Skip to content
Snippets Groups Projects
SMPFunction.cpp 286 KiB
Newer Older
	//  and used as a pointer, then hashed using an arithmetic operation. If the arithmetic
	//  operation always treats its source operands as NUMERIC and produces a NUMERIC
	//  result, e.g. SMP_BITWISE_XOR, then the USE of that pointer is NUMERIC within
	//  this xor instruction. If the DEF at the beginning of the SSA chain for the pointer
	//  is eventually marked as POINTER, then all USEs in the chain will be marked POINTER
	//  as well (see step 2 above). This inconsistency along the USE chain is perfectly
	//  acceptable in our type system. It is important to mark the USEs according to how
	//  we observe them being used, because consistent USEs will propagate back up to
	//  the DEF in step 3 above.

	bool changed;
clc5q's avatar
clc5q committed
	bool NewChange = false;
#if SMP_ANALYZE_INFER_TYPES_TIME
	bool DebugFlag2 = false;
	DebugFlag2 |= (0 == strcmp("Option", this->GetFuncName()));
	long NewChangeCount;
	long IterationCount = 0;
#endif
#if SMP_DEBUG_TYPE_INFERENCE
	DebugFlag |= (0 == strcmp("__libc_csu_init", this->GetFuncName()));
	list<SMPInstr *>::iterator InstIter;
	SMPInstr *CurrInst;
clc5q's avatar
clc5q committed
	set<DefOrUse, LessDefUse>::iterator CurrDef;
	set<DefOrUse, LessDefUse>::iterator NextDef;
	list<SMPBasicBlock *>::iterator BlockIter;
	SMPBasicBlock *CurrBlock;
	if (DebugFlag) {
		this->Dump();
	}
	// One time only: Set the types of immediate values, flags register, stack and frame
	//  pointers, and floating point registers.
		for (InstIter = this->Instrs.begin(); InstIter != this->Instrs.end(); ++InstIter) {
			CurrInst = (*InstIter);
				SMP_msg("SetImmedTypes for inst at %x: %s\n", CurrInst->GetAddr(), CurrInst->GetDisasm());
			CurrInst->SetImmedTypes(this->UseFP);
			// Infer signedness, bit width, and other info from the nature of the instruction
			//  (e.g. loads from stack locations whose signedness has been inferred earlier
			//  in FindOutGoingArgSize(), or inherently signed arithmetic opcodes like signed
			//  or unsigned multiplies and divides).
			CurrInst->MDSetWidthSignInfo(this->UseFP);
		}
		// Check for signedness inferences from conditional branches at the end of blocks.
		for (BlockIter = this->Blocks.begin(); BlockIter != this->Blocks.end(); ++BlockIter) {
			CurrBlock = (*BlockIter);
			CurrBlock->MarkBranchSignedness();
	// Iterate until no more changes: set types in DEF and USE lists based on RTL
clc5q's avatar
clc5q committed
	//  operators and the instruction category, SSA DEF-USE chains, etc.
#if SMP_ANALYZE_INFER_TYPES_TIME
		if (DebugFlag2)
			++IterationCount;
#endif
#if SMP_ANALYZE_INFER_TYPES_TIME
			if (DebugFlag2)
				NewChangeCount = 0;
#endif
			// Step one: Infer types within instructions, context free.
			// Step two, propagating DEF types to all USEs, happens within step one
			//  whenever a DEF type is set for the first time.
			for (InstIter = this->Instrs.begin(); InstIter != this->Instrs.end(); ++InstIter) {
				CurrInst = (*InstIter);
				if (DebugFlag) SMP_msg("Inferring types for %s\n", CurrInst->GetDisasm());
				NewChange = CurrInst->InferTypes();
				changed = (changed || NewChange);
#if SMP_ANALYZE_INFER_TYPES_TIME
				if (DebugFlag2 && NewChange) {
					ea_t InstAddr = CurrInst->GetAddr();
#endif
			}
#if SMP_ANALYZE_INFER_TYPES_TIME
			if (DebugFlag2) {
				SMP_msg(" InferTypes iteration: %ld NewChangeCount: %ld \n", IterationCount, NewChangeCount);
		if (DebugFlag) SMP_msg("Finished type inference steps 1 and 2.\n");
clc5q's avatar
clc5q committed
		// Step three: If all USEs of an SSA name have the same type, but the DEF has no
		//  type, then infer that the DEF must have the same type.
		this->TypedDefs = 0;
		this->UntypedDefs = 0;
		this->TypedPhiDefs = 0;
		this->UntypedPhiDefs = 0;
		// This step of the type inference might converge faster if we used a reverse iterator
		//  to go through the instructions, because we could infer a DEF, propagate it to
		//  the right hand side by making SMPInstr::InferOperatorType() public and calling it
		//  on the SMP_ASSIGN operator after we set the type of the left hand side (DEF). Any
		//  additional DEF inferences would be triggered mostly in the upwards direction by
		//  setting the type of one or more USEs in the current instruction. How much time gain
		//  could be achieved by doing this sequence is questionable.  !!!!****!!!!****
		for (InstIter = this->Instrs.begin(); InstIter != this->Instrs.end(); ++InstIter) {
			CurrInst = (*InstIter);
clc5q's avatar
clc5q committed
			// Find any DEF that still has type UNINIT.
			CurrDef = CurrInst->GetFirstDef();
			while (CurrDef != CurrInst->GetLastDef()) {
				// Set erase() and insert() are needed to change types of DEFs, so
				//  get hold of the next iterator value now.
				NextDef = CurrDef;
				++NextDef;
				if (UNINIT != CurrDef->GetType()) {
					++(this->TypedDefs);
				}
				else {
clc5q's avatar
clc5q committed
					op_t DefOp = CurrDef->GetOp();
					bool MemDef = (DefOp.type != o_reg);
					bool AliasedMemWrite = (MemDef && CurrDef->HasIndirectWrite());
					++(this->UntypedDefs);
					if (MDIsIndirectMemoryOpnd(DefOp, this->UseFP)  // relax this?
						|| (o_mem == DefOp.type)
						// Don't want to infer along DEF-USE chains for indirect
						//  memory accesses until we have alias analysis.
						++CurrDef;
						continue;
					}
clc5q's avatar
clc5q committed
					ea_t DefAddr = CurrInst->GetAddr();
					// Call inference method based on whether it is a block-local
					//  name or a global name.
					CurrBlock = CurrInst->GetBlock();
					if (CurrBlock->IsLocalName(DefOp)) {
clc5q's avatar
clc5q committed
						set<op_t, LessOp>::iterator NameIter;
						NameIter = CurrBlock->FindLocalName(DefOp);
						assert(CurrBlock->GetLastLocalName() != NameIter);
clc5q's avatar
clc5q committed
						unsigned int LocIndex = ExtractGlobalIndex(*NameIter);
						NewChange = CurrBlock->InferLocalDefType(DefOp, LocIndex, DefAddr);
						if (NewChange) {
							--(this->UntypedDefs);
							++(this->TypedDefs);
						}
clc5q's avatar
clc5q committed
						changed = (changed || NewChange);
					}
					else {
						// global name
						bool CallInst = ((CALL == CurrInst->GetDataFlowType())
							|| (INDIR_CALL == CurrInst->GetDataFlowType()));
						SMPOperandType DefType = UNINIT;
						DefType = this->InferGlobalDefType(DefOp,
							CurrDef->GetSSANum(), CurrBlock, CallInst, DefAddr);
						if (IsNotEqType(UNINIT, DefType)) {
							CurrDef = CurrInst->SetDefType(DefOp, DefType);
							--(this->UntypedDefs);
							++(this->TypedDefs);
							NewChange = true;
clc5q's avatar
clc5q committed
						changed = (changed || NewChange);
					} // end if local name ... else ...
				} // end if (UNINIT != CurrDef->GetType()) .. else ...
clc5q's avatar
clc5q committed
				CurrDef = NextDef;
			} // end while all DEFs in the DEF set
		} // end for all instructions
		if (DebugFlag) SMP_msg("Finished type inference step 3.\n");
		for (BlockIter = this->Blocks.begin(); BlockIter != this->Blocks.end(); ++BlockIter) {
			CurrBlock = (*BlockIter);
			changed |= CurrBlock->InferAllPhiDefTypes();
		if (DebugFlag) SMP_msg("Finished unconditional phi type inference.\n");

#if SMP_CONDITIONAL_TYPE_PROPAGATION
		if (!changed) { // Try conditional type propagation
			changed |= this->ConditionalTypePropagation();
#if SMP_DEBUG_TYPE_INFERENCE
			if (DebugFlag) {
				SMP_msg("changed = %d after conditional type propagation.\n", changed);
	} while (changed);
	// With type inference finished, infer signedness from the types, e.g.
	//  POINTER and CODEPOINTER types must be UNSIGNED.
	if (FirstIter) { // Don't want profiler-dependent signedness in the system yet.
		for (InstIter = this->Instrs.begin(); InstIter != this->Instrs.end(); ++InstIter) {
			(*InstIter)->InferSignednessFromSMPTypes(this->UsesFramePointer());
	// Record the meet of all register types that reach RETURN instructions.
	this->MDFindReturnTypes();
	return;
} // end of SMPFunction::InferTypes()
// determine signedness and width info for all operands
void SMPFunction::InferFGInfo(void) {
	bool changed, NewChange;
	unsigned short IterCount = 0;
	list<SMPInstr *>::iterator InstIter;
	list<SMPBasicBlock *>::iterator BlockIter;
	SMPBasicBlock *CurrBlock;
		for (InstIter = this->Instrs.begin(); InstIter != this->Instrs.end(); ++InstIter) {
			SMPInstr *CurrInst = (*InstIter);
			NewChange = CurrInst->InferFGInfo(IterCount);
			changed = (changed || NewChange);
		}
			for (BlockIter = this->Blocks.begin(); BlockIter != this->Blocks.end(); ++BlockIter) {
				CurrBlock = (*BlockIter);
#if STARS_AGGRESSIVE_SIGNEDNESS_PROPAGATION
		if (!changed) {
			changed = this->PropagateSignedness();
		}
#endif
	return;
} // end of SMPFunction::InferFGInfo()

// Apply the profiler information to this function once we've inferred everything we can about it.
void SMPFunction::ApplyProfilerInformation(ProfilerInformation* pi)
{
	assert(pi);

	// If no profiler annotations are available, save time.
	if (0 == pi->GetProfilerAnnotationCount())
		return;

	list<SMPInstr *>::iterator InstIter;
	set<DefOrUse, LessDefUse>::iterator CurrDef, NextDef;
	
	bool DebugFlag = false;
clc5q's avatar
clc5q committed
#if SMP_DEBUG_PROFILED_TYPE_INFERENCE
	DebugFlag |= (0 == strcmp("dohanoi", this->GetFuncName()));
#endif

	// for each instruction in this function 
	for (InstIter = this->Instrs.begin(); InstIter != this->Instrs.end(); ++InstIter) {
		SMPInstr *CurrInst = (*InstIter);
		// lookup whether a load at this instruction was profiled as always numeric 
		InstructionInformation* ii = pi->GetInfo(CurrInst->GetAddr());
clc5q's avatar
clc5q committed
		if (ii && DebugFlag)
			SMP_msg("Found instruction information for %x\n", CurrInst->GetAddr());
clc5q's avatar
clc5q committed
		if (ii && ii->isNumeric()) {
#if SMP_DEBUG_PROFILED_TYPE_INFERENCE
			SMP_msg("Found instruction information for %x and it's numeric!\n", CurrInst->GetAddr());
clc5q's avatar
clc5q committed
#endif
			CurrInst->UpdateMemLoadTypes((SMPOperandType)(NUMERIC|PROF_BASE));
		}

		// lookup whether this instruction has been profiled as an indirect call
clc5q's avatar
clc5q committed
		set<ea_t> indirect_call_targets = pi->GetIndirectCallTargets(CurrInst->GetAddr());
clc5q's avatar
clc5q committed
		for (set<ea_t>::iterator ict_iter = indirect_call_targets.begin();
			ict_iter != indirect_call_targets.end();
			++ict_iter)
clc5q's avatar
clc5q committed
			ea_t target = *ict_iter;
			if (vector_exists(target, IndirectCallTargets))
				IndirectCallTargets.push_back(target);
			if (vector_exists(target, AllCallTargets))
				AllCallTargets.push_back(target);

		}

}	// end of SMPFunction::ApplyProfilerInformation

// For the UNINIT type DEF DefOp, see if all its USEs have a single type.
//  If so, set the DEF to that type and return type,
//  else return UNINIT.
// If DefAddr == BADADDR, then the DEF is in a Phi function, not an instruction.
SMPOperandType SMPFunction::InferGlobalDefType(op_t DefOp, int SSANum, SMPBasicBlock *DefBlock, bool CallInst, ea_t DefAddr) {
	bool DebugFlag = false;
	bool FoundNumeric = false;
	bool FoundPointer = false;
	bool FoundUnknown = false;
	bool FoundUninit = false;
#if SMP_DEBUG_TYPE_INFERENCE
	DebugFlag |= (0 == strcmp("mem_init", this->GetFuncName()));
clc5q's avatar
clc5q committed

	if (DebugFlag) {
		SMP_msg("InferGlobalDefType for SSANum %d of ", SSANum);
clc5q's avatar
clc5q committed
		PrintOperand(DefOp);
	list<SMPInstr *>::iterator InstIter;
clc5q's avatar
clc5q committed

	assert(0 <= SSANum);
	set<DefOrUse, LessDefUse>::iterator CurrUse, CurrDef;
	// Go through all instructions in the block and find the instructions
clc5q's avatar
clc5q committed
	//  that have USEs of DefOp with SSANum. If all USEs in the chain have
	//  a single type (other than UNINIT), change the DEF type to match the
	//  USE type and set changed to true.
	SMPOperandType UseType = UNINIT;
	SMPOperandType PtrType = UNINIT;
clc5q's avatar
clc5q committed

	if (BADADDR == DefAddr) { // DEF is in a Phi function
		FoundDEF = true;
	}
	else { // DEF is in an instruction
		FoundDEF = false; // need to see the DefAddr first
	}

	for (InstIter = DefBlock->GetFirstInstr(); InstIter != DefBlock->GetLastInstr(); ++InstIter) {
		SMPInstr *CurrInst = (*InstIter);
		if ((!FoundDEF) && (DefAddr == CurrInst->GetAddr())) {
			FoundDEF = true;
		}
		else if (FoundDEF) {
			CurrDef = CurrInst->FindDef(DefOp);
			if (CurrDef != CurrInst->GetLastDef()) {
				// Found re-DEF of DefOp.
				DefEscapes = false;
			}
		}
		CurrUse = CurrInst->FindUse(DefOp);
		if (CurrUse != CurrInst->GetLastUse()) { // found a USE of DefOp
			if (CurrUse->GetSSANum() == SSANum) { // matched SSA number
				UseType = CurrUse->GetType();
				FoundNumeric |= (IsNumeric(UseType));
				FoundUnknown |= (IsUnknown(UseType));
				FoundUninit |= (IsEqType(UNINIT, UseType));
				if (IsDataPtr(UseType)) {
					if (FoundPointer) {
						if (IsNotEqType(PtrType, UseType)) {
clc5q's avatar
clc5q committed
#if SMP_DEBUG_TYPE_INFERENCE
							SMP_msg("WARNING: Differing ptr types in global chain:");
							SMP_msg(" Prev: %d Current: %d %s\n", PtrType, UseType,
clc5q's avatar
clc5q committed
#endif
							PtrType = POINTER;
						}
					else {
						FoundPointer = true;
						PtrType = UseType;
					}
				}
			} // end if matched SSA #
		} // end if found a USE of DefOp
	} // end for all instructions

	if (DefEscapes) { // did not find re-def
		DefEscapes = DefBlock->IsLiveOut(DefOp);
	}


	if (DefEscapes) { // Need to recurse into successor blocks
		list<SMPBasicBlock *>::iterator SuccIter;
		ea_t TempAddr;
		this->ResetProcessedBlocks(); // set up recursion
		for (SuccIter = DefBlock->GetFirstSucc(); SuccIter != DefBlock->GetLastSucc(); ++SuccIter) {
			SMPBasicBlock *CurrBlock = (*SuccIter);
			set<SMPPhiFunction, LessPhi>::iterator PhiIter = CurrBlock->FindPhi(DefOp);
			TempAddr = DefAddr;
			if (PhiIter != CurrBlock->GetLastPhi()) {
				TempAddr = BADADDR;  // signals that DefOp will get re-DEFed in a Phi function.
			}
			else if (BADADDR == TempAddr) { // was BADADDR coming in to this function
				// We don't want to pass BADADDR down the recursion chain, because it will be interpreted
				//  by each successor block to mean that DefOp was a Phi USE that got re-DEFed in a Phi function
				//  within itself. Pass the dummy address that indicates LiveIn to the block.
				TempAddr = CurrBlock->GetFirstAddr() - 1;
clc5q's avatar
clc5q committed
			}
			// Should we screen the recursive call below using CurrBlock->IsLiveIn(DefOp) for speed?  !!!!****!!!!
			CurrBlock->InferGlobalDefType(DefOp, SSANum, TempAddr, FoundNumeric, FoundPointer, FoundUnknown, FoundUninit, PtrType);
clc5q's avatar
clc5q committed
	}
	// Do we have a consistent type?
	// If we see any definite POINTER uses, we must set the DEF
	//  to type POINTER or a refinement of it.
	if (FoundPointer)
		UseType = PtrType;
	else if (FoundNumeric && !FoundUninit && !FoundUnknown)
		UseType = NUMERIC;
	else
		return UNINIT; // no POINTER, but no consistent type

	assert(UNINIT != UseType);
	if (DebugFlag) SMP_msg("Inferring global DEF of type %d\n", UseType);
	return UseType;
clc5q's avatar
clc5q committed
} // end of SMPFunction::InferGlobalDefType()

#define SMP_SIMPLE_CONDITIONAL_TYPE_PROPAGATION 1
#if SMP_SIMPLE_CONDITIONAL_TYPE_PROPAGATION
// The simple form of conditional type propagation observes that we
//  simply need to apply the meet operator over Phi function USEs and
//  then propagate any DEF type changes using PropagateGlobalDefType().
//  The outermost iteration over all type inference methods in InferTypes()
//  will take care of all the propagation that is handled by the work list
//  processing in the textbook algorithm.
// Iteration convergence might be slower in the simple approach, but the code
//  is much simpler to debug.
bool SMPFunction::ConditionalTypePropagation(void) {
	bool changed = false;
	SMPBasicBlock *CurrBlock;
	vector<SMPBasicBlock *>::iterator CurrRPO;
	set<SMPPhiFunction, LessPhi>::iterator CurrPhi;

	for (CurrRPO = this->RPOBlocks.begin(); CurrRPO != this->RPOBlocks.end(); ++CurrRPO) {
		CurrBlock = *CurrRPO;
		SMPOperandType MeetType;
		for (CurrPhi = CurrBlock->GetFirstPhi(); CurrPhi != CurrBlock->GetLastPhi(); ++CurrPhi) {
			MeetType = CurrPhi->ConditionalMeetType();
			// Here we use a straight equality test, not our macros,
			//  because we consider it a change if the MeetType is
			//  profiler derived and the DEFType is not.
			if (MeetType == CurrPhi->GetDefType())
				continue;
			// Change the DEF type to the MeetType and propagate.
			op_t DefOp = CurrPhi->GetAnyOp();
			bool IsMemOp = (o_reg != DefOp.type);
			CurrPhi = CurrBlock->SetPhiDefType(DefOp, MeetType);
			changed = true;
			this->ResetProcessedBlocks();
			changed |= CurrBlock->PropagateGlobalDefType(DefOp,
				MeetType, CurrPhi->GetDefSSANum(), IsMemOp);
		} // end for all phi functions in the current block
	} // end for all blocks

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

#else  // not SMP_SIMPLE_CONDITIONAL_TYPE_PROPAGATION

// Apply the SCC (Sparse Conditional Constant) propagation algorithm to
//  propagate types starting from unresolved Phi DEFs.
bool SMPFunction::ConditionalTypePropagation(void) {
	bool changed = false;

	// Collections of Phi functions and instructions that have a DEF
	//  with type UNINIT for the current global name.
	map<int, set<SMPPhiFunction, LessPhi>::iterator> UninitDEFPhis;
	vector<list<SMPInstr>::iterator> UninitDEFInsts;

	// Work lists of Phi functions and instructions that need to be processed
	//  according to the SCC algorithm.
	list<map<int, set<SMPPhiFunction, LessPhi>::iterator>::iterator> PhiWorkList;
	list<vector<list<SMPInstr>::iterator>::iterator> InstWorkList;

	// Iterate through all global names that are either (1) registers
	//  or (2) stack locations in SAFE functions.
	set<op_t, LessOp>::iterator CurrGlob;
	for (CurrGlob = this->GetFirstGlobalName(); CurrGlob != this->GetLastGlobalName(); ++CurrGlob) {
		op_t GlobalOp = *CurrGlob;
		list<SMPBasicBlock>::iterator CurrBlock;
		vector<list<SMPBasicBlock>::iterator>::iterator CurrRPO;
		if (MDIsIndirectMemoryOpnd(GlobalOp, this->UseFP))
			continue; // need alias analysis to process indirect accesses
		if ((GlobalOp.type != o_reg)
			&& (!((this->GetReturnAddressStatus() == FUNC_SAFE) && MDIsStackAccessOpnd(GlobalOp, this->UseFP))))
			continue; // not register, not safe stack access

		// Set up a map (indexed by SSANum) of iterators to Phi functions
		//  for the current global name that have UNINIT as the Phi DEF type.
		UninitDEFPhis.clear();
		UninitDEFInsts.clear();
		for (CurrRPO = this->RPOBlocks.begin(); CurrRPO != this->RPOBlocks.end(); ++CurrRPO) {
			CurrBlock = *CurrRPO;
			set<SMPPhiFunction, LessPhi>::iterator CurrPhi;
			CurrPhi = CurrBlock->FindPhi(GlobalOp);
			if (CurrPhi != CurrBlock->GetLastPhi()) {
				// Found Phi function for current global name.
				if (IsEqType(CurrPhi->GetDefType(), UNINIT)) {
					// Phi DEF is UNINIT; add Phi to the map.
					pair<int, set<SMPPhiFunction, LessPhi>::iterator> TempPair(CurrPhi->GetDefSSANum(), CurrPhi);
					bool Inserted = false;
					map<int, set<SMPPhiFunction, LessPhi>::iterator>::iterator WhereIns;
					pair<map<int, set<SMPPhiFunction, LessPhi>::iterator>::iterator, bool> Result(WhereIns, Inserted);
					Result = UninitDEFPhis.insert(TempPair);
					assert(Result.second == true);
				}
			}
		} // end for all blocks

		// If any Phi DEF had UNINIT as its type, set up a vector of
		//  iterators to instructions that have UNINIT as the DEF type
		//  for the current global name.
		if (UninitDEFPhis.empty())
			continue;
		list<SMPInstr *>::iterator InstIter;
		for (InstIter = this->Instrs.begin(); InstIter != this->Instrs.end(); ++InstIter) {
			SMPInstr *CurrInst = (*InstIter);
			set<DefOrUse, LessDefUse>::iterator CurrDef = CurrInst->FindDef(GlobalOp);
			if (CurrDef != CurrInst->GetLastDef()) {
				// Found DEF of current global name.
				if (IsEqType(UNINIT, CurrDef->GetType())) {
					UninitDEFInsts.push_back(CurrInst);
				}
			}
		} // end for all instructions

		// Put all UNINIT Phi DEFs that have at least one USE
		//  that is not UNINIT onto the PhiWorkList.
		map<int, set<SMPPhiFunction, LessPhi>::iterator>::iterator CurrUnPhi;
		for (CurrUnPhi = UninitDEFPhis.begin(); CurrUnPhi != UninitDEFPhis.end(); ++CurrUnPhi) {
			pair<int, set<SMPPhiFunction, LessPhi>::iterator> PhiDefPair(*CurrUnPhi);
			if (PhiDefPair.second->HasTypedUses()) {
				PhiWorkList.push_back(CurrUnPhi);
			}
		}

		// Iterate until both work lists are empty:
		while (!(PhiWorkList.empty() && InstWorkList.empty())) {
			// Process Phi items first.
			while (!PhiWorkList.empty()) {
				// If applying the meet operator over the Phi USE types
				//  would produce a new DEF type, change the DEF type and
				//  propagate it, adding Phi functions and instructions that
				//  received the propagated type to their respective work lists.
				map<int, set<SMPPhiFunction, LessPhi>::iterator>::iterator MapIter;
				MapIter = PhiWorkList.front();
				PhiWorkList.pop_front();  // remove from work list
				pair<int, set<SMPPhiFunction, LessPhi>::iterator> PhiDefPair;
				PhiDefPair.first = MapIter->first;
				PhiDefPair.second = MapIter->second;
				set<SMPPhiFunction, LessPhi>::iterator CurrPhi = PhiDefPair.second;
				SMPOperandType MeetType = CurrPhi->ConditionalMeetType();
				// Here we use a straight equality test, not our macros,
				//  because we consider it a change if the MeetType is
				//  profiler derived and the DEFType is not.
				if (MeetType == CurrPhi->GetDefType())
					continue;
				// At this point, we need to set the DEFType to the MeetType
				//  and propagate the change. We have a map of all the
				//  critical Phi functions for this global name, as well
				//  as a vector of the relevant instructions for this name.
				CurrPhi->SetDefType(MeetType);
				changed = true;
				int DefSSANum = CurrPhi->GetDefSSANum();
				map<int, set<SMPPhiFunction, LessPhi>::iterator>::iterator PhiIter;
				vector<list<SMPInstr>::iterator>::iterator InstIter;
				// Propagate to Phi functions first.
				for (PhiIter = UninitDEFPhis.begin(); PhiIter != UninitDEFPhis.end(); ++PhiIter) {
					if (DefSSANum == PhiIter->first)
						continue;  // Skip the Phi that we just changed
					for (size_t index = 0; index < PhiIter->second->GetPhiListSize(); ++index) {
						if (DefSSANum == PhiIter->second->GetUseSSANum(index)) {
							// Matched SSA # to USE. Propagate new type.
							PhiIter->second->SetRefType(index, MeetType);
							// Add this phi function to the work list.
							PhiWorkList.push_back(PhiIter);
						}
					}
				}
#define SMP_COND_TYPE_PROP_TO_INSTS 0
#if SMP_COND_TYPE_PROP_TO_INSTS
				// Propagate to instructions with uninit DEFs of global name.
				//  The idea is that the instructions that hold up type propagation
				//  are the ones that USE and then DEF the same global name.
				//  For example, "increment EAX" has to know the type of
				//  the USE of EAX in order to set the type of the DEF.
#endif
			} // end while the PhiWorkList is not empty
#if SMP_COND_TYPE_PROP_TO_INSTS
			// The PhiWorkList is empty at this point, so process
			//  instructions on the InstWorkList.
#endif
		} // end while both work lists are not empty

	} // end for all global names
	return changed;
} // end of SMPFunction::ConditionalTypePropagation()
#endif  // end if SMP_SIMPLE_CONDITIONAL_TYPE_PROPAGATION else ...

// Propagate signedness FG info from DEFs to USEs whenever there is no USE sign info.
bool SMPFunction::PropagateSignedness(void) {
	bool changed = false;
#if STARS_AGGRESSIVE_SIGNEDNESS_PROPAGATION
	map<int, struct FineGrainedInfo>::iterator UseFGIter, DefFGIter;
	list<SMPBasicBlock *>::iterator BlockIter;
	for (UseFGIter = this->GlobalUseFGInfoBySSA.begin(); UseFGIter != this->GlobalUseFGInfoBySSA.end(); ++UseFGIter) {
		struct FineGrainedInfo UseFG = UseFGIter->second;
		if (0 == (UseFG.SignMiscInfo & FG_MASK_SIGNEDNESS_BITS)) {
			// No signedness info. Propagate any signedness info from DEF.
			int UseHashValue = UseFGIter->first;
			unsigned short DefSignMask = this->GetDefSignMiscInfo(UseHashValue);
			DefSignMask &= FG_MASK_SIGNEDNESS_BITS;
			if (0 != DefSignMask) {
				// DEF has signedness info.
				UseFGIter->second.SignMiscInfo |= DefSignMask;
				changed = true;
			}
		}
	}
	// See if we have DEF signedness info for DEFs with no corresponding USE map entries.
	for (DefFGIter = this->GlobalDefFGInfoBySSA.begin(); DefFGIter != this->GlobalDefFGInfoBySSA.end(); ++DefFGIter) {
		struct FineGrainedInfo DefFG = DefFGIter->second;
		unsigned short DefSignMask = (DefFG.SignMiscInfo & FG_MASK_SIGNEDNESS_BITS);
		if (0 != DefSignMask) {
			// Has signedness info. See if USE has no entry.
			int DefHashValue = DefFGIter->first;
			UseFGIter = this->GlobalUseFGInfoBySSA.find(DefHashValue);
			if (UseFGIter == this->GlobalUseFGInfoBySSA.end()) {
				this->UpdateUseSignMiscInfo(DefHashValue, DefSignMask);
				changed = true;
			}
		}
	}
	// Do the same processsing for block-local registers.
	for (BlockIter = this->Blocks.begin(); BlockIter != this->Blocks.end(); ++BlockIter) {
		bool NewChange = (*BlockIter)->PropagateDEFSignedness();
		changed = changed || NewChange;
	}
#endif
	return changed;
} // end of SMPFunction::PropagateSignedness()

// Emit all annotations for the function, including all per-instruction
//  annotations.
void SMPFunction::EmitAnnotations(FILE *AnnotFile, FILE *InfoAnnotFile) {
	// Emit annotation for the function as a whole.
	list<SMPBasicBlock *>::iterator BlockIter;
	SMPBasicBlock *CurrBlock;
	bool FuncHasProblems = ((!this->AnalyzedSP) || (!this->HasGoodRTLs()) || (this->HasUnresolvedIndirectCalls())
		|| (this->HasUnresolvedIndirectJumps()) || (this->HasSharedChunks()));

	if (this->StaticFunc) {
		SMP_fprintf(AnnotFile,	"%10x %6zu FUNC LOCAL  %s ", this->FuncInfo.startEA,
			this->Size, this->GetFuncName());
		SMP_fprintf(AnnotFile,	"%10x %6zu FUNC GLOBAL %s ", this->FuncInfo.startEA,
			this->Size, this->GetFuncName());
	switch (this->GetReturnAddressStatus())
	}
	if (this->FuncInfo.does_return()) {
	if (this->IsLeaf())
	// store the return address
	SMP_fprintf(AnnotFile,"%10x ", this->FuncInfo.endEA - 1);

	if (this->IsLibFunc())
		SMP_fprintf(AnnotFile, "LIBRARY ");
	SMP_fprintf(AnnotFile, "\n");
jdh8d's avatar
jdh8d committed
	// Emit annotations about how to restore register values
	SMP_fprintf(AnnotFile, "%10x %6d FUNC FRAMERESTORE ", this->FuncInfo.startEA, 0);
jdh8d's avatar
jdh8d committed
	{
		SMP_fprintf(AnnotFile, "%d %d %d ", i, this->SavedRegLoc[i], this->ReturnRegTypes[i]);
jdh8d's avatar
jdh8d committed
	}
jdh8d's avatar
jdh8d committed

	SMP_fprintf(AnnotFile, "%10x %6d FUNC MMSAFENESS ", this->FuncInfo.startEA, 0);
clc5q's avatar
clc5q committed
	if (!IsSpecSafe())
clc5q's avatar
clc5q committed
	else if (!IsSafe())
clc5q's avatar
clc5q committed
	else {
		assert(IsSafe());
	// If function has problems that limited our analyses, emit an information annotation so that
	//  other tools can be aware of which analyses will be sound.
	if (FuncHasProblems) {
		SMP_fprintf(InfoAnnotFile,	"%10x %6zu FUNC PROBLEM %s ", this->FuncInfo.startEA,
			this->Size, this->GetFuncName());
		if (!this->AnalyzedSP) {
			SMP_fprintf(InfoAnnotFile, "STACKANALYSIS ");
		}
		if (this->HasSharedChunks()) {
		}
		if (this->HasUnresolvedIndirectJumps()) {
			SMP_fprintf(InfoAnnotFile, "JUMPUNRESOLVED ");
		}
		if (this->HasUnresolvedIndirectCalls()) {
			SMP_fprintf(InfoAnnotFile, "CALLUNRESOLVED ");
		}
		if (!this->HasGoodRTLs()) {
	// Loop through all instructions in the function.
	// Output optimization annotations for those
	//  instructions that do not require full computation
	//  of their memory metadata by the Memory Monitor SDT.
	list<SMPInstr *>::iterator InstIter = Instrs.begin();
	++InstIter;  // skip marker instruction
	bool AllocSeen = false; // Reached LocalVarsAllocInstr yet?
	bool DeallocTrigger = false;
	for ( ; InstIter != Instrs.end(); ++InstIter) {
		SMPInstr *CurrInst = (*InstIter);
		ea_t addr = CurrInst->GetAddr();
		SMP_fprintf(AnnotFile, "%10x %6d INSTR BELONGTO %x \n", addr, 0, GetStartAddr());
		if (this->LocalVarsAllocInstr == addr) {
			AllocSeen = true;
			if (this->NeedsStackReferent)
				this->EmitStackFrameAnnotations(AnnotFile, CurrInst);
				int OptType = CurrInst->GetOptType(); 
				if (5 == OptType) { // ADD or SUB
					// Prevent mmStrata from extending the caller's stack frame
					//  to include the new allocation.
					SMP_fprintf(AnnotFile, "%10x %6d INSTR LOCAL SafeFrameAlloc %s \n",
						addr, -1, CurrInst->GetDisasm());
				}
				else if (CurrInst->MDIsPushInstr()) {
					SMP_fprintf(AnnotFile, "%10x %6d INSTR LOCAL NoWarn %s \n",
				// mmStrata ignores the DATAREF annotations anyway, so even though
				//  they are not needed, emit them for use by Strata and other tools
				//  in other projects besides MEDS.
				this->EmitStackFrameAnnotations(AnnotFile, CurrInst);
		}
		// If this is the instruction which deallocated space
		//  for local variables, we set a flag to remind us to 
		//  emit an annotation on the next instruction.
		// mmStrata wants the instruction AFTER the
		//  deallocating instruction, so that it processes
		//  the deallocation after it happens. It inserts
		//  instrumentation before an instruction, not
		//  after, so it will insert the deallocating
		//  instrumentation before the first POP of callee-saved regs,
		//  if there are any, or before the return, otherwise.
		if (addr == this->LocalVarsDeallocInstr) {
			DeallocTrigger = true;
		}
		else if (DeallocTrigger) { // Time for annotation
			SMP_fprintf(AnnotFile,	"%10x %6d DEALLOC STACK esp - %d %s\n", addr,
				this->LocalVarsSize, this->LocalVarsSize, CurrInst->GetDisasm());
		if (this->StackPtrAnalysisSucceeded() && this->HasGoodRTLs() && !this->HasUnresolvedIndirectJumps() && !this->HasSharedChunks()) {
			CurrInst->EmitTypeAnnotations(this->UseFP, AllocSeen, this->NeedsStackReferent, AnnotFile, InfoAnnotFile);
			CurrInst->EmitIntegerErrorAnnotations(InfoAnnotFile);
			CurrInst->EmitAnnotations(this->UseFP, AllocSeen, this->NeedsStackReferent, AnnotFile, InfoAnnotFile);

		if (CurrInst->MDIsReturnInstr() && this->GetReturnAddressStatus() == FUNC_SAFE)
			CurrInst->EmitSafeReturn(AnnotFile);
	}  // end for all instructions

	// Loop through all basic blocks and emit profiling request annotations
	//  for those blocks that have unsafe memory writes in them.
	this->SafeBlocks = 0;
	this->UnsafeBlocks = 0;
	for (BlockIter = this->Blocks.begin(); BlockIter != this->Blocks.end(); ++BlockIter) {
		CurrBlock = (*BlockIter);
		if (CurrBlock->MaybeAliasedWrite()) {
			++(this->UnsafeBlocks);
			list<SMPInstr *>::iterator CurrInst;
			CurrInst = CurrBlock->GetFirstInstr();
			ea_t addr = (*CurrInst)->GetAddr();
			SMP_fprintf(AnnotFile,	"%10x %6d BLOCK PROFILECOUNT %s\n", addr,
				(*CurrInst)->GetCmd().size, (*CurrInst)->GetDisasm());
	return;
} // end of SMPFunction::EmitAnnotations()

// Debug output dump.
void SMPFunction::Dump(void) {
	list<SMPBasicBlock *>::iterator CurrBlock;
	SMP_msg("Debug dump for function: %s\n", this->GetFuncName());
	SMP_msg("UseFP: %d  LocalVarsAllocInstr: %x\n", this->UseFP,
		this->LocalVarsAllocInstr);
	for (size_t index = 0; index < this->IDom.size(); ++index) {
		SMP_msg("IDOM for %zu: %d\n", index, this->IDom.at(index));
	for (size_t index = 0; index < this->DomTree.size(); ++index) {
		list<int>::iterator DomIter;
		for (DomIter = this->DomTree.at(index).second.begin();
			DomIter != this->DomTree.at(index).second.end();
			++DomIter) {
	set<op_t, LessOp>::iterator NameIter;
	for (NameIter = this->GlobalNames.begin(); NameIter != this->GlobalNames.end(); ++NameIter) {
		SMP_msg("index: %d ", ExtractGlobalIndex(*NameIter));
		PrintListOperand(*NameIter);
	SMP_msg("Blocks each name is defined in: \n");
	for (size_t index = 0; index < this->BlocksDefinedIn.size(); ++index) {
		SMP_msg("Name index: %zu Blocks: ", index);
		list<int>::iterator BlockIter;
		for (BlockIter = this->BlocksDefinedIn.at(index).begin();
			BlockIter != this->BlocksDefinedIn.at(index).end();
			++BlockIter) {
	}
	for (CurrBlock = this->Blocks.begin(); CurrBlock != this->Blocks.end(); ++CurrBlock) {
		// Dump out the function number and data flow sets before the instructions.
		(*CurrBlock)->Dump();
	SMP_msg("End of debug dump for function: %s\n", this->GetFuncName());
	return;
} // end of SMPFunction::Dump()


// Analyzes the function to see if the return address can be marked as safe 
void SMPFunction::MarkFunctionSafe() {
clc5q's avatar
clc5q committed
#if SMP_DEBUG_FUNC
	SMP_msg(" Analyzing function %s and isLeaf = %d \n ", this->GetFuncName(), this->IsLeaf());
	bool HasStackPointerCopy = false;
	bool HasStackPointerPush = false;
	bool HasIndirectGlobalWrite = false;
	bool WritesAboveLocalFrame = false;		// Direct writes above local frame
	bool WritesAboveLocalFrameIndirect = false;	// Indirect writes above local frame
	bool HasIndexedStackWrite = false;
	bool HasIndirectWrite = false;
	bool HasBranchToFarChunk = false; // tail call jump, or perhaps shared chunks in function
	bool IsIndirectCallTarget = false; // could be called indirectly
	bool IsTailCallTarget = false; // could be called by jump instruction used as tail call
	bool HasNoCallers = this->AllCallSources.empty();

#if SMP_USE_SWITCH_TABLE_INFO
	if (this->UnresolvedIndirectJumps) {
#else
	if (this->IndirectJumps) {
clc5q's avatar
clc5q committed
#if SMP_DEBUG_FUNC
		SMP_msg("Function %s marked as unsafe due to indirect jumps\n", this->GetFuncName());

#if SMP_DECLARE_INDIRECT_TARGETS_UNSAFE
	ea_t FirstAddr = this->FirstEA;
	SMP_xref_t xrefs;
	for (bool ok = xrefs.SMP_first_to(FirstAddr, XREF_ALL); ok; ok = xrefs.SMP_next_to()) {
		ea_t FromAddr = xrefs.GetFrom();
		if (FromAddr != 0) {
			if (!xrefs.GetIscode()) { // found data xref
				IsIndirectCallTarget = true; // addr of func appears in data; assume indirect calls to func
			}
			else { // found code xref; see if it is a jump used as a tail call
				// These tail calls could be a problem for fast returns if they go from unsafe to safe functions.
				insn_t TempCmd;
				ulong TempFeatures;
				bool CmdOK = SMPGetCmd(FromAddr, TempCmd, TempFeatures);
				if (!CmdOK) {
					// Better be conservative and assume it could be a tail call.
					IsTailCallTarget = true;
					SMP_msg("ERROR: Could not decode instruction at %x from within MarkFunctionSafe(); assuming tail call\n", FromAddr);
				}
				else if (TempCmd.itype != MD_CALL_INSTRUCTION) { // not a call instruction; must be jump of some sort
					IsTailCallTarget = true;
				}
			}
		}
	}
	this->PossibleIndirectCallTarget = IsIndirectCallTarget;
	this->PossibleTailCallTarget = IsTailCallTarget;
	list<SMPInstr *>::iterator Instructions = Instrs.begin();
	SMPInstr *CurrInst;
	++Instructions;  // skip marker instruction
	// While processing the stack pointer writes, the prologue code for
	//  saving the frame register and allocating local variables needs to be
	//  handled.
	bool SaveEBP = false;
	bool XferESPtoEBP = false;
	for ( ; Instructions != Instrs.end(); ++Instructions) {
clc5q's avatar
clc5q committed
#if SMP_VERBOSE_DEBUG_FUNC 
		SMP_msg(" Total number of defs for this instruction %d\n", CurrInst->NumDefs());
		if (CurrInst->IsBranchToFarChunk()) {
			HasBranchToFarChunk = true;
		}
		if (!SaveEBP) { // still looking for "push ebp"
			if (CurrInst->MDIsPushInstr() && CurrInst->GetCmd().Operands[0].is_reg(R_bp)) {
				SaveEBP = true;
				continue;
			}
		}
		else if (!XferESPtoEBP) { // found "push ebp", looking for "mov ebp,esp"
			insn_t CurrCmd = CurrInst->GetCmd();
			if ((CurrCmd.itype == NN_mov)
					&& (CurrInst->GetFirstDef()->GetOp().is_reg(R_bp))
					&& (CurrInst->GetFirstUse()->GetOp().is_reg(R_sp))) {
				XferESPtoEBP = true;
				continue;
			}
		}
		ea_t address = CurrInst->GetAddr();
		if (address == this->LocalVarsAllocInstr ||
		    address == this->LocalVarsDeallocInstr)
			continue;

		if (CurrInst->MDIsStackPointerCopy(this->UseFP)) {
			if (NN_lea == CurrInst->GetCmd().itype) {
				// If an lea instruction loads an address above
				//  the stack frame, we must assume that writes
				//  above the stack frame could occur.
				op_t TempOp = CurrInst->GetLeaMemUseOp();
				if (this->WritesAboveLocalFrame(TempOp, CurrInst->AreUsesNormalized()))
					WritesAboveLocalFrameIndirect = true;