SMPFunction.cpp

					|| (this->LocalVarTable[i].offset < (long) this->OutgoingArgsSize))
					continue;
				qfprintf(AnnotFile, "%10x %6d DATAREF STACK %d esp + %d CHILDOF %d OFFSET %d LOCALVAR %s \n",
					addr, this->LocalVarTable[i].size, DataReferentID,
					this->LocalVarTable[i].offset, ParentReferentID,
					this->LocalVarTable[i].offset, this->LocalVarTable[i].VarName);
				++DataReferentID;
			}
		} // end if (this->AnalyzedSP and not Alloca .... )
#endif
	} // end if (0 < LocalVarsSize)
	return;
} // end of SMPFunction::EmitStackFrameAnnotations() 

// Main data flow analysis driver. Goes through the function and
//  fills all objects for instructions, basic blocks, and the function
//  itself.
void SMPFunction::Analyze(void) {
	list<SMPInstr>::iterator FirstInBlock = this->Instrs.end();
	   // For starting a basic block
	list<SMPInstr>::iterator LastInBlock = this->Instrs.end();
	   // Terminating a basic block

#if SMP_DEBUG_CONTROLFLOW
	msg("Entering SMPFunction::Analyze.\n");
#endif

	// Get some basic info from the FuncInfo structure.
	this->Size = this->FuncInfo.endEA - this->FuncInfo.startEA;
	this->UseFP = (0 != (this->FuncInfo.flags & (FUNC_FRAME | FUNC_BOTTOMBP)));
	this->StaticFunc = (0 != (this->FuncInfo.flags & FUNC_STATIC));
	get_func_name(this->FuncInfo.startEA, this->FuncName,
		sizeof(this->FuncName) - 1);
	this->BlockCount = 0;
	this->AnalyzedSP = this->FuncInfo.analyzed_sp();

#if SMP_DEBUG_CONTROLFLOW
	msg("SMPFunction::Analyze: got basic info.\n");
#endif

	// Cycle through all chunks that belong to the function.
	func_tail_iterator_t FuncTail(&(this->FuncInfo));
	size_t ChunkCounter = 0;
	for (bool ChunkOK = FuncTail.main(); ChunkOK; ChunkOK = FuncTail.next()) {
		const area_t &CurrChunk = FuncTail.chunk();
		++ChunkCounter;
		if (1 < ChunkCounter) {
			this->SharedChunks = true;
#if SMP_DEBUG_CHUNKS
			msg("Found tail chunk for %s at %x\n", this->FuncName, CurrChunk.startEA);
#endif
		}
		// Build the instruction and block lists for the function.
		for (ea_t addr = CurrChunk.startEA; addr < CurrChunk.endEA;
			addr = get_item_end(addr)) {
			flags_t InstrFlags = getFlags(addr);
			if (isHead(InstrFlags) && isCode(InstrFlags)) {
				SMPInstr CurrInst = SMPInstr(addr);
				// Fill in the instruction data members.
#if SMP_DEBUG_CONTROLFLOW
				msg("SMPFunction::Analyze: calling CurrInst::Analyze.\n");
#endif
				CurrInst.Analyze();
				if (SMPBinaryDebug) {
					msg("Disasm:  %s \n", CurrInst.GetDisasm());
				}
				if (this->AnalyzedSP) {
					// Audit the IDA SP analysis.
					sval_t sp_delta = get_spd(&(this->FuncInfo), addr);
					// sp_delta is difference between current value of stack pointer
					//  and value of the stack pointer coming into the function. It
					//  is updated AFTER each instruction. Thus, it should not get back
					//  above zero (e.g. to +4) until after a return instruction.
					if (sp_delta > 0) {
						// Stack pointer has underflowed, according to IDA's analysis,
						//  which is probably incorrect.
						this->AnalyzedSP = false;
						msg("Resetting AnalyzedSP to false for %s\n", this->GetFuncName());
						msg("Underflowing instruction: %s sp_delta: %d\n", CurrInst.GetDisasm(),
							sp_delta);
					}
				}
				if (CurrInst.GetDataFlowType() == INDIR_CALL)
					this->IndirectCalls = true;
				else if (CurrInst.GetDataFlowType() == CALL) {
					for (size_t i = 0; i < CurrInst.NumUses(); ++i) {
						optype_t OpType = CurrInst.GetUse(i).GetOp().type;
						if ((OpType == o_near) || (OpType == o_far)) {
							ea_t CallTarget = CurrInst.GetUse(i).GetOp().addr;
							this->DirectCallTargets.push_back(CallTarget);
						}
					}
				}

				// Before we insert the instruction into the instruction
				//  list, determine if it is a jump target that does not
				//  follow a basic block terminator. This is the special case
				//  of a CASE in a SWITCH that falls through into another
				//  CASE, for example. The first sequence of statements
				//  was not terminated by a C "break;" statement, so it
				//  looks like straight line code, but there is an entry
				//  point at the beginning of the second CASE sequence and
				//  we have to split basic blocks at the entry point.
				if ((FirstInBlock != this->Instrs.end())
					&& CurrInst.IsJumpTarget()) {
#if SMP_DEBUG_CONTROLFLOW
					msg("SMPFunction::Analyze: hit special jump target case.\n");
#endif
					LastInBlock = --(this->Instrs.end());
					SMPBasicBlock CurrBlock = SMPBasicBlock(this, FirstInBlock,
						LastInBlock);
					CurrBlock.Analyze();
					// If not the first chunk in the function, it is a shared
					//  tail chunk.
					if (ChunkCounter > 1) {
						CurrBlock.SetShared();
					}
					FirstInBlock = this->Instrs.end();
					LastInBlock = this->Instrs.end();
					this->Blocks.push_back(CurrBlock);
					this->BlockCount += 1;
				}

#if SMP_DEBUG_CONTROLFLOW
		msg("SMPFunction::Analyze: putting CurrInst on list.\n");
#endif
				// Insert instruction at end of list.
				this->Instrs.push_back(CurrInst);

				// Find basic block leaders and terminators.
				if (FirstInBlock == this->Instrs.end()) {
#if SMP_DEBUG_CONTROLFLOW
		msg("SMPFunction::Analyze: setting FirstInBlock.\n");
#endif
					FirstInBlock = --(this->Instrs.end());
				}
				if (CurrInst.IsBasicBlockTerminator()) {
#if SMP_DEBUG_CONTROLFLOW
		msg("SMPFunction::Analyze: found block terminator.\n");
#endif
					LastInBlock = --(this->Instrs.end());
					SMPBasicBlock CurrBlock = SMPBasicBlock(this, FirstInBlock, LastInBlock);
					CurrBlock.Analyze();
					// If not the first chunk in the function, it is a shared
					//  tail chunk.
					if (ChunkCounter > 1) {
						CurrBlock.SetShared();
					}
					FirstInBlock = this->Instrs.end();
					LastInBlock = this->Instrs.end();
					this->Blocks.push_back(CurrBlock);
					this->BlockCount += 1;

					// Is the instruction a branch to a target outside the function? If
					//  so, this function has shared tail chunks.
					if (CurrInst.IsBranchToFarChunk()) {
						this->SharedChunks = true;
					}
				}
			} // end if (isHead(InstrFlags) && isCode(InstrFlags)
		} // end for (ea_t addr = FuncInfo.startEA; ... )

		// Handle the special case in which a function does not terminate
		//  with a return instruction or any other basic block terminator.
		//  Sometimes IDA Pro sees a call to a NORET function and decides
		//  to not include the dead code after it in the function. That
		//  dead code includes the return instruction, so the function no
		//  longer includes a return instruction and terminates with a CALL.
		if (FirstInBlock != this->Instrs.end()) {
			LastInBlock = --(this->Instrs.end());
			SMPBasicBlock CurrBlock = SMPBasicBlock(this, FirstInBlock, LastInBlock);
			CurrBlock.Analyze();
			// If not the first chunk in the function, it is a shared
			//  tail chunk.
			if (ChunkCounter > 1) {
				CurrBlock.SetShared();
			}
			FirstInBlock = this->Instrs.end();
			LastInBlock = this->Instrs.end();
			this->Blocks.push_back(CurrBlock);
			this->BlockCount += 1;
		}
	} // end for (bool ChunkOK = ...)

#if KLUDGE_VFPRINTF_FAMILY
	if (0 != strstr(this->GetFuncName(), "printf")) {
		this->SharedChunks = true;
		msg("Kludging function %s\n", this->GetFuncName());
	}
#endif

	// Set up basic block links and map of instructions to blocks.
	if (!(this->HasSharedChunks())) {
		bool DumpFlag = false;
#if SMP_DEBUG_DATAFLOW
		DumpFlag |= (0 == strcmp("main", this->GetFuncName()));
		DumpFlag |= (0 == strcmp("dohanoi", this->GetFuncName()));
		DumpFlag |= (0 == strcmp("simtest", this->GetFuncName()));
#endif
		this->SetLinks();
#if SMP_COMPUTE_LVA_SSA
		list<SMPInstr>::iterator CurrInst;
		bool GoodRTL;
		for (CurrInst = this->Instrs.begin(); CurrInst != this->Instrs.end(); ++CurrInst) {
			// Build tree RTLs for the instruction.
			GoodRTL = CurrInst->BuildRTL();
#if SMP_DEBUG_BUILD_RTL
			if (!GoodRTL) {
				msg("ERROR: Cannot build RTL at %x for %s\n", CurrInst->GetAddr(), CurrInst->GetDisasm());
			}
#if 0
			else {
				msg("Built RTL at %x for %s\n", CurrInst->GetAddr(), CurrInst->GetDisasm());
			}
#endif
#endif
		}
		this->RPONumberBlocks();
		this->LiveVariableAnalysis();
		this->ComputeSSA();
		if (DumpFlag) 
			this->Dump();
#endif
	}

#if SMP_DEBUG_CONTROLFLOW
	msg("SMPFunction::Analyze: set stack frame info.\n");
#endif
	// Figure out the stack frame and related info.
	this->SetStackFrameInfo();

	return;
} // end of SMPFunction::Analyze()

// Compute SSA form data structures across the function.
void SMPFunction::ComputeSSA(void) {
#if SMP_DEBUG_DATAFLOW
	bool DumpFlag = (0 == strcmp("main", this->GetFuncName()));
	DumpFlag |= (0 == strcmp("dohanoi", this->GetFuncName()));
	DumpFlag |= (0 == strcmp("_init_proc", this->GetFuncName()));
#endif

#if 0
	if (DumpFlag)
		this->Dump();
#endif
	this->ComputeIDoms();
	this->ComputeDomFrontiers();
	this->ComputeGlobalNames();
	this->ComputeBlocksDefinedIn();

	this->InsertPhiFunctions();
	this->BuildDominatorTree();
	this->SSARenumber();
	list<SMPBasicBlock>::iterator CurrBlock;
	for (CurrBlock = this->Blocks.begin(); CurrBlock != this->Blocks.end(); ++CurrBlock) {
		CurrBlock->SetLocalNames();
		CurrBlock->SSALocalRenumber();
#if 1
		CurrBlock->MarkDeadRegs();
#endif
	}
#if SMP_DEBUG_DATAFLOW
	if (DumpFlag)
		this->Dump();
#endif
	return;
} // end of SMPFunction::ComputeSSA()

// Link basic blocks to their predecessors and successors, and build the map
//  of instruction addresses to basic blocks.
void SMPFunction::SetLinks(void) {
	list<SMPBasicBlock>::iterator CurrBlock;
#if SMP_DEBUG_DATAFLOW
	msg("SetLinks called for %s\n", this->GetFuncName());
#endif
	// First, set up the map of instructions to basic blocks.
	for (CurrBlock = this->Blocks.begin(); CurrBlock != this->Blocks.end(); ++CurrBlock) {
		list<list<SMPInstr>::iterator>::iterator CurrInst;
		for (CurrInst = CurrBlock->GetFirstInstr();
			CurrInst != CurrBlock->GetLastInstr();
			++CurrInst) {
				pair<ea_t, list<SMPBasicBlock>::iterator> MapItem((*CurrInst)->GetAddr(),CurrBlock);
				InstBlockMap.insert(MapItem);
		}
	}

#if SMP_DEBUG_DATAFLOW
	msg("SetLinks finished mapping: %s\n", this->GetFuncName());
#endif
	// Next, set successors of each basic block, also setting up the predecessors in the
	//  process.
	for (CurrBlock = this->Blocks.begin(); CurrBlock != this->Blocks.end(); ++CurrBlock) {
		list<SMPInstr>::iterator CurrInst = *(--(CurrBlock->GetLastInstr()));
		// Last instruction in block; set successors
		bool CallFlag = (CALL == CurrInst->GetDataFlowType());
		xrefblk_t CurrXrefs;
		for (bool ok = CurrXrefs.first_from(CurrInst->GetAddr(), XREF_ALL);
			ok;
			ok = CurrXrefs.next_from()) {
				if ((CurrXrefs.to != 0) && (CurrXrefs.iscode)) {
					// Found a code target, with its address in CurrXrefs.to
					if (CallFlag && (CurrXrefs.to != (CurrInst->GetAddr() + CurrInst->GetCmd().size))) {
						// A call instruction will have two targets: the fall through to the
						//  next instruction, and the called function. We want to link to the
						//  fall-through instruction, but not to the called function.
						// Some blocks end with a call just because the fall-through instruction
						//  is a jump target from elsewhere.
						continue;
					}
					map<ea_t, list<SMPBasicBlock>::iterator>::iterator MapEntry;
					MapEntry = this->InstBlockMap.find(CurrXrefs.to);
					if (MapEntry == this->InstBlockMap.end()) {
						msg("WARNING: addr %x not found in map for %s\n", CurrXrefs.to,
							this->GetFuncName());
						msg(" Referenced from %s\n", CurrInst->GetDisasm());
					}
					else {
						list<SMPBasicBlock>::iterator Target = MapEntry->second;
						// Make target block a successor of current block.
						CurrBlock->LinkToSucc(Target);
						// Make current block a predecessor of target block.
						Target->LinkToPred(CurrBlock);
					}
				}
		} // end for all xrefs
	} // end for all blocks

	// If we have any blocks that are all no-ops and have no predecessors, remove those
	//  blocks. They are dead and make the CFG no longer a lattice. Any blocks that have
	//  no predecessors but are not all no-ops should also be removed with a different
	//  log message.
	CurrBlock = this->Blocks.begin();
	++CurrBlock; // don't delete the top block, no matter what.
	while (CurrBlock != this->Blocks.end()) {
		if (CurrBlock->GetFirstPred() == CurrBlock->GetLastPred()) {
			if (CurrBlock->AllNops())
				msg("Removing all nops block at %x\n", CurrBlock->GetFirstAddr());
			else
				msg("Removing block with no predecessors at %x\n", CurrBlock->GetFirstAddr());
			// Remove this block from the predecessors list of its successors.
			list<list<SMPBasicBlock>::iterator>::iterator SuccIter;
			ea_t TempAddr = CurrBlock->GetFirstAddr();
			for (SuccIter = CurrBlock->GetFirstSucc(); SuccIter != CurrBlock->GetLastSucc(); ++SuccIter) {
				(*SuccIter)->ErasePred(TempAddr);
			}
			// Finally, remove the block from the blocks list.
			CurrBlock = this->Blocks.erase(CurrBlock);
			this->BlockCount -= 1;
		}
		else
			++CurrBlock;
	}

	return;
} // end of SMPFunction::SetLinks()

// Number all basic blocks in reverse postorder (RPO) and set RPOBlocks vector to
//  access them.
void SMPFunction::RPONumberBlocks(void) {
	bool DebugFlag = false;
#if SMP_DEBUG_DATAFLOW
	DebugFlag = (0 == strcmp("__ieee754_pow", this->GetFuncName()));
	if (DebugFlag) msg("Entered RPONumberBlocks\n");
#endif
	int CurrNum = 0;
	list<list<SMPBasicBlock>::iterator> WorkList;

	// Number the first block with 0.
	list<SMPBasicBlock>::iterator CurrBlock = this->Blocks.begin();
#if 0
	if (this->RPOBlocks.capacity() <= (size_t) this->BlockCount) {
		msg("Reserving %d RPOBlocks old value: %d\n", 2+this->BlockCount, this->RPOBlocks.capacity());
		this->RPOBlocks.reserve(2 + this->BlockCount);
		this->RPOBlocks.assign(2 + this->BlockCount, this->Blocks.end());
	}
#endif
	CurrBlock->SetNumber(CurrNum);
	this->RPOBlocks.push_back(CurrBlock);
	++CurrNum;
	// Push the first block's successors onto the work list.
	list<list<SMPBasicBlock>::iterator>::iterator CurrSucc = CurrBlock->GetFirstSucc();
	while (CurrSucc != CurrBlock->GetLastSucc()) {
		WorkList.push_back(*CurrSucc);
		++CurrSucc;
	}

	// Use the WorkList to iterate through all blocks in the function
	list<list<SMPBasicBlock>::iterator>::iterator CurrListItem = WorkList.begin();
	bool change;
	while (!WorkList.empty()) {
		change = false;
		while (CurrListItem != WorkList.end()) {
			if ((*CurrListItem)->GetNumber() != SMP_BLOCKNUM_UNINIT) {
				// Duplicates get pushed onto the WorkList because a block
				//  can be the successor of multiple other blocks. If it is
				//  already numbered, it is a duplicate and can be removed
				//  from the list.
				CurrListItem = WorkList.erase(CurrListItem);
				change = true;
				continue;
			}
			if ((*CurrListItem)->AllPredecessorsNumbered()) {
				// Ready to be numbered.
				(*CurrListItem)->SetNumber(CurrNum);
#if 0
				msg("Set RPO number %d\n", CurrNum);
				if (DebugFlag && (7 == CurrNum))
					this->Dump();
#endif
				this->RPOBlocks.push_back(*CurrListItem);
				++CurrNum;
				change = true;
				// Push its unnumbered successors onto the work list.
				CurrSucc = (*CurrListItem)->GetFirstSucc();
				while (CurrSucc != (*CurrListItem)->GetLastSucc()) {
					if ((*CurrSucc)->GetNumber() == SMP_BLOCKNUM_UNINIT)
						WorkList.push_back(*CurrSucc);
					++CurrSucc;
				}
				CurrListItem = WorkList.erase(CurrListItem);
			}
			else {
				++CurrListItem;
			}
		} // end while (CurrListItem != WorkList.end())
		if (change) {
			// Reset CurrListItem to beginning of work list for next iteration.
			CurrListItem = WorkList.begin();
		}
		else {
			// Loops can cause us to not be able to find a WorkList item that has
			//  all predecessors numbered. Take the WorkList item with the lowest address
			//  and number it so we can proceed.
			CurrListItem = WorkList.begin();
			ea_t LowAddr = (*CurrListItem)->GetFirstAddr();
			list<list<SMPBasicBlock>::iterator>::iterator SaveItem = CurrListItem;
			++CurrListItem;
			while (CurrListItem != WorkList.end()) {
				if (LowAddr > (*CurrListItem)->GetFirstAddr()) {
					SaveItem = CurrListItem;
					LowAddr = (*CurrListItem)->GetFirstAddr();
				}
				++CurrListItem;
			}
			// SaveItem should now be numbered.
			(*SaveItem)->SetNumber(CurrNum);
#if SMP_DEBUG_DATAFLOW
			msg("Picked LowAddr %x and set RPO number %d\n", LowAddr, CurrNum);
#endif
			this->RPOBlocks.push_back(*SaveItem);
			++CurrNum;
			// Push its unnumbered successors onto the work list.
			CurrSucc = (*SaveItem)->GetFirstSucc();
			while (CurrSucc != (*SaveItem)->GetLastSucc()) {
				if ((*CurrSucc)->GetNumber() == SMP_BLOCKNUM_UNINIT)
					WorkList.push_back(*CurrSucc);
				++CurrSucc;
			}
			CurrListItem = WorkList.erase(SaveItem);
			CurrListItem = WorkList.begin();
		} // end if (change) ... else ...
	} // end while work list is nonempty
	return;
} // end of SMPFunction::RPONumberBlocks()

// Perform live variable analysis on all blocks in the function.
// See chapter 9 of Cooper/Torczon, Engineering a Compiler, for the algorithm.
void SMPFunction::LiveVariableAnalysis(void) {
	list<SMPBasicBlock>::iterator CurrBlock;
#if SMP_DEBUG_DATAFLOW
	msg("LiveVariableAnalysis for %s\n", this->GetFuncName());
	bool DebugFlag = (0 == strcmp("__ieee754_pow", this->GetFuncName()));
#endif

	for (CurrBlock = this->Blocks.begin(); CurrBlock != this->Blocks.end(); ++CurrBlock) {
		// Initialize the Killed and UpwardExposed sets for each block.
		CurrBlock->InitKilledExposed();
	}

	bool changed;
	// Iterate over each block, updating LiveOut sets until no more changes are made.
	// NOTE: LVA is more efficient when computed over a reverse post-order list of blocks
	//  from the inverted CFG. We have an RPO list from the forward CFG.
#if 1
	do {
		changed = false;
		for (CurrBlock = this->Blocks.begin(); CurrBlock != this->Blocks.end(); ++CurrBlock) {
			changed |= CurrBlock->UpdateLiveOut();
		}
	} while (changed);
#else // Use reverse postorder
	do {
		changed = false;
		for (size_t index = 0; index < this->RPOBlocks.size(); ++index) {
			CurrBlock = this->RPOBlocks[index];
			changed |= CurrBlock->UpdateLiveOut();
		}
	} while (changed);
#endif
#if SMP_DEBUG_DATAFLOW
	if (DebugFlag) msg("Exiting LiveVariableAnalysis\n");
#endif
	return;
} // end of SMPFunction::LiveVariableAnalysis()

// Return the IDom index that is the end of the intersection prefix of the Dom sets of
//  the two blocks designated by the RPO numbers passed in.
// See Cooper & Torczon, "Engineering a Compiler" 1st edition figure 9.8.
int SMPFunction::IntersectDoms(int block1, int block2) const {
	int finger1 = block1;
	int finger2 = block2;
	while (finger1 != finger2) {
		while (finger1 > finger2)
			finger1 = this->IDom.at(finger1);
		while (finger2 > finger1)
			finger2 = this->IDom.at(finger2);
	}
	return finger1;
} // end of SMPFunction::IntersectDoms()

// Compute immediate dominators of all blocks into IDom[] vector.
void SMPFunction::ComputeIDoms(void) {
	bool DebugFlag = false;
#if SMP_DEBUG_DATAFLOW
	DebugFlag = (0 == strcmp("__ieee754_pow", this->GetFuncName()));
	if (DebugFlag) msg("Entered ComputeIDoms\n");
#endif
	// Initialize the IDom[] vector to uninitialized values for all blocks.
	this->IDom.reserve(this->BlockCount);
	this->IDom.assign(this->BlockCount, SMP_BLOCKNUM_UNINIT);
	if (DebugFlag) msg("BlockCount = %d\n", this->BlockCount);
	this->IDom[0] = 0; // Start block dominated only by itself
	bool changed;
	do {
		changed = false;
		for (size_t RPONum = 1; RPONum < (size_t) this->BlockCount; ++RPONum) {
			if (DebugFlag) msg("RPONum %d\n", RPONum);
			if (DebugFlag) {
				msg("RPOBlocks vector size: %d\n", this->RPOBlocks.size());
				for (size_t index = 0; index < this->RPOBlocks.size(); ++index) {
					msg("RPOBlocks entry %d is %d\n", index, RPOBlocks[index]->GetNumber());
				}
			}
			list<SMPBasicBlock>::iterator CurrBlock = this->RPOBlocks.at(RPONum);
			// if (DebugFlag) msg("CurrBlock: %x\n", CurrBlock._Ptr);
			list<list<SMPBasicBlock>::iterator>::iterator CurrPred;
			// Initialize NewIdom to the first processed predecessor of block RPONum.
			int NewIdom = SMP_BLOCKNUM_UNINIT;
			for (CurrPred = CurrBlock->GetFirstPred(); CurrPred != CurrBlock->GetLastPred(); ++CurrPred) {
				int PredNum = (*CurrPred)->GetNumber();
				if (DebugFlag) msg("Pred: %d\n", PredNum);
				// **!!** See comment below about unreachable blocks.
				if (SMP_BLOCKNUM_UNINIT == PredNum)
					continue;
				int PredIDOM = this->IDom.at(PredNum);
				if (DebugFlag) msg("Pred IDom: %d\n", PredIDOM);
				if (SMP_BLOCKNUM_UNINIT != PredIDOM) {
					NewIdom = PredNum;
					break;
				}
			}
			if (NewIdom == SMP_BLOCKNUM_UNINIT)
				msg("Failure on NewIdom in ComputeIDoms for %s\n", this->GetFuncName());
			assert(NewIdom != SMP_BLOCKNUM_UNINIT);
			// Loop through all predecessors of block RPONum except block NewIdom.
			//  Set NewIdom to the intersection of its Dom set and the Doms set of
			//  each predecessor that has had its Doms set computed.
			for (CurrPred = CurrBlock->GetFirstPred(); CurrPred != CurrBlock->GetLastPred(); ++CurrPred) {
				int PredNum = (*CurrPred)->GetNumber();
				if (DebugFlag) msg("PredNum: %d\n", PredNum);
				// **!!** We could avoid failure on unreachable basic blocks
				//  by executing a continue statement if PredNum is -1. Long term solution
				//  is to prune out unreachable basic blocks.
				if (PredNum == SMP_BLOCKNUM_UNINIT)
					continue;
				int PredIDOM = this->IDom.at(PredNum);
				if (DebugFlag) msg("PredIDOM: %d\n", PredIDOM);
				if ((SMP_BLOCKNUM_UNINIT == PredIDOM) || (NewIdom == PredIDOM)) {
					// Skip predecessors that have uncomputed Dom sets, or are the
					//  current NewIdom.
					continue;
				}
				if (DebugFlag) msg("Old NewIdom value: %d\n", NewIdom);
				NewIdom = this->IntersectDoms(PredNum, NewIdom);
				if (DebugFlag) msg("New NewIdom value: %d\n", NewIdom);
			}
			// If NewIdom is not the value currently in vector IDom[], update the
			//  vector entry and set changed to true.
			if (NewIdom != this->IDom.at(RPONum)) {
				if (DebugFlag) msg("IDOM changed from %d to %d\n", this->IDom.at(RPONum), NewIdom);
				this->IDom[RPONum] = NewIdom;
				changed = true;
			}
		}
	} while (changed);
	return;
} // end of SMPFunction::ComputeIDoms()

// Compute dominance frontier sets for each block.
void SMPFunction::ComputeDomFrontiers(void) {
	list<SMPBasicBlock>::iterator CurrBlock;
	for (CurrBlock = this->Blocks.begin(); CurrBlock != this->Blocks.end(); ++CurrBlock) {
		// We look only at join points in the CFG, as per Cooper/Torczon chapter 9.
		if (1 < CurrBlock->GetNumPreds()) { // join point; more than 1 predecessor
			int runner;
			list<list<SMPBasicBlock>::iterator>::iterator CurrPred;
			for (CurrPred = CurrBlock->GetFirstPred(); CurrPred != CurrBlock->GetLastPred(); ++CurrPred) {
				// For each predecessor, we run up the IDom[] vector and add CurrBlock to the
				//  DomFrontier for all blocks that are between CurrPred and IDom[CurrBlock],
				//  not including IDom[CurrBlock] itself.
				runner = (*CurrPred)->GetNumber();
				while (runner != this->IDom.at(CurrBlock->GetNumber())) {
					// Cooper/Harvey/Kennedy paper does not quite agree with the later
					//  text by Cooper/Torczon. Text says that the start node has no IDom
					//  in the example on pages 462-463, but it shows an IDOM for the
					//  root node in Figure 9.9 of value == itself. The first edition text
					//  on p.463 seems correct, as the start node dominates every node and
					//  thus should have no dominance frontier.
					if (SMP_TOP_BLOCK == runner)
						break;
					(*CurrPred)->AddToDomFrontier(CurrBlock->GetNumber());
					runner = this->IDom.at(runner);
				}
			} // end for all predecessors
		} // end if join point
	} // end for all blocks
	return;
} // end of SMPFunction::ComputeDomFrontiers()

// Compute the GlobalNames set, which includes all operands that are used in more than
//  one basic block. It is the union of all UpExposedSets of all blocks.
void SMPFunction::ComputeGlobalNames(void) {
	set<op_t, LessOp>::iterator SetIter;
	list<SMPBasicBlock>::iterator CurrBlock;
	unsigned int index = 0;
	if (this->Blocks.size() < 2)
		return; // cannot have global names if there is only one block

	bool DebugFlag = false;
#if SMP_DEBUG_DATAFLOW
	DebugFlag = (0 == strcmp("__ieee754_pow", this->GetFuncName()));
#endif

	for (CurrBlock = this->Blocks.begin(); CurrBlock != this->Blocks.end(); ++CurrBlock) {
		for (SetIter = CurrBlock->GetFirstUpExposed(); SetIter != CurrBlock->GetLastUpExposed(); ++SetIter) {
			op_t TempOp = *SetIter;
			// The GlobalNames set will have the complete collection of operands that we are
			//  going to number in our SSA computations. We now assign an operand number
			//  within the op_t structure for each, so that we can index into the
			//  BlocksUsedIn[] vector, for example. This operand number is not to be
			//  confused with SSA numbers.
			// We use the operand number field op_t.n for the lower 8 bits, and the offset
			//  fields op_t.offb:op_t.offo for the upper 16 bits. We are overwriting IDA
			//  values here, but operands in the data flow analysis sets should never be
			//  inserted back into the program anyway.
			SetGlobalIndex(&TempOp, index);

#if SMP_DEBUG_DATAFLOW
			msg("Global Name: ");
			PrintListOperand(TempOp);
#endif
			set<op_t, LessOp>::iterator AlreadyInSet;
			pair<set<op_t, LessOp>::iterator, bool> InsertResult;
			InsertResult = this->GlobalNames.insert(TempOp);
			if (!InsertResult.second) {
				// Already in GlobalNames, so don't assign an index number.
				;
#if SMP_DEBUG_DATAFLOW
				msg(" already in GlobalNames.\n");
#endif
			}
			else {
				++index;
#if SMP_DEBUG_DATAFLOW
				msg(" inserted as index %d\n", ExtractGlobalIndex(TempOp));
#endif
			}
		} // for each upward exposed item in the current block
	} // for each basic block

	assert(16777215 >= this->GlobalNames.size()); // index fits in 24 bits
	return;
} // end of SMPFunction::ComputeGlobalNames()

// For each item in GlobalNames, record the blocks that DEF the item.
void SMPFunction::ComputeBlocksDefinedIn(void) {
	// Loop through all basic blocks and examine all DEFs. For Global DEFs, record
	//  the block number in BlocksDefinedIn. The VarKillSet records DEFs without
	//  having to examine every instruction.
	list<SMPBasicBlock>::iterator CurrBlock;
	this->BlocksDefinedIn.clear();
	for (size_t i = 0; i < this->GlobalNames.size(); ++i) {
		list<int> TempList;
		this->BlocksDefinedIn.push_back(TempList);
	}
#if SMP_DEBUG_DATAFLOW
	msg("Number of GlobalNames: %d\n", this->GlobalNames.size());
	msg("Size of BlocksDefinedIn: %d\n", this->BlocksDefinedIn.size());
#endif
	for (CurrBlock = this->Blocks.begin(); CurrBlock != this->Blocks.end(); ++CurrBlock) {
		set<op_t, LessOp>::iterator KillIter;
		for (KillIter = CurrBlock->GetFirstVarKill(); KillIter != CurrBlock->GetLastVarKill(); ++KillIter) {
			// If killed item is not a block-local item (it is global), record it.
			set<op_t, LessOp>::iterator NameIter = this->GlobalNames.find(*KillIter);
			if (NameIter != this->GlobalNames.end()) { // found in GlobalNames set
				// We have a kill of a global name. Get index from three 8-bit fields.
				unsigned int index = ExtractGlobalIndex(*NameIter);
				if (index >= this->GlobalNames.size()) {
					// We are about to assert false.
					msg("ComputeBlocksDefinedIn: Bad index: %d limit: %d\n", index,
						this->GlobalNames.size());
					msg("Block number %d\n", CurrBlock->GetNumber());
					msg("Killed item: ");
					PrintListOperand(*KillIter);
					msg("\n");
					msg("This is a fatal error.\n");
				}
				assert(index < this->GlobalNames.size());
				// index is a valid subscript for the BlocksDefinedIn vector. Push the
				//  current block number onto the list of blocks that define this global name.
				this->BlocksDefinedIn[index].push_back(CurrBlock->GetNumber());
			}			
		}
	}
	return;
} // end of SMPFunction::ComputeBlocksDefinedIn()

// Compute the phi functions at the entry point of each basic block that is a join point.
void SMPFunction::InsertPhiFunctions(void) {
	set<op_t, LessOp>::iterator NameIter;
	list<int> WorkList;  // list of block numbers
	bool DebugFlag = false;
#if SMP_DEBUG_DATAFLOW
	DebugFlag = (0 == strcmp("__ieee754_pow", this->GetFuncName()));
#endif
	if (DebugFlag) msg("GlobalNames size: %d\n", this->GlobalNames.size());
	for (NameIter = this->GlobalNames.begin(); NameIter != this->GlobalNames.end(); ++NameIter) {
		int CurrNameIndex = (int) (ExtractGlobalIndex(*NameIter));
		if (DebugFlag) msg("CurrNameIndex: %d\n", CurrNameIndex);
#if 0
		DebugFlag = (DebugFlag && (6 == CurrNameIndex));
#endif
		// Initialize the work list to all blocks that define the current name.
		WorkList.clear();
		list<int>::iterator WorkIter;
		for (WorkIter = this->BlocksDefinedIn.at((size_t) CurrNameIndex).begin();
			WorkIter != this->BlocksDefinedIn.at((size_t) CurrNameIndex).end();
			++WorkIter) {
			WorkList.push_back(*WorkIter);
		}

		// Iterate through the work list, inserting phi functions for the current name
		//  into all the blocks in the dominance frontier of each work list block.
		//  Insert into the work list each block that had a phi function added.
		while (!WorkList.empty()) {
#if SMP_DEBUG_DATAFLOW
			msg("WorkList size: %d\n", WorkList.size());
#endif
			list<int>::iterator WorkIter = WorkList.begin();
			while (WorkIter != WorkList.end()) {
				set<int>::iterator DomFrontIter;
#if SMP_DEBUG_DATAFLOW
				msg("WorkIter: %d\n", *WorkIter);
#endif
				if (DebugFlag && (*WorkIter > this->BlockCount)) {
					msg("ERROR: WorkList block # %d out of range.\n", *WorkIter);
				}
				list<SMPBasicBlock>::iterator WorkBlock = this->RPOBlocks[*WorkIter];
				for (DomFrontIter = WorkBlock->GetFirstDomFrontier();
					DomFrontIter != WorkBlock->GetLastDomFrontier();
					++DomFrontIter) {
#if SMP_DEBUG_DATAFLOW
					msg("DomFront: %d\n", *DomFrontIter);
#endif
					if (DebugFlag && (*DomFrontIter > this->BlockCount)) {
						msg("ERROR: DomFront block # %d out of range.\n", *DomFrontIter);
					}
					list<SMPBasicBlock>::iterator PhiBlock = this->RPOBlocks[*DomFrontIter];
					// Before inserting a phi function for the current name in *PhiBlock,
					//  see if the current name is LiveIn for *PhiBlock. If not, there
					//  is no need for the phi function. This check is what makes the SSA
					//  a fully pruned SSA.
					if (PhiBlock->IsLiveIn(*NameIter)) {
						size_t NumPreds = PhiBlock->GetNumPreds();
						DefOrUse CurrRef(*NameIter);
						SMPPhiFunction CurrPhi(CurrNameIndex, CurrRef);
						for (size_t NumCopies = 0; NumCopies < NumPreds; ++NumCopies) {
							CurrPhi.PushBack(CurrRef); // inputs to phi
						}
						if (PhiBlock->AddPhi(CurrPhi)) {
							// If not already in Phi set, new phi function was inserted.
							WorkList.push_back(PhiBlock->GetNumber());
#if SMP_DEBUG_DATAFLOW
							msg("Added phi for name %d at top of block %d\n", CurrNameIndex, PhiBlock->GetNumber());
#endif
						}
					}
					else {
						if (DebugFlag) {
							msg("Global %d not LiveIn for block %d\n", CurrNameIndex, PhiBlock->GetNumber());
						}
					}
				} // end for all blocks in the dominance frontier
				// Remove current block number from the work list
				if (DebugFlag) {
					msg("Removing block %d from work list.\n", *WorkIter);
				}
				WorkIter = WorkList.erase(WorkIter);
			} // end for all block numbers in the work list
		} // end while the work list is not empty
		if (DebugFlag) msg("WorkList empty.\n");
	} // end for all elements of the GlobalNames set
	return;
} // end of SMPFunction::InsertPhiFunctions()

// Build the dominator tree.
void SMPFunction::BuildDominatorTree(void) {
	size_t index;
	// First, fill the DomTree vector with the parent numbers filled in and the child lists
	//  left empty.
	for (index = 0; index < this->IDom.size(); ++index) {
		pair<int, list<int> > DomTreeEntry;
		DomTreeEntry.first = this->IDom.at(index);
		DomTreeEntry.second.clear();
		this->DomTree.push_back(DomTreeEntry);
	}
	// Now, push the children onto the appropriate lists.
	for (index = 0; index < this->IDom.size(); ++index) {
		// E.g. if block 5 has block 3 as a parent, then we fetch the number 3
		//  using the expression this->DomTree.at(index).first, which was just
		//  initialized in the previous loop. Then we go to DomTree entry 3 and push
		//  the number 5 on its child list.
		int parent = this->DomTree.at(index).first;
		if (parent != index) // block can dominate itself, but not in DomTree!
			this->DomTree.at(parent).second.push_back((int) index);
	}

	return;
} // end of SMPFunction::BuildDominatorTree()

// Helper for SSA subscript renumbering: return the next SSA number for the global name
//  and increment the SSACounter to prepare the next number. Push the returned number onto
//  the SSAStack for the global name.
int SMPFunction::SSANewNumber(size_t GlobNameIndex) {
	int Subscript = this->SSACounter.at(GlobNameIndex);
	++(this->SSACounter[GlobNameIndex]);
	this->SSAStack[GlobNameIndex].push_back(Subscript);
	return Subscript;
} // end of SMPFunction::SSANewNumber()

// Main helper for SSA subscript renumbering. Renumber within block throughout its phi
//  functions, then its DEFs and USEs, then its phi successors. Recurse then on all
//  successors in the dominator tree.
void SMPFunction::SSARename(int BlockNumber) {
	assert(0 <= BlockNumber);
	assert(BlockNumber < this->BlockCount);
	list<SMPBasicBlock>::iterator CurrBlock = this->RPOBlocks.at((size_t) BlockNumber);

	bool DumpFlag = false;
#if SMP_DEBUG_DATAFLOW
	DumpFlag |=	(0 == strcmp("main", this->GetFuncName()));
	DumpFlag |= (0 == strcmp("dohanoi", this->GetFuncName()));
	DumpFlag |= (0 == strcmp("image_to_texture", this->GetFuncName()));
#endif

	if (DumpFlag) msg("Entered SSARename for block number %d\n", BlockNumber);

	// For each phi function at the top of the block, rename the DEF of the phi function
	//  using SSANewNumber() on the global name index.
	set<SMPPhiFunction, LessPhi>::iterator CurrPhi;
	list<SMPPhiFunction> TempPhiList;
	int GlobalNameIndex;
	for (CurrPhi = CurrBlock->GetFirstPhi(); CurrPhi != CurrBlock->GetLastPhi(); ++CurrPhi) {
		GlobalNameIndex = CurrPhi->GetIndex();
		assert(0 <= GlobalNameIndex);
		int NewSSANum = this->SSANewNumber((size_t) GlobalNameIndex);
#if 0
			// g++ is a pain in the neck and won't allow changes to the set item
			//  through CurrPhi, which it types as a const iterator, so this next line does
			//  not compile in g++.
		CurrPhi->SetSSADef(NewSSANum);
#else
		SMPPhiFunction TempPhi = (*CurrPhi);
		TempPhi.SetSSADef(NewSSANum);
		TempPhiList.push_back(TempPhi);
#endif
	}
	// Go back through the Phi function set and replace the items that need to be updated.
	//  Thank you g++ for being a pain.
	list<SMPPhiFunction>::iterator TempIter;
	for (TempIter = TempPhiList.begin(); TempIter != TempPhiList.end(); ++TempIter) {
		// Use the op_t from the first phi use, because they are all the same.
		bool Erased = CurrBlock->ErasePhi(TempIter->GetPhiRef(0).GetOp());
		assert(Erased);
		// Now we can add back the phi function that had the DEF SSA number changed.
		bool Added = CurrBlock->AddPhi(*TempIter);
		assert(Added);
	}
	TempPhiList.clear();
	if (DumpFlag) msg("Processed phi functions at top.\n");

	// For each instruction in the block, rename all global USEs and then all global DEFs.
	list<list<SMPInstr>::iterator>::iterator CurrInst;
	for (CurrInst = CurrBlock->GetFirstInstr(); CurrInst != CurrBlock->GetLastInstr(); ++CurrInst) {
		size_t RefIndex;
		for (RefIndex = 0; RefIndex < (*CurrInst)->NumUses(); ++RefIndex) {
			DefOrUse CurrUse = (*CurrInst)->GetUse(RefIndex);
			// See if Use is a global name.
			set<op_t, LessOp>::iterator GlobIter = this->GlobalNames.find(CurrUse.GetOp());
			if (GlobIter != this->GlobalNames.end()) { // found it
				unsigned int GlobIndex = ExtractGlobalIndex(*GlobIter);
				if (GlobIndex > this->SSAStack.size()) {
					// Get some debug info out to the log file before we crash.
					msg("Bad GlobIndex: %d\n", GlobIndex);
					msg("Error in function %s\n", this->GetFuncName());
					exit(EXIT_FAILURE);
				}
				// Set the SSA number for this use to the top of stack SSA # (back())
				int NewSSANum;
				if (this->SSAStack.at(GlobIndex).empty()) {
					// No top of stack entry to read.
#if SMP_DEBUG_DATAFLOW
					msg("WARNING: function %s : Use of uninitialized variable: ",
						this->GetFuncName());
					msg(" Variable: ");
					PrintListOperand(*GlobIter);
					msg(" Block number: %d Address: %x Instruction: %s\n", BlockNumber,
						(*CurrInst)->GetAddr(), (*CurrInst)->GetDisasm());
#endif
					NewSSANum = SMP_SSA_UNINIT;
				}
				else {
					NewSSANum = this->SSAStack.at(GlobIndex).back();
				}
				(*CurrInst)->SetUseSSA(RefIndex, NewSSANum);
			}
		} // end for all USEs
		for (RefIndex = 0; RefIndex < (*CurrInst)->NumDefs(); ++RefIndex) {
			DefOrUse CurrDef = (*CurrInst)->GetDef(RefIndex);
			// See if Def is a global name.
			set<op_t, LessOp>::iterator GlobIter = this->GlobalNames.find(CurrDef.GetOp());
			if (GlobIter != this->GlobalNames.end()) { // found it
				unsigned int GlobIndex = ExtractGlobalIndex(*GlobIter);
				// Set the SSA number for this DEF to the SSANewNumber top of stack
				(*CurrInst)->SetDefSSA(RefIndex, this->SSANewNumber(GlobIndex));
			}
		} //  end for all DEFs
	} // end for all instructions
	if (DumpFlag) msg("Processed all instructions.\n");

	// For all control flow graph (not dominator tree) successors, fill in the current
	//  (outgoing) SSA number in the corresponding USE slot in the phi function, for all
	//  global names appearing in phi functions.
	list<list<SMPBasicBlock>::iterator>::iterator SuccIter;
	for (SuccIter = CurrBlock->GetFirstSucc(); SuccIter != CurrBlock->GetLastSucc(); ++SuccIter) {
		// What position in the Preds list of this successor is CurrBlock?
		int ListPos = (*SuccIter)->GetPredPosition(BlockNumber);
		assert(0 <= ListPos);

		// Go through all phi functions in this successor. At ListPos position in the
		//  incoming arguments for that phi function, set the SSA number to the SSA number
		//  in the top of stack entry for the global name associated with that phi function.
		set<SMPPhiFunction, LessPhi>::iterator CurrPhi;
		for (CurrPhi = (*SuccIter)->GetFirstPhi(); CurrPhi != (*SuccIter)->GetLastPhi(); ++CurrPhi) {
			int GlobIndex = CurrPhi->GetIndex();
			int CurrSSA;
			if (this->SSAStack.at(GlobIndex).empty()) {
				// No top of stack entry to read.
#if SMP_DEBUG_DATAFLOW
				msg("WARNING: function %s : Path to use of uninitialized variable: ",
					this->GetFuncName());
				msg(" Variable: ");
				PrintListOperand(CurrPhi->GetAnyOp());
				msg(" Block number: %d Successor block number: %d\n", BlockNumber,
					(*SuccIter)->GetNumber());
#endif
				CurrSSA = SMP_SSA_UNINIT;
			}
			else {
				CurrSSA = this->SSAStack.at(GlobIndex).back();  // fetch from top of stack
			}
#if 0
			// g++ is a pain in the neck and won't allow changes to the set item
			//  through CurrPhi, which it types as a const iterator, so this next line does
			//  not compile in g++. C++ does not know how to distinguish between changing
			//  the field that ordering is based on, and other fields, so g++ has to be
			//  strict, I guess.
			CurrPhi->SetSSARef(ListPos, CurrSSA);
#else
			SMPPhiFunction TempPhi = (*CurrPhi);
			TempPhi.SetSSARef(ListPos, CurrSSA);
			TempPhiList.push_back(TempPhi);
			if (DumpFlag && (BlockNumber >= 3) && (BlockNumber <= 4)) {
				msg("BlockNumber: %d  ListPos: %d\n", BlockNumber, ListPos);
			}
#endif
		} // end for all phi functions in successor
		// Go back through the Phi function set and replace the items that need to be updated.
		//  Thank you g++ for being a pain.