Home | History | Annotate | Download | only in TableGen
      1 //===- DisassemblerEmitter.cpp - Generate a disassembler ------------------===//
      2 //
      3 //                     The LLVM Compiler Infrastructure
      4 //
      5 // This file is distributed under the University of Illinois Open Source
      6 // License. See LICENSE.TXT for details.
      7 //
      8 //===----------------------------------------------------------------------===//
      9 
     10 #include "DisassemblerEmitter.h"
     11 #include "CodeGenTarget.h"
     12 #include "X86DisassemblerTables.h"
     13 #include "X86RecognizableInstr.h"
     14 #include "ARMDecoderEmitter.h"
     15 #include "FixedLenDecoderEmitter.h"
     16 #include "llvm/TableGen/Error.h"
     17 #include "llvm/TableGen/Record.h"
     18 
     19 using namespace llvm;
     20 using namespace llvm::X86Disassembler;
     21 
     22 /// DisassemblerEmitter - Contains disassembler table emitters for various
     23 /// architectures.
     24 
     25 /// X86 Disassembler Emitter
     26 ///
     27 /// *** IF YOU'RE HERE TO RESOLVE A "Primary decode conflict", LOOK DOWN NEAR
     28 ///     THE END OF THIS COMMENT!
     29 ///
     30 /// The X86 disassembler emitter is part of the X86 Disassembler, which is
     31 /// documented in lib/Target/X86/X86Disassembler.h.
     32 ///
     33 /// The emitter produces the tables that the disassembler uses to translate
     34 /// instructions.  The emitter generates the following tables:
     35 ///
     36 /// - One table (CONTEXTS_SYM) that contains a mapping of attribute masks to
     37 ///   instruction contexts.  Although for each attribute there are cases where
     38 ///   that attribute determines decoding, in the majority of cases decoding is
     39 ///   the same whether or not an attribute is present.  For example, a 64-bit
     40 ///   instruction with an OPSIZE prefix and an XS prefix decodes the same way in
     41 ///   all cases as a 64-bit instruction with only OPSIZE set.  (The XS prefix
     42 ///   may have effects on its execution, but does not change the instruction
     43 ///   returned.)  This allows considerable space savings in other tables.
     44 /// - Six tables (ONEBYTE_SYM, TWOBYTE_SYM, THREEBYTE38_SYM, THREEBYTE3A_SYM,
     45 ///   THREEBYTEA6_SYM, and THREEBYTEA7_SYM contain the hierarchy that the
     46 ///   decoder traverses while decoding an instruction.  At the lowest level of
     47 ///   this hierarchy are instruction UIDs, 16-bit integers that can be used to
     48 ///   uniquely identify the instruction and correspond exactly to its position
     49 ///   in the list of CodeGenInstructions for the target.
     50 /// - One table (INSTRUCTIONS_SYM) contains information about the operands of
     51 ///   each instruction and how to decode them.
     52 ///
     53 /// During table generation, there may be conflicts between instructions that
     54 /// occupy the same space in the decode tables.  These conflicts are resolved as
     55 /// follows in setTableFields() (X86DisassemblerTables.cpp)
     56 ///
     57 /// - If the current context is the native context for one of the instructions
     58 ///   (that is, the attributes specified for it in the LLVM tables specify
     59 ///   precisely the current context), then it has priority.
     60 /// - If the current context isn't native for either of the instructions, then
     61 ///   the higher-priority context wins (that is, the one that is more specific).
     62 ///   That hierarchy is determined by outranks() (X86DisassemblerTables.cpp)
     63 /// - If the current context is native for both instructions, then the table
     64 ///   emitter reports a conflict and dies.
     65 ///
     66 /// *** RESOLUTION FOR "Primary decode conflict"S
     67 ///
     68 /// If two instructions collide, typically the solution is (in order of
     69 /// likelihood):
     70 ///
     71 /// (1) to filter out one of the instructions by editing filter()
     72 ///     (X86RecognizableInstr.cpp).  This is the most common resolution, but
     73 ///     check the Intel manuals first to make sure that (2) and (3) are not the
     74 ///     problem.
     75 /// (2) to fix the tables (X86.td and its subsidiaries) so the opcodes are
     76 ///     accurate.  Sometimes they are not.
     77 /// (3) to fix the tables to reflect the actual context (for example, required
     78 ///     prefixes), and possibly to add a new context by editing
     79 ///     lib/Target/X86/X86DisassemblerDecoderCommon.h.  This is unlikely to be
     80 ///     the cause.
     81 ///
     82 /// DisassemblerEmitter.cpp contains the implementation for the emitter,
     83 ///   which simply pulls out instructions from the CodeGenTarget and pushes them
     84 ///   into X86DisassemblerTables.
     85 /// X86DisassemblerTables.h contains the interface for the instruction tables,
     86 ///   which manage and emit the structures discussed above.
     87 /// X86DisassemblerTables.cpp contains the implementation for the instruction
     88 ///   tables.
     89 /// X86ModRMFilters.h contains filters that can be used to determine which
     90 ///   ModR/M values are valid for a particular instruction.  These are used to
     91 ///   populate ModRMDecisions.
     92 /// X86RecognizableInstr.h contains the interface for a single instruction,
     93 ///   which knows how to translate itself from a CodeGenInstruction and provide
     94 ///   the information necessary for integration into the tables.
     95 /// X86RecognizableInstr.cpp contains the implementation for a single
     96 ///   instruction.
     97 
     98 void DisassemblerEmitter::run(raw_ostream &OS) {
     99   CodeGenTarget Target(Records);
    100 
    101   OS << "/*===- TableGen'erated file "
    102      << "---------------------------------------*- C -*-===*\n"
    103      << " *\n"
    104      << " * " << Target.getName() << " Disassembler\n"
    105      << " *\n"
    106      << " * Automatically generated file, do not edit!\n"
    107      << " *\n"
    108      << " *===---------------------------------------------------------------"
    109      << "-------===*/\n";
    110 
    111   // X86 uses a custom disassembler.
    112   if (Target.getName() == "X86") {
    113     DisassemblerTables Tables;
    114 
    115     const std::vector<const CodeGenInstruction*> &numberedInstructions =
    116       Target.getInstructionsByEnumValue();
    117 
    118     for (unsigned i = 0, e = numberedInstructions.size(); i != e; ++i)
    119       RecognizableInstr::processInstr(Tables, *numberedInstructions[i], i);
    120 
    121     // FIXME: As long as we are using exceptions, might as well drop this to the
    122     // actual conflict site.
    123     if (Tables.hasConflicts())
    124       throw TGError(Target.getTargetRecord()->getLoc(),
    125                     "Primary decode conflict");
    126 
    127     Tables.emit(OS);
    128     return;
    129   }
    130 
    131   // ARM and Thumb have a CHECK() macro to deal with DecodeStatuses.
    132   if (Target.getName() == "ARM" ||
    133       Target.getName() == "Thumb") {
    134     FixedLenDecoderEmitter(Records,
    135                            "ARM",
    136                            "if (!Check(S, ", ")) return MCDisassembler::Fail;",
    137                            "S", "MCDisassembler::Fail",
    138                            "  MCDisassembler::DecodeStatus S = MCDisassembler::Success;\n(void)S;").run(OS);
    139     return;
    140   }
    141 
    142   FixedLenDecoderEmitter(Records, Target.getName()).run(OS);
    143 }
    144