[AMDGPU][SDAG] Only fold flat offsets if they are inbounds

For flat memory instructions where the address is supplied as a base address
register with an immediate offset, the memory aperture test ignores the
immediate offset. Currently, ISel does not respect that, which leads to
miscompilations where valid input programs crash when the address computation
relies on the immediate offset to get the base address in the proper memory
aperture. Global or scratch instructions are not affected.

This patch only selects flat instructions with immediate offsets from address
computations with the inbounds flag: If the address computation does not leave
the bounds of the allocated object, it cannot leave the bounds of the memory
aperture and is therefore safe to handle with an immediate offset.

It also adds the inbounds flag to DAG nodes resulting from transformations:
- Address computations resulting from getObjectPtrOffset. As far as I can tell,
  this function is only used to compute addresses within accessed memory ranges,
  e.g., for loads and stores that are split during legalization.
- Reassociated inbounds adds. If both involved operations are inbounds, then so
  are operations after the transformation.
- Address computations in the SelectionDAG lowering of the memcpy/move/set
  intrinsics. Base and result of the address arithmetic there are accessed, so
  the operation must be inbounds.

It might make sense to separate these changes into their own PR, but I don't
see a way to test them without adding a use of the inbounds SDAG flag.

Affected tests:
- CodeGen/AMDGPU/fold-gep-offset.ll: Offsets are no longer wrongly folded,
  added new positive tests where we still do fold them.
- Transforms/InferAddressSpaces/AMDGPU/flat_atomic.ll: Offset folding doesn't
  seem integral to this test, so the test is not changed to make offset folding
  still happen.
- CodeGen/AMDGPU/loop-prefetch-data.ll: loop-reduce prefers to base addresses
  on the potentially OOB addresses used for prefetching for memory accesses,
  that might be a separate issue to look into.
- Added memset tests to CodeGen/AMDGPU/memintrinsic-unroll.ll to make sure that
  offsets in the memset DAG lowering are still folded properly.

A similar patch for GlobalISel will follow.

Fixes SWDEV-516125.

Author: ritter-x2a

llvm/include/llvm/CodeGen/SelectionDAG.h

@@ -1069,23 +1069,27 @@ class SelectionDAG {
                              SDValue EVL);
 
   /// Returns sum of the base pointer and offset.
-  /// Unlike getObjectPtrOffset this does not set NoUnsignedWrap by default.
+  /// Unlike getObjectPtrOffset this does not set NoUnsignedWrap and InBounds by
+  /// default.
   SDValue getMemBasePlusOffset(SDValue Base, TypeSize Offset, const SDLoc &DL,
                                const SDNodeFlags Flags = SDNodeFlags());
   SDValue getMemBasePlusOffset(SDValue Base, SDValue Offset, const SDLoc &DL,
                                const SDNodeFlags Flags = SDNodeFlags());
 
   /// Create an add instruction with appropriate flags when used for
   /// addressing some offset of an object. i.e. if a load is split into multiple
-  /// components, create an add nuw from the base pointer to the offset.
+  /// components, create an add nuw inbounds from the base pointer to the
+  /// offset.
   SDValue getObjectPtrOffset(const SDLoc &SL, SDValue Ptr, TypeSize Offset) {
-    return getMemBasePlusOffset(Ptr, Offset, SL, SDNodeFlags::NoUnsignedWrap);
+    return getMemBasePlusOffset(
+        Ptr, Offset, SL, SDNodeFlags::NoUnsignedWrap | SDNodeFlags::InBounds);
   }
 
   SDValue getObjectPtrOffset(const SDLoc &SL, SDValue Ptr, SDValue Offset) {
     // The object itself can't wrap around the address space, so it shouldn't be
     // possible for the adds of the offsets to the split parts to overflow.
-    return getMemBasePlusOffset(Ptr, Offset, SL, SDNodeFlags::NoUnsignedWrap);
+    return getMemBasePlusOffset(
+        Ptr, Offset, SL, SDNodeFlags::NoUnsignedWrap | SDNodeFlags::InBounds);
   }
 
   /// Return a new CALLSEQ_START node, that starts new call frame, in which

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -1213,9 +1213,12 @@ SDValue DAGCombiner::reassociateOpsCommutative(unsigned Opc, const SDLoc &DL,
 
   if (DAG.isConstantIntBuildVectorOrConstantInt(N01)) {
     SDNodeFlags NewFlags;
-    if (N0.getOpcode() == ISD::ADD && N0->getFlags().hasNoUnsignedWrap() &&
-        Flags.hasNoUnsignedWrap())
-      NewFlags |= SDNodeFlags::NoUnsignedWrap;
+    if (N0.getOpcode() == ISD::ADD) {
+      if (N0->getFlags().hasNoUnsignedWrap() && Flags.hasNoUnsignedWrap())
+        NewFlags |= SDNodeFlags::NoUnsignedWrap;
+      if (N0->getFlags().hasInBounds() && Flags.hasInBounds())
+        NewFlags |= SDNodeFlags::InBounds;
+    }
 
     if (DAG.isConstantIntBuildVectorOrConstantInt(N1)) {
       // Reassociate: (op (op x, c1), c2) -> (op x, (op c1, c2))

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

@@ -8279,7 +8279,7 @@ static SDValue getMemcpyLoadsAndStores(
       if (Value.getNode()) {
         Store = DAG.getStore(
             Chain, dl, Value,
-            DAG.getMemBasePlusOffset(Dst, TypeSize::getFixed(DstOff), dl),
+            DAG.getObjectPtrOffset(dl, Dst, TypeSize::getFixed(DstOff)),
             DstPtrInfo.getWithOffset(DstOff), Alignment, MMOFlags, NewAAInfo);
         OutChains.push_back(Store);
       }
@@ -8304,14 +8304,14 @@ static SDValue getMemcpyLoadsAndStores(
 
       Value = DAG.getExtLoad(
           ISD::EXTLOAD, dl, NVT, Chain,
-          DAG.getMemBasePlusOffset(Src, TypeSize::getFixed(SrcOff), dl),
+          DAG.getObjectPtrOffset(dl, Src, TypeSize::getFixed(SrcOff)),
           SrcPtrInfo.getWithOffset(SrcOff), VT,
           commonAlignment(*SrcAlign, SrcOff), SrcMMOFlags, NewAAInfo);
       OutLoadChains.push_back(Value.getValue(1));
 
       Store = DAG.getTruncStore(
           Chain, dl, Value,
-          DAG.getMemBasePlusOffset(Dst, TypeSize::getFixed(DstOff), dl),
+          DAG.getObjectPtrOffset(dl, Dst, TypeSize::getFixed(DstOff)),
           DstPtrInfo.getWithOffset(DstOff), VT, Alignment, MMOFlags, NewAAInfo);
       OutStoreChains.push_back(Store);
     }
@@ -8448,7 +8448,7 @@ static SDValue getMemmoveLoadsAndStores(SelectionDAG &DAG, const SDLoc &dl,
 
     Value = DAG.getLoad(
         VT, dl, Chain,
-        DAG.getMemBasePlusOffset(Src, TypeSize::getFixed(SrcOff), dl),
+        DAG.getObjectPtrOffset(dl, Src, TypeSize::getFixed(SrcOff)),
         SrcPtrInfo.getWithOffset(SrcOff), *SrcAlign, SrcMMOFlags, NewAAInfo);
     LoadValues.push_back(Value);
     LoadChains.push_back(Value.getValue(1));
@@ -8463,7 +8463,7 @@ static SDValue getMemmoveLoadsAndStores(SelectionDAG &DAG, const SDLoc &dl,
 
     Store = DAG.getStore(
         Chain, dl, LoadValues[i],
-        DAG.getMemBasePlusOffset(Dst, TypeSize::getFixed(DstOff), dl),
+        DAG.getObjectPtrOffset(dl, Dst, TypeSize::getFixed(DstOff)),
         DstPtrInfo.getWithOffset(DstOff), Alignment, MMOFlags, NewAAInfo);
     OutChains.push_back(Store);
     DstOff += VTSize;
@@ -8595,7 +8595,7 @@ static SDValue getMemsetStores(SelectionDAG &DAG, const SDLoc &dl,
     assert(Value.getValueType() == VT && "Value with wrong type.");
     SDValue Store = DAG.getStore(
         Chain, dl, Value,
-        DAG.getMemBasePlusOffset(Dst, TypeSize::getFixed(DstOff), dl),
+        DAG.getObjectPtrOffset(dl, Dst, TypeSize::getFixed(DstOff)),
         DstPtrInfo.getWithOffset(DstOff), Alignment,
         isVol ? MachineMemOperand::MOVolatile : MachineMemOperand::MONone,
         NewAAInfo);

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

@@ -1744,72 +1744,82 @@ bool AMDGPUDAGToDAGISel::SelectFlatOffsetImpl(SDNode *N, SDValue Addr,
          isFlatScratchBaseLegal(Addr))) {
       int64_t COffsetVal = cast<ConstantSDNode>(N1)->getSExtValue();
 
-      const SIInstrInfo *TII = Subtarget->getInstrInfo();
-      if (TII->isLegalFLATOffset(COffsetVal, AS, FlatVariant)) {
-        Addr = N0;
-        OffsetVal = COffsetVal;
-      } else {
-        // If the offset doesn't fit, put the low bits into the offset field and
-        // add the rest.
-        //
-        // For a FLAT instruction the hardware decides whether to access
-        // global/scratch/shared memory based on the high bits of vaddr,
-        // ignoring the offset field, so we have to ensure that when we add
-        // remainder to vaddr it still points into the same underlying object.
-        // The easiest way to do that is to make sure that we split the offset
-        // into two pieces that are both >= 0 or both <= 0.
-
-        SDLoc DL(N);
-        uint64_t RemainderOffset;
-
-        std::tie(OffsetVal, RemainderOffset) =
-            TII->splitFlatOffset(COffsetVal, AS, FlatVariant);
-
-        SDValue AddOffsetLo =
-            getMaterializedScalarImm32(Lo_32(RemainderOffset), DL);
-        SDValue Clamp = CurDAG->getTargetConstant(0, DL, MVT::i1);
-
-        if (Addr.getValueType().getSizeInBits() == 32) {
-          SmallVector<SDValue, 3> Opnds;
-          Opnds.push_back(N0);
-          Opnds.push_back(AddOffsetLo);
-          unsigned AddOp = AMDGPU::V_ADD_CO_U32_e32;
-          if (Subtarget->hasAddNoCarry()) {
-            AddOp = AMDGPU::V_ADD_U32_e64;
-            Opnds.push_back(Clamp);
-          }
-          Addr = SDValue(CurDAG->getMachineNode(AddOp, DL, MVT::i32, Opnds), 0);
+      // Adding the offset to the base address in a FLAT instruction must not
+      // change the memory aperture in which the address falls. Therefore we can
+      // only fold offsets from inbounds GEPs into FLAT instructions.
+      bool IsInBounds = Addr->getFlags().hasInBounds();
+      if (COffsetVal == 0 || FlatVariant != SIInstrFlags::FLAT || IsInBounds) {
+        const SIInstrInfo *TII = Subtarget->getInstrInfo();
+        if (TII->isLegalFLATOffset(COffsetVal, AS, FlatVariant)) {
+          Addr = N0;
+          OffsetVal = COffsetVal;
         } else {
-          // TODO: Should this try to use a scalar add pseudo if the base address
-          // is uniform and saddr is usable?
-          SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32);
-          SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32);
-
-          SDNode *N0Lo = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
-                                                DL, MVT::i32, N0, Sub0);
-          SDNode *N0Hi = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
-                                                DL, MVT::i32, N0, Sub1);
-
-          SDValue AddOffsetHi =
-              getMaterializedScalarImm32(Hi_32(RemainderOffset), DL);
-
-          SDVTList VTs = CurDAG->getVTList(MVT::i32, MVT::i1);
-
-          SDNode *Add =
-              CurDAG->getMachineNode(AMDGPU::V_ADD_CO_U32_e64, DL, VTs,
-                                     {AddOffsetLo, SDValue(N0Lo, 0), Clamp});
-
-          SDNode *Addc = CurDAG->getMachineNode(
-              AMDGPU::V_ADDC_U32_e64, DL, VTs,
-              {AddOffsetHi, SDValue(N0Hi, 0), SDValue(Add, 1), Clamp});
-
-          SDValue RegSequenceArgs[] = {
-              CurDAG->getTargetConstant(AMDGPU::VReg_64RegClassID, DL, MVT::i32),
-              SDValue(Add, 0), Sub0, SDValue(Addc, 0), Sub1};
-
-          Addr = SDValue(CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, DL,
-                                                MVT::i64, RegSequenceArgs),
-                         0);
+          // If the offset doesn't fit, put the low bits into the offset field
+          // and add the rest.
+          //
+          // For a FLAT instruction the hardware decides whether to access
+          // global/scratch/shared memory based on the high bits of vaddr,
+          // ignoring the offset field, so we have to ensure that when we add
+          // remainder to vaddr it still points into the same underlying object.
+          // The easiest way to do that is to make sure that we split the offset
+          // into two pieces that are both >= 0 or both <= 0.
+
+          SDLoc DL(N);
+          uint64_t RemainderOffset;
+
+          std::tie(OffsetVal, RemainderOffset) =
+              TII->splitFlatOffset(COffsetVal, AS, FlatVariant);
+
+          SDValue AddOffsetLo =
+              getMaterializedScalarImm32(Lo_32(RemainderOffset), DL);
+          SDValue Clamp = CurDAG->getTargetConstant(0, DL, MVT::i1);
+
+          if (Addr.getValueType().getSizeInBits() == 32) {
+            SmallVector<SDValue, 3> Opnds;
+            Opnds.push_back(N0);
+            Opnds.push_back(AddOffsetLo);
+            unsigned AddOp = AMDGPU::V_ADD_CO_U32_e32;
+            if (Subtarget->hasAddNoCarry()) {
+              AddOp = AMDGPU::V_ADD_U32_e64;
+              Opnds.push_back(Clamp);
+            }
+            Addr =
+                SDValue(CurDAG->getMachineNode(AddOp, DL, MVT::i32, Opnds), 0);
+          } else {
+            // TODO: Should this try to use a scalar add pseudo if the base
+            // address is uniform and saddr is usable?
+            SDValue Sub0 =
+                CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32);
+            SDValue Sub1 =
+                CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32);
+
+            SDNode *N0Lo = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
+                                                  DL, MVT::i32, N0, Sub0);
+            SDNode *N0Hi = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
+                                                  DL, MVT::i32, N0, Sub1);
+
+            SDValue AddOffsetHi =
+                getMaterializedScalarImm32(Hi_32(RemainderOffset), DL);
+
+            SDVTList VTs = CurDAG->getVTList(MVT::i32, MVT::i1);
+
+            SDNode *Add =
+                CurDAG->getMachineNode(AMDGPU::V_ADD_CO_U32_e64, DL, VTs,
+                                       {AddOffsetLo, SDValue(N0Lo, 0), Clamp});
+
+            SDNode *Addc = CurDAG->getMachineNode(
+                AMDGPU::V_ADDC_U32_e64, DL, VTs,
+                {AddOffsetHi, SDValue(N0Hi, 0), SDValue(Add, 1), Clamp});
+
+            SDValue RegSequenceArgs[] = {
+                CurDAG->getTargetConstant(AMDGPU::VReg_64RegClassID, DL,
+                                          MVT::i32),
+                SDValue(Add, 0), Sub0, SDValue(Addc, 0), Sub1};
+
+            Addr = SDValue(CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, DL,
+                                                  MVT::i64, RegSequenceArgs),
+                           0);
+          }
         }
       }
     }