/// <summary>
/// Processes procedures individually, building complex expression
/// trees out of the simple, close-to-the-machine code generated by
/// the disassembly.
/// </summary>
/// <param name="rl"></param>
public void BuildExpressionTrees()
{
eventListener.ShowProgress("Building expressions.", 0, program.Procedures.Count);
foreach (var sst in this.ssts)
{
var ssa = sst.SsaState;
try
{
DumpWatchedProcedure("Before expression coalescing", ssa.Procedure);
// Procedures should be untangled from each other. Now process
// each one separately.
DeadCode.Eliminate(ssa);
// Build expressions. A definition with a single use can be subsumed
// into the using expression.
var coa = new Coalescer(ssa);
coa.Transform();
DeadCode.Eliminate(ssa);
var vp = new ValuePropagator(program.SegmentMap, ssa, program.CallGraph, dynamicLinker, eventListener);
vp.Transform();
DumpWatchedProcedure("After expression coalescing", ssa.Procedure);
var liv = new LinearInductionVariableFinder(
ssa,
new BlockDominatorGraph(
ssa.Procedure.ControlGraph,
ssa.Procedure.EntryBlock));
liv.Find();
foreach (var de in liv.Contexts)
{
var str = new StrengthReduction(ssa, de.Key, de.Value);
str.ClassifyUses();
str.ModifyUses();
}
DeadCode.Eliminate(ssa);
DumpWatchedProcedure("After strength reduction", ssa.Procedure);
// Definitions with multiple uses and variables joined by PHI functions become webs.
var web = new WebBuilder(program, ssa, program.InductionVariables, eventListener);
web.Transform();
ssa.ConvertBack(false);
DumpWatchedProcedure("After data flow analysis", ssa.Procedure);
}
catch (Exception ex)
{
eventListener.Error(
eventListener.CreateProcedureNavigator(program, ssa.Procedure),
ex,
"An internal error occurred while building the expressions of {0}",
ssa.Procedure.Name);
}
eventListener.Advance(1);
}
}
/// <summary>
/// This callback is called from the SccFinder, which passes it a list
/// of Procedures that form a SCC.
/// </summary>
/// <param name="procs"></param>
private void UntangleProcedureScc(IList <Procedure> procs)
{
this.sccProcs = procs.ToHashSet();
flow.CreateFlowsFor(procs);
// Convert all procedures in the SCC to SSA form and perform
// value propagation.
var ssts = procs.Select(ConvertToSsa).ToArray();
this.ssts.AddRange(ssts);
DumpWatchedProcedure("After extra stack vars", ssts);
// At this point, the computation of ProcedureFlow is possible.
var trf = new TrashedRegisterFinder(program, flow, ssts, this.eventListener);
trf.Compute();
// New stack based variables may be available now.
foreach (var sst in ssts)
{
var vp = new ValuePropagator(program.SegmentMap, sst.SsaState, program.CallGraph, dynamicLinker, this.eventListener);
vp.Transform();
sst.RenameFrameAccesses = true;
sst.Transform();
DumpWatchedProcedure("After extra stack vars", sst.SsaState.Procedure);
}
foreach (var ssa in ssts.Select(sst => sst.SsaState))
{
RemoveImplicitRegistersFromHellNodes(ssa);
var sac = new SegmentedAccessClassifier(ssa);
sac.Classify();
var prj = new ProjectionPropagator(ssa, sac);
prj.Transform();
DumpWatchedProcedure("After projection propagation", ssa.Procedure);
}
var uid = new UsedRegisterFinder(flow, procs, this.eventListener);
foreach (var sst in ssts)
{
var ssa = sst.SsaState;
RemovePreservedUseInstructions(ssa);
DeadCode.Eliminate(ssa);
uid.ComputeLiveIn(ssa, true);
var procFlow = flow[ssa.Procedure];
RemoveDeadArgumentsFromCalls(ssa.Procedure, procFlow, ssts);
DumpWatchedProcedure("After dead call argument removal", ssa.Procedure);
}
eventListener.Advance(procs.Count);
}
public void Transform()
{
CollectLiveOutStorages();
DumpLiveOut();
bool change;
do
{
if (eventListener.IsCanceled())
{
return;
}
change = false;
this.wl.AddRange(ssaStates);
while (wl.TryGetWorkItem(out SsaState ssa))
{
if (this.eventListener.IsCanceled())
{
return;
}
var vp = new ValuePropagator(program.SegmentMap, ssa, program.CallGraph, dynamicLinker, eventListener);
vp.Transform();
change |= RemoveUnusedDefinedValues(ssa, wl);
//DataFlowAnalysis.DumpWatchedProcedure("After RemoveUnusedDefinedValues", ssa.Procedure);
change |= RemoveLiveInStorages(ssa.Procedure, dataFlow[ssa.Procedure], wl);
//DataFlowAnalysis.DumpWatchedProcedure("After RemoveLiveInStorages", ssa.Procedure);
}
} while (change);
foreach (var proc in procToSsa.Keys)
{
var liveOut = CollectLiveOutStorages(proc);
var flow = this.dataFlow[proc];
flow.BitsLiveOut = SummarizeStorageBitranges(flow.BitsLiveOut.Concat(liveOut));
flow.grfLiveOut = SummarizeFlagGroups(liveOut);
}
}
/// <summary>
/// Processes procedures individually, building complex expression trees out
/// of the simple, close-to-the-machine code generated by the disassembly.
/// </summary>
/// <param name="rl"></param>
public void BuildExpressionTrees()
{
int i = 0;
foreach (Procedure proc in program.Procedures.Values)
{
eventListener.ShowProgress("Building complex expressions.", i, program.Procedures.Values.Count);
++i;
try
{
var larw = new LongAddRewriter(proc, program.Architecture);
larw.Transform();
Aliases alias = new Aliases(proc, program.Architecture, flow);
alias.Transform();
var doms = new DominatorGraph <Block>(proc.ControlGraph, proc.EntryBlock);
var sst = new SsaTransform(flow, proc, importResolver, doms);
var ssa = sst.SsaState;
var cce = new ConditionCodeEliminator(ssa.Identifiers, program.Platform);
cce.Transform();
//var cd = new ConstDivisionImplementedByMultiplication(ssa);
//cd.Transform();
DeadCode.Eliminate(proc, ssa);
var vp = new ValuePropagator(program.Architecture, ssa.Identifiers, proc);
vp.Transform();
DeadCode.Eliminate(proc, ssa);
// Build expressions. A definition with a single use can be subsumed
// into the using expression.
var coa = new Coalescer(proc, ssa);
coa.Transform();
DeadCode.Eliminate(proc, ssa);
var liv = new LinearInductionVariableFinder(
proc,
ssa.Identifiers,
new BlockDominatorGraph(proc.ControlGraph, proc.EntryBlock));
liv.Find();
foreach (KeyValuePair <LinearInductionVariable, LinearInductionVariableContext> de in liv.Contexts)
{
var str = new StrengthReduction(ssa, de.Key, de.Value);
str.ClassifyUses();
str.ModifyUses();
}
var opt = new OutParameterTransformer(proc, ssa.Identifiers);
opt.Transform();
DeadCode.Eliminate(proc, ssa);
// Definitions with multiple uses and variables joined by PHI functions become webs.
var web = new WebBuilder(proc, ssa.Identifiers, program.InductionVariables);
web.Transform();
ssa.ConvertBack(false);
}
catch (StatementCorrelatedException stex)
{
eventListener.Error(
eventListener.CreateBlockNavigator(program, stex.Statement.Block),
stex,
"An error occurred during data flow analysis.");
}
catch (Exception ex)
{
eventListener.Error(
new NullCodeLocation(proc.Name),
ex,
"An error occurred during data flow analysis.");
}
}
}
private void UntangleProcedureScc(IList <Procedure> procs)
{
if (procs.Count == 1)
{
var proc = procs[0];
Aliases alias = new Aliases(proc, program.Architecture, flow);
alias.Transform();
// Transform the procedure to SSA state. When encountering 'call' instructions,
// they can be to functions already visited. If so, they have a "ProcedureFlow"
// associated with them. If they have not been visited, or are computed destinations
// (e.g. vtables) they will have no "ProcedureFlow" associated with them yet, in
// which case the the SSA treats the call as a "hell node".
var doms = proc.CreateBlockDominatorGraph();
var sst = new SsaTransform(flow, proc, importResolver, doms);
var ssa = sst.SsaState;
// Propagate condition codes and registers. At the end, the hope is that
// all statements like (x86) mem[esp_42+4] will have been converted to
// mem[fp - 30]. We also hope that procedure constants kept in registers
// are propagated to the corresponding call sites.
var cce = new ConditionCodeEliminator(ssa.Identifiers, program.Platform);
cce.Transform();
var vp = new ValuePropagator(program.Architecture, ssa.Identifiers, proc);
vp.Transform();
// Now compute SSA for the stack-based variables as well. That is:
// mem[fp - 30] becomes wLoc30, while
// mem[fp + 30] becomes wArg30.
// This allows us to compute the dataflow of this procedure.
sst.RenameFrameAccesses = true;
sst.AddUseInstructions = true;
sst.Transform();
// Propagate those newly discovered identifiers.
vp.Transform();
// At this point, the computation of _actual_ ProcedureFlow should be possible.
var tid = new TrashedRegisterFinder2(program.Architecture, flow, proc, ssa.Identifiers, this.eventListener);
tid.Compute();
DeadCode.Eliminate(proc, ssa);
// Build expressions. A definition with a single use can be subsumed
// into the using expression.
var coa = new Coalescer(proc, ssa);
coa.Transform();
DeadCode.Eliminate(proc, ssa);
var liv = new LinearInductionVariableFinder(
proc,
ssa.Identifiers,
new BlockDominatorGraph(proc.ControlGraph, proc.EntryBlock));
liv.Find();
foreach (var de in liv.Contexts)
{
var str = new StrengthReduction(ssa, de.Key, de.Value);
str.ClassifyUses();
str.ModifyUses();
}
//var opt = new OutParameterTransformer(proc, ssa.Identifiers);
//opt.Transform();
DeadCode.Eliminate(proc, ssa);
// Definitions with multiple uses and variables joined by PHI functions become webs.
var web = new WebBuilder(proc, ssa.Identifiers, program.InductionVariables);
web.Transform();
ssa.ConvertBack(false);
}
else
{
throw new NotImplementedException();
}
}
/// <summary>
/// Converts all registers and stack accesses to SSA variables.
/// </summary>
/// <param name="proc"></param>
/// <returns>The SsaTransform for the procedure.</returns>
public SsaTransform ConvertToSsa(Procedure proc)
{
if (program.NeedsSsaTransform)
{
// Transform the procedure to SSA state. When encountering 'call'
// instructions, they can be to functions already visited. If so,
// they have a "ProcedureFlow" associated with them. If they have
// not been visited, or are computed destinations (e.g. vtables)
// they will have no "ProcedureFlow" associated with them yet, in
// which case the the SSA treats the call as a "hell node".
var sst = new SsaTransform(program, proc, sccProcs, dynamicLinker, this.ProgramDataFlow);
var ssa = sst.Transform();
DumpWatchedProcedure("After SSA", ssa.Procedure);
// Merge unaligned memory accesses.
var fuser = new UnalignedMemoryAccessFuser(ssa);
fuser.Transform();
// After value propagation expressions like (x86)
// mem[esp_42+4] will have been converted to mem[fp - 30].
// We also hope that procedure constants
// kept in registers are propagated to the corresponding call
// sites.
var vp = new ValuePropagator(program.SegmentMap, ssa, program.CallGraph, dynamicLinker, eventListener);
vp.Transform();
DumpWatchedProcedure("After first VP", ssa.Procedure);
// Fuse additions and subtractions that are linked by the carry flag.
var larw = new LongAddRewriter(ssa);
larw.Transform();
// Propagate condition codes and registers.
var cce = new ConditionCodeEliminator(ssa, program.Platform);
cce.Transform();
vp.Transform();
DumpWatchedProcedure("After CCE", ssa.Procedure);
// Now compute SSA for the stack-based variables as well. That is:
// mem[fp - 30] becomes wLoc30, while
// mem[fp + 30] becomes wArg30.
// This allows us to compute the dataflow of this procedure.
sst.RenameFrameAccesses = true;
sst.Transform();
DumpWatchedProcedure("After SSA frame accesses", ssa.Procedure);
var icrw = new IndirectCallRewriter(program, ssa, eventListener);
while (!eventListener.IsCanceled() && icrw.Rewrite())
{
vp.Transform();
sst.RenameFrameAccesses = true;
sst.Transform();
}
var fpuGuesser = new FpuStackReturnGuesser(ssa);
fpuGuesser.Rewrite();
// By placing use statements in the exit block, we will collect
// reaching definitions in the use statements.
sst.AddUsesToExitBlock();
sst.RemoveDeadSsaIdentifiers();
// Backpropagate stack pointer from procedure return.
var spBackpropagator = new StackPointerBackpropagator(ssa);
spBackpropagator.BackpropagateStackPointer();
DumpWatchedProcedure("After SP BP", ssa.Procedure);
// Propagate those newly created stack-based identifiers.
vp.Transform();
DumpWatchedProcedure("After VP2", ssa.Procedure);
return(sst);
}
else
{
// We are assuming phi functions are already generated.
var sst = new SsaTransform(program, proc, sccProcs, dynamicLinker, this.ProgramDataFlow);
return(sst);
}
}
/// <summary>
/// Processes procedures individually, building complex expression trees out
/// of the simple, close-to-the-machine code generated by the disassembly.
/// </summary>
/// <param name="rl"></param>
public void BuildExpressionTrees()
{
int i = 0;
foreach (Procedure proc in program.Procedures.Values)
{
if (eventListener.IsCanceled())
{
break;
}
eventListener.ShowProgress("Building complex expressions.", i, program.Procedures.Values.Count);
++i;
try
{
var sst = BuildSsaTransform(proc);
var ssa = sst.SsaState;
var fuser = new UnalignedMemoryAccessFuser(ssa);
fuser.Transform();
var vp = new ValuePropagator(program.SegmentMap, ssa, importResolver, eventListener);
sst.RenameFrameAccesses = true;
var icrw = new IndirectCallRewriter(program, ssa, eventListener);
while (!eventListener.IsCanceled() && icrw.Rewrite())
{
vp.Transform();
sst.Transform();
}
var cce = new ConditionCodeEliminator(ssa, program.Platform);
cce.Transform();
//var cd = new ConstDivisionImplementedByMultiplication(ssa);
//cd.Transform();
DeadCode.Eliminate(proc, ssa);
vp.Transform();
DeadCode.Eliminate(proc, ssa);
// Build expressions. A definition with a single use can be subsumed
// into the using expression.
var coa = new Coalescer(proc, ssa);
coa.Transform();
DeadCode.Eliminate(proc, ssa);
vp.Transform();
var liv = new LinearInductionVariableFinder(
proc,
ssa.Identifiers,
new BlockDominatorGraph(proc.ControlGraph, proc.EntryBlock));
liv.Find();
foreach (KeyValuePair <LinearInductionVariable, LinearInductionVariableContext> de in liv.Contexts)
{
var str = new StrengthReduction(ssa, de.Key, de.Value);
str.ClassifyUses();
str.ModifyUses();
}
var opt = new OutParameterTransformer(proc, ssa.Identifiers);
opt.Transform();
DeadCode.Eliminate(proc, ssa);
// Definitions with multiple uses and variables joined by PHI functions become webs.
var web = new WebBuilder(proc, ssa.Identifiers, program.InductionVariables);
web.Transform();
ssa.ConvertBack(false);
}
catch (StatementCorrelatedException stex)
{
eventListener.Error(
eventListener.CreateStatementNavigator(program, stex.Statement),
stex,
"An error occurred during data flow analysis.");
}
catch (Exception ex)
{
eventListener.Error(
new NullCodeLocation(proc.Name),
ex,
"An error occurred during data flow analysis.");
}
}
}
/// <summary>
/// Processes procedures individually, building complex expression
/// trees out of the simple, close-to-the-machine code generated by
/// the disassembly.
/// </summary>
/// <param name="rl"></param>
public void BuildExpressionTrees()
{
eventListener.ShowProgress("Building expressions.", 0, program.Procedures.Count);
foreach (var sst in this.ssts !)
{
var ssa = sst.SsaState;
try
{
if (program.User.AggressiveBranchRemoval)
{
// This ends up being very aggressive and doesn't replicate the original
// binary code. See discussion on https://github.com/uxmal/reko/issues/932
DumpWatchedProcedure("urb", "Before unreachable block removal", ssa.Procedure);
var urb = new UnreachableBlockRemover(ssa, eventListener);
urb.Transform();
}
DumpWatchedProcedure("precoa", "Before expression coalescing", ssa.Procedure);
// Procedures should be untangled from each other. Now process
// each one separately.
DeadCode.Eliminate(ssa);
// Build expressions. A definition with a single use can be subsumed
// into the using expression.
var coa = new Coalescer(ssa);
coa.Transform();
DeadCode.Eliminate(ssa);
var vp = new ValuePropagator(program.SegmentMap, ssa, program.CallGraph, dynamicLinker, eventListener);
vp.Transform();
DumpWatchedProcedure("postcoa", "After expression coalescing", ssa.Procedure);
var liv = new LinearInductionVariableFinder(
ssa,
new BlockDominatorGraph(
ssa.Procedure.ControlGraph,
ssa.Procedure.EntryBlock));
liv.Find();
foreach (var de in liv.Contexts)
{
var str = new StrengthReduction(ssa, de.Key, de.Value);
str.ClassifyUses();
str.ModifyUses();
}
DeadCode.Eliminate(ssa);
DumpWatchedProcedure("sr", "After strength reduction", ssa.Procedure);
// Definitions with multiple uses and variables joined by PHI functions become webs.
var web = new WebBuilder(program, ssa, program.InductionVariables, eventListener);
web.Transform();
ssa.ConvertBack(false);
DumpWatchedProcedure("dfa", "After data flow analysis", ssa.Procedure);
}
catch (Exception ex)
{
eventListener.Error(
eventListener.CreateProcedureNavigator(program, ssa.Procedure),
ex,
"An internal error occurred while building the expressions of {0}",
ssa.Procedure.Name);
}
eventListener.Advance(1);
}
}
/// <summary>
/// Converts all registers and stack accesses to SSA variables.
/// </summary>
/// <param name="proc"></param>
/// <returns>The SsaTransform for the procedure.</returns>
public SsaTransform ConvertToSsa(Procedure proc)
{
if (!program.NeedsSsaTransform)
{
// Some formats, like LLVM, already have phi functions.
var sst = new SsaTransform(program, proc, sccProcs !, dynamicLinker, this.ProgramDataFlow);
return(sst);
}
try
{
// Transform the procedure to SSA state. When encountering 'call'
// instructions, they can be to functions already visited. If so,
// they have a "ProcedureFlow" associated with them. If they have
// not been visited, or are computed destinations (e.g. vtables)
// they will have no "ProcedureFlow" associated with them yet, in
// which case the the SSA treats the call as a "hell node".
var sst = new SsaTransform(program, proc, sccProcs !, dynamicLinker, this.ProgramDataFlow);
var ssa = sst.Transform();
DumpWatchedProcedure("ssa", "After SSA", ssa);
// Merge unaligned memory accesses.
var fuser = new UnalignedMemoryAccessFuser(ssa);
fuser.Transform();
// Fuse additions and subtractions that are linked by the carry flag.
var larw = new LongAddRewriter(ssa, eventListener);
larw.Transform();
DumpWatchedProcedure("larw", "After long add rewriter", ssa);
// After value propagation expressions like (x86)
// mem[esp_42+4] will have been converted to mem[fp - 30].
// We also hope that procedure constants
// kept in registers are propagated to the corresponding call
// sites.
var vp = new ValuePropagator(program.SegmentMap, ssa, program.CallGraph, dynamicLinker, eventListener);
vp.Transform();
DumpWatchedProcedure("vp", "After first VP", ssa);
// Value propagation may uncover more opportunities.
larw = new LongAddRewriter(ssa, eventListener);
larw.Transform();
DumpWatchedProcedure("larw2", "After second long add rewriter", ssa);
// Eliminate condition codes by discovering uses of ccodes
// and replacing them with higher-level constructs.
var cce = new ConditionCodeEliminator(program, ssa, eventListener);
cce.Transform();
vp.Transform();
DumpWatchedProcedure("cce", "After CCE", ssa);
// Now compute SSA for the stack-based variables as well. That is:
// mem[fp - 30] becomes wLoc30, while
// mem[fp + 30] becomes wArg30.
// This allows us to compute the dataflow of this procedure.
sst.RenameFrameAccesses = true;
sst.Transform();
DumpWatchedProcedure("ssaframe", "After SSA frame accesses", ssa);
var icrw = new IndirectCallRewriter(program, ssa, eventListener);
while (!eventListener.IsCanceled() && icrw.Rewrite())
{
vp.Transform();
sst.RenameFrameAccesses = true;
sst.Transform();
}
var fpuGuesser = new FpuStackReturnGuesser(ssa, eventListener);
fpuGuesser.Transform();
DumpWatchedProcedure("fpug", "After FPU stack guesser", ssa);
// By placing use statements in the exit block, we will collect
// reaching definitions in the use statements.
sst.AddUsesToExitBlock();
sst.RemoveDeadSsaIdentifiers();
// Backpropagate stack pointer from procedure return.
var spBackpropagator = new StackPointerBackpropagator(ssa, eventListener);
spBackpropagator.BackpropagateStackPointer();
DumpWatchedProcedure("spbp", "After SP BP", ssa);
// Propagate those newly created stack-based identifiers.
vp.Transform();
DumpWatchedProcedure("vp2", "After VP2", ssa);
return(sst);
}
catch (Exception ex)
{
var nl = Environment.NewLine;
var banner = $"// {proc.Name} ==========={nl}{ex.Message}{nl}{ex.StackTrace}{nl}{nl}";
services.GetService <ITestGenerationService>()?
.ReportProcedure($"analysis_{99:00}_crash.txt", banner, proc);
throw;
}
}