[llvm] [RISCV][MachineCombiner] Add reassociation optimizations for RVV instructions (PR #88307)

[Min-Yih Hsu via llvm-commits]

================
@@ -1619,8 +1619,228 @@ static bool isFMUL(unsigned Opc) {
   }
 }
 
+bool RISCVInstrInfo::isVectorAssociativeAndCommutative(const MachineInstr &Inst,
+                                                       bool Invert) const {
+#define OPCODE_LMUL_CASE(OPC)                                                  \
+  case RISCV::OPC##_M1:                                                        \
+  case RISCV::OPC##_M2:                                                        \
+  case RISCV::OPC##_M4:                                                        \
+  case RISCV::OPC##_M8:                                                        \
+  case RISCV::OPC##_MF2:                                                       \
+  case RISCV::OPC##_MF4:                                                       \
+  case RISCV::OPC##_MF8
+
+#define OPCODE_LMUL_MASK_CASE(OPC)                                             \
+  case RISCV::OPC##_M1_MASK:                                                   \
+  case RISCV::OPC##_M2_MASK:                                                   \
+  case RISCV::OPC##_M4_MASK:                                                   \
+  case RISCV::OPC##_M8_MASK:                                                   \
+  case RISCV::OPC##_MF2_MASK:                                                  \
+  case RISCV::OPC##_MF4_MASK:                                                  \
+  case RISCV::OPC##_MF8_MASK
+
+  unsigned Opcode = Inst.getOpcode();
+  if (Invert) {
+    if (auto InvOpcode = getInverseOpcode(Opcode))
+      Opcode = *InvOpcode;
+    else
+      return false;
+  }
+
+  // clang-format off
+  switch (Opcode) {
+  default:
+    return false;
+  OPCODE_LMUL_CASE(PseudoVADD_VV):
+  OPCODE_LMUL_MASK_CASE(PseudoVADD_VV):
+  OPCODE_LMUL_CASE(PseudoVMUL_VV):
+  OPCODE_LMUL_MASK_CASE(PseudoVMUL_VV):
+    return true;
+  }
+  // clang-format on
+
+#undef OPCODE_LMUL_MASK_CASE
+#undef OPCODE_LMUL_CASE
+}
+
+bool RISCVInstrInfo::areRVVInstsReassociable(const MachineInstr &MI1,
+                                             const MachineInstr &MI2) const {
+  if (!areOpcodesEqualOrInverse(MI1.getOpcode(), MI2.getOpcode()))
+    return false;
+
+  assert(MI1.getMF() == MI2.getMF());
+  const MachineRegisterInfo *MRI = &MI1.getMF()->getRegInfo();
+  const TargetRegisterInfo *TRI = MRI->getTargetRegisterInfo();
+
+  // Make sure vtype operands are also the same.
+  const MCInstrDesc &Desc = get(MI1.getOpcode());
+  const uint64_t TSFlags = Desc.TSFlags;
+
+  auto checkImmOperand = [&](unsigned OpIdx) {
+    return MI1.getOperand(OpIdx).getImm() == MI2.getOperand(OpIdx).getImm();
+  };
+
+  auto checkRegOperand = [&](unsigned OpIdx) {
+    return MI1.getOperand(OpIdx).getReg() == MI2.getOperand(OpIdx).getReg();
+  };
+
+  // PassThru
+  if (!checkRegOperand(1))
----------------
mshockwave wrote:

Good point. I think this holds when Root (MI1) has noreg for passthrough. When Root has non-noreg passthrough and Prev (MI2) has noreg passthrough things get a little tricky, as we also need to check if Root's passthrough is actually Prev's result.
But in any case, I'm only adding the most basic version of reassociation optimization for RVV in this patch, so I put this idea in a TODO comment.

https://github.com/llvm/llvm-project/pull/88307