Back to Dotnet

Generating `amd64InstrDecode.h`

src/coreclr/debug/ee/amd64/gen_amd64InstrDecode/README.md

11.0.10012.5 KB
Original Source

Generating amd64InstrDecode.h

TL;DR - What do I do? Process

The following process was executed on an amd64 Linux host in this directory.

bash
# Create the program createOpcodes
gcc createOpcodes.cpp -o createOpcodes

# Execute the program to create opcodes.cpp
./createOpcodes > opcodes.cpp

# Compile opcodes.cpp to opcodes
gcc -g opcodes.cpp -o opcodes

# Disassemble opcodes
gdb opcodes -batch -ex "set disassembly-flavor intel" -ex "disass /r opcodes" > opcodes.intel

# Parse disassembly and generate code
# Build as a separate step so it will display build errors, if any.
../../../../../../dotnet.sh build
cat opcodes.intel | ../../../../../../dotnet.sh run > new_amd64InstrDecode.h

After checking it, copy the generated new_amd64InstrDecode.h to ../amd64InstrDecode.h.

This process can be run using the createTables.sh script in this directory.

Technical design

The primary purpose of amd64InstrDecode.h is to provide a reliable and accurate mechanism to implement the amd64 NativeWalker::DecodeInstructionForPatchSkip(..) function.

This function needs to be able to decode an arbitrary amd64 instruction. The decoder currently must be able to identify:

  • Whether the instruction includes an instruction pointer relative memory access (RIP relative addressing)
  • The location of the memory displacement within the instruction
  • The instruction length in bytes
  • The size of the memory operation in bytes

To get this right is complicated, because the amd64 instruction set is complicated.

A high level view of the amd64 instruction set can be seen by looking at:

AMD64 Architecture Programmer's Manual Volume 3: General-Purpose and System Instructions Section 1.1 Instruction Encoding Overview Figure 1-1. Instruction Encoding Syntax

also:

Intel(R) 64 and IA-32 Architectures Software Developer's Manual Volume 2: Instruction Set Reference, A-Z Chapter 2 Instruction Format

Also useful in the manuals:

AMD: Appendix A: Opcode and Operand Encodings Intel: Volume 2, Appendix A: Opcode Map

The general behavior of each instruction can be modified by many of the bytes in the 1-15 byte instruction (15 is the maximum byte length of an instruction).

This set of files generates a metadata table by extracting the data from sample instruction disassembly.

The process entails:

  • Generating a necessary set of instructions
  • Generating parsable disassembly for the instructions
  • Parsing the disassembly
  • Generating the tables

Generating a necessary set of instructions

What set of possible instruction encodings are needed to extract the information needed in the tables?

The necessary set

  • All instruction forms which use RIP relative memory accesses.
  • All combinations of modifier bits which affect the instruction form
    • presence and/or size of the memory access
    • size or presence of immediates
    • vector size bits

So with modrm.mod = 0, modrm.rm = 0x5 (RIP relative memory access) we need all combinations of:

  • opcodemap
  • opcode
  • modrm.reg
  • pp, W, L, L'L
  • Some combinations of vvvv
  • Optional prefixes: repe, repne, opSize

Padding

We will iterate through all the necessary set. Many of these combinations will lead to invalid/undefined encodings. This will cause the disassembler to give up and mark the disassemble as bad.

The disassembly will then resume trying to disassemble at the next boundary.

To make sure the disassembler attempts to disassemble every instruction, we need to make sure the preceding instruction is always valid and terminates at our desired instruction boundary.

Through examination of the Primary opcode map, it is observed that 0x50-0x5f are all 1 byte instructions. These become convenient padding.

After each necessary instruction we insert enough padding bytes to fill the maximum instruction length and leave at least one additional one byte instruction.

Fixed suffix

Using a fixed suffix makes disassembly parsing simpler.

After the modrm byte, the generated instructions always include a postamble,

C++
const char* postamble = "0x50, 0x51, 0x52, 0x53, 0x54, 0x55, 0x56, 0x57, 0x58, 0x59,\n";

This meets the padding consistency needs.

Ordering

As a convenience to the parser the encoded instructions are logically ordered. The ordering is generally as follows, but can vary slightly depending on the needs of the particular opcode map:

  • map
  • opcode
  • pp & some prefixes
  • modrm.reg
  • W, L, L'L, vvvv

This is to keep related instruction grouped together.

Encoding the instructions

The simplest way to get these instructions into an object file for disassembly is to place them into a C++ BYTE array.

The file createOpcodes.cpp is the source for a program which will generate opcodes.cpp

bash
# Create the program createOpcodes
gcc createOpcodes.cpp -o createOpcodes

# Execute the program to create opcodes.cpp
./createOpcodes > opcodes.cpp

opcodes.cpp will now be a C++ source file with uint8_t opcodes[] initialized with our set of necessary instructions and padding.

We need to compile this to an executable to prepare for disassembly.

bash
# Compile opcodes.cpp to opcodes
gcc -g opcodes.cpp -o opcodes

Generating parsable disassembly

In investigating the various disassembly formats, the intel disassembly format is superior to the att format. This is because the intel format clearly marks the instruction relative accesses and their sizes. For instance:

  • "BYTE PTR [rip+0x53525150]"
  • "WORD PTR [rip+0x53525150]"
  • "DWORD PTR [rip+0x53525150]"
  • "QWORD PTR [rip+0x53525150]"
  • "OWORD PTR [rip+0x53525150]"
  • "XMMWORD PTR [rip+0x53525150]"
  • "YMMWORD PTR [rip+0x53525150]"
  • "ZMMWORD PTR [rip+0x53525150]"
  • "FWORD PTR [rip+0x53525150]"
  • "TBYTE PTR [rip+0x53525150]"

Also it is important to have all the raw bytes in the disassembly. This allows accurately determining the instruction length.

It also helps identifying which instructions are from our needed set.

I happened to have used gdb as a disassembler.

bash
# Disassemble opcodes
gdb opcodes -batch -ex "set disassembly-flavor intel" -ex "disass /r opcodes" > opcodes.intel

Alternative disassemblers

It seems objdump could provide similar results. This is untested. The parser may need to be modified for subtle differences.

bash
objdump -D -M intel -b --insn-width=15 -j .data opcodes

The lldb parser aborts parsing when it observes bad instruction. It might be usable with additional python scripts.

Windows disassembler may also work. It has not been tried.

Parsing the disassembly

bash
# Parse disassembly and generate code
cat opcodes.intel | dotnet run > ../amd64InstrDecode.h

Finding relevant disassembly lines

We are not interested in all lines in the disassembly. The disassembler stray comments, recovery, and our padding introduce lines we need to ignore.

We filter out and ignore non-disassembly lines using a Regex for a disassembly line.

We expect the generated instruction samples to be in a group. The first instruction in the group is the only one we are interested in.

The group is terminated by a pair of instructions. The first terminal instruction must have 0x58 as the last byte in its encoding. The final terminal instruction must be a 0x59\tpop.

We continue parsing the first line of each group.

Ignoring bad encodings

Many encodings are not valid. For gdb, these instructions are marked (bad). We filter and ignore these.

Parsing the disassembly for each instruction sample

For each sample, we need to calculate the important properties:

  • mnemonic
  • raw encoding
  • disassembly (for debug)
  • Encoding
    • map
    • opcode position
  • Encoding Flags
    • pp, W, L, L'L, prefix and encoding flags
  • SuffixFlags
    • presence of RIP relative accesses
      • size of operation
      • position in the list of operands
      • number of immediate bytes
Supplementing the disassembler

In a few cases it was observed the disassembly of some memory operations did not include a size. These were manually researched. For the ones with reasonable sizes, these were added to a table to manually override these unknown sizes.

opCodeExt

To facilitate identifying sets of instructions, the tool creates an opCodeExt.

For the Primary map this is simply the encoded opcode from the instruction shifted left by 4 bits.

For the Secondary, F38, and F39 maps this is the encoded opcode from the instruction shifted left by 4 bits or'ed with a synthetic pp. The synthetic pp is constructed to match the rules of Table 1-22. VEX/XOP.pp Encoding from the AMD64 Architecture Programmer’s Manual Volume 3: General-Purpose and System Instructions. For the case where the opSize 0x66 prefix is present with a rep* prefix, the rep* prefix is used to encode pp.

For the VEX* maps this is the encoded opcode from the instruction shifted left by 4 bits or'ed with pp.

Identifying sets of instructions

For most instructions, the opCodeExt will uniquely identify the instruction.

For many instructions, modrm.reg is used to help uniquely identify the instruction. These instruction typically change mnemonic and behavior as modrm.reg changes. These become problematic, when the form of these instructions vary.

For a few other instructions the L, W, vvvv values may change the instruction behavior. Usually these do not change mnemonic.

The set of instructions is therefore usually grouped by the opcode map and opCodeExt generated above. For these a change in opCodeExt or map will start a new group.

For select problematic groups of modrm.reg sensitive instructions, a change in modrm.reg will start a new group.

For each set of instructions

  • Calculate the intersection and union of the SuffixFlags for the set.
  • The flags in the intersection are common to all instructions in the set.
  • The ones in the union, but not in the intersection vary within the set based on the encoding flags. These are the sometimesFlags
  • Determine the rules for the sometimesFlags. For each combination of sometimesFlags, check each rule by calling TestHypothesis. This determines if the rule corresponds to the set of observations.

Encode the rule as a string. The rule might encode that the W bit or L bit causes a different memory/immediate behavior for the particular <map, opCodeExt> entry.

Add the rule to the set of all observed rules. Add the set's rule with comment to a dictionary.

Generating the C++ code

At this point generating the code is rather simple.

Iterate through the set of rules to create an enumeration of InstrForm.

For each map iterate through the dictionary, filling missing instructions with an appropriate pattern for undefined instructions.

Encoding

The design uses a simple fully populated direct look up table to provide a nice simple means of looking up. This direct map approach is expected to consume ~10K bytes.

Other approaches like a sparse list may reduce total memory usage. The added complexity did not seem worth it.

Benefits

This approach is intended to reduce the human error introduced by manually parsing and encoding the various instruction forms from their respective descriptions.

Limitations

The approach of using a single object file as the source of disassembly samples, is restricted to a max compilation/link unit size. Early drafts were generating more instructions, and couldn't be compiled.

However, there is no restriction that all the samples must come from single executable. These could easily be separated by opcode map.

Risks

New instruction sets

This design is for existing instruction sets. New instruction sets will require more work.

Further this methodology uses the disassembler to generate the tables. Until a reasonably featured disassembler is created, the new instruction set can not be supported by this methodology.

The previous methodology of manually encoding these new instruction set would still be possible.

Disassembler errors

This design presumes the disassembler is correct. The specific version of the disassembler may have disassembly bugs. Using newer disassemblers would mitigate this to some extent.

Bugs

  • Inadequate samples. Are there other bits which modify instruction behavior which we missed?
  • Parser/Table generator implementation bugs. Does the parser do what it was intended to do?

Reasons to regenerate the file

Disassembler error discovered

Add a patch to the parser to workaround the bug and regenerate the table

Newer disassembler available

Regenerate and compare.

New debugger feature requires more metadata

Add new feature code, regenerate.