Newer
Older
// into the statement:
// mov eax,[esp+3Ch+var_30]
// Here, 3Ch == 60 decimal is the distance between ESP and EBP, and
// var_30 is defined to have the value -30h == -48 decimal. So, the
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
// "frame size" in IDA Pro is 60 bytes, and a certain local can be
// addressed in ESP-relative manner as shown, or as [ebp+var_30] for
// EBP-relative addressing. The interactive IDA user can then edit
// the name var_30 to something mnemonic, such as "virus_size", and IDA
// will replace all occurrences with the new name, so that code references
// automatically become [ebp+virus_size]. As the user proceeds
// interactively, he eventually produces very understandable code.
// This all makes sense for producing readable assembly text. However,
// our analyses have a compiler perspective as well as a memory access
// defense perspective. SMP distinguishes between callee saved regs,
// which should not be overwritten in the function body, and local
// variables, which can be written. We view the stack frame in logical
// pieces: here are the saved regs, here are the locals, here is the
// return address, etc. We don't care which direction from EBP the
// callee-saved registers lie; we don't want to lump them in with the
// local variables. We also don't like the fact that IDA Pro will take
// the function prologue code shown above and declare frregs=4 and
// frsize=60, because frsize no longer matches the stack allocation
// statement sub esp,34h == sub esp,52. We prefer frsize=52 and frregs=12.
// So, the task of this function is to fix these stack sizes in our
// private data members for the function, while leaving the IDA database
// alone because IDA needs to maintain its own definitions of these
// variables.
// Fixing means we will update the data members LocalVarsSize and
// CalleeSavedRegsSize.
// NOTE: This function is both machine dependent and platform dependent.
// The prologue and epilogue code generated by gcc-linux is as discussed
// above, while on Visual Studio and other Windows x86 compilers, the
// saving of registers other than EBP happens AFTER local stack allocation.
// A Windows version of the function would expect to see the pushing
// of ESI and EDI AFTER the sub esp,34h statement.
bool SMPFunction::MDFixFrameInfo(void) {
int SavedRegsSize = 0;
int OtherPushesSize = 0; // besides callee-saved regs
int NewLocalsSize = 0;
int OldFrameTotal = this->CalleeSavedRegsSize + this->LocalVarsSize;
bool Changed = false;
bool DebugFlag = (0 == strcmp("__libc_csu_init", this->GetFuncName()));
// Iterate through the first basic block in the function. If we find
// a frame allocating Instr in it, then we have local vars. If not,
// we don't, and LocalVarsSize should have been zero. Count the callee
// register saves leading up to the local allocation. Set data members
// according to what we found if the values of the data members would
// change.
SMPBasicBlock *CurrBlock = this->Blocks.front();
list<SMPInstr *>::iterator CurrIter = CurrBlock->GetFirstInstr();
#if SMP_USE_SSA_FNOP_MARKER
++CurrIter; // skip marker instruction
for ( ; CurrIter != CurrBlock->GetLastInstr(); ++CurrIter) {
SMPInstr *CurrInstr = (*CurrIter);
if (CurrInstr->MDIsPushInstr()) {
// We will make the gcc-linux assumption that a PUSH in
// the first basic block, prior to the stack allocating
// instruction, is a callee register save. To make this
// more robust, we ensure that the register is from
// the callee saved group of registers, and that it has
// not been defined thus far in the function (else it might
// be a push of an outgoing argument to a call that happens
// in the first block when there are no locals). **!!!!**
if (CurrInstr->MDUsesCalleeSavedReg()
&& !CurrInstr->HasSourceMemoryOperand()) {
SavedRegsSize += 4; // **!!** should check the size
clc5q
committed
if (DebugFlag) SMP_msg("libc_csu_init SavedRegsSize: %d %s\n", SavedRegsSize,
}
else {
// Pushes of outgoing args can be scheduled so that
// they are mixed with the pushes of callee saved regs.
OtherPushesSize += 4;
clc5q
committed
if (DebugFlag) SMP_msg("libc_csu_init OtherPushesSize: %d %s\n", OtherPushesSize,
}
}
else if (CurrInstr->MDIsFrameAllocInstr()) {
clc5q
committed
if (DebugFlag) SMP_msg("libc_csu_init allocinstr: %s\n", CurrInstr->GetDisasm());
SavedRegsSize += OtherPushesSize;
// Get the size being allocated.
set<DefOrUse, LessDefUse>::iterator CurrUse;
for (CurrUse = CurrInstr->GetFirstUse(); CurrUse != CurrInstr->GetLastUse(); ++CurrUse) {
// Find the immediate operand.
if (o_imm == CurrUse->GetOp().type) {
// Get its value into LocalVarsSize.
long AllocValue = (signed long) CurrUse->GetOp().value;
// One compiler might have sub esp,24 and another
// might have add esp,-24. Take the absolute value.
if (0 > AllocValue)
AllocValue = -AllocValue;
if (AllocValue != (long) this->LocalVarsSize) {
Changed = true;
#if SMP_DEBUG_FRAMEFIXUP
if (AllocValue + SavedRegsSize != OldFrameTotal)
clc5q
committed
SMP_msg("Total frame size changed: %s\n", this->GetFuncName());
#endif
this->LocalVarsSize = (asize_t) AllocValue;
this->CalleeSavedRegsSize = (ushort) SavedRegsSize;
NewLocalsSize = this->LocalVarsSize;
}
else { // Old value was correct; no change.
NewLocalsSize = this->LocalVarsSize;
if (SavedRegsSize != this->CalleeSavedRegsSize) {
this->CalleeSavedRegsSize = (ushort) SavedRegsSize;
Changed = true;
#if SMP_DEBUG_FRAMEFIXUP
clc5q
committed
SMP_msg("Only callee regs size changed: %s\n", this->GetFuncName());
#endif
}
}
} // end if (o_imm == ...)
} // end for all uses
break; // After frame allocation instr, we are done
} // end if (push) .. elsif frame allocating instr
} // end for all instructions in the first basic block
// If we did not find an allocating instruction, see if it would keep
// the total size the same to set LocalVarsSize to 0 and to set
// CalleeSavedRegsSize to SavedRegsSize. If so, do it. If not, we
// might be better off to leave the numbers alone.
if (!Changed && (NewLocalsSize == 0)) {
clc5q
committed
if (DebugFlag) SMP_msg("libc_csu_init OldFrameTotal: %d \n", OldFrameTotal);
if (OldFrameTotal == SavedRegsSize) {
this->LocalVarsSize = 0;
Changed = true;
}
#if SMP_DEBUG_FRAMEFIXUP
else {
clc5q
committed
SMP_msg("Could not update frame sizes: %s\n", this->GetFuncName());
}
#endif
}
#if SMP_DEBUG_FRAMEFIXUP
if ((0 < OtherPushesSize) && (0 < NewLocalsSize))
clc5q
committed
SMP_msg("Extra pushes found of size %d in %s\n", OtherPushesSize,
clc5q
committed
#if SMP_DEBUG_FRAMEFIXUP
if (Changed) {
SMP_msg("Fixed stack frame size info: %s\n", this->GetFuncName());
SMPBasicBlock *CurrBlock = this->Blocks.front();
SMP_msg("First basic block:\n");
for (list<SMPInstr *>::iterator CurrInstr = CurrBlock->GetFirstInstr();
CurrInstr != CurrBlock->GetLastInstr();
++CurrInstr) {
SMP_msg("%s\n", (*CurrInstr)->GetDisasm());
}
}
#endif
return Changed;
} // end of SMPFunction::MDFixFrameInfo()
// Some functions have difficult to find stack allocations. For example, in some
// version of glibc, strpbrk() zeroes out register ECX and then pushes it more than
// 100 times in order to allocate zero-ed out local vars space for a character translation
// table. We will use the stack pointer analysis of IDA to find out if there is a point
// in the first basic block at which the stack pointer reaches the allocation total
// that IDA is expecting for the local vars region.
// If so, we return the address of the instruction at which ESP reaches its value, else
// we return BADADDR.
ea_t SMPFunction::FindAllocPoint(asize_t OriginalLocSize) {
sval_t TargetSize = - ((sval_t) OriginalLocSize); // negate; stack grows down
#if SMP_DEBUG_FRAMEFIXUP
clc5q
committed
bool DebugFlag = (0 == strcmp("_dl_runtime_resolve", this->GetFuncName()));
clc5q
committed
SMP_msg("%s OriginalLocSize: %d\n", this->GetFuncName(), OriginalLocSize);
// Limit our analysis to the first basic block in the function.
list<SMPInstr *>::iterator InstIter = this->Instrs.begin();
#if SMP_USE_SSA_FNOP_MARKER
++InstIter; // skip marker instruction
for ( ; InstIter != this->Instrs.end(); ++InstIter) {
SMPInstr *CurrInst = (*InstIter);
ea_t addr = CurrInst->GetAddr();
// get_spd() returns a cumulative delta of ESP
clc5q
committed
sval_t sp_delta = get_spd(this->GetFuncInfo(), addr);
#if SMP_DEBUG_FRAMEFIXUP
if (DebugFlag)
clc5q
committed
SMP_msg("%s delta: %d at %x\n", this->GetFuncName(), sp_delta, addr);
if (sp_delta == TargetSize) { // <= instead of == here? **!!**
// Previous instruction hit the frame size.
if (InstIter == this->Instrs.begin()) {
return BADADDR; // cannot back up from first instruction
}
else {
ea_t PrevAddr = (*(--InstIter))->GetAddr();
#if SMP_USE_SSA_FNOP_MARKER
if ((*(this->Instrs.begin()))->GetAddr() == PrevAddr)
return BADADDR; // don't return marker instruction
else
return PrevAddr;
#else
return PrevAddr;
#endif
if (CurrInst->IsLastInBlock()) {
// It could be that the current instruction will cause the stack pointer
// delta to reach the TargetSize. sp_delta is not updated until after the
// current instruction, so we need to look ahead one instruction if the
// current block falls through. On the other hand, if the current block
// ends with a jump or return, we cannot hit TargetSize.
if (CurrInst->IsBasicBlockTerminator())
return BADADDR;
list<SMPInstr *>::iterator NextInstIter = InstIter;
++NextInstIter;
if (NextInstIter == this->Instrs.end())
return BADADDR;
sp_delta = get_spd(this->GetFuncInfo(), (*NextInstIter)->GetAddr());
if (sp_delta == TargetSize) {
// CurrInst will cause stack pointer delta to hit TargetSize.
return addr;
}
else {
return BADADDR;
}
} // end if LastInBlock
} // end for all instructions
} // end if (this->AnalyzedSP)
#if SMP_DEBUG_FRAMEFIXUP
else {
clc5q
committed
SMP_msg("AnalyzedSP is false for %s\n", this->GetFuncName());
}
#endif
return BADADDR;
} // end of SMPFunction::FindAllocPoint()
// IDA Pro is sometimes confused by a function that uses the frame pointer
// register for other purposes. For the x86, a function that uses EBP
// as a frame pointer would begin with: push ebp; mov ebp,esp to save
// the old value of EBP and give it a new value as a frame pointer. The
// allocation of local variable space would have to come AFTER the move
// instruction. A function that begins: push ebp; push esi; sub esp,24
// is obviously not using EBP as a frame pointer. IDA is apparently
// confused by the push ebp instruction being the first instruction
// in the function. We will reset UseFP to false in this case.
// The inverse problem happens with a function that begins with instructions
// other than push ebp; mov ebp,esp; ... etc. but eventually has those
// instructions in the first basic block. For example, a C compiler generates
// for the first block of main():
// lea ecx,[esp+arg0]
// and esp, 0xfffffff0
// push dword ptr [ecx-4]
// push ebp
// mov ebp,esp
// push ecx
// sub esp,<framesize>
//
// This function is obviously using EBP as a frame pointer, but IDA Pro marks
// the function as not using a frame pointer. We will reset UseFP to true in
// this case.
// NOTE: This logic should work for both Linux and Windows x86 prologues.
bool SMPFunction::MDFixUseFP(void) {
list<SMPInstr *>::iterator InstIter = this->Instrs.begin();
ea_t addr = (*InstIter)->GetAddr();
#if SMP_USE_SSA_FNOP_MARKER
++InstIter; // skip marker instruction
SMPInstr *CurrInst = (*InstIter);
if (!(this->UseFP)) {
// See if we can detect the instruction "push ebp" followed by the instruction
// "mov ebp,esp" in the first basic block. The instructions do not have to be
// consecutive. If we find them, we will reset UseFP to true.
bool FirstBlockProcessed = false;
bool EBPSaved = false;
bool ESPintoEBP = false;
do {
FirstBlockProcessed = CurrInst->IsLastInBlock();
if (!EBPSaved) { // still looking for "push ebp"
if (CurrInst->MDIsPushInstr() && CurrInst->GetCmd().Operands[0].is_reg(R_bp)) {
EBPSaved = true;
}
}
else if (!ESPintoEBP) { // 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))) {
ESPintoEBP = true;
FirstBlockProcessed = true; // exit loop
}
}
++InstIter;
CurrInst = (*InstIter);
addr = CurrInst->GetAddr();
// We must get EBP set to its frame pointer value before we reach the
// local frame allocation instruction (i.e. the subtraction of locals space
// from the stack pointer).
FirstBlockProcessed |= (addr >= this->LocalVarsAllocInstr);
} while (!FirstBlockProcessed);
// If we found ESPintoEBP, we also found EBPSaved first, and we need to change
// this->UseFP to true and return true. Otherwise, return false.
this->UseFP = ESPintoEBP;
clc5q
committed
if (ESPintoEBP)
clc5q
committed
SMP_msg("INFO: MDFixUseFP reset UseFP to true for %s\n", this->GetFuncName());
return ESPintoEBP;
} // end if (!(this->UseFP))
// At this point, this->UseFP must have been true on entry to this method and we will
// check whether it should be reset to false.
while (addr < this->LocalVarsAllocInstr) {
set<DefOrUse, LessDefUse>::iterator CurrDef = CurrInst->GetFirstDef();
while (CurrDef != CurrInst->GetLastDef()) {
if (CurrDef->GetOp().is_reg(R_bp))
return false; // EBP got set before locals were allocated
++InstIter;
CurrInst = (*InstIter);
addr = CurrInst->GetAddr();
}
// If we found no defs of the frame pointer before the local vars
// allocation, then the frame pointer register is not being used
// as a frame pointer, just as a general callee-saved register.
this->UseFP = false;
clc5q
committed
SMP_msg("INFO: MDFixUseFP reset UseFP to false for %s\n", this->GetFuncName());
return true;
} // end of SMPFunction::MDFixUseFP()
// Find the callee-saved reg offsets (negative offset from return address)
// for all registers pushed onto the stack before the stack frame allocation
// instruction.
void SMPFunction::MDFindSavedRegs(void) {
list<SMPInstr *>::iterator InstIter;
int RegIndex;
clc5q
committed
func_t *CurrFunc = SMP_get_func(this->GetStartAddr());
assert(NULL != CurrFunc);
for (InstIter = this->Instrs.begin(); InstIter != this->Instrs.end(); ++InstIter) {
SMPInstr *CurrInst = (*InstIter);
if (CurrInst->GetAddr() > this->LocalVarsAllocInstr)
break;
if (!(CurrInst->MDIsPushInstr()))
continue;
sval_t CurrOffset = get_spd(CurrFunc, CurrInst->GetAddr());
if (CurrInst->GetCmd().itype == NN_push) {
op_t PushedReg = CurrInst->GetPushedOpnd();
if (o_reg == PushedReg.type) {
RegIndex = (int) PushedReg.reg;
if (RegIndex > R_di) {
clc5q
committed
SMP_msg("WARNING: Skipping save of register %d\n", RegIndex);
continue;
}
if (this->SavedRegLoc.at((size_t) RegIndex) == 0) {
this->SavedRegLoc[(size_t) RegIndex] = CurrOffset - 4;
}
else {
clc5q
committed
SMP_msg("WARNING: Multiple saves of register %d\n", RegIndex);
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
}
} // end if register push operand
} // end if PUSH instruction
else if (NN_pusha == CurrInst->GetCmd().itype) {
// **!!** Handle pushes of all regs.
this->SavedRegLoc[(size_t) R_ax] = CurrOffset - 4;
this->SavedRegLoc[(size_t) R_cx] = CurrOffset - 8;
this->SavedRegLoc[(size_t) R_dx] = CurrOffset - 12;
this->SavedRegLoc[(size_t) R_bx] = CurrOffset - 16;
this->SavedRegLoc[(size_t) R_sp] = CurrOffset - 20;
this->SavedRegLoc[(size_t) R_bp] = CurrOffset - 24;
this->SavedRegLoc[(size_t) R_si] = CurrOffset - 28;
this->SavedRegLoc[(size_t) R_di] = CurrOffset - 32;
break; // all regs accounted for
}
else if (CurrInst->MDIsEnterInstr()) {
this->SavedRegLoc[(size_t) R_bp] = CurrOffset - 4;
}
} // end for all instructions
return;
} // end of SMPFunction::MDFindSavedRegs()
// Compute the ReturnRegTypes[] as the meet over all register types
// at all return instructions.
void SMPFunction::MDFindReturnTypes(void) {
list<SMPBasicBlock *>::iterator BlockIter;
SMPBasicBlock *CurrBlock;
list<SMPInstr *>::iterator InstIter;
vector<SMPOperandType> RegTypes;
SMPInstr *CurrInst;
for (BlockIter = this->Blocks.begin(); BlockIter != this->Blocks.end(); ++BlockIter) {
CurrBlock = (*BlockIter);
if (CurrBlock->HasReturn()) {
// Get the types of all registers at the RETURN point.
// Calculate the meet function over them.
InstIter = CurrBlock->GetLastInstr();
--InstIter;
assert(RETURN == CurrInst->GetDataFlowType());
set<DefOrUse, LessDefUse>::iterator CurrUse;
for (CurrUse = CurrInst->GetFirstUse();
CurrUse != CurrInst->GetLastUse();
++CurrUse) {
op_t UseOp = CurrUse->GetOp();
if ((o_reg != UseOp.type) || (R_di < UseOp.reg))
continue;
this->ReturnRegTypes[UseOp.reg]
= SMPTypeMeet(this->ReturnRegTypes.at(UseOp.reg),
CurrUse->GetType());
} // for all USEs in the RETURN instruction
} // end if current block has a RETURN
} // end for all blocks
return;
} // end of SMPFunction::MDFindReturnTypes()
// Determine local variable boundaries in the stack frame.
void SMPFunction::BuildLocalVarTable(void) {
// Currently we just use the info that IDA Pro has inferred from the direct
// addressing of stack locations.
this->SemiNaiveLocalVarID();
return;
} // end of SMPFunction::BuildLocalVarTable()
clc5q
committed
// Limit damage from garbage stack offset values produced by IDA Pro.
#define IDAPRO_KLUDGE_STACK_FRAME_SIZE_LIMIT 5000000
// Use the local variable offset list from IDA's stack frame structure to compute
// the table of local variable boundaries.
void SMPFunction::SemiNaiveLocalVarID(void) {
// NOTE: We use IDA Pro's offsets from this->FuncInfo (e.g. frsize) and NOT
// our own corrected values in our private data members. The offsets we
// read from the stack frame structure returned by get_frame() are consistent
// with other IDA Pro values, not with our corrected values.
list<SMPInstr *>::iterator InstIter;
bool DebugFlag = false;
this->LocalVarOffsetLimit = -20000;
#if SMP_DEBUG_STACK_GRANULARITY
DebugFlag |= (0 == strcmp("qSort3", this->GetFuncName()));
#endif
clc5q
committed
func_t *FuncPtr = SMP_get_func(this->FuncInfo.startEA);
if (NULL == FuncPtr) {
clc5q
committed
SMP_msg("ERROR in SMPFunction::SemiNaiveLocalVarID; no func ptr\n");
}
assert(NULL != FuncPtr);
struc_t *StackFrame = get_frame(FuncPtr);
if (NULL == StackFrame) {
clc5q
committed
SMP_msg("WARNING: No stack frame info from get_frame for %s\n", this->GetFuncName());
return;
}
member_t *Member = StackFrame->members;
for (size_t i = 0; i < StackFrame->memqty; ++i, ++Member) {
long offset;
if (NULL == Member) {
clc5q
committed
SMP_msg("NULL stack frame member pointer in %s\n", this->GetFuncName());
break;
}
get_member_name(Member->id, MemberName, MAXSMPVARSTR - 1);
if (MemberName == NULL) {
#if SMP_DEBUG_STACK_GRANULARITY
clc5q
committed
SMP_msg("NULL stack frame member in %s\n", this->GetFuncName());
continue;
}
if (Member->unimem()) {
// Not a separate variable; name for member of a union.
// The union itself should have a separate entry, so we skip this.
clc5q
committed
SMP_msg("STACK INFO: Skipping union member %s frame member %zu in stack frame for %s\n",
continue;
}
offset = (long) Member->get_soff(); // Would be 0 for union member, so we skipped them above.
if (DebugFlag) {
clc5q
committed
SMP_msg("%s local var %s at offset %ld\n", this->GetFuncName(), MemberName, offset);
clc5q
committed
if (offset > IDAPRO_KLUDGE_STACK_FRAME_SIZE_LIMIT) {
SMP_msg("ERROR: Rejected enormous stack offset %ld for var %s in func %s\n", offset, MemberName, this->GetFuncName());
continue;
}
struct LocalVar TempLocal;
TempLocal.offset = offset;
TempLocal.size = Member->eoff - Member->soff; // audit later
clc5q
committed
SMP_strncpy(TempLocal.VarName, MemberName, sizeof(TempLocal.VarName) - 1);
this->LocalVarTable.push_back(TempLocal);
if ((offset + (long) TempLocal.size) >= this->LocalVarOffsetLimit) {
this->LocalVarOffsetLimit = (long) (TempLocal.offset + TempLocal.size);
}
} // end for all stack frame members
// If AnalyzedSP is false, that is all we can do.
if (!this->AnalyzedSP) {
// No allocations; sometimes happens in library functions.
this->OutgoingArgsSize = 0;
this->MinStackDelta = 0;
this->AllocPointDelta = 0;
return;
}
// Calculate min and max stack point deltas.
this->MinStackDelta = 20000; // Final value should be negative or zero
this->MaxStackDelta = -1000; // Final value should be zero.
InstIter = this->Instrs.begin();
if ((*InstIter)->IsFloatNop())
++InstIter; // skip marker instruction
for ( ; InstIter != this->Instrs.end(); ++InstIter) {
SMPInstr *CurrInst = (*InstIter);
ea_t addr = CurrInst->GetAddr();
sval_t sp_delta = get_spd(this->GetFuncInfo(), addr);
if (sp_delta < this->MinStackDelta)
this->MinStackDelta = sp_delta;
if (sp_delta > this->MaxStackDelta)
this->MaxStackDelta = sp_delta;
if (addr == this->LocalVarsAllocInstr) {
// Total stack pointer delta is sp_delta for the next instruction,
// because IDA updates the sp delta AFTER each instruction.
list<SMPInstr *>::iterator NextInstIter = InstIter;
++NextInstIter;
sp_delta = get_spd(this->GetFuncInfo(), (*NextInstIter)->GetAddr());
this->AllocPointDelta = sp_delta;
}
}
// IDA Pro sometimes fails to add stack frame members for all incoming args, etc.
// Find and correct these omissions by examining stack accesses in instructions
// and extend the LocalVarTable to cover whatever is out of range.
if (!this->AuditLocalVarTable()) {
// Catastrophic error must have occurred, probably due to errors in IDA's
// stack pointer analysis, despite AnalyzedSP being true.
if (!(this->LocalVarTable.empty())) {
this->GoodLocalVarTable = true;
// Sort the LocalVarTable so that we do not depend on IDA Pro
// presenting the stack frame members in order.
std::sort(this->LocalVarTable.begin(), this->LocalVarTable.end(), LocalVarCompare);
}
#if SMP_DEBUG_STACK_GRANULARITY
clc5q
committed
SMP_msg("Computing %d local var sizes\n", this->LocalVarTable.size());
// Now we want to audit the size field for each local
if (this->GoodLocalVarTable) {
size_t VarLimit = this->LocalVarTable.size() - 1;
assert(this->LocalVarTable.size() > 0);
for (size_t VarIndex = 0; VarIndex < VarLimit; ++VarIndex) {
struct LocalVar TempLocEntry = this->LocalVarTable[VarIndex];
bool AboveLocalsRegion = (TempLocEntry.offset >= this->LocalVarsSize);
size_t TempSize = this->LocalVarTable[VarIndex + 1].offset
- TempLocEntry.offset;
int DiffSize = ((int) TempSize) - ((int) TempLocEntry.size);
// We don't have IDA Pro stack frame members for callee saved registers. This
// omission can make it seem that there is a gap between the uppermost local
// variable and the return address or saved frame pointer. Avoid expanding the
// last local variable into the callee saved registers region.
if (DiffSize > 0) { // We are expanding the size.
if (!AboveLocalsRegion && ((TempLocEntry.offset + TempLocEntry.size + DiffSize) > this->LocalVarsSize)) {
// Current local does not start above the locals region, but its new size will
// carry it into the locals region.
if ((TempLocEntry.offset + TempLocEntry.size) > this->LocalVarsSize) {
// Weird. It already overlapped the callee saved regs region.
clc5q
committed
SMP_msg("WARNING: Local var at offset %ld size %zu in %s extends above local vars region.\n",
TempLocEntry.offset, TempLocEntry.size, this->GetFuncName());
}
// Limit DiffSize to avoid overlapping callee saved regs.
DiffSize = this->LocalVarsSize - (TempLocEntry.offset + TempLocEntry.size);
if (DiffSize < 0)
DiffSize = 0; // started out positive, cap it at zero.
}
}
if (DiffSize < 0)
DiffSize = 0; // should not happen with sorted LocalVarTable unless duplicate entries.
if (DiffSize != 0) {
#if SMP_DEBUG_STACK_GRANULARITY
clc5q
committed
SMP_msg("STACK INFO: Adjusted size for stack frame member at %ld in %s\n",
#endif
this->LocalVarTable[VarIndex].size += DiffSize;
}
#if 0 // Using Member->eoff seems to be working for all members, including the last one.
#if SMP_DEBUG_STACK_GRANULARITY
clc5q
committed
SMP_msg("Computing last local var size for frsize %d\n", this->FuncInfo.frsize);
// Size of last local is total frsize minus savedregs in frame minus offset of last local
size_t SavedRegsSpace = 0; // portion of frsize that is saved regs, not locals.
if (this->CalleeSavedRegsSize > this->FuncInfo.frregs) {
// IDA Pro counts the save of EBP in frregs, but then EBP gets its new
// value and callee saved regs other than the old EBP push get counted
// in frsize rather than frregs. CalleeSavedRegsSize includes all saved
// regs on the stack, both above and below the current EBP offset.
// NOTE: For windows, this has to be done differently, as callee saved regs
// happen at the bottom of the local frame, not the top.
#if 0
SavedRegsSpace = this->CalleeSavedRegsSize - this->FuncInfo.frregs;
#else
SavedRegsSpace = this->FuncInfo.frsize - this->LocalVarsSize;
#endif
this->LocalVarTable.back().size = this->FuncInfo.frsize
- SavedRegsSpace - this->LocalVarTable.back().offset;
this->LocalVarOffsetLimit = this->LocalVarTable.back().offset
+ (adiff_t) this->LocalVarTable.back().size;
#if 0 // AboveLocalsSize is not a reliable number.
// IDA Pro can have difficulty with some irregular functions such as are found
// in the C startup code. The frsize value might be bogus. Just punt on the
// local variable ID if that is the case.
if ((this->LocalVarOffsetLimit - AboveLocalsSize) > (adiff_t) this->FuncInfo.frsize) {
this->LocalVarTable.clear();
this->GoodLocalVarTable = false;
clc5q
committed
SMP_msg("WARNING: Bad frsize %d for %s OffsetLimit: %d AboveLocalsSize: %d LocalVarsSize: %d ; abandoning SemiNaiveLocalVarID.\n",
this->FuncInfo.frsize, this->GetFuncName(), this->LocalVarOffsetLimit, AboveLocalsSize, this->LocalVarsSize);
return;
}
assert((this->LocalVarOffsetLimit - AboveLocalsSize) <= (adiff_t) this->FuncInfo.frsize);
// Find out how many of the locals are really outgoing args.
if (this->AnalyzedSP && !this->CallsAlloca && (BADADDR != this->LocalVarsAllocInstr)) {
this->FindOutgoingArgsSize();
}
else {
clc5q
committed
SMP_msg("FindOutgoingArgsSize not called for %s ", this->GetFuncName());
SMP_msg("AnalyzedSP: %d CallsAlloca: %d LocalVarsAllocInstr: %x \n",
this->AnalyzedSP, this->CallsAlloca, this->LocalVarsAllocInstr);
}
return;
} // end of SMPFunction::SemiNaiveLocalVarID()
// Check and correct the LocalVarTable derived from IDA Pro stack frame members.
// Examine each instruction and see if any stack accesses are beyond the LocalVarTable
// and create new entries in the LocalVarTable if so.
bool SMPFunction::AuditLocalVarTable(void) {
list<SMPInstr *>::iterator InstIter;
// We cannot depend on IDA Pro making Member
// entries for everything that is accessed on the stack.
// When an incoming arg is accessed but no Member is
// created, then LocalVarOffsetLimit will be too small
// and we will get ERROR messages. Just loop through the
// instructions, find offsets higher than the LocalVarTable
// currently holds, and add new entries to LocalVarTable to
// handle them.
// Iterate through all instructions and record stack frame accesses in the StackFrameMap.
InstIter = this->Instrs.begin();
if ((*InstIter)->IsFloatNop())
++InstIter; // skip marker instruction
for ( ; InstIter != this->Instrs.end(); ++InstIter) {
SMPInstr *CurrInst = (*InstIter);
ea_t InstAddr = CurrInst->GetAddr();
sval_t sp_delta = get_spd(this->GetFuncInfo(), InstAddr);
if (0 < sp_delta) {
// Stack underflow; about to assert
clc5q
committed
SMP_msg("Stack underflow at %x %s sp_delta: %d\n", CurrInst->GetAddr(),
CurrInst->GetDisasm(), sp_delta);
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
}
assert(0 >= sp_delta);
ea_t offset;
size_t DataSize;
bool UsedFramePointer;
bool IndexedAccess;
bool SignedMove;
bool UnsignedMove;
if (CurrInst->HasDestMemoryOperand()) {
// NOTE: We need to catch stack pushes here also (callee-saved regs). !!!!!*******!!!!!!!!
set<DefOrUse, LessDefUse>::iterator CurrDef;
for (CurrDef = CurrInst->GetFirstDef(); CurrDef != CurrInst->GetLastDef(); ++CurrDef) {
op_t TempOp = CurrDef->GetOp();
if (TempOp.type != o_phrase && TempOp.type != o_displ)
continue;
if (this->MDGetStackOffsetAndSize(CurrInst, TempOp, sp_delta, offset, DataSize, UsedFramePointer,
IndexedAccess, SignedMove, UnsignedMove)) {
SignedOffset = (int) offset;
if (IndexedAccess && ((0 > SignedOffset) || ((offset + DataSize) > this->StackFrameMap.size()))) {
continue; // Indexed expressions can be within frame but offset is outside frame
}
#if 0 // ls_O3.exe has IDA trouble on chunked function get_funky_string().
assert(0 <= SignedOffset);
#else
if (0 > SignedOffset) { // negative offset but not Indexed; very bad
clc5q
committed
SMP_msg("ERROR: Negative stack offset at %x in %s. Abandoning LocalVar ID.\n", CurrInst->GetAddr(), this->GetFuncName());
return false;
}
#endif
if ((SignedOffset + (long) DataSize) > this->LocalVarOffsetLimit) {
// Going out of range. Extend LocalVarTable.
struct LocalVar TempLocal;
char TempStr[20];
TempLocal.offset = (long) SignedOffset;
TempLocal.size = DataSize;
clc5q
committed
SMP_strncpy(TempLocal.VarName, "SMP_InArg", sizeof(TempLocal.VarName) - 1);
(void) SMP_snprintf(TempStr, 18, "%d", offset);
SMP_strncat(TempLocal.VarName, TempStr, sizeof(TempLocal.VarName) - 1);
this->LocalVarTable.push_back(TempLocal);
this->LocalVarOffsetLimit = (long) (SignedOffset + (long) DataSize);
}
}
}
}
if (CurrInst->HasSourceMemoryOperand()) {
set<DefOrUse, LessDefUse>::iterator CurrUse;
for (CurrUse = CurrInst->GetFirstUse(); CurrUse != CurrInst->GetLastUse(); ++CurrUse) {
op_t TempOp = CurrUse->GetOp();
clc5q
committed
if ((TempOp.type != o_phrase) && (TempOp.type != o_displ))
continue;
if (this->MDGetStackOffsetAndSize(CurrInst, TempOp, sp_delta, offset, DataSize, UsedFramePointer,
IndexedAccess, SignedMove, UnsignedMove)) {
SignedOffset = (int) offset;
if (IndexedAccess && ((0 > SignedOffset) || ((offset + DataSize) > this->StackFrameMap.size()))) {
continue; // Indexed expressions can be within frame but offset is outside frame
}
assert(0 <= SignedOffset);
if ((SignedOffset + (long) DataSize) > this->LocalVarOffsetLimit) {
// Going out of range. Extend LocalVarTable.
struct LocalVar TempLocal;
char TempStr[20];
TempLocal.offset = (long) SignedOffset;
TempLocal.size = DataSize;
clc5q
committed
SMP_strncpy(TempLocal.VarName, "SMP_InArg", sizeof(TempLocal.VarName) - 1);
(void) SMP_snprintf(TempStr, 18, "%d", offset);
SMP_strncat(TempLocal.VarName, TempStr, sizeof(TempLocal.VarName) - 1);
this->LocalVarTable.push_back(TempLocal);
this->LocalVarOffsetLimit = (long) (SignedOffset + (long) DataSize);
}
}
}
}
} // end for all instructions
// Fill in the gaps with new variables as well. SHOULD WE? WHY?
return true;
} // end of SMPFunction::AuditLocalVarTable()
// Determine how many bytes at the bottom of the stack frame (i.e. at bottom of
// this->LocalVarsSize) are used for outgoing args. This is the case when the cdecl
// calling convention is used, e.g. gcc/linux allocates local var space + out args space
// in a single allocation and then writes outarg values directly to ESP+0, ESP+4, etc.
void SMPFunction::FindOutgoingArgsSize(void) {
// Compute the lowest value reached by the stack pointer.
list<SMPInstr *>::iterator InstIter;
unsigned short BitWidthMask;
#if SMP_DEBUG_STACK_GRANULARITY
DebugFlag = (0 == strcmp("BZ2_blockSort", this->GetFuncName()));
this->OutgoingArgsComputed = true;
clc5q
committed
SMP_msg("DEBUG: Entered FindOutgoingArgsSize for %s\n", this->GetFuncName());
#if SMP_IDAPRO52_WORKAROUND
this->OutgoingArgsSize = 16;
return;
#if SMP_DEBUG_STACK_GRANULARITY
clc5q
committed
SMP_msg("AllocPointDelta: %d MinStackDelta: %d\n", this->AllocPointDelta, this->MinStackDelta);
if ((0 <= this->MinStackDelta) || (0 <= this->AllocPointDelta)) {
// No allocations; sometimes happens in library functions.
this->OutgoingArgsSize = 0;
this->MinStackDelta = 0;
this->AllocPointDelta = 0;
return;
}
assert(0 > this->MinStackDelta);
// Allocate a vector of stack frame entries, one for each byte of the stack frame.
// This will be our memory map for analyzing stack usage.
int limit = 0;
#if 1
if (this->LocalVarOffsetLimit > 0) {
if (limit < (this->LocalVarOffsetLimit + this->MinStackDelta)) {
// Make room for incoming args, other stuff above local vars.
limit = this->LocalVarOffsetLimit + this->MinStackDelta;
if (this->MinStackDelta < this->AllocPointDelta) {
// Also have stuff below alloc point to make room for.
limit += (this->AllocPointDelta - this->MinStackDelta);
}
#endif
for (int i = this->MinStackDelta; i < limit; ++i) {
struct StackFrameEntry TempEntry;
TempEntry.VarPtr = NULL;
TempEntry.offset = (long) i;
TempEntry.Read = false;
TempEntry.Written = false;
TempEntry.AddressTaken = false;
TempEntry.ESPRelativeAccess = false;
TempEntry.EBPRelativeAccess = false;
TempEntry.IndexedAccess = false;
this->StackFrameMap.push_back(TempEntry);
struct FineGrainedInfo TempFineGrained;
TempFineGrained.SignMiscInfo = 0;
TempFineGrained.SizeInfo = 0;
this->FineGrainedStackTable.push_back(TempFineGrained);
#if 0
for (int i = 0; i < this->LocalVarOffsetLimit; ++i) {
struct FineGrainedInfo TempFineGrained;
TempFineGrained.SignMiscInfo = 0;
TempFineGrained.SizeInfo = 0;
this->FineGrainedStackTable.push_back(TempFineGrained);
#endif
// Fill in the VarPtr fields for each StackFrameMap entry.
if (0 <= this->AllocPointDelta) {
clc5q
committed
SMP_msg("FATAL ERROR: AllocPointDelta = %d in %s\n", this->AllocPointDelta, this->GetFuncName());
assert(0 > this->AllocPointDelta);
for (size_t i = 0; i < this->LocalVarTable.size(); ++i) {
assert(this->LocalVarTable.at(i).offset >= 0);
// Picture that AllocPointDelta is -200, MinStackDelta is -210, and
// the LocalVarTable[i].offset is +8 (i.e. 8 bytes above alloc point).
// Then base = 8 + (-200 - -210) = 8 + 10 = 18, the proper offset into
// the StackFrameMap.
size_t base = (size_t) (this->LocalVarTable.at(i).offset
+ (this->AllocPointDelta - this->MinStackDelta));
size_t limit = base + this->LocalVarTable.at(i).size;
if (limit > this->StackFrameMap.size()) {
clc5q
committed
SMP_msg("ERROR: FindOutArgsSize: base = %zu limit = %zu StackFrameMap size = %zu in %s\n",
base, limit, this->StackFrameMap.size(), this->GetFuncName());
this->OutgoingArgsComputed = false;
this->OutgoingArgsSize = 0;
return;
}
assert(limit <= this->StackFrameMap.size());
for (size_t MapIndex = base; MapIndex < limit; ++MapIndex) {
this->StackFrameMap[MapIndex].VarPtr = &(this->LocalVarTable.at(i));
}
}
// Iterate through all instructions and record stack frame accesses in the StackFrameMap.
InstIter = this->Instrs.begin();
#if SMP_USE_SSA_FNOP_MARKER
if ((*InstIter)->IsFloatNop())
++InstIter; // skip marker instruction
for ( ; InstIter != this->Instrs.end(); ++InstIter) {
SMPInstr *CurrInst = (*InstIter);
ea_t InstAddr = CurrInst->GetAddr();
sval_t sp_delta = get_spd(this->GetFuncInfo(), InstAddr);
if (0 < sp_delta) {
// Stack underflow; about to assert
clc5q
committed
SMP_msg("FATAL ERROR: Stack underflow at %x %s sp_delta: %d\n", CurrInst->GetAddr(),
CurrInst->GetDisasm(), sp_delta);
}
assert(0 >= sp_delta);
ea_t offset;
size_t DataSize;
bool UsedFramePointer;
bool IndexedAccess;
bool SignedMove;
bool UnsignedMove;
if (CurrInst->HasDestMemoryOperand()) {
set<DefOrUse, LessDefUse>::iterator CurrDef;
for (CurrDef = CurrInst->GetFirstDef(); CurrDef != CurrInst->GetLastDef(); ++CurrDef) {
op_t TempOp = CurrDef->GetOp();
if (TempOp.type != o_phrase && TempOp.type != o_displ)
continue;
if (this->MDGetStackOffsetAndSize(CurrInst, TempOp, sp_delta, offset, DataSize, UsedFramePointer,
IndexedAccess, SignedMove, UnsignedMove)) {
SignedOffset = (int) offset;
if (IndexedAccess && ((0 > SignedOffset) || ((offset + DataSize) > this->StackFrameMap.size()))) {
continue; // Indexed expressions can be within frame even when offset is outside frame
}
assert(0 <= SignedOffset);
#if 0
if (offset >= this->FuncInfo.frsize)
continue; // limit processing to outgoing args and locals
if ((offset + DataSize) > this->StackFrameMap.size()) {
clc5q
committed
SMP_msg("ERROR: offset = %u DataSize = %zu FrameMapSize = %zu\n",
offset, DataSize, this->StackFrameMap.size());
}
assert((offset + DataSize) <= this->StackFrameMap.size());
for (int j = 0; j < (int) DataSize; ++j) {
this->StackFrameMap[offset + j].Written = true;
this->StackFrameMap[offset + j].IndexedAccess = IndexedAccess;
if (!UsedFramePointer) {
this->StackFrameMap[offset + j].ESPRelativeAccess = true;
}
else {
this->StackFrameMap[offset + j].EBPRelativeAccess = true;
BitWidthMask = ComputeOperandBitWidthMask(TempOp, DataSize);
this->FineGrainedStackTable.at(offset).SizeInfo |= BitWidthMask;
this->FineGrainedStackTable.at(offset).SignMiscInfo |= FG_MASK_WRITTEN;
if (IndexedAccess) {
this->FineGrainedStackTable.at(offset).SignMiscInfo |= FG_MASK_INDEXED_ACCESS;
}
if (!UsedFramePointer) {
this->FineGrainedStackTable.at(offset).SignMiscInfo |= FG_MASK_SP_RELATIVE;
}
else {
this->FineGrainedStackTable.at(offset).SignMiscInfo |= FG_MASK_FP_RELATIVE;
}
// We will process the signedness of stores later, so that loads can take precedence
// over stores in determining signedness.
} // end if MDGetStackOffsetAndSize()
} // end for all DEFs
} // end if DestMemoryOperand
if (CurrInst->HasSourceMemoryOperand()) {
set<DefOrUse, LessDefUse>::iterator CurrUse;
for (CurrUse = CurrInst->GetFirstUse(); CurrUse != CurrInst->GetLastUse(); ++CurrUse) {
op_t TempOp = CurrUse->GetOp();
if (TempOp.type != o_phrase && TempOp.type != o_displ)
continue;
if (this->MDGetStackOffsetAndSize(CurrInst, TempOp, sp_delta, offset, DataSize, UsedFramePointer,
IndexedAccess, SignedMove, UnsignedMove)) {
SignedOffset = (int) offset;
if (IndexedAccess && ((0 > SignedOffset) || ((SignedOffset + DataSize) > this->StackFrameMap.size()))) {
continue; // Indexed expressions can be within frame but offset is outside frame
}
assert(0 <= SignedOffset);
#if 0
if (offset >= this->FuncInfo.frsize)
continue; // limit processing to outgoing args and locals
#endif
if ((SignedOffset + DataSize) > this->StackFrameMap.size()) {
clc5q
committed
SMP_msg("ERROR: offset = %u DataSize = %zu FrameMapSize = %zu\n",
offset, DataSize, this->StackFrameMap.size());
assert((SignedOffset + DataSize) <= this->StackFrameMap.size());
for (int j = 0; j < (int) DataSize; ++j) {
this->StackFrameMap[offset + j].Read = true;
this->StackFrameMap[offset + j].IndexedAccess |= IndexedAccess;
if (!UsedFramePointer)
this->StackFrameMap[offset + j].ESPRelativeAccess = true;
else
this->StackFrameMap[offset + j].EBPRelativeAccess = true;
}
BitWidthMask = ComputeOperandBitWidthMask(TempOp, DataSize);
this->FineGrainedStackTable.at(offset).SizeInfo |= BitWidthMask;
this->FineGrainedStackTable.at(offset).SignMiscInfo |= FG_MASK_READ;
if (IndexedAccess) {
this->FineGrainedStackTable.at(offset).SignMiscInfo |= FG_MASK_INDEXED_ACCESS;
}
if (!UsedFramePointer) {
this->FineGrainedStackTable.at(offset).SignMiscInfo |= FG_MASK_SP_RELATIVE;
}
else {
this->FineGrainedStackTable.at(offset).SignMiscInfo |= FG_MASK_FP_RELATIVE;
}
if (SignedMove) {
this->FineGrainedStackTable.at(offset).SignMiscInfo |= FG_MASK_SIGNED;
}
else if (UnsignedMove) {
this->FineGrainedStackTable.at(offset).SignMiscInfo |= FG_MASK_UNSIGNED;
}
} // end if MDGetStackOffsetAndSize()
} // end if SourceMemoryOperand
// NOTE: Detect taking the address of stack locations. **!!**
} // end for all instructions
// If function is a leaf function, set OutgoingArgsSize to zero and return.
clc5q
committed
if (this->IsLeaf() && !(this->IsDirectlyRecursive())) {
this->OutgoingArgsSize = 0;
return;
}
// For non-leaf functions, set the OutgoingArgsSize to the write-only, ESP-relative
// region of the bottom of the StackFrameMap.
bool OutgoingArgsRegionFinished = false;
bool IndexedOutgoingArgs = false; // Any indexed accesses to outgoing args?
size_t FramePadSize = 0;
for (size_t MapIndex = 0; MapIndex < this->StackFrameMap.size(); ++MapIndex) {
// Some of the bottom of the stack frame might be below the local frame allocation.
// These are pushes that happened after allocation, etc. We skip over these
// locations and define the outgoing args region to start strictly at the bottom