MetalLibraryArchive
is a product of reverse-engineering Apple's metallib
file format.
You can use MetalLibraryArchive
to get the library type, target platform, Metal functions, etc., from a metallib
file.
The extracted information of a Metal function includes:
- Function name.
- Function type - vertex, fragment, kernel, extern, etc.
- Metal Shading Language version of the function.
- Bitcode of the function which can be converted into human-readable LLVM assembly language using llvm-dis.
- Source code of the function if the
metallib
is configured to include source code.
Available at: https://yuao.github.io/MetalLibraryExplorer
An executable target called "Explorer" is included in the package. "Explorer" is a GUI app which can open, unpack and disassemble (with the help of llvm-dis
) metallib
files.
Note llvm-dis
is not included, you can get a copy of the binary at https://github.com/llvm/llvm-project/releases
Use the "Disassembler" menu in the app to locate the llvm-dis
executable file.
You can also use MetalLibraryArchive
as a library:
import MetalLibraryArchive
let archive = try Archive(data: Data(contentsOf: metallibURL))
let libraryType = archive.libraryType
let functions = archive.functions
Byte Range | Type | Content |
---|---|---|
0...3 | FourCharCode | MTLB |
4...5 | UInt16 | Target platform |
6...9 | (UInt16, UInt16) | Version of the metallib file (major, minor) |
10 | UInt8 | Type of the metallib file |
11 | UInt8 | Target OS |
12...15 | (UInt16, UInt16) | Version of the target OS (major, minor) |
16...23 | UInt64 | Size of the metallib file |
24...39 | (UInt64, UInt64) | Offset and size of the function list |
40...55 | (UInt64, UInt64) | Offset and size of the public metadata section |
56...71 | (UInt64, UInt64) | Offset and size of the private metadata section |
72...87 | (UInt64, UInt64) | Offset and size of the bitcode section |
Target Platform | Value |
---|---|
macOS | 0x8001 (0x01,0x80) |
iOS | 0x0001 (0x01,0x00) |
metallib Type | Value |
---|---|
Executable | 0x00 |
Core Image | 0x01 |
Dynamic | 0x02 |
Symbol Companion | 0x03 |
Target OS | Value |
---|---|
Unknown | 0x00 |
macOS | 0x81 |
iOS | 0x82 |
tvOS | 0x83 |
watchOS | 0x84 |
bridgeOS (Probably) | 0x85 |
macCatalyst | 0x86 |
iOS Simulator | 0x87 |
tvOS Simulator | 0x88 |
watchOS Simulator | 0x89 |
Byte Range | Type | Content |
---|---|---|
0...3 | UInt32 | Entry count (the number of functions) |
4... | Tag Groups | Each tag group holds some information about a Metal function |
The number of tag groups equals to the number of functions.
Byte Range | Type | Content |
---|---|---|
0...3 | UInt32 | Size of the tag group |
4... | Tags |
Byte Range | Type | Content |
---|---|---|
0...3 | FourCharCode | Name of the tag |
4...5 | UInt16 | Size of the tag |
6... | Bytes | Content of the tag |
Name | Content Data Type | Content |
---|---|---|
NAME | NULL-terminated C-style string | Name of the function |
MDSZ | UInt64 | Size of the bitcode |
TYPE | UInt8 | Type of the function |
HASH | SHA256 Digest | Hash of the bitcode data (SHA256) |
OFFT | (UInt64, UInt64, UInt64) | Offsets of the information about this function in public metadata section, private metadata section and bitcode section |
SOFF | UInt64 | Offset of the source code archive of the function in embeded source code section |
VERS | (UInt16, UInt16, UInt16, UInt16) | Bitcode and language versions (air.major, air.minor, language.major, language.minor) |
LAYR | UInt8 | Metal type of the render_target_array_index (for layered rendering) |
TESS | UInt8 | Patch type and number of control points per-patch (for post-tessellation vertex function) |
ENDT | End of the tag group |
Function Type | Value | Note |
---|---|---|
Vertex | 0x00 | |
Fragment | 0x01 | |
Kernel | 0x02 | |
Unqualified | 0x03 | Functions in Metal dynamic library |
Visible | 0x04 | Functions with [[visible]] or [[stitchable]] attributes |
Extern | 0x05 | Extern functions complied with -fcikernel option |
Intersection | 0x06 |
Content of the TESS
tag:
// Patch types:
// - triangle: 1
// - quad: 2
let content: UInt8 = controlPointCount << 2 | patchType
Contains information about function constants, tessellation patches, return types, etc.
Tags: CNST
, VATT
, VATY
, RETR
, ARGR
, etc.
Contains pathes to the shader source (DEBI
tag) and .air
(DEPF
tag) files.
Only exists if FunctionListOffset + FunctionListSize + 4 != PublicMetadataOffset
Byte Range | Type | Content |
---|---|---|
FunctionListOffset + FunctionListSize + 4 ... |
Tags | Header extension tags |
Name | Type | Content |
---|---|---|
HDYN | (UInt64, UInt64) | Offset and size of the dynamic header section |
VLST | (UInt64, UInt64) | Offset and size of the exported variable list |
ILST | (UInt64, UInt64) | Offset and size of the imported symbol list |
HSRD/HSRC | (UInt64, UInt64) | Offset and size of the embeded source code section |
UUID | UUID | UUID of the Metal library. |
ENDT | End of the header extension |
Name | Content Data Type | Content |
---|---|---|
NAME | NULL-terminated C-style string | Install name of the library |
DYNL | NULL-terminated C-style string | Linked dynamic library |
Variable list and imported symbol list have structures that are similar to that of the function list.
Only exists if the metallib
build process is configured to include source code.
Byte Range | Type | Content |
---|---|---|
0...1 | UInt16 | Number of items in this section |
2...n | NULL-terminated C-style string | Link options of the metallib file |
n...m | NULL-terminated C-style string | Working directory |
m... | Tag Group | SARC tag |
Note "Working directory" only exists in HSRD
.
Note SARC
tag uses 4-bytes (UInt32
) content size.
Content of the SARC
tag:
Byte Range | Type | Content |
---|---|---|
0...n | NULL-terminated C-style string | ID of the source code archive |
n... | BZh | Bzip2 compressed source code archive |
Value | Type | Value | Type |
---|---|---|---|
0x00 | None | 0x01 | Struct |
0x02 | Array | 0x03 | Float |
0x04 | Float2 | 0x05 | Float3 |
0x06 | Float4 | 0x07 | Float2x2 |
0x08 | Float2x3 | 0x09 | Float2x4 |
0x0A | Float3x2 | 0x0B | Float3x3 |
0x0C | Float3x4 | 0x0D | Float4x2 |
0x0E | Float4x3 | 0x0F | Float4x4 |
0x10 | Half | 0x11 | Half2 |
0x12 | Half3 | 0x13 | Half4 |
0x14 | Half2x2 | 0x15 | Half2x3 |
0x16 | Half2x4 | 0x17 | Half3x2 |
0x18 | Half3x3 | 0x19 | Half3x4 |
0x1A | Half4x2 | 0x1B | Half4x3 |
0x1C | Half4x4 | 0x1D | Int |
0x1E | Int2 | 0x1F | Int3 |
0x20 | Int4 | 0x21 | UInt |
0x22 | UInt2 | 0x23 | UInt3 |
0x24 | UInt4 | 0x25 | Short |
0x26 | Short2 | 0x27 | Short3 |
0x28 | Short4 | 0x29 | UShort |
0x2A | UShort2 | 0x2B | UShort3 |
0x2C | UShort4 | 0x2D | Char |
0x2E | Char2 | 0x2F | Char3 |
0x30 | Char4 | 0x31 | UChar |
0x32 | UChar2 | 0x33 | UChar3 |
0x34 | UChar4 | 0x35 | Bool |
0x36 | Bool2 | 0x37 | Bool3 |
0x38 | Bool4 | 0x3A | Texture |
0x3B | Sampler | 0x3C | Pointer |
0x3E | R8Unorm | 0x3F | R8Snorm |
0x40 | R16Unorm | 0x41 | R16Snorm |
0x42 | RG8Unorm | 0x43 | RG8Snorm |
0x44 | RG16Unorm | 0x45 | RG16Snorm |
0x46 | RGBA8Unorm | 0x47 | RGBA8Unorm_sRGB |
0x48 | RGBA8Snorm | 0x49 | RGBA16Unorm |
0x4A | RGBA16Snorm | 0x4B | RGB10A2Unorm |
0x4C | RG11B10Float | 0x4D | RGB9E5Float |
0x4E | RenderPipeline | 0x4F | ComputePipeline |
0x50 | IndirectCommandBuffer | 0x51 | Long |
0x52 | Long2 | 0x53 | Long3 |
0x54 | Long4 | 0x55 | ULong |
0x56 | ULong2 | 0x57 | ULong3 |
0x58 | ULong4 | 0x59 | Double |
0x5A | Double2 | 0x5B | Double3 |
0x5C | Double4 | 0x5D | Float8 |
0x5E | Float16 | 0x5F | Half8 |
0x60 | Half16 | 0x61 | Int8 |
0x62 | Int16 | 0x63 | UInt8 |
0x64 | UInt16 | 0x65 | Short8 |
0x66 | Short16 | 0x67 | UShort8 |
0x68 | UShort16 | 0x69 | Char8 |
0x6A | Char16 | 0x6B | UChar8 |
0x6C | UChar16 | 0x6D | Long8 |
0x6E | Long16 | 0x6F | ULong8 |
0x70 | ULong16 | 0x71 | Double8 |
0x72 | Double16 | 0x73 | VisibleFunctionTable |
0x74 | IntersectionFunctionTable | 0x75 | PrimitiveAccelerationStructure |
0x76 | InstanceAccelerationStructure | 0x77 | Bool8 |
0x78 | Bool16 |
If you think there's a mistake, please open an issue. You can also choose to open a pull request with the failure test included.
This project would not have started without zhuowei's research that revealed the basic binary layout of a metallib
file, the function list as well as the bitcode section. Thanks, @zhuowei!
I tried to continue the research to get a complete structure of the metallib
file, but found it too hard to move forward based on guesswork alone. So I turned my attention to the Metal.framework
hoping to find out how the framework loads a metallib
file. Fortunately, it's not too hard after dragging Metal.framework/Metal
to Hopper Disassembler.
Metal.framework
uses MTLLibraryDataWithArchive::parseArchiveSync(...)
to load metallib
files. There is a lot of information hidden in the assembly of MTLLibraryDataWithArchive
. For example:
-
The file starts with
0x424c544d
(MTLB); The size of the file is recorded at offset0x10
.int __ZN25MTLLibraryDataWithArchive16parseArchiveSyncEPP7NSErrorb(void * * arg0, bool arg1) { r12 = rdx; r14 = arg1; r13 = arg0; (*(*arg0 + 0xb8))(arg0, 0x0); //LibraryWithFile::setPosition(...) r15 = r13 + 0x78; rbx = (*(*r13 + 0xc0))(r13, r15, 0x58); //LibraryWithFile::readBytes(...) rax = *r13; rax = (*(rax + 0xc8))(r13); //LibraryWithFile::getFileSize(...) // 0x424c544d - MTLB // File size field offset: 0x88 - 0x78 = 0x10 if (((rbx != 0x58) || (*(int32_t *)(r13 + 0x78) != 0x424c544d)) || (*(r13 + 0x88) != rax)) goto loc_6a65b; ... loc_6a65b: if (r14 == 0x0) goto loc_6a6c5; loc_6a660: rdx = @"Invalid library file"; ... }
-
A
Int16
value at offset0x4
is related to the target platform.loc_6a610: // 0x7c - 0x78 = 0x4 rax = *(int16_t *)(r13 + 0x7c) & 0xffff; if ((rax >= 0x0) || (r12 == 0x0)) goto loc_6a6ea; loc_6a627: if (r14 == 0x0) goto loc_6a6c5; loc_6a630: rdx = @"This library format is not supported on this platform (or was built with an old version of the tools)"; goto loc_6a689;
-
There is a "Header Extension Section" that contains information about "Dynamic Header Section", "Imported Symbol List" and "Variable List":
if (MTLLibraryDataWithArchive::parseHeaderExtension(r13, r13 + 0x100, r14) != 0x0) { if (MTLLibraryDataWithArchive::parseDynamicHeaderSection(r13) != 0x0) { if (MTLLibraryDataWithArchive::parseImportedSymbolListSection(r13) != 0x0) { rax = MTLLibraryDataWithArchive::parseVariableListSection(r13); } else { rax = 0x0; } } else { rax = 0x0; } } else { rax = 0x0; }
-
The bitcode is validated using SHA256.
int ____ZN25MTLLibraryDataWithArchive15validateBitCodeEmmPK6NSDataRK12MTLUINT256_t_block_invoke(int arg0) { ... CC_SHA256_Init(&var_B0); CC_SHA256_Update(&var_B0, r14, *(int32_t *)(r15 + 0x38)); CC_SHA256_Final(&var_48, &var_B0); ... }
-
A lot of FourCC codes:
// 0x454e4454 - ENDT loc_6a8bc: if (rax == 0x454e4454) goto loc_6a871; ... // 0x54595045 - TYPE loc_6a984: if (rax == 0x54595045) goto loc_6a9dc; ... // 0x44594e4c - DYNL loc_6ae5b: if (rax != 0x44594e4c) goto loc_6b002; ... // 0x56455253 - VERS loc_6b731: if (rax == 0x56455253) goto loc_6b81c;
After some digging around I was able to get an overview of the metallib
file's structure:
-
The file has a 88 bytes header that contains file version, target platform, library type, section indices, etc.
-
There are 4 sections recorded in the file header:
-
Function list
-
Public metadata
-
Private metadata
-
Bitcode modules
Each section is recorded with an offset and a size. This means sections can be non-contiguous, which allows Apple to introduce new sections in between without breaking the compatibility. And Apple did that exactly for the "header extension" section - it lies between the function list and the public metadata section.
-
-
Most of the sections (except the bitcode section) resemble a "tag" based structure:
-
FourCharCode is used as the tag's name/type.
-
An
UInt16
(in most cases) value of size follows the tag's name.The source archive data tag
SARC
unsurprisingly uses anUInt32
value for its size - a source archive can easily exceed 65KB. -
Tags are grouped:
-
Each group represents a set of properties of an item.
-
Tag group ends with an
ENDT
tag.
-
-
Next, I need to figure out what information each tag/field holds. This can be hard to get from the assembly of the Metal.framework
because:
-
Some fields may be designed purely for tooling or debugging, so
MTLLibraryDataWithArchive
may just ignore them. -
The assembly is platform dependent. For example, the iOS version of
MTLLibraryDataWithArchive
may only check whether themetallib
is built for iOS and cannot tell if the library is built for macOS. -
Some fields are just hard to analyze and follow. Examples:
-
There are 3 offsets in the
OFFT
tag of the function, where are they pointing to? and how are they finally used? -
What are the possible values of the function type? What does each value mean?
-
It seems that the quickest way to get this information is through experiments.
I started by manually compiling metal
files with different shaders, options and SDKs, then inspecting each field I was interested in. My desktop was quickly flooded with metallib
files and HexFiend windows, but I didn't find much useful information. I need something that can automatically build metallib
and presents me only the field that I'm interested in.
I came up with the "Test Driven Guessing":
-
Write a
metallib
parser based on the binary structure overview at hand. -
In the parser, log the value of a field/tag (or some related fields) that is currently unknown.
-
Create tests that produce
metallib
files using different kinds of shaders and compile options that may affect the value of the field, and use the parser to parse the file data. -
Run tests and analyze the log to make hypotheses.
-
Update the parser based on hypotheses.
-
Run tests again to verify.
After a few rounds, I was able to get the function type table, target OS table, and the meaning of 3 offsets in the OFFT
tag.
I also found a few things interesting in this process:
-
Metal does not support watchOS, however it is possible to build a
metallib
targeting watchOS. And Apple does include somemetallib
s in the watchOS SDK. (e.g.Xcode.app/Contents/Developer/Platforms/WatchOS.platform/Library/Developer/CoreSimulator/Profiles/Runtimes/watchOS.simruntime/Contents/Resources/RuntimeRoot/System/Library/Frameworks/CoreImage.framework/ci_filters.metallib
) -
Empty
metallib
s targeting old versions of iOS are mistakenly marked as targeting macOS. -
I cannot build a
metallib
that has the target OS value0x85
. At first I thought it might be reserved for the concealed realityOS, but later found out it is more likely for the bridgeOS.
Apr 10, 2022
Tags like LAYR
, VATY
, CNST
, etc., contain UInt8
values of Metal data types. The corresponding description for each data type value can be retrieved using a private class in Metal.framework - MTLTypeInternal
id value = [[NSClassFromString(@"MTLTypeInternal") alloc] initWithDataType:0x06];
NSLog(@"%@", value.description); // MTLDataTypeFloat4
I created a command line tool to generate the Metal data type table.
cd Utilities
swift run metal-data-type-tools gen-markdown --columns 2 # generate a markdown table
swift run metal-data-type-tools gen-swift # generate a Swift enum for Metal data types.
Mar 31, 2022
The air-lld
(Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/metal/ios/bin/air-lld
) also provides a lot of information about how the metallib
file is built. Some section names and descriptions are updated.
int __ZN4llvm3air20MetalLibObjectWriter5writeEv() {
r14 = rdi;
rax = llvm::air::MetalLibObjectWriter::writeHeader();
if (rax != 0x0) goto loc_1000351b9;
loc_100035135:
rax = llvm::air::MetalLibObjectWriter::writeFunctionList();
if (rax != 0x0) goto loc_1000351b9;
loc_100035141:
rax = llvm::air::MetalLibObjectWriter::writeHeaderExtension();
if (rax != 0x0) goto loc_1000351b9;
loc_10003514d:
rax = llvm::air::MetalLibObjectWriter::writePublicMetadata();
if (rax != 0x0) goto loc_1000351b9;
loc_100035159:
rax = llvm::air::MetalLibObjectWriter::writePrivateMetadata();
if (rax != 0x0) goto loc_1000351b9;
loc_100035165:
rax = llvm::air::MetalLibObjectWriter::writeModuleList();
if (rax != 0x0) goto loc_1000351b9;
loc_100035171:
rax = llvm::air::MetalLibObjectWriter::writeSources();
if (rax != 0x0) goto loc_1000351b9;
loc_10003517d:
rax = llvm::air::MetalLibObjectWriter::writeDynamicHeader();
if (rax != 0x0) goto loc_1000351b9;
loc_100035189:
rax = llvm::air::MetalLibObjectWriter::writeVariableList();
if (rax != 0x0) goto loc_1000351b9;
loc_100035195:
rax = llvm::air::MetalLibObjectWriter::writeImportedSymbolList();
if (rax != 0x0) goto loc_1000351b9;
loc_1000351a1:
rax = llvm::air::MetalLibObjectWriter::computeUUID();
if (rax != 0x0) goto loc_1000351b9;
loc_1000351ad:
rax = llvm::air::MetalLibObjectWriter::backpatchAllLocations();
if (rax == 0x0) goto loc_1000351c2;
loc_1000351b9:
rbx = rax;
goto loc_1000351bb;
loc_1000351bb:
rax = rbx;
return rax;
loc_1000351c2:
rbx = 0x0;
std::__1::system_category();
goto loc_1000351bb;
}