[SelectionDAG] Add space-optimized forms of OPC_CheckComplexPat (#73310)

We record the usage of each `ComplexPat` and sort the `ComplexPat`s
by usage.

For the top 8 `ComplexPat`s, we will emit a `OPC_CheckComplexPatN`
to save one byte.

Overall this reduces the llc binary size with all in-tree targets by
about 89K.

Author: wangpc-pp

llvm/include/llvm/CodeGen/SelectionDAGISel.h

@@ -207,6 +207,14 @@ class SelectionDAGISel : public MachineFunctionPass {
     OPC_CheckChild2CondCode,
     OPC_CheckValueType,
     OPC_CheckComplexPat,
+    OPC_CheckComplexPat0,
+    OPC_CheckComplexPat1,
+    OPC_CheckComplexPat2,
+    OPC_CheckComplexPat3,
+    OPC_CheckComplexPat4,
+    OPC_CheckComplexPat5,
+    OPC_CheckComplexPat6,
+    OPC_CheckComplexPat7,
     OPC_CheckAndImm,
     OPC_CheckOrImm,
     OPC_CheckImmAllOnesV,

llvm/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp

@@ -3358,8 +3358,18 @@ void SelectionDAGISel::SelectCodeCommon(SDNode *NodeToMatch,
         break;
       continue;
     }
-    case OPC_CheckComplexPat: {
-      unsigned CPNum = MatcherTable[MatcherIndex++];
+    case OPC_CheckComplexPat:
+    case OPC_CheckComplexPat0:
+    case OPC_CheckComplexPat1:
+    case OPC_CheckComplexPat2:
+    case OPC_CheckComplexPat3:
+    case OPC_CheckComplexPat4:
+    case OPC_CheckComplexPat5:
+    case OPC_CheckComplexPat6:
+    case OPC_CheckComplexPat7: {
+      unsigned CPNum = Opcode == OPC_CheckComplexPat
+                           ? MatcherTable[MatcherIndex++]
+                           : Opcode - OPC_CheckComplexPat0;
       unsigned RecNo = MatcherTable[MatcherIndex++];
       assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat");
 

llvm/test/TableGen/dag-isel-complexpattern.td

@@ -22,7 +22,7 @@ def CP32 : ComplexPattern<i32, 0, "SelectCP32">;
 def INSTR : Instruction {
 // CHECK-LABEL: OPC_CheckOpcode, TARGET_VAL(ISD::STORE)
 // CHECK: OPC_CheckTypeI32
-// CHECK: OPC_CheckComplexPat, /*CP*/0, /*#*/1, // SelectCP32:$
+// CHECK: OPC_CheckComplexPat0, /*#*/1, // SelectCP32:$
 // CHECK: Src: (st (add:{ *:[i32] } (CP32:{ *:[i32] }), (CP32:{ *:[i32] })), i64:{ *:[i64] }:$addr)
   let OutOperandList = (outs);
   let InOperandList = (ins GPR64:$addr);

llvm/utils/TableGen/DAGISelMatcherEmitter.cpp

@@ -63,7 +63,6 @@ class MatcherTableEmitter {
   StringMap<unsigned> PatternPredicateMap;
   std::vector<std::string> PatternPredicates;
 
-  DenseMap<const ComplexPattern*, unsigned> ComplexPatternMap;
   std::vector<const ComplexPattern*> ComplexPatterns;
 
 
@@ -84,8 +83,38 @@ class MatcherTableEmitter {
   }
 
 public:
-  MatcherTableEmitter(const CodeGenDAGPatterns &cgp)
-      : CGP(cgp), OpcodeCounts(Matcher::HighestKind + 1, 0) {}
+  MatcherTableEmitter(const Matcher *TheMatcher, const CodeGenDAGPatterns &cgp)
+      : CGP(cgp), OpcodeCounts(Matcher::HighestKind + 1, 0) {
+    // Record the usage of ComplexPattern.
+    DenseMap<const ComplexPattern *, unsigned> ComplexPatternUsage;
+
+    // Iterate the whole MatcherTable once and do some statistics.
+    std::function<void(const Matcher *)> Statistic = [&](const Matcher *N) {
+      while (N) {
+        if (auto *SM = dyn_cast<ScopeMatcher>(N))
+          for (unsigned I = 0; I < SM->getNumChildren(); I++)
+            Statistic(SM->getChild(I));
+        else if (auto *SOM = dyn_cast<SwitchOpcodeMatcher>(N))
+          for (unsigned I = 0; I < SOM->getNumCases(); I++)
+            Statistic(SOM->getCaseMatcher(I));
+        else if (auto *STM = dyn_cast<SwitchTypeMatcher>(N))
+          for (unsigned I = 0; I < STM->getNumCases(); I++)
+            Statistic(STM->getCaseMatcher(I));
+        else if (auto *CPM = dyn_cast<CheckComplexPatMatcher>(N))
+          ++ComplexPatternUsage[&CPM->getPattern()];
+        N = N->getNext();
+      }
+    };
+    Statistic(TheMatcher);
+
+    // Sort ComplexPatterns by usage.
+    std::vector<std::pair<const ComplexPattern *, unsigned>> ComplexPatternList(
+        ComplexPatternUsage.begin(), ComplexPatternUsage.end());
+    sort(ComplexPatternList,
+         [](const auto &A, const auto &B) { return A.second > B.second; });
+    for (const auto &ComplexPattern : ComplexPatternList)
+      ComplexPatterns.push_back(ComplexPattern.first);
+  }
 
   unsigned EmitMatcherList(const Matcher *N, const unsigned Indent,
                            unsigned StartIdx, raw_ostream &OS);
@@ -146,12 +175,7 @@ class MatcherTableEmitter {
     return Entry-1;
   }
   unsigned getComplexPat(const ComplexPattern &P) {
-    unsigned &Entry = ComplexPatternMap[&P];
-    if (Entry == 0) {
-      ComplexPatterns.push_back(&P);
-      Entry = ComplexPatterns.size();
-    }
-    return Entry-1;
+    return llvm::find(ComplexPatterns, &P) - ComplexPatterns.begin();
   }
 
   unsigned getNodeXFormID(Record *Rec) {
@@ -652,8 +676,13 @@ EmitMatcher(const Matcher *N, const unsigned Indent, unsigned CurrentIdx,
   case Matcher::CheckComplexPat: {
     const CheckComplexPatMatcher *CCPM = cast<CheckComplexPatMatcher>(N);
     const ComplexPattern &Pattern = CCPM->getPattern();
-    OS << "OPC_CheckComplexPat, /*CP*/" << getComplexPat(Pattern) << ", /*#*/"
-       << CCPM->getMatchNumber() << ',';
+    unsigned PatternNo = getComplexPat(Pattern);
+    if (PatternNo < 8)
+      OS << "OPC_CheckComplexPat" << PatternNo << ", /*#*/"
+         << CCPM->getMatchNumber() << ',';
+    else
+      OS << "OPC_CheckComplexPat, /*CP*/" << PatternNo << ", /*#*/"
+         << CCPM->getMatchNumber() << ',';
 
     if (!OmitComments) {
       OS << " // " << Pattern.getSelectFunc();
@@ -665,7 +694,7 @@ EmitMatcher(const Matcher *N, const unsigned Indent, unsigned CurrentIdx,
         OS << " + chain result";
     }
     OS << '\n';
-    return 3;
+    return PatternNo < 8 ? 2 : 3;
   }
 
   case Matcher::CheckAndImm: {
@@ -1267,7 +1296,7 @@ void llvm::EmitMatcherTable(Matcher *TheMatcher,
   OS << "#endif\n\n";
 
   BeginEmitFunction(OS, "void", "SelectCode(SDNode *N)", false/*AddOverride*/);
-  MatcherTableEmitter MatcherEmitter(CGP);
+  MatcherTableEmitter MatcherEmitter(TheMatcher, CGP);
 
   // First we size all the children of the three kinds of matchers that have
   // them. This is done by sharing the code in EmitMatcher(). but we don't