[SeparateConstOffsetFromGEP] Preserve inbounds flag based on ValueTracking and NUW (#130617)

If we know that the initial GEP was inbounds, and we change it to a
sequence of GEPs from the same base pointer where every offset is
non-negative, then the new GEPs are inbounds.

We can also preserve inbounds if the inbounds GEP and the involved additions are NUW.

For SWDEV-516125.

Author: ritter-x2a

llvm/lib/Transforms/Scalar/SeparateConstOffsetFromGEP.cpp

@@ -235,8 +235,10 @@ class ConstantOffsetExtractor {
   /// \p GEP The given GEP
   /// \p UserChainTail Outputs the tail of UserChain so that we can
   ///                  garbage-collect unused instructions in UserChain.
+  /// \p PreservesNUW  Outputs whether the extraction allows preserving the
+  ///                  GEP's nuw flag, if it has one.
   static Value *Extract(Value *Idx, GetElementPtrInst *GEP,
-                        User *&UserChainTail);
+                        User *&UserChainTail, bool &PreservesNUW);
 
   /// Looks for a constant offset from the given GEP index without extracting
   /// it. It returns the numeric value of the extracted constant offset (0 if
@@ -778,17 +780,45 @@ Value *ConstantOffsetExtractor::removeConstOffset(unsigned ChainIndex) {
   return NewBO;
 }
 
+/// A helper function to check if reassociating through an entry in the user
+/// chain would invalidate the GEP's nuw flag.
+static bool allowsPreservingNUW(const User *U) {
+  if (const BinaryOperator *BO = dyn_cast<BinaryOperator>(U)) {
+    // Binary operations need to be effectively add nuw.
+    auto Opcode = BO->getOpcode();
+    if (Opcode == BinaryOperator::Or) {
+      // Ors are only considered here if they are disjoint. The addition that
+      // they represent in this case is NUW.
+      assert(cast<PossiblyDisjointInst>(BO)->isDisjoint());
+      return true;
+    }
+    return Opcode == BinaryOperator::Add && BO->hasNoUnsignedWrap();
+  }
+  // UserChain can only contain ConstantInt, CastInst, or BinaryOperator.
+  // Among the possible CastInsts, only trunc without nuw is a problem: If it
+  // is distributed through an add nuw, wrapping may occur:
+  // "add nuw trunc(a), trunc(b)" is more poisonous than "trunc(add nuw a, b)"
+  if (const TruncInst *TI = dyn_cast<TruncInst>(U))
+    return TI->hasNoUnsignedWrap();
+  return isa<CastInst>(U) || isa<ConstantInt>(U);
+}
+
 Value *ConstantOffsetExtractor::Extract(Value *Idx, GetElementPtrInst *GEP,
-                                        User *&UserChainTail) {
+                                        User *&UserChainTail,
+                                        bool &PreservesNUW) {
   ConstantOffsetExtractor Extractor(GEP->getIterator());
   // Find a non-zero constant offset first.
   APInt ConstantOffset =
       Extractor.find(Idx, /* SignExtended */ false, /* ZeroExtended */ false,
                      GEP->isInBounds());
   if (ConstantOffset == 0) {
     UserChainTail = nullptr;
+    PreservesNUW = true;
     return nullptr;
   }
+
+  PreservesNUW = all_of(Extractor.UserChain, allowsPreservingNUW);
+
   // Separates the constant offset from the GEP index.
   Value *IdxWithoutConstOffset = Extractor.rebuildWithoutConstOffset();
   UserChainTail = Extractor.UserChain.back();
@@ -1052,6 +1082,10 @@ bool SeparateConstOffsetFromGEP::splitGEP(GetElementPtrInst *GEP) {
     }
   }
 
+  // Track information for preserving GEP flags.
+  bool AllOffsetsNonNegative = AccumulativeByteOffset >= 0;
+  bool AllNUWPreserved = true;
+
   // Remove the constant offset in each sequential index. The resultant GEP
   // computes the variadic base.
   // Notice that we don't remove struct field indices here. If LowerGEP is
@@ -1070,15 +1104,19 @@ bool SeparateConstOffsetFromGEP::splitGEP(GetElementPtrInst *GEP) {
       // uses the variadic part as the new index.
       Value *OldIdx = GEP->getOperand(I);
       User *UserChainTail;
-      Value *NewIdx =
-          ConstantOffsetExtractor::Extract(OldIdx, GEP, UserChainTail);
+      bool PreservesNUW;
+      Value *NewIdx = ConstantOffsetExtractor::Extract(
+          OldIdx, GEP, UserChainTail, PreservesNUW);
       if (NewIdx != nullptr) {
         // Switches to the index with the constant offset removed.
         GEP->setOperand(I, NewIdx);
         // After switching to the new index, we can garbage-collect UserChain
         // and the old index if they are not used.
         RecursivelyDeleteTriviallyDeadInstructions(UserChainTail);
         RecursivelyDeleteTriviallyDeadInstructions(OldIdx);
+        AllOffsetsNonNegative =
+            AllOffsetsNonNegative && isKnownNonNegative(NewIdx, *DL);
+        AllNUWPreserved &= PreservesNUW;
       }
     }
   }
@@ -1099,12 +1137,35 @@ bool SeparateConstOffsetFromGEP::splitGEP(GetElementPtrInst *GEP) {
   // inbounds keyword is not present, the offsets are added to the base
   // address with silently-wrapping two's complement arithmetic".
   // Therefore, the final code will be a semantically equivalent.
-  //
-  // TODO(jingyue): do some range analysis to keep as many inbounds as
-  // possible. GEPs with inbounds are more friendly to alias analysis.
-  // TODO(gep_nowrap): Preserve nuw at least.
   GEPNoWrapFlags NewGEPFlags = GEPNoWrapFlags::none();
-  GEP->setNoWrapFlags(GEPNoWrapFlags::none());
+
+  // If the initial GEP was inbounds/nusw and all variable indices and the
+  // accumulated offsets are non-negative, they can be added in any order and
+  // the intermediate results are in bounds and don't overflow in a nusw sense.
+  // So, we can preserve the inbounds/nusw flag for both GEPs.
+  bool CanPreserveInBoundsNUSW = AllOffsetsNonNegative;
+
+  // If the initial GEP was NUW and all operations that we reassociate were NUW
+  // additions, the resulting GEPs are also NUW.
+  if (GEP->hasNoUnsignedWrap() && AllNUWPreserved) {
+    NewGEPFlags |= GEPNoWrapFlags::noUnsignedWrap();
+    // If the initial GEP additionally had NUSW (or inbounds, which implies
+    // NUSW), we know that the indices in the initial GEP must all have their
+    // signbit not set. For indices that are the result of NUW adds, the
+    // add-operands therefore also don't have their signbit set. Therefore, all
+    // indices of the resulting GEPs are non-negative -> we can preserve
+    // the inbounds/nusw flag.
+    CanPreserveInBoundsNUSW |= GEP->hasNoUnsignedSignedWrap();
+  }
+
+  if (CanPreserveInBoundsNUSW) {
+    if (GEP->isInBounds())
+      NewGEPFlags |= GEPNoWrapFlags::inBounds();
+    else if (GEP->hasNoUnsignedSignedWrap())
+      NewGEPFlags |= GEPNoWrapFlags::noUnsignedSignedWrap();
+  }
+
+  GEP->setNoWrapFlags(NewGEPFlags);
 
   // Lowers a GEP to either GEPs with a single index or arithmetic operations.
   if (LowerGEP) {