[AMDGPU][SDAG] Handle ISD::PTRADD in various special cases (#145330)

ritter-x2a · web-flow · commit a2dcc88f39a8 · 2025-09-19T10:19:38.000+02:00
There are more places in SIISelLowering.cpp and AMDGPUISelDAGToDAG.cpp
that check for ISD::ADD in a pointer context, but as far as I can tell
those are only relevant for 32-bit pointer arithmetic (like frame
indices/scratch addresses and LDS), for which we don't enable PTRADD
generation yet.

For SWDEV-516125.
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -8600,7 +8600,7 @@ static bool isMemSrcFromConstant(SDValue Src, ConstantDataArraySlice &Slice) {
   GlobalAddressSDNode *G = nullptr;
   if (Src.getOpcode() == ISD::GlobalAddress)
     G = cast<GlobalAddressSDNode>(Src);
-  else if (Src.getOpcode() == ISD::ADD &&
+  else if (Src->isAnyAdd() &&
            Src.getOperand(0).getOpcode() == ISD::GlobalAddress &&
            Src.getOperand(1).getOpcode() == ISD::Constant) {
     G = cast<GlobalAddressSDNode>(Src.getOperand(0));
diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -638,8 +638,14 @@ bool TargetLowering::ShrinkDemandedOp(SDValue Op, unsigned BitWidth,
       // operands on the new node are also disjoint.
       SDNodeFlags Flags(Op->getFlags().hasDisjoint() ? SDNodeFlags::Disjoint
                                                      : SDNodeFlags::None);
+      unsigned Opcode = Op.getOpcode();
+      if (Opcode == ISD::PTRADD) {
+        // It isn't a ptradd anymore if it doesn't operate on the entire
+        // pointer.
+        Opcode = ISD::ADD;
+      }
       SDValue X = DAG.getNode(
-          Op.getOpcode(), dl, SmallVT,
+          Opcode, dl, SmallVT,
           DAG.getNode(ISD::TRUNCATE, dl, SmallVT, Op.getOperand(0)),
           DAG.getNode(ISD::TRUNCATE, dl, SmallVT, Op.getOperand(1)), Flags);
       assert(DemandedSize <= SmallVTBits && "Narrowed below demanded bits?");
@@ -2860,6 +2866,11 @@ bool TargetLowering::SimplifyDemandedBits(
       return TLO.CombineTo(Op, And1);
     }
     [[fallthrough]];
+  case ISD::PTRADD:
+    if (Op.getOperand(0).getValueType() != Op.getOperand(1).getValueType())
+      break;
+    // PTRADD behaves like ADD if pointers are represented as integers.
+    [[fallthrough]];
   case ISD::ADD:
   case ISD::SUB: {
     // Add, Sub, and Mul don't demand any bits in positions beyond that
@@ -2969,9 +2980,9 @@ bool TargetLowering::SimplifyDemandedBits(
 
     if (Op.getOpcode() == ISD::MUL) {
       Known = KnownBits::mul(KnownOp0, KnownOp1);
-    } else { // Op.getOpcode() is either ISD::ADD or ISD::SUB.
+    } else { // Op.getOpcode() is either ISD::ADD, ISD::PTRADD, or ISD::SUB.
       Known = KnownBits::computeForAddSub(
-          Op.getOpcode() == ISD::ADD, Flags.hasNoSignedWrap(),
+          Op.getOpcode() != ISD::SUB, Flags.hasNoSignedWrap(),
           Flags.hasNoUnsignedWrap(), KnownOp0, KnownOp1);
     }
     break;
@@ -5679,7 +5690,7 @@ bool TargetLowering::isGAPlusOffset(SDNode *WN, const GlobalValue *&GA,
     return true;
   }
 
-  if (N->getOpcode() == ISD::ADD) {
+  if (N->isAnyAdd()) {
     SDValue N1 = N->getOperand(0);
     SDValue N2 = N->getOperand(1);
     if (isGAPlusOffset(N1.getNode(), GA, Offset)) {
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
@@ -1531,7 +1531,7 @@ bool AMDGPUDAGToDAGISel::SelectMUBUF(SDValue Addr, SDValue &Ptr, SDValue &VAddr,
       C1 = nullptr;
   }
 
-  if (N0.getOpcode() == ISD::ADD) {
+  if (N0->isAnyAdd()) {
     // (add N2, N3) -> addr64, or
     // (add (add N2, N3), C1) -> addr64
     SDValue N2 = N0.getOperand(0);
@@ -1993,7 +1993,7 @@ bool AMDGPUDAGToDAGISel::SelectGlobalSAddr(SDNode *N, SDValue Addr,
   }
 
   // Match the variable offset.
-  if (Addr.getOpcode() == ISD::ADD) {
+  if (Addr->isAnyAdd()) {
     LHS = Addr.getOperand(0);
 
     if (!LHS->isDivergent()) {
@@ -2495,7 +2495,7 @@ bool AMDGPUDAGToDAGISel::SelectSMRDBaseOffset(SDNode *N, SDValue Addr,
 
   SDValue N0, N1;
   // Extract the base and offset if possible.
-  if (CurDAG->isBaseWithConstantOffset(Addr) || Addr.getOpcode() == ISD::ADD) {
+  if (Addr->isAnyAdd() || CurDAG->isADDLike(Addr)) {
     N0 = Addr.getOperand(0);
     N1 = Addr.getOperand(1);
   } else if (getBaseWithOffsetUsingSplitOR(*CurDAG, Addr, N0, N1)) {
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -11298,7 +11298,7 @@ SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
     SDValue VOffset;
     // Try to split SAddr and VOffset. Global and LDS pointers share the same
     // immediate offset, so we cannot use a regular SelectGlobalSAddr().
-    if (Addr->isDivergent() && Addr.getOpcode() == ISD::ADD) {
+    if (Addr->isDivergent() && Addr->isAnyAdd()) {
       SDValue LHS = Addr.getOperand(0);
       SDValue RHS = Addr.getOperand(1);
 
@@ -12916,8 +12916,7 @@ SDValue SITargetLowering::performSHLPtrCombine(SDNode *N, unsigned AddrSpace,
 
   // We only do this to handle cases where it's profitable when there are
   // multiple uses of the add, so defer to the standard combine.
-  if ((N0.getOpcode() != ISD::ADD && N0.getOpcode() != ISD::OR) ||
-      N0->hasOneUse())
+  if ((!N0->isAnyAdd() && N0.getOpcode() != ISD::OR) || N0->hasOneUse())
     return SDValue();
 
   const ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(N1);
@@ -12956,6 +12955,8 @@ SDValue SITargetLowering::performSHLPtrCombine(SDNode *N, unsigned AddrSpace,
       N->getFlags().hasNoUnsignedWrap() &&
       (N0.getOpcode() == ISD::OR || N0->getFlags().hasNoUnsignedWrap()));
 
+  // Use ISD::ADD even if the original operation was ISD::PTRADD, since we can't
+  // be sure that the new left operand is a proper base pointer.
   return DAG.getNode(ISD::ADD, SL, VT, ShlX, COffset, Flags);
 }
 
diff --git a/llvm/test/CodeGen/AMDGPU/ptradd-sdag-mubuf.ll b/llvm/test/CodeGen/AMDGPU/ptradd-sdag-mubuf.ll
@@ -5,50 +5,26 @@
 ; Test PTRADD handling in AMDGPUDAGToDAGISel::SelectMUBUF.
 
 define amdgpu_kernel void @v_add_i32(ptr addrspace(1) %out, ptr addrspace(1) %in) {
-; GFX6_PTRADD-LABEL: v_add_i32:
-; GFX6_PTRADD:       ; %bb.0:
-; GFX6_PTRADD-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x0
-; GFX6_PTRADD-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX6_PTRADD-NEXT:    s_mov_b32 s7, 0x100f000
-; GFX6_PTRADD-NEXT:    s_mov_b32 s10, 0
-; GFX6_PTRADD-NEXT:    s_mov_b32 s11, s7
-; GFX6_PTRADD-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX6_PTRADD-NEXT:    v_mov_b32_e32 v1, s3
-; GFX6_PTRADD-NEXT:    v_add_i32_e32 v0, vcc, s2, v0
-; GFX6_PTRADD-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; GFX6_PTRADD-NEXT:    s_mov_b32 s8, s10
-; GFX6_PTRADD-NEXT:    s_mov_b32 s9, s10
-; GFX6_PTRADD-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64 glc
-; GFX6_PTRADD-NEXT:    s_waitcnt vmcnt(0)
-; GFX6_PTRADD-NEXT:    buffer_load_dword v0, v[0:1], s[8:11], 0 addr64 offset:4 glc
-; GFX6_PTRADD-NEXT:    s_waitcnt vmcnt(0)
-; GFX6_PTRADD-NEXT:    s_mov_b32 s6, -1
-; GFX6_PTRADD-NEXT:    s_mov_b32 s4, s0
-; GFX6_PTRADD-NEXT:    s_mov_b32 s5, s1
-; GFX6_PTRADD-NEXT:    v_add_i32_e32 v0, vcc, v2, v0
-; GFX6_PTRADD-NEXT:    buffer_store_dword v0, off, s[4:7], 0
-; GFX6_PTRADD-NEXT:    s_endpgm
-;
-; GFX6_LEGACY-LABEL: v_add_i32:
-; GFX6_LEGACY:       ; %bb.0:
-; GFX6_LEGACY-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x0
-; GFX6_LEGACY-NEXT:    s_mov_b32 s7, 0x100f000
-; GFX6_LEGACY-NEXT:    s_mov_b32 s10, 0
-; GFX6_LEGACY-NEXT:    s_mov_b32 s11, s7
-; GFX6_LEGACY-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX6_LEGACY-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX6_LEGACY-NEXT:    s_mov_b64 s[8:9], s[2:3]
-; GFX6_LEGACY-NEXT:    v_mov_b32_e32 v1, 0
-; GFX6_LEGACY-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64 glc
-; GFX6_LEGACY-NEXT:    s_waitcnt vmcnt(0)
-; GFX6_LEGACY-NEXT:    buffer_load_dword v0, v[0:1], s[8:11], 0 addr64 offset:4 glc
-; GFX6_LEGACY-NEXT:    s_waitcnt vmcnt(0)
-; GFX6_LEGACY-NEXT:    s_mov_b32 s6, -1
-; GFX6_LEGACY-NEXT:    s_mov_b32 s4, s0
-; GFX6_LEGACY-NEXT:    s_mov_b32 s5, s1
-; GFX6_LEGACY-NEXT:    v_add_i32_e32 v0, vcc, v2, v0
-; GFX6_LEGACY-NEXT:    buffer_store_dword v0, off, s[4:7], 0
-; GFX6_LEGACY-NEXT:    s_endpgm
+; GFX6-LABEL: v_add_i32:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x0
+; GFX6-NEXT:    s_mov_b32 s7, 0x100f000
+; GFX6-NEXT:    s_mov_b32 s10, 0
+; GFX6-NEXT:    s_mov_b32 s11, s7
+; GFX6-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX6-NEXT:    s_mov_b64 s[8:9], s[2:3]
+; GFX6-NEXT:    v_mov_b32_e32 v1, 0
+; GFX6-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    buffer_load_dword v0, v[0:1], s[8:11], 0 addr64 offset:4 glc
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    s_mov_b32 s6, -1
+; GFX6-NEXT:    s_mov_b32 s4, s0
+; GFX6-NEXT:    s_mov_b32 s5, s1
+; GFX6-NEXT:    v_add_i32_e32 v0, vcc, v2, v0
+; GFX6-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX6-NEXT:    s_endpgm
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr inbounds i32, ptr addrspace(1) %in, i32 %tid
   %b_ptr = getelementptr i32, ptr addrspace(1) %gep, i32 1
@@ -60,4 +36,5 @@ define amdgpu_kernel void @v_add_i32(ptr addrspace(1) %out, ptr addrspace(1) %in
 }
 
 ;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
-; GFX6: {{.*}}
+; GFX6_LEGACY: {{.*}}
+; GFX6_PTRADD: {{.*}}
diff --git a/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll b/llvm/test/CodeGen/AMDGPU/ptradd-sdag-optimizations.ll
