitsZero) == bitsZero)) { if (verbose) { printf(" Disabled by jitNoCSE2 Ones/Zeros mask\n"); } return true; } } else if ((jitNoCSE2 & 0xF000000) == 0xE000000) { unsigned totalCSEMask = totalCSEcount & 0xFFF; unsigned disableMask = jitNoCSE2 & 0xFFF; disableMask >>= (totalCSEMask % 12); if (disableMask & 1) { if (verbose) { printf(" Disabled by jitNoCSE2 rotating disable mask\n"); } return true; } } else if (jitNoCSE2 <= totalCSEcount) { if (verbose) { printf(" Disabled by jitNoCSE2 > totalCSEcount\n"); } return true; } } return false; } #endif void Compiler::optOptimizeCSEs() { #ifdef DEBUG if (verbose) { printf("\n*************** In optOptimizeCSEs()\n"); printf("Blocks/Trees at start of optOptimizeCSE phase\n"); fgDispBasicBlocks(true); } #endif // DEBUG optCSECandidateCount = 0; optCSEstart = lvaCount; #if FEATURE_VALNUM_CSE INDEBUG(optEnsureClearCSEInfo()); optOptimizeValnumCSEs(); EndPhase(PHASE_OPTIMIZE_VALNUM_CSES); #endif // FEATURE_VALNUM_CSE } /***************************************************************************** * * Cleanup after CSE to allow us to run more than once. */ void Compiler::optCleanupCSEs() { // We must clear the BBF_VISITED and BBF_MARKED flags // for (BasicBlock* block = fgFirstBB; block; block = block->bbNext) { // And clear all the "visited" bits on the block // block->bbFlags &= ~(BBF_VISITED | BBF_MARKED); /* Walk the statement trees in this basic block */ GenTree* stmt; // Initialize 'stmt' to the first non-Phi statement stmt = block->FirstNonPhiDef(); for (; stmt; stmt = stmt->gtNext) { noway_assert(stmt->gtOper == GT_STMT); /* We must clear the gtCSEnum field */ for (GenTree* tree = stmt->gtStmt.gtStmtExpr; tree; tree = tree->gtPrev) { tree->gtCSEnum = NO_CSE; } } } } #ifdef DEBUG /***************************************************************************** * * Ensure that all the CSE information in the IR is initialized the way we expect it, * before running a CSE phase. This is basically an assert that optCleanupCSEs() is not needed. */ void Compiler::optEnsureClearCSEInfo() { for (BasicBlock* block = fgFirstBB; block; block = block->bbNext) { assert((block->bbFlags & (BBF_VISITED | BBF_MARKED)) == 0); /* Walk the statement trees in this basic block */ GenTree* stmt; // Initialize 'stmt' to the first non-Phi statement stmt = block->FirstNonPhiDef(); for (; stmt; stmt = stmt->gtNext) { assert(stmt->gtOper == GT_STMT); for (GenTree* tree = stmt->gtStmt.gtStmtExpr; tree; tree = tree->gtPrev) { assert(tree->gtCSEnum == NO_CSE); } } } } #endif // DEBUG /*****************************************************************************/ #endif // FEATURE_ANYCSE /*****************************************************************************/