llvm-ir
seeks to provide a Rust-y representation of LLVM IR.
It's based on the idea that an LLVM Instruction
shouldn't be an opaque
datatype, but rather an enum
with variants like Add
, Call
, and
Store
.
Likewise, types like BasicBlock
, Function
, and Module
should be
Rust structs containing as much information as possible.
Unlike other safe LLVM bindings such as inkwell
, llvm-ir
does not rely
on continuous FFI to the LLVM API.
It uses the LLVM API only for its initial parsing step, to pull in all the
data it needs to construct its rich representation of LLVM IR.
Once llvm-ir
creates a Module
data structure by parsing an LLVM file
(using the excellent llvm-sys
low-level LLVM bindings), it drops the LLVM
FFI objects and makes no further FFI calls.
This allows you to work with the resulting LLVM IR in pure safe Rust.
llvm-ir
is intended for consumption of LLVM IR, and not necessarily
production of LLVM IR (yet).
That is, it is aimed at program analysis and related applications which want
to read and analyze LLVM IR.
In the future, perhaps llvm-ir
could be able to output its Module
s back
into LLVM files, or even send them directly to the LLVM library for compiling.
If this interests you, contributions are welcome!
(Or in the meantime, check out inkwell
for a different safe interface for
producing LLVM IR.)
But if you're looking for a nice read-oriented representation of LLVM IR for
working in pure Rust, that's exactly what llvm-ir
can provide today.
This crate is on crates.io, so you can simply
add it as a dependency in your Cargo.toml
, selecting the feature corresponding
to the LLVM version you want:
[dependencies]
llvm-ir = { version = "0.11.1", features = ["llvm-18"] }
Currently, the supported LLVM versions are llvm-9
, llvm-10
, llvm-11
,
llvm-12
, llvm-13
, llvm-14
, llvm-15
, llvm-16
, llvm-17
, and llvm-18
.
Then, the easiest way to get started is to parse some existing LLVM IR into
this crate's data structures.
To do this, you need LLVM bitcode (*.bc
) or text-format IR (*.ll
) files.
If you currently have C/C++ sources (say, source.c
), you can generate
*.bc
files with clang
's -c
and -emit-llvm
flags:
clang -c -emit-llvm source.c -o source.bc
Alternately, to compile Rust sources to LLVM bitcode, you can use rustc
's
--emit=llvm-bc
flag.
In either case, once you have a bitcode file, then you can use llvm-ir
's
Module::from_bc_path
function:
use llvm_ir::Module;
let module = Module::from_bc_path("path/to/my/file.bc")?;
or if you have a text-format IR file, you can use Module::from_ir_path()
.
You may also be interested in the llvm-ir-analysis
crate, which computes
control-flow graphs, dominator trees, etc for llvm-ir
functions.
Documentation for llvm-ir
can be found on docs.rs,
or of course you can generate local documentation with cargo doc --open
.
The documentation includes links to relevant parts of the LLVM documentation
when appropriate.
Note that some data structures differ slightly depending on your choice of
LLVM version. The docs.rs documentation is generated with the llvm-10
feature; for other LLVM versions, you can get appropriate documentation with
cargo doc --features=llvm-<x> --open
where <x>
is the LLVM version you're
using.
Starting with llvm-ir
0.7.0, LLVM versions are selected by a Cargo feature
flag. This means that a single crate version can be used for any supported LLVM
version. Currently, llvm-ir
supports LLVM versions 9 through 18, selected by
feature flags llvm-9
through llvm-18
.
You should select the LLVM version corresponding to the version of the LLVM library you are linking against (i.e., that is available on your system.) Newer LLVMs should be able to read bitcode produced by older LLVMs, so you should be able to use this crate to parse bitcode older than the LLVM version you select via crate feature, even bitcode produced by LLVMs older than LLVM 9. However, this is not extensively tested by us.
llvm-ir
works on stable Rust. As of this writing, it requires Rust 1.65+.
For development or debugging, you may want LLVM text-format (*.ll
) files in
addition to *.bc
files.
For C/C++ sources, you can generate these by passing -S -emit-llvm
to
clang
, instead of -c -emit-llvm
.
E.g.,
clang -S -emit-llvm source.c -o source.ll
For Rust sources, you can use rustc
's --emit=llvm-ir
flag.
Additionally, you may want to pass the -g
flag to clang
, clang++
, or
rustc
when generating bitcode.
This will generate LLVM bitcode with debuginfo, which will ensure that
Instruction
s, Terminator
s, GlobalVariable
s, and Function
s
have valid DebugLoc
s attached. (See the HasDebugLoc
trait.)
Also note that these DebugLoc
s are only available in LLVM 9 and newer;
previous versions of LLVM had a bug in this interface in the C API which
would cause segfaults.
A few features of LLVM IR are not yet represented in llvm-ir
's data
structures.
Most notably, llvm-ir
recovers debug-location metadata (for mapping back to
source locations), but makes no attempt to recover any other debug metadata.
LLVM files containing metadata can still be parsed in with no problems, but
the resulting Module
structures will not contain any of the metadata,
except debug locations.
A few other features are missing from llvm-ir
's data structures because
getters for them are missing from the LLVM C API and the Rust llvm-sys
crate, only being present in the LLVM C++ API.
These include but are not limited to:
- the "fast-math flags" on various floating-point operations
- contents of inline assembly functions
- information about the clauses in the variadic
LandingPad
instruction - information about the operands of a
BlockAddress
constant expression - information about
TargetExtType
types - the "prefix data" associated with a function
- the values of constant integers which are larger than 64 bits (and don't fit in 64 bits) -- see #5
- the "other labels" reachable from a
CallBr
terminator (which was introduced in LLVM 9) - (LLVM 16 and lower -- fixed in LLVM 17 and later) the
nsw
andnuw
flags onAdd
,Sub
,Mul
, andShl
, and likewise theexact
flag onUDiv
,SDiv
,LShr
, andAShr
. The C API has functionality to create new instructions specifying values of these flags, but not to query the values of these flags on existing instructions. - (LLVM 9 and lower -- fixed in LLVM 10 and later) the opcode for the
AtomicRMW
instruction, i.e.,Xchg
,Add
,Max
,Min
, and the like.
More discussion about this is in LLVM bug #42692. Any contributions to filling these gaps in the C API are greatly appreciated!
llvm-ir
took its original inspiration from the llvm-hs-pure
Haskell package.
Most of the data structures in the original release of llvm-ir
were
essentially translations from Haskell to Rust of the data structures in
llvm-hs-pure
(with some tweaks).
llvm-ir
0.7.0 includes several fairly major changes from previous
versions, which are outlined here.
- LLVM versions are now selected via Cargo features. You must select exactly
one of the features
llvm-8
,llvm-9
, orllvm-10
. Previously, we had the0.6.x
branch for LLVM 10, the0.5.x
branch for LLVM 9, and didn't officially support LLVM 8. Now, a single release supports LLVM 8, 9, and 10.- (Note: Versions of this crate beyond 0.7.0 have added support for later LLVM versions as well. For instance, 0.7.3 and later also support LLVM 11; and 0.7.5 and later also support LLVM 12. Crate version 0.11.0 removed support for LLVM 8.)
FunctionAttribute
andParameterAttribute
are now proper enums with descriptive variants such asNoInline
,StackProtect
, etc. Previously, attributes were opaque numeric codes which were difficult to interpret.- Several changes to improve runtime performance and especially memory
consumption, particularly when parsing large LLVM modules. This involves a
number of breaking changes to the public interface:
- Most users of
Type
now own aTypeRef
rather than aType
directly. This includesOperand::LocalOperand
,GlobalVariable
, many variants ofInstruction
, many variants ofConstant
, and some variants ofType
itself, among others. See the documentation onTypeRef
. - Similarly, most users of
Constant
now own aConstantRef
rather than aConstant
directly. See the documentation onConstantRef
. - To get the type of
Typed
objects, the provided.get_type()
method now requires an additional argument; most users will probably prefermodule.type_of()
(ormodule.types.type_of()
). Type::NamedStructType
no longer carries a weak reference to its inner type; instead, you can look up the name usingmodule.types.named_struct_def()
to get the definition for any named struct type in the module.
- Most users of
- The required Rust version increased from 1.36+ to 1.39+.
- (Note: Versions of this crate beyond 0.7.0 have increased this requirement further. For the current required Rust version, see "Compatibility" above.)