Make LastStores a proper lattice

cranelift/codegen/src/egraph/mod.rs

@@ -1,7 +1,7 @@
 //! Support for egraphs represented in the DataFlowGraph.
 
 use crate::FxHashSet;
-use crate::alias_analysis::{AliasAnalysis, LastStores, OptResult};
+use crate::alias_analysis::{AliasAnalysis, MemoryState, OptResult};
 use crate::branch_to_trap::BranchToTrapAnalysis;
 use crate::ctxhash::{CtxEq, CtxHash, NullCtx};
 use crate::cursor::{Cursor, CursorPosition, FuncCursor};
@@ -155,7 +155,7 @@ where
     /// build a post-dominator tree for dead-store elimination.
     cfg: &'opt ControlFlowGraph,
     pub(crate) alias_analysis: &'opt mut AliasAnalysis<'analysis>,
-    pub(crate) alias_analysis_state: &'opt mut LastStores,
+    pub(crate) alias_analysis_state: &'opt mut MemoryState,
     pub(crate) branch_to_trap_analysis: &'opt mut BranchToTrapAnalysis,
     ctrl_plane: &'opt mut ControlPlane,
     // Held locally during optimization of one node (recursively):
@@ -923,11 +923,14 @@ impl<'a> EgraphPass<'a> {
             {
                 gvn_map_blocks.pop();
                 gvn_map.decrement_depth();
+                self.alias_analysis.pop_scope();
             }
 
             gvn_map.increment_depth();
             gvn_map_blocks.push(block);
 
+            let mut alias_analysis_state = self.alias_analysis.push_scope(&*self.cfg, block);
+
             // Check that `gvn_map_blocks` is the path from this block up to the
             // root in the dominator tree.
             debug_assert_eq!(gvn_map_blocks, {
@@ -945,8 +948,6 @@ impl<'a> EgraphPass<'a> {
             trace!("Processing block {}", block);
             cursor.set_position(CursorPosition::Before(block));
 
-            let mut alias_analysis_state = self.alias_analysis.block_starting_state(block);
-
             for &param in cursor.func.dfg.block_params(block) {
                 trace!("creating initial singleton eclass for blockparam {}", param);
                 value_to_opt_value[param] = param;

cranelift/filetests/filetests/alias/fence-fallback-across-join.clif

@@ -0,0 +1,77 @@
+test optimize precise-output
+set opt_level=speed
+target aarch64
+
+function %fence_fallback_survives_join(i64, i32) -> i32, i32 {
+    region0 = 0 "R0"
+
+block0(v0: i64, v1: i32):
+    v3 = load.i32 notrap aligned region0 v0
+    brif v1, block1, block2
+
+block1:
+    jump block3
+
+block2:
+    jump block3
+
+block3:
+    v4 = load.i32 notrap aligned region0 v0
+    return v3, v4
+}
+
+; function %fence_fallback_survives_join(i64, i32) -> i32, i32 fast {
+;     region0 = 0 "R0"
+;
+; block0(v0: i64, v1: i32):
+;     v3 = load.i32 notrap aligned region0 v0
+;     brif v1, block1, block2
+;
+; block1:
+;     jump block3
+;
+; block2:
+;     jump block3
+;
+; block3:
+;     return v3, v3
+; }
+
+function %unknown_does_not_take_the_fence_fallback(i64, i32, i32) -> i32, i32 {
+    region0 = 0 "R0"
+
+block0(v0: i64, v1: i32, v2: i32):
+    v3 = load.i32 notrap aligned region0 v0
+    brif v1, block1, block2
+
+block1:
+    ;; This store prevents `block3` from reusing `v3`.
+    store notrap aligned region0 v2, v0
+    jump block3
+
+block2:
+    jump block3
+
+block3:
+    v4 = load.i32 notrap aligned region0 v0
+    return v3, v4
+}
+
+; function %unknown_does_not_take_the_fence_fallback(i64, i32, i32) -> i32, i32 fast {
+;     region0 = 0 "R0"
+;
+; block0(v0: i64, v1: i32, v2: i32):
+;     v3 = load.i32 notrap aligned region0 v0
+;     brif v1, block1, block2
+;
+; block1:
+;     store.i32 notrap aligned region0 v2, v0
+;     jump block3
+;
+; block2:
+;     jump block3
+;
+; block3:
+;     v4 = load.i32 notrap aligned region0 v0
+;     return v3, v4
+; }

cranelift/filetests/filetests/alias/join-order-independence.clif

@@ -0,0 +1,85 @@
+test optimize precise-output
+set opt_level=speed
+target aarch64
+
+;; The last-store analysis must not depend on the order in which the worklist
+;; happens to visit blocks. These two functions differ only in the order of the
+;; `brif`'s two arms, so they must optimize identically.
+
+function %arms_in_order(i64, i32, i32) -> i32, i32 {
+    region0 = 0 "R0"
+
+block0(v0: i64, v1: i32, v2: i32):
+    store notrap aligned region0 v1, v0
+    brif v1, block3, block1
+
+block1:
+    store notrap aligned region0 v2, v0
+    jump block3
+
+block3:
+    v3 = load.i32 notrap aligned region0 v0
+    jump block4
+
+block4:
+    v4 = load.i32 notrap aligned region0 v0
+    return v3, v4
+}
+
+; function %arms_in_order(i64, i32, i32) -> i32, i32 fast {
+;     region0 = 0 "R0"
+;
+; block0(v0: i64, v1: i32, v2: i32):
+;     store notrap aligned region0 v1, v0
+;     brif v1, block3, block1
+;
+; block1:
+;     store.i32 notrap aligned region0 v2, v0
+;     jump block3
+;
+; block3:
+;     v3 = load.i32 notrap aligned region0 v0
+;     jump block4
+;
+; block4:
+;     return v3, v3
+; }
+
+function %arms_swapped(i64, i32, i32) -> i32, i32 {
+    region0 = 0 "R0"
+
+block0(v0: i64, v1: i32, v2: i32):
+    store notrap aligned region0 v1, v0
+    brif v1, block1, block3
+
+block1:
+    store notrap aligned region0 v2, v0
+    jump block3
+
+block3:
+    v3 = load.i32 notrap aligned region0 v0
+    jump block4
+
+block4:
+    v4 = load.i32 notrap aligned region0 v0
+    return v3, v4
+}
+
+; function %arms_swapped(i64, i32, i32) -> i32, i32 fast {
+;     region0 = 0 "R0"
+;
+; block0(v0: i64, v1: i32, v2: i32):
+;     store notrap aligned region0 v1, v0
+;     brif v1, block1, block3
+;
+; block1:
+;     store.i32 notrap aligned region0 v2, v0
+;     jump block3
+;
+; block3:
+;     v3 = load.i32 notrap aligned region0 v0
+;     jump block4
+;
+; block4:
+;     return v3, v3
+; }

cranelift/filetests/filetests/alias/join-single-predecessor.clif

@@ -0,0 +1,57 @@
+test optimize precise-output
+set opt_level=speed
+target aarch64
+
+function %single_pred_inherits(i64, i32, i32, i32) -> i32, i32 {
+    region0 = 0 "R0"
+
+block0(v0: i64, v1: i32, v2: i32, v3: i32):
+    store notrap aligned region0 v1, v0
+    jump block1
+
+block1:
+    brif v2, block2, block3
+
+block3:
+    jump block2
+
+block2:
+    v10 = load.i32 notrap aligned region0 v0
+    jump block4
+
+block4:
+    ;; `block4` has exactly one predecessor, `block2`, so at the fixpoint its
+    ;; region0 slot must equal `block2`'s outgoing value, and `v11` must
+    ;; therefore be forwarded to `v10`.
+    v11 = load.i32 notrap aligned region0 v0
+    store notrap aligned region0 v3, v0
+    brif v1, block1, block5
+
+block5:
+    return v10, v11
+}
+
+; function %single_pred_inherits(i64, i32, i32, i32) -> i32, i32 fast {
+;     region0 = 0 "R0"
+;
+; block0(v0: i64, v1: i32, v2: i32, v3: i32):
+;     store notrap aligned region0 v1, v0
+;     jump block1
+;
+; block1:
+;     brif.i32 v2, block2, block3
+;
+; block3:
+;     jump block2
+;
+; block2:
+;     v10 = load.i32 notrap aligned region0 v0
+;     jump block4
+;
+; block4:
+;     store.i32 notrap aligned region0 v3, v0
+;     brif.i32 v1, block1, block5
+;
+; block5:
+;     return v10, v10
+; }

cranelift/filetests/filetests/alias/merge-token-identity.clif

@@ -0,0 +1,63 @@
+test optimize precise-output
+set opt_level=speed
+target aarch64
+
+;; `v3` and `v4` sit after two independent joins, with `block4`'s store in
+;; between.
+function %two_independent_joins(i64, i32, i32) -> i32, i32 {
+    region0 = 0 "R0"
+
+block0(v0: i64, v1: i32, v2: i32):
+    brif v1, block1, block2
+
+block1:
+    store notrap aligned region0 v2, v0
+    jump block3
+
+block2:
+    jump block3
+
+block3:
+    v3 = load.i32 notrap aligned region0 v0
+    brif v1, block4, block5
+
+block4:
+    store notrap aligned region0 v1, v0
+    jump block6
+
+block5:
+    jump block6
+
+block6:
+    v4 = load.i32 notrap aligned region0 v0
+    return v3, v4
+}
+
+; function %two_independent_joins(i64, i32, i32) -> i32, i32 fast {
+;     region0 = 0 "R0"
+;
+; block0(v0: i64, v1: i32, v2: i32):
+;     brif v1, block1, block2
+;
+; block1:
+;     store.i32 notrap aligned region0 v2, v0
+;     jump block3
+;
+; block2:
+;     jump block3
+;
+; block3:
+;     v3 = load.i32 notrap aligned region0 v0
+;     brif.i32 v1, block4, block5
+;
+; block4:
+;     store.i32 notrap aligned region0 v1, v0
+;     jump block6
+;
+; block5:
+;     jump block6
+;
+; block6:
+;     v4 = load.i32 notrap aligned region0 v0
+;     return v3, v4
+; }

tests/disas/gc/array-copy-with-fuel.wat

@@ -93,8 +93,6 @@
 ;; @002b                               brif.i32 v6, block4, block7
 ;;
 ;;                                 block4:
-;;                                     v143 = load.i32 notrap aligned region6 v162
-;;                                     v145 = load.i32 notrap aligned region7 v163
 ;; @002b                               v89 = icmp.i64 ult v45, v69
 ;; @002b                               v94 = iadd.i64 v45, v170
 ;; @002b                               v95 = iadd.i64 v69, v170