[TableGen] Slim down the data structures in xxxGenInstrInfo.inc, step 1

[Paul-C-Anagnostopoulos]

Summary

This simple patch reduces the size of the Offsets and OpcodeOperandTypes tables in the xxxGenInstrInfo.inc files. It is the first step in trying to reduce the size of data structures in those files.

Not until I made these changes did I realize that these two data structures are used only by the AVR target. But perhaps it will come into play in future targets.

[craig.topper]

Does any in tree target use this function? I don't see any defines of GET_INSTRINFO_OPERAND_TYPE or calls to getOperandType. AVR only defines GET_INSTRINFO_OPERAND_TYPES_ENUM

[lattner]

Looks reasonable to me, Paul mentions that only AVR is using it in tree.

[Paul-C-Anagnostopoulos]

No, Craig is correct. AVR uses the GET_INSTRINFO_OPERAND_TYPES_ENUM data, but not the GET_INSTRINFO_OPERAND_TYPE data. My search was not specific enough.

Why do you suppose someone wrote this part of InstrInfoEmitter? Could it be used by an out-of-tree target? Could it once have been used but no longer?

[lattner]

I'd look at git blame, see who contributed it. We can then contact them, or just propose deletion of the code on llvm-dev and see if anyone speaks up :)

[Paul-C-Anagnostopoulos]

Here are the Phabricator discussions. Daniel Sanders says that he knows of out-of-tree users of these data structures.

https://reviews.llvm.org/D63320

https://reviews.llvm.org/rL366278

[dsanders]

In D93208#2455809, @Paul-C-Anagnostopoulos wrote:

Here are the Phabricator discussions. Daniel Sanders says that he knows of out-of-tree users of these data structures.

https://reviews.llvm.org/D63320

https://reviews.llvm.org/rL366278

Yep, one bit I can share is that we use it to describe kinds of immediates. For example our MIRFormatter uses it to print named constants, bitfields, or plain immediates depending on this table.

[Paul-C-Anagnostopoulos]

@dsanders Which target is this?

And you're saying that you use the OpcodeOperandTypes table, not just the OperandType enum?

[dsanders]

In D93208#2456177, @Paul-C-Anagnostopoulos wrote:

@dsanders Which target is this?

It's an out-of-tree proprietary target

And you're saying that you use the OpcodeOperandTypes table, not just the OperandType enum?

That's right, we use getOperandType() which uses that table

[Paul-C-Anagnostopoulos]

Okay then, I will clean up and then push this revision.

[Paul-C-Anagnostopoulos]

Pushed.

[dsanders]

Thanks!

llvm/utils/TableGen/InstrInfoEmitter.cpp

void InstrInfoEmitter::emitOperandTypeMappings(
   OS << "#endif // GET_INSTRINFO_OPERAND_TYPES_ENUM\n\n";
 
   OS << "#ifdef GET_INSTRINFO_OPERAND_TYPE\n";
   OS << "#undef GET_INSTRINFO_OPERAND_TYPE\n";
   OS << "namespace llvm {\n";
   OS << "namespace " << Namespace << " {\n";
   OS << "LLVM_READONLY\n";
   OS << "static int getOperandType(uint16_t Opcode, uint16_t OpIdx) {\n";
-  // TODO: Factor out instructions with same operands to compress the tables.
+  // TODO: Factor out duplicate operand lists to compress the tables.
   if (!NumberedInstructions.empty()) {
     std::vector<int> OperandOffsets;
     std::vector<Record *> OperandRecords;
     int CurrentOffset = 0;
     for (const CodeGenInstruction *Inst : NumberedInstructions) {
       OperandOffsets.push_back(CurrentOffset);
       for (const auto &Op : Inst->Operands) {
         const DagInit *MIOI = Op.MIOperandInfo;
         if (!MIOI || MIOI->getNumArgs() == 0) {
           // Single, anonymous, operand.
           OperandRecords.push_back(Op.Rec);
           ++CurrentOffset;
         } else {
           for (Init *Arg : make_range(MIOI->arg_begin(), MIOI->arg_end())) {
             OperandRecords.push_back(cast<DefInit>(Arg)->getDef());
             ++CurrentOffset;
           }
         }
       }
     }
 
-    // Emit the table of offsets for the opcode lookup.
-    OS << "  const int Offsets[] = {\n";
+    // Emit the table of offsets (indexes) into the operand type table.
+    // Size the unsigned integer offset to save space.
+    assert(OperandRecords.size() <= UINT32_MAX &&
+               "Too many operands for offset table");
+    OS << ((OperandRecords.size() <= UINT16_MAX) ? "  const uint16_t"
+                                                 : "  const uint32_t");
+    OS << " Offsets[] = {\n";
     for (int I = 0, E = OperandOffsets.size(); I != E; ++I)
       OS << "    " << OperandOffsets[I] << ",\n";
     OS << "  };\n";
 
     // Add an entry for the end so that we don't need to special case it below.
     OperandOffsets.push_back(OperandRecords.size());
+
     // Emit the actual operand types in a flat table.
-    OS << "  const int OpcodeOperandTypes[] = {\n    ";
+    // Size the signed integer operand type to save space.
+    assert(EnumVal <= INT16_MAX &&
+               "Too many operand types for operand types table");
+    OS << ((EnumVal <= INT8_MAX) ? "  const int8_t" : "  const int16_t");
+    OS << " OpcodeOperandTypes[] = {\n    ";
     for (int I = 0, E = OperandRecords.size(), CurOffset = 1; I != E; ++I) {
       // We print each Opcode's operands in its own row.
       if (I == OperandOffsets[CurOffset]) {
         OS << "\n    ";
         // If there are empty rows, mark them with an empty comment.
         while (OperandOffsets[++CurOffset] == I)
           OS << "/**/\n    ";
       }