SMPFunction.cpp

					if (Status != DefPhi->GetDefMetadata()) {
						DefPhi = CurrBlock->SetPhiDefMetadata(UseOp, Status);
						changed = true;
						// If the Phi DEF has live metadata, then the Phi
						//  USEs each have live metadata. Propagate.
						int UseSSANum;
						for (size_t index = 0; index < DefPhi->GetPhiListSize(); ++index) {
							UseSSANum = DefPhi->GetUseSSANum(index);
							// UseSSANum can be -1 in some cases because
							//  we conservatively make EAX and EDX be USEs
							//  of all return instructions, when the function
							//  might have a void return type, making it
							//  appear as if an uninitialized EAX or EDX
							//  could make it to the return block.
							if (0 <= UseSSANum) {
								changed |= this->PropagateGlobalMetadata(UseOp,
									Status, UseSSANum);
							}
						}
					}
					FoundDef = true;
					break;
				}
			}
		} // end for all blocks
	} // end if !FoundDef

	if (!FoundDef) {
		msg("ERROR: Could not find DEF of SSANum %d for: ");
		PrintOperand(UseOp);
		msg(" in function %s\n", this->GetFuncName());
	}

	return changed;
} // end of SMPFunction::PropagateGlobalMetadata()

// Find consecutive DEFs of the same type and mark the second one redundant.
void SMPFunction::FindRedundantMetadata(void) {
	list<SMPBasicBlock>::iterator CurrBlock;
	bool changed = false;

	for (CurrBlock = this->Blocks.begin(); CurrBlock != this->Blocks.end(); ++CurrBlock) {
		changed |= CurrBlock->FindRedundantLocalMetadata(this->SafeFunc);
	}
	return;
} // end of SMPFunction::FindRedundantMetadata()

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

#if 0
	DumpFlag |= (0 == strcmp("_nl_find_msg", this->GetFuncName()));
	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 (DebugFlag) CurrBlock->Dump();

#if SMP_FULL_LIVENESS_ANALYSIS
		CurrBlock->CreateGlobalChains();
#endif
#if 1
		CurrBlock->MarkDeadRegs();
#endif
	}
#if SMP_DEBUG_DATAFLOW
	if (DumpFlag)
		this->Dump();
#endif
	return;
} // end of SMPFunction::ComputeSSA()

// Find memory writes (DEFs) with possible aliases
void SMPFunction::AliasAnalysis(void) {
	// First task: Mark which memory DEFs MIGHT be aliased because an
	//  indirect memory write occurs somewhere in the DEF-USE chain.
	//  Memory DEF-USE chains with no possible aliasing can be subjected
	//  to type inference and type-based optimizing annotations, e.g. a
	//  register spill to memory followed by retrieval from spill memory
	//  followed by NUMERIC USEs should be typed as a continuous NUMERIC
	//  chain if there is no possibility of aliasing.

	// Preparatory step: For each indirect write, mark all def-use chains
	//  (maintained at the basic block level) that include the indirect
	//  write instruction. If there are no indirect writes in the function,
	//  leave all DEFs marked as unaliased and exit.
	if (!(this->HasIndirectWrites))
		return;

	list<SMPBasicBlock>::iterator CurrBlock;
	for (CurrBlock = this->Blocks.begin(); CurrBlock != this->Blocks.end(); ++CurrBlock) {
		list<list<SMPInstr>::iterator>::iterator CurrInst;
		for (CurrInst = CurrBlock->GetFirstInstr();
			CurrInst != CurrBlock->GetLastInstr();
			++CurrInst) {
			if ((*CurrInst)->HasIndirectMemoryWrite()) {
				CurrBlock->MarkIndWriteChains((*CurrInst)->GetAddr());
				// Until we get true aliasing analysis, any indirect write
				//  is classified as may-be-aliased.
				CurrBlock->SetMaybeAliased(true);
			}
		} // end for all insts in block
	} // end for all blocks in function

	// Step one: Find only the memory DEFs to start with.
	list<SMPInstr>::iterator CurrInst;
	bool FoundIndWrite = false;
	for (CurrInst = this->Instrs.begin(); CurrInst != this->Instrs.end(); ++CurrInst) {
		if (CurrInst->HasDestMemoryOperand()) {
			// Starting with the DEF instruction, traverse the control flow
			//  until we run into (A) the re-definition of the operand, including
			//  a re-definition of any of its addressing registers, or (B) an
			//  indirect write. Return false if condition A terminates the
			//  search, and true if condition B terminates the search.
			this->ResetProcessedBlocks();
			op_t MemDefOp = CurrInst->MDGetMemDefOp();
			assert(o_void != MemDefOp.type);
			set<DefOrUse, LessDefUse>::iterator CurrMemDef = CurrInst->FindDef(MemDefOp);
			assert(CurrMemDef != CurrInst->GetLastDef());
			int SSANum = CurrMemDef->GetSSANum();
			FoundIndWrite = this->FindPossibleChainAlias(CurrInst, MemDefOp, SSANum);
			if (FoundIndWrite) {
				// Mark the DEF as aliased.
				CurrMemDef = CurrInst->SetDefIndWrite(CurrMemDef->GetOp(), true);
				break; // Don't waste time after first alias found
			}
		} // end if inst has dest memory operand
	} // end for all instructions

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

// Does the DefOp DEF_USE chain have an indirect mem write starting at CurrInst?
bool SMPFunction::FindPossibleChainAlias(list<SMPInstr>::iterator CurrInst, op_t DefOp, int SSANum) {
	// Starting with the DEF instruction, traverse the control flow
	//  until we run into (A) the re-definition of the operand, including
	//  a re-definition of any of its addressing registers, or (B) an
	//  indirect write. Return false if condition A terminates the
	//  search, and true if condition B terminates the search.
	SMPBasicBlock *CurrBlock = CurrInst->GetBlock();
	if (!(CurrBlock->IsProcessed())) {
		CurrBlock->SetProcessed(true);
	}
	else
		return false; // block already processed

	// Proceed by cases:
	ea_t DefAddr = CurrInst->GetAddr();
	// Case 1: Local name. Return the IndWrite flag for the local Def-Use
	//  chain begun by CurrInst.
	if (CurrBlock->IsLocalName(DefOp)) {
		return CurrBlock->GetLocalDUChainIndWrite(DefOp, SSANum);
	}

	// Case 2: Global name.
	// Case 2A: If Def-Use chain within this block for this memory operand
	//  has its IndWrite flag set to true, then stop and return true.
	else if (CurrBlock->GetGlobalDUChainIndWrite(DefOp, DefAddr)) {
		return true;
	}

	// Case 2B: Else if Def-Use chain is not the last chain in this block
	//  for this operand, then there must be a later redefinition of the
	//  memory operand (with new SSA number assigned) later in this block.
	//  Because we did not fall into case 2A, we know there is no IndWrite
	//  within the current memory operand's chain, so we return false.
	else if (CurrBlock->IsLastGlobalChain(DefOp, DefAddr)) {
		return false;
	}

	// Case 2C: Else if current memory operand is NOT LiveOut, even though
	//  this is the last def-use chain in the block, then there is no more
	//  traversing of the control flow graph to be done. The chain has ended
	//  without encountering an IndWrite, so return false.
	else if (!(CurrBlock->IsLiveOut(DefOp))) {
		return false;
	}

	// Case 2D: We have passed all previous checks, so we must have a memory
	//  operand that reaches the end of the block without encountering an
	//  IndWrite and is LiveOut. Its may-alias status will be determined by
	//  following the control flow graph for all successor blocks and examining
	//  the def-use chains in those blocks.
	list<list<SMPBasicBlock>::iterator>::iterator SuccBlock;
	SuccBlock = CurrBlock->GetFirstSucc();
	bool FoundIndWrite = false;
	do {
		ea_t FirstAddr = (*SuccBlock)->GetFirstAddr() - 1;
		FoundIndWrite = this->FindChainAliasHelper((*SuccBlock), DefOp, DefAddr);
		++SuccBlock;
	} while (!FoundIndWrite && (SuccBlock != CurrBlock->GetLastSucc()));

	return FoundIndWrite;
} // end of SMPFunction::FindPossibleChainAlias()

// recursive helper for global DU-chains that traverse CFG
bool SMPFunction::FindChainAliasHelper(list<SMPBasicBlock>::iterator CurrBlock, op_t DefOp, ea_t DefAddr) {
	if (!(CurrBlock->IsProcessed())) {
		CurrBlock->SetProcessed(true);
	}
	else
		return false; // block already processed

	// Proceed by global cases:
	// Case 2: Global name.
	// Case 2A: If Def-Use chain within this block for this memory operand
	//  has its IndWrite flag set to true, then stop and return true.
	if (CurrBlock->GetGlobalDUChainIndWrite(DefOp, DefAddr)) {
		return true;
	}

	// Case 2B: Else if Def-Use chain is not the last chain in this block
	//  for this operand, then there must be a later redefinition of the
	//  memory operand (with new SSA number assigned) later in this block.
	//  Because we did not fall into case 2A, we know there is no IndWrite
	//  within the current memory operand's chain, so we return false.
	else if (CurrBlock->IsLastGlobalChain(DefOp, DefAddr)) {
		return false;
	}

	// Case 2C: Else if current memory operand is NOT LiveOut, even though
	//  this is the last def-use chain in the block, then there is no more
	//  traversing of the control flow graph to be done. The chain has ended
	//  without encountering an IndWrite, so return false.
	else if (!(CurrBlock->IsLiveOut(DefOp))) {
		return false;
	}

	// Case 2D: We have passed all previous checks, so we must have a memory
	//  operand that reaches the end of the block without encountering an
	//  IndWrite and is LiveOut. Its may-alias status will be determined by
	//  following the control flow graph for all successor blocks and examining
	//  the def-use chains in those blocks.
	list<list<SMPBasicBlock>::iterator>::iterator SuccBlock;
	SuccBlock = CurrBlock->GetFirstSucc();
	bool FoundIndWrite = false;
	do {
		ea_t FirstAddr = (*SuccBlock)->GetFirstAddr() - 1;
		FoundIndWrite = this->FindChainAliasHelper((*SuccBlock), DefOp, DefAddr);
		++SuccBlock;
	} while (!FoundIndWrite && (SuccBlock != CurrBlock->GetLastSucc()));

    return FoundIndWrite;
} // end of SMPFunction::FindChainAliasHelper()


// 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()));
		bool CondTailCall = false;
		if (CurrBlock->HasReturn()) {
			if (!(CurrInst->IsCondTailCall())) {
				// We either have a return instruction or an unconditional
				//  tail call instruction. We don't want to link to the
				//  tail call target, and there is no link for a return
				continue;
			}
			else {
				// We have a conditional tail call. We don't want to
				//  link to the tail call target, but we do want fall
				//  through to the next instruction.
				CondTailCall = true;
			}
		}

		// Last instruction in block; set successors
		bool CallFlag = (CALL == CurrInst->GetDataFlowType());
		bool IndirCallFlag = (INDIR_CALL == CurrInst->GetDataFlowType());
		bool TailCallFlag = CondTailCall && CurrInst->IsCondTailCall();
		bool IndirJumpFlag = (INDIR_JUMP == CurrInst->GetDataFlowType());
		xrefblk_t CurrXrefs;
		bool LinkedToTarget = false;
		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 || IndirCallFlag || TailCallFlag) 
						&& (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);
						LinkedToTarget = true;
#if SMP_USE_SWITCH_TABLE_INFO
						if (IndirJumpFlag) {
#if SMP_DEBUG_SWITCH_TABLE_INFO
							msg("Switch table link: jump at %x target at %x\n",
								CurrInst->GetAddr(), CurrXrefs.to);
#else
							;
#endif
						}
#endif
					}
				}
		} // end for all xrefs
		if (IndirJumpFlag && (!LinkedToTarget)) {
			this->UnresolvedIndirectJumps = true;
			msg("WARNING: Unresolved indirect jump at %x\n", CurrInst->GetAddr());
		}
		else if (IndirCallFlag && (!LinkedToTarget)) {
			this->UnresolvedIndirectCalls = true;
			msg("WARNING: Unresolved indirect call at %x\n", CurrInst->GetAddr());
		}
	} // 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.
	// NOTE: Prior to construction of hell nodes in functions with unresolved indirect jumps,
	//  we cannot conclude that a block with no predecessors is unreachable. Also, the block
	//  order might be such that removal of a block makes an already processed block
	//  unreachable, so we have to iterate until there are no more changes.
#if SMP_USE_SWITCH_TABLE_INFO
	if (!(this->HasUnresolvedIndirectJumps())) {
#else
	if (!(this->HasIndirectJumps())) {
#endif
		bool changed;
		do {
			changed = false;
			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);
					}
					// Remove the unreachable instructions from the function inst list.
					list<list<SMPInstr>::iterator>::iterator InstIter;
					for (InstIter = CurrBlock->GetFirstInstr(); InstIter != CurrBlock->GetLastInstr(); ++InstIter) {
						list<SMPInstr>::iterator DummyIter = this->Instrs.erase(*InstIter);
					}
					// Finally, remove the block from the blocks list.
					CurrBlock = this->Blocks.erase(CurrBlock);
					this->BlockCount -= 1;
					changed = true;
				}
				else
					++CurrBlock;
			} // end while all blocks after the first one
		} while (changed);
	} // end if not indirect jumps

	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

	// Prior to construction of hell nodes for functions with indirect jumps, there
	//  could still be unnumbered blocks because they appear to be unreachable
	//  (no predecessors from SetLinks() because they are reached only via indirect
	//  jumps). We need to number these and push them on the RPOBlocks vector so
	//  that the vector contains all the blocks.
	if (this->HasIndirectJumps()) {
		for (CurrBlock = this->Blocks.begin(); CurrBlock != this->Blocks.end(); ++CurrBlock) {
			if (SMP_BLOCKNUM_UNINIT == CurrBlock->GetNumber()) {
				msg("Numbering indirectly reachable block at %x\n", CurrBlock->GetFirstAddr());
				CurrBlock->SetNumber(CurrNum);
				this->RPOBlocks.push_back(CurrBlock);
				++CurrNum;
			}
		}
	}
	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());
#endif
	bool DebugFlag = (0 == strcmp("_IO_file_close_mmap", this->GetFuncName()));

	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, so it is just as
	//  good to simply iterate through the blocks in layout order.
#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_USE_SSA_FNOP_MARKER
	// Create DEFs in the marker instruction for all names in the LiveInSet
	//  of the first block. These are the names for the function that
	//  would otherwise look like USEs of uninitialized variables later.
	// Note that the LiveVariableAnalysis work does not actually populate
	//  a LiveInSet for the first block, so we simulate it with its
	//  dataflow equation, UpExposed union (LiveOut minus VarKill).
	set<op_t, LessOp>::iterator UpExposedIter, LiveOutIter;
	list<SMPInstr>::iterator MarkerInst = this->Instrs.begin();
	for (UpExposedIter = this->Blocks.begin()->GetFirstUpExposed();
		UpExposedIter != this->Blocks.begin()->GetLastUpExposed();
		++UpExposedIter) {
			// Add DEF with SSANum of 0.
			MarkerInst->AddDef(*UpExposedIter, UNINIT, 0);
			// Add to the VarKill set.
			this->Blocks.begin()->AddVarKill(*UpExposedIter);
	}
	for (LiveOutIter = this->Blocks.begin()->GetFirstLiveOut();
		LiveOutIter != this->Blocks.begin()->GetLastLiveOut();
		++LiveOutIter) {
			if (!(this->Blocks.begin()->IsVarKill(*LiveOutIter))) {
				// Add DEF with SSANum of 0.
				MarkerInst->AddDef(*LiveOutIter, UNINIT, 0);
				// Add to the VarKill set.
				this->Blocks.begin()->AddVarKill(*LiveOutIter);
			}
	}
#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());
				if (this->HasIndirectJumps()) {
					// Might be reachable only through indirect jumps.
					NewIdom = 0; // make it dominated by entry block
				}
			}
			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;
	list<SMPBasicBlock>::iterator RunnerBlock;
	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;
					RunnerBlock = this->RPOBlocks.at(runner);
					RunnerBlock->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 != (int) 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
	}