ir-builder

A Haskell library for constructing LLVM IR programmatically. It provides a monadic DSL for building modules, functions, blocks, and instructions, along with a pure renderer that serializes the result to LLVM assembly.

Overview

The library is structured as a pipeline:

IRBuilder (monadic DSL) → IRModule (data types) → renderModule → Text

You describe an IR module using the IRBuilder monad. When you run compileModule, the builder produces an IRModule value. renderModule then converts that value to the textual LLVM IR format that can be passed to llc, clang, or lli.

import LLVM.IR
example :: Text
example = compileModule "example" $ do
 define i32 "add_one" [(i32, "x")] LExternal [] $ do
 beginBlock "entry"
 r <- add i32 (OLocal i32 "x") (OConstant (CInt 32 1))
 ret r

Building

Requires Stack.

stack build # build the library
stack test # run the test suite

Quick example: Hello, World!

{-# LANGUAGE OverloadedStrings #-}
module Examples.HelloWorld (helloWorld) where
mport LLVM.IR
helloWorld :: IRBuilder ()
helloWorld = do
 declare "puts" i32 [ptr]
 emitGlobal (IRString LPrivate ".str" "Hello, World!0円")
 define i32 "main" [] LExternal [] $ do
 beginBlock "entry"
 r1 <- gep (TArray 14 i8) (OGlobal TPtr ".str")
 [OConstant (CInt 32 0), OConstant (CInt 32 0)]
 callVoid NoTail i32 (OGlobal i32 "puts") [r1]
 ret (OConstant (CInt 32 0))

Running compileModule "hello_world" helloWorld produces:

@.str = private constant [14 x i8] c"Hello, World!00円"
declare i32 @puts(ptr)
define i32 @main() {
entry:
 %1 = getelementptr [14 x i8], ptr @.str, i32 0, i32 0
 call i32 @puts(ptr %1)
 ret i32 0
}

Quick example: Iterative factorial with mdo

The IRBuilder monad supports MonadFix. You can use mdo to refer to SSA values before they are emitted, which is necessary for phi nodes in loops:

{-# LANGUAGE OverloadedStrings #-}
{-# LANGUAGE RecursiveDo #-}
module Examples.Factorial (factorialModule) where
import LLVM.IR
factorialModule :: IRBuilder ()
factorialModule = do
 declareVarArg "printf" i32 [ptr]
 emitGlobal (IRString LPrivate ".fmt" "fact(5) = %ld\n0円")
 define i64 "fact" [(i64, "n")] LExternal [] $ mdo
 beginBlock "entry"
 br "loop"
 beginBlock "loop"
 accPhi <- phi i64 [(OConstant (CInt 64 1), "entry"), (newAcc, "body")]
 nPhi <- phi i64 [(OLocal i64 "n", "entry"), (newN, "body")]
 cond <- icmp ICmpSGt i64 nPhi (OConstant (CInt 64 0))
 condbr cond "body" "exit"
 beginBlock "body"
 newAcc <- mul i64 accPhi nPhi
 newN <- sub i64 nPhi (OConstant (CInt 64 1))
 br "loop"
 beginBlock "exit"
 ret accPhi
 define i32 "main" [] LExternal [] $ do
 beginBlock "entry"
 result <- call NoTail i64 (OGlobal i64 "fact") [OConstant (CInt 64 5)]
 callVoid NoTail i32 (OGlobal i32 "printf") [OGlobal ptr ".fmt", result]
 ret (OConstant (CInt 32 0))

Patterns

Conditional with phi merge

When both branches converge at a single exit point, phi selects the appropriate value based on which predecessor block control arrived from:

define i32 "max" [(i32, "a"), (i32, "b")] LExternal [] $ do
 beginBlock "entry"
 cond <- icmp ICmpSGt i32 (OLocal i32 "a") (OLocal i32 "b")
 condbr cond "then_" "else_"
 beginBlock "then_"
 br "merge"
 beginBlock "else_"
 br "merge"
 beginBlock "merge"
 r <- phi i32 [(OLocal i32 "a", "then_"), (OLocal i32 "b", "else_")]
 ret r

define i32 @max(i32 %a, i32 %b) {
entry:
 %1 = icmp sgt i32 %a, %b
 br i1 %1, label %then_, label %else_
then_:
 br label %merge
else_:
 br label %merge
merge:
 %2 = phi i32 [ %a, %then_ ], [ %b, %else_ ]
 ret i32 %2
}

Switch

switch dispatches on an integer value with a mandatory default case and zero or more literal arms. Each arm is a (IRConstant, IRName) pair:

define i32 "classify" [(i32, "n")] LExternal [] $ do
 beginBlock "entry"
 switch (OLocal i32 "n") "default_"
 [ (CInt 32 0, "zero")
 , (CInt 32 1, "one")
 ]
 beginBlock "zero"
 ret (OConstant (CInt 32 10))
 beginBlock "one"
 ret (OConstant (CInt 32 20))
 beginBlock "default_"
 ret (OConstant (CInt 32 (-1)))

Type system

LLVM types are represented by the IRType data type:

LLVM.IRType.Constructors exports shorthand aliases: i1, i8, i16, i32, i64, i128, float, double, ptr, void, and constructor helpers struct, array, vector, fun, named.

Linkage

IRLinkage controls symbol visibility and is the fourth argument to define:

Function attributes

IRAttribute values form the fifth argument to define as a list:

Instruction set

All instruction-emitting functions are in LLVM.IRInstruction.Constructors and share the constraint (MonadIRBuilder m). Terminator-emitting functions are in LLVM.IRTerminator.Constructors.

Arithmetic

add, sub, mul, sdiv, udiv, srem, urem :: IRType -> IROperand -> IROperand -> m IROperand

Bitwise

and, or, xor, shl, lshr, ashr :: IRType -> IROperand -> IROperand -> m IROperand

Floating-point

fadd, fsub, fmul, fdiv :: IRType -> IROperand -> IROperand -> m IROperand
fneg :: IRType -> IROperand -> m IROperand

Comparison

icmp :: IRICmpCond -> IRType -> IROperand -> IROperand -> m IROperand
fcmp :: IRFCmpCond -> IRType -> IROperand -> IROperand -> m IROperand

Memory

alloca :: IRType -> IROperand -> m IROperand -- alloca <type>, <count>
load :: IRType -> IROperand -> m IROperand -- load <type>, ptr <ptr>
store :: IROperand -> IROperand -> m () -- store <val>, ptr <ptr>
gep :: IRType -> IROperand -> [IROperand] -> m IROperand

Type casts

bitcast :: IROperand -> IRType -> m IROperand
sext :: IROperand -> IRType -> m IROperand
zext :: IROperand -> IRType -> m IROperand
trunc :: IROperand -> IRType -> m IROperand
inttoptr :: IROperand -> IRType -> m IROperand
ptrtoint :: IROperand -> IRType -> m IROperand

Calls

call :: IRTailMarker -> IRType -> IROperand -> [IROperand] -> m IROperand
callVoid :: IRTailMarker -> IRType -> IROperand -> [IROperand] -> m ()

IRTailMarker is one of NoTail, Tail, or MustTail.

Miscellaneous

phi :: IRType -> [(IROperand, IRName)] -> m IROperand
select :: IRType -> IROperand -> IROperand -> IROperand -> m IROperand
freeze :: IROperand -> m IROperand

Aggregates

extractValue :: IRType -> IROperand -> [Int] -> m IROperand
insertValue :: IROperand -> IROperand -> [Int] -> m IROperand

Vectors

extractElement :: IRType -> IROperand -> IROperand -> m IROperand
insertElement :: IROperand -> IROperand -> IROperand -> m IROperand
shuffleVector :: IRType -> IROperand -> IROperand -> [Int] -> m IROperand

Use -1 in the mask to indicate an undef slot in shuffleVector.

Atomics

atomicRMW :: IRAtomicOrdering -> IRAtomicOp -> IROperand -> IROperand -> m IROperand
cmpXchg :: IRAtomicOrdering -> IRAtomicOrdering -> IROperand -> IROperand -> IROperand -> m IROperand
fence :: IRAtomicOrdering -> m ()

IRAtomicOrdering: Unordered, Monotonic, Acquire, Release, AcqRel, SeqCst.
IRAtomicOp: ARMWAdd, ARMWSub, ARMWAnd, ARMWOr, ARMWXor, ARMWXchg, ARMWMax, ARMWMin, ARMWUMax, ARMWUMin, ARMWFAdd, ARMWFSub, ARMWFMax, ARMWFMin, ARMWNand.

cmpXchg returns an operand of type { <ty>, i1 }.

Terminators

ret :: IROperand -> m ()
retVoid :: m ()
br :: IRName -> m ()
condbr :: IROperand -> IRName -> IRName -> m ()
switch :: IROperand -> IRName -> [(IRConstant, IRName)] -> m ()
unreachable :: m ()

Globals and declarations

-- Declare an external function
declare :: IRName -> IRType -> [IRType] -> m ()
-- Declare a variadic external function
declareVarArg :: IRName -> IRType -> [IRType] -> m ()
-- Emit a named type declaration (%Name = type { ... })
emitTypeDecl :: IRName -> IRType -> m ()
-- Emit any IRGlobal directly
emitGlobal :: IRGlobal -> m ()

IRGlobal constructors:

Annotations

Instructions can carry an IRAnnotation metadata value. The (<##>) operator attaches a comment to an instruction result:

r <- add i32 a b <##> "compute sum"

This renders as a trailing inline comment on the same line as the instruction:

%1 = add i32 %a, %b ; compute sum

Verification

Two post-hoc checks are available on a compiled IRModule:

verifyModule :: IRModule -> Either String ()
typeCheckModule :: IRModule -> Either String ()

verifyModule checks structural properties: no duplicate block labels, no duplicate SSA names within a function, and all branch targets are defined blocks.

typeCheckModule checks that operand types are consistent with instruction semantics: arithmetic operands match the declared type, phi incoming values match the declared type, load/store pointer operands are ptr, alloca counts are integers, sign-extension/truncation constraints hold, and so on.

Neither function is called automatically by compileModule — you invoke them when needed.

Module reference

Dependencies

• base
• text — Text is used throughout for names and rendered output
• bytestring — string literal globals
• containers — Set in verification
• mtl — StateT, ExceptT, MonadFix