src/cpu/x86/vm/c1_LIRAssembler_x86.cpp

rev 8802 : G1 performance improvements: card batching, joining, sorting, prefetching and write barrier fence elision and simplification based on a global syncrhonization using handshakes piggybacking on thread-local safepoints.
rev 8804 : Further JPRT improvements

@@ -24,10 +24,11 @@
 
 #include "precompiled.hpp"
 #include "asm/macroAssembler.hpp"
 #include "asm/macroAssembler.inline.hpp"
 #include "c1/c1_Compilation.hpp"
+#include "c1/c1_CodeStubs.hpp"
 #include "c1/c1_LIRAssembler.hpp"
 #include "c1/c1_MacroAssembler.hpp"
 #include "c1/c1_Runtime1.hpp"
 #include "c1/c1_ValueStack.hpp"
 #include "ci/ciArrayKlass.hpp"

@@ -511,11 +512,11 @@
 }
 
 
 // This is the fast version of java.lang.String.compare; it has not
 // OSR-entry and therefore, we generate a slow version for OSR's
-void LIR_Assembler::emit_string_compare(LIR_Opr arg0, LIR_Opr arg1, LIR_Opr dst, CodeEmitInfo* info) {
+void LIR_Assembler::emit_string_compare(LIR_Opr arg0, LIR_Opr arg1, LIR_Opr dst, C1ThreadLocalSafepoint *tls_stub, CodeEmitInfo* info) {
   __ movptr (rbx, rcx); // receiver is in rcx
   __ movptr (rax, arg1->as_register());
 
   // Get addresses of first characters from both Strings
   __ load_heap_oop(rsi, Address(rax, java_lang_String::value_offset_in_bytes()));

@@ -581,19 +582,19 @@
 
   // strings are equal up to min length
 
   __ bind(noLoop);
   __ pop(rax);
-  return_op(LIR_OprFact::illegalOpr);
+  return_op(LIR_OprFact::illegalOpr, tls_stub);
 
   __ bind(haveResult);
   // leave instruction is going to discard the TOS value
   __ mov (rax, rcx); // result of call is in rax,
 }
 
 
-void LIR_Assembler::return_op(LIR_Opr result) {
+void LIR_Assembler::return_op(LIR_Opr result, C1ThreadLocalSafepoint *code_stub) {
   assert(result->is_illegal() || !result->is_single_cpu() || result->as_register() == rax, "word returns are in rax,");
   if (!result->is_illegal() && result->is_float_kind() && !result->is_xmm_register()) {
     assert(result->fpu() == 0, "result must already be on TOS");
   }
 

@@ -602,37 +603,62 @@
 
   bool result_is_oop = result->is_valid() ? result->is_oop() : false;
 
   // Note: we do not need to round double result; float result has the right precision
   // the poll sets the condition code, but no data registers
+
+  if (!ThreadLocalSafepoints) {
   AddressLiteral polling_page(os::get_polling_page(), relocInfo::poll_return_type);
 
   if (Assembler::is_polling_page_far()) {
     __ lea(rscratch1, polling_page);
     __ relocate(relocInfo::poll_return_type);
     __ testl(rax, Address(rscratch1, 0));
   } else {
     __ testl(rax, polling_page);
   }
+  } else {
+#ifdef _LP64
+    code_stub->set_safepoint_pc(__ pc());
+    __ relocate(relocInfo::poll_return_type);
+    __ testb(Address(r15_thread, Thread::yieldpoint_offset()), 2);
+    __ jcc(Assembler::equal, *code_stub->entry());
+#else
+    ShouldNotReachHere();
+#endif
+  }
   __ ret(0);
 }
 
 
-int LIR_Assembler::safepoint_poll(LIR_Opr tmp, CodeEmitInfo* info) {
+int LIR_Assembler::safepoint_poll(LIR_Opr tmp, C1ThreadLocalSafepoint *code_stub, CodeEmitInfo* info) {
   AddressLiteral polling_page(os::get_polling_page(), relocInfo::poll_type);
   guarantee(info != NULL, "Shouldn't be NULL");
   int offset = __ offset();
+  if (!ThreadLocalSafepoints) {
   if (Assembler::is_polling_page_far()) {
     __ lea(rscratch1, polling_page);
     offset = __ offset();
     add_debug_info_for_branch(info);
     __ relocate(relocInfo::poll_type);
     __ testl(rax, Address(rscratch1, 0));
   } else {
     add_debug_info_for_branch(info);
     __ testl(rax, polling_page);
   }
+  } else {
+#ifdef _LP64
+    add_debug_info_for_branch(info);
+    code_stub->set_safepoint_pc(__ pc());
+    __ relocate(relocInfo::poll_type);
+    __ testb(Address(r15_thread, Thread::yieldpoint_offset()), 1);
+    __ jcc(Assembler::equal, *code_stub->entry());
+#else
+    ShouldNotReachHere();
+#endif
+  }
+
   return offset;
 }
 
 
 void LIR_Assembler::move_regs(Register from_reg, Register to_reg) {