[llvm] Use computeConstantRange to improve llvm.objectsize computation

Using computeConstantRange, it is possible to compute valuable information for
allocation functions, GEP and alloca, even in the presence of some dynamic
information.

llvm.objectsize plays an important role in _FORTIFY_SOURCE definitions,
so improving its diagnostic in turns improves the security of compiled
application.

As a side note, as a result of recent optimization improvements, clang no
longer passes https://github.com/serge-sans-paille/builtin_object_size-test-suite
This commit restores the situation and greatly improves the scope of
code handled by the static version of __builtin_object_size.

Author: serge-sans-paille

llvm/include/llvm/Analysis/MemoryBuiltins.h

@@ -32,6 +32,7 @@ class AAResults;
 class Argument;
 class ConstantPointerNull;
 class DataLayout;
+class DominatorTree;
 class ExtractElementInst;
 class ExtractValueInst;
 class GEPOperator;
@@ -160,8 +161,10 @@ struct ObjectSizeOpts {
   /// though they can't be evaluated. Otherwise, null is always considered to
   /// point to a 0 byte region of memory.
   bool NullIsUnknownSize = false;
-  /// If set, used for more accurate evaluation
+  /// If set, used for more accurate evaluation.
   AAResults *AA = nullptr;
+  /// If set, used for more accurate evaluation.
+  DominatorTree *DT = nullptr;
 };
 
 /// Compute the size of the object pointed by Ptr. Returns true and the
@@ -186,6 +189,12 @@ Value *lowerObjectSizeCall(
     const TargetLibraryInfo *TLI, AAResults *AA, bool MustSucceed,
     SmallVectorImpl<Instruction *> *InsertedInstructions = nullptr);
 
+Value *lowerObjectSizeCall(
+    IntrinsicInst *ObjectSize, const DataLayout &DL,
+    const TargetLibraryInfo *TLI, AAResults *AA, DominatorTree *DT,
+    bool MustSucceed,
+    SmallVectorImpl<Instruction *> *InsertedInstructions = nullptr);
+
 /// SizeOffsetType - A base template class for the object size visitors. Used
 /// here as a self-documenting way to handle the values rather than using a
 /// \p std::pair.

llvm/include/llvm/IR/Value.h

@@ -723,12 +723,16 @@ class Value {
       bool AllowInvariantGroup = false,
       function_ref<bool(Value &Value, APInt &Offset)> ExternalAnalysis =
           nullptr) const;
-  Value *stripAndAccumulateConstantOffsets(const DataLayout &DL, APInt &Offset,
-                                           bool AllowNonInbounds,
-                                           bool AllowInvariantGroup = false) {
+
+  Value *stripAndAccumulateConstantOffsets(
+      const DataLayout &DL, APInt &Offset, bool AllowNonInbounds,
+      bool AllowInvariantGroup = false,
+      function_ref<bool(Value &Value, APInt &Offset)> ExternalAnalysis =
+          nullptr) {
     return const_cast<Value *>(
         static_cast<const Value *>(this)->stripAndAccumulateConstantOffsets(
-            DL, Offset, AllowNonInbounds, AllowInvariantGroup));
+            DL, Offset, AllowNonInbounds, AllowInvariantGroup,
+            ExternalAnalysis));
   }
 
   /// This is a wrapper around stripAndAccumulateConstantOffsets with the

llvm/lib/Analysis/MemoryBuiltins.cpp

@@ -25,6 +25,7 @@
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/DataLayout.h"
 #include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/Dominators.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/GlobalAlias.h"
 #include "llvm/IR/GlobalVariable.h"
@@ -589,19 +590,28 @@ Value *llvm::lowerObjectSizeCall(IntrinsicInst *ObjectSize,
                                  const TargetLibraryInfo *TLI,
                                  bool MustSucceed) {
   return lowerObjectSizeCall(ObjectSize, DL, TLI, /*AAResults=*/nullptr,
-                             MustSucceed);
+                             /*DT=*/nullptr, MustSucceed);
 }
 
 Value *llvm::lowerObjectSizeCall(
     IntrinsicInst *ObjectSize, const DataLayout &DL,
     const TargetLibraryInfo *TLI, AAResults *AA, bool MustSucceed,
     SmallVectorImpl<Instruction *> *InsertedInstructions) {
+  return lowerObjectSizeCall(ObjectSize, DL, TLI, AA, /*DT=*/nullptr,
+                             MustSucceed, InsertedInstructions);
+}
+
+Value *llvm::lowerObjectSizeCall(
+    IntrinsicInst *ObjectSize, const DataLayout &DL,
+    const TargetLibraryInfo *TLI, AAResults *AA, DominatorTree *DT,
+    bool MustSucceed, SmallVectorImpl<Instruction *> *InsertedInstructions) {
   assert(ObjectSize->getIntrinsicID() == Intrinsic::objectsize &&
          "ObjectSize must be a call to llvm.objectsize!");
 
   bool MaxVal = cast<ConstantInt>(ObjectSize->getArgOperand(1))->isZero();
   ObjectSizeOpts EvalOptions;
   EvalOptions.AA = AA;
+  EvalOptions.DT = DT;
 
   // Unless we have to fold this to something, try to be as accurate as
   // possible.
@@ -714,14 +724,54 @@ OffsetSpan ObjectSizeOffsetVisitor::computeImpl(Value *V) {
   // readjust the APInt as we pass it upwards in order for the APInt to match
   // the type the caller passed in.
   APInt Offset(InitialIntTyBits, 0);
+
   V = V->stripAndAccumulateConstantOffsets(
       DL, Offset, /* AllowNonInbounds */ true, /* AllowInvariantGroup */ true);
 
+  // Give it another try with approximated analysis. We don't start with this
+  // one because stripAndAccumulateConstantOffsets behaves differently wrt.
+  // overflows if we provide an external Analysis.
+  if(isa<GetElementPtrInst>(V)) {
+    // External Analysis used to compute the Min/Max value of individual Offsets
+    // within a GEP.
+    auto OffsetRangeAnalysis = [this, V](Value &VOffset, APInt &Offset) {
+      if (auto *C = dyn_cast<ConstantInt>(&VOffset)) {
+        Offset = C->getValue();
+        return true;
+      }
+      if (Options.EvalMode != ObjectSizeOpts::Mode::Min &&
+          Options.EvalMode != ObjectSizeOpts::Mode::Max) {
+        return false;
+      }
+      ConstantRange CR = computeConstantRange(
+          &VOffset, /*ForSigned*/ true, /*UseInstrInfo*/ true, /*AC=*/nullptr,
+          /*CtxtI=*/dyn_cast<Instruction>(V), /*DT=*/Options.DT);
+      if (CR.isFullSet())
+        return false;
+
+      if (Options.EvalMode == ObjectSizeOpts::Mode::Min) {
+        Offset = CR.getSignedMax();
+        // Upper bound actually unknown.
+        if (Offset.isMaxSignedValue())
+          return false;
+      } else {
+        Offset = CR.getSignedMin();
+        // Lower bound actually unknown.
+        if (Offset.isMinSignedValue())
+          return false;
+      }
+      return true;
+    };
+
+    V = V->stripAndAccumulateConstantOffsets(
+        DL, Offset, /* AllowNonInbounds */ true, /* AllowInvariantGroup */ true,
+        /*ExternalAnalysis=*/OffsetRangeAnalysis);
+  }
+
   // Later we use the index type size and zero but it will match the type of the
   // value that is passed to computeImpl.
   IntTyBits = DL.getIndexTypeSizeInBits(V->getType());
   Zero = APInt::getZero(IntTyBits);
-
   OffsetSpan ORT = computeValue(V);
 
   bool IndexTypeSizeChanged = InitialIntTyBits != IntTyBits;
@@ -801,6 +851,26 @@ OffsetSpan ObjectSizeOffsetVisitor::visitAllocaInst(AllocaInst &I) {
 
     return Overflow ? ObjectSizeOffsetVisitor::unknown()
                     : OffsetSpan(Zero, alignAndCap(Size, I.getAlign()));
+  } else {
+    ConstantRange CR =
+        computeConstantRange(ArraySize, /*ForSigned*/ false,
+                             /*UseInstrInfo*/ true, /*AC=*/nullptr,
+                             /*CtxtI=*/&I, /*DT=*/Options.DT);
+    if (CR.isFullSet())
+      return ObjectSizeOffsetVisitor::unknown();
+    APInt Bound;
+    if (Options.EvalMode == ObjectSizeOpts::Mode::Max) {
+      Bound = CR.getUnsignedMax();
+      // Upper bound actually unknown.
+      if (Bound.isMaxValue())
+        return ObjectSizeOffsetVisitor::unknown();
+    } else {
+      Bound = CR.getUnsignedMin();
+      // Lower bound actually unknown.
+      if (Bound.isMinValue())
+        return ObjectSizeOffsetVisitor::unknown();
+    }
+    return OffsetSpan(Zero, alignAndCap(Bound, I.getAlign()));
   }
   return ObjectSizeOffsetVisitor::unknown();
 }
@@ -818,7 +888,32 @@ OffsetSpan ObjectSizeOffsetVisitor::visitArgument(Argument &A) {
 }
 
 OffsetSpan ObjectSizeOffsetVisitor::visitCallBase(CallBase &CB) {
-  if (std::optional<APInt> Size = getAllocSize(&CB, TLI)) {
+  auto Mapper = [&CB, this](const Value *V) -> const Value * {
+    if (!V->getType()->isIntegerTy())
+      return V;
+    if (isa<ConstantInt>(V))
+      return V;
+    ConstantRange CR = computeConstantRange(
+        V, /*ForSigned*/ false, /*UseInstrInfo*/ true, /*AC=*/nullptr,
+        /*CtxtI=*/&CB, /*DT=*/Options.DT);
+    if (CR.isFullSet())
+      return V;
+
+    APInt Bound;
+    if (Options.EvalMode == ObjectSizeOpts::Mode::Max) {
+      Bound = CR.getUnsignedMax();
+      // Upper bound actually unknown.
+      if (Bound.isMaxValue())
+        return V;
+    } else {
+      Bound = CR.getUnsignedMin();
+      // Lower bound actually unknown.
+      if (Bound.isMinValue())
+        return V;
+    }
+    return ConstantInt::get(V->getType(), Bound);
+  };
+  if (std::optional<APInt> Size = getAllocSize(&CB, TLI, Mapper)) {
     // Very large unsigned value cannot be represented as OffsetSpan.
     if (Size->isNegative())
       return ObjectSizeOffsetVisitor::unknown();

llvm/lib/Transforms/InstCombine/InstructionCombining.cpp

@@ -3318,8 +3318,9 @@ Instruction *InstCombinerImpl::visitAllocSite(Instruction &MI) {
       if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
         if (II->getIntrinsicID() == Intrinsic::objectsize) {
           SmallVector<Instruction *> InsertedInstructions;
-          Value *Result = lowerObjectSizeCall(
-              II, DL, &TLI, AA, /*MustSucceed=*/true, &InsertedInstructions);
+          Value *Result =
+              lowerObjectSizeCall(II, DL, &TLI, AA, &DT,
+                                  /*MustSucceed=*/true, &InsertedInstructions);
           for (Instruction *Inserted : InsertedInstructions)
             Worklist.add(Inserted);
           replaceInstUsesWith(*I, Result);

llvm/lib/Transforms/Scalar/LowerConstantIntrinsics.cpp

@@ -143,7 +143,7 @@ bool llvm::lowerConstantIntrinsics(Function &F, const TargetLibraryInfo &TLI,
       IsConstantIntrinsicsHandled++;
       break;
     case Intrinsic::objectsize:
-      NewValue = lowerObjectSizeCall(II, DL, &TLI, true);
+      NewValue = lowerObjectSizeCall(II, DL, &TLI, /*AA=*/nullptr, DT, true);
       LLVM_DEBUG(dbgs() << "Folding " << *II << " to " << *NewValue << "\n");
       ObjectSizeIntrinsicsHandled++;
       break;

llvm/test/Transforms/LowerConstantIntrinsics/builtin-object-size-range.ll

@@ -0,0 +1,93 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -passes=lower-constant-intrinsics -S < %s | FileCheck %s
+
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+declare i64 @llvm.objectsize.i64.p0(ptr, i1 immarg, i1 immarg, i1 immarg)
+declare noalias ptr @malloc(i64 noundef) #0
+
+define i64 @select_alloc_size(i1 %cond) {
+; CHECK-LABEL: @select_alloc_size(
+; CHECK-NEXT:    [[SIZE:%.*]] = select i1 [[COND:%.*]], i64 3, i64 4
+; CHECK-NEXT:    [[PTR:%.*]] = alloca i8, i64 [[SIZE]], align 1
+; CHECK-NEXT:    [[RES:%.*]] = select i1 [[COND]], i64 4, i64 3
+; CHECK-NEXT:    ret i64 [[RES]]
+;
+  %size = select i1 %cond, i64 3, i64 4
+  %ptr = alloca i8, i64 %size
+  %objsize_max = call i64 @llvm.objectsize.i64.p0(ptr %ptr, i1 false, i1 true, i1 false)
+  %objsize_min = call i64 @llvm.objectsize.i64.p0(ptr %ptr, i1 true, i1 true, i1 false)
+  %res = select i1 %cond, i64 %objsize_max, i64 %objsize_min
+  ret i64 %res
+}
+
+define i64 @select_malloc_size(i1 %cond) {
+; CHECK-LABEL: @select_malloc_size(
+; CHECK-NEXT:    [[SIZE:%.*]] = select i1 [[COND:%.*]], i64 3, i64 4
+; CHECK-NEXT:    [[PTR:%.*]] = call noalias ptr @malloc(i64 noundef [[SIZE]])
+; CHECK-NEXT:    [[RES:%.*]] = select i1 [[COND]], i64 4, i64 3
+; CHECK-NEXT:    ret i64 [[RES]]
+;
+  %size = select i1 %cond, i64 3, i64 4
+  %ptr = call noalias ptr @malloc(i64 noundef %size)
+  %objsize_max = call i64 @llvm.objectsize.i64.p0(ptr %ptr, i1 false, i1 true, i1 false)
+  %objsize_min = call i64 @llvm.objectsize.i64.p0(ptr %ptr, i1 true, i1 true, i1 false)
+  %res = select i1 %cond, i64 %objsize_max, i64 %objsize_min
+  ret i64 %res
+}
+
+define i64 @select_gep_offset(i1 %cond) {
+; CHECK-LABEL: @select_gep_offset(
+; CHECK-NEXT:    [[PTR:%.*]] = alloca i8, i64 10, align 1
+; CHECK-NEXT:    [[OFFSET:%.*]] = select i1 [[COND:%.*]], i64 3, i64 4
+; CHECK-NEXT:    [[PTR_SLIDE:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 [[OFFSET]]
+; CHECK-NEXT:    [[RES:%.*]] = select i1 [[COND]], i64 7, i64 6
+; CHECK-NEXT:    ret i64 [[RES]]
+;
+  %ptr = alloca i8, i64 10
+  %offset = select i1 %cond, i64 3, i64 4
+  %ptr.slide = getelementptr inbounds i8, ptr %ptr, i64 %offset
+  %objsize_max = call i64 @llvm.objectsize.i64.p0(ptr %ptr.slide, i1 false, i1 true, i1 false)
+  %objsize_min = call i64 @llvm.objectsize.i64.p0(ptr %ptr.slide, i1 true, i1 true, i1 false)
+  %res = select i1 %cond, i64 %objsize_max, i64 %objsize_min
+  ret i64 %res
+}
+
+define i64 @select_gep_neg_offset(i1 %cond) {
+; CHECK-LABEL: @select_gep_neg_offset(
+; CHECK-NEXT:    [[PTR:%.*]] = alloca i8, i64 10, align 1
+; CHECK-NEXT:    [[PTR_SLIDE_1:%.*]] = getelementptr inbounds i8, ptr [[PTR]], i64 5
+; CHECK-NEXT:    [[OFFSET:%.*]] = select i1 [[COND:%.*]], i64 -3, i64 -4
+; CHECK-NEXT:    [[PTR_SLIDE_2:%.*]] = getelementptr inbounds i8, ptr [[PTR_SLIDE_1]], i64 [[OFFSET]]
+; CHECK-NEXT:    [[RES:%.*]] = select i1 [[COND]], i64 9, i64 8
+; CHECK-NEXT:    ret i64 [[RES]]
+;
+  %ptr = alloca i8, i64 10
+  %ptr.slide.1 = getelementptr inbounds i8, ptr %ptr, i64 5
+  %offset = select i1 %cond, i64 -3, i64 -4
+  %ptr.slide.2 = getelementptr inbounds i8, ptr %ptr.slide.1, i64 %offset
+  %objsize_max = call i64 @llvm.objectsize.i64.p0(ptr %ptr.slide.2, i1 false, i1 true, i1 false)
+  %objsize_min = call i64 @llvm.objectsize.i64.p0(ptr %ptr.slide.2, i1 true, i1 true, i1 false)
+  %res = select i1 %cond, i64 %objsize_max, i64 %objsize_min
+  ret i64 %res
+}
+
+define i64 @select_gep_offsets(i1 %cond) {
+; CHECK-LABEL: @select_gep_offsets(
+; CHECK-NEXT:    [[PTR:%.*]] = alloca [10 x i8], i64 2, align 1
+; CHECK-NEXT:    [[OFFSET:%.*]] = select i1 [[COND:%.*]], i32 0, i32 1
+; CHECK-NEXT:    [[PTR_SLIDE:%.*]] = getelementptr inbounds [10 x i8], ptr [[PTR]], i32 [[OFFSET]], i32 5
+; CHECK-NEXT:    [[RES:%.*]] = select i1 [[COND]], i64 15, i64 5
+; CHECK-NEXT:    ret i64 [[RES]]
+;
+  %ptr = alloca [10 x i8], i64 2
+  %offset = select i1 %cond, i32 0, i32 1
+  %ptr.slide = getelementptr inbounds [10 x i8], ptr %ptr, i32 %offset, i32 5
+  %objsize_max = call i64 @llvm.objectsize.i64.p0(ptr %ptr.slide, i1 false, i1 true, i1 false)
+  %objsize_min = call i64 @llvm.objectsize.i64.p0(ptr %ptr.slide, i1 true, i1 true, i1 false)
+  %res = select i1 %cond, i64 %objsize_max, i64 %objsize_min
+  ret i64 %res
+}
+
+attributes #0 = { nounwind allocsize(0) }