Module std::intrinsics::mir

source ·
🔬This is a nightly-only experimental API. (custom_mir)
Expand description

Rustc internal tooling for hand-writing MIR.

If for some reasons you are not writing rustc tests and have found yourself considering using this feature, turn back. This is exceptionally unstable. There is no attempt at all to make anything work besides those things which the rustc test suite happened to need. If you make a typo you’ll probably ICE. Really, this is not the solution to your problems. Consider instead supporting the stable MIR project group.

The documentation for this module describes how to use this feature. If you are interested in hacking on the implementation, most of that documentation lives at rustc_mir_build/src/build/custom/mod.rs.

Typical usage will look like this:

#![feature(core_intrinsics, custom_mir)]
#![allow(internal_features)]

use core::intrinsics::mir::*;

#[custom_mir(dialect = "built")]
pub fn simple(x: i32) -> i32 {
    mir!(
        let temp2: i32;

        {
            let temp1 = x;
            Goto(my_second_block)
        }

        my_second_block = {
            temp2 = Move(temp1);
            RET = temp2;
            Return()
        }
    )
}
Run

The custom_mir attribute tells the compiler to treat the function as being custom MIR. This attribute only works on functions - there is no way to insert custom MIR into the middle of another function. The dialect and phase parameters indicate which version of MIR you are inserting here. Generally you’ll want to use #![custom_mir(dialect = "built")] if you want your MIR to be modified by the full MIR pipeline, or #![custom_mir(dialect = "runtime", phase = "optimized")] if you don’t.

The input to the mir! macro is:

  • An optional return type annotation in the form of type RET = ...;. This may be required if the compiler cannot infer the type of RET.
  • A possibly empty list of local declarations. Locals can also be declared inline on assignments via let. Type inference generally works. Shadowing does not.
  • A list of basic blocks. The first of these is the start block and is where execution begins. All blocks other than the start block need to be given a name, so that they can be referred to later.
    • Each block is a list of semicolon terminated statements, followed by a terminator. The syntax for the various statements and terminators is designed to be as similar as possible to the syntax for analogous concepts in native Rust. See below for a list.

§Examples

#![feature(core_intrinsics, custom_mir)]
#![allow(internal_features)]

use core::intrinsics::mir::*;

#[custom_mir(dialect = "built")]
pub fn choose_load(a: &i32, b: &i32, c: bool) -> i32 {
    mir!(
        {
            match c {
                true => t,
                _ => f,
            }
        }

        t = {
            let temp = a;
            Goto(load_and_exit)
        }

        f = {
            temp = b;
            Goto(load_and_exit)
        }

        load_and_exit = {
            RET = *temp;
            Return()
        }
    )
}

#[custom_mir(dialect = "built")]
fn unwrap_unchecked<T>(opt: Option<T>) -> T {
    mir!({
        RET = Move(Field(Variant(opt, 1), 0));
        Return()
    })
}

#[custom_mir(dialect = "runtime", phase = "optimized")]
fn push_and_pop<T>(v: &mut Vec<T>, value: T) {
    mir!(
        let _unused;
        let popped;

        {
            Call(_unused = Vec::push(v, value), pop)
        }

        pop = {
            Call(popped = Vec::pop(v), drop)
        }

        drop = {
            Drop(popped, ret)
        }

        ret = {
            Return()
        }
    )
}

#[custom_mir(dialect = "runtime", phase = "optimized")]
fn annotated_return_type() -> (i32, bool) {
    mir!(
        type RET = (i32, bool);
        {
            RET.0 = 1;
            RET.1 = true;
            Return()
        }
    )
}
Run

We can also set off compilation failures that happen in sufficiently late stages of the compiler:

ⓘ
#![feature(core_intrinsics, custom_mir)]

extern crate core;
use core::intrinsics::mir::*;

#[custom_mir(dialect = "built")]
fn borrow_error(should_init: bool) -> i32 {
    mir!(
        let temp: i32;

        {
            match should_init {
                true => init,
                _ => use_temp,
            }
        }

        init = {
            temp = 0;
            Goto(use_temp)
        }

        use_temp = {
            RET = temp;
            Return()
        }
    )
}
Run
error[E0381]: used binding is possibly-uninitialized
  --> test.rs:24:13
   |
8  | /     mir!(
9  | |         let temp: i32;
10 | |
11 | |         {
...  |
19 | |             temp = 0;
   | |             -------- binding initialized here in some conditions
...  |
24 | |             RET = temp;
   | |             ^^^^^^^^^^ value used here but it is possibly-uninitialized
25 | |             Return()
26 | |         }
27 | |     )
   | |_____- binding declared here but left uninitialized

error: aborting due to previous error

For more information about this error, try `rustc --explain E0381`.

§Syntax

The lists below are an exhaustive description of how various MIR constructs can be created. Anything missing from the list should be assumed to not be supported, PRs welcome.

§Locals
  • The _0 return local can always be accessed via RET.
  • Arguments can be accessed via their regular name.
  • All other locals need to be declared with let somewhere and then can be accessed by name.
§Places
  • Locals implicit convert to places.
  • Field accesses, derefs, and indexing work normally.
  • Fields in variants can be accessed via the Variant and Field associated functions, see their documentation for details.
§Operands
  • Places implicitly convert to Copy operands.
  • Move operands can be created via Move.
  • Const blocks, literals, named constants, and const params all just work.
  • Static and StaticMut can be used to create &T and *mut Ts to statics. These are constants in MIR and the only way to access statics.
§Statements
§Rvalues
  • Operands implicitly convert to Use rvalues.
  • &, &mut, addr_of!, and addr_of_mut! all work to create their associated rvalue.
  • Discriminant, Len, and CopyForDeref have associated functions.
  • Unary and binary operations use their normal Rust syntax - a * b, !c, etc.
  • The binary operation Offset can be created via Offset.
  • Checked binary operations are represented by wrapping the associated binop in Checked.
  • Array repetition syntax ([foo; 10]) creates the associated rvalue.
§Terminators

Custom MIR does not currently support cleanup blocks or non-trivial unwind paths. As such, there are no resume and abort terminators, and terminators that might unwind do not have any way to indicate the unwind block.

  • Goto, Return, Unreachable and Drop have associated functions.
  • match some_int_operand becomes a SwitchInt. Each arm should be literal => basic_block
    • The exception is the last arm, which must be _ => basic_block and corresponds to the otherwise branch.
  • Call has an associated function as well. The third argument of this function is a normal function call expression, for example my_other_function(a, 5).

Macros

  • mirExperimental
    Macro for generating custom MIR.
  • placeExperimental
    Helper macro that allows you to treat a value expression like a place expression.

Structs

  • BasicBlockExperimental
    Type representing basic blocks.

Functions