jit: restore recursive unroll state transitions (#1176)

* jit: restore recursive unroll state transitions

* wasm: keep recursive bridges compiled

Author: youknowone

majit/majit-backend-wasm/src/lib.rs

@@ -100,16 +100,17 @@ fn diag_bump(i: usize) {
 // A source token is compiled before a later guard may become a CA bridge.
 // Freeze modest room for that bridge at first compilation; a later trace that
 // exceeds either bound is declined rather than changing the live frame's
-// offsets. The fib CA path measured raw maxima of 9 positional slots and 15
-// Ref homes, but that is not a bridge fail-arg census for the full suite; keep
-// both existing floors until such a census establishes smaller safe bounds.
+// offsets. The recursive-unroll fib CA bridge needs more than 64 Ref homes:
+// declining it leaves the recursive return guard permanently blackholed and
+// turns every later invocation into a host round-trip. Keep enough bounded
+// per-token reserve for that bridge and the existing full-suite shapes.
 // A CALL_ASSEMBLER target must retain enough frozen spill/home geometry for a
 // later exit bridge.  nbody's callee bridge needs more Ref homes than the old
 // 16-slot floor; declining it turns every CA invocation into a blackhole.  The
 // larger fixed reserve keeps the bridge in compiled wasm and is still bounded
 // per compiled token.
 const FROZEN_CHAIN_VALUE_SLOTS: usize = 64;
-const FROZEN_CHAIN_REF_HOMES: usize = 64;
+const FROZEN_CHAIN_REF_HOMES: usize = 128;
 
 /// An op whose result advances loop-carried state. A value produced inside the
 /// re-running region by arithmetic or by a heap load is fresh on each pass, so

majit/majit-metainterp/src/pyjitpl.rs

@@ -13831,11 +13831,11 @@ impl<M: Clone> MetaInterp<M> {
 
         // pyjitpl.py:1404 `dont_trace_here(greenboxes)` — the only `&mut`,
         // flagged by `decide_recursive_inline` when recursion reached
-        // `max_unroll_recursion`. This is the ONLY place the recursion bound
-        // applies it: pyre's FBW walker bounds self-recursive inlining with
-        // its own counter (`pyre-jit-trace` `jitcode_dispatch::inline_call`,
-        // against `FBW_MAX_INLINE_RECURSION`) and applies no side effect
-        // there, so do not read this comment as "the tracer does it too".
+        // `max_unroll_recursion`.  The production FBW walker mirrors both the
+        // counter and this side effect in
+        // `pyre-jit-trace/jitcode_dispatch/inline_call.rs`; keeping the state
+        // transition on both tracing paths prevents sibling branches from
+        // restarting recursive unrolling after the first bound hit.
         if should_disable {
             self.warm_state.disable_noninlinable_function(callee_key);
         }

pyre/bench/fib_recursive.cranelift.jitstats

@@ -1,4 +1,4 @@
-bridges_compiled=3
+bridges_compiled=8
 descr_set_absent=0
 descr_set_ambiguous=0
 descr_set_stale_absent=0
@@ -8,7 +8,7 @@ fbw_rolled_back_with_effects=0
 fbw_store_journal_rollback_failed=0
 field_pos_attached_misplaced=0
 field_pos_spec_misplaced=0
-guard_failures=406
+guard_failures=1645
 internal_compile_panics=0
 loops_aborted=0
 loops_compiled=1

pyre/bench/fib_recursive.dynasm.jitstats

@@ -1,4 +1,4 @@
-bridges_compiled=3
+bridges_compiled=8
 descr_set_absent=0
 descr_set_ambiguous=0
 descr_set_stale_absent=0
@@ -8,7 +8,7 @@ fbw_rolled_back_with_effects=0
 fbw_store_journal_rollback_failed=0
 field_pos_attached_misplaced=0
 field_pos_spec_misplaced=0
-guard_failures=406
+guard_failures=1645
 internal_compile_panics=0
 loops_aborted=0
 loops_compiled=1

pyre/bench/fib_recursive.wasm.jitstats

@@ -1,4 +1,4 @@
-bridges_compiled=3
+bridges_compiled=8
 descr_set_absent=0
 descr_set_ambiguous=0
 descr_set_stale_absent=0
@@ -8,7 +8,8 @@ fbw_rolled_back_with_effects=0
 fbw_store_journal_rollback_failed=0
 field_pos_attached_misplaced=0
 field_pos_spec_misplaced=0
-guard_failures=406
+guard_failures=1645
 internal_compile_panics=0
 loops_aborted=0
 loops_compiled=1
+retraces_compiled=0

pyre/bench/synth/bound_method_builtin_fold.wasm.jitstats

@@ -8,7 +8,8 @@ fbw_rolled_back_with_effects=0
 fbw_store_journal_rollback_failed=0
 field_pos_attached_misplaced=0
 field_pos_spec_misplaced=0
-guard_failures=459
+guard_failures=458
 internal_compile_panics=0
 loops_aborted=0
 loops_compiled=7
+retraces_compiled=0

pyre/bench/synth/bridge_recursion_overflow.cranelift.jitstats

@@ -1,4 +1,4 @@
-bridges_compiled=3
+bridges_compiled=4
 descr_set_absent=0
 descr_set_ambiguous=0
 descr_set_stale_absent=0
@@ -8,7 +8,7 @@ fbw_rolled_back_with_effects=0
 fbw_store_journal_rollback_failed=0
 field_pos_attached_misplaced=0
 field_pos_spec_misplaced=0
-guard_failures=798
+guard_failures=962
 internal_compile_panics=0
 loops_aborted=0
 loops_compiled=1

pyre/bench/synth/bridge_recursion_overflow.dynasm.jitstats

@@ -1,4 +1,4 @@
-bridges_compiled=3
+bridges_compiled=4
 descr_set_absent=0
 descr_set_ambiguous=0
 descr_set_stale_absent=0
@@ -8,7 +8,7 @@ fbw_rolled_back_with_effects=0
 fbw_store_journal_rollback_failed=0
 field_pos_attached_misplaced=0
 field_pos_spec_misplaced=0
-guard_failures=798
+guard_failures=962
 internal_compile_panics=0
 loops_aborted=0
 loops_compiled=1

pyre/bench/synth/bridge_recursion_overflow.wasm.jitstats

@@ -1,4 +1,4 @@
-bridges_compiled=3
+bridges_compiled=4
 descr_set_absent=0
 descr_set_ambiguous=0
 descr_set_stale_absent=0
@@ -8,7 +8,7 @@ fbw_rolled_back_with_effects=0
 fbw_store_journal_rollback_failed=0
 field_pos_attached_misplaced=0
 field_pos_spec_misplaced=0
-guard_failures=798
+guard_failures=962
 internal_compile_panics=0
 loops_aborted=0
 loops_compiled=1

pyre/bench/synth/ca_bridge_multiframe_resume_double_call.cranelift.jitstats

@@ -1,4 +1,4 @@
-bridges_compiled=16
+bridges_compiled=24
 descr_set_absent=0
 descr_set_ambiguous=0
 descr_set_stale_absent=0
@@ -8,7 +8,7 @@ fbw_rolled_back_with_effects=0
 fbw_store_journal_rollback_failed=0
 field_pos_attached_misplaced=0
 field_pos_spec_misplaced=0
-guard_failures=2592
+guard_failures=3437
 internal_compile_panics=0
 loops_aborted=0
 loops_compiled=3