[AArch64][GlobalISel] Refactor Combine G_CONCAT_VECTOR (#80866)

The combine now works using tablegen and checks if new instruction is
legal before creating it.

Author: ChuongGoh-arm

llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h

@@ -224,22 +224,18 @@ class CombinerHelper {
   /// - concat_vector(undef, undef) => undef
   /// - concat_vector(build_vector(A, B), build_vector(C, D)) =>
   ///   build_vector(A, B, C, D)
-  ///
-  /// \pre MI.getOpcode() == G_CONCAT_VECTORS.
-  bool tryCombineConcatVectors(MachineInstr &MI);
+  /// ==========================================================
   /// Check if the G_CONCAT_VECTORS \p MI is undef or if it
   /// can be flattened into a build_vector.
-  /// In the first case \p IsUndef will be true.
-  /// In the second case \p Ops will contain the operands needed
-  /// to produce the flattened build_vector.
+  /// In the first case \p Ops will be empty
+  /// In the second case \p Ops will contain the operands
+  /// needed to produce the flattened build_vector.
   ///
   /// \pre MI.getOpcode() == G_CONCAT_VECTORS.
-  bool matchCombineConcatVectors(MachineInstr &MI, bool &IsUndef,
-                                 SmallVectorImpl<Register> &Ops);
-  /// Replace \p MI with a flattened build_vector with \p Ops or an
-  /// implicit_def if IsUndef is true.
-  void applyCombineConcatVectors(MachineInstr &MI, bool IsUndef,
-                                 const ArrayRef<Register> Ops);
+  bool matchCombineConcatVectors(MachineInstr &MI, SmallVector<Register> &Ops);
+  /// Replace \p MI with a flattened build_vector with \p Ops
+  /// or an implicit_def if \p Ops is empty.
+  void applyCombineConcatVectors(MachineInstr &MI, SmallVector<Register> &Ops);
 
   /// Try to combine G_SHUFFLE_VECTOR into G_CONCAT_VECTORS.
   /// Returns true if MI changed.

llvm/include/llvm/Target/GlobalISel/Combine.td

@@ -1253,6 +1253,14 @@ def match_ors : GICombineRule<
         [{ return Helper.matchOr(*${root}, ${matchinfo}); }]),
   (apply [{ Helper.applyBuildFn(*${root}, ${matchinfo}); }])>;
 
+// Combines concat operations
+def concat_matchinfo : GIDefMatchData<"SmallVector<Register>">;
+def combine_concat_vector : GICombineRule<
+  (defs root:$root, concat_matchinfo:$matchinfo),
+  (match (wip_match_opcode G_CONCAT_VECTORS):$root,
+        [{ return Helper.matchCombineConcatVectors(*${root}, ${matchinfo}); }]),
+  (apply [{ Helper.applyCombineConcatVectors(*${root}, ${matchinfo}); }])>;
+
 // FIXME: These should use the custom predicate feature once it lands.
 def undef_combines : GICombineGroup<[undef_to_fp_zero, undef_to_int_zero,
                                      undef_to_negative_one,
@@ -1326,11 +1334,12 @@ def all_combines : GICombineGroup<[trivial_combines, insert_vec_elt_combines,
     intdiv_combines, mulh_combines, redundant_neg_operands,
     and_or_disjoint_mask, fma_combines, fold_binop_into_select,
     sub_add_reg, select_to_minmax, redundant_binop_in_equality,
-    fsub_to_fneg, commute_constant_to_rhs, match_ands, match_ors]>;
+    fsub_to_fneg, commute_constant_to_rhs, match_ands, match_ors, 
+    combine_concat_vector]>;
 
 // A combine group used to for prelegalizer combiners at -O0. The combines in
 // this group have been selected based on experiments to balance code size and
 // compile time performance.
 def optnone_combines : GICombineGroup<[trivial_combines,
     ptr_add_immed_chain, combines_for_extload,
-    not_cmp_fold, opt_brcond_by_inverting_cond]>;
+    not_cmp_fold, opt_brcond_by_inverting_cond, combine_concat_vector]>;

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

@@ -222,21 +222,11 @@ void CombinerHelper::applyCombineCopy(MachineInstr &MI) {
   replaceRegWith(MRI, DstReg, SrcReg);
 }
 
-bool CombinerHelper::tryCombineConcatVectors(MachineInstr &MI) {
-  bool IsUndef = false;
-  SmallVector<Register, 4> Ops;
-  if (matchCombineConcatVectors(MI, IsUndef, Ops)) {
-    applyCombineConcatVectors(MI, IsUndef, Ops);
-    return true;
-  }
-  return false;
-}
-
-bool CombinerHelper::matchCombineConcatVectors(MachineInstr &MI, bool &IsUndef,
-                                               SmallVectorImpl<Register> &Ops) {
+bool CombinerHelper::matchCombineConcatVectors(MachineInstr &MI,
+                                               SmallVector<Register> &Ops) {
   assert(MI.getOpcode() == TargetOpcode::G_CONCAT_VECTORS &&
          "Invalid instruction");
-  IsUndef = true;
+  bool IsUndef = true;
   MachineInstr *Undef = nullptr;
 
   // Walk over all the operands of concat vectors and check if they are
@@ -246,6 +236,8 @@ bool CombinerHelper::matchCombineConcatVectors(MachineInstr &MI, bool &IsUndef,
     Register Reg = MO.getReg();
     MachineInstr *Def = MRI.getVRegDef(Reg);
     assert(Def && "Operand not defined");
+    if (!MRI.hasOneNonDBGUse(Reg))
+      return false;
     switch (Def->getOpcode()) {
     case TargetOpcode::G_BUILD_VECTOR:
       IsUndef = false;
@@ -275,10 +267,21 @@ bool CombinerHelper::matchCombineConcatVectors(MachineInstr &MI, bool &IsUndef,
       return false;
     }
   }
+
+  // Check if the combine is illegal
+  LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
+  if (!isLegalOrBeforeLegalizer(
+          {TargetOpcode::G_BUILD_VECTOR, {DstTy, MRI.getType(Ops[0])}})) {
+    return false;
+  }
+
+  if (IsUndef)
+    Ops.clear();
+
   return true;
 }
-void CombinerHelper::applyCombineConcatVectors(
-    MachineInstr &MI, bool IsUndef, const ArrayRef<Register> Ops) {
+void CombinerHelper::applyCombineConcatVectors(MachineInstr &MI,
+                                               SmallVector<Register> &Ops) {
   // We determined that the concat_vectors can be flatten.
   // Generate the flattened build_vector.
   Register DstReg = MI.getOperand(0).getReg();
@@ -289,9 +292,9 @@ void CombinerHelper::applyCombineConcatVectors(
   // checking that at all Ops are undef.  Alternatively, we could have
   // generate a build_vector of undefs and rely on another combine to
   // clean that up.  For now, given we already gather this information
-  // in tryCombineConcatVectors, just save compile time and issue the
+  // in matchCombineConcatVectors, just save compile time and issue the
   // right thing.
-  if (IsUndef)
+  if (Ops.empty())
     Builder.buildUndef(NewDstReg);
   else
     Builder.buildBuildVector(NewDstReg, Ops);

llvm/lib/Target/AArch64/AArch64Combine.td

@@ -288,6 +288,6 @@ def AArch64PostLegalizerCombiner
                         constant_fold_binops, identity_combines,
                         ptr_add_immed_chain, overlapping_and,
                         split_store_zero_128, undef_combines,
-                        select_to_minmax, or_to_bsp,
+                        select_to_minmax, or_to_bsp, combine_concat_vector,
                         commute_constant_to_rhs]> {
 }

llvm/lib/Target/AArch64/GISel/AArch64O0PreLegalizerCombiner.cpp

@@ -91,8 +91,6 @@ bool AArch64O0PreLegalizerCombinerImpl::tryCombineAll(MachineInstr &MI) const {
 
   unsigned Opc = MI.getOpcode();
   switch (Opc) {
-  case TargetOpcode::G_CONCAT_VECTORS:
-    return Helper.tryCombineConcatVectors(MI);
   case TargetOpcode::G_SHUFFLE_VECTOR:
     return Helper.tryCombineShuffleVector(MI);
   case TargetOpcode::G_MEMCPY_INLINE:

llvm/lib/Target/AArch64/GISel/AArch64PreLegalizerCombiner.cpp

@@ -720,8 +720,6 @@ bool AArch64PreLegalizerCombinerImpl::tryCombineAll(MachineInstr &MI) const {
 
   unsigned Opc = MI.getOpcode();
   switch (Opc) {
-  case TargetOpcode::G_CONCAT_VECTORS:
-    return Helper.tryCombineConcatVectors(MI);
   case TargetOpcode::G_SHUFFLE_VECTOR:
     return Helper.tryCombineShuffleVector(MI);
   case TargetOpcode::G_UADDO:

llvm/lib/Target/AMDGPU/AMDGPUPreLegalizerCombiner.cpp

@@ -106,8 +106,6 @@ bool AMDGPUPreLegalizerCombinerImpl::tryCombineAll(MachineInstr &MI) const {
     return true;
 
   switch (MI.getOpcode()) {
-  case TargetOpcode::G_CONCAT_VECTORS:
-    return Helper.tryCombineConcatVectors(MI);
   case TargetOpcode::G_SHUFFLE_VECTOR:
     return Helper.tryCombineShuffleVector(MI);
   }