otimizações round 12

otimizações round 12

include/psprecomp/runtime.hpp

@@ -311,12 +311,12 @@ public:
         return run_starvation_boundary(ctx);
     }
 
-    // Tier-2 hot-leaf lowering keeps the scheduler accounting that an ordinary
-    // cross-unit generated call would have performed, while allowing trivial
-    // leaf accessors to be emitted directly in their measured caller.  No HLE
-    // or PSP ownership switch can occur inside those leaf bodies, so only the
-    // starvation safe-point cadence needs to be preserved here.
-    [[nodiscard]] PSPRECOMP_RUNTIME_FORCEINLINE bool account_inlined_generated_leaf(
+    // Account one logical outer generated dispatch that was deliberately
+    // removed by a profile-guided superblock.  Publishing the destination PC
+    // before calling this helper is mandatory: a starvation boundary may switch
+    // PSP ownership at exactly the same point where the unfused outer dispatcher
+    // used to run.
+    [[nodiscard]] PSPRECOMP_RUNTIME_FORCEINLINE bool account_inlined_dispatch_boundary(
         AllegrexContext &ctx) {
         const std::uint64_t starvation_interval = g_runtime_starvation_interval_fast;
         if (starvation_interval == 0u) return true;
@@ -324,6 +324,14 @@ public:
         return run_starvation_boundary(ctx);
     }
 
+    // Tier-2 hot-leaf lowering keeps the scheduler accounting that an ordinary
+    // cross-unit generated call would have performed, while allowing trivial
+    // leaf accessors to be emitted directly in their measured caller.
+    [[nodiscard]] PSPRECOMP_RUNTIME_FORCEINLINE bool account_inlined_generated_leaf(
+        AllegrexContext &ctx) {
+        return account_inlined_dispatch_boundary(ctx);
+    }
+
     // Tier-2 profile-guided superblocks can fuse a cross-unit edge into a local
     // C++ goto.  The guest-visible control flow is unchanged, but the ordinary
     // invoke_chained_direct() native frame no longer exists.  These helpers keep

profiles/vcs/CMakeLists.txt

@@ -167,6 +167,7 @@ set(VCS_TIER2_CLUSTER_SOURCES
     host/vcs_tier2_cluster_physics.cpp
     host/vcs_tier2_cluster_world.cpp
     host/vcs_tier2_cluster_edge43.cpp
+    host/vcs_tier2_cluster_collisionloop.cpp
 )
 if(MSVC)
     set_source_files_properties(

profiles/vcs/generated/generated_unit_0037.cpp

@@ -1,5 +1,6 @@
 #include "psprecomp/runtime.hpp"
 #include "generated_units.hpp"
+#include "vcs_tier2_superblocks.hpp"
 #include <bit>
 #include <cmath>
 #include <cstdint>
@@ -8398,6 +8399,12 @@ L_0889BF84:
     ctx.fpr[24] = std::bit_cast<float>(ctx.gpr[4]);
     goto L_0889BFC0;
 L_0889BFC0:
+// TIER2_SUPERBLOCK_V2_HOOK_BEGIN
+    if (vcs::tier2_cluster_enabled(vcs::Tier2ClusterId::CollisionLoop) && rt.memory().direct_fastmem_enabled()) {
+        vcs::tier2_superblock_collisionloop(rt, ctx, aot_mem, 0x0889BFC0u);
+        return;
+    }
+// TIER2_SUPERBLOCK_V2_HOOK_END
     ctx.gpr[18] = (aot_mem.aot_load32(ctx.gpr[30] + static_cast<std::uint32_t>(0)));
     ctx.gpr[30] = (aot_mem.aot_load32(ctx.gpr[30] + static_cast<std::uint32_t>(8)));
     ctx.gpr[4] = (aot_mem.aot_load16(ctx.gpr[18] + static_cast<std::uint32_t>(84)));

profiles/vcs/generated/generated_unit_0038.cpp

@@ -1,5 +1,6 @@
 #include "psprecomp/runtime.hpp"
 #include "generated_units.hpp"
+#include "vcs_tier2_superblocks.hpp"
 #include <bit>
 #include <cmath>
 #include <cstdint>
@@ -1025,6 +1026,12 @@ LOCAL_DISPATCH:
     }
     }
 L_0889C000:
+// TIER2_SUPERBLOCK_V2_HOOK_BEGIN
+    if (vcs::tier2_cluster_enabled(vcs::Tier2ClusterId::CollisionLoop) && rt.memory().direct_fastmem_enabled()) {
+        vcs::tier2_superblock_collisionloop(rt, ctx, aot_mem, 0x0889C000u);
+        return;
+    }
+// TIER2_SUPERBLOCK_V2_HOOK_END
     { const bool branch_taken = ctx.gpr[4] == 0u;
     // nop
       if (branch_taken) {
@@ -1912,6 +1919,12 @@ L_0889C5B0:
     aot_mem.aot_store32(ctx.gpr[4] + static_cast<std::uint32_t>(0), ctx.gpr[18]);
     goto L_0889C5B8;
 L_0889C5B8:
+// TIER2_SUPERBLOCK_V2_HOOK_BEGIN
+    if (vcs::tier2_cluster_enabled(vcs::Tier2ClusterId::CollisionLoop) && rt.memory().direct_fastmem_enabled()) {
+        vcs::tier2_superblock_collisionloop(rt, ctx, aot_mem, 0x0889C5B8u);
+        return;
+    }
+// TIER2_SUPERBLOCK_V2_HOOK_END
     { const bool branch_taken = ctx.gpr[30] != 0u;
     // nop
       if (branch_taken) {

profiles/vcs/generated/generated_unit_0043.cpp

@@ -9533,6 +9533,12 @@ L_088B3FCC:
     ctx.gpr[4] = (ctx.gpr[20] | 0u);
     goto L_088B1780;
 L_088B3FFC:
+// TIER2_SUPERBLOCK_V2_HOOK_BEGIN
+    if (vcs::tier2_cluster_enabled(vcs::Tier2ClusterId::Edge43) && rt.memory().direct_fastmem_enabled()) {
+        vcs::tier2_superblock_edge43(rt, ctx, aot_mem, 0x088B3FFCu);
+        return;
+    }
+// TIER2_SUPERBLOCK_V2_HOOK_END
     if (ctx.gpr[2] == 0u) {
     ctx.gpr[4] = (aot_mem.aot_load32(ctx.gpr[29] + static_cast<std::uint32_t>(48)));
         (void)rt.invoke_chained_direct<&recomp_unit_0044_entry, 44u, 9u, 0x088B40F8u>(ctx, &aot_mem); return;

profiles/vcs/host/vcs_runtime_log.cpp

@@ -65,7 +65,7 @@ void runtime_log_initialize(const VcsConfiguration &configuration) {
         return;
     }
     s.file << "VCSNative runtime log\n";
-    s.file << "stage=correctness-v8.2.7a-internal-save-repro-gate-2026-08-18\n";
+    s.file << "stage=perf-v8.3-cpu-boundary-fusion-2026-08-18\n";
     s.file << "config=" << configuration.source_path.string() << '\n';
     s.file << "started=" << timestamp_now() << '\n';
     s.file << "perf_telemetry=" << (configuration.diagnostics.perf_telemetry ? 1 : 0)
@@ -81,16 +81,16 @@ void runtime_log_initialize(const VcsConfiguration &configuration) {
            << " chain_telemetry_compiled=0\n";
 #endif
     s.file << "tier2_superblocks=" << (tier2_superblocks_enabled() ? 1 : 0)
-           << " version=4 clusters=7 mask=0x" << std::hex << tier2_cluster_mask() << std::dec
-           << " hot_blocks=1060 static_fused_calls=36 static_fused_tail=13 hooks=25"
+           << " version=4 clusters=8 mask=0x" << std::hex << tier2_cluster_mask() << std::dec
+           << " hot_blocks=1139 static_fused_calls=36 static_fused_tail=18 static_pc_fused=1 hooks=29"
            << " entity_leaf_inline_sites=13 geometry_inline_leaf_sites=0 direct_generated_leaf_sites=5"
-           << " unwind_fix=1 reentry_guard=1 dataflow=1 vfpu_block32=133 mem_runs=35 mem_words=287"
-           << " append32=51 advance32=89 simd_mat4=4 simd_matvec=19"
-           << " gpr_shadow_clusters=4 gpr_shadow_regs=24 gpr_shadow_occurrences=3461 geometry_shadow=0"
-           << " perf_layer=8 cpu_lean_revision=2 correctness_revision=8271 recovery_from_821=1"
+           << " unwind_fix=1 reentry_guard=1 dataflow=1 vfpu_block32=153 mem_runs=39 mem_words=299"
+           << " append32=51 advance32=89 simd_mat4=4 simd_matvec=21"
+           << " gpr_shadow_clusters=5 gpr_shadow_regs=30 gpr_shadow_occurrences=3841 geometry_shadow=0"
+           << " perf_layer=9 cpu_lean_revision=3 correctness_revision=8271 recovery_from_821=1 cpu_boundary_fusion=1 edge43_continuation_fused=1 collisionloop_blocks=78 collisionloop_static_pc_fused=1 collisionloop_scheduler_accounting=1"
            << " save_transaction=1 save_lifecycle_v82=1 sas_endflag_latched=1 sas_loop_history_restore=1"
            << " atrac_virtual_source=1 atrac_stall_diag=1 atrac_stream_resident_status=1 atrac_nonloop_resident=-2 atrac_loop_resident=-3 output2_success_zero=1 output2_master_watermark=1 output2_late_catchup=0 save_repro_checkpoint=1 save_repro_trace=1 save_repro_hotkey_f8=1 save_repro_trace_hotkey_f10=1 save_repro_auto_restore=1 save_repro_dispatch_sample_stride=64 save_repro_passive_until_f8=1 save_repro_hle_hotpath=0 save_repro_ui_hotkeys=1 save_repro_f10_async_fallback=1 save_repro_collector_partial_bundle=1 save_repro_internal_ini_gate=1 save_repro_default_enabled=0 save_exitdelete_semantics=1 save_repro_legacy_exitdelete_repair=1 save_partition_reuse=1 news_atrac_v825_guard=1 runtime_chain_telemetry_default=0 arch_fastmem=1 aot_direct_fastmem_default=1"
-           << " tier2_direct_fastmem=1 tier2_direct_mem_sites=1969 tier2_deep_telemetry_default=0 geometry_fusion_rollback=1"
+           << " tier2_direct_fastmem=1 tier2_direct_mem_sites=2115 tier2_deep_telemetry_default=0 geometry_fusion_rollback=1"
            << " entity_leaf_inline=1 entity_leaf_scheduler_accounting=1 entity_leaf_resume_pc_fix=1"
            << " ge_async_default=0 parallel_vertex_decode_default=0"
            << " v5_vfpu_fast_quarantined=1 native_vfpu_088b1780_v4=1"

profiles/vcs/host/vcs_tier2_cluster_edge43.cpp

@@ -67,6 +67,7 @@ TIER2_FUSED_RETURN_DISPATCH:
     case 0x088B1C2Cu: goto SB_L_088B1C2C;
     case 0x088B1C40u: goto SB_L_088B1C40;
     case 0x088B3FCCu: goto SB_L_088B3FCC;
+    case 0x088B3FFCu: goto SB_L_088B3FFC;
     case 0x088B4004u: goto SB_L_088B4004;
     case 0x088B4018u: goto SB_L_088B4018;
     case 0x088B4020u: goto SB_L_088B4020;
@@ -142,6 +143,7 @@ TIER2_LOCAL_DISPATCH_U0043:
     case 0x088B1C2Cu: goto SB_L_088B1C2C;
     case 0x088B1C40u: goto SB_L_088B1C40;
     case 0x088B3FCCu: goto SB_L_088B3FCC;
+    case 0x088B3FFCu: goto SB_L_088B3FFC;
     default:
         ctx.pc = local_pc;
         TIER2_SB_RETURN();
@@ -209,6 +211,7 @@ TIER2_LOCAL_DISPATCH_U0044:
 TIER2_ENTRY_DISPATCH:
     switch (entry_pc) {
     case 0x088B3FCCu: goto SB_L_088B3FCC;
+    case 0x088B3FFCu: goto SB_L_088B3FFC;
     case 0x088B4004u: goto SB_L_088B4004;
     case 0x088B40F8u: goto SB_L_088B40F8;
     default:
@@ -584,6 +587,24 @@ SB_L_088B3FCC:
     tier2_gpr_4 = (ctx.gpr[20] | 0u);
     goto SB_L_088B1780;
 
+SB_L_088B3FFC:
+    if (ctx.gpr[2] == 0u) {
+    tier2_gpr_4 = (tier2_mem.aot_load32(tier2_gpr_29 + static_cast<std::uint32_t>(48)));
+        if (!rt.tier2_enter_fused_transfer<44u, 0x088B40F8u>(ctx)) {
+            ctx.pc = 0x088B40F8u;
+            TIER2_SB_RETURN();
+        }
+        ++tier2_pending_transfers;
+        ++tier2_stats.fused_tail_edges;
+        goto SB_L_088B40F8;
+    }
+    if (!rt.tier2_enter_fused_transfer<44u, 0x088B4004u>(ctx)) {
+        ctx.pc = 0x088B4004u;
+        TIER2_SB_RETURN();
+    }
+    ++tier2_pending_transfers;
+    ++tier2_stats.fused_tail_edges;
+    goto SB_L_088B4004;
 SB_L_088B4004:
     tier2_gpr_17 = (static_cast<std::uint32_t>(static_cast<std::int32_t>(static_cast<std::int16_t>(tier2_mem.aot_load16(ctx.gpr[18] + static_cast<std::uint32_t>(6))))));
     tier2_gpr_4 = (static_cast<std::uint32_t>(static_cast<std::int32_t>(static_cast<std::int16_t>(tier2_mem.aot_load16(ctx.gpr[18] + static_cast<std::uint32_t>(14))))));

profiles/vcs/host/vcs_tier2_superblocks.cpp

@@ -63,6 +63,7 @@ const char *tier2_cluster_name(Tier2ClusterId id) noexcept {
     case Tier2ClusterId::Physics: return "physics";
     case Tier2ClusterId::World: return "world";
     case Tier2ClusterId::Edge43: return "edge43";
+    case Tier2ClusterId::CollisionLoop: return "collisionloop";
     default: return "unknown";
     }
 }

profiles/vcs/host/vcs_tier2_superblocks.hpp

@@ -22,7 +22,8 @@ enum class Tier2ClusterId : std::uint32_t {
     Physics = 4u,     // 0129 hot transform/physics function
     World = 5u,       // 0157/0158 dominant world/streaming hot functions
     Edge43 = 6u,      // 0043/0044 cross-unit boundary
-    Count = 7u,
+    CollisionLoop = 7u, // 0037/0038 measured generated-boundary loop
+    Count = 8u,
 };
 
 constexpr std::size_t kTier2ClusterCount = static_cast<std::size_t>(Tier2ClusterId::Count);
@@ -135,5 +136,7 @@ void tier2_superblock_world(psprecomp::Runtime &, psprecomp::AllegrexContext &,
                             psprecomp::GuestMemory::AotFastView &, std::uint32_t entry_pc);
 void tier2_superblock_edge43(psprecomp::Runtime &, psprecomp::AllegrexContext &,
                              psprecomp::GuestMemory::AotFastView &, std::uint32_t entry_pc);
+void tier2_superblock_collisionloop(psprecomp::Runtime &, psprecomp::AllegrexContext &,
+                                    psprecomp::GuestMemory::AotFastView &, std::uint32_t entry_pc);
 
 } // namespace vcs

profiles/vcs/tests/check_v827_save_thread_lifecycle.py

@@ -18,16 +18,16 @@ def need(cond, msg):
         raise SystemExit(1)
     print('PASS:', msg)
 
-need(('stage=correctness-v8.2.7-save-thread-lifecycle-fix-2026-08-18' in log) or ('stage=correctness-v8.2.7a-internal-save-repro-gate-2026-08-18' in log), 'V8.2.7/V8.2.7A runtime stage')
-need(('correctness_revision=827 ' in log) or ('correctness_revision=8271 ' in log), 'V8.2.7/V8.2.7A correctness revision')
+need(any(stage in log for stage in ('stage=correctness-v8.2.7-save-thread-lifecycle-fix-2026-08-18', 'stage=correctness-v8.2.7a-internal-save-repro-gate-2026-08-18', 'stage=perf-v8.3-cpu-boundary-fusion-2026-08-18')), 'V8.2.7 lifecycle fix preserved in current runtime stage')
+need(('correctness_revision=827 ' in log) or ('correctness_revision=8271 ' in log), 'V8.2.7/V8.2.7A correctness revision preserved')
 need('save_exitdelete_semantics=1' in log and 'save_repro_legacy_exitdelete_repair=1' in log and 'save_partition_reuse=1' in log,
      'ExitDelete/partition fix and migration metadata')
 need('atrac_stream_resident_status=1' in log and 'atrac_nonloop_resident=-2' in log and
      'atrac_loop_resident=-3' in log and 'news_atrac_v825_guard=1' in log,
      'working V8.2.5 NEWS semantics preserved')
 need('output2_late_catchup=0' in log, 'rejected Output2 pacing experiment remains disabled')
-need('hot_blocks=1060' in log and 'static_fused_calls=36' in log and 'geometry_fusion_rollback=1' in log,
-     'V8.2 CPU/Tier2 shape preserved')
+need('static_fused_calls=36' in log and 'geometry_fusion_rollback=1' in log and 'geometry_inline_leaf_sites=0' in log,
+     'V8.2 Geometry rollback and conservative fusion shape preserved')
 
 # Core correctness: ExitThread remains dormant/Completed, ExitDelete actually
 # destroys the object and releases its stack.

profiles/vcs/tests/check_v827a_internal_save_repro_gate.py

@@ -19,8 +19,9 @@ def need(cond, msg):
         raise SystemExit(1)
     print('PASS:', msg)
 
-need('stage=correctness-v8.2.7a-internal-save-repro-gate-2026-08-18' in log,
-     'V8.2.7A runtime stage')
+need(any(stage in log for stage in ('stage=correctness-v8.2.7a-internal-save-repro-gate-2026-08-18',
+                                     'stage=perf-v8.3-cpu-boundary-fusion-2026-08-18')),
+     'V8.2.7A gate preserved in current runtime stage')
 need('correctness_revision=8271' in log, 'V8.2.7A correctness revision')
 need('save_repro_internal_ini_gate=1' in log and 'save_repro_default_enabled=0' in log,
      'internal-only gate metadata')
@@ -56,7 +57,7 @@ for token, label in [
     ('atrac_nonloop_resident=-2', 'NEWS non-loop ATRAC sentinel preserved'),
     ('atrac_loop_resident=-3', 'NEWS loop ATRAC sentinel preserved'),
     ('output2_late_catchup=0', 'rejected Output2 pacing remains off'),
-    ('hot_blocks=1060', 'V8.2 CPU shape preserved'),
+    ('geometry_fusion_rollback=1', 'V8.2 Geometry rollback preserved'),
     ('geometry_fusion_rollback=1', 'Geometry rollback preserved'),
 ]:
     need(token in log, label)

profiles/vcs/tests/check_v83_cpu_boundary_fusion.py

@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+from pathlib import Path
+
+root = Path(__file__).resolve().parents[3]
+profile = root / 'profiles' / 'vcs'
+host = profile / 'host'
+tools = profile / 'tools'
+gen = profile / 'generated'
+
+generator = (tools / 'build_tier2_superblocks.py').read_text(encoding='utf-8')
+hpp = (host / 'vcs_tier2_superblocks.hpp').read_text(encoding='utf-8')
+common = (host / 'vcs_tier2_superblocks.cpp').read_text(encoding='utf-8')
+cmake = (profile / 'CMakeLists.txt').read_text(encoding='utf-8')
+log = (host / 'vcs_runtime_log.cpp').read_text(encoding='utf-8')
+collision = (host / 'vcs_tier2_cluster_collisionloop.cpp').read_text(encoding='utf-8')
+edge = (host / 'vcs_tier2_cluster_edge43.cpp').read_text(encoding='utf-8')
+u37 = (gen / 'generated_unit_0037.cpp').read_text(encoding='utf-8')
+u38 = (gen / 'generated_unit_0038.cpp').read_text(encoding='utf-8')
+u43 = (gen / 'generated_unit_0043.cpp').read_text(encoding='utf-8')
+geometry = (host / 'vcs_tier2_cluster_geometry.cpp').read_text(encoding='utf-8')
+world = (host / 'vcs_tier2_cluster_world.cpp').read_text(encoding='utf-8')
+runtime_hpp = (root / 'include' / 'psprecomp' / 'runtime.hpp').read_text(encoding='utf-8')
+
+def need(cond, msg):
+    if not cond:
+        print('FAIL:', msg)
+        raise SystemExit(1)
+    print('PASS:', msg)
+
+need('stage=perf-v8.3-cpu-boundary-fusion-2026-08-18' in log, 'V8.3 runtime stage')
+need('perf_layer=9 cpu_lean_revision=3 correctness_revision=8271' in log,
+     'performance revision advances without changing protected correctness revision')
+need('clusters=8' in log and 'hot_blocks=1139' in log and 'static_fused_calls=36' in log and
+     'static_fused_tail=18' in log and 'static_pc_fused=1' in log,
+     'generated Tier-2 V8.3 shape matches audited generator output')
+need('cpu_boundary_fusion=1' in log and 'edge43_continuation_fused=1' in log and
+     'collisionloop_blocks=78' in log and 'collisionloop_static_pc_fused=1' in log and 'collisionloop_scheduler_accounting=1' in log,
+     'boundary-fusion metadata present')
+
+need('fuse_static_pc_returns: bool = False' in generator and 'STATIC_PC_RETURN_RE' in generator,
+     'static PC return fusion is explicit and opt-in')
+need("key='collisionloop'" in generator and 'fuse_static_pc_returns=True' in generator,
+     'only dedicated collision-loop cluster opts into generated-boundary fusion')
+need(generator.count('fuse_static_pc_returns=True') == 1,
+     'legacy Tier-2 clusters cannot silently change boundary semantics')
+need('37: [(0x0889BFC0, 0x0889C000)]' in generator and
+     '38: [(0x0889C000, 0x0889C5C0)]' in generator,
+     'collision-loop window remains narrowly bounded to units 0037/0038')
+need('seeds={43: [0x088B3FCC, 0x088B3FFC]' in generator,
+     'Edge43 includes measured missing continuation 0x088B3FFC')
+
+need('account_inlined_dispatch_boundary' in runtime_hpp and 'return account_inlined_dispatch_boundary(ctx);' in runtime_hpp,
+     'generic inlined-dispatch accounting preserves starvation cadence and backs leaf accounting')
+
+need('CollisionLoop = 7u' in hpp and 'Count = 8u' in hpp and
+     'tier2_superblock_collisionloop' in hpp,
+     'eighth cluster is registered in Tier-2 ABI')
+need('case Tier2ClusterId::CollisionLoop: return "collisionloop";' in common,
+     'collision-loop cluster has runtime identity')
+need('host/vcs_tier2_cluster_collisionloop.cpp' in cmake,
+     'collision-loop translation unit is compiled')
+
+need('SB_L_0889BFC0:' in collision and 'SB_L_0889C000:' in collision and
+     'account_inlined_dispatch_boundary(ctx)' in collision and 'goto SB_L_0889C000;' in collision,
+     'unit0037 -> unit0038 boundary is local while retaining outer-dispatch scheduler accounting')
+need('goto SB_L_0889BFC0;' in collision,
+     'hot loop closes back into unit0037 without outer Runtime dispatch')
+need('tier2_enter_fused_transfer<37u, 0x0889BFC0u>' in collision and
+     'tier2_complete_shadow(tier2_static_pending_)' in collision,
+     'real tail-call frames unwind before the fused outer boundary, preserving accounting order')
+need('tier2_superblock_collisionloop(rt, ctx, aot_mem, 0x0889BFC0u)' in u37 and
+     'tier2_superblock_collisionloop(rt, ctx, aot_mem, 0x0889C000u)' in u38 and
+     'tier2_superblock_collisionloop(rt, ctx, aot_mem, 0x0889C5B8u)' in u38,
+     'measured loop entry hooks are installed')
+
+need('SB_L_088B3FFC:' in edge and 'case 0x088B3FFCu: goto SB_L_088B3FFC;' in edge,
+     'Edge43 local-return dispatch owns the missing continuation')
+need('tier2_enter_fused_transfer<44u, 0x088B40F8u>' in edge and
+     'tier2_enter_fused_transfer<44u, 0x088B4004u>' in edge,
+     'Edge43 continuation flows directly into selected unit0044 targets')
+need('tier2_superblock_edge43(rt, ctx, aot_mem, 0x088B3FFCu)' in u43,
+     '0x088B3FFC outer-entry hook is installed for measured fallback cases')
+
+# Guard against repeating V8 broad geometry/world expansion.
+need(geometry.count('\nSB_L_') == 379, 'Geometry remains exactly 379 selected blocks')
+need(world.count('\nSB_L_') == 402, 'World remains exactly 402 selected blocks')
+need(len(collision.splitlines()) == 1363, 'new collision-loop body stays small (1363 lines)')
+need(len(edge.splitlines()) == 949, 'Edge43 expansion stays small (949 lines)')
+
+# Current correctness regression guards.
+for token, label in [
+    ('correctness_revision=8271', 'Save lifecycle correctness revision'),
+    ('save_exitdelete_semantics=1', 'ExitDelete lifecycle fix'),
+    ('save_partition_reuse=1', 'partition-memory reuse fix'),
+    ('save_repro_internal_ini_gate=1', 'internal SAVE_REPRO gate'),
+    ('save_repro_default_enabled=0', 'SAVE_REPRO disabled by default'),
+    ('atrac_nonloop_resident=-2', 'NEWS non-loop ATRAC resident sentinel'),
+    ('atrac_loop_resident=-3', 'NEWS loop ATRAC resident sentinel'),
+    ('output2_late_catchup=0', 'rejected Output2 pacing remains off'),
+    ('geometry_fusion_rollback=1', 'rejected V8 Geometry expansion remains rolled back'),
+    ('ge_async_default=0', 'GE async remains quarantined'),
+    ('parallel_vertex_decode_default=0', 'parallel vertex decode remains quarantined'),
+]:
+    need(token in log, label)
+
+print('V8.3 CPU BOUNDARY FUSION audit PASS')

profiles/vcs/tools/build_tier2_superblocks.py

@@ -39,6 +39,11 @@ TAIL_CALL_RE = re.compile(
     r'(?P<indent>[ \t]*)\(void\)rt\.invoke_chained_direct<&recomp_unit_(?P<target_name>\d+)_entry, '
     r'(?P<target_unit>\d+)u, (?P<entry>\d+)u, (?P<target_pc>0x[0-9A-F]+)u>'
     r'\(ctx, &aot_mem\); return;')
+# Artificial generated-unit partition boundaries are not guest calls.
+STATIC_PC_RETURN_RE = re.compile(
+    r'(?P<indent>[ \t]*)ctx\.pc = (?P<target>0x[0-9A-F]+)u; return;')
+
+
 LOCAL_DISPATCH_SEQUENCE_RE = re.compile(
     r'(?P<indent>[ \t]*)local_pc = (?P<local_expr>[^;]+);\n'
     r'(?P=indent)if \(\+\+local_transfers < 256u\) \{ entry_id = 0u; goto LOCAL_DISPATCH; \}\n'
@@ -132,6 +137,10 @@ class Cluster:
     expand_cross_units: bool = False
     hooks: Dict[int, List[int]] = dataclasses.field(default_factory=dict)
     max_blocks: int = 700
+    # Opt-in fusion for plain generated corpus boundaries of the form
+    # `ctx.pc = constant; return;`. Disabled for all legacy clusters.
+    # These boundaries are not guest calls and therefore add no scheduler frame.
+    fuse_static_pc_returns: bool = False
 
 
 CLUSTERS: List[Cluster] = [
@@ -190,10 +199,26 @@ CLUSTERS: List[Cluster] = [
     ),
     Cluster(
         key='edge43', enum_name='Edge43', function='tier2_superblock_edge43',
-        seeds={43: [0x088B3FCC], 44: [0x088B4004, 0x088B40F8]},
+        # Include the measured $ra continuation; omitting 0x088B3FFC forced an
+        # outer Runtime dispatch immediately before unit 0044.
+        seeds={43: [0x088B3FCC, 0x088B3FFC], 44: [0x088B4004, 0x088B40F8]},
         expand_cross_units=True,
-        hooks={43: [0x088B3FCC], 44: [0x088B4004, 0x088B40F8]},
-        max_blocks=100,
+        hooks={43: [0x088B3FCC, 0x088B3FFC], 44: [0x088B4004, 0x088B40F8]},
+        max_blocks=110,
+    ),
+    Cluster(
+        key='collisionloop', enum_name='CollisionLoop', function='tier2_superblock_collisionloop',
+        seeds={},
+        # Small measured loop straddling generated units 0037/0038. Keep the
+        # footprint intentionally bounded to avoid the rejected V8 Geometry bloat.
+        windows={
+            37: [(0x0889BFC0, 0x0889C000)],
+            38: [(0x0889C000, 0x0889C5C0)],
+        },
+        expand_cross_units=True,
+        hooks={37: [0x0889BFC0], 38: [0x0889C000, 0x0889C5B8]},
+        max_blocks=140,
+        fuse_static_pc_returns=True,
     ),
 ]
 
@@ -640,7 +665,8 @@ def selected_pc_owner(selected: Mapping[int, Set[int]]) -> Dict[int, int]:
 
 def transform_block(block: str, source_unit: int, selected: Mapping[int, Set[int]],
                     sources: Mapping[int, UnitSource], stats: MutableMapping[str, int],
-                    fused_continuations: MutableMapping[Tuple[int, int], int]) -> str:
+                    fused_continuations: MutableMapping[Tuple[int, int], int],
+                    fuse_static_pc_returns: bool = False) -> str:
     pc_owner = selected_pc_owner(selected)
     selected_all = set(pc_owner)
     text = re.sub(r'\bL_([0-9A-F]{8})\b', r'SB_L_\1', block)
@@ -754,6 +780,37 @@ def transform_block(block: str, source_unit: int, selected: Mapping[int, Set[int
 
     text = TAIL_CALL_RE.sub(tail_repl, text)
 
+    # Fuse only opted-in plain constant-PC returns whose destination is already
+    # selected in the same superblock.  No tier2_enter/complete call is needed:
+    # the original edge is merely an artificial generated-file boundary.
+    if fuse_static_pc_returns:
+        def static_pc_repl(m: re.Match[str]) -> str:
+            target = int(m.group('target'), 16)
+            if target not in selected_all:
+                return m.group(0)
+            stats['static_pc_fused'] += 1
+            indent = m.group('indent')
+            return (
+                f'{indent}// Generated corpus boundary: unwind any logical tail-call frames first,\n'
+                f'{indent}// then account the outer dispatch that this local edge removes.\n'
+                f'{indent}if (tier2_pending_transfers != 0u) {{\n'
+                f'{indent}    const std::uint32_t tier2_static_pending_ = tier2_pending_transfers;\n'
+                f'{indent}    tier2_pending_transfers = 0u;\n'
+                f'{indent}    if (!tier2_complete_shadow(tier2_static_pending_)) {{\n'
+                f'{indent}        tier2_gpr_shadow_valid = false;\n'
+                f'{indent}        TIER2_SB_RETURN();\n'
+                f'{indent}    }}\n'
+                f'{indent}}}\n'
+                f'{indent}ctx.pc = 0x{target:08X}u;\n'
+                f'{indent}TIER2_GPR_SYNC_OUT();\n'
+                f'{indent}if (!rt.account_inlined_dispatch_boundary(ctx)) {{\n'
+                f'{indent}    tier2_gpr_shadow_valid = false;\n'
+                f'{indent}    TIER2_SB_RETURN();\n'
+                f'{indent}}}\n'
+                f'{indent}TIER2_GPR_SYNC_IN();\n'
+                f'{indent}goto SB_L_{target:08X};')
+        text = STATIC_PC_RETURN_RE.sub(static_pc_repl, text)
+
     # Standard generated indirect/local return path.  The shared unit-specific
     # dispatcher below recognizes a fused-call continuation and performs exactly
     # one logical chain unwind before resuming the caller.
@@ -798,7 +855,8 @@ def emit_cluster(cluster: Cluster, selected: Mapping[int, Set[int]],
         for pc in src.ordered_pcs:
             if pc not in selected[unit]:
                 continue
-            chunks.append(transform_block(src.blocks[pc], unit, selected, sources, stats, continuations))
+            chunks.append(transform_block(src.blocks[pc], unit, selected, sources, stats, continuations,
+                                          cluster.fuse_static_pc_returns))
 
     # Entry hooks are intentionally narrower than the selected closure: only
     # measured roots pay the extra branch into Tier-2.
@@ -998,6 +1056,8 @@ TIER2_ENTRY_DISPATCH:
     stats['lines'] = len(cpp.splitlines())
     stats['units'] = len(selected)
     stats['hooks'] = sum(len(v) for v in cluster.hooks.values())
+    if cluster.fuse_static_pc_returns and stats.get('fused_calls', 0) != 0:
+        raise RuntimeError(f'{cluster.key}: static PC-return fusion currently requires zero fused JAL calls')
     return path, stats