1 //! This optimization replaces stack accesses with SSA variables and removes dead stores when possible.
3 //! # Undefined behaviour
5 //! This optimization is based on the assumption that stack slots which don't have their address
6 //! leaked through `stack_addr` are only accessed using `stack_load` and `stack_store` in the
7 //! function which has the stack slots. This optimization also assumes that stack slot accesses
8 //! are never out of bounds. If these assumptions are not correct, then this optimization may remove
9 //! `stack_store` instruction incorrectly, or incorrectly use a previously stored value as the value
10 //! being loaded by a `stack_load`.
12 use std::collections::BTreeMap;
16 use rustc_data_structures::fx::FxHashSet;
18 use cranelift_codegen::cursor::{Cursor, FuncCursor};
19 use cranelift_codegen::ir::immediates::Offset32;
20 use cranelift_codegen::ir::{InstructionData, Opcode, ValueDef};
22 use crate::prelude::*;
24 /// Workaround for `StackSlot` not implementing `Ord`.
25 #[derive(Copy, Clone, PartialEq, Eq)]
26 struct OrdStackSlot(StackSlot);
28 impl fmt::Debug for OrdStackSlot {
29 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
30 write!(f, "{:?}", self.0)
34 impl PartialOrd for OrdStackSlot {
35 fn partial_cmp(&self, rhs: &Self) -> Option<std::cmp::Ordering> {
36 self.0.as_u32().partial_cmp(&rhs.0.as_u32())
40 impl Ord for OrdStackSlot {
41 fn cmp(&self, rhs: &Self) -> std::cmp::Ordering {
42 self.0.as_u32().cmp(&rhs.0.as_u32())
46 #[derive(Debug, Default)]
47 struct StackSlotUsage {
48 stack_addr: FxHashSet<Inst>,
49 stack_load: FxHashSet<Inst>,
50 stack_store: FxHashSet<Inst>,
54 fn potential_stores_for_load(&self, ctx: &Context, load: Inst) -> Vec<Inst> {
59 match spatial_overlap(&ctx.func, store, load) {
60 SpatialOverlap::No => false, // Can never be the source of the loaded value.
61 SpatialOverlap::Partial | SpatialOverlap::Full => true,
65 match temporal_order(ctx, store, load) {
66 TemporalOrder::NeverBefore => false, // Can never be the source of the loaded value.
67 TemporalOrder::MaybeBefore | TemporalOrder::DefinitivelyBefore => true,
70 .collect::<Vec<Inst>>()
73 fn potential_loads_of_store(&self, ctx: &Context, store: Inst) -> Vec<Inst> {
78 match spatial_overlap(&ctx.func, store, load) {
79 SpatialOverlap::No => false, // Can never be the source of the loaded value.
80 SpatialOverlap::Partial | SpatialOverlap::Full => true,
84 match temporal_order(ctx, store, load) {
85 TemporalOrder::NeverBefore => false, // Can never be the source of the loaded value.
86 TemporalOrder::MaybeBefore | TemporalOrder::DefinitivelyBefore => true,
89 .collect::<Vec<Inst>>()
92 fn remove_unused_stack_addr(func: &mut Function, inst: Inst) {
93 func.dfg.detach_results(inst);
94 func.dfg.replace(inst).nop();
97 fn remove_unused_load(func: &mut Function, load: Inst) {
98 func.dfg.detach_results(load);
99 func.dfg.replace(load).nop();
102 fn remove_dead_store(&mut self, func: &mut Function, store: Inst) {
103 func.dfg.replace(store).nop();
104 self.stack_store.remove(&store);
107 fn change_load_to_alias(&mut self, func: &mut Function, load: Inst, value: Value) {
108 let loaded_value = func.dfg.inst_results(load)[0];
109 let loaded_type = func.dfg.value_type(loaded_value);
111 if func.dfg.value_type(value) == loaded_type {
112 func.dfg.detach_results(load);
113 func.dfg.replace(load).nop();
114 func.dfg.change_to_alias(loaded_value, value);
116 func.dfg.replace(load).bitcast(loaded_type, value);
119 self.stack_load.remove(&load);
123 struct OptimizeContext<'a> {
124 ctx: &'a mut Context,
125 stack_slot_usage_map: BTreeMap<OrdStackSlot, StackSlotUsage>,
128 impl<'a> OptimizeContext<'a> {
129 fn for_context(ctx: &'a mut Context) -> Self {
130 ctx.flowgraph(); // Compute cfg and domtree.
132 // Record all stack_addr, stack_load and stack_store instructions.
133 let mut stack_slot_usage_map = BTreeMap::<OrdStackSlot, StackSlotUsage>::new();
135 let mut cursor = FuncCursor::new(&mut ctx.func);
136 while let Some(_block) = cursor.next_block() {
137 while let Some(inst) = cursor.next_inst() {
138 match cursor.func.dfg[inst] {
139 InstructionData::StackLoad {
140 opcode: Opcode::StackAddr,
145 .entry(OrdStackSlot(stack_slot))
146 .or_insert_with(StackSlotUsage::default)
150 InstructionData::StackLoad {
151 opcode: Opcode::StackLoad,
156 .entry(OrdStackSlot(stack_slot))
157 .or_insert_with(StackSlotUsage::default)
161 InstructionData::StackStore {
162 opcode: Opcode::StackStore,
168 .entry(OrdStackSlot(stack_slot))
169 .or_insert_with(StackSlotUsage::default)
180 stack_slot_usage_map,
185 pub(super) fn optimize_function(
187 clif_comments: &mut crate::pretty_clif::CommentWriter,
189 combine_stack_addr_with_load_store(&mut ctx.func);
191 let mut opt_ctx = OptimizeContext::for_context(ctx);
193 // FIXME Repeat following instructions until fixpoint.
195 remove_unused_stack_addr_and_stack_load(&mut opt_ctx);
197 #[cfg(debug_assertions)]
199 for (&OrdStackSlot(stack_slot), usage) in &opt_ctx.stack_slot_usage_map {
200 clif_comments.add_comment(stack_slot, format!("used by: {:?}", usage));
204 for (stack_slot, users) in opt_ctx.stack_slot_usage_map.iter_mut() {
205 if users.stack_addr.is_empty().not() {
206 // Stack addr leaked; there may be unknown loads and stores.
207 // FIXME use stacked borrows to optimize
211 for load in users.stack_load.clone().into_iter() {
212 let potential_stores = users.potential_stores_for_load(&opt_ctx.ctx, load);
214 #[cfg(debug_assertions)]
215 for &store in &potential_stores {
216 clif_comments.add_comment(
219 "Potential store -> load forwarding {} -> {} ({:?}, {:?})",
220 opt_ctx.ctx.func.dfg.display_inst(store, None),
221 opt_ctx.ctx.func.dfg.display_inst(load, None),
222 spatial_overlap(&opt_ctx.ctx.func, store, load),
223 temporal_order(&opt_ctx.ctx, store, load),
228 match *potential_stores {
230 #[cfg(debug_assertions)]
231 clif_comments.add_comment(load, format!("[BUG?] Reading uninitialized memory"));
234 if spatial_overlap(&opt_ctx.ctx.func, store, load) == SpatialOverlap::Full
235 && temporal_order(&opt_ctx.ctx, store, load)
236 == TemporalOrder::DefinitivelyBefore =>
238 // Only one store could have been the origin of the value.
239 let stored_value = opt_ctx.ctx.func.dfg.inst_args(store)[0];
241 #[cfg(debug_assertions)]
243 .add_comment(load, format!("Store to load forward {} -> {}", store, load));
245 users.change_load_to_alias(&mut opt_ctx.ctx.func, load, stored_value);
247 _ => {} // FIXME implement this
251 for store in users.stack_store.clone().into_iter() {
252 let potential_loads = users.potential_loads_of_store(&opt_ctx.ctx, store);
254 #[cfg(debug_assertions)]
255 for &load in &potential_loads {
256 clif_comments.add_comment(
259 "Potential load from store {} <- {} ({:?}, {:?})",
260 opt_ctx.ctx.func.dfg.display_inst(load, None),
261 opt_ctx.ctx.func.dfg.display_inst(store, None),
262 spatial_overlap(&opt_ctx.ctx.func, store, load),
263 temporal_order(&opt_ctx.ctx, store, load),
268 if potential_loads.is_empty() {
269 // Never loaded; can safely remove all stores and the stack slot.
270 // FIXME also remove stores when there is always a next store before a load.
272 #[cfg(debug_assertions)]
273 clif_comments.add_comment(
276 "Remove dead stack store {} of {}",
277 opt_ctx.ctx.func.dfg.display_inst(store, None),
282 users.remove_dead_store(&mut opt_ctx.ctx.func, store);
286 if users.stack_store.is_empty() && users.stack_load.is_empty() {
287 opt_ctx.ctx.func.stack_slots[stack_slot.0].size = 0;
292 fn combine_stack_addr_with_load_store(func: &mut Function) {
293 // Turn load and store into stack_load and stack_store when possible.
294 let mut cursor = FuncCursor::new(func);
295 while let Some(_block) = cursor.next_block() {
296 while let Some(inst) = cursor.next_inst() {
297 match cursor.func.dfg[inst] {
298 InstructionData::Load {
299 opcode: Opcode::Load,
304 if cursor.func.dfg.ctrl_typevar(inst) == types::I128
305 || cursor.func.dfg.ctrl_typevar(inst).is_vector()
307 continue; // WORKAROUD: stack_load.i128 not yet implemented
309 if let Some((stack_slot, stack_addr_offset)) =
310 try_get_stack_slot_and_offset_for_addr(cursor.func, addr)
312 if let Some(combined_offset) = offset.try_add_i64(stack_addr_offset.into())
314 let ty = cursor.func.dfg.ctrl_typevar(inst);
315 cursor.func.dfg.replace(inst).stack_load(
323 InstructionData::Store {
324 opcode: Opcode::Store,
329 if cursor.func.dfg.ctrl_typevar(inst) == types::I128
330 || cursor.func.dfg.ctrl_typevar(inst).is_vector()
332 continue; // WORKAROUND: stack_store.i128 not yet implemented
334 if let Some((stack_slot, stack_addr_offset)) =
335 try_get_stack_slot_and_offset_for_addr(cursor.func, addr)
337 if let Some(combined_offset) = offset.try_add_i64(stack_addr_offset.into())
339 cursor.func.dfg.replace(inst).stack_store(
353 fn remove_unused_stack_addr_and_stack_load(opt_ctx: &mut OptimizeContext<'_>) {
354 // FIXME incrementally rebuild on each call?
355 let mut stack_addr_load_insts_users = FxHashMap::<Inst, FxHashSet<Inst>>::default();
357 let mut cursor = FuncCursor::new(&mut opt_ctx.ctx.func);
358 while let Some(_block) = cursor.next_block() {
359 while let Some(inst) = cursor.next_inst() {
360 for &arg in cursor.func.dfg.inst_args(inst) {
361 if let ValueDef::Result(arg_origin, 0) = cursor.func.dfg.value_def(arg) {
362 match cursor.func.dfg[arg_origin].opcode() {
363 Opcode::StackAddr | Opcode::StackLoad => {
364 stack_addr_load_insts_users
366 .or_insert_with(FxHashSet::default)
376 #[cfg(debug_assertions)]
377 for inst in stack_addr_load_insts_users.keys() {
378 let mut is_recorded_stack_addr_or_stack_load = false;
379 for stack_slot_users in opt_ctx.stack_slot_usage_map.values() {
380 is_recorded_stack_addr_or_stack_load |= stack_slot_users.stack_addr.contains(inst)
381 || stack_slot_users.stack_load.contains(inst);
383 assert!(is_recorded_stack_addr_or_stack_load);
386 // Replace all unused stack_addr and stack_load instructions with nop.
387 let mut func = &mut opt_ctx.ctx.func;
389 for stack_slot_users in opt_ctx.stack_slot_usage_map.values_mut() {
392 .drain_filter(|inst| {
393 stack_addr_load_insts_users
395 .map(|users| users.is_empty())
398 .for_each(|inst| StackSlotUsage::remove_unused_stack_addr(&mut func, inst));
402 .drain_filter(|inst| {
403 stack_addr_load_insts_users
405 .map(|users| users.is_empty())
408 .for_each(|inst| StackSlotUsage::remove_unused_load(&mut func, inst));
412 fn try_get_stack_slot_and_offset_for_addr(
415 ) -> Option<(StackSlot, Offset32)> {
416 if let ValueDef::Result(addr_inst, 0) = func.dfg.value_def(addr) {
417 if let InstructionData::StackLoad {
418 opcode: Opcode::StackAddr,
421 } = func.dfg[addr_inst]
423 return Some((stack_slot, offset));
429 #[derive(Copy, Clone, Debug, PartialEq, Eq)]
430 enum SpatialOverlap {
436 fn spatial_overlap(func: &Function, src: Inst, dest: Inst) -> SpatialOverlap {
437 fn inst_info(func: &Function, inst: Inst) -> (StackSlot, Offset32, u32) {
438 match func.dfg[inst] {
439 InstructionData::StackLoad {
440 opcode: Opcode::StackAddr,
444 | InstructionData::StackLoad {
445 opcode: Opcode::StackLoad,
449 | InstructionData::StackStore {
450 opcode: Opcode::StackStore,
454 } => (stack_slot, offset, func.dfg.ctrl_typevar(inst).bytes()),
455 _ => unreachable!("{:?}", func.dfg[inst]),
459 debug_assert_ne!(src, dest);
461 let (src_ss, src_offset, src_size) = inst_info(func, src);
462 let (dest_ss, dest_offset, dest_size) = inst_info(func, dest);
464 if src_ss != dest_ss {
465 return SpatialOverlap::No;
468 if src_offset == dest_offset && src_size == dest_size {
469 return SpatialOverlap::Full;
472 let src_end: i64 = src_offset.try_add_i64(i64::from(src_size)).unwrap().into();
473 let dest_end: i64 = dest_offset
474 .try_add_i64(i64::from(dest_size))
477 if src_end <= dest_offset.into() || dest_end <= src_offset.into() {
478 return SpatialOverlap::No;
481 SpatialOverlap::Partial
484 #[derive(Copy, Clone, Debug, PartialEq, Eq)]
486 /// `src` will never be executed before `dest`.
489 /// `src` may be executed before `dest`.
492 /// `src` will always be executed before `dest`.
493 /// There may still be other instructions in between.
497 fn temporal_order(ctx: &Context, src: Inst, dest: Inst) -> TemporalOrder {
498 debug_assert_ne!(src, dest);
500 if ctx.domtree.dominates(src, dest, &ctx.func.layout) {
501 TemporalOrder::DefinitivelyBefore
502 } else if ctx.domtree.dominates(src, dest, &ctx.func.layout) {
503 TemporalOrder::NeverBefore
505 TemporalOrder::MaybeBefore