Skip to content

Commit

Permalink
feat(perf): Remove redundant inc rc without instructions between (noi…
Browse files Browse the repository at this point in the history 
…r-lang#6183)

# Description

## Problem\*

Part of general effort to reduce Brillig bytecode sizes. 

## Summary\*

We often have patterns like such (example from
`brillig_rc_regression_6123`):
```
brillig fn main f0 {
  b0():
    inc_rc [Field 0, Field 0]
    inc_rc [Field 0, Field 0]
    inc_rc [Field 0, Field 0]
    inc_rc [Field 0, Field 0]
    inc_rc [Field 0, Field 0]
    v4 = allocate
    store [Field 0, Field 0] at v4
    v5 = allocate
    store u32 0 at v5
    inc_rc [Field 0, Field 0]
    v7 = allocate
    store [Field 0, Field 0] at v7
    inc_rc [Field 0, Field 0]
    inc_rc [Field 0, Field 0]
    jmp b1(u32 0)
```
It is usually due to each impl method that mutably borrows self placing
inc RCs on the internal arrays of the struct for which it is
implementing the method. Without any instructions in between these
`inc_rc` instructions, we can safely collapse the inc_rcs into a single
one.

The SSA was `brillig_rc_regression_6123` now looks like such:
```
brillig fn main f0 {
  b0():
    inc_rc [Field 0, Field 0]
    v4 = allocate
    store [Field 0, Field 0] at v4
    v5 = allocate
    store u32 0 at v5
    inc_rc [Field 0, Field 0]
    v7 = allocate
    store [Field 0, Field 0] at v7
    inc_rc [Field 0, Field 0]
    jmp b1(u32 0)
```

This PR does a slight refactor as the `track_inc_rcs_to_remove` was
starting to take a lot of parameters. I made a new `RcTracker` struct.
Its state is simply updated per instruction as previously and then the
instructions it marks for removal are including at the end of performing
DIE on a block.

## Additional Context

## Documentation\*

Check one:
- [X] No documentation needed.
- [ ] Documentation included in this PR.
- [ ] **[For Experimental Features]** Documentation to be submitted in a
separate PR.

# PR Checklist\*

- [X] I have tested the changes locally.
- [X] I have formatted the changes with [Prettier](https://prettier.io/)
and/or `cargo fmt` on default settings.
  • Loading branch information
@vezenovm
vezenovm authored Sep 30, 2024
1 parent 1ac980b commit be9dcfe
Showing 1 changed file with 119 additions and 91 deletions.
210 changes: 119 additions & 91 deletions compiler/noirc_evaluator/src/ssa/opt/die.rs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -106,19 +106,7 @@ impl Context {

let instructions_len = block.instructions().len();

// We can track IncrementRc instructions per block to determine whether they are useless.
// IncrementRc and DecrementRc instructions are normally side effectual instructions, but we remove
// them if their value is not used anywhere in the function. However, even when their value is used, their existence
// is pointless logic if there is no array set between the increment and the decrement of the reference counter.
// We track per block whether an IncrementRc instruction has a paired DecrementRc instruction
// with the same value but no array set in between.
// If we see an inc/dec RC pair within a block we can safely remove both instructions.
let mut rcs_with_possible_pairs: HashMap<Type, Vec<RcInstruction>> = HashMap::default();
let mut rc_pairs_to_remove = HashSet::default();
// We also separately track all IncrementRc instructions and all arrays which have been mutably borrowed.
// If an array has not been mutably borrowed we can then safely remove all IncrementRc instructions on that array.
let mut inc_rcs: HashMap<ValueId, HashSet<InstructionId>> = HashMap::default();
let mut borrowed_arrays: HashSet<ValueId> = HashSet::default();
let mut rc_tracker = RcTracker::default();

// Indexes of instructions that might be out of bounds.
// We'll remove those, but before that we'll insert bounds checks for them.
Expand Down Expand Up @@ -147,32 +135,11 @@ impl Context {
}
}

self.track_inc_rcs_to_remove(
*instruction_id,
function,
&mut rcs_with_possible_pairs,
&mut rc_pairs_to_remove,
&mut inc_rcs,
&mut borrowed_arrays,
);
rc_tracker.track_inc_rcs_to_remove(*instruction_id, function);
}

let non_mutated_arrays =
inc_rcs
.keys()
.filter_map(|value| {
if !borrowed_arrays.contains(value) {
Some(&inc_rcs[value])
} else {
None
}
})
.flatten()
.copied()
.collect::<Vec<_>>();

self.instructions_to_remove.extend(non_mutated_arrays);
self.instructions_to_remove.extend(rc_pairs_to_remove);
self.instructions_to_remove.extend(rc_tracker.get_non_mutated_arrays());
self.instructions_to_remove.extend(rc_tracker.rc_pairs_to_remove);

// If there are some instructions that might trigger an out of bounds error,
// first add constrain checks. Then run the DIE pass again, which will remove those
Expand All @@ -196,59 +163,6 @@ impl Context {
false
}

fn track_inc_rcs_to_remove(
&self,
instruction_id: InstructionId,
function: &Function,
rcs_with_possible_pairs: &mut HashMap<Type, Vec<RcInstruction>>,
inc_rcs_to_remove: &mut HashSet<InstructionId>,
inc_rcs: &mut HashMap<ValueId, HashSet<InstructionId>>,
borrowed_arrays: &mut HashSet<ValueId>,
) {
let instruction = &function.dfg[instruction_id];
// DIE loops over a block in reverse order, so we insert an RC instruction for possible removal
// when we see a DecrementRc and check whether it was possibly mutated when we see an IncrementRc.
match instruction {
Instruction::IncrementRc { value } => {
if let Some(inc_rc) = pop_rc_for(*value, function, rcs_with_possible_pairs) {
if !inc_rc.possibly_mutated {
inc_rcs_to_remove.insert(inc_rc.id);
inc_rcs_to_remove.insert(instruction_id);
}
}

inc_rcs.entry(*value).or_default().insert(instruction_id);
}
Instruction::DecrementRc { value } => {
let typ = function.dfg.type_of_value(*value);

// We assume arrays aren't mutated until we find an array_set
let dec_rc =
RcInstruction { id: instruction_id, array: *value, possibly_mutated: false };
rcs_with_possible_pairs.entry(typ).or_default().push(dec_rc);
}
Instruction::ArraySet { array, .. } => {
let typ = function.dfg.type_of_value(*array);
if let Some(dec_rcs) = rcs_with_possible_pairs.get_mut(&typ) {
for dec_rc in dec_rcs {
dec_rc.possibly_mutated = true;
}
}

borrowed_arrays.insert(*array);
}
Instruction::Store { value, .. } => {
// We are very conservative and say that any store of an array value means it has the potential
// to be mutated. This is done due to the tracking of mutable borrows still being per block.
let typ = function.dfg.type_of_value(*value);
if matches!(&typ, Type::Array(..) | Type::Slice(..)) {
borrowed_arrays.insert(*value);
}
}
_ => {}
}
}

/// Returns true if an instruction can be removed.
///
/// An instruction can be removed as long as it has no side-effects, and none of its result
Expand Down Expand Up @@ -601,6 +515,103 @@ fn apply_side_effects(
(lhs, rhs)
}

#[derive(Default)]
struct RcTracker {
// We can track IncrementRc instructions per block to determine whether they are useless.
// IncrementRc and DecrementRc instructions are normally side effectual instructions, but we remove
// them if their value is not used anywhere in the function. However, even when their value is used, their existence
// is pointless logic if there is no array set between the increment and the decrement of the reference counter.
// We track per block whether an IncrementRc instruction has a paired DecrementRc instruction
// with the same value but no array set in between.
// If we see an inc/dec RC pair within a block we can safely remove both instructions.
rcs_with_possible_pairs: HashMap<Type, Vec<RcInstruction>>,
rc_pairs_to_remove: HashSet<InstructionId>,
// We also separately track all IncrementRc instructions and all arrays which have been mutably borrowed.
// If an array has not been mutably borrowed we can then safely remove all IncrementRc instructions on that array.
inc_rcs: HashMap<ValueId, HashSet<InstructionId>>,
mut_borrowed_arrays: HashSet<ValueId>,
// The SSA often creates patterns where after simplifications we end up with repeat
// IncrementRc instructions on the same value. We track whether the previous instruction was an IncrementRc,
// and if the current instruction is also an IncrementRc on the same value we remove the current instruction.
// `None` if the previous instruction was anything other than an IncrementRc
previous_inc_rc: Option<ValueId>,
}

impl RcTracker {
fn track_inc_rcs_to_remove(&mut self, instruction_id: InstructionId, function: &Function) {
let instruction = &function.dfg[instruction_id];

if let Instruction::IncrementRc { value } = instruction {
if let Some(previous_value) = self.previous_inc_rc {
if previous_value == *value {
self.rc_pairs_to_remove.insert(instruction_id);
}
}
self.previous_inc_rc = Some(*value);
} else {
self.previous_inc_rc = None;
}

// DIE loops over a block in reverse order, so we insert an RC instruction for possible removal
// when we see a DecrementRc and check whether it was possibly mutated when we see an IncrementRc.
match instruction {
Instruction::IncrementRc { value } => {
if let Some(inc_rc) =
pop_rc_for(*value, function, &mut self.rcs_with_possible_pairs)
{
if !inc_rc.possibly_mutated {
self.rc_pairs_to_remove.insert(inc_rc.id);
self.rc_pairs_to_remove.insert(instruction_id);
}
}

self.inc_rcs.entry(*value).or_default().insert(instruction_id);
}
Instruction::DecrementRc { value } => {
let typ = function.dfg.type_of_value(*value);

// We assume arrays aren't mutated until we find an array_set
let dec_rc =
RcInstruction { id: instruction_id, array: *value, possibly_mutated: false };
self.rcs_with_possible_pairs.entry(typ).or_default().push(dec_rc);
}
Instruction::ArraySet { array, .. } => {
let typ = function.dfg.type_of_value(*array);
if let Some(dec_rcs) = self.rcs_with_possible_pairs.get_mut(&typ) {
for dec_rc in dec_rcs {
dec_rc.possibly_mutated = true;
}
}

self.mut_borrowed_arrays.insert(*array);
}
Instruction::Store { value, .. } => {
// We are very conservative and say that any store of an array value means it has the potential
// to be mutated. This is done due to the tracking of mutable borrows still being per block.
let typ = function.dfg.type_of_value(*value);
if matches!(&typ, Type::Array(..) | Type::Slice(..)) {
self.mut_borrowed_arrays.insert(*value);
}
}
_ => {}
}
}

fn get_non_mutated_arrays(&self) -> HashSet<InstructionId> {
self.inc_rcs
.keys()
.filter_map(|value| {
if !self.mut_borrowed_arrays.contains(value) {
Some(&self.inc_rcs[value])
} else {
None
}
})
.flatten()
.copied()
.collect()
}
}
#[cfg(test)]
mod test {
use std::sync::Arc;
Expand Down Expand Up @@ -902,10 +913,27 @@ mod test {
assert_eq!(main.dfg[main.entry_block()].instructions().len(), 6);

// We expect the output to be unchanged
// Expected output:
//
// acir(inline) fn main f0 {
// b0(v0: [u32; 2]):
// inc_rc v0
// v3 = array_set v0, index u32 0, value u32 1
// inc_rc v0
// v4 = array_get v3, index u32 1
// return v4
// }
let ssa = ssa.dead_instruction_elimination();
let main = ssa.main();

assert_eq!(main.dfg[main.entry_block()].instructions().len(), 6);
let instructions = main.dfg[main.entry_block()].instructions();
// We expect only the repeated inc_rc instructions to be collapsed into a single inc_rc.
assert_eq!(instructions.len(), 4);

assert!(matches!(&main.dfg[instructions[0]], Instruction::IncrementRc { .. }));
assert!(matches!(&main.dfg[instructions[1]], Instruction::ArraySet { .. }));
assert!(matches!(&main.dfg[instructions[2]], Instruction::IncrementRc { .. }));
assert!(matches!(&main.dfg[instructions[3]], Instruction::ArrayGet { .. }));
}

#[test]
Expand Down

0 comments on commit be9dcfe

Please sign in to comment.