[regalloc] Enable spill placer for non loop-top phis

thibaudmichaud · V8 LUCI CQ · commit 51d206741627 · 2026-01-22T06:50:39.000-08:00
The spill placer is only enabled for loop-top phis because we did not see benefits for other kinds of live ranges at the time. Other ranges are guarded by "--stress-turbo-late-spilling". However, we just observed that the JetStream dotnet-interp-wasm benchmark contains an expensive register spill at the start of each interpreter loop iteration for a non-phi range, and greatly benefits from moving the spill down into the (rare) successor blocks that actually require the value to be on the stack. Enable the flag by default and rename it to "--turbo-always-optimize-spills". This is expected to slightly increase code size in some cases, but this should never increase the number of spills on any control-flow path. R=dlehmann@chromium.org Bug: 475502210 Change-Id: I91b31c632da7ec192d3a41a61a79d561b13fcf18 Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/7510687 Reviewed-by: Daniel Lehmann <dlehmann@chromium.org> Commit-Queue: Thibaud Michaud <thibaudm@chromium.org> Cr-Commit-Position: refs/heads/main@{#104853}
diff --git a/BUILD.gn b/BUILD.gn
@@ -2707,7 +2707,7 @@ template("run_mksnapshot") {
                                         "root_out_dir") + "/mksnapshot",
                          root_build_dir),
       "--turbo_instruction_scheduling",
-      "--stress-turbo-late-spilling",
+      "--turbo-always-optimize-spills",
 
       # In cross builds, the snapshot may be generated for both the host and
       # target toolchains.  The same host binary is used to generate both, so
diff --git a/src/compiler/backend/spill-placer.cc b/src/compiler/backend/spill-placer.cc
@@ -38,13 +38,13 @@ void SpillPlacer::Add(TopLevelLiveRange* range) {
   // - If the value is defined in a deferred block, then the logic to select
   //   the earliest deferred block as the insertion point would cause
   //   incorrect behavior, so the value must be spilled at the definition.
-  // - We haven't seen any indication of performance improvements from seeking
-  //   optimal spilling positions except on loop-top phi values, so spill
-  //   any value that isn't a loop-top phi at the definition to avoid
-  //   increasing the code size for no benefit.
+  // - Late spilling was only enabled for loop phis initially. However, we also
+  // saw benefits in some wasm benchmarks when enabling it for other ranges, so
+  // the "--turbo-always-optimize-spills" flag is enabled by default now. It can
+  // be cleaned up if this is stable and does not cause regressions.
   if (range->GetSpillMoveInsertionLocations(data()) == nullptr ||
       range->spilled() || top_start_block->IsDeferred() ||
-      (!v8_flags.stress_turbo_late_spilling && !range->is_loop_phi())) {
+      (!v8_flags.turbo_always_optimize_spills && !range->is_loop_phi())) {
     range->CommitSpillMoves(data(), spill_operand);
     return;
   }
diff --git a/src/flags/flag-definitions.h b/src/flags/flag-definitions.h
@@ -1367,7 +1367,7 @@ DEFINE_BOOL_READONLY(opt, false,
 #endif  // V8_ENABLE_TURBOFAN
 
 DEFINE_BOOL(
-    stress_turbo_late_spilling, false,
+    turbo_always_optimize_spills, true,
     "optimize placement of all spill instructions, not just loop-top phis")
 
 DEFINE_BOOL(turbo_wasm_address_reassociation, true,
