MSAN: Allow emitting checks for struct types

[guiand]

Summary

This is a precursor to the general eager-checks patch.

[eugenis]

llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp

1155

this can be improved now by moving the origin update code behind the shadow check

1409

extra "ShadowItem ="

1410

I don't think this is possible

1413

I'd prefer if this generated less pointless comparisons.

It would be nice if CreateICmpNE(i1 x, i1 false) automatically folded to "x". I think this is a reasonable change to make in IRBuilder, even though this is probably a very unlikely situation.

As an alternative, make convertShadowToScalar return i1 and remove icmp in the callers (call it something like convertShadowToBool). Check Combiner class how not to emit a "false" constant. As for the Maybe* callers, it's ok to not handle aggregates there.

[guiand]

llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp

1155

This check covers arrays as well, so I don't think I can get rid of it completely but I can change it to check for isArrayType(). I wrote some code to let convertShadowToScalar handle Array types as well, but I thought I would leave that to a separate patch.

[guiand]

Addressed comments

[eugenis]

llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp

1155

Ack.

llvm/test/Instrumentation/MemorySanitizer/check-struct.ll

10

Please check that the test passes when LLVM is build without assertions. AFAIR temp names are lost in that case.

19

There is still this unnecessary icmp.

[guiand]

Don't emit unnecessary icmp

[guiand]

llvm/test/Instrumentation/MemorySanitizer/check-struct.ll

10

As far as I can tell it's passing without assertions enabled, yeah.

[guiand]

Fixed msan_x86_bts_asm test after optimizing away icmp i1 %x, false

[eugenis]

Please upload with context.

[guiand]

Reuploaded with context

[eugenis]

LGTM with 2 comments

llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp

1414

Mention that the result may not be the same bit width as the original type in the comment (this is the usual property of shadow types).

llvm/test/Instrumentation/MemorySanitizer/check-struct.ll

20

could this line say

; CHECK: br i1 [[F1_OR]]

?

[vitalybuka]

llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp

1415

this code do not uses {} for one line blocks

1428

do you see a difference without this check?

1430

this comment is not very useful

[guiand]

llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp

1428

If you look at msan_x86_bts_asm.ll, this check eliminates an unnecessary icmp in some cases.

llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp

     const DataLayout &DL = F.getParent()->getDataLayout();
     const Align OriginAlignment = std::max(kMinOriginAlignment, Alignment);
     unsigned StoreSize = DL.getTypeStoreSize(Shadow->getType());
-    if (Shadow->getType()->isAggregateType()) {
+    if (Shadow->getType()->isArrayTy()) {
       paintOrigin(IRB, updateOrigin(Origin, IRB), OriginPtr, StoreSize,
                   OriginAlignment);
     } else {
-      Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
+      Value *ConvertedShadow = convertShadowToScalar(Shadow, IRB);
       if (auto *ConstantShadow = dyn_cast<Constant>(ConvertedShadow)) {
         if (ClCheckConstantShadow && !ConstantShadow->isZeroValue())
           paintOrigin(IRB, updateOrigin(Origin, IRB), OriginPtr, StoreSize,

                             IRB.CreatePointerCast(Addr, IRB.getInt8PtrTy()),
                             Origin});
       } else {
-        Value *Cmp = IRB.CreateICmpNE(
-            ConvertedShadow, getCleanShadow(ConvertedShadow), "_mscmp");
+        Value *Cmp = convertToBool(ConvertedShadow, IRB, "_mscmp");
         Instruction *CheckTerm = SplitBlockAndInsertIfThen(
             Cmp, &*IRB.GetInsertPoint(), false, MS.OriginStoreWeights);
         IRBuilder<> IRBNew(CheckTerm);

                            bool AsCall) {
     IRBuilder<> IRB(OrigIns);
     LLVM_DEBUG(dbgs() << "  SHAD0 : " << *Shadow << "\n");
-    Value *ConvertedShadow = convertToShadowTyNoVec(Shadow, IRB);
+    Value *ConvertedShadow = convertShadowToScalar(Shadow, IRB);
     LLVM_DEBUG(dbgs() << "  SHAD1 : " << *ConvertedShadow << "\n");
 
     if (auto *ConstantShadow = dyn_cast<Constant>(ConvertedShadow)) {

                                                 ? Origin
                                                 : (Value *)IRB.getInt32(0)});
     } else {
-      Value *Cmp = IRB.CreateICmpNE(ConvertedShadow,
-                                    getCleanShadow(ConvertedShadow), "_mscmp");
+      Value *Cmp = convertToBool(ConvertedShadow, IRB, "_mscmp");
       Instruction *CheckTerm = SplitBlockAndInsertIfThen(
           Cmp, OrigIns,
           /* Unreachable */ !MS.Recover, MS.ColdCallWeights);

     return ty;
   }
 
+  /// Extract combined shadow of struct elements as a bool
+  Value *collapseStructShadow(StructType *Struct, Value *Shadow,
+                              IRBuilder<> &IRB) {
+    Value *FalseVal = IRB.getIntN(/* width */ 1, /* value */ 0);
+    Value *Aggregator = FalseVal;
+
+    for (unsigned Idx = 0; Idx < Struct->getNumElements(); Idx++) {
+      // Combine by ORing together each element's bool shadow
+      Value *ShadowItem = IRB.CreateExtractValue(Shadow, Idx);
+      Value *ShadowInner = convertShadowToScalar(ShadowItem, IRB);
+      Value *ShadowBool = convertToBool(ShadowInner, IRB);
+
+      if (Aggregator != FalseVal)
+        Aggregator = IRB.CreateOr(Aggregator, ShadowBool);
+      else
+        Aggregator = ShadowBool;
+    }
+
+    return Aggregator;
+  }
+
   /// Convert a shadow value to it's flattened variant.
-  Value *convertToShadowTyNoVec(Value *V, IRBuilder<> &IRB) {
+  Value *convertShadowToScalar(Value *V, IRBuilder<> &IRB) {
+    if (StructType *Struct = dyn_cast<StructType>(V->getType())) {
+      return collapseStructShadow(Struct, V, IRB);
+    }
     Type *Ty = V->getType();
     Type *NoVecTy = getShadowTyNoVec(Ty);
     if (Ty == NoVecTy) return V;
     return IRB.CreateBitCast(V, NoVecTy);
   }
 
+  // Convert a scalar value to an i1 by comparing with 0
+  Value *convertToBool(Value *V, IRBuilder<> &IRB, const Twine &name = "") {
+    Type *VTy = V->getType();
+    assert(VTy->isIntegerTy());
+    if (VTy->getIntegerBitWidth() == 1)
+      // Convert bool to bool is a nop
+      return V;
+    return IRB.CreateICmpNE(V, ConstantInt::get(VTy, 0), name);
+  }
+
   /// Compute the integer shadow offset that corresponds to a given
   /// application address.
   ///

     if (!InsertChecks) return;
 #ifndef NDEBUG
     Type *ShadowTy = Shadow->getType();
-    assert((isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy)) &&
-           "Can only insert checks for integer and vector shadow types");
+    assert(
+        (isa<IntegerType>(ShadowTy) || isa<VectorType>(ShadowTy) ||
+         isa<StructType>(ShadowTy)) &&
+        "Can only insert checks for integer, vector, and struct shadow types");
 #endif
     InstrumentationList.push_back(
         ShadowOriginAndInsertPoint(Shadow, Origin, OrigIns));

           Constant *ConstOrigin = dyn_cast<Constant>(OpOrigin);
           // No point in adding something that might result in 0 origin value.
           if (!ConstOrigin || !ConstOrigin->isNullValue()) {
-            Value *FlatShadow = MSV->convertToShadowTyNoVec(OpShadow, IRB);
+            Value *FlatShadow = MSV->convertShadowToScalar(OpShadow, IRB);
             Value *Cond =
                 IRB.CreateICmpNE(FlatShadow, MSV->getCleanShadow(FlatShadow));
             Origin = IRB.CreateSelect(Cond, OpOrigin, Origin);

llvm/test/Instrumentation/MemorySanitizer/check-struct.ll

+; RUN: opt < %s -msan-check-access-address=0 -msan-track-origins=1 -S -passes='module(msan-module),function(msan)' 2>&1 | \
+; RUN:   FileCheck -allow-deprecated-dag-overlap -check-prefixes=CHECK,CHECK-ORIGINS %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; CHECK-LABEL: @main
+define { i32, i8 } @main() sanitize_memory {
+; CHECK: %p = inttoptr i64 0 to { i32, i8 }*
+  %p = inttoptr i64 0 to { i32, i8 } *
+; CHECK: %o = load { i32, i8 }, { i32, i8 }* %p
+  %o = load { i32, i8 }, { i32, i8 } *%p
+; CHECK: [[FIELD0:%.+]] = extractvalue { i32, i8 } %_msld, 0
+; CHECK: [[F0_POISONED:%.+]] = icmp ne i32 [[FIELD0]]
+; CHECK: [[FIELD1:%.+]] = extractvalue { i32, i8 } %_msld, 1
+; CHECK: [[F1_POISONED:%.+]] = icmp ne i8 [[FIELD1]]
+; CHECK: [[F1_OR:%.+]] = or i1 [[F0_POISONED]], [[F1_POISONED]]
+; CHECK-NOT: icmp ne i1 {{.*}}, false
+; CHECK: br i1 %
+; CHECK: call void @__msan_warning
+; CHECK: ret { i32, i8 } %o
+  ret { i32, i8 } %o
+}

llvm/test/Instrumentation/MemorySanitizer/msan_x86_bts_asm.ll

 ; CHECK: [[MSPROP:%.*]] = trunc i8 [[MSLD]] to i1
 
 ; Is the shadow poisoned?
-; CHECK: [[MSCMP:%.*]] = icmp ne i1 [[MSPROP]], false
-; CHECK: br i1 [[MSCMP]], label %[[IFTRUE:.*]], label {{.*}}
+; CHECK: br i1 [[MSPROP]], label %[[IFTRUE:.*]], label {{.*}}
 
 ; If yes, raise a warning.
 ; CHECK: [[IFTRUE]]: