For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo function

inlined_assembly

def inlined_assembly[asm: StringSpan[ImmStaticOrigin], result_type: TrivialRegisterPassable, *types: AnyType, *, constraints: StringSpan[ImmStaticOrigin], has_side_effect: Bool = True](*args: *types.values) -> result_type

Generates inline assembly code with the given constraints and arguments.

This function allows embedding raw assembly instructions directly into Mojo code, providing fine-grained control over hardware operations. It uses LLVM-style inline assembly syntax and constraint strings.

The assembly string uses $0, $1, etc. to reference operands. Output operands (including the return value) are numbered first, followed by input operands.

Example:

from std.sys import inlined_assembly

# Convert bfloat16 to float32 on NVIDIA GPU using PTX assembly.
# "$0" is the output (float32), "$1" is the input (int16 bitcast of bf16).
var my_bf16_as_int16 = Int16(0x3F80)  # Example bf16 bit pattern
var result = inlined_assembly[
    "cvt.f32.bf16 $0, $1;",
    Float32,
    constraints="=f,h",
    has_side_effect=False,
](my_bf16_as_int16)

# Execute a no-op sleep instruction on AMD GPU (no return value).
inlined_assembly[
    "s_sleep 0",
    NoneType,
    constraints="",
    has_side_effect=True,
]()

Parameters:

asm (StringSpan[ImmStaticOrigin]): The assembly instruction string. Use $0, $1, etc. to reference operands, where $0 is the output (if any) and subsequent numbers are inputs in order.
result_type (TrivialRegisterPassable): The return type of the assembly operation. Use NoneType for assembly that produces no result.
*types (AnyType): The types of the input arguments.
constraints (StringSpan[ImmStaticOrigin]): LLVM-style constraint string specifying register allocation and operand placement. The output constraint comes first (prefixed with =), followed by input constraints separated by commas. Available constraints are target-specific; refer to LLVM's inline assembly documentation and your target's backend for valid options.
has_side_effect (Bool): If True (default), the assembly is treated as having side effects and won't be optimized away or reordered. Set to False for pure computations to enable compiler optimizations.

Args:

*args (*types.values): The input arguments to pass to the assembly instruction.

Returns:

result_type: The result of the assembly operation, or None if result_type is NoneType.