Add support to ValueTracking for determining that a pointer is non-null

by virtue of inbounds GEPs that preclude a null pointer.

This is a very common pattern in the code generated by std::vector and
other standard library routines which use allocators that test for null
pervasively. This is one step closer to teaching Clang+LLVM to be able
to produce an empty function for:

  void f() {
    std::vector<int> v;
    v.push_back(1);
    v.push_back(2);
    v.push_back(3);
    v.push_back(4);
  }

Which is related to getting them to completely fold SmallVector
push_back sequences into constants when inlining and other optimizations
make that a possibility.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@169573 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Analysis/ValueTracking.cpp

@@ -862,6 +862,72 @@ bool llvm::isPowerOfTwo(Value *V, const DataLayout *TD, bool OrZero,
   return false;
 }
 
+/// \brief Test whether a GEP's result is known to be non-null.
+///
+/// Uses properties inherent in a GEP to try to determine whether it is known
+/// to be non-null.
+///
+/// Currently this routine does not support vector GEPs.
+static bool isGEPKnownNonNull(GEPOperator *GEP, const DataLayout *DL,
+                              unsigned Depth) {
+  if (!GEP->isInBounds() || GEP->getPointerAddressSpace() != 0)
+    return false;
+
+  // FIXME: Support vector-GEPs.
+  assert(GEP->getType()->isPointerTy() && "We only support plain pointer GEP");
+
+  // If the base pointer is non-null, we cannot walk to a null address with an
+  // inbounds GEP in address space zero.
+  if (isKnownNonZero(GEP->getPointerOperand(), DL, Depth))
+    return true;
+
+  // Past this, if we don't have DataLayout, we can't do much.
+  if (!DL)
+    return false;
+
+  // Walk the GEP operands and see if any operand introduces a non-zero offset.
+  // If so, then the GEP cannot produce a null pointer, as doing so would
+  // inherently violate the inbounds contract within address space zero.
+  for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
+       GTI != GTE; ++GTI) {
+    // Struct types are easy -- they must always be indexed by a constant.
+    if (StructType *STy = dyn_cast<StructType>(*GTI)) {
+      ConstantInt *OpC = cast<ConstantInt>(GTI.getOperand());
+      unsigned ElementIdx = OpC->getZExtValue();
+      const StructLayout *SL = DL->getStructLayout(STy);
+      uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
+      if (ElementOffset > 0)
+        return true;
+      continue;
+    }
+
+    // If we have a zero-sized type, the index doesn't matter. Keep looping.
+    if (DL->getTypeAllocSize(GTI.getIndexedType()) == 0)
+      continue;
+
+    // Fast path the constant operand case both for efficiency and so we don't
+    // increment Depth when just zipping down an all-constant GEP.
+    if (ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand())) {
+      if (!OpC->isZero())
+        return true;
+      continue;
+    }
+
+    // We post-increment Depth here because while isKnownNonZero increments it
+    // as well, when we pop back up that increment won't persist. We don't want
+    // to recurse 10k times just because we have 10k GEP operands. We don't
+    // bail completely out because we want to handle constant GEPs regardless
+    // of depth.
+    if (Depth++ >= MaxDepth)
+      continue;
+
+    if (isKnownNonZero(GTI.getOperand(), DL, Depth))
+      return true;
+  }
+
+  return false;
+}
+
 /// isKnownNonZero - Return true if the given value is known to be non-zero
 /// when defined.  For vectors return true if every element is known to be
 /// non-zero when defined.  Supports values with integer or pointer type and
@@ -881,6 +947,13 @@ bool llvm::isKnownNonZero(Value *V, const DataLayout *TD, unsigned Depth) {
   if (Depth++ >= MaxDepth)
     return false;
 
+  // Check for pointer simplifications.
+  if (V->getType()->isPointerTy()) {
+    if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
+      if (isGEPKnownNonNull(GEP, TD, Depth))
+        return true;
+  }
+
   unsigned BitWidth = getBitWidth(V->getType(), TD);
 
   // X | Y != 0 if X != 0 or Y != 0.

test/Transforms/InstSimplify/compare.ll

@@ -165,6 +165,46 @@ entry:
   ret i1 %cmp
 }
 
+define i1 @gep13(i8* %ptr) {
+; CHECK: @gep13
+; We can prove this GEP is non-null because it is inbounds.
+  %x = getelementptr inbounds i8* %ptr, i32 1
+  %cmp = icmp eq i8* %x, null
+  ret i1 %cmp
+; CHECK-NEXT: ret i1 false
+}
+
+define i1 @gep14({ {}, i8 }* %ptr) {
+; CHECK: @gep14
+; We can't simplify this because the offset of one in the GEP actually doesn't
+; move the pointer.
+  %x = getelementptr inbounds { {}, i8 }* %ptr, i32 0, i32 1
+  %cmp = icmp eq i8* %x, null
+  ret i1 %cmp
+; CHECK-NOT: ret i1 false
+}
+
+define i1 @gep15({ {}, [4 x {i8, i8}]}* %ptr, i32 %y) {
+; CHECK: @gep15
+; We can prove this GEP is non-null even though there is a user value, as we
+; would necessarily violate inbounds on one side or the other.
+  %x = getelementptr inbounds { {}, [4 x {i8, i8}]}* %ptr, i32 0, i32 1, i32 %y, i32 1
+  %cmp = icmp eq i8* %x, null
+  ret i1 %cmp
+; CHECK-NEXT: ret i1 false
+}
+
+define i1 @gep16(i8* %ptr, i32 %a) {
+; CHECK: @gep16
+; We can prove this GEP is non-null because it is inbounds and because we know
+; %b is non-zero even though we don't know its value.
+  %b = or i32 %a, 1
+  %x = getelementptr inbounds i8* %ptr, i32 %b
+  %cmp = icmp eq i8* %x, null
+  ret i1 %cmp
+; CHECK-NEXT: ret i1 false
+}
+
 define i1 @zext(i32 %x) {
 ; CHECK: @zext
   %e1 = zext i32 %x to i64