[llvm] r191652 - Filter out repeated sections from the X86 disassembler modRMTable. Saves about ~43K from a released build. Unfortunately the disassembler tables are still upwards of 800K.

Craig Topper craig.topper at gmail.com
Sun Sep 29 23:23:19 PDT 2013


Author: ctopper
Date: Mon Sep 30 01:23:19 2013
New Revision: 191652

URL: http://llvm.org/viewvc/llvm-project?rev=191652&view=rev
Log:
Filter out repeated sections from the X86 disassembler modRMTable. Saves about ~43K from a released build. Unfortunately the disassembler tables are still upwards of 800K.

Modified:
    llvm/trunk/utils/TableGen/X86DisassemblerTables.cpp
    llvm/trunk/utils/TableGen/X86DisassemblerTables.h

Modified: llvm/trunk/utils/TableGen/X86DisassemblerTables.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/utils/TableGen/X86DisassemblerTables.cpp?rev=191652&r1=191651&r2=191652&view=diff
==============================================================================
--- llvm/trunk/utils/TableGen/X86DisassemblerTables.cpp (original)
+++ llvm/trunk/utils/TableGen/X86DisassemblerTables.cpp Mon Sep 30 01:23:19 2013
@@ -265,35 +265,6 @@ static inline const char* stringForOpera
   }
 }
 
-void DisassemblerTables::emitOneID(raw_ostream &o, unsigned &i, InstrUID id,
-                                   bool addComma) const {
-  if (id)
-    o.indent(i * 2) << format("0x%hx", id);
-  else
-    o.indent(i * 2) << 0;
-
-  if (addComma)
-    o << ", ";
-  else
-    o << "  ";
-
-  o << "/* ";
-  o << InstructionSpecifiers[id].name;
-  o << "*/";
-
-  o << "\n";
-}
-
-/// emitEmptyTable - Emits the modRMEmptyTable, which is used as a ID table by
-///   all ModR/M decisions for instructions that are invalid for all possible
-///   ModR/M byte values.
-///
-/// @param o        - The output stream on which to emit the table.
-/// @param i        - The indentation level for that output stream.
-static void emitEmptyTable(raw_ostream &o, unsigned &i) {
-  o.indent(i * 2) << "0x0, /* EmptyTable */\n";
-}
-
 /// getDecisionType - Determines whether a ModRM decision with 255 entries can
 ///   be compacted by eliminating redundant information.
 ///
@@ -393,6 +364,7 @@ DisassemblerTables::~DisassemblerTables(
 
 void DisassemblerTables::emitModRMDecision(raw_ostream &o1, raw_ostream &o2,
                                            unsigned &i1, unsigned &i2,
+                                           unsigned &ModRMTableNum,
                                            ModRMDecision &decision) const {
   static uint32_t sTableNumber = 0;
   static uint32_t sEntryNumber = 1;
@@ -411,44 +383,56 @@ void DisassemblerTables::emitModRMDecisi
     return;
   }
 
-  o1 << "/* Table" << sTableNumber << " */\n";
-  i1++;
+  std::vector<unsigned> ModRMDecision;
 
   switch (dt) {
     default:
       llvm_unreachable("Unknown decision type");
     case MODRM_ONEENTRY:
-      emitOneID(o1, i1, decision.instructionIDs[0], true);
+      ModRMDecision.push_back(decision.instructionIDs[0]);
       break;
     case MODRM_SPLITRM:
-      emitOneID(o1, i1, decision.instructionIDs[0x00], true); // mod = 0b00
-      emitOneID(o1, i1, decision.instructionIDs[0xc0], true); // mod = 0b11
+      ModRMDecision.push_back(decision.instructionIDs[0x00]);
+      ModRMDecision.push_back(decision.instructionIDs[0xc0]);
       break;
     case MODRM_SPLITREG:
       for (unsigned index = 0; index < 64; index += 8)
-        emitOneID(o1, i1, decision.instructionIDs[index], true);
+        ModRMDecision.push_back(decision.instructionIDs[index]);
       for (unsigned index = 0xc0; index < 256; index += 8)
-        emitOneID(o1, i1, decision.instructionIDs[index], true);
+        ModRMDecision.push_back(decision.instructionIDs[index]);
       break;
     case MODRM_SPLITMISC:
       for (unsigned index = 0; index < 64; index += 8)
-        emitOneID(o1, i1, decision.instructionIDs[index], true);
+        ModRMDecision.push_back(decision.instructionIDs[index]);
       for (unsigned index = 0xc0; index < 256; ++index)
-        emitOneID(o1, i1, decision.instructionIDs[index], true);
+        ModRMDecision.push_back(decision.instructionIDs[index]);
       break;
     case MODRM_FULL:
       for (unsigned index = 0; index < 256; ++index)
-        emitOneID(o1, i1, decision.instructionIDs[index], true);
+        ModRMDecision.push_back(decision.instructionIDs[index]);
       break;
   }
 
-  i1--;
+  unsigned &EntryNumber = ModRMTable[ModRMDecision];
+  if (EntryNumber == 0) {
+    EntryNumber = ModRMTableNum;
+
+    ModRMTableNum += ModRMDecision.size();
+    o1 << "/* Table" << EntryNumber << " */\n";
+    i1++;
+    for (std::vector<unsigned>::const_iterator I = ModRMDecision.begin(),
+           E = ModRMDecision.end(); I != E; ++I) {
+      o1.indent(i1 * 2) << format("0x%hx", *I) << ", /* "
+                        << InstructionSpecifiers[*I].name << " */\n";
+    }
+    i1--;
+  }
 
   o2.indent(i2) << "{ /* struct ModRMDecision */" << "\n";
   i2++;
 
   o2.indent(i2) << stringForDecisionType(dt) << "," << "\n";
-  o2.indent(i2) << sEntryNumber << " /* Table" << sTableNumber << " */\n";
+  o2.indent(i2) << EntryNumber << " /* Table" << EntryNumber << " */\n";
 
   i2--;
   o2.indent(i2) << "}";
@@ -482,6 +466,7 @@ void DisassemblerTables::emitModRMDecisi
 
 void DisassemblerTables::emitOpcodeDecision(raw_ostream &o1, raw_ostream &o2,
                                             unsigned &i1, unsigned &i2,
+                                            unsigned &ModRMTableNum,
                                             OpcodeDecision &decision) const {
   o2.indent(i2) << "{ /* struct OpcodeDecision */" << "\n";
   i2++;
@@ -493,7 +478,8 @@ void DisassemblerTables::emitOpcodeDecis
 
     o2 << "/* 0x" << format("%02hhx", index) << " */" << "\n";
 
-    emitModRMDecision(o1, o2, i1, i2, decision.modRMDecisions[index]);
+    emitModRMDecision(o1, o2, i1, i2, ModRMTableNum,
+                      decision.modRMDecisions[index]);
 
     if (index <  255)
       o2 << ",";
@@ -509,6 +495,7 @@ void DisassemblerTables::emitOpcodeDecis
 
 void DisassemblerTables::emitContextDecision(raw_ostream &o1, raw_ostream &o2,
                                              unsigned &i1, unsigned &i2,
+                                             unsigned &ModRMTableNum,
                                              ContextDecision &decision,
                                              const char* name) const {
   o2.indent(i2) << "static const struct ContextDecision " << name << " = {\n";
@@ -522,7 +509,8 @@ void DisassemblerTables::emitContextDeci
     o2 << " */";
     o2 << "\n";
 
-    emitOpcodeDecision(o1, o2, i1, i2, decision.opcodeDecisions[index]);
+    emitOpcodeDecision(o1, o2, i1, i2, ModRMTableNum,
+                       decision.opcodeDecisions[index]);
 
     if (index + 1 < IC_max)
       o2 << ", ";
@@ -711,13 +699,14 @@ void DisassemblerTables::emitContextTabl
 }
 
 void DisassemblerTables::emitContextDecisions(raw_ostream &o1, raw_ostream &o2,
-                                             unsigned &i1, unsigned &i2) const {
-  emitContextDecision(o1, o2, i1, i2, *Tables[0], ONEBYTE_STR);
-  emitContextDecision(o1, o2, i1, i2, *Tables[1], TWOBYTE_STR);
-  emitContextDecision(o1, o2, i1, i2, *Tables[2], THREEBYTE38_STR);
-  emitContextDecision(o1, o2, i1, i2, *Tables[3], THREEBYTE3A_STR);
-  emitContextDecision(o1, o2, i1, i2, *Tables[4], THREEBYTEA6_STR);
-  emitContextDecision(o1, o2, i1, i2, *Tables[5], THREEBYTEA7_STR);
+                                              unsigned &i1, unsigned &i2,
+                                              unsigned &ModRMTableNum) const {
+  emitContextDecision(o1, o2, i1, i2, ModRMTableNum, *Tables[0], ONEBYTE_STR);
+  emitContextDecision(o1, o2, i1, i2, ModRMTableNum, *Tables[1], TWOBYTE_STR);
+  emitContextDecision(o1, o2, i1, i2, ModRMTableNum, *Tables[2], THREEBYTE38_STR);
+  emitContextDecision(o1, o2, i1, i2, ModRMTableNum, *Tables[3], THREEBYTE3A_STR);
+  emitContextDecision(o1, o2, i1, i2, ModRMTableNum, *Tables[4], THREEBYTEA6_STR);
+  emitContextDecision(o1, o2, i1, i2, ModRMTableNum, *Tables[5], THREEBYTEA7_STR);
 }
 
 void DisassemblerTables::emit(raw_ostream &o) const {
@@ -736,11 +725,17 @@ void DisassemblerTables::emit(raw_ostrea
   emitContextTable(o, i2);
   o << "\n";
 
+  unsigned ModRMTableNum = 0;
+
   o << "static const InstrUID modRMTable[] = {\n";
   i1++;
-  emitEmptyTable(o1, i1);
+  std::vector<unsigned> EmptyTable(1, 0);
+  ModRMTable[EmptyTable] = ModRMTableNum;
+  ModRMTableNum += EmptyTable.size();
+  o1 << "/* EmptyTable */\n";
+  o1.indent(i1 * 2) << "0x0,\n";
   i1--;
-  emitContextDecisions(o1, o2, i1, i2);
+  emitContextDecisions(o1, o2, i1, i2, ModRMTableNum);
 
   o << o1.str();
   o << "  0x0\n";

Modified: llvm/trunk/utils/TableGen/X86DisassemblerTables.h
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/utils/TableGen/X86DisassemblerTables.h?rev=191652&r1=191651&r2=191652&view=diff
==============================================================================
--- llvm/trunk/utils/TableGen/X86DisassemblerTables.h (original)
+++ llvm/trunk/utils/TableGen/X86DisassemblerTables.h Mon Sep 30 01:23:19 2013
@@ -20,6 +20,7 @@
 #include "X86DisassemblerShared.h"
 #include "X86ModRMFilters.h"
 #include "llvm/Support/raw_ostream.h"
+#include <map>
 #include <vector>
 
 namespace llvm {
@@ -41,28 +42,16 @@ private:
   /// [5] three-byte opcodes of the form 0f a7 __
   ContextDecision* Tables[6];
 
+  // Table of ModRM encodings.
+  typedef std::map<std::vector<unsigned>, unsigned> ModRMMapTy;
+  mutable ModRMMapTy ModRMTable;
+
   /// The instruction information table
   std::vector<InstructionSpecifier> InstructionSpecifiers;
 
   /// True if there are primary decode conflicts in the instruction set
   bool HasConflicts;
 
-  /// emitOneID - Emits a table entry for a single instruction entry, at the
-  ///   innermost level of the structure hierarchy.  The entry is printed out
-  ///   in the format "nnnn, /* MNEMONIC */" where nnnn is the ID in decimal,
-  ///   the comma is printed if addComma is true, and the menonic is the name
-  ///   of the instruction as listed in the LLVM tables.
-  ///
-  /// @param o        - The output stream to print the entry on.
-  /// @param i        - The indentation level for o.
-  /// @param id       - The unique ID of the instruction to print.
-  /// @param addComma - Whether or not to print a comma after the ID.  True if
-  ///                    additional items will follow.
-  void emitOneID(raw_ostream &o,
-                 uint32_t &i,
-                 InstrUID id,
-                 bool addComma) const;
-
   /// emitModRMDecision - Emits a table of entries corresponding to a single
   ///   ModR/M decision.  Compacts the ModR/M decision if possible.  ModR/M
   ///   decisions are printed as:
@@ -91,12 +80,11 @@ private:
   /// @param o2       - The output stream to print the decision structure to.
   /// @param i1       - The indentation level to use with stream o1.
   /// @param i2       - The indentation level to use with stream o2.
+  /// @param ModRMTableNum - next table number for adding to ModRMTable.
   /// @param decision - The ModR/M decision to emit.  This decision has 256
   ///                   entries - emitModRMDecision decides how to compact it.
-  void emitModRMDecision(raw_ostream &o1,
-                         raw_ostream &o2,
-                         uint32_t &i1,
-                         uint32_t &i2,
+  void emitModRMDecision(raw_ostream &o1, raw_ostream &o2,
+                         unsigned &i1, unsigned &i2, unsigned &ModRMTableNum,
                          ModRMDecision &decision) const;
 
   /// emitOpcodeDecision - Emits an OpcodeDecision and all its subsidiary ModR/M
@@ -120,12 +108,11 @@ private:
   /// @param o2       - The output stream for the decision structure itself.
   /// @param i1       - The indent level to use with stream o1.
   /// @param i2       - The indent level to use with stream o2.
+  /// @param ModRMTableNum - next table number for adding to ModRMTable.
   /// @param decision - The OpcodeDecision to emit along with its subsidiary
   ///                    structures.
-  void emitOpcodeDecision(raw_ostream &o1,
-                          raw_ostream &o2,
-                          uint32_t &i1,
-                          uint32_t &i2,
+  void emitOpcodeDecision(raw_ostream &o1, raw_ostream &o2,
+                          unsigned &i1, unsigned &i2, unsigned &ModRMTableNum,
                           OpcodeDecision &decision) const;
 
   /// emitContextDecision - Emits a ContextDecision and all its subsidiary
@@ -155,15 +142,13 @@ private:
   /// @param o2       - The output stream to print the decision structure to.
   /// @param i1       - The indent level to use with stream o1.
   /// @param i2       - The indent level to use with stream o2.
+  /// @param ModRMTableNum - next table number for adding to ModRMTable.
   /// @param decision - The ContextDecision to emit along with its subsidiary
   ///                   structures.
   /// @param name     - The name for the ContextDecision.
-  void emitContextDecision(raw_ostream &o1,
-                           raw_ostream &o2,
-                           uint32_t &i1,
-                           uint32_t &i2,
-                           ContextDecision &decision,
-                           const char* name) const;
+  void emitContextDecision(raw_ostream &o1, raw_ostream &o2,
+                           unsigned &i1, unsigned &i2, unsigned &ModRMTableNum,
+                           ContextDecision &decision, const char* name) const;
 
   /// emitInstructionInfo - Prints the instruction specifier table, which has
   ///   one entry for each instruction, and contains name and operand
@@ -194,7 +179,7 @@ private:
   /// @param o  - The output stream to which the instruction table should be
   ///             written.
   /// @param i  - The indent level for use with the stream.
-  void emitInstructionInfo(raw_ostream &o, uint32_t &i) const;
+  void emitInstructionInfo(raw_ostream &o, unsigned &i) const;
 
   /// emitContextTable - Prints the table that is used to translate from an
   ///   instruction attribute mask to an instruction context.  This table is
@@ -220,10 +205,10 @@ private:
   /// @param o2 - The output stream to print the decision structures to.
   /// @param i1 - The indent level to use with stream o1.
   /// @param i2 - The indent level to use with stream o2.
-  void emitContextDecisions(raw_ostream &o1,
-                            raw_ostream &o2,
-                            uint32_t &i1,
-                            uint32_t &i2) const;
+  /// @param ModRMTableNum - next table number for adding to ModRMTable.
+  void emitContextDecisions(raw_ostream &o1, raw_ostream &o2,
+                            unsigned &i1, unsigned &i2,
+                            unsigned &ModRMTableNum) const;
 
   /// setTableFields - Uses a ModRMFilter to set the appropriate entries in a
   ///   ModRMDecision to refer to a particular instruction ID.





More information about the llvm-commits mailing list