6791178: Specialize for zero as the compressed oop vm heap base

Summary: Use zero based compressed oops if java heap is below 32gb and unscaled compressed oops if java heap is below 4gb.
Reviewed-by: never, twisti, jcoomes, coleenp

5 files changed, 60 insertions, 19 deletions

diff --git a/src/cpu/sparc/vm/assembler_sparc.cpp b/src/cpu/sparc/vm/assembler_sparc.cpp
index d9a7cf076..f31d9765a 100644
--- a/src/cpu/sparc/vm/assembler_sparc.cpp
+++ b/src/cpu/sparc/vm/assembler_sparc.cpp
@@ -4316,7 +4316,13 @@ void MacroAssembler::store_heap_oop(Register d, const Address& a, int offset) {
 
 void MacroAssembler::encode_heap_oop(Register src, Register dst) {
   assert (UseCompressedOops, "must be compressed");
+  assert (Universe::heap() != NULL, "java heap should be initialized");
+  assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
   verify_oop(src);
+  if (Universe::narrow_oop_base() == NULL) {
+    srlx(src, LogMinObjAlignmentInBytes, dst);
+    return;
+  }
   Label done;
   if (src == dst) {
     // optimize for frequent case src == dst
@@ -4338,26 +4344,39 @@ void MacroAssembler::encode_heap_oop(Register src, Register dst) {
 
 void MacroAssembler::encode_heap_oop_not_null(Register r) {
   assert (UseCompressedOops, "must be compressed");
+  assert (Universe::heap() != NULL, "java heap should be initialized");
+  assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
   verify_oop(r);
-  sub(r, G6_heapbase, r);
+  if (Universe::narrow_oop_base() != NULL)
+    sub(r, G6_heapbase, r);
   srlx(r, LogMinObjAlignmentInBytes, r);
 }
 
 void MacroAssembler::encode_heap_oop_not_null(Register src, Register dst) {
   assert (UseCompressedOops, "must be compressed");
+  assert (Universe::heap() != NULL, "java heap should be initialized");
+  assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
   verify_oop(src);
-  sub(src, G6_heapbase, dst);
-  srlx(dst, LogMinObjAlignmentInBytes, dst);
+  if (Universe::narrow_oop_base() == NULL) {
+    srlx(src, LogMinObjAlignmentInBytes, dst);
+  } else {
+    sub(src, G6_heapbase, dst);
+    srlx(dst, LogMinObjAlignmentInBytes, dst);
+  }
 }
 
 // Same algorithm as oops.inline.hpp decode_heap_oop.
 void  MacroAssembler::decode_heap_oop(Register src, Register dst) {
   assert (UseCompressedOops, "must be compressed");
-  Label done;
+  assert (Universe::heap() != NULL, "java heap should be initialized");
+  assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
   sllx(src, LogMinObjAlignmentInBytes, dst);
-  bpr(rc_nz, true, Assembler::pt, dst, done);
-  delayed() -> add(dst, G6_heapbase, dst); // annuled if not taken
-  bind(done);
+  if (Universe::narrow_oop_base() != NULL) {
+    Label done;
+    bpr(rc_nz, true, Assembler::pt, dst, done);
+    delayed() -> add(dst, G6_heapbase, dst); // annuled if not taken
+    bind(done);
+  }
   verify_oop(dst);
 }
 
@@ -4366,8 +4385,11 @@ void  MacroAssembler::decode_heap_oop_not_null(Register r) {
   // pd_code_size_limit.
   // Also do not verify_oop as this is called by verify_oop.
   assert (UseCompressedOops, "must be compressed");
+  assert (Universe::heap() != NULL, "java heap should be initialized");
+  assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
   sllx(r, LogMinObjAlignmentInBytes, r);
-  add(r, G6_heapbase, r);
+  if (Universe::narrow_oop_base() != NULL)
+    add(r, G6_heapbase, r);
 }
 
 void  MacroAssembler::decode_heap_oop_not_null(Register src, Register dst) {
@@ -4375,14 +4397,17 @@ void  MacroAssembler::decode_heap_oop_not_null(Register src, Register dst) {
   // pd_code_size_limit.
   // Also do not verify_oop as this is called by verify_oop.
   assert (UseCompressedOops, "must be compressed");
+  assert (Universe::heap() != NULL, "java heap should be initialized");
+  assert (LogMinObjAlignmentInBytes == Universe::narrow_oop_shift(), "decode alg wrong");
   sllx(src, LogMinObjAlignmentInBytes, dst);
-  add(dst, G6_heapbase, dst);
+  if (Universe::narrow_oop_base() != NULL)
+    add(dst, G6_heapbase, dst);
 }
 
 void MacroAssembler::reinit_heapbase() {
   if (UseCompressedOops) {
     // call indirectly to solve generation ordering problem
-    Address base(G6_heapbase, (address)Universe::heap_base_addr());
+    Address base(G6_heapbase, (address)Universe::narrow_oop_base_addr());
     load_ptr_contents(base, G6_heapbase);
   }
 }
diff --git a/src/cpu/sparc/vm/sparc.ad b/src/cpu/sparc/vm/sparc.ad
index f9631ddf6..03c79b5a2 100644
--- a/src/cpu/sparc/vm/sparc.ad
+++ b/src/cpu/sparc/vm/sparc.ad
@@ -547,7 +547,11 @@ int MachCallDynamicJavaNode::ret_addr_offset() {
     int v_off = entry_offset*wordSize + vtableEntry::method_offset_in_bytes();
     int klass_load_size;
     if (UseCompressedOops) {
-      klass_load_size = 3*BytesPerInstWord; // see MacroAssembler::load_klass()
+      assert(Universe::heap() != NULL, "java heap should be initialized");
+      if (Universe::narrow_oop_base() == NULL)
+        klass_load_size = 2*BytesPerInstWord; // see MacroAssembler::load_klass()
+      else
+        klass_load_size = 3*BytesPerInstWord;
     } else {
       klass_load_size = 1*BytesPerInstWord;
     }
@@ -1601,9 +1605,11 @@ void MachUEPNode::format( PhaseRegAlloc *ra_, outputStream *st ) const {
   st->print_cr("\nUEP:");
 #ifdef    _LP64
   if (UseCompressedOops) {
+    assert(Universe::heap() != NULL, "java heap should be initialized");
     st->print_cr("\tLDUW   [R_O0 + oopDesc::klass_offset_in_bytes],R_G5\t! Inline cache check - compressed klass");
     st->print_cr("\tSLL    R_G5,3,R_G5");
-    st->print_cr("\tADD    R_G5,R_G6_heap_base,R_G5");
+    if (Universe::narrow_oop_base() != NULL)
+      st->print_cr("\tADD    R_G5,R_G6_heap_base,R_G5");
   } else {
     st->print_cr("\tLDX    [R_O0 + oopDesc::klass_offset_in_bytes],R_G5\t! Inline cache check");
   }
@@ -2502,7 +2508,11 @@ encode %{
       __ load_klass(O0, G3_scratch);
       int klass_load_size;
       if (UseCompressedOops) {
-        klass_load_size = 3*BytesPerInstWord;
+        assert(Universe::heap() != NULL, "java heap should be initialized");
+        if (Universe::narrow_oop_base() == NULL)
+          klass_load_size = 2*BytesPerInstWord;
+        else
+          klass_load_size = 3*BytesPerInstWord;
       } else {
         klass_load_size = 1*BytesPerInstWord;
       }
diff --git a/src/cpu/sparc/vm/stubGenerator_sparc.cpp b/src/cpu/sparc/vm/stubGenerator_sparc.cpp
index 9b4981dff..5ef4f23ab 100644
--- a/src/cpu/sparc/vm/stubGenerator_sparc.cpp
+++ b/src/cpu/sparc/vm/stubGenerator_sparc.cpp
@@ -2942,14 +2942,15 @@ class StubGenerator: public StubCodeGenerator {
     StubRoutines::_atomic_add_ptr_entry      = StubRoutines::_atomic_add_entry;
     StubRoutines::_fence_entry               = generate_fence();
 #endif  // COMPILER2 !=> _LP64
-
-    StubRoutines::Sparc::_partial_subtype_check                = generate_partial_subtype_check();
   }
 
 
   void generate_all() {
     // Generates all stubs and initializes the entry points
 
+    // Generate partial_subtype_check first here since its code depends on
+    // UseZeroBaseCompressedOops which is defined after heap initialization.
+    StubRoutines::Sparc::_partial_subtype_check                = generate_partial_subtype_check();
     // These entry points require SharedInfo::stack0 to be set up in non-core builds
     StubRoutines::_throw_AbstractMethodError_entry         = generate_throw_exception("AbstractMethodError throw_exception",          CAST_FROM_FN_PTR(address, SharedRuntime::throw_AbstractMethodError),  false);
     StubRoutines::_throw_IncompatibleClassChangeError_entry= generate_throw_exception("IncompatibleClassChangeError throw_exception", CAST_FROM_FN_PTR(address, SharedRuntime::throw_IncompatibleClassChangeError),  false);
diff --git a/src/cpu/sparc/vm/vm_version_sparc.cpp b/src/cpu/sparc/vm/vm_version_sparc.cpp
index e2866cb43..96d399fc6 100644
--- a/src/cpu/sparc/vm/vm_version_sparc.cpp
+++ b/src/cpu/sparc/vm/vm_version_sparc.cpp
@@ -72,6 +72,9 @@ void VM_Version::initialize() {
         FLAG_SET_ERGO(bool, UseCompressedOops, false);
       }
     }
+    // 32-bit oops don't make sense for the 64-bit VM on sparc
+    // since the 32-bit VM has the same registers and smaller objects.
+    Universe::set_narrow_oop_shift(LogMinObjAlignmentInBytes);
 #endif // _LP64
 #ifdef COMPILER2
     // Indirect branch is the same cost as direct
diff --git a/src/cpu/sparc/vm/vtableStubs_sparc.cpp b/src/cpu/sparc/vm/vtableStubs_sparc.cpp
index df9262ccb..ce2c8532d 100644
--- a/src/cpu/sparc/vm/vtableStubs_sparc.cpp
+++ b/src/cpu/sparc/vm/vtableStubs_sparc.cpp
@@ -221,13 +221,15 @@ int VtableStub::pd_code_size_limit(bool is_vtable_stub) {
     if (is_vtable_stub) {
       // ld;ld;ld,jmp,nop
       const int basic = 5*BytesPerInstWord +
-                        // shift;add for load_klass
-                        (UseCompressedOops ? 2*BytesPerInstWord : 0);
+                        // shift;add for load_klass (only shift with zero heap based)
+                        (UseCompressedOops ?
+                         ((Universe::narrow_oop_base() == NULL) ? BytesPerInstWord : 2*BytesPerInstWord) : 0);
       return basic + slop;
     } else {
       const int basic = (28 LP64_ONLY(+ 6)) * BytesPerInstWord +
-                        // shift;add for load_klass
-                        (UseCompressedOops ? 2*BytesPerInstWord : 0);
+                        // shift;add for load_klass (only shift with zero heap based)
+                        (UseCompressedOops ?
+                         ((Universe::narrow_oop_base() == NULL) ? BytesPerInstWord : 2*BytesPerInstWord) : 0);
       return (basic + slop);
     }
   }