[ARM] Add IR intrinsics for MVE VLD[24] and VST[24].

The VST2 and VST4 instructions take two or four vector registers as
input, and store part of each register to memory in an interleaved
pattern. They come in variants indicating which part of each register
they store (VST20 and VST21; VST40 to VST43 inclusive); the intention
is that issuing each of those variants in turn has the combined effect
of loading or storing the whole set of registers to a memory block of
equal size. The corresponding VLD2 and VLD4 instructions load from
memory in the same interleaved format: each one overwrites only part
of its output register set, and again, the idea is that if you use
VLD4{0,1,2,3} or VLD2{0,1} together, you end up having written to the
whole of each register.

I've implemented the stores and loads quite differently. The loads
were easiest to implement as a single intrinsic that expands to all
four VLD4x instructions or both VLD2x, delivering four complete output
registers. (Implementing each individual load as a separate
instruction taking four input registers to partially overwrite is
possible in theory, but pointless, and when I tried it, I found it
would need extra work to get the register allocation not to be
horrible.) Since that intrinsic delivers multiple outputs, it has to
be instruction-selected in custom C++.

But the store instructions are easier to model individually, because
they don't overwrite any register at all and you can write a DAG Isel
pattern in Tablegen for each one.

Hence, my new intrinsic `int_arm_mve_vld4q` expands to four load
instructions, delivers four full output vectors, and is handled by C++
code, whereas `int_arm_mve_vst4q` expands to just one store
instruction, takes four input vectors and a constant indicating which
lanes to store, and is handled entirely in Tablegen. (And similarly
for vld2q/vst2q.) This is asymmetric, but it was the easiest way to do
each one.

Reviewers: dmgreen, miyuki, ostannard

Subscribers: kristof.beyls, hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D68700

2 files changed, 97 insertions, 0 deletions

diff --git a/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp b/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
index 6fe5e593314..59acc34906e 100644
--- a/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -232,6 +232,14 @@ private:
   void SelectMVE_VADCSBC(SDNode *N, uint16_t OpcodeWithCarry,
                          uint16_t OpcodeWithNoCarry, bool Add, bool Predicated);
 
+  /// SelectMVE_VLD - Select MVE interleaving load intrinsics. NumVecs
+  /// should be 2 or 4. The opcode array specifies the instructions
+  /// used for 8, 16 and 32-bit lane sizes respectively, and each
+  /// pointer points to a set of NumVecs sub-opcodes used for the
+  /// different stages (e.g. VLD20 versus VLD21) of each load family.
+  void SelectMVE_VLD(SDNode *N, unsigned NumVecs,
+                     const uint16_t *const *Opcodes);
+
   /// SelectVLDDup - Select NEON load-duplicate intrinsics.  NumVecs
   /// should be 1, 2, 3 or 4.  The opcode array specifies the instructions used
   /// for loading D registers.
@@ -2449,6 +2457,47 @@ void ARMDAGToDAGISel::SelectMVE_VADCSBC(SDNode *N, uint16_t OpcodeWithCarry,
   CurDAG->SelectNodeTo(N, Opcode, N->getVTList(), makeArrayRef(Ops));
 }
 
+void ARMDAGToDAGISel::SelectMVE_VLD(SDNode *N, unsigned NumVecs,
+                                    const uint16_t *const *Opcodes) {
+  EVT VT = N->getValueType(0);
+  SDLoc Loc(N);
+
+  const uint16_t *OurOpcodes;
+  switch (VT.getVectorElementType().getSizeInBits()) {
+  case 8:
+    OurOpcodes = Opcodes[0];
+    break;
+  case 16:
+    OurOpcodes = Opcodes[1];
+    break;
+  case 32:
+    OurOpcodes = Opcodes[2];
+    break;
+  default:
+    llvm_unreachable("bad vector element size in SelectMVE_VLD");
+  }
+
+  EVT DataTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64, NumVecs * 2);
+  EVT ResultTys[] = {DataTy, MVT::Other};
+
+  auto Data = SDValue(
+      CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, Loc, DataTy), 0);
+  SDValue Chain = N->getOperand(0);
+  for (unsigned Stage = 0; Stage < NumVecs; ++Stage) {
+    SDValue Ops[] = {Data, N->getOperand(2), Chain};
+    auto LoadInst =
+        CurDAG->getMachineNode(OurOpcodes[Stage], Loc, ResultTys, Ops);
+    Data = SDValue(LoadInst, 0);
+    Chain = SDValue(LoadInst, 1);
+  }
+
+  for (unsigned i = 0; i < NumVecs; i++)
+    ReplaceUses(SDValue(N, i),
+                CurDAG->getTargetExtractSubreg(ARM::qsub_0 + i, Loc, VT, Data));
+  ReplaceUses(SDValue(N, NumVecs), Chain);
+  CurDAG->RemoveDeadNode(N);
+}
+
 void ARMDAGToDAGISel::SelectVLDDup(SDNode *N, bool IsIntrinsic,
                                    bool isUpdating, unsigned NumVecs,
                                    const uint16_t *DOpcodes,
@@ -4182,6 +4231,31 @@ void ARMDAGToDAGISel::Select(SDNode *N) {
                    IntNo == Intrinsic::arm_mve_vldr_gather_base_wb_predicated);
       return;
     }
+
+    case Intrinsic::arm_mve_vld2q: {
+      static const uint16_t Opcodes8[] = {ARM::MVE_VLD20_8, ARM::MVE_VLD21_8};
+      static const uint16_t Opcodes16[] = {ARM::MVE_VLD20_16,
+                                           ARM::MVE_VLD21_16};
+      static const uint16_t Opcodes32[] = {ARM::MVE_VLD20_32,
+                                           ARM::MVE_VLD21_32};
+      static const uint16_t *const Opcodes[] = {Opcodes8, Opcodes16, Opcodes32};
+      SelectMVE_VLD(N, 2, Opcodes);
+      return;
+    }
+
+    case Intrinsic::arm_mve_vld4q: {
+      static const uint16_t Opcodes8[] = {ARM::MVE_VLD40_8, ARM::MVE_VLD41_8,
+                                          ARM::MVE_VLD42_8, ARM::MVE_VLD43_8};
+      static const uint16_t Opcodes16[] = {ARM::MVE_VLD40_16, ARM::MVE_VLD41_16,
+                                           ARM::MVE_VLD42_16,
+                                           ARM::MVE_VLD43_16};
+      static const uint16_t Opcodes32[] = {ARM::MVE_VLD40_32, ARM::MVE_VLD41_32,
+                                           ARM::MVE_VLD42_32,
+                                           ARM::MVE_VLD43_32};
+      static const uint16_t *const Opcodes[] = {Opcodes8, Opcodes16, Opcodes32};
+      SelectMVE_VLD(N, 4, Opcodes);
+      return;
+    }
     }
     break;
   }
diff --git a/llvm/lib/Target/ARM/ARMInstrMVE.td b/llvm/lib/Target/ARM/ARMInstrMVE.td
index 7d49df3d0c0..e43d64393a6 100644
--- a/llvm/lib/Target/ARM/ARMInstrMVE.td
+++ b/llvm/lib/Target/ARM/ARMInstrMVE.td
@@ -4352,6 +4352,29 @@ foreach wb = [MVE_vldst24_writeback<
                      "vst" # n.nvecs # stage # "." # s.lanesize>;
 }
 
+multiclass MVE_vst24_patterns<int lanesize, ValueType VT> {
+  foreach stage = [0,1] in
+    def : Pat<(int_arm_mve_vst2q i32:$addr,
+               (VT MQPR:$v0), (VT MQPR:$v1), (i32 stage)),
+              (!cast<Instruction>("MVE_VST2"#stage#"_"#lanesize)
+               (REG_SEQUENCE QQPR, VT:$v0, qsub_0, VT:$v1, qsub_1),
+               t2_addr_offset_none:$addr)>;
+
+  foreach stage = [0,1,2,3] in
+    def : Pat<(int_arm_mve_vst4q i32:$addr,
+               (VT MQPR:$v0), (VT MQPR:$v1),
+               (VT MQPR:$v2), (VT MQPR:$v3), (i32 stage)),
+              (!cast<Instruction>("MVE_VST4"#stage#"_"#lanesize)
+               (REG_SEQUENCE QQQQPR, VT:$v0, qsub_0, VT:$v1, qsub_1,
+                                     VT:$v2, qsub_2, VT:$v3, qsub_3),
+               t2_addr_offset_none:$addr)>;
+}
+defm : MVE_vst24_patterns<8, v16i8>;
+defm : MVE_vst24_patterns<16, v8i16>;
+defm : MVE_vst24_patterns<32, v4i32>;
+defm : MVE_vst24_patterns<16, v8f16>;
+defm : MVE_vst24_patterns<32, v4f32>;
+
 // end of MVE interleaving load/store
 
 // start of MVE predicable load/store