Share the preamble in switch interpreter.
Move the preamble code to the main loop to avoid replicating it.
This halves per-file compile time to 0.5 min.
It saves 225k(arm32) and 348k(arm64) from libart.so.
Performance difference is within noise (golem using the switch).
Test: "test.py -b -r --interpreter --host" with switch interpreter.
Change-Id: I2615b3568eb79560d2d342ac3dd28f6269ee5e74
diff --git a/runtime/interpreter/interpreter_switch_impl-inl.h b/runtime/interpreter/interpreter_switch_impl-inl.h
index 4125a0a..bf12a72 100644
--- a/runtime/interpreter/interpreter_switch_impl-inl.h
+++ b/runtime/interpreter/interpreter_switch_impl-inl.h
@@ -160,7 +160,7 @@
}
// Code to run before each dex instruction.
- ALWAYS_INLINE WARN_UNUSED bool PreambleSave(JValue* save_ref)
+ ALWAYS_INLINE WARN_UNUSED bool Preamble()
REQUIRES_SHARED(Locks::mutator_lock_) {
/* We need to put this before & after the instrumentation to avoid having to put in a */
/* post-script macro. */
@@ -168,6 +168,9 @@
return false;
}
if (UNLIKELY(instrumentation->HasDexPcListeners())) {
+ uint8_t opcode = inst->Opcode(inst_data);
+ bool is_move_result_object = (opcode == Instruction::MOVE_RESULT_OBJECT);
+ JValue* save_ref = is_move_result_object ? &ctx->result_register : nullptr;
if (UNLIKELY(!DoDexPcMoveEvent(self,
Accessor(),
shadow_frame,
@@ -2587,16 +2590,27 @@
shadow_frame.SetDexPC(dex_pc);
TraceExecution(shadow_frame, inst, dex_pc);
inst_data = inst->Fetch16(0);
+ {
+ bool exit_loop = false;
+ InstructionHandler<do_access_check, transaction_active> handler(
+ ctx, instrumentation, self, shadow_frame, dex_pc, inst, inst_data, exit_loop);
+ if (!handler.Preamble()) {
+ if (UNLIKELY(exit_loop)) {
+ return;
+ }
+ if (UNLIKELY(interpret_one_instruction)) {
+ break;
+ }
+ continue;
+ }
+ }
switch (inst->Opcode(inst_data)) {
#define OPCODE_CASE(OPCODE, OPCODE_NAME, pname, f, i, a, e, v) \
case OPCODE: { \
bool exit_loop = false; \
InstructionHandler<do_access_check, transaction_active> handler( \
ctx, instrumentation, self, shadow_frame, dex_pc, inst, inst_data, exit_loop); \
- constexpr bool is_move_result_object = (OPCODE == Instruction::MOVE_RESULT_OBJECT); \
- if (handler.PreambleSave(is_move_result_object ? &ctx->result_register : nullptr)) { \
- handler.OPCODE_NAME(); \
- } \
+ handler.OPCODE_NAME(); \
/* TODO: Advance 'inst' here, instead of explicitly in each handler */ \
if (UNLIKELY(exit_loop)) { \
return; \
@@ -2607,12 +2621,13 @@
#undef OPCODE_CASE
}
if (UNLIKELY(interpret_one_instruction)) {
- // Record where we stopped.
- shadow_frame.SetDexPC(inst->GetDexPc(insns));
- ctx->result = ctx->result_register;
- return;
+ break;
}
}
+ // Record where we stopped.
+ shadow_frame.SetDexPC(inst->GetDexPc(insns));
+ ctx->result = ctx->result_register;
+ return;
} // NOLINT(readability/fn_size)
} // namespace interpreter
