[AMDGPU][SDAG] Initial support for ISD::PTRADD (#141725)

Enable generation of PTRADD SelectionDAG nodes for pointer arithmetic for SI,
for now behind an internal CLI option. Also add basic patterns to match these
nodes. Optimizations will come in follow-up PRs. Basic tests for SDAG codegen
with PTRADD are in test/CodeGen/AMDGPU/ptradd-sdag.ll

Only affects 64-bit address spaces for now, since the immediate use case only
affects the flat address space.

For SWDEV-516125.

Author: ritter-x2a

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -61,6 +61,14 @@ static cl::opt<bool> UseDivergentRegisterIndexing(
     cl::desc("Use indirect register addressing for divergent indexes"),
     cl::init(false));
 
+// TODO: This option should be removed once we switch to always using PTRADD in
+// the SelectionDAG.
+static cl::opt<bool> UseSelectionDAGPTRADD(
+    "amdgpu-use-sdag-ptradd", cl::Hidden,
+    cl::desc("Generate ISD::PTRADD nodes for 64-bit pointer arithmetic in the "
+             "SelectionDAG ISel"),
+    cl::init(false));
+
 static bool denormalModeIsFlushAllF32(const MachineFunction &MF) {
   const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
   return Info->getMode().FP32Denormals == DenormalMode::getPreserveSign();
@@ -10457,6 +10465,11 @@ SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
   }
 }
 
+bool SITargetLowering::shouldPreservePtrArith(const Function &F,
+                                              EVT PtrVT) const {
+  return UseSelectionDAGPTRADD && PtrVT == MVT::i64;
+}
+
 // The raw.(t)buffer and struct.(t)buffer intrinsics have two offset args:
 // offset (the offset that is included in bounds checking and swizzling, to be
 // split between the instruction's voffset and immoffset fields) and soffset

llvm/lib/Target/AMDGPU/SIISelLowering.h

@@ -260,6 +260,8 @@ class SITargetLowering final : public AMDGPUTargetLowering {
 
   bool shouldExpandVectorDynExt(SDNode *N) const;
 
+  bool shouldPreservePtrArith(const Function &F, EVT PtrVT) const override;
+
 private:
   // Analyze a combined offset from an amdgcn_s_buffer_load intrinsic and store
   // the three offsets (voffset, soffset and instoffset) into the SDValue[3]

llvm/lib/Target/AMDGPU/SIInstructions.td

@@ -1376,6 +1376,20 @@ def : GCNPat <
       (i32 (V_MOV_B32_e32 (i32 0))), sub1)
 >;
 
+//===----------------------------------------------------------------------===//
+// PTRADD Patterns
+//===----------------------------------------------------------------------===//
+
+// GlobalISel shouldn't generate 64-bit addition pseudos.
+let GISelShouldIgnore = 1 in {
+def : GCNPat<
+  (DivergentBinFrag<ptradd> i64:$src0, i64:$src1),
+  (V_ADD_U64_PSEUDO $src0, $src1)>;
+def : GCNPat<
+  (UniformBinFrag<ptradd> i64:$src0, i64:$src1),
+  (S_ADD_U64_PSEUDO $src0, $src1)>;
+}
+
 /********** ============================================ **********/
 /********** Extraction, Insertion, Building and Casting  **********/
 /********** ============================================ **********/