[NewPM] Introduce PreserveCFG check

[yrouban]

Summary

Check that all passes, which report they preserve CFG, are really preserving CFG.
A new standard instrumentation is introduced. It can be switched on/off by the flag verify-cfg-preserved, which is on by default for debug builds.

[fedor.sergeev]

I'm not sure if introducing base for standard instrumentation classes (Instrument/Instruments) is a good change or not.
What problem does it solve?
Anyway that part seems to be rather orthogonal to the checker itself.
I would like to have it separated - either introduce Instrument/Instruments thing before the checker and then use it,
OR implement the checker the same way as other standard instrumentations are done (just as separate members) and only then refactor it all into Instrument/Instruments.

[yrouban]

In D81558#2085042, @fedor.sergeev wrote:

I'm not sure if introducing base for standard instrumentation classes (Instrument/Instruments) is a good change or not.
What problem does it solve?
Anyway that part seems to be rather orthogonal to the checker itself.
I would like to have it separated - either introduce Instrument/Instruments thing before the checker and then use it,
OR implement the checker the same way as other standard instrumentations are done (just as separate members) and only then refactor it all into Instrument/Instruments.

Ok.
I do not like that the current design makes me expose of a new standard instrumentation class declaration for all users of StandardInstrumentations. I believe it can be done internally inside StandardInstrumentations.cpp. That is why I proposed the Instrument base class.

I'm still not sure that this patch is needed. So, please feel free to comment. I used this patch once to see if there are any PreservedCFG violating passes except InstCombine. I did not find any. This does not prove there is none, but some level of confidence it gives.

llvm/lib/Passes/StandardInstrumentations.cpp

268

I realized that it is unsafe to dereference any of the saved BasicBlock* at the validation check.
But derefrence is only needed to print the difference when we are goin to crash anyway.
Another problem is that we might think that there are no difference if CFG has changed but BasicBlock* happened to be the same. It seems this patch needs to track values.

[yrouban]

Removed the class Instrument and moved the class declaration of PreservedCFGCheckerInstrumentation from cpp to h file.
Relaxed CFG check for successors order (to allow branch successors swap).
Made CFG diff printing safe for possible block removal.
Fixed push-pop CFG to/from stack and a parallel stack with pass names for assertions.
Rebased.

[yrouban]

Rebased on top of D84772. So D84826 is not needed.
Fixed safety of printCFGdiff() to handle removed blocks.

[kuhar]

I think this deserves some extra documentation that explains what it means to preserve CFG in plain English. For example, should we also care about preserving the number of return blocks?

llvm/include/llvm/Passes/StandardInstrumentations.h

77

CFG is a multigraph. A basic block can have duplicate successors, i.e., there may be many outgoing edges to the same basic block.
Shouldn't this be a map<BB, set<pair<BB, count>> or map<BB, multiset<BB, count>>?

77

Is tracking BasicBlock* enough? What if blocks get deleted and the tracked pointers become invalid? Does it make sense to use WeakVH instead?

[yrouban]

llvm/include/llvm/Passes/StandardInstrumentations.h

77

Ok. I had in mind checking PostDomTree invariant that must be insensitive to parallel edges.
I will try to make it stricter as you suggest.

77

To not complicate things with WeakVH I decided to not dereference the blocks collected before the checked pass unless I see the same pointer in the CFG after the pass. See the printCFGdiff().
WeakVH would allow to check accurately, but could not be used as a set/map key.

[kuhar]

llvm/include/llvm/Passes/StandardInstrumentations.h

77

AFAIK in C++, when pointee's lifetime ends, the pointer becomes invalid and you cannot use it in any meaningful way, even when you don't dereference it. It'd be incorrect to print it, compare with another pointers, etc. This would also make debugging very confusing.

[yrouban]

Defined a separate class named CFG.
CFG defines the notion PreserveCFG for the multigraph of BasicBlocks with unordered successors.
CFG can safely track and print CFG changes. To guard CFG from accessing deleted blocks PoisoningVH is used and is extended with one getter.
The cfg diff printing function is rewritten.
Rebased.

Here is an example of an output of a unexpected CFG change detected after SimplifyCFG mistakenly reported PreservedCFG.

$ opt -passes=simplifycfg -S -verify-cfg-preserved --debug-only=stdinstrumentations

before SimplifyCFG

define i32 @test5(i32 %A) {
  switch i32 %A, label %return [
    i32 2, label %return
    i32 10, label %2
    i32 9, label %1
    i32 8, label %2
  ]

  ret i32 1

  ret i32 2

return:
  ret i32 0
}

after SimplifyCFG

define i32 @test5(i32 %A) {
  switch i32 %A, label %return [
    i32 8, label %2
    i32 10, label %2
    i32 9, label %1
  ]

1:                                                ; preds = %return, %2, %0
  %merge = phi i32 [ 1, %0 ], [ 2, %2 ], [ 0, %return ]
  ret i32 %merge

2:                                                ; preds = %0, %0
  br label %1

return:                                           ; preds = %0
  br label %1
}

Mocked error

Error: SimplifyCFGPass reported it preserved CFG, but changes detected:
In function @test5
Different number of non-leaf basic blocks: before=1, after=3
Non-leaf block return<0x23931c0> is added (1 successors)
Different successors of block entry<0x2393130> (unordered):
- before (3): return<0x23931c0>(2), unnamed_1<0x2393990>, unnamed_2<0x23938d0>(2),
- after (3): return<0x23931c0>, unnamed_1<0x2393990>, unnamed_2<0x23938d0>(2),
Non-leaf block unnamed_2<0x23938d0> is added (1 successors)
LLVM ERROR: Preserved CFG changed by SimplifyCFGPass

[kuhar]

This looks very useful. Does it currently report any errors when enabled?

llvm/include/llvm/Passes/StandardInstrumentations.h

86

nit: NeedsGruard sounds like an external property that you need to satisfy for some reason. In this case, it seems to me that it's just an extra level of validation that we may want, right?
If that's the case, I'd consider renaming this to something like: UseGuargs, GuardBasicBlocks, TrackBBLifetime, etc.

llvm/lib/Passes/StandardInstrumentations.cpp

42

How much overhead does this have? It it's not negligible, we may want to enable it when expensive checks are enabled.

336

nit: !empty() or size() > 0

338

same as above

[yrouban]

In D81558#2249775, @kuhar wrote:

This looks very useful. Does it currently report any errors when enabled?

None.

llvm/lib/Passes/StandardInstrumentations.cpp

42

By default I proposed to enable it only for debug version. Do we track compile time for them?
I suggest that we leave this as is with additional comment to enable it for the expansive checks only if found a performance impact.
Basically for every function pass it runs twice over function blocks and edges and allocates linear memory. Should not be very big.

[kuhar]

llvm/lib/Passes/StandardInstrumentations.cpp

42

In general, I think we do care about it too. Could you run opt on some bigger .bc file to make sure there's little noticeable difference? I posted some .bc files in https://reviews.llvm.org/D83089.

[kuhar]

llvm/include/llvm/Passes/StandardInstrumentations.h

70–71

nit: could you move these just next to the forward-declaration for Module?

[yrouban]

I will collect performance data and fix for release builds.

llvm/include/llvm/IR/ValueHandle.h

506 ↗

(On Diff #289100)

bug: Poisoned is defined only for debug builds. I will define a custom CallbackVH extension.

[yrouban]

Implemented custom BBGuard class instead of PoisoningVH which does not track poisoning in release builds.
Removed DEBUG_TYPE and LLVM_DEBUG uses. Unexpected gaph diffs are printed unconditionally.
Addressed comments.
I will provide performance data soon.

[yrouban]

a minor fix in BBGuard.

[kuhar]

llvm/include/llvm/Passes/StandardInstrumentations.h

80

nit: s/struct/class?

82

nit: is this function used?

[yrouban]

I see no impact on CPU time with the following timed runs (both with and without assertions):
$ time opt -O3 FILE.bc --disable-output -verify-cfg-preserved=VCP

LLVM_ENABLE_ASSERTIONS=ON:

FILE	verify-cfg-preserved	Run 1 [s]	Run 2 [s]	Run 3 [s]	Run 4 [s]	Run 5 [s]	Avg	Std/Avg %	DeltaAvg%
llvm-as.bc	false	549.010	544.270	550.010	547.620	547.360	547.654	0.40
	true	548.640	547.670	547.690	548.900	548.680	548.316	0.11	0.12
sqlite3.bc	false	58.120	58.180	58.210	58.130	58.190	58.166	0.07
	true	58.230	58.120	58.130	58.240	58.210	58.186	0.10	0.03
wasm-as.bc	false	168.880	169.490	168.610	169.770	168.150	168.980	0.39
	true	168.660	169.730	169.360	169.620	168.500	169.174	0.33	0.11
wasm-opt.bc	false	181.860	180.550	181.340	180.790	181.910	181.290	0.34
	true	181.240	181.620	181.180	181.780	180.640	181.292	0.24	0.00

LLVM_ENABLE_ASSERTIONS=OFF:

FILE	verify-cfg-preserved	Run 1 [s]	Run 2 [s]	Run 3 [s]	Run 4 [s]	Run 5 [s]	Avg	Std/Avg %	DeltaAvg%
llvm-as.bc	false	408.650	408.730	410.100	406.650	410.110	408.848	0.35
	true	408.980	408.260	408.410	409.180	408.600	408.686	0.09	-0.04
sqlite3.bc	false	39.720	39.770	39.790	39.720	39.780	39.756	0.08
	true	39.690	39.730	39.740	39.760	39.750	39.734	0.07	-0.06
wasm-as.bc	false	117.270	117.460	117.450	117.410	117.250	117.368	0.09
	true	117.640	117.100	117.400	117.210	117.310	117.332	0.18	-0.03
wasm-opt.bc	false	125.900	126.030	125.910	126.020	126.180	126.008	0.09
	true	125.940	125.930	126.100	125.980	125.920	125.974	0.06	-0.03

[kuhar]

Thanks for the table, this looks good to me now, modulo the two remaining nits. I'd be more comfortable if somebody familiar with the new pass manager reviewed this too.

[yrouban]

In D81558#2085042, @fedor.sergeev wrote:

Fedor, could you please review the New Pass Manager part of this patch as I addressed your comments?

[fedor.sergeev]

NewPM use seems to be fine.
A few relatively minor comments that I would like to see addressed.

llvm/include/llvm/Passes/StandardInstrumentations.h

79

Are we ok with adding a generic 'CFG' class directly into llvm namespace by StandardInstrumentations.h?
If it is only for CFGChecker introduced here then either put it into Checker itself or into a nested namespace named appropriately
(e.g. cfg_checker_detail).

If you can see this class being generally useful then it deserves its own header/location.

110–111

It seems that GraphStack and PassStack are always pushed/popped synchronously.
Why dont you keep a single vector of (StringRef, Optional<CFG>) pair to make this couple more obvious?

llvm/lib/Passes/StandardInstrumentations.cpp

428

This seems to be a more new-PM-idiomatic way of checking:

PassPA.allAnalysesInSetPreserved<CFGAnalyses>()

[yrouban]

llvm/include/llvm/Passes/StandardInstrumentations.h

79

That is why I tried introducing the class Instrument that would encapsulate all declarations and implementations of PrintIRInstrumentation, PrintPassInstrumentation, TimePassesHandler, OptNoneInstrumentation, PreservedCFGCheckerInstrumentation in one cpp file :)

[yrouban]

Addressed Fedor's comments.

[fedor.sergeev]

LGTM

llvm/include/llvm/Passes/StandardInstrumentations.h

 /// This header defines a class that provides bookkeeping for all standard
 /// (i.e in-tree) pass instrumentations.
 ///
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_PASSES_STANDARDINSTRUMENTATIONS_H
 #define LLVM_PASSES_STANDARDINSTRUMENTATIONS_H
 
+#include "llvm/ADT/SmallSet.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/IR/PassInstrumentation.h"
 #include "llvm/IR/PassTimingInfo.h"
+#include "llvm/Support/CommandLine.h"
 
 #include <string>
 #include <utility>
 
 namespace llvm {
 
 class Module;
 
public:
   void registerCallbacks(PassInstrumentationCallbacks &PIC);
 
 private:
   bool skip(StringRef PassID, Any IR);
 
   bool DebugLogging;
 };
 
+class BasicBlock;
+class Function;
+
+class PreservedCFGCheckerInstrumentation {
+public:
+  // CFG is a map from BasicBlock to a set of its successors. It is important
+  // that the set is unordered because order of succesors may be changed.
+  using CFG = DenseMap<const BasicBlock *, SmallSet<const BasicBlock *, 4>>;
+
+private:
+  SmallVector<Optional<CFG>, 8> GraphStackBefore;
+  SmallVector<StringRef, 8> PassStack;
+
+  CFG buildGraph(const Function *F);
+  void printCFGdiff(const CFG &Before, const CFG &After) const;
+
+public:
+  static cl::opt<bool> VerifyPreservedCFG;
+  void registerCallbacks(PassInstrumentationCallbacks &PIC);
+};
+
 /// This class provides an interface to register all the standard pass
 /// instrumentations and manages their state (if any).
 class StandardInstrumentations {
   PrintIRInstrumentation PrintIR;
   TimePassesHandler TimePasses;
   OptNoneInstrumentation OptNone;
+  PreservedCFGCheckerInstrumentation PreservedCFGChecker;
 
 public:
   StandardInstrumentations(bool DebugLogging)
       : PrintIR(), TimePasses(), OptNone(DebugLogging) {}
 
   void registerCallbacks(PassInstrumentationCallbacks &PIC);
 
   TimePassesHandler &getTimePasses() { return TimePasses; }
 };
 } // namespace llvm
 
 #endif

llvm/lib/Passes/StandardInstrumentations.cpp

 #include "llvm/IR/PassInstrumentation.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/FormatVariadic.h"
 #include "llvm/Support/raw_ostream.h"
 
 using namespace llvm;
 
+#define DEBUG_TYPE "stdinstrumentations"
+
 // TODO: remove once all required passes are marked as such.
 static cl::opt<bool>
     EnableOptnone("enable-npm-optnone", cl::init(false),
                   cl::desc("Enable skipping optional passes optnone functions "
                            "under new pass manager"));
 
+cl::opt<bool> PreservedCFGCheckerInstrumentation::VerifyPreservedCFG(
+    "verify-cfg-preserved", cl::Hidden,
+#ifdef NDEBUG
+    cl::init(false));
+#else
+    cl::init(true));
+#endif
+
 namespace {
 
 /// Extracting Module out of \p IR unit. Also fills a textual description
 /// of \p IR for use in header when printing.
 Optional<std::pair<const Module *, std::string>> unwrapModule(Any IR) {
   if (any_isa<const Module *>(IR))
     return std::make_pair(any_cast<const Module *>(IR), std::string());
 
PIC.registerAfterPassCallback(
         });
     PIC.registerAfterPassInvalidatedCallback(
         [this](StringRef P, const PreservedAnalyses &) {
           this->printAfterPassInvalidated(P);
         });
   }
 }
 
 void OptNoneInstrumentation::registerCallbacks(
     PassInstrumentationCallbacks &PIC) {
   PIC.registerBeforePassCallback(
       [this](StringRef P, Any IR) { return this->skip(P, IR); });
 }
 
 bool OptNoneInstrumentation::skip(StringRef PassID, Any IR) {
   if (!EnableOptnone)
     return true;
   const Function *F = nullptr;
   if (any_isa<const Function *>(IR)) {
     F = any_cast<const Function *>(IR);
   } else if (any_isa<const Loop *>(IR)) {
     F = any_cast<const Loop *>(IR)->getHeader()->getParent();
   }
   if (F && F->hasOptNone()) {
     if (DebugLogging)
       dbgs() << "Skipping pass: " << PassID << " (optnone)\n";
     return false;
   }
   return true;
 }
 
+PreservedCFGCheckerInstrumentation::CFG
+PreservedCFGCheckerInstrumentation::buildGraph(const Function *F) {
+  CFG Result;
+  for (const auto &BB : *F)
+    Result[&BB].insert(succ_begin(&BB), succ_end(&BB));
+  return Result;
+}
+
+void PreservedCFGCheckerInstrumentation::printCFGdiff(
+    const PreservedCFGCheckerInstrumentation::CFG &Before,
+    const PreservedCFGCheckerInstrumentation::CFG &After) const {
+  if (Before.size() != After.size())
+    dbgs() << "Different number of basic blocks: before=" << Before.size()
+           << ", after=" << After.size() << "\n";
+  for (auto &BB : Before) {
+    auto BA = After.find(BB.first);
+    if (BA == After.end())
+      // Must not access BB as it is likely to be deleted. Just print its ptr.
+      dbgs() << "Block @" << BB.first << " is removed\n";
+  }
+
+  for (auto &BA : After) {
+    auto BB = Before.find(BA.first);
+    if (BB == Before.end()) {
+      dbgs() << "Block " << BA.first->getName() << " is added\n";
+      continue;
+    }
+
+    if (BB->second == BA.second)
+      continue;
+
+    dbgs() << "Different successors of block " << BA.first->getName() << ":\n";
+    dbgs() << "- before: ";
+    for (auto Succ : BB->second)
+      dbgs() << Succ->getName() << ", ";
+    dbgs() << "\n";
+    dbgs() << "-  after: ";
+    for (auto Succ : BA.second)
+      dbgs() << Succ->getName() << ", ";
+    dbgs() << "\n";
+  }
+}
+
+void PreservedCFGCheckerInstrumentation::registerCallbacks(
+    PassInstrumentationCallbacks &PIC) {
+  if (!VerifyPreservedCFG)
+    return;
+
+  PIC.registerBeforePassCallback([this](StringRef P, Any IR) {
+    PassStack.emplace_back(P);
+    if (any_isa<const Function *>(IR))
+      GraphStackBefore.emplace_back(buildGraph(any_cast<const Function *>(IR)));
+    else
+      GraphStackBefore.emplace_back(None);
+    return true;
+  });
+
+  PIC.registerAfterPassSkippedCallback([this](StringRef P, Any IR) {
+    (void)GraphStackBefore.pop_back_val();
+    auto PassName = PassStack.pop_back_val();
+    assert(PassName == P && "Before and After callbacks must correspond");
+  });
+
+  PIC.registerAfterPassInvalidatedCallback(
+      [this](StringRef P, const PreservedAnalyses &PassPA) {
+        (void)GraphStackBefore.pop_back_val();
+        auto PassName = PassStack.pop_back_val();
+        assert(PassName == P && "Before and After callbacks must correspond");
+      });
+
+  PIC.registerAfterPassCallback([this](StringRef P, Any IR,
+                                       const PreservedAnalyses &PassPA) {
+    auto PassName = PassStack.pop_back_val();
+    auto GraphBefore = GraphStackBefore.pop_back_val();
+    assert(PassName == P && "Before and After callbacks must correspond");
+
+    if (!PassPA.allAnalysesInSetPreserved(CFGAnalyses::ID()))
+      return;
+
+    if (any_isa<const Function *>(IR)) {
+      assert(GraphBefore && "Must be built in BeforePassCallback");
+      CFG GraphAfter = buildGraph(any_cast<const Function *>(IR));
+      if (GraphAfter == *GraphBefore)
+        return;
+
+      LLVM_DEBUG(
+          dbgs() << "Error: " << P
+                 << " reported it preserved CFG, but a change is detected:\n");
+      LLVM_DEBUG(printCFGdiff(*GraphBefore, GraphAfter));
+      report_fatal_error(Twine("Preserved CFG changed by ", P));
+    }
+  });
+}
+
 void StandardInstrumentations::registerCallbacks(
     PassInstrumentationCallbacks &PIC) {
   PrintIR.registerCallbacks(PIC);
   TimePasses.registerCallbacks(PIC);
   OptNone.registerCallbacks(PIC);
+  PreservedCFGChecker.registerCallbacks(PIC);
 }