8331341: secondary_super_cache does not scale well: C1 and interpreter

src/hotspot/cpu/aarch64/aarch64.ad

@@ -16044,9 +16044,10 @@ instruct ShouldNotReachHere() %{
 instruct partialSubtypeCheck(iRegP_R4 sub, iRegP_R0 super, iRegP_R2 temp, iRegP_R5 result, rFlagsReg cr)
 %{
  match(Set result (PartialSubtypeCheck sub super));
+ predicate(!UseSecondarySupersTable);
  effect(KILL cr, KILL temp);
 
- ins_cost(1100); // slightly larger than the next version
+ ins_cost(20 * INSN_COST); // slightly larger than the next version
  format %{ "partialSubtypeCheck $result, $sub, $super" %}
 
  ins_encode(aarch64_enc_partial_subtype_check(sub, super, temp, result));
@@ -16056,6 +16057,35 @@ instruct partialSubtypeCheck(iRegP_R4 sub, iRegP_R0 super, iRegP_R2 temp, iRegP_
  ins_pipe(pipe_class_memory);
 %}
 
+// Two versions of partialSubtypeCheck, both used when we need to
+// search for a super class in the secondary supers array. The first
+// is used when we don't know _a priori_ the class being searched
+// for. The second, far more common, is used when we do know: this is
+// used for instanceof, checkcast, and any case where C2 can determine
+// it by constant propagation.
+
+instruct partialSubtypeCheckVarSuper(iRegP_R4 sub, iRegP_R0 super, vRegD_V0 vtemp, iRegP_R5 result,
+ iRegP_R1 tempR1, iRegP_R2 tempR2, iRegP_R3 tempR3,
+ rFlagsReg cr)
+%{
+ match(Set result (PartialSubtypeCheck sub super));
+ predicate(UseSecondarySupersTable);
+ effect(KILL cr, TEMP tempR1, TEMP tempR2, TEMP tempR3, TEMP vtemp);
+
+ ins_cost(10 * INSN_COST); // slightly larger than the next version
+ format %{ "partialSubtypeCheck $result, $sub, $super" %}
+
+ ins_encode %{
+ bool success = false;
+ __ lookup_secondary_supers_table_var($sub$$Register, $super$$Register,
+ $tempR1$$Register, $tempR2$$Register, $tempR3$$Register,
+ $vtemp$$FloatRegister,
+ $result$$Register, /*L_success*/nullptr);
+ %}
+
+ ins_pipe(pipe_class_memory);
+%}
+
 instruct partialSubtypeCheckConstSuper(iRegP_R4 sub, iRegP_R0 super_reg, immP super_con, vRegD_V0 vtemp, iRegP_R5 result,
  iRegP_R1 tempR1, iRegP_R2 tempR2, iRegP_R3 tempR3,
  rFlagsReg cr)
@@ -16064,18 +16094,19 @@ instruct partialSubtypeCheckConstSuper(iRegP_R4 sub, iRegP_R0 super_reg, immP su
  predicate(UseSecondarySupersTable);
  effect(KILL cr, TEMP tempR1, TEMP tempR2, TEMP tempR3, TEMP vtemp);
 
- ins_cost(700); // smaller than the next version
+ ins_cost(5 * INSN_COST); // smaller than the next version
  format %{ "partialSubtypeCheck $result, $sub, $super_reg, $super_con" %}
 
  ins_encode %{
  bool success = false;
  u1 super_klass_slot = ((Klass*)$super_con$$constant)->hash_slot();
  if (InlineSecondarySupersTest) {
- success = __ lookup_secondary_supers_table($sub$$Register, $super_reg$$Register,
- $tempR1$$Register, $tempR2$$Register, $tempR3$$Register,
- $vtemp$$FloatRegister,
- $result$$Register,
- super_klass_slot);
+ success =
+ __ lookup_secondary_supers_table_const($sub$$Register, $super_reg$$Register,
+ $tempR1$$Register, $tempR2$$Register, $tempR3$$Register,
+ $vtemp$$FloatRegister,
+ $result$$Register,
+ super_klass_slot);
  } else {
  address call = __ trampoline_call(RuntimeAddress(StubRoutines::lookup_secondary_supers_table_stub(super_klass_slot)));
  success = (call != nullptr);
@@ -16089,21 +16120,6 @@ instruct partialSubtypeCheckConstSuper(iRegP_R4 sub, iRegP_R0 super_reg, immP su
  ins_pipe(pipe_class_memory);
 %}
 
-instruct partialSubtypeCheckVsZero(iRegP_R4 sub, iRegP_R0 super, iRegP_R2 temp, iRegP_R5 result, immP0 zero, rFlagsReg cr)
-%{
- match(Set cr (CmpP (PartialSubtypeCheck sub super) zero));
- effect(KILL temp, KILL result);
-
- ins_cost(1100); // slightly larger than the next version
- format %{ "partialSubtypeCheck $result, $sub, $super == 0" %}
-
- ins_encode(aarch64_enc_partial_subtype_check(sub, super, temp, result));
-
- opcode(0x0); // Don't zero result reg on hit
-
- ins_pipe(pipe_class_memory);
-%}
-
 // Intrisics for String.compareTo()
 
 instruct string_compareU(iRegP_R1 str1, iRegI_R2 cnt1, iRegP_R3 str2, iRegI_R4 cnt2,

src/hotspot/cpu/aarch64/c1_Runtime1_aarch64.cpp

@@ -842,7 +842,13 @@ OopMapSet* Runtime1::generate_code_for(StubID id, StubAssembler* sasm) {
  __ ldp(r4, r0, Address(sp, (sup_k_off) * VMRegImpl::stack_slot_size));
 
  Label miss;
- __ check_klass_subtype_slow_path(r4, r0, r2, r5, nullptr, &miss);
+ __ check_klass_subtype_slow_path(/*sub_klass*/r4,
+ /*super_klass*/r0,
+ /*temp_reg*/r2,
+ /*temp2_reg*/r5,
+ /*L_success*/nullptr,
+ /*L_failure*/&miss);
+ // Need extras for table lookup: r1, r3, vtemp
 
  // fallthrough on success:
  __ mov(rscratch1, 1);