[compiler] Clear stale data for ZeroExtendsWord32ToWord64

sethbrenith · V8 LUCI CQ · commit 780d5608bb8a · 2024-08-07T15:57:04.000Z
The first call to ZeroExtendsWord32ToWord64 produces a correct result, but leaves some incorrect values in phi_states_. To avoid incorrect behavior, we should clear those values when starting anew. I think that the performance impact of this change on compilation time should be small, because calls to ZeroExtendsWord32ToWord64 are infrequent. Here is a histogram showing, per function compiled in Octane, how often this new code is run: 0: 74.7% 1: 13.1% 2: 6.3% 3: 2.5% 4 or 5: 1.7% 6 to 9: 0.9% 11 to 33: 0.8% Bug: 356196918 Change-Id: I00a9e74652025bf8a32cb083a6e01c0273e44043 Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/5766478 Commit-Queue: Seth Brenith <seth.brenith@microsoft.com> Reviewed-by: Nico Hartmann <nicohartmann@chromium.org> Cr-Commit-Position: refs/heads/main@{#95528}
diff --git a/src/compiler/backend/instruction-selector.cc b/src/compiler/backend/instruction-selector.cc
@@ -5734,6 +5734,14 @@ bool InstructionSelectorT<Adapter>::ZeroExtendsWord32ToWord64(
   const int kMaxRecursionDepth = 100;
 
   if (this->IsPhi(node)) {
+    // Intermediate results from previous calls are not necessarily correct.
+    if (recursion_depth == 0) {
+      static_assert(sizeof(Upper32BitsState) == 1);
+      memset(phi_states_.data(),
+             static_cast<int>(Upper32BitsState::kNotYetChecked),
+             phi_states_.size());
+    }
+
     Upper32BitsState current = phi_states_[this->id(node)];
     if (current != Upper32BitsState::kNotYetChecked) {
       return current == Upper32BitsState::kUpperBitsGuaranteedZero;
