mz_expr/relation/canonicalize.rs
1// Copyright Materialize, Inc. and contributors. All rights reserved.
2//
3// Use of this software is governed by the Business Source License
4// included in the LICENSE file.
5//
6// As of the Change Date specified in that file, in accordance with
7// the Business Source License, use of this software will be governed
8// by the Apache License, Version 2.0.
9
10//! Utility functions to transform parts of a single `MirRelationExpr`
11//! into canonical form.
12
13use std::cmp::Ordering;
14use std::collections::{BTreeMap, BTreeSet};
15use std::rc::Rc;
16
17use itertools::Itertools;
18use mz_ore::soft_assert_or_log;
19use mz_repr::{SqlColumnType, SqlScalarType};
20
21use crate::visit::Visit;
22use crate::{MirScalarExpr, UnaryFunc, VariadicFunc, func};
23
24/// Canonicalize equivalence classes of a join and expressions contained in them.
25///
26/// `input_types` can be the [SqlColumnType]s of the join or the [SqlColumnType]s of
27/// the individual inputs of the join in order.
28///
29/// This function:
30/// * simplifies expressions to involve the least number of non-literal nodes.
31/// This ensures that we only replace expressions by "even simpler"
32/// expressions and that repeated substitutions reduce the complexity of
33/// expressions and a fixed point is certain to be reached. Without this
34/// rule, we might repeatedly replace a simple expression with an equivalent
35/// complex expression containing that (or another replaceable) simple
36/// expression, and repeat indefinitely.
37/// * reduces all expressions contained in `equivalences`.
38/// * Does everything that [canonicalize_equivalence_classes] does.
39pub fn canonicalize_equivalences<'a, I>(
40 equivalences: &mut Vec<Vec<MirScalarExpr>>,
41 input_column_types: I,
42) where
43 I: Iterator<Item = &'a Vec<SqlColumnType>>,
44{
45 let column_types = input_column_types
46 .flat_map(|f| f.clone())
47 .collect::<Vec<_>>();
48 // Calculate the number of non-leaves for each expression.
49 let mut to_reduce = equivalences
50 .drain(..)
51 .filter_map(|mut cls| {
52 let mut result = cls
53 .drain(..)
54 .map(|expr| (rank_complexity(&expr), expr))
55 .collect::<Vec<_>>();
56 result.sort();
57 result.dedup();
58 if result.len() > 1 { Some(result) } else { None }
59 })
60 .collect::<Vec<_>>();
61
62 let mut expressions_rewritten = true;
63 while expressions_rewritten {
64 expressions_rewritten = false;
65 for i in 0..to_reduce.len() {
66 // `to_reduce` will be borrowed as immutable, so in order to modify
67 // elements of `to_reduce[i]`, we are going to pop them out of
68 // `to_reduce[i]` and put the modified version in `new_equivalence`,
69 // which will then replace `to_reduce[i]`.
70 let mut new_equivalence = Vec::with_capacity(to_reduce[i].len());
71 while let Some((_, mut popped_expr)) = to_reduce[i].pop() {
72 #[allow(deprecated)]
73 popped_expr.visit_mut_post_nolimit(&mut |e: &mut MirScalarExpr| {
74 // If a simpler expression can be found that is equivalent
75 // to e,
76 if let Some(simpler_e) = to_reduce.iter().find_map(|cls| {
77 if cls.iter().skip(1).position(|(_, expr)| e == expr).is_some() {
78 Some(cls[0].1.clone())
79 } else {
80 None
81 }
82 }) {
83 // Replace e with the simpler expression.
84 *e = simpler_e;
85 expressions_rewritten = true;
86 }
87 });
88 popped_expr.reduce(&column_types);
89 new_equivalence.push((rank_complexity(&popped_expr), popped_expr));
90 }
91 new_equivalence.sort();
92 new_equivalence.dedup();
93 to_reduce[i] = new_equivalence;
94 }
95 }
96
97 // Map away the complexity rating.
98 *equivalences = to_reduce
99 .drain(..)
100 .map(|mut cls| cls.drain(..).map(|(_, expr)| expr).collect::<Vec<_>>())
101 .collect::<Vec<_>>();
102
103 canonicalize_equivalence_classes(equivalences);
104}
105
106/// Canonicalize only the equivalence classes of a join.
107///
108/// This function:
109/// * ensures the same expression appears in only one equivalence class.
110/// * ensures the equivalence classes are sorted and dedupped.
111/// ```rust
112/// use mz_expr::MirScalarExpr;
113/// use mz_expr::canonicalize::canonicalize_equivalence_classes;
114///
115/// let mut equivalences = vec![
116/// vec![MirScalarExpr::column(1), MirScalarExpr::column(4)],
117/// vec![MirScalarExpr::column(3), MirScalarExpr::column(5)],
118/// vec![MirScalarExpr::column(0), MirScalarExpr::column(3)],
119/// vec![MirScalarExpr::column(2), MirScalarExpr::column(2)],
120/// ];
121/// let expected = vec![
122/// vec![MirScalarExpr::column(0),
123/// MirScalarExpr::column(3),
124/// MirScalarExpr::column(5)],
125/// vec![MirScalarExpr::column(1), MirScalarExpr::column(4)],
126/// ];
127/// canonicalize_equivalence_classes(&mut equivalences);
128/// assert_eq!(expected, equivalences)
129/// ````
130pub fn canonicalize_equivalence_classes(equivalences: &mut Vec<Vec<MirScalarExpr>>) {
131 let mut uf = BTreeMap::new();
132 for class in equivalences.iter_mut() {
133 let mut iter = class.drain(..);
134 if let Some(first) = iter.next() {
135 let head = Rc::new(first);
136 for rest in iter {
137 uf.union(&head, &Rc::new(rest));
138 }
139 }
140 }
141
142 let mut eqs: BTreeMap<Rc<MirScalarExpr>, BTreeSet<Rc<MirScalarExpr>>> = BTreeMap::new();
143 for (k, v) in uf {
144 eqs.entry(v).or_default().insert(Rc::clone(&k));
145 }
146
147 let classes = eqs.into_values().collect::<Vec<_>>();
148 equivalences.resize(classes.len(), Vec::new());
149 equivalences
150 .iter_mut()
151 .zip_eq(classes)
152 .for_each(|(equivalence, class)| {
153 equivalence.extend(
154 class
155 .into_iter()
156 .map(|e| Rc::try_unwrap(e).expect("there to be only one strong ref")),
157 );
158 });
159
160 equivalences.retain(|es| es.len() > 1);
161 equivalences.sort();
162}
163
164/// Gives a relative complexity ranking for an expression. Higher numbers mean
165/// greater complexity.
166///
167/// Currently, this method weighs literals as the least complex and weighs all
168/// other expressions by the number of non-literals. In the future, we can
169/// change how complexity is ranked so that repeated substitutions would result
170/// in arriving at "better" fixed points. For example, we could try to improve
171/// performance by ranking expressions by their estimated computation time.
172///
173/// To ensure we arrive at a fixed point after repeated substitutions, valid
174/// complexity rankings must fulfill the following property:
175/// For any expression `e`, there does not exist a SQL function `f` such
176/// that `complexity(e) >= complexity(f(e))`.
177///
178/// For ease of intuiting the fixed point that we will arrive at after
179/// repeated substitutions, it is nice but not required that complexity
180/// rankings additionally fulfill the following property:
181/// If expressions `e1` and `e2` are such that
182/// `complexity(e1) < complexity(e2)` then for all SQL functions `f`,
183/// `complexity(f(e1)) < complexity(f(e2))`.
184fn rank_complexity(expr: &MirScalarExpr) -> usize {
185 if expr.is_literal() {
186 // literals are the least complex
187 return 0;
188 }
189 let mut non_literal_count = 1;
190 expr.visit_pre(|e| {
191 if !e.is_literal() {
192 non_literal_count += 1
193 }
194 });
195 non_literal_count
196}
197
198/// Applies a flat_map on a Vec, and overwrites the vec with the result.
199fn flat_map_modify<T, I, F>(v: &mut Vec<T>, f: F)
200where
201 F: FnMut(T) -> I,
202 I: IntoIterator<Item = T>,
203{
204 let mut xx = v.drain(..).flat_map(f).collect();
205 v.append(&mut xx);
206}
207
208/// Canonicalize predicates of a filter.
209///
210/// This function reduces and canonicalizes the structure of each individual
211/// predicate. Then, it transforms predicates of the form "A and B" into two: "A"
212/// and "B". Afterwards, it reduces predicates based on information from other
213/// predicates in the set. Finally, it sorts and deduplicates the predicates.
214///
215/// Additionally, it also removes IS NOT NULL predicates if there is another
216/// null rejecting predicate for the same sub-expression.
217pub fn canonicalize_predicates(
218 predicates: &mut Vec<MirScalarExpr>,
219 column_types: &[SqlColumnType],
220) {
221 soft_assert_or_log!(
222 predicates
223 .iter()
224 .all(|p| p.typ(column_types).scalar_type == SqlScalarType::Bool),
225 "cannot canonicalize predicates that are not of type bool"
226 );
227
228 // 1) Reduce each individual predicate.
229 predicates.iter_mut().for_each(|p| p.reduce(column_types));
230
231 // 2) Split "A and B" into two predicates: "A" and "B"
232 // Relies on the `reduce` above having flattened nested ANDs.
233 flat_map_modify(predicates, |p| {
234 if let MirScalarExpr::CallVariadic {
235 func: VariadicFunc::And,
236 exprs,
237 } = p
238 {
239 exprs
240 } else {
241 vec![p]
242 }
243 });
244
245 // 3) Make non-null requirements explicit as predicates in order for
246 // step 4) to be able to simplify AND/OR expressions with IS NULL
247 // sub-predicates. This redundancy is removed later by step 5).
248 let mut non_null_columns = BTreeSet::new();
249 for p in predicates.iter() {
250 p.non_null_requirements(&mut non_null_columns);
251 }
252 predicates.extend(non_null_columns.iter().map(|c| {
253 MirScalarExpr::column(*c)
254 .call_unary(UnaryFunc::IsNull(func::IsNull))
255 .call_unary(UnaryFunc::Not(func::Not))
256 }));
257
258 // 4) Reduce across `predicates`.
259 // If a predicate `p` cannot be null, and `f(p)` is a nullable bool
260 // then the predicate `p & f(p)` is equal to `p & f(true)`, and
261 // `!p & f(p)` is equal to `!p & f(false)`. For any index i, the `Vec` of
262 // predicates `[p1, ... pi, ... pn]` is equivalent to the single predicate
263 // `pi & (p1 & ... & p(i-1) & p(i+1) ... & pn)`. Thus, if `pi`
264 // (resp. `!pi`) cannot be null, it is valid to replace with `true` (resp.
265 // `false`) every subexpression in `(p1 & ... & p(i-1) & p(i+1) ... & pn)`
266 // that is equal to `pi`.
267
268 // If `p` is null and `q` is a nullable bool, then `p & q` can be either
269 // `null` or `false` depending on what `q`. Our rendering pipeline treats
270 // both as "remove this row." Thus, in the specific context of filter
271 // predicates, it is acceptable to make the aforementioned substitution
272 // even if `pi` can be null.
273
274 // Note that this does some dedupping of predicates since if `p1 = p2`
275 // then this reduction process will replace `p1` with true.
276
277 // Maintain respectively:
278 // 1) A list of predicates for which we have checked for matching
279 // subexpressions
280 // 2) A list of predicates for which we have yet to do so.
281 let mut completed = Vec::new();
282 let mut todo = Vec::new();
283 // Seed `todo` with all predicates.
284 std::mem::swap(&mut todo, predicates);
285
286 while let Some(predicate_to_apply) = todo.pop() {
287 // Helper method: for each predicate `p`, see if all other predicates
288 // (a.k.a. the union of todo & completed) contains `p` as a
289 // subexpression, and replace the subexpression accordingly.
290 // This method lives inside the loop because in order to comply with
291 // Rust rules that only one mutable reference to `todo` can be held at a
292 // time.
293 let mut replace_subexpr_other_predicates =
294 |expr: &MirScalarExpr, constant_bool: &MirScalarExpr| {
295 // Do not replace subexpressions equal to `expr` if `expr` is a
296 // literal to avoid infinite looping.
297 if !expr.is_literal() {
298 for other_predicate in todo.iter_mut() {
299 replace_subexpr_and_reduce(
300 other_predicate,
301 expr,
302 constant_bool,
303 column_types,
304 );
305 }
306 for other_idx in (0..completed.len()).rev() {
307 if replace_subexpr_and_reduce(
308 &mut completed[other_idx],
309 expr,
310 constant_bool,
311 column_types,
312 ) {
313 // If a predicate in the `completed` list has
314 // been simplified, stick it back into the `todo` list.
315 todo.push(completed.remove(other_idx));
316 }
317 }
318 }
319 };
320 // Meat of loop starts here. If a predicate p is of the form `!q`, replace
321 // every instance of `q` in every other predicate with `false.`
322 // Otherwise, replace every instance of `p` in every other predicate
323 // with `true`.
324 if let MirScalarExpr::CallUnary {
325 func: UnaryFunc::Not(func::Not),
326 expr,
327 } = &predicate_to_apply
328 {
329 replace_subexpr_other_predicates(expr, &MirScalarExpr::literal_false())
330 } else {
331 replace_subexpr_other_predicates(&predicate_to_apply, &MirScalarExpr::literal_true());
332 }
333 completed.push(predicate_to_apply);
334 }
335
336 // 5) Remove redundant !isnull/isnull predicates after performing the replacements
337 // in the loop above.
338 std::mem::swap(&mut todo, &mut completed);
339 while let Some(predicate_to_apply) = todo.pop() {
340 // Remove redundant !isnull(x) predicates if there is another predicate
341 // that evaluates to NULL when `x` is NULL.
342 if let Some(operand) = is_not_null(&predicate_to_apply) {
343 if todo
344 .iter_mut()
345 .chain(completed.iter_mut())
346 .any(|p| is_null_rejecting_predicate(p, &operand))
347 {
348 // skip this predicate
349 continue;
350 }
351 } else if let MirScalarExpr::CallUnary {
352 func: UnaryFunc::IsNull(func::IsNull),
353 expr,
354 } = &predicate_to_apply
355 {
356 if todo
357 .iter_mut()
358 .chain(completed.iter_mut())
359 .any(|p| is_null_rejecting_predicate(p, expr))
360 {
361 completed.push(MirScalarExpr::literal_false());
362 break;
363 }
364 }
365 completed.push(predicate_to_apply);
366 }
367
368 if completed.iter().any(|p| {
369 (p.is_literal_false() || p.is_literal_null()) &&
370 // This extra check is only needed if we determine that the soft-assert
371 // at the top of this function would ever fail for a good reason.
372 p.typ(column_types).scalar_type == SqlScalarType::Bool
373 }) {
374 // all rows get filtered away if any predicate is null or false.
375 *predicates = vec![MirScalarExpr::literal_false()]
376 } else {
377 // Remove any predicates that have been reduced to "true"
378 completed.retain(|p| !p.is_literal_true());
379 *predicates = completed;
380 }
381
382 // 6) Sort and dedup predicates.
383 predicates.sort_by(compare_predicates);
384 predicates.dedup();
385}
386
387/// Replace any matching subexpressions in `predicate`, and if `predicate` has
388/// changed, reduce it. Return whether `predicate` has changed.
389fn replace_subexpr_and_reduce(
390 predicate: &mut MirScalarExpr,
391 replace_if_equal_to: &MirScalarExpr,
392 replace_with: &MirScalarExpr,
393 column_types: &[SqlColumnType],
394) -> bool {
395 let mut changed = false;
396 #[allow(deprecated)]
397 predicate.visit_mut_pre_post_nolimit(
398 &mut |e| {
399 // The `cond` of an if statement is not visited to prevent `then`
400 // or `els` from being evaluated before `cond`, resulting in a
401 // correctness error.
402 if let MirScalarExpr::If { then, els, .. } = e {
403 return Some(vec![then, els]);
404 }
405 None
406 },
407 &mut |e| {
408 if e == replace_if_equal_to {
409 *e = replace_with.clone();
410 changed = true;
411 } else if let MirScalarExpr::CallBinary {
412 func: r_func,
413 expr1: r_expr1,
414 expr2: r_expr2,
415 } = replace_if_equal_to
416 {
417 if let Some(negation) = r_func.negate() {
418 if let MirScalarExpr::CallBinary {
419 func: l_func,
420 expr1: l_expr1,
421 expr2: l_expr2,
422 } = e
423 {
424 if negation == *l_func && l_expr1 == r_expr1 && l_expr2 == r_expr2 {
425 *e = MirScalarExpr::CallUnary {
426 func: UnaryFunc::Not(func::Not),
427 expr: Box::new(replace_with.clone()),
428 };
429 changed = true;
430 }
431 }
432 }
433 }
434 },
435 );
436 if changed {
437 predicate.reduce(column_types);
438 }
439 changed
440}
441
442/// Returns the inner operand if the given predicate is an IS NOT NULL expression.
443fn is_not_null(predicate: &MirScalarExpr) -> Option<MirScalarExpr> {
444 if let MirScalarExpr::CallUnary {
445 func: UnaryFunc::Not(func::Not),
446 expr,
447 } = &predicate
448 {
449 if let MirScalarExpr::CallUnary {
450 func: UnaryFunc::IsNull(func::IsNull),
451 expr,
452 } = &**expr
453 {
454 return Some((**expr).clone());
455 }
456 }
457 None
458}
459
460/// Whether the given predicate evaluates to NULL when the given operand expression is NULL.
461#[inline(always)]
462fn is_null_rejecting_predicate(predicate: &MirScalarExpr, operand: &MirScalarExpr) -> bool {
463 propagates_null_from_subexpression(predicate, operand)
464}
465
466fn propagates_null_from_subexpression(expr: &MirScalarExpr, operand: &MirScalarExpr) -> bool {
467 if operand == expr {
468 true
469 } else if let MirScalarExpr::CallVariadic { func, exprs } = &expr {
470 func.propagates_nulls()
471 && (exprs
472 .iter()
473 .any(|e| propagates_null_from_subexpression(e, operand)))
474 } else if let MirScalarExpr::CallBinary { func, expr1, expr2 } = &expr {
475 func.propagates_nulls()
476 && (propagates_null_from_subexpression(expr1, operand)
477 || propagates_null_from_subexpression(expr2, operand))
478 } else if let MirScalarExpr::CallUnary { func, expr } = &expr {
479 func.propagates_nulls() && propagates_null_from_subexpression(expr, operand)
480 } else {
481 false
482 }
483}
484
485/// Comparison method for sorting predicates by their complexity, measured by the total
486/// number of non-literal expression nodes within the expression.
487fn compare_predicates(x: &MirScalarExpr, y: &MirScalarExpr) -> Ordering {
488 (rank_complexity(x), x).cmp(&(rank_complexity(y), y))
489}
490
491/// For each equivalence class, it finds the simplest expression, which will be the canonical one.
492/// Returns a Map that maps from each expression in each equivalence class to the canonical
493/// expression in the same equivalence class.
494pub fn get_canonicalizer_map(
495 equivalences: &Vec<Vec<MirScalarExpr>>,
496) -> BTreeMap<MirScalarExpr, MirScalarExpr> {
497 let mut canonicalizer_map = BTreeMap::new();
498 for equivalence in equivalences {
499 // The unwrap is ok, because a join equivalence class can't be empty.
500 let canonical_expr = equivalence
501 .iter()
502 .min_by(|a, b| compare_predicates(*a, *b))
503 .unwrap();
504 for e in equivalence {
505 if e != canonical_expr {
506 canonicalizer_map.insert(e.clone(), canonical_expr.clone());
507 }
508 }
509 }
510 canonicalizer_map
511}
512
513/// A trait for a union-find data structure.
514pub trait UnionFind<T> {
515 /// Sets `self[x]` to the root from `x`, and returns a reference to the root.
516 fn find<'a>(&'a mut self, x: &T) -> Option<&'a T>;
517 /// Ensures that `x` and `y` have the same root.
518 fn union(&mut self, x: &T, y: &T);
519}
520
521impl<T: Clone + Ord> UnionFind<T> for BTreeMap<T, T> {
522 fn find<'a>(&'a mut self, x: &T) -> Option<&'a T> {
523 if !self.contains_key(x) {
524 None
525 } else {
526 if self[x] != self[&self[x]] {
527 // Path halving
528 let mut y = self[x].clone();
529 while y != self[&y] {
530 let grandparent = self[&self[&y]].clone();
531 *self.get_mut(&y).unwrap() = grandparent;
532 y.clone_from(&self[&y]);
533 }
534 *self.get_mut(x).unwrap() = y;
535 }
536 Some(&self[x])
537 }
538 }
539
540 fn union(&mut self, x: &T, y: &T) {
541 match (self.find(x).is_some(), self.find(y).is_some()) {
542 (true, true) => {
543 if self[x] != self[y] {
544 let root_x = self[x].clone();
545 let root_y = self[y].clone();
546 self.insert(root_x, root_y);
547 }
548 }
549 (false, true) => {
550 self.insert(x.clone(), self[y].clone());
551 }
552 (true, false) => {
553 self.insert(y.clone(), self[x].clone());
554 }
555 (false, false) => {
556 self.insert(x.clone(), x.clone());
557 self.insert(y.clone(), x.clone());
558 }
559 }
560 }
561}